# Decision and Authorization Latency in NASA Programs: A Cliometric Analysis of Program Cost, Schedule, and Mission Cadence, 1958 to 2026

**A Doctoral Dissertation**

**Candidate:** PHD-08
**Program:** COLLEGIUM 1st Battalion
**Category (NORTH STAR / JPL):** Mission Program Execution Management
**Method:** Quantitative cliometric time-series and program-phase panel regression with program and era fixed effects
**Status:** Design-stage. Every estimate is expected or illustrative, never executed. Every citation is real, with a clickable DOI or resolvable URL.

**Hall-of-Shoulders anchors:** Angus Maddison (consistent constant-unit measurement as the precondition of comparison); Robert Fogel (the explicit, bounded counterfactual); Douglass North (institutions as transaction-cost structures, with adaptive efficiency and path dependence). Adjacent anchors carried for depth: Oliver Williamson, Elinor Ostrom, and the staggered-treatment econometrics frontier (Goodman-Bacon; Callaway and Sant'Anna; de Chaisemartin and D'Haultfoeuille; Imai and Kim).

**Date:** 2026-06-15


## Abstract

The stewardship of a national space program rests, in the end, on the care with which public resources are committed and the speed with which sound decisions are reached. This dissertation asks whether the time the National Aeronautics and Space Administration takes to decide and authorize program actions is a measurable variable across its history, and whether longer latency is associated with worse program performance. The motivating observation is widely shared but rarely measured: NASA programs frequently exceed their cost and schedule baselines, and practitioners often attribute part of this to slow internal decision and authorization processes rather than to engineering difficulty alone. The NASA cost-and-schedule literature is mature on the technical and estimating side, and the public-administration literature has measured administrative burden and its association with organizational performance, but the two have never been joined in a long-run quantitative study of one agency.

The study constructs a program-phase panel covering NASA programs from 1958 to 2026, defines and measures administrative decision-and-authorization latency from documentary records, and proposes to estimate its association with three outcomes: cost growth relative to baseline, schedule slip relative to baseline, and realized mission cadence. The single falsifiable contribution is stated as a null hypothesis that authorization latency has no association with cost, schedule, or cadence, against the alternative that longer latency is associated with greater cost growth, more schedule slip, and lower cadence. Identification uses program and era fixed effects to remove time-invariant program characteristics and common shocks, relies on within-program and within-era variation in latency, and addresses the residual endogeneity of latency to phase-specific difficulty with an instrumental-variable strategy using authorizing-office workload and appropriations-calendar timing. Where discrete administrative regimes are the object of inference, heterogeneity-robust staggered-difference-in-differences estimators replace the naive two-way fixed-effects estimator, with a Goodman-Bacon decomposition reported as a diagnostic.

The dissertation reports the full research design, the data construction, and a pre-registered analysis plan with a binding decision rule. Consistent with honest scientific practice, it does not report estimates from the full dataset as if executed; expected signs and an illustrative specification are clearly labeled as design-stage, and the illustrative coefficient table is left unpopulated by design. The first contribution stands regardless of the regression result: a consistent, documentary-rule-based series of NASA authorization latency does not currently exist, and constructing it is a prerequisite for anyone who wishes to argue about whether the institution's own speed is part of the cost-and-schedule problem. The methodological discipline follows the quantitative economic history tradition: state the proposition quantitatively, build the measures from primary records, bound the estimate, and let the data falsify the hypothesis.


## Table of Contents

- Abstract
- List of Tables
- Chapter 1. Introduction
- Chapter 2. Theoretical Framework
- Chapter 3. Literature Review
- Chapter 4. Data and Measurement
- Chapter 5. Research Design and Identification
- Chapter 6. Analysis Plan and Expected Results
- Chapter 7. Discussion
- Chapter 8. Conclusion
- References
- Appendix A. Variable-Construction Codebook
- Appendix B. Program-Phase Inclusion List and Coverage Table
- Appendix C. Era-Regime Definition Table
- Appendix D. Robustness-Battery Specification List
- Appendix E. Pre-Registration Record


## List of Tables

- Table 3.1. Synthesis map of the field: branch, contribution, what each leaves open, and representative corpus evidence (Chapter 3, Section 3.6.1).
- Table 4.1. Measurement table: operational definition, primary source, and scale for every variable in the panel (Chapter 4, Section 4.8).
- Table 6.1. Illustrative coefficient table, deliberately unpopulated by design (Chapter 6, Section 6.7).
- Table A.1. Variable dictionary (Appendix A).
- Table B.1. Coverage by era and resolution (Appendix B).
- Table C.1. Era-regime mapping (Appendix C).
- Table D.1. Pre-specified robustness checks (Appendix D).

No figures are included; the dissertation is a design-stage document and its illustrative coefficient table is left unpopulated by design.


## Chapter 1. Introduction

## 1.1 The chapter thesis, stated first

An agency entrusted with the public's exploration of space owes its sponsors a clear account of where its time and money go. This dissertation makes one claim in service of that account, and designs one test of it. The claim is that the time the National Aeronautics and Space Administration takes to decide and to authorize program actions, what this study names decision-and-authorization latency, is a measurable variable across the agency's history from 1958 to 2026, and that this variable can be entered into a falsifiable test of whether longer latency is associated with worse program execution outcomes. The test is a program-phase panel regression, with program and era fixed effects and an instrumental-variable defense, of latency against three execution outcomes: cost growth relative to baseline, schedule slip relative to baseline, and realized mission cadence. The single falsifiable contribution is the null hypothesis that latency has no association with these outcomes, against the alternative that longer latency raises cost growth and schedule slip and lowers cadence. The methodological wager is that a problem practitioners have asserted for decades, namely that NASA programs overrun in part because the institution is slow to act, has never been measured as a quantity, and that measuring it is worth doing whether or not the regression ultimately rejects the null.

The order of that paragraph is deliberate, and it is the order this chapter and the dissertation follow throughout. The answer comes first; the development comes after. A reader should be able to state the contribution of this work after the first paragraph, then spend the remaining pages learning why the contribution is well-posed, how it is identified, what it requires of the data, and what would falsify it. The dissertation is written in the design stage. It reports no executed estimates from the full dataset. Where directional expectations or illustrative output formats appear, they are labeled as expectations or illustrations and must not be read as findings. This is not a hedge. It is the central discipline of quantitative economic history, the tradition this study works inside: state the proposition as a quantity, build the measure from primary records, bound the estimate, and let the data decide. The contribution that survives regardless of the regression result is the measure itself, because a consistent long-run series of NASA decision-and-authorization latency does not currently exist, and constructing one is a prerequisite for anyone, inside or outside the agency, who wishes to argue about whether the institution's own speed is part of the cost-and-schedule problem.

The remainder of this chapter develops that thesis in the order a falsifiable contribution should be developed. Section 1.2 frames the problem as a current state, a desired state, a gap, and a consequence of leaving the gap unaddressed. Section 1.3 supplies the institutional and historical context that makes the problem specific to NASA rather than generic to government. Section 1.4 breaks the research question into its explicit parts. Section 1.5 states the single falsifiable contribution as competing hypotheses, reproduced verbatim from the approved proposal so that no later chapter can quietly reword the contribution. Section 1.6 establishes significance for NASA, for the Jet Propulsion Laboratory, and for the named stakeholders in program execution management. Section 1.7 fixes the scope and the delimitations, including an explicit decision about what this dissertation is not. Section 1.8 defines the key terms with the precision the later measurement chapters require. Section 1.9 provides a roadmap of the eight chapters and the back matter.

## 1.2 The problem in full

### 1.2.1 Current state

NASA programs have a long and well-documented record of exceeding the cost and the schedule estimates set at their formal commitment points. This is not a contested premise; it is the settled finding of the agency's own cost-estimating community. The foundational work in that community used historical NASA cost and schedule growth to set the reserve guidelines that programs are required to carry, and it did so precisely because growth is large enough and regular enough to be a planning parameter rather than an exception [\[1\]](#ref-1). The same community has built and maintained datasets of historical mass, power, schedule, and cost growth for NASA science instruments and spacecraft, finding consistent positive growth between early formulation and launch [\[1\]](#ref-1). The National Research Council, convened to examine the problem at the agency level, produced a consensus study cataloging the management, technical, and budgetary causes of cost growth on NASA Earth and space science missions, which is itself evidence that the pattern is durable enough to warrant a National Academies study [\[6\]](#ref-6). Independent financial analysis of the contractor side of the problem reports that, in one recent accounting, fifteen NASA development-phase projects accounted for roughly twelve billion dollars in cost overruns and a cumulative twenty-eight years of schedule delay [\[34\]](#ref-34). A recurring strand of agency self-assessment goes further and asks whether NASA's planning models, applied honestly, would ever have approved the programs the agency actually flew, given that estimated costs became expected budgets and real budgets rarely matched expectations [\[123\]](#ref-123).

That is the current state in the plainest terms available: the empirical regularity of cost growth and schedule slip is established, measured, and institutionalized into the agency's reserve policy. The evidence is the cost-estimating datasets and the National Academies study, and the finding carries weight because it is reproduced across the agency's own cost community, an external academies panel, and independent financial analysis, so it is not an artifact of one method or one author. These sources measure growth, not its causes, and where they assign causes they do so chiefly to technical and contractual parameters. A careful reader might object that growth is an estimating problem rather than an execution problem; that distinction is exactly the one this dissertation is built to sharpen, and it is addressed below.
### 1.2.2 Desired state

The desired state has two components, one descriptive and one inferential. Descriptively, the field should possess a consistent, constant-unit series of NASA decision-and-authorization latency, built by a single documentary rule applied identically from 1958 to 2026, so that the agency's speed of acting can be set beside its cost growth and schedule slip on the same footing across eras. Inferentially, the field should possess a pre-registered, heterogeneity-robust panel test of whether that latency variable is associated with the three execution outcomes net of program-specific difficulty and era-specific conditions. Both components are absent today. The descriptive series does not exist. The inferential test has not been posed for NASA as a single agency across its full history. The desired state, in short, is to answer the question "does the institution's own decision speed move its program outcomes" with a number and an interval rather than an anecdote.

### 1.2.3 The gap

The gap is structural, and it sits between two mature literatures that have never been joined for this purpose. The first is the NASA cost-and-schedule literature, which is technically sophisticated and models cost growth as a function of technical parameters such as instrument mass and power, technology readiness, mission class, and contract type [\[1\]](#ref-1). This literature treats schedule as an outcome and sometimes as a driver of cost, and it has examined how policy changes move cost and schedule growth, which is direct evidence that administrative regime and not only engineering moves the outcomes [\[4\]](#ref-4). What it does not do is isolate administrative decision-and-authorization latency as a measured explanatory variable, and it does not assemble a single consistent panel spanning the agency's full history. The second literature is the public-administration tradition on red tape and administrative burden, which has developed validated constructs and measures for the procedural rules and clearances that organizations impose and has related them to organizational performance [\[7\]](#ref-7), [\[8\]](#ref-8). This literature establishes that administrative process is measurable and that it matters for outcomes in general, but it has not been applied to NASA as a single long-run case with program-level panel data. The gap, stated precisely, is the absence of any study that measures administrative latency inside NASA from documentary records, assembles it into a program panel covering 1958 to 2026, and tests whether it is associated with cost growth, schedule slip, and cadence. The two literatures themselves establish the gap: a variable named in the narrative of one literature and measured in the methods of the other has never been carried across the bridge between them for this agency.

### 1.2.4 Consequence of inaction

The consequence of leaving the gap open is concrete and ongoing. The practitioner community continues to attribute a portion of cost growth and schedule slip to slow decision-making, and the agency's own program framing makes the attribution explicit, describing the interval between sensing a condition, understanding its implications, authorizing a response, and executing that response as a binding constraint on program execution, with decision speed ranging from autonomous systems acting in milliseconds to administrative and congressional authorizations that take months. That practitioner framing is treated here as the narrative this dissertation tests, not as a measurement. Two examples are cited as evidence for the attribution. The Constellation Program, NASA's effort to replace the Space Shuttle and return humans to the Moon, was managed against an explicit life-cycle cost model because affordability was a first-order policy concern, and it consumed on the order of nine billion dollars over roughly five years and flew no operational missions before its cancellation; the cost figure is drawn from agency program-record material and the program's own Life-cycle Cost Analysis Model documentation rather than from a peer-reviewed estimate, and it is treated here as program-record-derived [\[102\]](#ref-102). The interval between the first and second crewed flights of the Space Launch System spanned roughly forty-one months, during which an upgraded upper stage and a second mobile launcher consumed resources and were subsequently terminated; this interval is taken from program records and agency framing and is likewise marked as program-record-derived, not as a peer-reviewed citation. The point of naming these examples is not to lean on them. It is the opposite. They are vivid anecdotes, not measurements, and the consequence of inaction is that the agency keeps reasoning from anecdotes of this kind without being able to say whether the pattern they suggest survives measurement. Until latency is measured, NASA cannot tell whether investing in faster, lower-layer authorization is a better or worse lever on cost and schedule than investing in technical risk reduction, which is the lever the existing literature already emphasizes. Confidence in this consequence claim is high, because both the attribution narrative and the inability to test it are directly observable in the agency's own framing and in the literature's silence on a measured latency variable.

## 1.3 Institutional and historical context

The problem is general to large public undertakings, but its specific form is institutional, and the institution has a history that the measurement must respect. The theoretical reason to expect the institution's own process to matter, and to matter durably, comes from Douglass North's distinction between institutions, the rules of the game, and organizations, the players who operate within them, and from his location of economic performance in the transaction costs that institutions raise or lower [\[18\]](#ref-18). Decision-and-authorization latency is, in North's vocabulary, a transaction cost internal to a public organization: the elapsed time cost of measuring a proposed action against the rules and of obtaining authorization to proceed. North's further concepts of adaptive efficiency and path dependence supply the historical mechanism and the reason for treating eras as distinct. Once a set of review and authorization rules is in place, it generates increasing returns; the organization and its contractors adapt to it, the cost of changing it rises, and the agency can lock onto a persistent and even inefficient decision trajectory [\[18\]](#ref-18). This is the theoretical reason latency might be both large and durable across NASA's history, and it is the reason era fixed effects are not a statistical convenience but a substantive requirement: rule regimes are real features of the historical record, not noise.

The history itself spans regimes that any honest measure must hold apart. The agency that authorized Apollo against an open-ended national commitment is not, institutionally, the agency that authorized the Space Shuttle against a cost-recovery promise, which is not the agency that authorized the Constellation Program against an explicit affordability constraint and a parametric life-cycle cost model [\[102\]](#ref-102), which is not the agency that today authorizes the Space Launch System and the Artemis campaign under continuing scrutiny of procurement strategy [\[34\]](#ref-34). The management-control practices of the Apollo era have themselves been studied as a distinct regime with lessons for later programs, which shows that the rules of decision and authorization were specific to their time and not constant across the agency's life [\[25\]](#ref-25). Layered on top of these program-era regimes is the federal budget and appropriations cycle, an annual rhythm that paces authorization for reasons that have nothing to do with the engineering difficulty of any particular program and everything to do with the calendar of congressional action [\[129\]](#ref-129). The appropriations cycle matters twice over to this study: it is part of the institutional context that makes latency durable, and, as Chapter 5 develops, its timing is a candidate source of variation in latency that is plausibly independent of a program's technical difficulty, which makes it useful for identification.

The broader economic context sharpens why the question is worth posing now. NASA operates under a long secular decline in its share of the federal budget and within a space sector whose economics have shifted toward commercial actors and platform strategies, a shift that economists have begun to analyze as a genuine reordering of who bears cost and risk in space activity [\[32\]](#ref-32). Comparative work on bespoke versus platform strategies, using a reference-class dataset of 203 space missions from 1963 to 2021, finds that repeatable platform approaches were markedly cheaper and faster than one-off bespoke approaches, with NASA's bespoke history serving as one pole of the comparison [\[59\]](#ref-59). That comparison is corroborative context, not this study's claim; this study does not test bespoke against platform. But it locates the present work in a moment when the agency's cost and schedule performance is under both economic and political pressure, and when the question of whether the agency's own decision processes are part of the performance problem is no longer academic. The institutional context, in sum, gives the study a theory of why latency should be expected to matter and to persist, a periodization of regimes the measure must respect, and a contemporary stake in the answer.

## 1.4 The research questions, broken out

The dissertation is organized around one research question with three explicit parts. Stating them separately prevents the common failure in which a single omnibus question is answered for one of its components and asserted for the others.

The overarching question, fixed verbatim from the shared design bible, is: Is the time NASA takes to decide and authorize program actions a measurable variable across the agency's history from 1958 to 2026, and is longer decision-and-authorization latency associated with worse program performance, measured as cost growth, schedule slip, and mission cadence?

This decomposes into the following explicit questions.

First, the measurement question. Can decision-and-authorization latency be constructed for NASA programs from documentary records, by a single rule applied identically across eras, at a resolution coarse enough to span the full 1958-to-2026 history and fine enough, in the modern subperiod, to resolve individual decision events? This question is logically prior to the others. If latency cannot be measured consistently, the inferential questions cannot be posed at all, and the answer to this first question is the contribution that stands regardless of the regression result.

Second, the cost-growth question. Net of program fixed effects, era fixed effects, and technical and programmatic controls, is longer authorization latency associated with greater cost growth, measured as the actual phase cost minus the baseline phase cost divided by the baseline, in constant fiscal-year dollars?

Third, the schedule-slip question. Net of the same controls and fixed effects, is longer authorization latency associated with greater schedule slip, measured as the actual phase duration minus the baseline phase duration divided by the baseline duration?

Fourth, the cadence question. At the era and program-family level, is higher latency per decision associated with lower realized mission cadence, measured as operational mission events per unit time within a defined program family and as the interval between successive flight or delivery events?

Fifth, the identification question, which is not separable from the four above but must be posed explicitly because the central threat is endogeneity. Can the association between latency and the outcomes be identified as something other than a reflection of program difficulty, given that harder programs may both take longer to authorize and overrun more? This question is answered by the design rather than by the data: program fixed effects, era fixed effects, an instrumental-variable strategy in the spirit of the procurement-competence literature [\[12\]](#ref-12), and the use of latency measured early in each phase. The dissertation reports both the fixed-effects and the instrumented estimates so that a reader can see exactly how much any conclusion depends on the instrument.

The five questions share one structure. Each is a quantity to be measured or a coefficient to be estimated, each carries a stated control set and fixed-effects structure, and each is falsifiable. None is rhetorical.

## 1.5 The single falsifiable contribution

The contribution is one testable proposition, stated as competing hypotheses. These are reproduced verbatim from the approved proposal and the design bible. No chapter of this dissertation may reword them.

**H0 (null).** Administrative decision-and-authorization latency in NASA programs has no association with program cost growth, schedule slip, or mission cadence, after accounting for program and era fixed effects and technical controls.

**H1 (alternative).** Longer administrative decision-and-authorization latency is associated with greater program cost growth, more schedule slip, and lower mission cadence, after accounting for program and era fixed effects and technical controls.

Under the notation fixed for the whole dissertation, the baseline specification for program i, phase p, and era t is
\[
\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t) \qquad\qquad (1)
\]

where \(\alpha_i\) are program fixed effects, \(\delta_t\) are era fixed effects, \(X\) is the vector of technical and programmatic controls, \(\varepsilon\) is the error clustered by program, and \(\beta\) is the coefficient of interest. Under H1, \(\beta\) is positive when the outcome is cost growth or schedule slip and negative when the outcome is mission cadence. This sign convention is restated in every chapter that touches the specification, so that the direction of the predicted effect is never left implicit.

The proposition is falsifiable in the strict sense, and the conditions of its falsification are fixed in advance rather than left to discretion. The contribution fails if the latency coefficient is statistically indistinguishable from zero in the preferred specification; if its sign reverses between the fixed-effects and the instrumented specifications; if it vanishes under the heterogeneity-robust estimators or under the alternative latency resolution; or if the apparent association is fully explained by baseline gaming or by reverse causation. This pre-registered decision rule is binding. The study is designed so that the null is a real possibility and not a straw figure, because latency is plausibly endogenous to program difficulty. In that case a raw positive correlation would reflect difficulty rather than process, and the design must defeat that reading before any rejection of the null can be claimed. This is a single falsifiable contribution rather than a collection of loosely related findings, because all five research questions resolve into the sign and significance of one coefficient, \(\beta\), estimated for three outcomes under one decision rule. Either the data move \(\beta\) away from zero in the predicted direction and the rule rejects H0, or they do not and the rule retains it. Both outcomes are informative, which is the mark of a well-posed test.

Calibrated confidence, and a clear account of what would change it, is part of the standard this dissertation holds itself to. At the design stage, the appropriate confidence in any directional claim about \(\beta\) is low, because no estimate from the full panel has been produced. Confidence rises with the construction and validation of the latency series against published cost-growth figures for overlapping programs, the demonstration that an instrument for latency is both relevant and defensible, and the stability of the coefficient across the robustness battery. Confidence falls with failure of the instrument's first stage, sensitivity of the sign to the latency resolution, or evidence that baselines were set in anticipation of slow authorization. The dissertation reports each of these checks rather than asserting the conclusion.

## 1.6 Significance for NASA, JPL, and the named stakeholders

The significance of the study follows a single causal logic, and stating that logic as a mechanism rather than as a correlation is required by the argument standard of this work. The driver is the agency's fragmented decision authority and its sequential, multi-layer authorization gates, paced by an annual appropriations cycle [\[129\]](#ref-129). The mechanism is that each pending action accrues elapsed time, the latency this study measures, during which standing program costs continue to be incurred and program requirements drift. The observable effect is that measured latency per program-phase rises. The operational consequence is higher cost growth and schedule slip and fewer flight or delivery events per period, which is lower cadence. The strategic implication is that a portion of NASA's cost and schedule performance would then be a controllable process variable, distinct from the irreducible difficulty of spaceflight, and therefore a lever available to program execution management. This is the chain that gives the study its stakes, and it is stated as a chain, driver to mechanism to observable effect to operational consequence to strategic implication, precisely so that a reader can attack any link. Where the dissertation can establish only correlation rather than the full causal chain, it downgrades the confidence of the claim and falls back on the instrumental-variable and early-in-phase-latency design as the response; it does not present correlation dressed as cause.

For NASA as an agency, the significance under the alternative is that compressing authorization latency would be a lever on cost and schedule that is independent of technology investment and of contract structure, the two levers the cost-growth literature already emphasizes [\[1\]](#ref-1), [\[4\]](#ref-4). Under the null, the significance is equally real and is symmetric: a null result would discipline the practitioner narrative that attributes overruns to slow decisions, and would direct reform effort toward the technical and estimating factors the existing literature documents rather than toward process speed [\[6\]](#ref-6). The study is constructed so that either result is a usable finding for the agency.

For the Jet Propulsion Laboratory, the significance is sharper because of the structure of its portfolio. JPL's programs are concentrated in deep-space and planetary missions with long development cycles, and the systems-engineering record of NASA deep-space missions shows that cost overruns and schedule impacts often surface late, as detailed knowledge of advanced and heritage systems accumulates over years [\[42\]](#ref-42). In a long-cycle program, authorization events accumulate over many years, and if latency per decision carries a cost-and-schedule penalty, that penalty compounds across the life of a planetary mission in a way it does not for a short program. The implication for JPL management, conditional on the alternative holding, is that investment in faster, lower-layer authorization may return cost and schedule savings comparable to investment in technical risk reduction, and that the return would be largest precisely on the long-cycle missions JPL specializes in.

For the named stakeholders in program execution management, which is the NORTH STAR and JPL category this dissertation serves, the significance is that the study would give the red-tape and administrative-burden constructs a concrete, dollar-and-month-denominated instantiation inside a single agency over a long horizon [\[7\]](#ref-7), [\[8\]](#ref-8). Procurement-competence research using federal contract data has already shown, with an instrumental-variable design, that the administrative side of program execution causally affects delays and cost overruns [\[12\]](#ref-12), and related work finds that ideological misalignment between political appointees and career procurement officers raises cost overruns and delays through a morale mechanism [\[53\]](#ref-53). This study would extend that body of evidence from the contract level to the program level and from a cross-section of agencies to one agency's full history. The qualifier carried throughout is that these stakes are conditional: they are what the finding would mean if the test rejects the null in the predicted direction, and they are stated as conditionals, not as realized results.

## 1.7 Scope and delimitations

The scope of the dissertation is bounded deliberately, and several boundaries are worth stating explicitly so that the study is not read as claiming more than it tests.

The study is confined to NASA. Its findings, whichever way they fall, pertain to the National Aeronautics and Space Administration and may not generalize to other agencies or to commercial programs with different authorization structures. The procurement-competence evidence from broader federal contracting [\[12\]](#ref-12) is used as corroboration of the mechanism's plausibility, not as a basis for statistical generalization from NASA to government at large.

The study is confined to the period from the agency's inception in 1958 through 2026. Coverage is dense and high-resolution for the period covered by independent major-project assessments and by the published cost-growth datasets, and it is sparser and coarser for the earliest decades, where baselines were defined less formally. The panel is therefore unbalanced by construction, and the two-resolution latency design, a coarse milestone-to-milestone measure for the full span and a fine key-decision-point measure for the modern subperiod, is the response to changing documentary density rather than an afterthought.

The study is a design-stage proposal and analysis plan. It does not report estimates from the full assembled dataset. Directional expectations are stated so that the test is interpretable, and an illustrative output table appears in the analysis-plan chapter with its cells deliberately left unpopulated. Reporting fabricated coefficients as if they were real would violate the falsifiability standard the dissertation is built on, and the refusal to populate the illustrative table is a methodological choice, not an omission.

The study is delimited away from enterprise-architecture modeling. This is a cliometric, econometric study. Its objects are programs, phases, documentary decision events, and panel estimates. It is not the design or delivery of a capability, a system, a data or service exchange, or an enterprise architecture to be fielded. The dissertation does not construct a strategic-objective-to-capability-to-operational-activity-to-system-function-to-data-exchange traceability, and it does not adopt the vocabulary of architecture frameworks. The single architecture-adjacent statement the study permits itself is the institutional-design implication discussed in the concluding chapters, that faster, lower-layer authorization may serve as a program-execution-management lever, and that statement is expressed in management terms, not in architecture-layer terms.

Three further delimitations bound the constructs. Authorization latency is operationalized from documentary records and may not capture the full administrative-time concept; the two-resolution measure and the sensitivity analyses test whether the operationalization drives the results rather than assuming it does not. Mission cadence is the most definition-sensitive of the three outcomes, depending on how a program family is bounded, and is therefore reported under multiple family definitions rather than under a single one asserted to be correct. Cost baselines are themselves chosen by the agency and may be set optimistically, a selection problem the megaproject literature emphasizes [\[13\]](#ref-13); the study tests for baseline gaming rather than ignoring it.

## 1.8 Definitions of key terms

The following definitions are fixed for the whole dissertation and are stated here at the precision the measurement and design chapters require.

**Decision-and-authorization latency** (the explanatory variable, often shortened to latency). The elapsed time, in months, between a documented trigger event, the point at which a decision or authorization becomes due, and the documented authorization event that resolves it. For each program-phase, latency is taken as the median elapsed time across the authorization events in that phase. It is constructed by a single documentary rule applied identically across all eras. Because documentary detail differs across the agency's history, latency is built at two resolutions: a coarse measure, defined from milestone-to-milestone intervals and available for the full 1958-to-2026 span, and a fine measure, defined from individual key-decision-point records and available only for the modern subperiod. Results are reported separately at each resolution.

**Cost growth** (outcome 1). The actual phase cost minus the baseline phase cost, divided by the baseline, with both expressed in constant fiscal-year dollars using the NASA New Start Inflation Index. Deflation to constant dollars is required before any cross-era comparison and is stated wherever such comparison appears.

**Schedule slip** (outcome 2). The actual phase duration minus the baseline phase duration, divided by the baseline duration, measured in months.

**Mission cadence** (outcome 3). At the era level, the number of operational mission events per unit time within a defined program family; at the program level, the interval between successive flight or delivery events. Reported under multiple family definitions because no single family definition is uniquely correct.

**Program-phase observation** (the unit of analysis). One observation per NASA program per major lifecycle phase for which a baseline and an actual are documented, so that a program with formulation, development, and operations phases contributes multiple observations. The panel dimension is the program, indexed i; the time dimension is the phase sequence p and the era t.

**Program fixed effects** (\(\alpha_i\)). Terms that absorb all time-invariant characteristics of a program, including unobserved difficulty that does not change across the program's phases.

**Era fixed effects** (\(\delta_t\)). Terms that absorb common shocks and rule regimes affecting all programs in a given period, including agency reorganizations, appropriations-cycle conditions, and administrative-rule changes. Their inclusion is justified by North's path-dependence argument that authorization regimes are durable historical features [\[18\]](#ref-18).
**Controls** (the vector X). Technical controls following the cost-estimating literature, namely instrument or spacecraft mass and power, technology readiness level at commitment, mission class, and contract type [\[1\]](#ref-1); and programmatic controls, namely the number of external partners and a funding-instability index constructed from year-over-year deviations between requested and appropriated funds [\[129\]](#ref-129).

**Cliometric.** Pertaining to the application of explicit, quantitative, hypothesis-testing methods to the historical record. The term carries three commitments invoked identically throughout the dissertation: consistent constant-unit measurement before any cross-era comparison, after Maddison [\[16\]](#ref-16); the bounded counterfactual estimated as a range rather than a point, after Fogel's tradition of social-saving estimation [\[17\]](#ref-17); and the treatment of administrative process as an institutional transaction cost, after North [\[18\]](#ref-18).

## 1.9 Roadmap of the dissertation

The dissertation proceeds in eight chapters and a back matter. Its structure mirrors the order of this introduction: contribution first, then the apparatus that makes the contribution defensible.

Chapter 2 builds the theoretical framework. It develops the cliometric lens through its three commitments: Maddison's requirement of consistent constant-unit measurement as the precondition of any cross-era comparison [\[16\]](#ref-16); Fogel's insistence on naming and costing the next-best alternative and reporting a bounded range rather than a single causal point [\[17\]](#ref-17); and North's institutional theory, in which latency is the internal transaction cost of measuring an action against the rules and obtaining authorization, with adaptive efficiency and path dependence explaining why latency is durable and why eras must be held apart [\[18\]](#ref-18). The chapter closes by stating the latency-to-outcome relationship as a transaction-cost mechanism.

Chapter 3 reviews the two literatures and shows why neither closes the question. It treats the NASA cost-and-schedule tradition and what it models and omits [\[1\]](#ref-1), [\[4\]](#ref-4); the public-administration literature on red tape, administrative burden, and validated rule measures, with the nuance that effects are real but conditional [\[7\]](#ref-7), [\[8\]](#ref-8); the procurement-competence evidence as the nearest external precedent for a causal administrative effect [\[12\]](#ref-12); the megaproject and optimism-bias literature on systematic, fat-tailed overrun and its decision-process causes [\[13\]](#ref-13); and the NASA program-management and management-control literature, including the Apollo lessons [\[25\]](#ref-25). It ends with a synthesis that locates the precise unfilled gap.

Chapter 4 specifies the data and the measurement. It names the documented sources, defines the program-phase unit and the unbalanced panel, and lays out the single documentary rule for constructing latency at both resolutions, the construction of the three outcomes including constant-dollar deflation, the control set, the 1958-to-2026 coverage with its changing documentary baseline, and the data limitations.

Chapter 5 develops the research design and identification. It defines the estimator, restates the specification in the fixed notation, and addresses the staggered-regime concern by using heterogeneity-robust estimators rather than a naive two-way fixed-effects difference-in-differences for any discrete-regime analysis, with a Goodman-Bacon decomposition reported as a diagnostic [\[21\]](#ref-21), [\[22\]](#ref-22), [\[23\]](#ref-23), [\[24\]](#ref-24). It then sets out the instrumental-variable strategy for endogeneity, the threats to internal, external, construct, and statistical-conclusion validity, and the named rival explanations [\[33\]](#ref-33).

Chapter 6 is the pre-registered analysis plan. It states the five-step estimation procedure, from assembling and validating the panel through describing before estimating, the fixed-effects baseline, the endogeneity step, and the robustness and heterogeneity battery, and it fixes the binding decision rule and the deliberately unpopulated illustrative table. It reports no executed estimates.

Chapter 7 is the discussion. It works through the implications under the alternative and under the null symmetrically, weighs the rival explanations even if the predicted pattern appears, treats external validity with the procurement evidence as the nearest benchmark [\[12\]](#ref-12), and restates the binding falsification conditions, closing with the single permitted institutional-design implication.

Chapter 8 concludes by restating the threefold contribution, the first part of which stands even if the hypothesis is not confirmed: the consistent long-run NASA latency series, the joining of the two literatures, and the pre-registered falsifiable hypothesis with its fixed decision rule and bounded counterfactual. It reaffirms the honest design-stage posture and points to future work, including executing the panel and extending the design to comparator agencies.

The back matter supplies the full reference list and the appendices: the variable-construction codebook, the program-phase inclusion and coverage table, the era-regime definition table, the robustness-battery specification list, and the time-stamped pre-registration record.

The document closes on the same claim it opens with. The contribution is a measurable variable and a falsifiable test of it; the measurement stands regardless of the regression result; and the discipline throughout is that of quantitative economic history: state the proposition as a quantity, build the measure from primary records, bound the estimate, and let the data decide.



## Chapter 2. Theoretical Framework

## 2.0 The chapter thesis

Decision-and-authorization latency inside NASA is best understood as an internal transaction cost, and the historical record of one agency can be read as a cliometric series only if three theoretical commitments govern the reading from the start: measurement must be put on a single, constant-unit standard before any cross-era claim is made; the counterfactual that "the program would have cost less or flown sooner under faster authorization" must be named and bounded rather than asserted; and the durability of slow authorization across decades must be explained by an institutional mechanism rather than treated as accident. This chapter builds that framework. It develops the cliometric measurement discipline of Angus Maddison, the bounded-counterfactual discipline of Robert Fogel, and the institutional transaction-cost theory of Douglass North and Oliver Williamson into a single conceptual model. The model yields the testable mechanism the empirical chapters operationalize: fragmented, sequential authorization gates raise the elapsed time each pending program action waits, that elapsed time is the latency variable, and longer latency is predicted to be associated with greater cost growth, greater schedule slip, and lower mission cadence, net of program and era fixed effects. The framework is the source of the sign predictions in the specification, and it is also the source of the discipline that keeps those predictions honest: every claim in this chapter is paired with its reasoning, its limits, and the evidence that would lower its confidence, because a framework that cannot be wrong cannot do work for a falsifiable dissertation.

The problem this chapter addresses is theoretical, not yet empirical. The *current state* is that the NASA cost-and-schedule literature has a sophisticated theory of why programs grow technically but no theory of administrative time as a measured cause, while the public-administration literature has a validated theory of administrative process but no application of it to one agency across a long horizon. The *desired state* is a single conceptual model in which administrative latency is a transaction cost, comparable across eras, with an explicit counterfactual and an explicit reason for its persistence. The *gap* is that no existing framework joins quantitative economic history's measurement and counterfactual discipline to institutional transaction-cost theory and points the joined apparatus at NASA program execution. The *consequence* of leaving the gap open is that the empirical work would have no principled rule for constructing latency across seven decades, no defensible way to interpret a coefficient as something other than a raw correlation, and no theory of why the coefficient, if found, would be a controllable lever rather than a fixed feature of spaceflight. This chapter closes the gap before the data chapters begin.

## 2.1 The cliometric method: quantification as the precondition of comparison

The first commitment of the framework is that no two periods of NASA's history can be compared until they are placed on a transparent, replicable measurement standard expressed in constant units. This is the central methodological insight of Angus Maddison's program in quantitative economic history, and it is the precondition for the entire dissertation rather than one technique among several.

Maddison's life work was the construction of long-run national accounts in common purchasing-power units so that the growth of economies separated by centuries and continents could be set against one another on the same footing [\[26\]](#ref-26). The Maddison Project continues that construction, maintaining historical gross domestic product series built so that one period can be compared with another without the comparison being contaminated by changing prices, changing definitions, or changing measurement conventions [\[16\]](#ref-16). The methodological core is not the particular numbers; it is the rule that comparison is meaningless until a single standard is imposed, and that the standard must be stated openly enough that another scholar can reproduce it. The cliometric tradition more broadly, surveyed by Diebolt and Haupert, is the application of explicit, replicable measurement and hypothesis-testing to the historical record, and it treats the construction of a consistent series as itself a scholarly product, not merely a step toward regression [\[137\]](#ref-137). North's own programmatic statement of the task of economic history places the same emphasis on measuring structure and performance before theorizing about cause [\[138\]](#ref-138).

The transfer to this study is direct and load-bearing. NASA program cost cannot be compared between a 1962 program and a 2024 program until both are deflated to constant fiscal-year dollars, and the dissertation fixes the deflator as the NASA New Start Inflation Index so that the rule is single and stated. Schedule cannot be compared until slip is measured against a consistently defined baseline rather than against whatever each program office happened to call its plan. The explanatory variable itself, authorization latency, cannot be compared across eras until it is measured by one documentary rule applied identically everywhere. The bible definition of latency, the elapsed months between a documented trigger event and the documented authorization event that resolves it, taken as the median across a phase's authorization events, is a Maddisonian construction: a single rule, applied to every program-phase, producing a number that means the same thing in 1962 and in 2024.

Consistent constant-unit measurement is therefore a necessary condition for the dissertation's comparisons. The Maddison and cliometric record bears this out, showing that cross-era comparison is defensible only after constant units are imposed [\[16\]](#ref-16), [\[26\]](#ref-26), [\[137\]](#ref-137), and the reason is a general methodological principle: quantities measured under different conventions are not comparable, so any cross-era regression on inconsistently measured variables estimates an artifact of the conventions rather than a relationship in the world. The cost-estimating literature itself already deflates to constant dollars when it compares growth across programs [\[1\]](#ref-1), which shows the discipline is not a foreign imposition but the established practice in the very community whose data this study extends. One qualification must be protected: constant-unit measurement is necessary but not sufficient, because a perfectly consistent measure of the wrong construct would still mislead, which is why the construct-validity defense of latency in the research-design chapter is a separate argument and not subsumed here. A reader might object that documentary density changes so much across seven decades that no single rule can yield truly comparable numbers; the framework's response, developed in the data chapter and previewed here, is the two-resolution design, a coarse milestone-to-milestone latency measure available for the full span and a fine key-decision-point measure available only for the modern subperiod, with results reported separately so that any finding is shown not to be an artifact of changing documentary resolution.

Two further Maddisonian rules structure the work and are stated here so the later chapters can invoke them by name. The first is the rule of quantifying proximate sources before reasoning about ultimate causes. The dissertation measures the proximate variable, latency, and its proximate correlates, cost growth, schedule slip, and cadence, before drawing any conclusion about ultimate institutional causes such as the design of the appropriations process or the fragmentation of decision authority. This ordering is not pedantry; it protects against the common failure mode in which a grand causal story is told and then decorated with whatever numbers are at hand. The second is that the long horizon disciplines short-run extrapolation. By spanning 1958 to 2026 rather than a recent window, the study prevents its conclusions from being an artifact of one budget era, and it allows era fixed effects to separate the conditions of, for example, the Apollo build-up from the conditions of the post-Shuttle period. The descriptive long-run latency series, reported before any estimation, is in this sense the first deliverable of the dissertation and stands as a contribution whether or not the regression rejects the null, exactly as Maddison's GDP series stand as contributions independent of any particular growth regression run on them.

## 2.2 The counterfactual and the bounded estimate
The second commitment is that the claim animating the dissertation is a counterfactual claim, and that a counterfactual claim is unmeasured until the next-best alternative is named, costed, and reported as a bounded range rather than a single point. This is the discipline Robert Fogel brought to economic history, and it governs how the dissertation interprets its central coefficient.

Fogel's contribution was to insist that a statement of the form "outcome Y could not have occurred without X" remains an assertion until one constructs the world in which X never happened and measures what Y would have been in that world. His estimate of the social saving of railroads was built by constructing the counterfactual economy in which the railroad was never built and the transport task fell to canals and wagons, and he reported the saving as a bounded range conditioned on stated assumptions rather than as a single confident number. The transfer of this method to other backward economies, including Herranz-Loncán's reconstruction of the railroad's impact in nineteenth-century Spain, shows the discipline operating as a portable technique: name the alternative, cost it with a defensible shadow price, and report bounds [\[17\]](#ref-17). The bounds serve epistemic honesty. A single point estimate of a counterfactual quantity overstates what the method can know, because the counterfactual world is not observed and its reconstruction depends on assumptions that are themselves uncertain.

The claim that a NASA program "would have cost less or flown sooner under faster authorization" is exactly this kind of counterfactual, and it inherits exactly this discipline. The dissertation cannot run the counterfactual experimentally; it cannot rerun the James Webb Space Telescope's development under a compressed authorization regime and observe the cost. What it can do is approximate the counterfactual through within-program comparison, asking how a program's cost growth differs across its own phases as its own latency varies, and through within-era comparison, asking how programs that experienced different latency in the same rule regime differ in outcome. This is the feasible counterfactual, and it is the reason the identification strategy in the bible specification rests on within-program and within-era variation absorbed by the program fixed effects \(\alpha_i\) and the era fixed effects \(\delta_t\). The Fogel discipline also dictates the reporting standard: the central estimate of the latency coefficient \(\beta\) is to be reported as a bounded range conditioned on the stated assumptions, not as a single causal point, and the analysis-plan chapter fixes this as the required output format.

The dissertation's contribution is therefore a counterfactual and must be bounded, not pointed. The Fogel-tradition record supports this, showing that counterfactual quantities are credibly estimated only when the alternative is named and the answer is bounded [\[17\]](#ref-17), and the reason is that an unobserved counterfactual world is reconstructed under uncertainty, so a single point understates that uncertainty while a range represents it. The modern panel-econometrics insight, developed in the research-design chapter, reinforces the point: heterogeneous treatment effects across programs and eras make a single summary estimand fragile, which is the contemporary statistical analogue of Fogel's insistence on bounds [\[21\]](#ref-21), [\[22\]](#ref-22), [\[23\]](#ref-23). Bounding does not rescue an estimate from a bad identification strategy; bounds around a confounded estimate are still confounded, which is why the counterfactual discipline of this section is paired with, and does not replace, the endogeneity treatment of the design chapter. One might object that bounded ranges are less actionable for program managers than point estimates; the answer is that a falsely precise point estimate that does not survive its own assumptions is worse than useless to a manager because it invites a decision the evidence cannot support, so the honest range is the more, not less, useful management input.

A subtle consequence of the Fogel discipline is the treatment of schedule and cadence as outcomes to be measured rather than as benefits to be assumed. Fogel valued time saved with a defensible shadow price rather than asserting that saved time was valuable by definition. The dissertation accordingly treats schedule slip and mission cadence as dependent variables, measured against baselines and family definitions, and does not assume in advance that compressing latency is valuable; whether it is valuable depends on whether the data show latency moving these outcomes. This is what keeps the study a test rather than an advocacy exercise. The null hypothesis is a genuine possibility precisely because the framework refuses to assume the benefit it is trying to measure.

## 2.3 Institutions and transaction costs: latency as the price of authorization

The third commitment is the theoretical heart of the framework: decision-and-authorization latency is an internal transaction cost. This move converts a vague practitioner intuition, "the agency is slow," into a variable with a theory behind it, and it draws on Douglass North's institutional economics and Oliver Williamson's transaction-cost economics.

North's foundational distinction is between institutions, the rules of the game, and organizations, the players who operate within them [\[18\]](#ref-18). He located economic performance in the transaction costs that institutions raise or lower, where transaction costs are the costs of measuring the valuable attributes of what is being exchanged and of enforcing agreements about it. His central historical argument is that moving from personal to impersonal exchange across distance and time multiplies these measuring-and-enforcing costs, and that institutions exist precisely to lower them; his later restatement of institutional economic theory sharpens the point that the cost structure of a transaction is set by the institutional environment, not by the technology of production alone [\[18\]](#ref-18), [\[91\]](#ref-91), [\[138\]](#ref-138). Williamson operationalized the transaction-cost idea for the firm, mapping transactions onto governance structures, markets, hybrids, and hierarchies, and arguing that each governance structure economizes on a different mix of transaction costs depending on asset specificity, uncertainty, and frequency [\[27\]](#ref-27), [\[28\]](#ref-28), [\[121\]](#ref-121). His analysis of public and private bureaucracies treats the public bureau as a governance structure whose distinctive feature is a set of procedural safeguards that raise the cost of internal transactions in exchange for accountability and probity [\[131\]](#ref-131). The transaction-cost framework has been carried into the analysis of procurement and contract governance specifically, where the costs of measuring, authorizing, and enforcing a public purchase are shown to be a determinant of procurement structure and outcome [\[81\]](#ref-81), [\[84\]](#ref-84), [\[97\]](#ref-97), [\[113\]](#ref-113), [\[119\]](#ref-119).

The framework's defining claim is therefore stated precisely: authorization latency is, in North's and Williamson's vocabulary, the internal transaction cost of measuring a proposed program action against the agency's rules and obtaining authorization to proceed. When a program proposes to move from formulation to development, or to rebaseline after a problem, or to exercise a contract option, the action must be measured against the applicable rules, routed through the layers empowered to authorize it, and resolved before execution can continue. The elapsed time that this measuring-and-authorizing consumes is latency, and it is a cost in exactly North's sense: a cost imposed by the institutional environment, distinct from the engineering cost of the action itself. This is the theoretical reason the dissertation expects latency to matter for cost, schedule, and cadence, and the reason latency is conceptually separable from technical difficulty, which is the separation the empirical design must then defend.

Authorization latency is, on this account, an internal transaction cost. The North and Williamson literatures establish the foundation, treating the procedural cost of authorizing an action within a hierarchy as a recognized and studied transaction cost [\[18\]](#ref-18), [\[27\]](#ref-27), [\[131\]](#ref-131), reinforced by the procurement-governance applications that measure such costs in public purchasing [\[81\]](#ref-81), [\[97\]](#ref-97), [\[113\]](#ref-113). The transaction-cost premise carries the logic: the cost of completing an exchange or an internal action is set by the institutional structure that governs measuring and authorizing it, so a procedural authorization gate is, by construction, a transaction cost. The empirical support is direct: Decarolis and colleagues, using federal procurement data, find that more competent procurement bureaus cause significant reductions in delays and cost overruns, which shows that the administrative, transaction-cost side of program execution moves cost and schedule outcomes [\[12\]](#ref-12). Not all of authorization time is a deadweight transaction cost; some authorization delay buys genuine option value by catching errors before they propagate, a point Williamson's accountability rationale for public-bureau procedure makes explicit [\[131\]](#ref-131), so the framework predicts a relationship between latency and outcomes but does not claim that all latency is waste. The obvious objection is that latency might be merely a symptom of difficult programs rather than an independent cost, in which case it would carry no causal content; this is the endogeneity threat, and it is so central that the framework names it here and the design chapter answers it with program fixed effects and an instrumental-variable strategy in the spirit of Decarolis [\[12\]](#ref-12).

The transaction-cost literature also tells the framework how to read the convergence of its corroborating evidence, which is a question of interpretation rather than of citation count. The make-or-buy and contract-governance studies in the corpus do not study NASA, and none of them measures authorization latency by the dissertation's documentary rule; what they establish, taken together, is the prior probability that procedural transaction costs are real and consequential in public organizations. Poppo and Zenger's test of competing theories of the firm shows that transaction-cost determinants predict make-or-buy decisions in information services better than the alternatives, which is evidence that the transaction-cost construct has empirical bite rather than being a tautology [\[119\]](#ref-119). Canitez and Celebi show the same determinants shaping procurement structure in public transport [\[113\]](#ref-113), Neal and colleagues show transaction-cost reasoning governing evidence-purchasing decisions in public school districts [\[97\]](#ref-97), and Landucci and Ciarallo operationalize transaction-cost economics directly for contract governance using procurement data [\[81\]](#ref-81). The interpretive point is convergence across sectors: when an independent body of studies in transport, education, and supply chains all find that procedural transaction costs move organizational decisions and outcomes, the prior that such costs also operate inside NASA is raised, even though none of these studies is about NASA. This convergence does not prove the NASA-specific claim; it is corroboration, not generalization, and the framework treats it as raising a prior rather than as evidence for the coefficient. The single piece of evidence that bears directly on the cost-and-schedule outcome variables is Decarolis and colleagues' finding that administrative competence causally reduces delays and overruns in federal procurement, which is why that study, and not the sectoral make-or-buy studies, anchors the mechanism claim [\[12\]](#ref-12).

The transaction-cost framing also clarifies what the controls in the bible specification are doing. The vector \(X(i,p,t)\) includes technical controls, instrument and spacecraft mass and power, technology readiness at commitment, mission class, and contract type, drawn from the cost-estimating literature [\[1\]](#ref-1), [\[19\]](#ref-19), and programmatic controls including the number of external partners and a funding-instability index. In transaction-cost terms, the technical controls hold constant the engineering difficulty of the action, the partner count holds constant the number of parties whose agreement must be measured and enforced (a Williamsonian determinant of transaction cost), and the funding-instability index holds constant the uncertainty in the institutional environment that North identifies as a transaction-cost amplifier. The controls are not an atheoretical kitchen sink; each corresponds to a transaction-cost determinant the theory names, which is why holding them constant isolates the part of latency that is procedural rather than substantive.

## 2.4 Adaptive efficiency and path dependence: why latency is large and durable

A theory that latency is a transaction cost is incomplete without an account of why that cost is large and why it persists across decades rather than being competed away. North supplies this account through the concepts of adaptive efficiency and path dependence, and it is the theoretical justification for the era fixed effects \(\delta_t\) in the bible specification.

North's argument is that institutions, once established, generate increasing returns. A set of review-and-authorization rules creates an organizational ecology around itself: roles are defined by the rules, expertise accumulates in operating them, complementary procedures are built on top of them, and the constituencies that benefit from the rules acquire an interest in their continuation [\[18\]](#ref-18), [\[91\]](#ref-91). The consequence is that the cost of changing the rules rises over time even when the rules are inefficient, so an organization can lock onto a persistent decision trajectory that no single actor chose and no single actor can easily reverse. This is path dependence: the present rule regime is shaped by the sequence of past regimes, and the menu of feasible reforms is narrowed by the accumulated investment in the existing arrangement. North contrasts this with adaptive efficiency, the capacity of an institutional framework to learn and to discard what does not work, and his historical claim is that adaptive efficiency is rare and that lock-in is common. Williamson's analysis of public bureaucracy reinforces the durability claim from a different angle: the procedural safeguards that define the public bureau are deliberately hard to change because their entire function is to constrain discretion, so a public agency's authorization procedures are durable by design, not by accident [\[131\]](#ref-131). Ostrom's study of long-enduring institutions for collective action shows the same lesson from the side of success, that durable institutional arrangements are sustained by the very feedback structures that make them resistant to quick change [\[31\]](#ref-31).

The transfer is the theoretical reason the dissertation expects authorization latency to be both substantial and persistent within a rule regime, and to shift in discrete steps when a regime changes, rather than to drift smoothly. NASA's authorization architecture has been reformed in identifiable regimes, the introduction and revision of the agency's program-management requirements, the adoption of joint cost-and-schedule confidence-level policy, and the standing-review-board structure among them, and the path-dependence account predicts that within each regime latency is relatively stable and across regimes it jumps. This prediction has a direct methodological consequence: era fixed effects \(\delta_t\) are necessary, not optional, because they absorb the rule-regime level of latency and force identification of the latency coefficient onto within-regime variation. Without era fixed effects, a regression would conflate the durable, regime-set level of latency with the program-specific variation that the framework holds to be the informative signal.

Authorization latency is large and durable because authorization rules are path-dependent and resist change. North's increasing-returns account of institutions [\[18\]](#ref-18), [\[91\]](#ref-91) and Williamson's account of the deliberately rigid public bureau [\[131\]](#ref-131) establish this, with Ostrom's enduring-institutions evidence as corroboration [\[31\]](#ref-31). The increasing-returns mechanism carries the logic: once a rule generates a supporting ecology of roles, expertise, and constituencies, the cost of changing it rises, so inefficient rules persist. NASA's authorization architecture is in fact reformed in discrete, infrequent regimes rather than continuously, which is the observable signature the path-dependence account predicts and which the era-regime definition table in the backmatter codifies. Path dependence explains durability but not the specific magnitude of latency in any one regime, which is an empirical quantity to be measured, not deduced; the framework predicts the shape of the latency series (stable within regime, stepped across regimes) without predicting its levels. The obvious objection, that observed latency persistence might reflect persistent program difficulty rather than persistent rules, is met by the specification: program fixed effects absorb time-invariant program difficulty while era fixed effects absorb the rule regime, so the two confounds are addressed by distinct components of the design, and the design chapter's staggered-regime analysis tests the regime-change prediction directly.

A second consequence of the path-dependence account concerns how era regimes enter the empirical design and why the framework insists they enter as discrete steps rather than as a continuous time trend. If authorization rules changed continuously and smoothly, a linear era trend would suffice to absorb their influence; but the North and Williamson accounts predict the opposite, that rules are stable for long stretches and then change in identifiable reforms, because the increasing-returns mechanism makes incremental change costly and concentrates change into episodic regime shifts [\[18\]](#ref-18), [\[131\]](#ref-131). This is why the dissertation defines eras as rule regimes, mapped explicitly in the backmatter era-regime table to $\delta_t$, rather than as calendar decades or as a smooth trend. The choice is theoretical, not merely a modeling convenience: it follows from the path-dependence claim that the relevant temporal structure of latency is the regime, the durable rule set, and not the year. It also carries a methodological cost the framework acknowledges, because regime boundaries are themselves a construct that the analyst draws, and a poorly drawn boundary would misallocate variation between $\delta_t$ and $\beta$. The research-design chapter addresses this by testing sensitivity to alternative regime definitions, so that the era-regime construct is defended rather than assumed, in the same spirit that the latency construct is defended at two resolutions.

The path-dependence account also supplies the mechanism by which a finding, if obtained, would be actionable. If latency is durable because rules are sticky, then compressing latency requires changing rules, which is costly but feasible and within the agency's control. This is the bridge from the theoretical framework to the single permitted institutional-design implication the discussion chapter draws, that faster, lower-layer authorization is a management lever, stated in program-execution-management terms. The framework earns that implication here by establishing that latency is a rule-set property the agency can alter, not a fixed cost of spaceflight.

## 2.5 The rhetoric-of-significance caution and the reporting standard

A framework built on measurement and bounded estimation must also fix a standard for what counts as a finding, and here the framework adopts the caution, long pressed within cliometrics, that statistical significance is not substantive significance. This commitment disciplines the reporting standard the analysis-plan chapter then enforces.

The cliometric tradition, in its self-critical mode surveyed by Diebolt and Haupert, has repeatedly distinguished a coefficient that is statistically distinguishable from zero from a coefficient whose magnitude matters for the question at hand [\[137\]](#ref-137). North's framing of the task of economic history as the measurement of structure and performance carries the same implication: the object is to learn how much structure shapes performance, and how much is a question of magnitude rather than of whether a coefficient clears a significance threshold [\[138\]](#ref-138). The caution bites acutely here because the panel is unbalanced and the number of program clusters is modest, conditions under which conventional standard errors can mislead in either direction. A latency coefficient could be statistically significant and economically trivial, a fraction of a percent of cost growth per month of latency that no manager would notice; or it could be economically large and statistically fragile, an effect worth acting on but estimated imprecisely from a short series. The framework requires that both the magnitude and the precision be reported and interpreted, and it forbids the substitution of a significance star for a substantive judgment.

The reporting standard must therefore privilege substantive magnitude and bounded precision over bare significance. The cliometric self-critique and North's structure-and-performance framing establish the principle [\[137\]](#ref-137), [\[138\]](#ref-138), and the reason is that a hypothesis about whether latency is a *material* lever on program performance is answered by magnitude, so a report that gives only the sign and significance of a coefficient leaves the question asked unanswered. The bounded-estimate discipline of Section 2.2, which already commits the study to reporting ranges, supports this, as does the heterogeneous-effects econometrics of the design chapter, which shows that a single significant summary coefficient can mask offsetting effects across programs and eras [\[21\]](#ref-21), [\[23\]](#ref-23). This caution does not license discarding significance entirely; the pre-registered decision rule still requires the coefficient to be statistically distinguishable from zero in the same direction across specifications, so significance is necessary but not sufficient for a finding. One might object that emphasizing magnitude invites post hoc judgments about what magnitude is "large enough"; the framework forecloses this by requiring that the magnitude be interpreted against a stated reference, the historical distribution of cost growth itself, so that "large" means "large relative to the growth the agency already experiences," a benchmark fixed before estimation rather than chosen to fit a result.

This caution is the reason the illustrative coefficient table in the analysis-plan chapter is left deliberately unpopulated and the expected signs are labeled as directional expectations, not findings. Reporting invented coefficients, even as placeholders dressed as results, would violate both the bounded-estimate discipline and the significance caution at once, because it would assert a magnitude and an implied precision that no estimation produced. The refusal to populate the table is the reporting standard of this section made concrete.

## 2.6 The conceptual model: latency as a transaction-cost mechanism
The four anchors combine into a single conceptual model that the empirical chapters test. This section states the model as an explicit causal chain, attaches the framework's components to each link, and fixes the sign predictions that the bible specification estimates.

The causal mechanism is stated as a chain from driver to strategic implication, and each causal claim names its mechanism rather than resting on correlation. The *driver* is the institutional structure of NASA authorization: fragmented decision authority distributed across program, center, and headquarters layers, sequential authorization gates that must be cleared in order, and an annual appropriations cycle that paces the budget actions on which each phase depends. This structure is a transaction-cost environment in North's and Williamson's sense, and it is durable in North's path-dependent sense [\[18\]](#ref-18), [\[27\]](#ref-27), [\[91\]](#ref-91), [\[131\]](#ref-131). The *mechanism* is that each pending program action accrues elapsed time, the latency, while it is measured against the rules and routed for authorization, and during that elapsed time the program's standing costs continue to accrue and its requirements are exposed to drift. The *observable effect* is that measured authorization latency per program-phase rises when the structure imposes more or slower gates. The *operational consequence* is that higher latency is associated with greater cost growth, because standing costs accumulate during waiting and delay invites requirements change; with greater schedule slip, because authorization latency is itself a component of elapsed schedule and the path from latency to slip is the most direct of the three; and with lower mission cadence, because higher latency per decision lengthens the interval between successive flight or delivery events within a program family. The *strategic implication* is that a portion of NASA's cost and schedule performance is a controllable process variable, distinct from the irreducible difficulty of spaceflight, and therefore a lever for program execution management, including at the Jet Propulsion Laboratory, whose long-cycle deep-space programs accumulate authorization events over many years.

The model maps onto the bible specification exactly. For program-phase observation indexed by program \(i\), phase \(p\), and era \(t\), the baseline specification is:

\[
\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t) \qquad\qquad (1)
\]

where Outcome is in turn cost growth, schedule slip, and cadence; Latency is authorization latency in months; \(X\) is the vector of technical and programmatic controls; \(\alpha_i\) are program fixed effects; \(\delta_t\) are era fixed effects; and \(\varepsilon\) is the error clustered by program. The coefficient of interest is \(\beta\). The framework supplies the sign predictions under the alternative hypothesis directly: \(\beta > 0\) for cost growth and for schedule slip, because the mechanism accrues standing cost and elapsed time during waiting, and \(\beta < 0\) for cadence, because higher latency per decision lengthens the interval between flight or delivery events. These are the predictions stated, verbatim in convention, in the bible, and the model derives them rather than restating them. Each component of the specification corresponds to a component of the model: \(\alpha_i\) removes the time-invariant program difficulty that the transaction-cost framing holds to be conceptually separable from procedural latency; \(\delta_t\) removes the durable, path-dependent rule-regime level of latency that the North account predicts; \(X\) holds constant the transaction-cost determinants the Williamson account names; and \(\beta\) isolates the within-program, within-era association between procedural latency and outcome that is the object of the test.

The hypotheses the model is built to test are fixed and restated here without alteration. The null, H0, is that administrative decision-and-authorization latency in NASA programs has no association with program cost growth, schedule slip, or mission cadence, after accounting for program and era fixed effects and technical controls. The alternative, H1, is that longer administrative decision-and-authorization latency is associated with greater program cost growth, more schedule slip, and lower mission cadence, after accounting for program and era fixed effects and technical controls. The framework makes H0 a genuine possibility rather than a straw figure: because latency may be endogenous to program difficulty, a raw positive correlation between latency and cost growth could reflect difficulty rather than process, and the transaction-cost separation the model asserts is a hypothesis about the world that the data may refuse to confirm. The model is therefore falsifiable in the strict sense the dissertation requires; if the estimated \(\beta\) is statistically indistinguishable from zero in the preferred specification, or reverses sign between the fixed-effects and instrumented specifications, or vanishes under the heterogeneity-robust estimators or the alternative latency resolution, the contribution fails by the pre-registered decision rule.

The framework also states its confidence calibration explicitly, as the logic band requires. The measurement contribution, the construction of a consistent constant-unit latency series across 1958 to 2026, is held with *high* confidence at the design stage, because it rests on the established Maddisonian discipline and on documentary sources whose existence is known [\[16\]](#ref-16), [\[26\]](#ref-26); what would lower this confidence is a discovery that the documentary record is too sparse in the early decades to support even the coarse latency measure, which the two-resolution design is built to detect. The causal claim that latency *moves* the three outcomes is held with no more than *moderate* confidence at the design stage, because it depends on an identification strategy that has not yet been executed and on an instrumental-variable argument whose strength is not yet measured; the analogous federal-procurement evidence raises this confidence by showing the mechanism operating in a related setting [\[12\]](#ref-12), and what would raise it further is a strong first stage and a coefficient stable across the fixed-effects and instrumented specifications, while what would lower it is a weak instrument, sign reversal across specifications, or evidence that the association is explained by baseline gaming or reverse causation. The framework will not claim more than the design-stage evidence grade supports, and it labels every directional expectation as an expectation, never as a result.

The argument of the dissertation runs through this model and is stated here in the order the later chapters develop it. The problem is real: NASA programs systematically grow in cost and slip in schedule, established by the cost-and-schedule literature [\[1\]](#ref-1), [\[5\]](#ref-5), [\[130\]](#ref-130) and corroborated by independent financial analysis of NASA contractors [\[34\]](#ref-34). The problem is material: overruns are large and consequential, with vivid program-record cases such as the Constellation cost history [\[102\]](#ref-102) and the broader megaproject evidence on fat-tailed overrun [\[13\]](#ref-13), [\[30\]](#ref-30). The design addresses the causal mechanism: latency is operationalized from documentary authorization records and entered net of program and era fixed effects with an instrument for office workload and appropriations timing, grounded in the transaction-cost theory of this chapter and the procurement-competence precedent [\[12\]](#ref-12), [\[18\]](#ref-18). It improves on the alternatives: the design uses heterogeneity-robust staggered estimators rather than naive two-way fixed effects, with a Goodman-Bacon diagnostic [\[21\]](#ref-21), [\[22\]](#ref-22), [\[23\]](#ref-23). The residual risk is acceptable and managed: endogeneity, baseline gaming, reverse causation, and construct fragility are pre-named with explicit checks, including the instrumental-variable strategy, the baseline-conservatism test, early-in-phase latency measurement, multiple cadence definitions, and the two latency resolutions [\[12\]](#ref-12), [\[33\]](#ref-33). This chapter establishes the first and third of these points, that the problem is real and that the design addresses the mechanism, by grounding both in the transaction-cost theory; the remaining chapters develop the others.

One scope decision is stated explicitly to close the framework. The architecture-traceability layer that would trace a strategic objective through a capability to an operational activity, a system function, and a data or service exchange is out of scope for this dissertation and is deliberately omitted. This is a cliometric econometric study whose objects are programs, phases, documentary authorization events, and panel estimates; it does not design, field, or deliver a capability, a system, or a data exchange. Forcing enterprise-architecture vocabulary onto a pure econometric and institutional-economics contribution would be a category error, and the framework therefore expresses its single institutional-design implication, faster and lower-layer authorization as a program-execution-management lever, in management terms rather than architecture-layer terms. The model of this chapter is complete as an econometric and institutional-economic object, and it requires no architectural scaffolding to be tested.

## 2.7 Summary of the framework

The framework assembled in this chapter converts a widely shared but unmeasured practitioner intuition into a falsifiable conceptual model through four theoretical commitments. From Maddison and the cliometric tradition it takes the commitment to constant-unit, single-rule measurement as the precondition of any cross-era comparison, which fixes the latency construction rule and the deflation standard and makes the long-run latency series a standalone contribution [\[16\]](#ref-16), [\[26\]](#ref-26), [\[137\]](#ref-137), [\[138\]](#ref-138). From Fogel it takes the commitment to naming and bounding the counterfactual, which fixes the within-program and within-era comparison as the feasible counterfactual and the bounded range as the required reporting format [\[17\]](#ref-17). From North and Williamson it takes the commitment to treating authorization latency as an internal transaction cost, the price of measuring an action against the rules and obtaining authorization, which gives the latency variable its theory, separates it conceptually from engineering difficulty, and grounds the controls in named transaction-cost determinants [\[18\]](#ref-18), [\[27\]](#ref-27), [\[28\]](#ref-28), [\[81\]](#ref-81), [\[84\]](#ref-84), [\[91\]](#ref-91), [\[97\]](#ref-97), [\[113\]](#ref-113), [\[119\]](#ref-119), [\[121\]](#ref-121), [\[131\]](#ref-131). From North's adaptive-efficiency and path-dependence analysis, reinforced by Williamson and Ostrom, it takes the account of why latency is large and durable, which justifies the era fixed effects and supplies the mechanism by which a finding would be an actionable lever rather than a fixed cost [\[18\]](#ref-18), [\[31\]](#ref-31), [\[131\]](#ref-131). The cliometric significance caution fixes the reporting standard that privileges substantive magnitude and bounded precision over bare statistical significance [\[137\]](#ref-137). The result is the conceptual model of Section 2.6: a transaction-cost mechanism running from fragmented, path-dependent authorization structure through accrued latency to cost growth, schedule slip, and cadence, mapped exactly onto the bible specification with sign predictions $\beta > 0$ for cost growth and schedule slip and $\beta < 0$ for cadence, and held to the design-stage confidence calibration that labels every prediction an expectation and protects the genuine possibility of the null. The data, measurement, and research-design chapters that follow operationalize this model; the analysis-plan chapter pre-registers the test of it; and the discussion chapter interprets either outcome of that test against the framework built here.



## Chapter 3. Literature Review

## 3.0 The chapter's answer, stated first

The literature relevant to this dissertation divides into two mature bodies of work that have never been joined for a single agency over the long run, and the precise shape of that disjunction is the gap the dissertation occupies. On one side stands a technically sophisticated NASA cost-and-schedule-growth literature, produced largely by the agency's cost-estimating community, which models program overrun as a function of technical and contractual parameters and treats schedule as an outcome or as a secondary driver of cost. On the other side stands a public-administration and economics literature on red tape, administrative burden, and bureaucratic competence, which has constructed validated measures of the procedural delay that organizations impose and has, in at least one rigorous case, demonstrated a causal link from administrative competence to delays and cost overruns in government procurement. A third, megaproject literature bridges the two by showing that large public projects overrun systematically and fat-tailedly, and that decision processes, optimism bias, and strategic misrepresentation are part of the cause rather than engineering difficulty alone. A fourth, smaller NASA program-management literature documents the agency's own reflections on management control and program execution.

The synthesis this chapter reaches, and defends claim by claim, is that the first body of work has the right dependent variables (NASA cost growth, schedule slip, cadence) but omits administrative decision-and-authorization latency as a measured explanatory variable; the second body has the right explanatory construct (procedural delay as a measurable organizational attribute) but has never been applied to NASA as a long-run program panel; the third supplies both a basis for expecting a latency effect and a methodological caution about how that effect can be confounded by baseline gaming; and the fourth supplies the institutional texture but not the quantification. The propositions that close the chapter follow directly: a consistent, documentary-rule-based latency series for NASA does not exist and is worth constructing in its own right, and a pre-registered panel test of whether longer latency raises cost growth and schedule slip and lowers cadence is both feasible and unprecedented. The gap that defines the dissertation is the precise non-overlap of two large literatures, and establishing that non-overlap requires reading both in substance, which is why this chapter runs long.

The remainder of the chapter is organized thematically. Section 3.1 reviews the NASA cost-and-schedule tradition and states what it leaves open. Section 3.2 reviews the public-administration literature on red tape, administrative burden, and validated measurement. Section 3.3 treats the procurement-competence evidence, the nearest external precedent for the dissertation's causal claim. Section 3.4 reviews the megaproject and optimism-bias literature, including the reference-class-forecasting debate and the strategic-misrepresentation account. Section 3.5 covers the NASA program-management and management-control literature. Section 3.6 synthesizes the four into a complete map of the field and states the unfilled gap and the propositions that follow. Throughout, the sign convention from the dissertation's notation is held fixed: under the alternative hypothesis the latency coefficient \(\beta\) is positive for cost growth and schedule slip and negative for mission cadence.

## 3.1 The NASA cost-and-schedule-growth literature

### 3.1.1 Current state, desired state, and the gap this section frames

The current state of knowledge about NASA cost and schedule growth is rich on the technical and estimating side and silent on administrative latency. The desired state, for this dissertation, is a literature that treats the time the agency takes to decide and authorize as a measured regressor on equal footing with mass, power, technology readiness, and contract type. The gap is the distance between the two, and the consequence of leaving it unaddressed is that the agency continues to attribute a portion of overrun to slow decision-making without ever having measured the decision-time variable. This section establishes the gap by reading the estimating literature on its own terms.

### 3.1.2 The reserve-guideline and growth-distribution tradition

The foundational empirical regularity is that NASA programs grow in cost and slip in schedule between their early formulation and their launch, and that the growth is large enough and regular enough to be treated as a distribution from which reserve guidelines can be set. Emmons, Bitten, and Freaner used historical NASA cost and schedule growth to derive future program and project reserve guidelines, and, for the purposes of this dissertation, they attempted to separate growth attributable to external programmatic reasons from growth attributable to internal technical reasons [\[1\]](#ref-1). The method was to assemble a historical sample of completed programs, compute growth relative to baseline at defined points, and fit reserve recommendations to the observed distribution. The limitation, read from the vantage of the present study, is that the "external programmatic" category, which is exactly where authorization latency lives, was acknowledged as a source of growth but was not itself measured as a continuous variable. The claim this dissertation advances is therefore not a contradiction of Emmons and colleagues but an extension of their own taxonomy: the external-programmatic residual they identified is the quantity the present study proposes to measure directly. This work is the conceptual ancestor of the dissertation because the separation of programmatic from technical causes is precisely the partition the latency variable operationalizes, and subsequent agency work continued to use that partition. Emmons and colleagues did not claim their programmatic category was causal, only that it was a source of variance, so the dissertation must supply its own identification rather than borrow theirs.

The growth-distribution tradition matured across a sequence of Aerospace Conference papers. Bitten and colleagues documented historical mass, power, schedule, and cost growth for NASA science instruments, establishing that growth is pervasive across the technical parameters and not confined to cost [\[2\]](#ref-2), and a companion analysis extended the same documentation to NASA spacecraft at the system level [\[3\]](#ref-3). Kipp and colleagues examined the impact of instrument schedule growth specifically on mission cost and schedule growth, which is direct evidence that schedule, the dissertation's second outcome, is itself a driver of cost and not merely a passive byproduct [\[130\]](#ref-130). These papers share a method, the assembly of historical growth datasets keyed to technical milestones, and a common limitation for present purposes: the regressors are technical (mass, power, instrument count) and the administrative process that paces the schedule is treated as background. The interpretation the dissertation draws is that the schedule-to-cost linkage these papers document is the mechanical channel through which latency, if it raises schedule, would also raise cost; the convergence of the instrument-growth and spacecraft-growth findings raises confidence (to moderate-high) that schedule is a live transmission path, while leaving the upstream question of what drives schedule unanswered.

The schedule-cost linkage is made explicit by Majerowicz and Shinn, who examined the relationship between schedule delays and cost overruns directly and argued that schedule is not a passive accounting residual but an active cost driver: programs that slip incur standing costs (the cost of keeping a development workforce and facilities in place) for the additional duration, so that a month of slip translates mechanically into additional dollars even when no scope changes [\[5\]](#ref-5). The method is the analysis of paired schedule-and-cost outcomes across programs; the finding, that schedule matters in the causal sense and not only as a correlate, is the single most important corpus support for the dissertation's claim that its second outcome (schedule slip) and its first outcome (cost growth) are linked through a standing-cost channel. The limitation is that Majerowicz and Shinn relate schedule to cost but do not ask what drives schedule, which is exactly the upstream question the dissertation answers by introducing latency. The dissertation's interpretation closes the loop the estimating literature leaves open: if latency raises schedule slip (the most direct prediction, because authorization latency is itself a component of elapsed schedule), and if schedule slip raises cost through the standing-cost channel Majerowicz and Shinn document, then latency raises cost through a fully specified two-step mechanism rather than through an unexplained correlation. This is why the dissertation predicts that the schedule coefficient will be the largest and most robust of the three: the mechanism connecting latency to schedule is the most direct, and the schedule-to-cost step is independently documented.

### 3.1.3 The policy-and-jointness evidence that administrative regime moves outcomes
Two strands within this literature come closer to the dissertation's thesis than the rest, and both warrant substantive treatment because they show that administrative arrangements, not technology alone, move cost and schedule. Bitten and colleagues examined the effect of policy changes on NASA science mission cost and schedule growth and found that the policy regime is associated with shifts in the growth distribution [\[4\]](#ref-4). The method compared growth outcomes across policy eras; the finding that the distribution moves with policy is the single most direct piece of corpus evidence that an administrative variable, here the policy regime rather than latency itself, affects the dissertation's outcomes. The limitation is that "policy change" is a coarse, era-level treatment rather than a program-phase-level continuous measure of how long authorization took, and the study did not isolate the mechanism through which policy moved the outcomes. The dissertation's contribution can be read as refining this coarse era-level finding into a fine, continuous, program-phase-level latency measure, and as supplying the mechanism (elapsed authorization time during which standing costs accrue and requirements drift) that the policy-change study left implicit.

The jointness evidence points the same way. Dwyer and colleagues studied the cost impacts of jointness through an in-depth case study of the NPOESS program and showed how organizational complexity, induced by joint governance across agencies, enabled, sustained, and induced cost growth over time [\[115\]](#ref-115). The mechanism Dwyer names, organizational complexity multiplying the coordination and authorization burden, is a transaction-cost mechanism in all but name, and it is the strongest single corpus instance of an administrative-structure variable driving cost through a coordination channel. The limitation is that it is a single case with a semi-quantitative framework, so it cannot speak to the general distribution of the effect across the agency's history. On the dissertation's reading, NPOESS is a vivid instance of the proposed mechanism operating, and the case-study evidence raises the prior probability that latency matters while leaving the population-level magnitude to be estimated; confidence drawn from a single semi-quantitative case is low on magnitude and moderate on existence.

The jointness channel rewards a closer look, because it is the corpus's most explicit aerospace statement of how an administrative arrangement converts into cost growth over time. Jointness, the pooling of two agencies' requirements into one program, was adopted on the theory that shared development would reduce total cost; the case found the opposite, that the joint structure induced complexity that enabled, sustained, and then induced further cost growth across the program's life [\[115\]](#ref-115). The driver is administrative (a governance decision to combine requirements), the intermediate quantity is coordination and authorization burden (more parties must agree before any action proceeds), and the observable is cost growth that compounds over the development horizon. This is the dissertation's hypothesized latency mechanism observed in one program, with the caveat that the framework is semi-quantitative and the counterfactual (the same program developed without jointness) is reconstructed rather than observed. The dissertation generalizes the channel and supplies the population-level estimate the case cannot, the division of labor between case evidence and panel estimation that the research-design chapter formalizes.

Dwyer, Szajnfarber, Cameron, and Crawley extended this case-level finding into a generalizable analytical model, published in Acta Astronautica, that links organizational and programmatic structure to cost growth trajectories on joint programs [\[140\]](#ref-140). The model formalizes three mechanisms through which jointness drives cost growth over time: complexity induced by integrating divergent agency requirements, governance friction that multiplies the number of authorization parties and raises the cost of each decision, and an enabling dynamic in which complexity and friction compound each other as the program matures. The contribution relative to the NPOESS case study is that the model generalizes the mechanism from one program's history to a structural account applicable to any multi-agency program, and it does so in the language of transaction-cost economics, making explicit what the NPOESS analysis left implicit. For the dissertation, this paper is the nearest published precedent for reading joint-governance complexity as a form of authorization latency; the governance-friction mechanism Dwyer and colleagues name is, in the dissertation's vocabulary, a structural driver of higher per-decision latency, because more parties must authorize each action and each additional party extends the authorization interval. The dissertation's partner-count control in vector X is the empirical instrument for holding this structural driver constant, so that the surviving latency coefficient captures variation in authorization time that jointness complexity cannot explain [\[140\]](#ref-140). Confidence that the Dwyer et al. model raises the prior for the dissertation's latency hypothesis is moderate to high; it is corroboration at the mechanism level, not a population-level estimate of magnitude.

### 3.1.4 The estimating-model and contract-structure tradition

A parallel strand builds parametric cost-estimating relationships and asks whether contract structure curbs growth. Stahl's survey of cost models for space telescopes shows how parametric cost-estimating relationships are constructed from historical data and where they mislead, and it is the canonical statement of how the cost-estimating community models cost as a function of technical drivers [\[19\]](#ref-19). Rusnock and colleagues sought to identify and quantify factors contributing to military and civil space system cost and schedule growth, using logistic and multiple regression across twenty-one Department of Defense and seventy-one NASA programs, and found, among other results, that firm-fixed-price contracts and longer program-manager tenure predict lower cost growth for military systems, and that smaller NASA programs grow differently than larger ones [\[101\]](#ref-101). This is the closest the estimating literature comes to the dissertation's method, a multi-program regression of growth on program attributes, and the finding that program-manager tenure matters is a faint signal that an administrative-continuity variable can move outcomes. The limitation is that tenure is a proxy for stability, not a measure of authorization latency, and the regressors remain dominated by technical and contractual variables. A 2021 Aerospace Conference analysis of the effectiveness of firm-fixed-price spacecraft contracts in curbing cost growth tested the contract-structure lever directly [\[20\]](#ref-20). The interpretation the dissertation draws across this strand is that the estimating community has already established the regression-on-program-attributes method and has already found that some non-technical attributes (contract type, manager tenure) move outcomes; the dissertation extends this established method to the one non-technical attribute the community has not measured, authorization latency.

### 3.1.5 The agency-level diagnosis and the cost-history framing

Two corpus items diagnose the problem at the agency level. The National Research Council's consensus study on controlling cost growth of NASA Earth and space science missions catalogs the management, technical, and budgetary causes of growth and is the authoritative agency-level statement that growth is multi-causal [\[6\]](#ref-6). Its method is expert consensus over case evidence; its value is that it names budgetary and management causes alongside technical ones, which licenses the dissertation's premise that non-technical causes are in play; its limitation is that a consensus catalog is not a quantification, so it identifies candidate causes without measuring their relative weight. Crooke and colleagues, in a study of evolving management strategies to improve NASA flagship cost and schedule performance using LUVOIR as a case, frame the problem from inside the agency, acknowledging the growing frustration with development cost and schedule overruns even on missions that perform exquisitely on orbit, and proposing management-strategy responses [\[90\]](#ref-90). The Rich and colleagues financial analysis of NASA contractors notes that a Government Accountability Office report attributed twelve billion dollars in cost overruns and twenty-eight cumulative years of schedule delay to fifteen development-phase projects, and examines whether contractor behavior and contract structure explain the pattern using prime-contractor data from 2008 to 2024 [\[34\]](#ref-34). The interpretation is that the agency-level diagnostic literature has established materiality (the dollar and schedule magnitudes are large) and multi-causality (budgetary and management causes are named), which together establish that the problem is both real and material while leaving the specific administrative-latency channel unmeasured.

### 3.1.6 What the NASA cost-and-schedule literature leaves open

Across this body of work, cost and schedule are the dependent variables and technical and contractual parameters are the regressors. Administrative latency is acknowledged in narrative (the "external programmatic" residual of Emmons and colleagues, the budgetary causes of the National Research Council, the organizational complexity of Dwyer and colleagues, the policy-regime effect of Bitten and colleagues) but is never constructed as a measured explanatory variable, and no study assembles a single consistent panel spanning the agency's full 1958-to-2026 history. The gap this section establishes is therefore specific and not rhetorical: the right outcomes are measured, the right method (multi-program regression) is in use, non-technical attributes are already known to move outcomes, and the one non-technical attribute the practitioner narrative emphasizes most, decision-and-authorization latency, is the one the literature has not measured. Confidence in this gap statement is high, because it rests not on the absence of relevant work but on a reading of what the relevant work measures and omits.

## 3.2 The public-administration literature on red tape and administrative burden

### 3.2.1 The construct that the cost literature lacks

If the NASA cost literature supplies the outcomes and lacks the explanatory construct, the public-administration literature supplies exactly that construct: a validated, measurable account of the procedural rules, clearances, and delays that organizations impose, and their relation to performance. Red tape and administrative burden are operationalized constructs with measurement instruments and an empirical record, and that record both motivates and disciplines the dissertation's latency variable. A measured construct in one domain can be transported to another if the operationalization is faithful, and the validated-scale tradition reviewed below supplies the foundation. The section keeps one limit in view rather than burying it: the public-administration evidence finds red-tape effects to be real but conditional and sometimes weaker than conventional wisdom assumes, so the dissertation must not assume a large effect a priori.

### 3.2.2 The Bozeman-Rainey-Pandey measurement program

The modern study of red tape begins with the recognition that the rules an organization imposes can be perceived, reported, and compared across organizations. Rainey, Pandey, and Bozeman, in their research note on public and private managers' perceptions of red tape, established that managers in public and private organizations differ in their reported experience of procedural constraint, and that this perception is a measurable organizational attribute rather than an undifferentiated complaint [\[7\]](#ref-7). The method was survey-based perceptual measurement; the contribution was to make red tape a variable; the limitation, acknowledged within the tradition, is that perceptual measures conflate the objective burden with the respondent's tolerance for it. For the dissertation this matters because it motivates a deliberately objective, documentary operationalization of latency (elapsed months between documented trigger and documented resolution) rather than a perceptual one, sidestepping the perception-versus-reality confound that the survey tradition wrestles with. The interpretation is that the public-administration field discovered the construct and its measurement problem at once, and the dissertation's documentary-time measure is a response to the measurement problem the field identified.

### 3.2.3 The empirical-performance finding and its protected qualifier

The single most consequential finding for the dissertation's framing is Brewer and Walker's empirical analysis of the impact of red tape on governmental performance, which found that red tape's effects on performance are real but more nuanced and conditional than the conventional wisdom that red tape is uniformly harmful [\[8\]](#ref-8). The method was an empirical analysis relating measured red tape to performance across governmental units; the finding was a qualified, conditional effect rather than a uniform negative one. This is the qualifier the chapter's central argument must protect: the dissertation predicts that longer latency is associated with worse outcomes, but Brewer and Walker's result warns that the association may be conditional on context, may be smaller than practitioners assume, and may even reverse in settings where procedure adds value. The dissertation honors this qualifier in two ways previewed here and developed in the research-design chapter: it pre-registers the null as a genuine possibility rather than a straw figure, and it frames its expectation as a directional prediction to be tested rather than a magnitude to be confirmed. A complementary natural quasi-experiment by Ran and colleagues, exploiting the gradual rollout of administrative licensing centers in some Chinese cities to study red-tape reform, found that reducing red tape improved corporate social performance by lowering efficiency-based transaction costs, and that the effect was heterogeneous across reform characteristics [\[107\]](#ref-107). The interpretation across these two studies is that the best-identified evidence (a staggered-rollout quasi-experiment) finds a real but heterogeneous red-tape effect operating specifically through a transaction-cost channel, exactly the mechanism the dissertation posits, and the heterogeneity finding raises confidence that the effect exists while warning that its magnitude is context-dependent.

### 3.2.4 Administrative burden and the green-tape refinement

The construct was extended in two directions that bear on the dissertation. Moynihan, Herd, and Harvey's work on administrative exclusion and the hidden costs of welfare claiming reframed procedure as a cost borne by the participants in a process, not only by the organization imposing it, and showed that procedural burden can have large allocative consequences [\[9\]](#ref-9). The method was case and administrative-data analysis; the contribution was to make the participant-borne cost of procedure visible; the limitation for the dissertation is that the welfare-claiming setting differs from program authorization, so the transfer is conceptual rather than direct. DeHart-Davis's theory of green tape, by contrast, drew a distinction the dissertation must respect: not all rules are red tape, and effective organizational rules (green tape) can improve rather than degrade performance [\[10\]](#ref-10). The green-tape theory is the strongest internal caution in the public-administration literature against assuming that all authorization delay is waste. The dissertation's interpretation is that some authorization latency is green tape (review that catches errors and prevents larger downstream losses) and some is red tape (delay that adds no value), and that the empirical test cannot distinguish the two a priori; this is precisely why the dissertation tests an association rather than asserting that all latency is harmful, and why retaining the null would be informative rather than a failure of the design. Confidence that latency is uniformly harmful is therefore set low at the design stage; the data must adjudicate.

The green-tape distinction also clarifies what the dissertation's latency coefficient actually estimates, worth stating precisely because it bears on interpretation. The coefficient beta is the average association between an additional month of authorization latency and the outcome, across the mix of green and red tape that the agency's processes actually contain. If the agency's authorization processes are predominantly green tape, the value-adding review catches enough downstream error that the net association is small or even favorable, and beta will be near zero or wrong-signed, which would retain the null. If they are predominantly red tape, the delay adds no offsetting value and beta carries the predicted sign. The estimated beta is therefore not a measure of how harmful delay is in principle but a measure of the net character of NASA's actual authorization regime, which is exactly the policy-relevant quantity: an agency does not choose between latency and no latency, it chooses how much of its authorization process to streamline, and the average net effect is what tells it whether streamlining pays. This reframing, which the green-tape theory makes available, is why the dissertation insists that retaining the null is a substantive finding (the regime is on balance value-adding or neutral) rather than a null result in the dismissive sense. Confidence in this interpretive framing is high because it follows deductively from the green-tape and red-tape distinction; the empirical value of beta remains, by design, unknown until the panel is estimated.

### 3.2.5 Measurement instruments and the validated-scale frontier

The field's measurement program produced validated instruments that the dissertation draws on for construct discipline even though it uses a documentary rather than perceptual measure. Borry's three-item red tape scale offered a compact, validated perceptual instrument for measuring red tape, demonstrating that the construct can be captured reliably with few items [\[11\]](#ref-11). A more recent exploratory study by Sarahadil and colleagues on the concept of red tape and efficiency among corporate and government managers reaffirmed that excessive bureaucracy, rigid procedures, and unnecessary administrative burden are associated with slow decision-making and reduced responsiveness, and explored how an efficient bureaucratic culture might be developed [\[70\]](#ref-70). Spenkuch, Teso, and Xu's study of ideology and performance in public organizations, while not a red-tape study per se, supplies a rigorous identified finding directly adjacent to the dissertation's claim: exploiting presidential transitions as a source of within-bureaucrat variation in political alignment, they found that procurement contracts overseen by ideologically misaligned officers exhibit greater cost overruns and delays, consistent with a morale mechanism [\[53\]](#ref-53). The method (within-bureaucrat variation from a transition shock) is a model for the dissertation's own within-program identification, and the finding that an administrative-personnel variable causally moves delays and overruns in federal procurement is strong corroboration that the administrative side of program execution affects the dissertation's outcomes. The interpretation across the measurement-and-evidence strand is that the construct is measurable (validated scales exist), that an identified design can recover an administrative effect on delays and overruns (Spenkuch and colleagues), and that the effect operates through plausible mechanisms (morale, transaction cost); confidence that an administrative-process variable can move cost and delay in federal government is therefore moderate-high, with the caveat that none of this work measures NASA or latency specifically.

### 3.2.6 What the public-administration literature leaves open

This literature supplies the explanatory construct the cost literature lacks and supplies an empirical precedent (Spenkuch and colleagues; Ran and colleagues) that an administrative variable can causally move delays and overruns. What it leaves open is NASA and the long run. No study in this tradition constructs a program-level panel of a single technical agency across its full history, and none operationalizes the specific construct of decision-and-authorization latency from documentary records. The gap is therefore the mirror image of Section 3.1's gap: the right construct exists but has never been applied to the right setting. Confidence in this gap statement is high.

## 3.3 The procurement-competence evidence: the nearest external precedent
### 3.3.1 Why this evidence is load-bearing

The dissertation's causal claim, that longer authorization latency raises cost and schedule, requires a precedent showing that the administrative side of public program execution causally affects cost and schedule rather than merely correlates with them. The procurement literature supplies the single nearest such precedent, and this section treats it as load-bearing rather than illustrative. Bureaucratic competence has been shown to causally reduce delays and cost overruns in government procurement, and this is the strongest available external basis for expecting a latency effect in NASA.

### 3.3.2 The Decarolis instrumental-variable finding

Decarolis, Giuffrida, Iossa, Mollisi, and Spagnolo used United States federal procurement data and an instrumental-variable strategy to show that more competent procurement bureaus cause significant reductions in delays and cost overruns [\[12\]](#ref-12). The method is the one the dissertation will emulate: contract-level data, an explicit instrument to address the endogeneity of competence to contract difficulty, and outcomes (delays, overruns) that map onto the dissertation's schedule and cost outcomes. The finding is causal, not associational, because of the instrument. The limitation, which the dissertation respects, is that the setting is general federal procurement rather than NASA, and that competence is a different construct from latency, though the two are linked (a competent bureau presumably resolves authorizations faster). Decarolis and colleagues establish, with credible identification, that the administrative quality of a procurement organization causally moves the very outcomes the dissertation studies, which transforms the dissertation's hypothesis from speculation into a domain-specific test of a mechanism already demonstrated elsewhere. This is the chapter's strongest single piece of support. Even so, external evidence cannot fix the NASA-specific magnitude; it can only raise the prior that a non-zero effect exists. Confidence that an administrative-organizational variable causally moves federal procurement cost and delay is high; confidence that the same magnitude transfers to NASA authorization latency is low and must be estimated.

### 3.3.3 The methodological lesson the dissertation borrows

Beyond its substantive finding, the procurement-competence evidence supplies a methodological template: the instrument. Because competence and difficulty are jointly determined (hard contracts may be assigned to better bureaus, or may simply take longer), Decarolis and colleagues needed an instrument plausibly unrelated to contract-specific difficulty. The dissertation faces the identical problem one level down. Latency and program-phase difficulty are jointly determined, and the dissertation borrows the identical solution, an instrument (authorizing-office workload; appropriations-calendar timing) plausibly unrelated to the technical difficulty of the specific phase. The Spenkuch within-bureaucrat design [\[53\]](#ref-53) reinforces the template by showing a second route to identification, exploiting an external shock (a presidential transition) that moves the administrative variable without moving program difficulty. The external literature has already solved, twice and credibly, the identification problem the dissertation must solve, so the dissertation inherits a tested identification strategy rather than inventing one. This raises confidence in the feasibility of the dissertation's design from moderate to high, conditional on the instruments being defensible in the NASA setting, which the research-design chapter addresses.

The dissertation cannot simply reuse Decarolis's competence measure, and the reason it must build its own latency variable sharpens the contribution. Competence, as Decarolis and colleagues operationalize it, is a bureau-level attribute inferred from procurement outcomes and personnel characteristics; it is a property of the organization. Latency, as this dissertation operationalizes it, is an event-level attribute measured directly from the elapsed time between documented trigger and documented resolution; it is a property of the decision. The two are related (a more competent bureau presumably resolves authorizations faster, so competence should reduce latency) but they are not the same variable, and the difference matters for policy. A competence finding tells an agency to hire or train better administrators; a latency finding tells an agency which specific authorization steps consume time and are candidates for compression. The dissertation's latency measure is therefore closer to the operational lever a program-execution manager can pull, which is why it constructs the event-level variable rather than importing the organization-level one. The procurement literature supplies the evidence that the channel exists and the template for identifying it; the dissertation supplies the operationally actionable measure the procurement literature did not need.

### 3.3.4 What the procurement literature leaves open

The procurement-competence evidence is the dissertation's nearest external precedent and its methodological template, but it is not the dissertation's test. It does not measure NASA, does not measure authorization latency specifically, and does not span a single agency's full history. Its role, stated precisely, is corroboration and method transfer, not generalization in the reverse direction: the dissertation will use agreement between its NASA-specific finding and the broader procurement result to argue that the mechanism is general while keeping the magnitude NASA-specific. Confidence in this delimitation is high.

## 3.4 The megaproject and optimism-bias literature

### 3.4.1 The bridge between the two literatures

The megaproject literature bridges the NASA cost tradition and the public-administration tradition by establishing three things the dissertation needs: that large public projects overrun systematically and fat-tailedly, that decision processes (not only engineering) cause the overrun, and that a specific confound, optimistic and strategically misrepresented baselines, threatens any naive correlation between process and outcome. This section reviews the literature as a body of corroborative evidence and as a source of one decisive methodological caution. Consistent with the expansion plan's instruction to avoid over-weighting non-aerospace sectors, it treats the large construction-and-energy overrun literature as a body of evidence cited through representative instances rather than enumerated case by case.

### 3.4.2 The systematic-overrun finding and its fat tail

Flyvbjerg and colleagues' synthesis of what is known about cost overrun established that overrun is the statistical norm, not the exception, across large projects, and that the distribution of overrun is fat-tailed, so that average overrun understates the risk of catastrophic overrun [\[13\]](#ref-13). The companion patterns-causes-cures analysis located part of the cause in decision processes and incentives rather than in engineering [\[14\]](#ref-14), and the earlier megaproject-policy-and-planning statement identified pervasive misinformation about costs, benefits, and risks as the central problem, with political-economic explanations (deliberate misrepresentation to win project approval) best accounting for the evidence [\[79\]](#ref-79). Cantarelli and colleagues examined cost overruns in large-scale transportation infrastructure and grounded the explanations in theory [\[15\]](#ref-15). The method across this strand is large-sample comparison of forecast against outturn cost; the finding is systematic overrun with a fat tail, reproduced across samples; the limitation for the dissertation is sectoral, since most of the evidence is transport and construction, not aerospace. The megaproject literature establishes that the problem is material, with very high confidence in the general project population, and supplies a strong prior that NASA, as a producer of bespoke megaprojects, sits in the same fat-tailed distribution. The supporting construction-and-energy overrun evidence (representative instances span Asian infrastructure [\[57\]](#ref-57), low-income-economy infrastructure [\[60\]](#ref-60), and hydropower [\[50\]](#ref-50)) is cited here as a body confirming the generality of systematic overrun, not as aerospace-specific evidence.

Two features of this body of evidence matter for the dissertation beyond the bare fact of overrun, and both are read here as method signals rather than as substantive aerospace findings. First, a recurring result across the cross-sectoral studies is that procurement and stakeholder-coordination deficiencies, not only engineering uncertainty, rank among the leading proximate causes of delay and overrun. Studies of procurement deficiency and resulting delay in public works find that the administrative side of project execution is repeatedly implicated when overrun causes are decomposed [\[82\]](#ref-82), and infrastructure analyses that separate causes attribute a substantial share to planning, procurement, and approval delay rather than to construction difficulty [\[60\]](#ref-60), [\[83\]](#ref-83). The dissertation does not cite these as evidence about NASA; it cites them as a body to establish that, wherever large public projects have been decomposed, an administrative-process channel reliably appears among the causes, which raises the prior that the same channel is present in NASA and is worth measuring directly. Second, the cross-sectoral literature increasingly uses regression and survival methods on project panels, which establishes that the dissertation's analytic approach is the field's normal science rather than a novelty; the methodological novelty the dissertation claims is the variable (latency) and the setting (one agency, long run), not the estimator. Confidence that an administrative-process channel is a general feature of large-public-project overrun is high on the breadth of this evidence; confidence that its NASA-specific magnitude can be read off the cross-sectoral studies is nil, which is why the dissertation measures it rather than importing it.

### 3.4.3 The bespoke-versus-platform comparison using NASA directly

One megaproject study speaks to NASA directly and is therefore weighted more heavily. Ansar and Flyvbjerg's comparison of bespoke versus platform strategies used a reference-class dataset of 203 space missions spanning 1963 to 2021, of which 181 were NASA and 22 were SpaceX, and found that the platform (repeatable) strategy was roughly ten times cheaper and twice as fast as the bespoke strategy, at lower risk of failure [\[59\]](#ref-59). The method is reference-class comparison on a large mission dataset; the finding is that repeatability, an organizational and procurement attribute, dominates bespoke development on cost and speed. This is the megaproject literature's most NASA-relevant result and it directly implicates an organizational variable (bespoke versus platform) in NASA cost and schedule outcomes. The limitation is that the comparison is between two organizations with many confounded differences, so the bespoke-versus-platform contrast cannot be cleanly separated from the NASA-versus-SpaceX contrast. This study raises the prior that organizational-process variables drive NASA cost and schedule outcomes substantially, while the confounding warns that a clean within-NASA design (which the dissertation supplies through program and era fixed effects) is needed to isolate the specific latency channel from the broader organizational contrast. Confidence that an organizational-process variable moves NASA mission cost and speed is, on this evidence, moderate-high.

### 3.4.4 Optimism bias, the planning fallacy, and reference-class forecasting

The behavioral account of why baselines are wrong is central to the dissertation's treatment of confounding. The literature attributes baseline inaccuracy to two distinct sources: cognitive optimism bias (the planning fallacy, in which estimators sincerely underestimate cost and time) and strategic misrepresentation (in which estimators deliberately understate cost to secure approval). Flyvbjerg's adjudication of the planning fallacy against Hirschman's Hiding Hand found that benefit overruns outweigh cost overruns in only about one fifth of projects, so the optimistic-baseline pattern is typical and the benevolent Hiding Hand is not [\[68\]](#ref-68). Flyvbjerg's reference-class-forecasting program proposed correcting baselines by anchoring them to the outturn distribution of comparable past projects rather than to a bottom-up estimate [\[30\]](#ref-30), and the approach has been applied across sectors, with representative applications to Hong Kong roadworks [\[64\]](#ref-64), fusion power-plant cost [\[52\]](#ref-52), and public housing in small island developing states [\[47\]](#ref-47). Lovallo, Cristofaro, and Flyvbjerg's three-stage governing-large-projects framework organized the behavioral and agency causes and the reference-class remedy into a forecasting-organizing-executing process [\[55\]](#ref-55). The dissertation draws a twofold interpretation. Substantively, optimism bias and strategic misrepresentation are the named mechanisms by which baselines become unreliable, which is the dissertation's third rival explanation (optimistic baselines) and the reason it includes a baseline-conservatism check. Methodologically, the dissertation's own design is a within-agency analogue of reference-class thinking: by comparing a program's phases against each other and programs within an era, it anchors its inference to a comparable reference set rather than to a single optimistic point.

The behavioral literature also bears, less obviously, on the dissertation's third outcome, mission cadence, through the platform-versus-bespoke contrast. Ansar and Flyvbjerg's finding that the platform strategy delivered missions twice as fast and ten times cheaper [\[59\]](#ref-59) is, at bottom, a cadence finding: a repeatable platform produces more flight events per period than a sequence of bespoke developments, because each new instance reuses the authorized design rather than re-running the full authorization gauntlet. Read through the dissertation's mechanism, the platform advantage is partly a latency advantage, because a platform amortizes the decision-and-authorization cost across many instances while a bespoke program pays it afresh each time. The dissertation does not test this hypothesis directly (it studies within-NASA latency, not the NASA-versus-platform contrast), but it serves as a coherence check: the dissertation's predicted negative latency-cadence association is consistent with the megaproject literature's strongest NASA-relevant result, which strengthens the case for the cadence prediction while leaving its magnitude to estimation. Confidence that higher per-decision latency depresses cadence is, on this combined reasoning, moderate at the design stage.

### 3.4.5 The contrary evidence on reference-class forecasting and on overrun optimism

Honest review requires the contrary cases, and the corpus contains them. Themsen's study of the processes of public megaproject cost estimation found, contrary to the dominant claim, that reference-class forecasting did not lead to more accurate cost estimates in the cases examined, because the estimation process itself was captured by the same organizational incentives reference-class forecasting was meant to neutralize [\[76\]](#ref-76). Christensen's analysis of cost-overrun optimism, using defense-acquisition data, examined whether program advocacy leads to suppression of adverse cost information and tested the generality of the A-12 program's "abiding cultural problems" against sixty-four completed acquisition contracts [\[72\]](#ref-72). The corrective methods the megaproject literature proposes are themselves vulnerable to the organizational incentives they target, which strengthens the dissertation's decision to measure an objective documentary variable (latency) rather than to rely on estimate-correction methods, and which shows that baseline gaming is a live threat the design must test rather than assume away. Confidence that baseline gaming is a real threat to any process-outcome correlation is, on this evidence, high, which is exactly why the dissertation pre-registers a baseline-conservatism check.

### 3.4.6 What the megaproject literature leaves open

The megaproject literature establishes materiality (systematic fat-tailed overrun), implicates decision processes and organizational form in the cause, supplies a direct NASA-relevant comparison, and names the confounds (optimism bias, strategic misrepresentation) that threaten the dissertation's inference. What it leaves open is the within-NASA, latency-specific, long-run quantification. Its NASA-relevant result [\[59\]](#ref-59) is a between-organization comparison, not a within-NASA panel; its behavioral results are about baseline formation, not about authorization latency; its methodological caution (baseline gaming) is a threat the dissertation must address rather than a test the dissertation can borrow. The gap is therefore that the megaproject literature has done the most to motivate the dissertation's hypothesis and the most to warn against confounding it, without supplying the specific within-agency latency measurement. Confidence is high.

## 3.5 The NASA program-management and management-control literature
### 3.5.1 The institutional texture the quantitative literatures lack

A smaller body of work documents NASA's own reflections on program and project management, and it supplies the institutional texture that the cost-estimating and public-administration literatures abstract away. The claim of this section is modest: this literature establishes that decision-and-authorization processes are a recognized object of management attention inside the agency, which licenses the dissertation's premise that latency is a real institutional phenomenon worth measuring, while contributing little quantification of its own.

### 3.5.2 The Apollo management-control case and the program-control tradition

Tucker and Alewine's study of the roles of management control, drawing lessons from the Apollo program, examined how management-control systems functioned in NASA's most celebrated program and what those systems contributed to its execution [\[25\]](#ref-25). The method is historical-archival case analysis; the contribution is to show that management control, the apparatus of review, authorization, and accountability, was integral to program execution and not incidental to it; the limitation is that a single celebrated case cannot establish the distribution of authorization latency or its association with outcomes across the agency. Hoban and colleagues' "Readings in program control" assembled the agency's own accumulated wisdom on cost estimating, planning and scheduling, cost overruns, and how program-control techniques contributed to a NASA development project's success [\[49\]](#ref-49), and the companion collection on issues in NASA program and project management gathered veteran managers' lessons and guiding principles [\[112\]](#ref-112). The interpretation across this strand is that the agency has long treated decision-and-authorization processes as a controllable management variable, which is the practitioner intuition the dissertation proposes to test; the corpus evidence here establishes that the variable is institutionally real and managerially salient, while supplying narrative rather than measurement. Confidence that authorization-and-control processes are a recognized, managed institutional phenomenon at NASA is high; confidence in any quantified effect from this literature is low because the literature is narrative.

### 3.5.3 The defense-acquisition adjacency

The defense-acquisition literature is adjacent and corroborative. Chadwick's overview of defense acquisition for Congress describes the system of systems, the requirements, resource-allocation, and acquisition processes, through which the Department of Defense plans and authorizes major capability development, tracing its evolution from the Packard Commission and the Goldwater-Nichols reorganization [\[118\]](#ref-118). The interpretation is that defense acquisition is a sibling authorization system whose documented complexity corroborates the dissertation's premise that multi-layer authorization is a real and consequential feature of public technical-program execution; the Christensen defense-acquisition cost-optimism finding [\[72\]](#ref-72) and the Spenkuch federal-procurement finding [\[53\]](#ref-53) are the quantified results adjacent to this institutional description. The limitation is that defense acquisition is not NASA, so it corroborates rather than tests. Confidence in the corroboration is moderate-high.

### 3.5.4 What the NASA program-management literature leaves open

This literature establishes that authorization and management control are real, recognized, managed features of NASA program execution, which is necessary groundwork for the dissertation's premise. What it leaves open is everything quantitative: it does not measure latency, does not assemble a panel, and does not estimate an association. Its role is to ground the construct institutionally, not to test it. Confidence is high.

## 3.6 Synthesis: the unfilled gap and the propositions that follow

### 3.6.1 A complete map of the field

The four bodies of work map onto three branches that, taken together, cover the relevant field and individually leave the dissertation's question open. The technical-and-estimating branch (Section 3.1) measures the outcomes and the technical drivers but omits administrative latency as a regressor. The administrative-process branch (Sections 3.2, 3.3, and the behavioral parts of 3.4) measures the explanatory construct and, in the procurement case, identifies its causal effect on cost and delay, but never in NASA over the long run. The methodological-and-institutional branch (the megaproject confound discussion in 3.4 and the program-management literature in 3.5) supplies the warnings (baseline gaming, optimism bias) and the institutional grounding but neither the outcomes nor the explanatory measure assembled into a test. The branches do not overlap in what they contribute, and together they cover what the dissertation needs, yet their intersection, a within-NASA, long-run, latency-specific, identified panel test, is empty. This is the gap, stated as the empty cell of a complete partition rather than as a rhetorical absence.

The following synthesis table compiles the reading. Each row names a branch, what it contributes, what it leaves open, and the representative corpus evidence, so that the gap is visible as the conjunction of the open columns.

| Branch | What it contributes | What it leaves open | Representative corpus evidence |
|---|---|---|---|
| Technical / estimating (3.1) | The outcomes (cost growth, schedule slip, cadence); the regression-on-program-attributes method; evidence that some non-technical attributes (contract type, manager tenure, policy regime) move outcomes | Administrative latency as a measured regressor; a single consistent 1958-2026 panel | [\[1\]](#ref-1), [\[2\]](#ref-2), [\[3\]](#ref-3), [\[4\]](#ref-4), [\[6\]](#ref-6), [\[19\]](#ref-19), [\[20\]](#ref-20), [\[34\]](#ref-34), [\[90\]](#ref-90), [\[101\]](#ref-101), [\[115\]](#ref-115), [\[130\]](#ref-130) |
| Administrative process (3.2, 3.3) | The explanatory construct (red tape, administrative burden, bureaucratic competence); validated measurement; an identified causal effect on delays and overruns in procurement | Application to NASA; the long run; the specific latency operationalization | [\[7\]](#ref-7), [\[8\]](#ref-8), [\[9\]](#ref-9), [\[10\]](#ref-10), [\[11\]](#ref-11), [\[12\]](#ref-12), [\[53\]](#ref-53), [\[70\]](#ref-70), [\[107\]](#ref-107) |
| Methodological / institutional (3.4, 3.5) | Materiality of systematic fat-tailed overrun; the named confounds (optimism bias, strategic misrepresentation, baseline gaming); institutional grounding of authorization as a managed variable | An assembled within-NASA latency test; quantification of the institutional narrative | [\[13\]](#ref-13), [\[14\]](#ref-14), [\[15\]](#ref-15), [\[25\]](#ref-25), [\[30\]](#ref-30), [\[49\]](#ref-49), [\[55\]](#ref-55), [\[59\]](#ref-59), [\[68\]](#ref-68), [\[72\]](#ref-72), [\[76\]](#ref-76), [\[79\]](#ref-79), [\[112\]](#ref-112), [\[118\]](#ref-118) |

The empty intersection of the three "what it leaves open" columns is the dissertation's contribution space.

### 3.6.2 The named causal mechanism the synthesis implies

The synthesis is not merely that a gap exists but that the literatures, read together, imply a specific causal mechanism that the dissertation will test rather than assume. The mechanism, stated as a chain, is the following. The driver is fragmented decision authority and sequential, multi-layer authorization gates paced by an annual appropriations cycle, a structure documented institutionally in the program-management literature [\[49\]](#ref-49), [\[112\]](#ref-112) and theorized as a transaction-cost structure in the institutional economics the dissertation's framework chapter develops. The mechanism is that each pending action accrues elapsed time, the authorization latency variable, during which standing program costs continue to be incurred and requirements drift, a channel corroborated by the jointness-and-complexity evidence [\[115\]](#ref-115) and the schedule-to-cost linkage [\[130\]](#ref-130). The observable effect is that measured latency per program-phase rises. The operational consequence is higher cost growth and schedule slip and fewer flight or delivery events per period, the dissertation's three outcomes, with the cost and schedule channels supported by the procurement-competence causal evidence [\[12\]](#ref-12), [\[53\]](#ref-53). The strategic implication is that a portion of NASA cost and schedule performance is a controllable process variable, distinct from irreducible engineering difficulty. The synthesis also names where only correlation, not causation, is presently available: most of the NASA-specific evidence (3.1) is associational, the identified causal evidence (3.3) is non-NASA, and the megaproject literature (3.4) warns that baseline gaming can manufacture a spurious process-outcome correlation. The mechanism is therefore stated with calibrated confidence: existence of an administrative effect on cost and delay in government is high-confidence (multiple identified designs), but its NASA-specific magnitude and its specific latency channel are low-confidence at the design stage and are exactly what the test is constructed to estimate.

The synthesis also explains, from the institutional reading the literature supports, why latency should be expected to be durable across NASA's history rather than a transient feature of any one administration, and this durability is what makes the long-run panel both feasible and necessary. Authorization regimes, once established, generate increasing returns: the agency and its contractors adapt their processes to the existing review and approval structure, the cost of altering that structure rises as more actors depend on it, and the regime persists even when a faster arrangement would be available. This is the path-dependence reasoning that the framework chapter develops formally and that the program-management literature corroborates in narrative, where veteran managers describe authorization practices that have outlasted the programs that spawned them [\[49\]](#ref-49), [\[112\]](#ref-112). The methodological consequence for this dissertation is the necessity of era fixed effects: because authorization regimes switch at identifiable points (a reorganization, a procurement-policy reform, an administrative-rule change) and then persist, the era term delta_t absorbs the regime-level component of latency, and the latency coefficient beta is identified from the within-era, within-program variation that remains. Without era fixed effects, a finding that high-latency eras also had high cost growth could merely reflect a common era condition (a period of budget instability, for instance) rather than a latency effect, which is the common-cause rival the discussion chapter weighs. The synthesis therefore does more than locate a gap: it specifies, from the institutional theory the literature supports, both the mechanism to be tested and the identification structure required to test it cleanly, which is the bridge from this chapter's literature reading to the next chapters' measurement and design.

### 3.6.3 The propositions that follow

Two propositions follow from the synthesis, and they are the propositions the rest of the dissertation develops and tests.

The first proposition is a measurement proposition and stands independent of any regression result. A consistent, documentary-rule-based series of NASA decision-and-authorization latency can be constructed across 1958 to 2026 from budget records, NASA Technical Reports Server documentation, and Government Accountability Office assessments, and no such series currently exists in the literatures reviewed. The construction is feasible because every component (baselines, milestones, authorization events) is documented somewhere in the corpus's named sources, and the estimating literature has already demonstrated the ability to assemble historical growth datasets [\[1\]](#ref-1), [\[2\]](#ref-2), [\[3\]](#ref-3). Documentary density changes across eras, which the dissertation's two-resolution (coarse and fine) design accommodates. This proposition satisfies the dissertation's standard that the contribution survives even if the hypothesis is not confirmed, because a consistent long-run latency series is a public good for program-execution management regardless of the regression outcome.

The second proposition is the falsifiable test proposition, stated in the dissertation's fixed notation. For program i, phase p, and era t, the relationship \(\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t)\) can be estimated with program fixed effects \(\alpha_i\) and era fixed effects \(\delta_t\), and the literature gap implies that this estimation has never been performed for NASA. Under the alternative hypothesis, \(\beta\) is positive for cost growth and schedule slip and negative for cadence; under the null, \(\beta\) is statistically indistinguishable from zero. The synthesis supplies the case for expecting a non-zero \(\beta\) (the procurement-competence and ideology-misalignment identified findings, the jointness mechanism, the red-tape-reform quasi-experiment) and supplies, with equal force, the reasons for taking the null seriously (the conditional and heterogeneous red-tape effects, the green-tape distinction, and the baseline-gaming confound). The chapter has held to one point throughout: this is a directional test of an association under an identification strategy, not a foregone confirmation: the green-tape theory [\[10\]](#ref-10) and the conditional Brewer-Walker finding [\[8\]](#ref-8) mean that retaining the null would be a substantive result, namely that NASA authorization latency, once program difficulty and era are accounted for, does not move the outcomes, and that reform should redirect to the technical and estimating factors the literature of Section 3.1 documents.

The literature reviewed here also fixes, in advance, the conditions under which the second proposition would be judged false, which is the discipline that distinguishes a pre-registered test from a search for a confirming specification. Three of those conditions are drawn directly from the cautions the literature supplies. First, the megaproject literature's documentation of baseline gaming and strategic misrepresentation [\[68\]](#ref-68), [\[76\]](#ref-76), [\[79\]](#ref-79) means that an apparent latency effect must survive a check that it is not an artifact of optimistically set baselines; if baseline conservatism, not latency, explains the outcome, the contribution fails. Second, the public-administration literature's finding of conditional and sometimes reversed red-tape effects [\[8\]](#ref-8), [\[10\]](#ref-10) means that a latency coefficient that vanishes or reverses under reasonable alternative specifications cannot be treated as confirmation. Third, the procurement and ideology evidence [\[12\]](#ref-12), [\[53\]](#ref-53) establishes that credible identification of an administrative effect requires an instrument or a within-unit shock, so a latency association that holds under fixed effects but collapses once the endogeneity of latency to phase-specific difficulty is addressed would fail the standard those studies set. The dissertation accordingly treats the literature not only as support for its hypothesis but as the source of the binding falsification conditions, which is the most demanding use a literature review can be put to: the same studies that make the latency effect plausible also specify exactly what evidence would refute it. This is the discipline carried to its conclusion. The design improves on the alternatives only because the alternatives (baseline gaming, conditional effects, endogeneity) are named from the literature and pre-committed as falsification routes rather than discovered after the fact.

### 3.6.4 How the gap positions the dissertation

The literature review's closing claim is that the dissertation occupies a position defined precisely by the non-overlap of two mature literatures and bridged by a third. It inherits the outcomes and the regression method from the NASA cost literature, the explanatory construct and the identification template from the public-administration and procurement literatures, the materiality and the confound catalog from the megaproject literature, and the institutional grounding from the NASA program-management literature. It contributes the one thing none of them contributes: a consistent long-run NASA latency series and a pre-registered, identified, heterogeneity-robust panel test of whether that latency moves cost growth, schedule slip, and mission cadence. The argument the dissertation carries is, by the end of this chapter, established on its first two points (the problem is real, established in 3.1.5 and 3.4.2; the problem is material, established in 3.1.5 and 3.4.2) and is set up on the remaining three (that the design addresses the mechanism, improves on the alternatives, and manages residual risk), which the data, measurement, research-design, and analysis-plan chapters develop. The honest design-stage posture is maintained without exception: no estimate is reported here as executed, every expected sign is labeled as a directional prediction under the alternative hypothesis, and the conditional and contrary evidence is given its full weight so that the null remains a genuine scientific possibility rather than a straw figure to be knocked down.

The chapter therefore closes where it opened: the relevant literature is not thin but disjoint, the gap is the empty intersection of a complete partition of the field, and the two propositions that follow, the constructible latency series and the falsifiable panel test, are the contributions the dissertation now proceeds to design and defend.
## Chapter 4. Data and Measurement

## 4.0 The chapter thesis, stated first

A consistent, documentary-rule-based series of NASA decision-and-authorization latency can be built from three named bodies of primary record, and every variable in the panel test can be operationalized by a single rule applied identically from 1958 to 2026. That is the claim this chapter defends. The measurement is the load-bearing contribution of the dissertation, prior to and independent of any estimate the regression eventually returns. If the panel is later assembled and the latency coefficient turns out to be statistically indistinguishable from zero, the null is retained and the causal hypothesis fails, but the constant-unit latency series remains a usable object that did not previously exist. The standard this chapter holds itself to is therefore not the standard of a results chapter, which reports what was found, but the standard of a measurement chapter, which establishes that what will be measured is the same thing in every era and that the rule producing it is transparent enough for another researcher to reproduce. The discipline is Maddison's: no two periods can be compared until a transparent, replicable measurement standard exists, expressed in constant units, so that one period can be set against another on the same footing [\[16\]](#ref-16), [\[26\]](#ref-26).

The problem this chapter addresses is concrete. The agency's documentary record is real, voluminous, and primary, but it was never created to support a longitudinal measure of authorization latency. Budget justifications, program commitment documents, Standing Review Board products, and Government Accountability Office assessments each record decision and authorization events, but they record them in formats that changed across seventy years, at densities that rose sharply over time, and under baseline definitions that were tightened repeatedly. The current state is a record that is rich but heterogeneous; the desired state is a single program-phase panel in which latency, cost growth, schedule slip, and mission cadence each mean exactly one thing; the gap is the set of construction rules that convert the first into the second; and the consequence of leaving the gap open is that any number anyone reports for NASA authorization latency is, at present, an anecdote with a date attached rather than a measurement. This chapter closes that gap at the level of design. It names each dataset and states its provenance, access, coverage, unit of analysis, and known biases; it operationalizes every variable in a measurement table; and it specifies the data-quality, validation, and ethics procedures that make the resulting series defensible. Consistent with the honest design-stage posture that governs the whole dissertation, no constructed series is reported here as if executed. Every illustrative interval or example value is labeled as such.

Two scope decisions carry forward from the bible and are restated here so the chapter cannot drift. First, the unit of analysis is the program-phase observation, the panel dimension is the program indexed by i, and the time dimension is the phase sequence p and the era t, exactly as in the specification \(\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t)\). Second, architecture-traceability vocabulary is out of scope. This is a cliometric econometric study whose objects are programs, phases, documentary events, and panel estimates. No capability, system, or data-service exchange is being fielded, so no DoDAF or BEA traceability table appears in this chapter, and none should be forced onto it.

## 4.1 The named datasets

The study draws on three documented bodies of primary record. Each is described below by provenance, the mechanism of access, coverage in time and program scope, the unit at which it records information, and the biases a careful constructor must correct for. The three are complementary rather than redundant: the agency budget and program records supply the baselines and the decision-event trail, NTRS supplies the technical parameters and much of the process documentation, and the GAO assessments supply an independent, consistently formatted measure of cost and schedule performance for the modern portion of the panel.

### 4.1.1 NASA historical budgets and program records

**Provenance.** The first dataset is the agency's own financial and programmatic record: annual budget estimates and the congressional justifications that accompany them, program and project commitment documents, and the lifecycle-review products that document formal decision events. These are the records in which a program's baseline cost and schedule are established at its formal commitment point and against which its realized cost and schedule are later booked. They are also the records in which the decision and authorization events live: key decision points, program commitment reviews, confirmation reviews, and the budget actions that authorize each successive phase. The governing process documents that define what these events are and when they occur are themselves in the documentary record. The agency's lifecycle-review architecture is specified in the Standing Review Board Handbook and in the program-management procedural requirements whose practical application is described in the NTRS record on reducing NPR 7120.5D to practice and in the earlier account of the NPG 7120.5A electronic review process [\[80\]](#ref-80), [\[94\]](#ref-94), [\[95\]](#ref-95). The Joint Confidence Level policy, which governs how cost and schedule commitments are set at the seventy-percent confidence level for many modern projects, supplies both a baseline-definition rule and a dated authorization event [\[86\]](#ref-86).

**Access.** Budget justifications and many commitment documents are public, distributed through the agency's budget pages and through congressional records; older material is concentrated in agency history collections and the National Archives. Access is therefore uneven: the modern record is online and machine-readable in part, the mid-period record is in scanned documents of variable quality, and the earliest record requires archival retrieval. This unevenness is not a defect to be hidden but a coverage fact to be carried explicitly into the two-resolution latency design of Section 4.3.

**Coverage.** In principle the budget and program record spans the full 1958-to-2026 horizon, because the agency has produced budget estimates every year of its existence. In practice the density and formality of the decision-event record rise sharply over time. The formal key-decision-point taxonomy and the program commitment review are creatures of the modern program-management procedural regime; the earliest decades recorded decisions in less standardized forms. Coverage of the explanatory variable is therefore dense and fine in the modern subperiod and sparse and coarse in the early decades, which is exactly the asymmetry the coarse-versus-fine latency measure is built to accommodate.

**Unit of analysis.** These records are organized by program and by fiscal year, and within a program by lifecycle phase and review milestone. They map cleanly onto the program-phase observation that is the panel's unit, because baselines and actuals are booked at defined milestones rather than only at program end, which is the same lifecycle logic the cost-estimating literature uses to measure growth between milestones [\[1\]](#ref-1), [\[2\]](#ref-2), [\[3\]](#ref-3).

**Known biases.** Three biases matter. First, baselines are chosen by the agency and may be set optimistically, a selection problem the megaproject literature documents in detail and which the reference-class-forecasting program was designed to counter [\[13\]](#ref-13), [\[30\]](#ref-30). An optimistically low baseline inflates measured growth without any change in the underlying program, so baseline choice must be treated as endogenous, not as ground truth. Second, the budget record reflects requested-versus-appropriated dynamics that are partly exogenous to the program, so a phase can slip for appropriations reasons that have nothing to do with the program's own decision processes; this is the motivation for the funding-instability control of Section 4.5 and for the appropriations-calendar instrument discussed in the design chapter. Third, decision events are not always dated precisely in the early record, so some latency values must be bounded rather than point-identified, a fact taken up in Section 4.7.

### 4.1.2 NASA Technical Reports Server (NTRS)

**Provenance.** The second dataset is the NASA Technical Reports Server, the agency's openly accessible repository of technical reports, conference papers, and management analyses. NTRS holds the documentary trail that supplies both technical parameters for the control vector and process documentation for the latency construction. The cost-estimation and control records, the lifecycle-cost-analysis-model documentation, and the operations-support estimation reports are representative of the kind of primary record NTRS makes available for variable construction [\[65\]](#ref-65), [\[102\]](#ref-102), [\[122\]](#ref-122). The Standing Review Board Handbook and the NPR 7120.5 practice records, already cited above, are NTRS holdings as well [\[80\]](#ref-80), [\[94\]](#ref-94), [\[95\]](#ref-95). The more recent objectives-driven, risk-informed, case-assured approach to safety and mission success documents the contemporary review philosophy under which modern decision events are generated [\[106\]](#ref-106).

**Access.** NTRS is a public web repository with a citation API and bulk-export capability, which makes it the most directly machine-accessible of the three datasets. Each holding carries a stable citation identifier; the corpus records the NTRS citation URLs for the process documents used here. Access is therefore reliable and reproducible, which is why NTRS is the backbone of the technical-control extraction and a major source for the process taxonomy.

**Coverage.** NTRS coverage is broad across the agency's history but uneven by program and by document type. Cost-estimation methodology reports, systems-engineering lessons-learned, and review-process documentation are well represented; the granular, dated decision-event records needed for the fine latency resolution are concentrated in the modern era, consistent with the budget-record asymmetry above. NTRS coverage of technical parameters such as mass, power, and technology readiness is strongest for science instruments and spacecraft, which is the same population the Aerospace Conference cost-growth datasets cover, enabling the validation cross-check of Section 4.6 [\[2\]](#ref-2), [\[3\]](#ref-3), [\[130\]](#ref-130).

**Unit of analysis.** NTRS records are documents, not program-phases. The constructor's task is therefore an extraction task: each relevant document is read to yield either a technical parameter attached to a program-phase or a dated decision event attached to a program-phase. The document is the source unit; the program-phase is the target unit; the codebook of Section 4.3 specifies the mapping.

**Known biases.** NTRS holdings are self-selected by what the agency chose to document and release. Programs that produced more formal technical reporting are over-represented relative to programs that did not, which can correlate with program size and era. This is a coverage-density bias rather than a measurement bias on any single observation, and it is handled by the unbalanced-panel structure and by reporting results at both latency resolutions so that a finding cannot be an artifact of the denser modern documentation. A second bias is that lessons-learned and methodology reports are written with a purpose, often to advocate a practice, so their narrative framing of decision speed must be treated as context, not as a latency measurement; only dated events are coded.

### 4.1.3 Government Accountability Office major-project assessments

**Provenance.** The third dataset is the Government Accountability Office's recurring assessment of NASA major projects, an independent series that reports cost and schedule performance against baseline for the agency's largest projects. Its value is precisely its independence and its formatting stability: the GAO applies a consistent definition of baseline and of growth across report years, which makes the series a clean external anchor for the modern portion of the panel and a cross-check on cost and schedule growth derived from agency records.

**Access.** The GAO assessments are public and distributed as numbered reports through the GAO website. They do not carry digital object identifiers; they are cited by GAO report number and gao.gov URL. This is an honest gap in the present corpus, which is built around DOI-resolvable records: the GAO NASA major-project assessment series is named in the dissertation charge and used here as a named dataset, but it is not yet represented as discrete citable entries in research/corpus.jsonl. Before the panel is assembled, at least two specific GAO NASA assessment reports must be added to the corpus by report number and URL, together with the GAO methodology appendix that defines its baseline and growth conventions, so that the validation step of Section 4.6 cites the precise documents it reproduces. The dissertation flags this as the highest-priority corpus-completion item and does not paper over it with a substitute citation.

**Coverage.** The GAO series covers the agency's largest projects in the modern era, beginning well after the agency's founding, and it does not reach the early decades at all. Its coverage is therefore a dense, high-quality slice of the modern subperiod rather than a full-history source. This is complementary to the budget and NTRS records, which reach back further but at lower and more variable resolution. The panel is unbalanced by construction in part because the GAO anchor exists only for the recent portion.

**Unit of analysis.** The GAO reports at the project level and, within a project, against the baseline established at the project's commitment point. This maps onto the program-phase observation for the development phase of major projects, which is the phase the GAO most consistently tracks. It does not by itself supply the multi-phase structure for older or smaller programs, which must come from the budget and NTRS records.
**Known biases.** Two biases attend the GAO series. First, it covers only major projects above a reporting threshold, so it is a selected, large-program sample; using it as the sole source would bias the panel toward big programs, which is why it anchors and cross-checks rather than defines the panel. Second, the GAO's baseline is the agency's committed baseline, so the optimistic-baseline selection problem flows through the GAO numbers as well. The GAO's independence corrects for measurement and reporting error, not for the agency's baseline-setting incentives, which remain a threat addressed in Section 4.7 and in the design chapter's baseline-conservatism test.

### 4.1.4 How the three datasets combine

The three sources are joined on the program-phase key. For a given program-phase, the budget and program record supplies the baseline and actual cost and schedule and the dated decision events; NTRS supplies the technical parameters and, in the modern era, additional dated review events; the GAO assessment, where it exists, supplies an independent confirmation of cost and schedule growth. Where two sources report the same quantity, the constructor records both and reconciles them by a fixed precedence rule stated in the codebook, with disagreements logged rather than silently averaged. This triangulation realizes the validation-against-known-values discipline: the panel is built so that, for the overlapping modern programs where the GAO and Aerospace Conference figures exist, the constructed cost and schedule growth can be checked against published values before any estimate is attempted [\[1\]](#ref-1), [\[3\]](#ref-3), [\[130\]](#ref-130). Confidence in the dataset architecture is high, because all three sources are primary, public, and independently produced, and because the join key is the same lifecycle milestone structure the cost-estimating community already uses. Adding the GAO discrete records and a documented New Start Inflation Index source to the corpus would raise that confidence further; the discovery that early-record decision events are too sparsely dated to support even the coarse latency resolution would lower it, a risk the next section confronts directly.

## 4.2 Unit of analysis and the unbalanced panel

The unit of analysis is the program-phase observation. Each NASA program contributes one observation per major lifecycle phase for which a baseline and an actual are documented, so a program with formulation, development, and operations phases contributes three observations, and a program documented only at the development phase contributes one. The panel dimension is the program i; the time dimension is the phase sequence p within a program and the calendar era t in which each phase milestone falls. This is the structure the specification assumes, and choosing the program-phase rather than the whole program as the unit is a deliberate identification choice, not a convenience.

The program-phase unit is what makes the within-program identification in the fixed-effects design possible, and the reasoning is mechanical. Program fixed effects alpha_i absorb everything fixed across a program's life, including the part of program difficulty that does not change from phase to phase. That absorption is useful only if a program contributes more than one observation, because a program that appears once provides no within-program variation for alpha_i to condition on. The standard panel-data result is that fixed effects identify a coefficient from within-unit variation, so the unit must be chosen at a level that delivers repeated observations per unit [\[93\]](#ref-93), [\[116\]](#ref-116), [\[126\]](#ref-126), [\[128\]](#ref-128). The cost-estimating literature's own lifecycle logic reinforces the choice, since it measures growth between defined milestones rather than only at program end, so the phase-level observation is the natural one for this domain [\[1\]](#ref-1), [\[2\]](#ref-2), [\[3\]](#ref-3). One consequence matters: a program observed in only one phase still enters the panel and contributes to the era comparison through delta_t, but it does not contribute to within-program identification of beta, so the effective sample for the within-program channel is smaller than the raw observation count. A critic might object that phases within a program are not independent, since a slip in formulation can mechanically push development; the response is that standard errors are clustered by program precisely to allow arbitrary within-program correlation across phases, so the dependence is accommodated rather than assumed away.

The panel is unbalanced by construction. Three structural facts produce the imbalance. First, programs differ in how many phases they ran and how many of those phases left a documented baseline and actual, so the number of observations per program varies. Second, the documentary density rises over time, so modern programs contribute more, and more finely dated, observations than early ones. Third, the GAO anchor exists only for major projects in the modern era, so the densest, most externally validated observations are concentrated in the recent decades. The study treats the imbalance as a feature to be reported, not a flaw to be balanced away by dropping observations, because dropping early or sparse observations to force balance would discard exactly the long-run variation the cliometric design exists to exploit. The cost of retaining the imbalance is borne in the inference: the modest number of program clusters and the uneven cluster sizes require the wild-cluster bootstrap and the alternative-clustering checks specified in the design chapter, so that conclusions are not driven by a few large modern programs.

## 4.3 Constructing authorization latency

Authorization latency is the explanatory variable, and its construction is the chapter's most consequential measurement task, because the entire contribution rests on whether latency can be measured by one rule across seventy years. The construct is defined exactly as in the bible: latency is the elapsed time, in months, between a documented trigger event, the point at which a decision or authorization becomes due, and the documented authorization event that resolves it; for each program-phase, latency is taken as the median elapsed time across the authorization events in that phase. The median rather than the mean is chosen because the distribution of authorization intervals is expected to be right-skewed, with a few very long waits, and the median is the more stable central measure under skew and under the bounded-event problem discussed below.

### 4.3.1 The single documentary rule

The governing principle is Maddison's single-rule standard: the same operational definition of a trigger event and a resolution event must be applied identically in 1962 and in 2024, or the resulting series is not comparable across eras [\[16\]](#ref-16), [\[26\]](#ref-26). The rule is stated as a taxonomy. A trigger event is any documented point at which an authorization becomes due: the completion of a lifecycle review that gates the next phase, a program commitment review that must be ratified before commitment, a confirmation review, a rebaselining trigger, or a budget action that requires authorization before funds obligate. A resolution event is the documented authorization that closes the trigger: the signed decision memorandum, the approved key decision point, the ratified commitment, or the enacted budget authority. The latency for a trigger-resolution pair is the months between them; the latency for a program-phase is the median across the pairs in that phase. The events themselves are defined by the agency's own process documents, so the rule is anchored to the institution's stated procedure rather than to the constructor's judgment: the Standing Review Board Handbook defines the review architecture, the NPR 7120.5 practice records define the lifecycle reviews and their gating role, and the Joint Confidence Level policy defines the commitment-setting event for modern projects [\[80\]](#ref-80), [\[86\]](#ref-86), [\[94\]](#ref-94), [\[95\]](#ref-95). Older eras are mapped onto the same taxonomy by identifying their functional equivalents, the milestone reviews and authorization actions that played the gating role before the modern nomenclature existed, with the mapping recorded explicitly in the codebook so a reader can see how a 1965 milestone was classed as the equivalent of a modern gating review.

### 4.3.2 The two-resolution design

Because the documentary detail available in 1962 differs from that available in 2024, a single resolution cannot serve the full span without either discarding the early decades or pretending to a precision the early record does not support. The study therefore constructs latency at two resolutions and reports results separately at each. The coarse measure is defined from milestone-to-milestone intervals and is available for the full 1958-to-2026 span: it uses the major phase-gating milestones that are documented even in the early record, so it can be computed everywhere but at lower temporal precision. The fine measure is defined from individual key-decision-point records and is available only for the modern subperiod where those records exist: it resolves latency at the level of individual authorization actions within a phase. The claim that the two-resolution design protects the contribution rests on a specific logic. If a latency effect appears only in the fine measure and the modern subperiod, it could be an artifact of denser documentation in recent decades; if it appears in the coarse measure across the full span as well, that artifact explanation is much weaker. Reporting both resolutions is therefore a falsification test built into the measurement, not merely a robustness courtesy, and the bible's falsification clause makes a result that vanishes under the alternative resolution a condition for retaining the null. Confidence in the coarse measure is moderate for the early decades, because milestone dates are recoverable but baselines were less formal then; confidence in the fine measure is high for the modern subperiod, because key-decision-point records are dated and standardized. The evidence that would raise early-decade confidence is the recovery of additional dated milestone records from archival budget documents; the evidence that would lower it is the discovery that early milestones cannot be consistently classed against the modern taxonomy.

### 4.3.3 The bounded-latency case

Some decision and authorization events are recorded without precise dates, especially in the early record. For these, latency cannot be point-identified; it must be bounded. The study assigns each such event a lower and an upper bound from the surrounding dated events, in the manner of Fogel's bounded social-saving estimate, which reports a range conditioned on stated assumptions rather than a single point when the counterfactual cannot be pinned exactly [\[17\]](#ref-17). A program-phase whose median latency depends on a bounded event therefore carries a bounded latency value, and the analysis chapter propagates these bounds into the reported estimates as ranges rather than collapsing them to a midpoint that would assert a false precision. This measurement decision has a direct inferential payoff: it keeps the early observations in the panel, preserving the long-run span, while making the resulting imprecision visible rather than hidden.

### 4.3.4 Endogeneity is a design problem, not a measurement problem

A caution belongs here so the measurement is not over-claimed. Measuring latency cleanly does not make it exogenous. The central threat, that harder phases both take longer to authorize and overrun more, is a property of how latency is generated, not of how it is measured, and it is handled in the design chapter by program and era fixed effects and by the instrumental-variable strategy in the spirit of Decarolis and colleagues, who use a comparable approach to isolate the administrative channel in federal procurement [\[12\]](#ref-12). The measurement choice that supports that identification is to record, for each authorization event, the point within the phase at which the trigger arose, so that latency measured early in a phase, before trouble has had time to accumulate, can be distinguished from latency measured late. Early-in-phase latency is less likely to be a consequence of an unfolding overrun and is therefore the cleaner input to the causal test. The measurement chapter's job is to make that distinction recordable; the design chapter's job is to use it.

## 4.4 Constructing the outcomes

The three outcomes are cost growth, schedule slip, and mission cadence. Each is defined by a fixed rule, and the cross-era comparison rule, constant-dollar deflation, is stated wherever cost enters.

### 4.4.1 Cost growth

Cost growth for a program-phase is the actual phase cost minus the baseline phase cost, divided by the baseline, with both figures expressed in constant fiscal-year dollars. The deflation is not optional and is not cosmetic: comparing a 1965 overrun to a 2015 overrun in nominal dollars would conflate inflation with growth and would violate the constant-unit standard the entire measurement rests on [\[16\]](#ref-16). The deflator is the NASA New Start Inflation Index, the agency's standard index for placing aerospace program costs on a common real footing. An honest corpus note is required: the New Start Inflation Index is referenced operationally throughout the cost-estimating literature used here but is not yet present in research/corpus.jsonl as a discrete citable source, and the index source document, by NASA index version and URL, must be added before the cost-growth series is finalized so that the deflation is fully reproducible. The cost figures themselves come from the budget and program record and, for modern major projects, are cross-checked against the GAO assessment and against the published Aerospace Conference cost-growth datasets for overlapping programs [\[1\]](#ref-1), [\[3\]](#ref-3), [\[77\]](#ref-77), [\[130\]](#ref-130). Confidence in the cost-growth measure is high in the modern subperiod, where multiple independent sources can be reconciled, and moderate in the early decades, where the baseline definition was less formal and fewer cross-checks exist.

### 4.4.2 Schedule slip

Schedule slip for a program-phase is the actual phase duration minus the baseline phase duration, divided by the baseline duration, measured in months. Schedule is the outcome most directly connected to latency, because authorization latency is itself a component of elapsed schedule, which is why the bible expects the schedule-slip coefficient to be the largest and most robust of the three under H1. This proximity is a measurement hazard as well as a substantive expectation: if latency and schedule slip share a mechanical component, the schedule-slip regression risks regressing a quantity partly on itself. The measurement response is to define the baseline phase duration from the committed schedule at the phase's commitment point and to define latency from authorization intervals that are conceptually distinct from the planned phase work, so that the two are not constructed from the same arithmetic. The schedule figures come from the same milestone record that supplies the latency events, and the relationship between schedule delay and cost overrun documented by Majerowicz and Shinn is the substantive reason schedule is carried as a distinct outcome rather than folded into cost [\[5\]](#ref-5).

### 4.4.3 Mission cadence

Mission cadence is the most construct-fragile outcome and is measured at two levels. At the era level, cadence is the number of operational mission events per unit time within a defined program family; at the program level, it is the interval between successive flight or delivery events. The fragility is definitional: cadence depends on how a program family is bounded, and no single family definition is uniquely correct. A "Mars program" cadence and a "planetary science" cadence are different numbers computed from overlapping events. The study's response is to report cadence under multiple family definitions rather than to privilege one, so that any cadence finding can be inspected for sensitivity to the family boundary. This is the same multiple-definition discipline the bible requires, and it is the reason the falsification clause names a cadence result that survives only under one family definition as a fragile result. There is no single published NASA cadence series by family; the cadence variable is constructed here from flight and delivery records under the alternative family definitions, which is a definitional construction task rather than a citation gap. Confidence in the cadence measure is moderate at best, lower than for cost and schedule, and this is stated rather than smoothed over: cadence is reported as the most tentative of the three outcomes, and a cadence result is weighted accordingly in the discussion.
## 4.5 Controls

The control vector X separates the latency channel from the technical and programmatic factors the existing literature already credits with moving cost and schedule. Two families of control enter.

The technical controls follow the cost-estimating literature directly. They are instrument or spacecraft mass and power, technology readiness level at the commitment point, mission class, and contract type. These are the regressors the parametric cost-estimating relationships use, and including them is what lets the latency coefficient be read as conditional on technical difficulty rather than as a proxy for it [\[1\]](#ref-1), [\[2\]](#ref-2), [\[3\]](#ref-3), [\[19\]](#ref-19), [\[20\]](#ref-20). Mass and power are recovered from NTRS technical records and the Aerospace Conference datasets; technology readiness at commitment is recorded from the commitment-review documentation; mission class and contract type are recorded from the program record, with contract type mattering because the firm-fixed-price literature finds contract structure itself bears on cost growth [\[20\]](#ref-20), [\[85\]](#ref-85). These controls are chosen because they are the variables the domain's own estimating models treat as the primary technical drivers, so omitting them would leave the latency coefficient open to the charge that it is absorbing technical difficulty. Technology readiness and mission class are coarser in the early record, so their measurement precision, like latency's, is era-dependent.

The programmatic controls are the number of external partners and a funding-instability index. The partner count captures the coordination load of multi-organization programs, a dimension the jointness literature shows can drive cost independently of the engineering, as in the NPOESS case where institutional jointness carried its own cost penalty [\[115\]](#ref-115). The funding-instability index is constructed from year-over-year deviations between requested and appropriated funds, drawn from the budget and appropriations record, and it absorbs the part of schedule slip and cost growth caused by the appropriations cycle rather than by the program's own decisions [\[88\]](#ref-88), [\[129\]](#ref-129). This control does double duty. Substantively it removes a confounder; methodologically it is the conceptual neighbor of the appropriations-calendar instrument the design chapter uses, since the timing of authorization relative to the appropriations calendar paces decisions for budget-dependent reasons that are plausibly independent of a specific phase's technical difficulty. The mechanism is explicit. Driver: an annual appropriations cycle subject to continuing resolutions and delayed enactment. Mechanism: budget actions that gate a phase wait on enacted authority, accruing elapsed time independent of engineering. Observable effect: measured latency and schedule slip rise in unstable-funding years. Operational consequence: cost and schedule growth that the funding-instability control attributes to the budget environment rather than to the agency's discretionary decision speed. Naming this mechanism keeps the latency coefficient from silently claiming credit for delay that the appropriations process, not internal authorization, produced.

## 4.6 Coverage, the changing documentary baseline, and validation against known values

Coverage is the full set of NASA programs with documented baselines from agency inception in 1958 through 2026. It is dense and high-resolution for the period covered by the GAO major-project assessments and by the Aerospace Conference cost-growth datasets, and sparser and coarser for the earliest decades, where baselines were defined less formally. The panel is unbalanced by construction for the reasons given in Section 4.2.

The central coverage hazard is that the definition of a baseline changed over time, becoming more formal and more conservative as the program-management procedural regime matured and as the Joint Confidence Level policy moved commitment-setting toward an explicit confidence standard [\[86\]](#ref-86), [\[95\]](#ref-95). A baseline set informally in 1965 and a baseline set at the seventy-percent joint confidence level in 2015 are not the same object, and a cost-growth number computed against each is not strictly comparable. The study's response is the two-resolution latency design combined with explicit era fixed effects: delta_t absorbs the common shift in baseline-setting practice within an era, so the latency coefficient is identified from variation net of the era-wide change in how baselines were defined. This is the measurement-side justification for a design-side choice, and it is the reason the chapter and the design chapter are tightly coupled.

Validation against known values is the quality gate that turns the constructed series from an assertion into a defensible measurement. The procedure is to reproduce published figures where published figures exist before trusting the constructed figures where they do not. For the overlapping modern programs, the constructed cost growth and schedule growth must match, within a stated tolerance, the values published in the GAO assessments and in the Aerospace Conference cost-growth datasets [\[1\]](#ref-1), [\[3\]](#ref-3), [\[130\]](#ref-130). Where the constructed value and the published value disagree beyond tolerance, the discrepancy is investigated and resolved before the observation is admitted, and the resolution is logged. This is the operational meaning of the describe-before-estimate and reproduce-known-values discipline: the panel earns trust on the cases that can be checked, which licenses its use on the cases that cannot. The claim that validation supports the early, uncheckable observations is qualified honestly. Validation can confirm that the construction rule reproduces known modern values, which raises confidence that the same rule applied to early observations is at least internally consistent, but it cannot independently verify an early value for which no published comparison exists. That residual uncertainty is carried as the bounded-latency range and the moderate-confidence label on the early decades, not hidden.

## 4.7 Limitations

Four data limitations are acknowledged at the outset, each paired with the construction response that contains it.

First, the documentary definition of a baseline changed over time, so the earliest observations carry more construction uncertainty than the modern ones. The two-resolution latency measure and the era fixed effects are the responses; the residual uncertainty is reported as the moderate-confidence label on the early decades rather than smoothed into a false uniformity [\[16\]](#ref-16).

Second, decision and authorization events are not always recorded with dates, so latency for some program-phases must be bounded rather than point-identified. The bounded-latency procedure of Section 4.3.3 keeps these observations in the panel while making their imprecision visible, in the Fogel bounded-estimate manner, and the analysis propagates the bounds into reported ranges [\[17\]](#ref-17).

Third, cost and schedule baselines are chosen by the agency and may be set optimistically, so measured growth can reflect baseline choice rather than process or engineering. This is the selection problem the megaproject and optimism-bias literatures document, and it is the reason baseline choice is treated as endogenous and tested by the baseline-conservatism check in the design chapter, which examines whether the ratio of held reserves to baseline correlates with latency [\[13\]](#ref-13), [\[14\]](#ref-14), [\[30\]](#ref-30). The limitation is real and is not fully resolved by measurement alone; it is bounded by the design.

Fourth, cadence depends on program-family definition, and no single definition is uniquely correct, so cadence is the most construct-fragile outcome. Reporting cadence under multiple family definitions is the response, and a cadence result that holds under only one definition is, by the bible's falsification clause, treated as fragile rather than confirmatory.

A fifth, corpus-level limitation is stated for completeness and transparency. Two named sources used operationally in this chapter are not yet discrete, citable entries in the corpus: the GAO NASA major-project assessment series, which carries GAO report numbers rather than DOIs, and the NASA New Start Inflation Index source document. Both must be added before the data chapter and the backmatter are finalized, the GAO series with at least two specific reports and its methodology appendix, the index with its NASA version and URL. Naming this gap here, rather than substituting a convenient citation, is itself part of the measurement discipline: the series is only reproducible if the exact source of the deflator and the exact GAO reports being reproduced are pinned in the bibliography.

## 4.8 Measurement table

The following table operationalizes every variable in the panel. Each row states the construct, its operational definition, its primary source, and its scale of measurement. The table is the chapter's single reference object for the design and analysis chapters that follow; the notation matches the specification \(\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t)\) exactly. No value in this table is an estimate; the table defines how each quantity is measured, not what it equals.

| Construct (role) | Operational definition | Primary source | Scale |
|---|---|---|---|
| Authorization latency (explanatory; Latency) | Median elapsed months between a documented trigger event (authorization becomes due) and the documented resolution event, across a program-phase's authorization events; coarse (milestone-to-milestone, full span) and fine (key-decision-point records, modern subperiod) resolutions reported separately; bounded where events are undated | Budget and program records; NTRS process records [\[80\]](#ref-80), [\[86\]](#ref-86), [\[94\]](#ref-94), [\[95\]](#ref-95) | Continuous, months (>= 0); bounded interval where undated |
| Cost growth (outcome 1) | (actual phase cost - baseline phase cost) / baseline phase cost, both in constant FY dollars via the NASA New Start Inflation Index | Budget and program records; GAO assessments; Aerospace Conference datasets [\[1\]](#ref-1), [\[3\]](#ref-3), [\[77\]](#ref-77), [\[130\]](#ref-130) | Continuous, fraction of baseline (real) |
| Schedule slip (outcome 2) | (actual phase duration - baseline phase duration) / baseline phase duration | Milestone record in budget and program records [\[5\]](#ref-5) | Continuous, fraction of baseline duration (months-based) |
| Mission cadence (outcome 3) | Era level: operational mission events per unit time within a program family; program level: interval between successive flight/delivery events; reported under multiple family definitions | Flight and delivery records (constructed) | Era: events per period (count rate); program: months between events |
| Mass (control) | Spacecraft or instrument mass at commitment | NTRS technical records; Aerospace Conference datasets [\[2\]](#ref-2), [\[3\]](#ref-3) | Continuous, kilograms |
| Power (control) | Spacecraft or instrument power at commitment | NTRS technical records; Aerospace Conference datasets [\[2\]](#ref-2), [\[3\]](#ref-3) | Continuous, watts |
| Technology readiness level (control) | TRL recorded at the commitment point | Commitment-review documentation [\[1\]](#ref-1) | Ordinal, TRL 1 to 9 |
| Mission class (control) | Risk/class category at commitment | Program record | Categorical |
| Contract type (control) | Prime contract structure (e.g., cost-type vs. firm-fixed-price) | Program record [\[20\]](#ref-20), [\[85\]](#ref-85) | Categorical |
| Number of external partners (control) | Count of distinct external partner organizations on the program | Program record; jointness documentation [\[115\]](#ref-115) | Count (>= 0) |
| Funding-instability index (control) | Constructed from year-over-year deviations between requested and appropriated funds across the phase | Budget and appropriations record [\[88\]](#ref-88), [\[129\]](#ref-129) | Continuous index (>= 0) |
| Program fixed effect (\(\alpha_i\)) | Program identity | Panel structure | Categorical (one per program i) |
| Era fixed effect (\(\delta_t\)) | Era / rule-regime period | Era-regime definition table (codebook) | Categorical (one per era t) |

## 4.9 Data quality, validation, and ethics

### 4.9.1 Data-quality procedures

The quality of the constructed panel is governed by four procedures, each stated so it can be audited. First, every variable is produced by the single documentary rule recorded in the variable-construction codebook, so that two constructors applying the codebook to the same records would produce the same values; inter-coder agreement on a sample of program-phases is the check on whether the rule is specified tightly enough, and disagreements drive codebook refinement rather than ad hoc adjudication. Second, every source-to-target mapping is logged: for each program-phase value, the source document and the extraction decision are recorded, so the provenance of any number is recoverable. Third, where two sources report the same quantity, both are recorded and reconciled by a fixed precedence rule, with disagreements beyond tolerance logged and investigated, never silently averaged. Fourth, the bounded-latency and bounded-value cases are flagged in the panel so that the analysis can treat point-identified and bounded observations distinctly. These procedures realize the constant-rule standard at the level of the individual datum, which is where the Maddison discipline ultimately has to hold if the cross-era comparison is to mean anything [\[16\]](#ref-16), [\[26\]](#ref-26).

### 4.9.2 Validation against known values

Validation, detailed in Section 4.6, is the load-bearing quality gate. The constructed cost and schedule growth for overlapping modern programs must reproduce the published GAO and Aerospace Conference figures within a stated tolerance before the construction rule is trusted on programs and eras where no published comparison exists [\[1\]](#ref-1), [\[3\]](#ref-3), [\[130\]](#ref-130). The logic is that a rule that reproduces known values has earned the right to be applied to unknown ones; a rule that fails the reproduction test is corrected before the panel is used. This is the measurement analogue of the dissertation's broader falsifiability posture: the construction is exposed to a test it could fail, and passing that test is what licenses the inference, not the constructor's confidence.

### 4.9.3 Ethics and access

The ethics profile of this study is favorable and is stated plainly. The data are documentary records about programs and institutional processes, not human subjects; there are no individuals whose privacy is at stake, no personally identifiable information, and no protected populations. The records are public or archivally accessible: budget justifications and GAO reports are public, NTRS is an open repository, and the early material is held in public archives. There is therefore no licensing barrier and no proprietary-data restriction that would prevent another researcher from reproducing the panel from the same sources, which is itself an ethical strength, because reproducibility is the condition under which the falsifiability claim is meaningful. The one ethical obligation that does bind is accurate attribution and faithful representation of the source record: decision events must be coded from what the documents actually say and date, not from the constructor's expectation of how slow a process should have been, and narrative framings in advocacy-oriented reports must be treated as context rather than as measurements. This obligation is operationalized by the source-logging and inter-coder procedures above, so that the ethics of faithful representation are enforced by the same mechanisms that enforce data quality.
## 4.10 Chapter synthesis

This chapter establishes that the design reaches the causal mechanism by measuring its central variable, authorization latency, from primary documentary records under a single rule, with the technical and programmatic confounders measured alongside it so the latency channel can be separated from them. That is the measurement layer's share of the argument, and the boundary is worth marking so the structure of the case stays visible. The chapter does not by itself establish that the problem is real or material; those points rest on the cost-growth and megaproject evidence developed in the literature chapter. The support is the three named datasets, their documented provenance and access, and the variable-construction rules and table above, applied under the cliometric measurement standard: a single, transparent, constant-unit rule applied identically across eras is the precondition for any cross-era comparison [\[16\]](#ref-16), [\[17\]](#ref-17), [\[18\]](#ref-18), [\[26\]](#ref-26). Three independent primary sources converge on the same program-phase key, and the construction must reproduce known published values before it is trusted on unknown ones [\[1\]](#ref-1), [\[3\]](#ref-3), [\[130\]](#ref-130). The construction is design-stage: no series is reported here as executed, the early decades carry moderate rather than high confidence, and two named sources, the GAO assessment series and the New Start Inflation Index document, must still be added to the corpus as discrete records. A careful examiner would press that clean measurement does not make latency exogenous. The point is conceded and routed to its proper home: exogeneity is a design problem solved by fixed effects and instruments in the next chapter, not a measurement problem solved here, and the measurement's contribution to that solution is the recordable early-in-phase-latency distinction and the funding-instability control that the design chapter then exploits [\[12\]](#ref-12), [\[115\]](#ref-115), [\[129\]](#ref-129).

Overall confidence in the dataset and measurement design is high for the modern subperiod and moderate for the early decades, and that asymmetry is built into the two-resolution reporting rather than averaged away. The single most valuable thing this chapter produces, independent of any regression, is the prospect of a consistent long-run series of NASA authorization latency, a measurement that does not currently exist and that is worth constructing whether the eventual coefficient rejects or retains the null. That is the chapter's answer to its own opening thesis: the series can be built by one rule from primary record, the rule is specified tightly enough to reproduce and to audit, and the honest residual uncertainties are named and bounded rather than concealed. The data chapter therefore hands the design chapter a panel whose every variable means one thing, whose imprecision is visible where it exists, and whose construction is reproducible from public sources, the necessary foundation for the falsifiable test the dissertation pre-registers.



## Chapter 5. Research Design and Identification

## 5.0 The chapter's answer

The design of this dissertation identifies the within-program, within-era association between administrative decision-and-authorization latency and three program-execution outcomes (cost growth, schedule slip, and mission cadence) by absorbing time-invariant program difficulty and common era shocks into fixed effects, by instrumenting the residual phase-specific variation in latency with two sources that move authorization timing for reasons unrelated to the engineering difficulty of the specific phase, and by replacing the naive two-way fixed-effects difference-in-differences estimator with heterogeneity-robust estimators wherever a discrete administrative regime is the object of inference. That is the whole claim of this chapter, stated first. Everything that follows develops it: the estimator and why it is chosen, the specifications written out in the fixed notation of the bible, the identification assumptions argued formally rather than asserted, the validity of each instrument tested rather than assumed, the four families of validity threat with their mitigations, the robustness battery, a power and minimum-detectable-effect analysis appropriate to a panel with a modest number of program clusters, the pre-registration commitment that binds the analysis in advance, and the computational and software plan that makes the whole procedure reproducible.

The chapter is written at design stage. No coefficient reported here is an executed estimate. Where a number appears, it is an illustrative input to a power calculation or a synthetic placeholder, and it is labeled as such. The discipline is deliberate, and it follows directly from the falsifiability standard the dissertation is built on: a design that quietly reported invented magnitudes as if they were findings would forfeit the very property that makes the contribution worth defending. The identification logic, by contrast, can and must be argued in full now, because the credibility of the eventual estimate rests entirely on whether the design was specified honestly before the data spoke.

The problem this chapter must solve, framed as current state, desired state, gap, and consequence, is the following. The current state of the question is that the practitioner attribution of NASA cost growth and schedule slip to slow internal decision-making is an unmeasured correlation at best and an anecdote at worst, while the nearest rigorous external evidence (that administrative competence causally reduces procurement delays and overruns in United States federal contracting) comes from a setting with thousands of contract-level observations and a credible instrument [\[12\]](#ref-12), [\[53\]](#ref-53). The desired state is an estimate of the latency-to-outcome relationship inside NASA that survives the obvious objection that harder programs both wait longer and overrun more. The gap is a research design that separates the process signal from the difficulty confound in a single-agency panel that is unbalanced by construction, spans rule regimes that switch on at staggered times, and contains far fewer independent clusters than the procurement studies. The consequence of leaving the gap unaddressed is that the agency continues to reform on the basis of a narrative it has never tested, unable to say whether process-speed reform or technical-risk reduction is the better lever. This chapter is the design that closes the gap.

## 5.1 The estimator and why it is chosen

The primary estimator is a linear panel regression with two-way fixed effects: program fixed effects and era fixed effects, estimated by within transformation and reported with standard errors clustered by program. The choice is not automatic, and it is defended here against the two natural alternatives, pooled cross-section and random effects, for reasons specific to this problem rather than generic.

The two-way fixed-effects panel estimator is the correct workhorse for identifying the latency association in these data. The central confound is unobserved program difficulty, which is plausibly persistent across a program's lifecycle phases but heterogeneous across programs, and the central nuisance is the era-level rule regime, which is common to all programs active in a period but differs across periods. A fixed-effects within estimator removes any regressor's correlation with unit-specific and period-specific unobservables by differencing them out, so that identification comes only from variation in latency that occurs within a program across its phases and within an era across programs, net of both sets of effects. The panel-methods literature is explicit that fixed effects is the appropriate default precisely when measured explanatory variables are correlated with unit-specific unobservables that threaten causal inference, the situation here [\[93\]](#ref-93), [\[126\]](#ref-126). Fixed effects buys this protection only against time-invariant confounders; difficulty that varies by phase is not removed, and that residual is exactly what the instrumental-variable layer of Section 5.4 must address. One objection deserves a full hearing: a vocal strand of the methods literature argues that a correctly specified within-between random-effects model (the Mundlak formulation) can deliver everything fixed effects promises and more, including time-invariant covariates and random slopes [\[116\]](#ref-116), [\[128\]](#ref-128). That argument is taken seriously rather than dismissed. The dissertation's response is twofold. First, the within-between specification will be estimated as a robustness check and its between-program component reported, because the contrast between the within and between effects of latency is itself substantively interesting (it is the difference between "a program that waits longer than its own norm overruns more" and "programs that wait longer on average overrun more"). Second, the headline estimate remains the within (fixed-effects) estimate, because the dissertation's identifying claim is explicitly a within-program one and the between-program comparison is the channel most contaminated by the difficulty confound. The Mundlak equivalence does not weaken the fixed-effects choice; it clarifies what the fixed-effects estimate is conditioning on and supplies a principled robustness lane.

Pooled ordinary least squares is rejected for the obvious reason that it leaves the difficulty confound entirely in the error and would deliver an estimate of beta that is the sum of the causal channel and the selection of difficult programs into long latency. Random effects without the Mundlak correction is rejected because it imposes orthogonality between latency and the program effect, an assumption the difficulty confound makes untenable; the Hausman logic that distinguishes the two is well rehearsed in the panel literature the design draws on [\[93\]](#ref-93). The choice of fixed effects is therefore made for cause, against named alternatives, not a default reached for convenience.

A further reason the fixed-effects within estimator is the right primary tool is specific to the institutional theory that motivates the study. North's account of why authorization regimes are large and durable is an account of persistence: once a set of review and authorization rules is in place, it generates increasing returns, organizations adapt to it, and the cost of changing it rises, so a program inherits a latency level that is substantially a property of the regime it was born into rather than of its own engineering [\[18\]](#ref-18), [\[91\]](#ref-91). If that persistence operates mostly at the program level (a program adopts a posture toward its reviewers and carries it through its phases), then the program fixed effect is not merely a statistical nuisance absorber but the empirical counterpart of the path-dependent program-level component North's theory predicts, and the within variation that remains after differencing it out is the cleaner object: the phase-to-phase movement in latency that a program-specific posture cannot explain. The estimator choice and the institutional theory are therefore mutually reinforcing, which is why the design treats the program effect as substantively meaningful and reports the within-between decomposition (Section 5.8) rather than discarding the between component as uninterpretable.

Two further estimation choices follow from the structure of the outcome. Because cost growth and schedule slip are bounded-below ratios (a phase cannot lose more than its baseline) with right skew, the design pre-commits to reporting the linear within estimate as the headline and a log-transformed-outcome within estimate as a robustness lane, so that a finding driven by a handful of extreme-growth phases in the right tail is visible rather than hidden inside a mean effect. And because the panel is unbalanced and some outcomes are censored or partially observed for the earliest decades, the design notes but does not adopt nonparametric panel estimators for censored dependent variables as a sensitivity check on the most heavily censored cadence outcome, citing the relevant inference results so the choice is traceable [\[125\]](#ref-125). Neither lane changes the headline estimator; both make the headline's robustness to functional form and censoring legible.

## 5.2 The specification

The baseline specification is the equation fixed in the bible and used identically across every chapter of this dissertation. For program-phase observation indexed by program i, phase p, and era t:

\[
\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t) \qquad\qquad (1)
\]

where Outcome is in turn cost growth, schedule slip, and mission cadence; Latency(i,p,t) is authorization latency in months, taken as the median elapsed time across the authorization events in that program-phase; X(i,p,t) is the vector of technical and programmatic controls (instrument or spacecraft mass and power, technology readiness level at commitment, mission class, contract type, number of external partners, and the funding-instability index); \(\alpha_i\) are program fixed effects; \(\delta_t\) are era fixed effects; and \(\varepsilon(i,p,t)\) is the error term, clustered by program. The coefficient of interest is \(\beta\). The sign convention is fixed and restated wherever \(\beta\) is interpreted: under H1, \(\beta\) is positive for cost growth and for schedule slip, and negative for mission cadence; under H0, \(\beta\) is statistically indistinguishable from zero for all three outcomes.

Three specification points require elaboration because they are where a careless implementation would go wrong.

First, the outcome is the same continuous, baseline-normalized quantity defined in Chapter 4, and it is estimated separately for each of the three outcomes rather than stacked, so that beta carries a distinct interpretation per outcome and the standard errors are not artificially shared across outcomes. Cost growth and schedule slip are dimensionless fractions of baseline (with cost expressed in constant fiscal-year dollars via the NASA New Start Inflation Index, per the measurement discipline of Chapter 4); cadence is in events per period at the era level and in inter-event months at the program level, which is why cadence is estimated under its own specification with era fixed effects and program-family definitions rather than mechanically inside the same program-phase frame.

Second, the controls in X are entered to absorb the technical drivers the cost-estimating literature has already shown to move cost and schedule (mass, power, technology readiness, mission class, contract type), so that beta is the latency association net of those drivers rather than a proxy for them [\[1\]](#ref-1), [\[19\]](#ref-19), [\[101\]](#ref-101). This is a substantive design choice, not a mechanical one: omitting these controls would let latency stand in for instrument complexity, and the estimate would then conflate "waiting longer" with "building something harder." The funding-instability index, constructed from year-over-year deviations between requested and appropriated funds, is included in X precisely because budget instability is a candidate common cause of both high latency and high cost growth, and conditioning on it closes one of the three named rival explanations directly inside the regression.

Third, inference. Standard errors are clustered by program to allow arbitrary correlation across a program's phases, because the same program's phase-level errors are surely correlated (a program with an optimistic culture or a difficult subsystem carries that through its phases). But program-clustered inference with a modest number of programs raises the small-cluster problem: cluster-robust variance estimators are biased toward zero, and conventional critical values over-reject, when the number of clusters is not large. The design therefore reports, alongside the conventional cluster-robust intervals, wild-cluster bootstrap intervals, the standard remedy for inference with few clusters and the one that restores correct size when the asymptotic-in-clusters approximation is poor. The number of program clusters is the binding constraint on statistical-conclusion validity in this study, and Section 5.7 quantifies what it implies for detectable effects; flagging it here, in the specification, is deliberate.

The cadence outcome requires a specification distinct from the cost-growth and schedule-slip outcomes, and the distinction is stated explicitly to prevent the error of forcing all three into a single program-phase frame. Cadence is not a per-phase ratio; it is a rate (operational mission events per unit time within a program family at the era level) and an interval (months between successive flight or delivery events at the program level). The era-level cadence specification therefore replaces the program fixed effect with a program-family fixed effect and retains the era fixed effect, so that the latency coefficient is identified from within-family, within-era variation in the family's average authorization latency. Because event counts are non-negative integers, the design pre-commits to a count-data specification (a fixed-effects Poisson or negative-binomial form) for the rate outcome alongside the linear form, since a linear model on a count can produce nonsensical negative fitted rates and understate the influence of high-cadence families; the linear and count estimates are reported together and divergence is flagged. The program-level inter-event specification, by contrast, treats the months-between-events as a duration and is the more construct-fragile of the two, which is the operational reason cadence is reported under multiple family definitions throughout the robustness battery. Stating the cadence specification separately here, rather than letting the bible equation appear to cover it mechanically, is the honest representation of what the third outcome actually requires.

## 5.3 Identification, the staggered-regime concern, and heterogeneity-robust estimators
The identifying variation has been named: within-program variation in latency across phases, and within-era variation in latency across programs. The formal identification assumption for the fixed-effects baseline is conditional mean independence of the phase-specific error from latency given the program effect, the era effect, and the controls. Stated plainly: once we hold a program's fixed characteristics, an era's common conditions, and the technical controls constant, the remaining variation in how long authorization took for a given phase is assumed uncorrelated with the remaining shock to that phase's cost growth, schedule slip, or cadence. This assumption is strong, it is not testable directly, and the honest posture of this chapter is to state it as an assumption, argue for its plausibility, name the conditions under which it fails, and build the instrumental-variable layer of Section 5.4 as the response to the most consequential failure mode. Confidence in the bare fixed-effects identification is therefore moderate, not high; the instrumented design is what raises it.

A distinct and more technical threat arises whenever the object of inference shifts from the continuous latency variable to the effect of a discrete administrative-regime change, for example a reorganization that compresses reporting layers or a procurement-reform regime that changes how authorizations route. Because such regimes switch on at staggered calendar times across programs, the design is, for those analyses, a staggered-adoption difference-in-differences design, and the modern econometric literature has established that the two-way fixed-effects estimator is not a safe default in that setting.

For any analysis of a discrete administrative-regime change, the two-way fixed-effects difference-in-differences estimator must be replaced by a heterogeneity-robust estimator. The canonical decomposition result shows why: the two-way fixed-effects difference-in-differences estimand is a weighted average of all possible two-group, two-period comparisons, and some of those comparisons use already-treated units as controls for later-treated units, which can attach negative weights to legitimate treatment effects and produce an estimate with the wrong sign even when every underlying effect is positive [\[21\]](#ref-21), [\[29\]](#ref-29). When treatment effects are heterogeneous across cohorts or dynamic over time, as administrative-regime effects on long-lived programs certainly are, the negative-weighting pathology is not a remote possibility but the expected behavior of the estimator, and a large simulation and survey literature confirms that the bias is material in exactly the heterogeneous, staggered settings this design contains [\[104\]](#ref-104), [\[135\]](#ref-135), [\[136\]](#ref-136). A family of estimators robust to multiple periods and heterogeneous, dynamic effects now exists and is well validated: the group-time average-treatment-effect estimator [\[22\]](#ref-22); the intertemporal estimators of de Chaisemartin and D'Haultfoeuille [\[23\]](#ref-23), [\[109\]](#ref-109); the efficient imputation estimator for event studies [\[133\]](#ref-133); and the synthetic difference-in-differences approach for sparse, staggered adoption [\[99\]](#ref-99), [\[110\]](#ref-110). The two-way Mundlak equivalence further clarifies which extended specifications recover heterogeneity-robust estimands and which do not [\[132\]](#ref-132), and the event-study identification results discipline the choice of binning and distributed-lag structure [\[134\]](#ref-134). These estimators are deployed only for the discrete-regime analyses; they are not a replacement for the continuous-latency baseline, where the negative-weighting problem does not arise in the same form. One might object that because the headline variable is continuous latency rather than a binary treatment, the staggered-difference-in-differences machinery is unnecessary altogether. The dissertation rejects that shortcut. Even continuous treatments admit multiple, sometimes misleading, interpretations under two-way fixed effects, and the continuous-treatment difference-in-differences results are explicit that parallel-trends-type assumptions and selection-bias caveats carry over [\[100\]](#ref-100). The continuous framing is therefore the first line of defense against the worst of the staggered-binary problem, not a license to ignore it.

The practical protocol follows from this. The continuous-latency specification of Section 5.2 is estimated first and is the headline. The Goodman-Bacon decomposition is computed and reported as a diagnostic wherever a regime indicator enters, so that the reader can see how much weight the two-way fixed-effects comparison places on forbidden already-treated-as-control contrasts [\[29\]](#ref-29). Where the decomposition reveals meaningful negative weighting, the regime effect is re-estimated with the Callaway and Sant'Anna group-time estimator and the de Chaisemartin and D'Haultfoeuille intertemporal estimator, and the three estimates are reported side by side [\[22\]](#ref-22), [\[23\]](#ref-23). The Imai and Kim results on when two-way fixed-effects regression does and does not identify a causal effect with panel data govern the interpretation of the baseline throughout, supplying the formal conditions under which the within estimator carries a causal reading [\[24\]](#ref-24). The mechanism by which a regime would matter is named and not left as a black box: a regime that compresses authorization layers (driver) reduces the elapsed time each pending action accrues (mechanism), which lowers measured latency per program-phase (observable effect), which under H1 should reduce cost growth and schedule slip and raise cadence (operational consequence), implying that the regime is a controllable lever on execution performance (strategic implication). Stating the chain makes clear that the regime analysis tests the same causal pathway as the continuous-latency analysis, through a different door.

The two specifications identify different estimands, and naming each one explicitly is a prerequisite for reading any coefficient correctly. For the continuous-latency specification, the identifying variation is within-program and within-era; the estimand is accordingly a variance-weighted average of the effect of latency across programs and phases, where program-phase observations with larger within-unit variation in latency receive more weight. That weighting reflects the identifying variation, not an analyst choice, and it means the headline coefficient answers a specific and bounded question: what is the expected change in the outcome for a program-phase that experiences more latency than that same program's other phases and more latency than other programs in the same era, holding observable covariates constant. The answer to that question is meaningful and is the dissertation's primary target, but it is not the population-average effect across all programs, nor a specific program's effect, nor the effect that a uniform administrative intervention would induce across the portfolio. For the discrete-regime specifications, where programs adopt a new authorization architecture at staggered calendar dates, the design commits to group-time average treatment effects as the primary estimand, as formalized by Callaway and Sant'Anna: the average effect of adoption for programs that adopted in a given calendar cohort, measured at a given number of periods after adoption [\[22\]](#ref-22). Group-time effects are more interpretable than the two-way fixed-effects summary in this setting for a precise structural reason: the TWFE estimand aggregates over all cohort-period cells and attaches implicit weights to those cells that, under treatment-effect heterogeneity, can be negative for some already-treated cohorts [\[21\]](#ref-21). A summary coefficient carrying negative weights on some comparisons can have the opposite sign from every individual group-time effect, and in a setting where administrative-regime effects on long-lived programs are expected to vary by era, program maturity, and institutional context, negative weighting is not a tail risk but the likely behavior of the TWFE estimator. Group-time effects are therefore recovered separately and aggregated to an overall summary only when pre-tests and event-study patterns support treating the heterogeneity as negligible; where the pattern reveals meaningful variation, the dissertation reports the disaggregated group-time effects as the primary evidence and declines to collapse them into a single number that would obscure the variation. This commitment is fixed before the estimates are seen so the reporting choice cannot be influenced by which presentation looks more favorable.

## 5.4 Endogeneity and the instrumental-variable strategy

The central internal threat is that latency is endogenous to phase-specific program difficulty. Program fixed effects remove the part of difficulty that is constant across a program's life, but a program can encounter a genuinely hard phase, and a hard phase may both take longer to authorize (because the action is more consequential and the reviewers more cautious) and overrun more (because the engineering is harder). If so, the within-program correlation between latency and the outcomes reflects difficulty, not process, and the fixed-effects beta is biased away from zero in the direction H1 predicts, manufacturing a spurious confirmation. This is the single most dangerous failure mode in the entire design, and it is addressed head-on with an instrumental-variable strategy modeled on the procurement literature that solved the analogous problem.

Residual phase-specific endogeneity of latency can be addressed by instrumenting latency with two sources of variation that move authorization timing for reasons orthogonal to the engineering difficulty of the specific phase. The procurement-competence literature faced the identical confound (more competent bureaus might handle easier contracts) and resolved it with an instrument for bureaucratic competence, recovering a causal effect of administrative quality on delays and overruns [\[12\]](#ref-12), and an independent design exploiting presidential transitions as a source of within-bureaucrat variation in political alignment likewise isolated an administrative channel into procurement cost overruns and delays [\[53\]](#ref-53). The logic is the standard one: an instrument that shifts how long an authorization takes, but that does not enter the outcome equation except through latency and is uncorrelated with the phase-specific difficulty shock, identifies the causal effect of latency on the outcome even when latency is otherwise endogenous. The two candidate instruments are constructed from sources the documentary record supplies and the institutional structure makes plausible. The first is the contemporaneous workload of the authorizing office, measured as the number of other programs awaiting authorization at the same office in the same period; the institutional reason a given phase waits longer when the office is congested is queueing, not the difficulty of that phase, and the Standing Review Board and Key Decision Point machinery documented in the NASA process records makes the queue a real and datable object [\[80\]](#ref-80), [\[94\]](#ref-94), [\[95\]](#ref-95). The second is the timing of the action relative to the annual appropriations calendar; budget-dependent authorizations are paced by the appropriations cycle and by continuing-resolution conditions for reasons entirely exogenous to any single program's engineering, and the budget-and-appropriations-cycle literature establishes that this pacing is a structural feature of the federal process rather than a program-specific choice [\[88\]](#ref-88), [\[129\]](#ref-129). Instrument validity is argued and tested, not assumed, and the design reports both the fixed-effects and the instrumented estimates so the reader can see exactly how much the conclusion moves when the instrument is imposed. Each instrument's exclusion restriction has a credible objection, and the design pre-commits to the test that addresses it.

For the authorizing-office-workload instrument, the threat to exclusion is that office congestion correlates with an agency-wide busy period that itself drives cost growth (a common-cause story). The design addresses this in three ways: the era fixed effects already absorb agency-wide period conditions; the funding-instability control absorbs the budget-shock channel through which a busy period would most plausibly raise cost; and a placebo test regresses the outcome on the instrument among program-phases whose authorization had already cleared the congested office, where the instrument should have no effect if the exclusion restriction holds. For the appropriations-calendar instrument, the threat is that the appropriations cycle drives cost growth directly through funding timing rather than only through authorization latency. The same funding-instability control absorbs the direct-funding channel, and the design additionally reports the estimate restricted to program-phases funded from prior-year balances, for which the appropriations-calendar instrument shifts authorization timing without shifting the money, sharpening the exclusion case.

The relevance condition deserves its own argument, because an instrument that fails relevance is worse than no instrument: weak instruments amplify any small violation of exclusion and bias the instrumented estimate toward the very confounded fixed-effects estimate the instrument was meant to escape. The relevance case for each instrument is institutional, not merely statistical. Authorizing-office workload is relevant because the documented review machinery is a genuine queue: Standing Review Boards, Key Decision Points, and Joint Confidence Level analyses are scheduled events that compete for a finite reviewing capacity, so when the office is congested a given phase's authorization waits longer for reasons of throughput [\[80\]](#ref-80), [\[86\]](#ref-86), [\[95\]](#ref-95). The appropriations-calendar instrument is relevant because budget-dependent authorizations cannot legally proceed until the enabling appropriation exists, so the position of an action within the annual cycle, and whether the period falls under a continuing resolution, mechanically shifts when authorization can occur [\[88\]](#ref-88), [\[129\]](#ref-129). In both cases the relevance is a feature of how the institution actually authorizes, which is why the first stage is expected to be strong; but expectation is not evidence, so the design reports the realized first stage and treats relevance as something the data must demonstrate.

First-stage strength is reported and not assumed: the design reports the first-stage F statistic (and its heteroskedasticity- and cluster-robust analogue) for each instrument and for the two jointly, and treats weak-instrument diagnostics as a gate that must be passed before the instrumented estimate is interpreted. Because exclusion restrictions are never beyond doubt, the design also implements sensitivity analysis under possibly-invalid instruments rather than treating validity as binary: the formal results on causal inference with imperfect instrumental variables bound how much the estimate can move under a stated degree of instrument invalidity, and the corresponding tooling makes the bounds computable [\[111\]](#ref-111), [\[127\]](#ref-127). With two instruments for one endogenous regressor the model is over-identified, which the design exploits in two ways: an over-identification test (the joint null that both instruments are valid) is reported as a falsification probe, and, because such tests have low power and a passed test does not prove validity, the imperfect-instrument bounds are reported regardless of the over-identification test's outcome, so that validity is treated as a continuous degree of belief rather than a binary verdict. Confidence in the causal reading of beta is therefore conditioned explicitly on the instruments: high where the first stage is strong and the imperfect-instrument bounds exclude zero, moderate where the first stage is adequate but the bounds are wide, and low where the instrument is weak, in which case the design reports the fixed-effects estimate as a within-program association and downgrades the causal claim accordingly. This calibration is not a hedge; it is the honest mapping from the strength of the identification to the strength of the claim, and it is fixed in advance so it cannot be adjusted after the estimates are seen.

## 5.5 Threats to validity

The design is organized around the four classical validity types, and each threat is paired with the specific mitigation the design pre-commits to. The framing follows the rival-explanations discipline of case-design methodology: a threat is not retired by being named, only by being given a check that could in principle fail [\[33\]](#ref-33).

### 5.5.1 Internal validity

The principal internal threat is the endogeneity of latency to phase-specific difficulty, treated in full in Section 5.4 through the fixed-effects-plus-instrument design. A second internal threat is baseline gaming. If an agency, anticipating slow authorization, sets an optimistic baseline so that the program looks affordable at commitment, then baseline choice and latency are jointly determined, and measured cost growth and schedule slip partly reflect the baseline rather than the process. The mechanism is well documented outside NASA: optimism bias and strategic misrepresentation systematically depress baselines on large public projects [\[13\]](#ref-13), [\[30\]](#ref-30). The design's check is a baseline-conservatism test: it proxies baseline conservatism by the ratio of held reserves to baseline (the cost-estimating community's own confidence-level construct, in which baseline plus reserves is targeted near the fiftieth percentile and reserves plus unallocated future expenses near the seventieth) and tests whether that ratio correlates with latency [\[85\]](#ref-85), [\[86\]](#ref-86). If conservatism and latency are uncorrelated, the baseline-gaming channel is unlikely to be driving the result; if they correlate, the design conditions on the reserve ratio and reports how much beta moves. A third internal threat is reverse causation: a program already overrunning may generate more authorization events and longer latency as a consequence of trouble. The design's responses are to measure latency early in each phase, before most of the phase's cost growth has accrued, and to lean on the instruments, which shift latency for reasons that cannot be a downstream consequence of the phase's own overrun.

### 5.5.2 External validity

The findings pertain to NASA, a single agency with a particular authorization architecture, and the design does not claim statistical generalization to other agencies or to commercial programs with different decision structures. The external-validity posture is therefore explicit and modest: the dissertation frames its claims as NASA-specific and uses the federal-procurement evidence only as corroboration of the mechanism, not as a basis for extrapolating the magnitude. The nearest external benchmark is the procurement-competence result, which finds that administrative competence causally reduces delays and overruns across United States federal contracting [\[12\]](#ref-12), reinforced by the ideology-and-performance evidence on procurement officers [\[53\]](#ref-53) and by the broader transaction-cost reading of public bureaucracy that treats administrative time as a real cost of internal governance [\[81\]](#ref-81), [\[131\]](#ref-131). Agreement between this study's NASA-specific finding and that broader body would strengthen the case that the mechanism is general while leaving the magnitude NASA-specific; disagreement would not falsify the NASA estimate but would caution against transporting it. The platform-versus-bespoke comparison that contrasts NASA's and a commercial operator's space-mission cost, speed, and scalability is a further external touchstone for the cadence outcome in particular, and it is read as context rather than as a comparator the NASA panel must reproduce [\[59\]](#ref-59).

### 5.5.3 Construct validity

The central construct, authorization latency, is operationalized from documentary records, and the operationalization may not capture the full administrative-time concept a practitioner has in mind. The design's defense is the two-resolution measure introduced in Chapter 4: a coarse, milestone-to-milestone latency available for the full 1958-to-2026 span, and a fine, individual-Key-Decision-Point latency available only for the modern subperiod. Every result is reported separately at each resolution, so that any finding which appears only at one resolution is flagged as resolution-dependent rather than presented as robust. The construct most at risk is mission cadence, which depends on how a program family is defined and for which no single definition is uniquely correct; the design reports cadence under multiple family definitions and treats a cadence result that survives only one definition as fragile. Construct validity for latency is further supported by the documentary grounding of the trigger and resolution events in the actual review machinery (Key Decision Points, Program Commitment Reviews, Standing Review Board reviews, Joint Confidence Level analyses), so that the measure tracks events the institution itself records and acts upon rather than an analyst's reconstruction [\[80\]](#ref-80), [\[86\]](#ref-86), [\[95\]](#ref-95).

### 5.5.4 Statistical-conclusion validity

The threats here are the unbalanced panel, the modest number of program clusters, and the consequent fragility of conventional inference. The unbalanced structure is intrinsic (early decades carry coarser, sparser baselines) and is handled by estimators that accommodate unbalanced panels and by reporting whether results hold on the balanced modern subpanel. The small-cluster problem is addressed by the wild-cluster bootstrap alongside conventional cluster-robust intervals, and the design reports estimates under alternative clustering (by program, by era, and two-way) and under alternative fixed-effects structures, so that no conclusion rests on a single inference choice. The handbook treatment of regression and causal inference with complex, clustered, and longitudinal samples governs these choices [\[93\]](#ref-93), [\[117\]](#ref-117). The design also pre-commits to reporting effect sizes with intervals rather than significance stars alone, because the substantive question is whether latency moves cost and schedule by an amount that matters operationally, not merely whether a coefficient clears a threshold; a coefficient can be statistically distinguishable from zero and operationally trivial, and the reverse, and the reporting standard is built to keep that distinction visible.

## 5.6 Pattern-matching and rival explanations

The design adopts the pattern-matching logic of case research: a predicted pattern of results is specified in advance, and the analysis asks whether the observed pattern matches it or matches a named rival instead [\[33\]](#ref-33). The predicted pattern under H1 is specific and jointly testable: the latency coefficient is positive for cost growth, positive and largest and most robust for schedule slip (because authorization latency is itself a component of elapsed schedule and the mechanism is the most direct of the three), and negative for cadence. A result that confirmed cost growth but not schedule slip would be internally suspicious, because schedule slip is the outcome with the most direct mechanical link to latency, and the design treats such a pattern as evidence against the causal reading rather than as partial confirmation.

Three rival explanations are specified in advance and matched against, even if the predicted association is found. The first rival is reverse causation, whose predicted signature is that the latency association is strong for late-phase latency (latency measured after most overrun has accrued) but weak for early-phase latency; the design's early-in-phase measurement and the instruments are the discriminating tests. The second rival is common cause through an era condition such as budget instability, whose signature is that the latency association shrinks toward zero once the funding-instability index and era fixed effects are included; the design conditions on both inside the baseline, so a result that survives is already net of this rival. The third rival is optimistic baselines, whose signature is that the latency association is fully explained by baseline conservatism; the baseline-conservatism test of Section 5.5.1 is the discriminator. The design's commitment is that each rival has a pre-specified observable signature and a check that could in principle vindicate it, so that retaining H1 means the data declined to match any of the three named rivals, not merely that H1 was assumed.

A fourth rival, weaker than the three above but worth pre-naming, is that the latency construct is itself a proxy for organizational complexity rather than for administrative time per se: a program that spans many agencies or centers may both accrue longer authorization latency (more parties to clear) and overrun more (more interfaces to manage), so that latency stands in for jointness rather than for waiting. The mechanism is documented: jointness induces technical and organizational complexity that drives cost growth, as the in-depth NPOESS case study shows by tracing how collaborating-agency interactions multiplied complexity and cost [\[115\]](#ref-115). The design's check is the number-of-external-partners control already in X: conditioning on partner count, the latency coefficient is identified net of the jointness channel, and a result that survives is net of this fourth rival as well. The transaction-cost reading clarifies why the partner-count control is the right instrument-free defense here: each additional party to an authorization is a transaction whose measurement and enforcement cost the institution must bear, so partner count is the observable counterpart of the transaction-cost complexity that jointness creates [\[81\]](#ref-81), [\[113\]](#ref-113), [\[131\]](#ref-131). Conditioning on it does not eliminate the possibility that latency and jointness are entangled, but it forces any surviving latency effect to be variation in administrative time that partner count cannot explain, which is precisely the construct the dissertation intends to measure.
## 5.7 Power and minimum-detectable-effect analysis

A design-stage power analysis matters here because the binding constraint on this study is not the number of program-phase observations but the number of independent program clusters. A design that cannot detect an operationally meaningful effect with the clusters available would risk retaining H0 for lack of power rather than for absence of effect. The analysis below is illustrative: every numerical input is a stated assumption, not an estimate from the data, and its purpose is to map the cluster count and effect size into a minimum detectable effect so the reader can judge feasibility before any estimation.

The relevant power quantity for a clustered panel is the minimum detectable effect (MDE), the smallest true beta the design can reject the null against with a target power (conventionally eighty percent) at a chosen significance level (conventionally five percent, two-sided). For a within estimator with cluster-robust inference, the MDE scales with the residual standard deviation of the outcome after the fixed effects and controls, inversely with the residual standard deviation of latency after the fixed effects and controls (the within variation the estimate exploits), and inversely with the square root of the effective number of clusters adjusted for the intra-cluster correlation of both latency and the outcome. The design computes the MDE under a grid of assumptions rather than a single point, because each input is uncertain at design stage.

To make the exercise concrete and reproducible, the design pre-specifies the grid. Outcome residual standard deviation for cost growth is varied across plausible values consistent with the published NASA cost-growth distributions (the cost-estimating literature reports that the large majority of missions experience cost growth from preliminary design review to launch, with substantial dispersion, so a residual standard deviation on the order of a few tens of percent of baseline is the assumed range, taken as an illustrative input and not as an estimate) [\[1\]](#ref-1), [\[85\]](#ref-85). Within-program residual variation in latency is varied across a range that brackets both a regime in which programs differ little from their own latency norm across phases (small within variation, large MDE) and a regime in which phases differ substantially (large within variation, small MDE). The effective cluster count is varied across the plausible range of distinct NASA programs with documented baselines, deflated by an assumed intra-cluster correlation to give the effective rather than nominal cluster count. The output of the grid is a table, computed at design stage, of the MDE for beta for each outcome under each combination, expressed in the natural units of the design: the change in cost growth (as a fraction of baseline) per one-month increase in authorization latency that the design could detect with eighty percent power.

The intra-cluster correlation term in the MDE formula deserves emphasis because it is the quantity that punishes this design most. In a clustered panel, the effective sample size for inference is not the count of program-phase observations but a design-effect-deflated count: the nominal observation count divided by a factor that grows with the product of the average cluster size and the intra-cluster correlation of the outcome. A program that contributes many phases adds far less independent information than its phase count suggests if its phases are highly correlated, which they are when a program-level posture drives both latency and outcome. The practical consequence is that adding phases to existing programs barely improves power, while adding programs improves it substantially, and the design states this asymmetry plainly so that no reader mistakes a large program-phase count for a large effective sample. The same logic is why the power grid varies the effective rather than the nominal cluster count and why the headline power figure is the bootstrap-adjusted, design-effect-deflated MDE rather than a naive calculation on the raw observation count.

The interpretation of that table is pre-committed and is itself a contribution of the design stage. If the MDE under the central grid cell is smaller than the effect size that would be operationally meaningful (for instance, smaller than the per-month cost-growth effect that would make process-speed reform competitive with technical-risk reduction as a lever), the design is adequately powered and a retained null would be informative. If the MDE under the central grid cell is larger than the operationally meaningful effect, the design is underpowered against the effects that matter, and the pre-registered response is explicit: restrict the headline estimate to the modern, fine-resolution subpanel where within-latency variation and cluster richness are greatest, report the full panel as descriptive rather than inferential, and state plainly that the full-span test is powered only against large effects. The wild-cluster bootstrap is folded into the power analysis as well, because the small-cluster correction widens intervals and therefore raises the MDE; reporting the bootstrap-adjusted MDE rather than the asymptotic one is the honest figure, and the design commits to the bootstrap-adjusted MDE as the headline power number. Confidence in the power assessment is moderate at design stage, by construction, because the inputs are assumptions; it becomes high only once the assembled panel supplies the realized residual variances, at which point the pre-registered grid is re-evaluated at the realized values before any hypothesis test is run.

## 5.8 The robustness battery

The robustness battery is specified in advance and is not a menu from which favorable results will be selected after the fact; the pre-registration of Section 5.9 fixes which checks are reported regardless of outcome. The battery has seven components, each tied to a specific threat it tests.

First, the two latency resolutions. Every headline estimate is re-estimated at both the coarse (milestone-to-milestone, full-span) and the fine (Key-Decision-Point, modern-subperiod) latency measure, and a result that holds only at one resolution is reported as resolution-dependent (tests construct validity of latency).

Second, the multiple cadence definitions. The cadence outcome is re-estimated under each pre-specified program-family definition, and a cadence result that survives only one definition is flagged as fragile (tests construct validity of cadence).

Third, the heterogeneity-robust estimators. Wherever a discrete administrative regime enters, the two-way fixed-effects estimate is accompanied by the Callaway and Sant'Anna group-time estimate, the de Chaisemartin and D'Haultfoeuille intertemporal estimate, and the Goodman-Bacon decomposition diagnostic, and divergence among them is reported rather than suppressed [\[22\]](#ref-22), [\[23\]](#ref-23), [\[29\]](#ref-29), [\[109\]](#ref-109), [\[133\]](#ref-133) (tests the staggered-regime threat).

Fourth, the instrument battery. Each outcome is reported as the fixed-effects estimate, the just-identified instrumented estimate for each instrument separately, and the over-identified instrumented estimate using both instruments, with first-stage strength and the imperfect-instrument sensitivity bounds for each [\[12\]](#ref-12), [\[111\]](#ref-111), [\[127\]](#ref-127) (tests endogeneity).

Fifth, the inference variants. Each headline estimate is reported with conventional program-clustered intervals, wild-cluster bootstrap intervals, and at least one alternative clustering (era, two-way), so that no conclusion depends on a single inference choice (tests statistical-conclusion validity).

Sixth, the specification variants. The within-between (Mundlak) random-effects specification is estimated alongside the fixed-effects baseline, reporting the between-program latency effect explicitly, and the control set is varied (technical-only, programmatic-only, full) so the reader can see the sensitivity of \(\beta\) to conditioning [\[116\]](#ref-116), [\[128\]](#ref-128) (tests the modeling choice and confound conditioning).

Seventh, the bounded-range reporting. Following the counterfactual discipline of quantitative economic history, central estimates are reported as bounded ranges conditioned on stated assumptions rather than as single points, in the manner of the social-saving tradition that reports upper and lower bounds rather than a single causal number [\[16\]](#ref-16), [\[17\]](#ref-17). The bound is constructed from the span of the instrumented and fixed-effects estimates and from the imperfect-instrument sensitivity analysis, so that the headline magnitude is a defensible interval, not a point that overstates precision.

The institutional reason the battery is this extensive, and not a token gesture, is that the durability of NASA's authorization regimes across decades (the path-dependence and increasing-returns property that makes latency large and persistent) means that any single specification is at risk of confounding a rule regime with a program-specific pattern, and only a battery that varies the era treatment, the latency resolution, the estimator, and the inference can separate the two [\[18\]](#ref-18), [\[91\]](#ref-91). The battery is the operational expression of the bible's institutional anchor, not an add-on to it.

## 5.9 The pre-registration commitment

The analysis is pre-registered, and the pre-registration is binding rather than aspirational. Before any estimation on the assembled panel, the dissertation records, with a time stamp, the following fixed elements: the hypotheses H0 and H1 in the exact wording of the bible; the baseline specification in the fixed notation; the definitions of latency (both resolutions), cost growth, schedule slip, and cadence (all family definitions); the control set; the instrument definitions and their validity tests; the heterogeneity-robust estimators and the Goodman-Bacon diagnostic; the robustness battery of Section 5.8; the power grid of Section 5.7; and the decision rule below.

The decision rule is fixed in advance precisely to prevent specification search. H0 is rejected only if the latency coefficient is statistically distinguishable from zero at conventional levels and carries the predicted sign (positive for cost growth and schedule slip, negative for cadence) in the same direction across the fixed-effects and instrumented specifications, and survives the robustness battery. The contribution fails (H0 is retained) if the coefficient is statistically indistinguishable from zero in the preferred specification, or reverses sign between the fixed-effects and instrumented specifications, or vanishes under the heterogeneity-robust estimators or the alternative latency resolution, or is fully explained by baseline gaming or by reverse causation. These falsification conditions are stated as binding gates, not as discretionary considerations, and the pre-registration record is what makes them binding: a reader can check the executed analysis against the time-stamped commitment and detect any deviation. The deliberate refusal to populate the illustrative coefficient table of the prospectus with invented numbers is part of this same commitment; reporting fabricated coefficients as if real would violate the falsifiability standard the design exists to protect, and the table is left as placeholders by design until the pre-registered procedure is executed on the assembled panel.

One pre-registered allowance is recorded honestly: the power grid of Section 5.7 is re-evaluated at the realized residual variances once the panel is assembled, before any hypothesis test, and if the design proves underpowered against operationally meaningful effects on the full span, the headline shifts to the modern fine-resolution subpanel and the full-span analysis is reported as descriptive. This is not a degree of freedom that can rescue a null into a rejection; it can only narrow the inferential claim, never widen it, and it is recorded in advance so that the narrowing, if it occurs, is visibly a pre-committed contingency rather than a post hoc retreat.

## 5.10 Computational and software plan

The design is reproducible by construction, and the computational plan is specified at the same level of detail as the econometrics, because a pre-registered analysis that cannot be reproduced is only half pre-registered.

The estimation environment is open-source and version-pinned. Panel fixed-effects estimation, cluster-robust and wild-cluster-bootstrap inference, and the within-between specification are implemented in a high-fixed-effects regression framework with reproducible random seeds for the bootstrap. The heterogeneity-robust difference-in-differences estimators are implemented from the reference implementations the methods authors provide: the group-time estimator of Callaway and Sant'Anna, the intertemporal estimators of de Chaisemartin and D'Haultfoeuille, the efficient imputation estimator of Borusyak, Jaravel, and Spiess, and the Goodman-Bacon decomposition, each cited to its source so the implementation is traceable [\[22\]](#ref-22), [\[23\]](#ref-23), [\[29\]](#ref-29), [\[109\]](#ref-109), [\[133\]](#ref-133). Instrumental-variable estimation uses standard two-stage least squares with cluster-robust first-stage diagnostics, and the imperfect-instrument sensitivity analysis is implemented with the dedicated tooling for causal inference under possibly-invalid instruments [\[111\]](#ref-111), [\[127\]](#ref-127). The synthetic difference-in-differences robustness lane for sparse, staggered adoption uses its reference implementation [\[99\]](#ref-99), [\[110\]](#ref-110).

Three reproducibility commitments are made explicit. First, every figure, table, and reported coefficient is produced by a script from the raw assembled panel with no manual intervention between data and output, so that the entire result set regenerates from a single command. Second, the variable-construction codebook (the single documentary rule for latency, the trigger-and-resolution event taxonomy, the coarse-versus-fine resolution mapping) is itself code, applied identically across all eras in keeping with the Maddison consistent-measurement standard, so that the construction of latency is auditable and not a matter of analyst judgment exercised case by case [\[16\]](#ref-16). Third, the pre-registration record, the analysis scripts, the power grid, and the constructed panel are versioned together, so that the time-stamped commitment and the executed analysis can be diffed against each other. The computational plan thus closes the loop opened by the pre-registration: the design is not only specified in advance but specified in a form that lets any reader confirm the executed analysis is the one that was pre-registered.

## 5.11 How this chapter advances the argument
This chapter shows that the design reaches the causal mechanism and improves on the alternatives. Two of the dissertation's load-bearing claims belong to earlier chapters: that the problem is real and material is established in the introduction and literature chapters, and that the variables are measurable from primary records is established in the data chapter. This chapter establishes that the design addresses the causal mechanism (fixed effects remove the persistent difficulty confound, instruments address the residual phase-specific endogeneity, and the named mechanism from fragmented authority through accrued latency to cost, schedule, and cadence is tested through both the continuous-latency and discrete-regime doors), and that it improves on the alternatives (heterogeneity-robust estimators replace the demonstrably biased two-way fixed-effects default for regime analyses, and the within estimator is chosen against pooled and naive random-effects alternatives for cause). The question of residual risk is carried by the threats-to-validity, power, and pre-registration sections, which name every material residual risk (endogeneity, baseline gaming, reverse causation, construct fragility, small-cluster inference, underpowering) and attach to each a pre-committed check or contingency, so that the residual risk is not eliminated (it cannot be, in an observational single-agency design) but is bounded, disclosed, and managed.

The honest summary of the chapter's confidence is graded, not uniform. Confidence that the fixed-effects baseline removes time-invariant difficulty is high, because that is what the within transformation does by construction. Confidence that the instrumented estimate recovers a causal effect is conditional, ranging from high to low depending on the realized first-stage strength and the imperfect-instrument bounds, and the design commits to reporting that conditionality rather than collapsing it. Confidence that the staggered-regime analyses avoid the negative-weighting pathology is high, because the heterogeneity-robust estimators are designed for that purpose and the Goodman-Bacon diagnostic makes the pathology visible if it appears. Confidence in the power of the full-span test is moderate at design stage and is made high or low only by the realized variances, with a pre-committed contingency either way. What would raise the overall confidence is what the design pre-registers to collect: a strong first stage, imperfect-instrument bounds that exclude zero, agreement across the two latency resolutions and the heterogeneity-robust estimators, and a realized power grid that clears the operationally meaningful effect size. What would lower it, and would be reported rather than hidden, is a weak first stage, wide sensitivity bounds, resolution-dependence, or a realized MDE above the meaningful threshold. The design is built so that either the raising or the lowering is legible to the reader, which is the property a falsifiable contribution requires of its research design.



## Chapter 6. Analysis Plan and Expected Results

## 6.0 The answer this chapter gives

This chapter fixes, in advance and in writing, exactly what will be done to the assembled program-phase panel, exactly how the result will be read, and exactly what reading would falsify the contribution. That is its whole purpose, and the answer is worth stating before it is developed. The answer is a pre-registered five-step procedure with a single binding decision rule: the procedure assembles and validates the panel against published figures, describes every variable by era before any model is fit, estimates the fixed-effects baseline of the bible specification for each of the three outcomes, re-estimates under an instrumental-variable strategy for endogeneity, and then subjects the result to a robustness and heterogeneity battery; the decision rule rejects the null only if the latency coefficient is statistically distinguishable from zero, carries the predicted sign in the same direction across the fixed-effects and instrumented specifications, and survives the robustness battery, and it retains the null otherwise. The chapter also states the directional expectations under the alternative, with a named mechanism behind each expected sign, and it leaves the illustrative coefficient table deliberately unpopulated. Nothing in this chapter is an executed estimate. Every number that appears is labeled either as a target drawn from the published literature against which a constructed measure will be validated, or as an illustrative placeholder with no empirical content. The discipline is the discipline of the design stage: write the plan so completely that, once the data are in hand, no remaining degree of freedom lets a result be manufactured.

The problem this chapter addresses can be stated in the current-state, desired-state, gap, consequence frame that governs the dissertation. The current state of the practitioner conversation about NASA program execution is that overruns are attributed to slow decision-making in narrative and anecdote, and the current state of the quantitative literature is that no consistent latency series exists and no analysis plan has been committed to in advance of estimation, so that any future regression of latency on cost growth would be exposed to the charge of specification search after the fact. The desired state is a procedure specified so completely before estimation that the result, whatever it is, is credible because the analyst could not have steered toward it. The gap is the absence of such a pre-registered plan for this specific panel, with this specific explanatory variable, under the specific endogeneity and heterogeneity threats that Chapters 4 and 5 identified. Leaving the gap open carries a consequence: a latency study would inherit the credibility problem that the megaproject literature attributes to optimistic forecasting and strategic misrepresentation [\[13\]](#ref-13), [\[14\]](#ref-14), where the analyst's discretion, not the data, determines the headline. This chapter closes the gap by removing the analyst's discretion in advance.

The chapter proceeds in the order in which the work will actually be done. Sections 6.1 through 6.5 walk through the five estimation steps, each developed as a substantive design decision rather than a checklist item. Section 6.6 states the pre-registered decision rule verbatim and enumerates the falsification conditions. Section 6.7 states the expected signs as directional expectations with mechanisms, explicitly not as findings, and presents the unpopulated illustrative table together with the methodological argument for refusing to fill it. Section 6.8 describes the illustrative simulation that will be used to check the estimation code and to calibrate statistical power before the real data are touched, Section 6.9 describes how the event-study and latency-profile output will be interpreted, and Section 6.10 fixes the pre-registration record and the analytic sequencing that make the plan binding. Throughout, three commitments of quantitative economic history govern the design: consistent single-rule measurement before comparison, in the Maddison tradition [\[16\]](#ref-16), [\[26\]](#ref-26); the bounded counterfactual reported as a range rather than a point, in the Fogel tradition [\[17\]](#ref-17); and the institutional reading of latency as a durable transaction cost that motivates era fixed effects, in the North tradition [\[18\]](#ref-18).

The citation form the plan uses warrants one clarification before the steps begin. Every reference in this dissertation, whether a seed reference that anchors the design or a corpus-expansion reference that supplies supporting breadth, is cited by a single bracketed number that resolves to one entry in the consolidated reference list, so that the provenance of every claim is traceable to one real record. The numbering runs in a single sequence: the twenty-five seed references occupy [\[1\]](#ref-1) through [\[25\]](#ref-25), so that [\[1\]](#ref-1) is Emmons, Bitten, and Freaner on historical NASA cost and schedule growth and [\[12\]](#ref-12) is Decarolis and colleagues on bureaucratic competence and procurement outcomes; the Hall-of-Shoulders anchor references follow; and the corpus-expansion references continue the sequence. No claim carries more than one citation form, and every bracketed number in the body resolves to exactly one numbered entry in the reference list.

## 6.1 Step one: assemble and validate the panel

The first step is to assemble the program-phase panel from the three named sources of Chapter 4, the NASA historical budget and program records, the NASA Technical Reports Server documentation, and the Government Accountability Office major-project assessments, and then to validate the constructed cost and schedule measures against figures that already exist in the published record. This step supports the conclusion that the constructed measures reproduce known values where known values exist, because several of the outcome quantities the panel constructs have already been computed and published, independently, by the cost-estimating community. The connecting principle is elementary: a new instrument earns trust by reproducing a known reading before it is asked to produce an unknown one. The Maddison standard rests on the same logic, in which a long-run series is credible only because the same construction rule, applied to a period whose value is independently known, returns that known value [\[16\]](#ref-16), [\[26\]](#ref-26). Validation is possible only over the overlap set, the programs for which a published cost-growth or schedule-growth figure exists, and one must guard against the possibility that a measure reproducing published values on the modern, well-documented overlap might still drift on the sparse early decades where no external check exists; the response to that concern is carried by the two-resolution latency design of Chapter 4 and by the descriptive diagnostics of Step two.

Concretely, the validation targets are the historical cost-and-schedule growth figures reported by the Aerospace Conference cost-estimating series and the GAO assessments. Emmons, Bitten, and Freaner constructed distributions of NASA cost and schedule growth in order to set reserve guidelines, and they separated growth into externally programmatic and internally technical components [\[1\]](#ref-1); the subsequent Aerospace Conference papers reported historical mass, power, schedule, and cost growth for NASA science instruments and for NASA spacecraft [\[2\]](#ref-2), [\[3\]](#ref-3), and related instrument schedule growth to mission cost and schedule growth [\[130\]](#ref-130). Majerowicz and Shinn reported the relationship between schedule delays and cost overruns directly [\[5\]](#ref-5). The National Research Council's consensus study tabulated cost-growth magnitudes at the agency level for Earth and space science missions [\[6\]](#ref-6). These are the published readings the constructed panel must reproduce. The validation procedure is therefore mechanical and pre-specified: for every program-phase that also appears in one of these published sources, compute the panel's cost-growth and schedule-slip values, compare them to the published values, and report the distribution of discrepancies. The acceptance criterion, fixed here, is that the panel's reconstructed growth figures fall within the rounding and definitional tolerance of the published figures for the overlap set; material and systematic discrepancy on the overlap set is treated as a construction error to be diagnosed and corrected before any model is fit, not as a finding.

This validation step is interpreted, not merely listed. What the convergence of the Aerospace Conference series, the NRC study, and the GAO assessments means for the present design is that the outcome side of the panel is anchored to an external, multiply-reported standard, so that the novel and unanchored part of the construction is isolated to the explanatory variable, authorization latency, which no prior source has built. That isolation is deliberate. If the outcomes are validated against published figures and only latency is new, then any association the later steps find between latency and the outcomes cannot be an artifact of mismeasuring the outcomes, because the outcomes were forced to match independent measurements first. Confidence that the validation step is feasible is high, because the published growth figures exist and are stable across report years; confidence that the validation will succeed without correction is moderate, because baseline definitions drifted across eras and the earliest observations carry construction uncertainty that the overlap set cannot fully discipline. Evidence that would raise the latter confidence is a wide overlap set extending into the earlier decades; evidence that would lower it is systematic discrepancy concentrated in any single era, which would flag an era-specific construction problem.

## 6.2 Step two: describe before estimating

The second step is to report the univariate distributions of authorization latency, cost growth, schedule slip, and mission cadence, by era, with constant-dollar deflation applied to cost via the NASA New Start Inflation Index, before any regression is estimated. This descriptive step is itself a contribution and not a preliminary throat-clearing exercise, because a consistent long-run series of NASA authorization latency does not currently exist. The cost-estimating literature has measured cost and schedule growth but has never constructed a documentary-rule-based latency series spanning 1958 to 2026 [\[1\]](#ref-1), [\[2\]](#ref-2), [\[3\]](#ref-3), [\[5\]](#ref-5), [\[6\]](#ref-6), and the public-administration literature has measured red tape and administrative burden with perceptual scales and procedural counts in other organizational settings but has never applied a single consistent latency rule to one agency across its full history [\[7\]](#ref-7), [\[8\]](#ref-8), [\[10\]](#ref-10), [\[11\]](#ref-11). The Maddison precept governs the order of work: quantification of the proximate variable is the precondition of any comparison or causal claim, so that building and displaying the latency series is logically prior to, and independently valuable from, any regression that uses it [\[16\]](#ref-16), [\[26\]](#ref-26). The Maddison Project's own practice bears this out, treating the construction and publication of the constant-unit GDP series as a contribution in its own right, separate from any growth-accounting exercise built on top of it [\[16\]](#ref-16), [\[26\]](#ref-26). The descriptive series inherits the construction uncertainty of the early decades and the two-resolution split, so the coarse and fine latency series are described separately and never spliced into a single line that would imply a continuity the documents do not support.

The descriptive output is specified in advance to forestall selective reporting. For each era regime defined in Chapter 5 and mapped to the era fixed effect, the analysis will report the central tendency and dispersion of latency at both the coarse milestone-to-milestone resolution and, where available, the fine key-decision-point resolution, alongside the era distributions of cost growth in constant dollars, schedule slip, and cadence under each program-family definition. The describe-before-estimate rule also fixes the order of operations: the descriptive tables are produced and inspected before the estimation code is run on the real outcomes, so that the analyst sees the marginal distributions, the missingness pattern, and the unbalanced structure of the panel before seeing any conditional association. This ordering guards against the optimism-bias and hindsight pathologies that the megaproject literature documents, in which analysts unconsciously steer toward an expected result once they have seen it [\[13\]](#ref-13), [\[14\]](#ref-14), [\[30\]](#ref-30). By committing to describe first, the design ensures that the shape of the latency distribution, including any fat tail analogous to the fat-tailed overrun distributions that Flyvbjerg reports for megaprojects [\[13\]](#ref-13), is documented as a descriptive fact rather than discovered as a convenient one.

The interpretive payoff of Step two is that it converts the practitioner anecdote into a measured marginal distribution. The Constellation and Space Launch System examples invoked in Chapter 1 are, as that chapter conceded, vivid anecdotes rather than measurements; the Constellation life-cycle cost record is a single program's documentary trace [\[102\]](#ref-102), and the inter-flight interval figure is program-record-derived rather than peer-reviewed. Step two replaces the anecdote with the full distribution of latency across all program-phases in each era, so that the question shifts from whether one program waited a long time to whether long waits are systematic, era-patterned, and co-distributed with poor outcomes. Confidence that this descriptive contribution will stand is very high, because it depends only on the documentary construction and the deflation, both of which are mechanical once the construction rule is fixed; it does not depend on any regression succeeding. This is the sense in which the first part of the contribution stands even if the null is ultimately retained, a point the introduction and conclusion both carry.

## 6.3 Step three: the fixed-effects baseline

The third step estimates the bible specification for each of the three outcomes. For program i, phase p, and era t, the specification is, verbatim from the shared notation:

\[
\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t) \qquad\qquad (1)
\]

where Outcome is in turn cost growth, schedule slip, and cadence; Latency is authorization latency in months; X is the vector of technical and programmatic controls of Chapter 4 (mass, power, technology readiness level at commitment, mission class, contract type, number of external partners, and the funding-instability index); \(\alpha_i\) are program fixed effects; \(\delta_t\) are era fixed effects; \(\varepsilon\) is the error; and \(\beta\) is the coefficient of interest. Under the alternative hypothesis, \(\beta\) is positive for cost growth and schedule slip and negative for cadence. Standard errors are clustered by program to allow arbitrary within-program correlation across phases, and, because the number of program clusters is modest, inference is reported additionally under the wild-cluster bootstrap so that the clustered standard error is not the sole basis for a rejection decision.

Step three identifies the latency coefficient from within-program and within-era variation, net of time-invariant program difficulty and common era shocks. Program fixed effects absorb every characteristic of a program that does not change across its phases, including the unobserved baseline difficulty that the cost-estimating literature treats as the dominant driver of growth [\[1\]](#ref-1), [\[2\]](#ref-2), [\[19\]](#ref-19), and era fixed effects absorb the rule regimes and appropriations conditions that move all programs in a period together. The standard panel-identification argument carries the logic: once the two-way fixed effects are partialled out, the residual variation in latency is variation within a program across its phases and within an era across programs, which is the variation the design intends to exploit [\[116\]](#ref-116), [\[126\]](#ref-126), [\[128\]](#ref-128). The panel-econometric literature on the choice between fixed and random effects and on what fixed-effects regression does and does not identify supplies the foundation [\[116\]](#ref-116), [\[117\]](#ref-117), [\[125\]](#ref-125), [\[126\]](#ref-126), [\[128\]](#ref-128). Program fixed effects, as Chapter 5 notes, remove only the fixed component of difficulty, not difficulty that varies by phase, so the baseline coefficient is interpreted as conditional on the maintained assumption that phase-varying difficulty is uncorrelated with latency, an assumption that Step four exists to relax. The design must also answer the objection that the two-way fixed-effects estimand can be contaminated when treatment timing is staggered and effects are heterogeneous [\[21\]](#ref-21), [\[24\]](#ref-24), [\[109\]](#ref-109); the response, developed in Step five and in Chapter 5, is that latency enters as a continuous within-program variable rather than a binary staggered treatment, which sidesteps the worst of the negative-weighting problem, and that any discrete-regime analysis uses heterogeneity-robust estimators rather than naive two-way fixed effects [\[22\]](#ref-22), [\[23\]](#ref-23), [\[29\]](#ref-29).

Three design decisions internal to Step three are fixed here. First, the three outcomes are estimated separately rather than stacked, because cost growth and schedule slip are phase-level fractions while cadence is most naturally defined at the era level within a program family, and forcing them into one system would impose a cross-outcome error structure the data do not warrant. Second, the cadence regression uses era fixed effects and family definitions rather than program fixed effects, because cadence is a family-level rate, not a program-phase fraction, a distinction the illustrative table preserves. Third, the technical and programmatic controls are entered in a fixed order with the funding-instability index always present, because that index is the control most directly aimed at the common-cause rival in which a period of budget turbulence drives both latency and cost growth without latency causing cost growth; the appropriations-cycle literature supplies the construction logic for that index [\[88\]](#ref-88), [\[129\]](#ref-129). Confidence that the baseline can be estimated is high; confidence that the baseline coefficient is the causal effect is low to moderate by design, because the endogeneity threat is real and unaddressed until Step four, and the design states this rather than concealing it.

The choice to enter latency as a continuous regressor, rather than to dichotomize it into high and low latency, is itself a substantive identification decision and is fixed here with its reasoning. A binary high-latency treatment switching on at staggered calendar times across programs is exactly the configuration in which the two-way fixed-effects difference-in-differences estimand decomposes into a weighted average of two-group comparisons that can carry negative weights, the pathology that Goodman-Bacon characterizes and that the heterogeneity-robust literature was developed to repair [\[21\]](#ref-21), [\[22\]](#ref-22), [\[23\]](#ref-23), [\[29\]](#ref-29), [\[109\]](#ref-109), [\[135\]](#ref-135). By keeping latency continuous, the design exploits the full within-program and within-era gradient of the variable and avoids manufacturing a staggered binary treatment where the underlying construct is a continuous duration; the continuous-treatment difference-in-differences literature confirms that the continuous form is the cleaner object when the treatment is naturally a dose rather than a switch [\[100\]](#ref-100). This decision does not eliminate the heterogeneity concern, because heterogeneous slopes across programs and eras can still bias a pooled continuous coefficient, but it removes the most acute negative-weighting failure and confines the residual heterogeneity to a form the Step-five battery can probe. One caveat is that the continuous specification assumes a roughly linear latency-outcome relationship over the observed range; the design tests this assumption by re-estimating with latency entered flexibly, for example in deciles or with a spline, and reports whether the linear coefficient is a faithful summary of a possibly nonlinear profile, since a fat-tailed latency distribution analogous to the fat-tailed overrun distributions of the megaproject literature [\[13\]](#ref-13) could place most of the effect in the upper tail.

The clustering and inference choices are equally pre-committed, because with a modest number of program clusters the standard cluster-robust variance estimator can understate uncertainty and over-reject. The design clusters by program as the primary specification, to allow arbitrary within-program correlation across a program's phases, and reports the wild-cluster bootstrap as the primary basis for the rejection decision rather than the asymptotic clustered standard error, because the asymptotic approximation is unreliable at the realized cluster count. It additionally reports inference under alternative clustering, for example by era and two-way by program and era, so that the rejection decision is shown to be insensitive to the clustering dimension. The panel-econometric literature on the choice and interpretation of fixed-effects models, and on inference with time-series cross-section data, governs these choices and is cited as their backing [\[93\]](#ref-93), [\[116\]](#ref-116), [\[117\]](#ref-117), [\[125\]](#ref-125), [\[126\]](#ref-126), [\[128\]](#ref-128). The reason this matters for the present study is that the panel is unbalanced by construction, dense in the modern GAO-assessed era and sparse in the earliest decades, so the cluster count and the imbalance interact, and the simulation of Section 6.8 is what calibrates how the wild-cluster bootstrap behaves under that structure before any real inference is reported.
## 6.4 Step four: addressing endogeneity

The fourth step re-estimates the specification under an instrumental-variable strategy and reports the first-stage strength alongside a comparison of the fixed-effects and instrumented coefficients. The instrumented estimate identifies the effect of latency once the portion of latency driven by phase-specific difficulty has been purged. The central internal threat, named in Chapter 5, is that harder phases both take longer to authorize and overrun more, so the raw within-program association could reflect difficulty rather than process. Standard instrumental-variable logic answers it: if an instrument moves latency but is plausibly unrelated to the technical difficulty of the specific phase, the component of latency it isolates is exogenous to difficulty, and the coefficient it identifies is the latency effect rather than the difficulty effect. Decarolis and colleagues support the approach. They used an instrumental-variable strategy on contract-level federal procurement data to show that bureaucratic competence causally reduces delays and cost overruns, establishing both that the administrative side of execution moves outcomes and that an instrument can credibly isolate it [\[12\]](#ref-12). An instrument is only as good as its exclusion restriction, which cannot be tested directly and must be argued. The design therefore reports both the fixed-effects and the instrumented estimates, so the reader can see how much the conclusion depends on the instrument, and it treats the instrument's validity as argued and tested rather than assumed.

Two candidate instruments are pre-specified, exactly as in Chapter 5. The first is the contemporaneous workload of the authorizing office, measured as the number of other programs awaiting authorization at the same office in the same period. The mechanism is administrative congestion: when an authorizing office is processing many pending actions, each action waits longer for reasons of queue length rather than the technical difficulty of any one phase, so office workload shifts latency through a channel plausibly orthogonal to phase difficulty. The second is the timing of the action relative to the appropriations calendar, which paces authorization for budget-dependent reasons that are independent of program difficulty; the budget-and-appropriations-cycle literature supplies the institutional grounding for this instrument, documenting how the annual cycle and its recurrent delays structure the timing of budget-dependent actions [\[88\]](#ref-88), [\[129\]](#ref-129). The design reports the first-stage F-statistic and the partial correlation of each instrument with latency. Because an instrument may be only approximately valid, it also applies tools designed for inference under possibly invalid instruments, so a marginal exclusion violation is bounded rather than ignored [\[111\]](#ref-111), [\[127\]](#ref-127). The design must answer the objection that office workload could itself be correlated with a difficult era in which many hard programs are authorized together. The response is that era fixed effects already absorb the common era component, so the instrument is asked to supply only within-era, cross-program variation in congestion, which is the variation least confounded with any single phase's difficulty.

The validity of each instrument is argued and probed rather than asserted, and the design fixes the probes in advance. The exclusion restriction for the office-workload instrument is that the number of other programs awaiting authorization at the same office in the same period affects a given phase's outcomes only through that phase's latency, and not through any channel correlated with the phase's own difficulty. The design probes this by testing whether office workload predicts the technical controls of the focal phase, since a workload measure correlated with the focal phase's mass, power, or technology readiness would signal that congestion is proxying for a difficulty wave rather than for queue length. The exclusion restriction for the appropriations-timing instrument is that the position of an action within the annual budget cycle affects outcomes only through the latency the cycle imposes, and not through a contemporaneous shock to program difficulty. The design probes this by exploiting the institutional fact that the appropriations calendar is set by congressional process rather than by any single program's engineering, a fact the budget-cycle literature documents [\[88\]](#ref-88), [\[129\]](#ref-129), and by testing the instrument's balance against the focal-phase controls. Because no exclusion restriction can be verified directly, the design also reports second-stage inference under tools built for possibly invalid instruments, so a marginal violation of either restriction is bounded rather than assumed away, and the sensitivity of the coefficient to the admissible degree of invalidity is reported as part of the result [\[111\]](#ref-111), [\[127\]](#ref-127). Over-identification, when both instruments are used jointly, supplies a further internal check: agreement of the two instruments estimated separately, and a passing over-identification test when they are combined, is evidence that neither is grossly violating its exclusion restriction, while disagreement is reported as a limit on the strength of the causal reading rather than resolved by selecting the more convenient instrument.

The interpretation of Step four is pre-committed. If the instrumented coefficient retains the sign and rough magnitude of the fixed-effects coefficient and the first stage is strong, the design reads this as evidence that the baseline association is not merely difficulty masquerading as process. If the instrumented coefficient collapses toward zero or reverses while the first stage is strong, the design reads this as evidence that the baseline association was substantially driven by phase-specific difficulty, which under the decision rule counts against the alternative. If the first stage is weak, the instrumented estimate is uninformative and is reported as such, with the fixed-effects estimate carrying the inference under its stated, weaker assumption. A weak instrument is never permitted to manufacture a rejection, and the first-stage F-statistic threshold below which the instrument is declared weak is fixed before estimation, so the weak-instrument judgment is not itself a discretionary choice made after seeing the second stage. Confidence in the instrumental-variable step is moderate. It would be raised by a strong, theoretically clean first stage, by agreement between the two instruments estimated separately, and by a passing over-identification test; it would be lowered by a weak first stage, by disagreement between the instruments, or by sensitivity of the second-stage coefficient to the possibly-invalid-instrument bounds.

## 6.5 Step five: robustness and heterogeneity

The fifth step subjects the baseline and instrumented results to a battery of robustness and heterogeneity checks, and reports central estimates as bounded ranges conditioned on stated assumptions rather than as single points. The conclusion, in either direction, should survive reasonable variation in the analytic choices, so that it is a property of the data and not of one specification. The catalog of fragilities identified across Chapters 4 and 5 sets the agenda: the two latency resolutions, the multiple cadence-family definitions, the staggered-regime concern, the modest cluster count, and the unbalanced panel. The reasoning is that a finding appearing only under one resolution, one cadence definition, or one estimator is not a finding about NASA but a finding about a choice, and the design's job is to separate the two. The bounded-range reporting standard follows the Fogel discipline, in which the central quantity is reported as an upper-and-lower-bounded range conditioned on explicit assumptions rather than as a single causal point, because the counterfactual cannot be run experimentally and the honest object is an interval [\[17\]](#ref-17). No robustness battery is exhaustive; the design pre-commits to a specific, enumerated set, so the battery itself is not a venue for post hoc search.

The battery is enumerated here and fixed. First, every estimate is produced at both the coarse and the fine latency resolution, and a result is credited only if it holds at the resolution for which it is defined; the fine resolution is available only for the modern subperiod, so a result present at the coarse full-span resolution but absent at the fine modern resolution is reported as exactly that, not averaged away. Second, the cadence regressions are re-estimated under each program-family definition, because cadence is the most construct-fragile outcome and the design reported it under alternatives by construction. Third, for any analysis that estimates the effect of a discrete administrative-regime change on latency or outcomes, the design uses the heterogeneity-robust estimators of Callaway and Sant'Anna and of de Chaisemartin and D'Haultfoeuille rather than the two-way fixed-effects difference-in-differences estimator, and reports the Goodman-Bacon decomposition as a diagnostic of how much the naive estimand would have weighted each comparison, including any negative weights [\[21\]](#ref-21), [\[22\]](#ref-22), [\[23\]](#ref-23), [\[29\]](#ref-29). The modern staggered-difference-in-differences literature is the source for this estimator menu and for the event-study and continuous-treatment variants the design will draw on [\[99\]](#ref-99), [\[100\]](#ref-100), [\[104\]](#ref-104), [\[109\]](#ref-109), [\[110\]](#ref-110), [\[132\]](#ref-132), [\[133\]](#ref-133), [\[134\]](#ref-134), [\[135\]](#ref-135), [\[136\]](#ref-136). Fourth, inference is reported under alternative clustering and under the wild-cluster bootstrap given the modest number of program clusters, and under alternative fixed-effects structures, so that the rejection decision is not specification-dependent. The Imai and Kim result on when two-way fixed-effects regression identifies a causal effect guides the interpretation of the fixed-effects specification and its maintained assumptions throughout [\[24\]](#ref-24).

The interpretation of Step five is that the robustness battery is where the falsification conditions of the decision rule are operationalized. A latency coefficient that is significant and correctly signed in the baseline but vanishes under the heterogeneity-robust estimators, or reverses between latency resolutions, or appears only under one cadence-family definition, does not reject the null under the pre-registered rule; it is recorded as a fragile association that fails the standard. Confidence that the battery is the right battery is high, because it is constructed to map one-to-one onto the named threats of Chapter 5 rather than to maximize the chance of a rejection. The bounded-range reporting means that the deliverable of the whole procedure, for each outcome, is a stated interval for β conditioned on the resolution, the cadence definition, the estimator, and the clustering, accompanied by the sign and the survival record across the battery, rather than a single starred coefficient.

## 6.6 The pre-registered decision rule and the falsification conditions

The decision rule is stated here in the form fixed by the shared bible, and it is fixed in advance precisely to prevent specification search after the data are seen. The null hypothesis is rejected only if the latency coefficient is statistically distinguishable from zero at conventional levels and carries the predicted sign in the same direction across the fixed-effects and instrumented specifications and survives the robustness battery of Step five. If the coefficient is indistinguishable from zero, or changes sign between specifications, or vanishes under the heterogeneity-robust estimators or under the alternative latency resolution, the null is retained and the contribution's hypothesis test fails. This conjunctive structure is the point: the rule requires statistical significance, the correct sign, consistency across the fixed-effects and instrumented estimates, and survival of robustness simultaneously, so that no single specification can carry a rejection on its own.

The falsification conditions follow directly and are enumerated so that they are binding rather than discretionary. The contribution's hypothesis is falsified by any one of the following outcomes: a latency coefficient indistinguishable from zero in the preferred specification; a coefficient whose sign reverses between the fixed-effects and instrumented specifications; a coefficient that vanishes under the heterogeneity-robust estimators or under the alternative latency resolution; or a demonstration that the apparent association is fully explained by baseline gaming or by reverse causation. The baseline-gaming check is the test, specified in Chapter 5, of whether baseline conservatism, proxied by the ratio of held reserves to baseline, correlates with latency in a way that would make measured growth an artifact of optimistic baseline choice rather than of process; the megaproject and optimism-bias literature is the reason this rival is taken seriously, since it documents that public-project baselines are set optimistically and that strategic misrepresentation is a systematic feature of large-project forecasting [\[13\]](#ref-13), [\[14\]](#ref-14), [\[30\]](#ref-30), [\[72\]](#ref-72). The reverse-causation check is the use of latency measured early in each phase, before trouble has had time to generate authorization events, together with the instrumental-variable strategy, so that the design does not credit an association in which overrunning programs generate latency rather than latency generating overruns.

The honesty of this rule rests on a single feature: a decision rule fixed before estimation removes the analyst's discretion over what counts as a confirming result, the discretion that the optimism-bias literature identifies as the channel through which forecasts and evaluations drift toward the analyst's prior [\[13\]](#ref-13), [\[14\]](#ref-14), [\[30\]](#ref-30). The broader methodological position is that a falsifiable proposition is one for which the conditions of failure are specified in advance and could actually occur; the design satisfies this because the null is a real possibility, given that latency may be endogenous to difficulty and that any raw correlation could reflect difficulty rather than process. Pre-registration disciplines the confirmatory test but does not forbid exploratory analysis; any analysis beyond the five pre-registered steps is labeled exploratory and is excluded from the rejection decision. One might object that a conjunctive rule is too demanding and risks a false retention of the null. That risk is acknowledged and accepted as the deliberate price of credibility, because the design would rather retain a true null than reject a false one, since a spurious rejection in this literature would feed exactly the unmeasured practitioner narrative the dissertation exists to discipline. Either outcome is informative under this rule, which is the mark of a well-posed test.

## 6.7 Expected signs, mechanisms, and the deliberately unpopulated table

This section states the directional expectations under the alternative hypothesis, with a named causal mechanism behind each, so that the test is interpretable. These are expectations, not findings. No estimate has been produced from the full dataset, and the directional statements below describe what the alternative predicts, not what the data have shown.

The expected sign of the latency coefficient in the cost-growth regression is positive. The mechanism, stated as a chain, is that fragmented decision authority and sequential, multi-layer authorization gates, paced by an annual appropriations cycle, cause each pending action to accrue elapsed time during which standing program costs continue to be incurred and requirements drift; the observable effect is higher measured latency per program-phase; the operational consequence is larger cost growth relative to baseline, because time-dependent standing costs accumulate during the waiting interval and because delay opens a window for requirements change. This is not a bare correlation. The transaction-cost reading of North supplies the driver, in which the time to measure a proposed action against the institution's rules and to obtain authorization is itself a cost internal to the organization [\[18\]](#ref-18), and the cost-estimating literature supplies the observable channel, in which schedule-related growth and cost growth move together [\[5\]](#ref-5), [\[130\]](#ref-130). The Decarolis procurement evidence supplies the external precedent that administrative process moves cost outcomes in government contracting [\[12\]](#ref-12). Confidence in the predicted sign is moderate; it would be raised by agreement between the fixed-effects and instrumented estimates and by survival under both latency resolutions, and it would be lowered if the coefficient depended on the standing-cost channel being mismeasured.

The expected sign of the latency coefficient in the schedule-slip regression is also positive, and this is expected to be the largest and most robust of the three associations. The mechanism is the most direct of the three, because authorization latency is itself a component of elapsed schedule: time spent waiting for authorization is time the phase has not progressed, so the path from latency to schedule slip is partly mechanical and partly behavioral, with the behavioral part being the requirements drift and rework that delay invites. One qualification matters, and the interpretation must carry it: because latency is partly definitionally inside elapsed schedule, the schedule-slip association is the one most exposed to the objection that it is tautological rather than causal, and the design's response is to measure latency early in the phase and to instrument it, so the estimated association reflects the behavioral and standing-cost channels rather than the pure accounting identity. Confidence in the predicted sign for schedule slip is moderate to high on the direction and lower on the interpretation as causal rather than partly mechanical; the early-in-phase measurement and the instrument are what would raise the interpretive confidence.

The expected sign of the latency coefficient in the cadence regression is negative. The mechanism is that higher latency per decision, accumulated across the many authorization events of a program family over time, lengthens the interval between successive flight or delivery events and therefore lowers the realized number of operational mission events per period within that family. The strategic implication, stated in program-execution-management terms and not in any architecture-layer vocabulary, is that a portion of NASA's realized mission cadence would then be a controllable process variable distinct from irreducible engineering difficulty, a lever especially relevant to the long-cycle deep-space programs in which authorization events accumulate over many years. Confidence in the cadence sign is low to moderate, lower than for the other two outcomes, because cadence is the most construct-fragile outcome and depends on the program-family definition; the multiple-family reporting is the response, and agreement of the negative sign across family definitions is the evidence that would raise confidence, while sign instability across definitions is the evidence that would lower it and, under the decision rule, would count against rejecting the null on cadence.

When the procedure is eventually executed, the reporting object for each outcome is a bounded range rather than a single coefficient, in the Fogel manner, and stating this now keeps the expected-sign language of this section from later being misread as licensing point estimates [\[17\]](#ref-17). Fogel's discipline, carried throughout the dissertation, is that a counterfactual claim of the form "the program would have cost less or flown sooner under faster authorization" cannot be settled by a single causal point, because the counterfactual world cannot be run experimentally; the honest object is an interval conditioned on stated assumptions, with an upper and a lower bound that name the assumptions under which each is reached. The within-program and within-era comparison the fixed-effects design supports is the feasible approximation to that unrunnable experiment, and the bounded range, conditioned on the latency resolution, the cadence-family definition, the estimator, and the clustering, is what the procedure will deliver for each β. The expected signs of this section therefore predict the side of zero on which the bounded range will fall under the alternative, not a magnitude; predicting a magnitude in advance of estimation would be exactly the fabrication the design refuses.

The illustrative table below shows the shape of the output the procedure will produce. The numbers are not present. The cells that would hold coefficients and intervals are left as placeholders, and this is a methodological choice, not an omission. Reporting fabricated coefficients as if they were estimates would violate the falsifiability standard on which the dissertation is built, and would reproduce in miniature exactly the strategic-misrepresentation pathology the megaproject literature documents [\[13\]](#ref-13), [\[14\]](#ref-14). The table is therefore presented as a specified-but-unpopulated artifact, to be completed only after the panel is assembled and the five-step procedure is executed.

| Outcome | Latency coefficient (per +1 month) | Cluster-robust interval | Specification |
|---|---|---|---|
| Cost growth (fraction of baseline) | [sign expected +; not estimated] | [to be estimated] | Program + era FE, IV |
| Schedule slip (fraction of baseline) | [sign expected +; not estimated] | [to be estimated] | Program + era FE, IV |
| Mission cadence (events per period) | [sign expected −; not estimated] | [to be estimated] | Era FE, family-defined |

The sign-convention reminder is restated so it travels with the table: under the alternative hypothesis, β is positive for cost growth and schedule slip and negative for cadence; under the null, β is statistically indistinguishable from zero for all three. The bracketed sign annotations record the prediction, not a result, and the interval columns are empty by design because there is no executed estimate to report.

## 6.8 Design of the illustrative simulation

Before the real panel is touched, the estimation pipeline will be exercised on a synthetic panel generated under known parameters, and this section specifies that simulation. The purpose is twofold and strictly methodological: to verify that the estimation code recovers a known coefficient when one is built into the data, and to calibrate the statistical power of the design given the unbalanced panel and the modest number of program clusters. The synthetic data carry no empirical content about NASA; they exist only to test the machinery and to bound the design's ability to detect an effect of a given size. Every quantity in this section is illustrative by construction.

The simulation generates a program-phase panel that mirrors the structural features of the real panel without claiming its values. It draws a set of synthetic programs, assigns each an unobserved fixed difficulty as a program-level intercept, assigns each phase to one of the era regimes with a regime-level intercept, and generates synthetic latency with three components: a program component correlated with the synthetic difficulty, to reproduce the endogeneity threat; an era component, to reproduce the regime structure; and an idiosyncratic component, including the congestion and appropriations-timing variation that the instruments are meant to exploit, to reproduce the exogenous part the instrumental-variable strategy targets. Outcomes are then generated from the bible specification with a known β imposed, plus the synthetic fixed effects and a clustered error. The first test is recovery: with the known β imposed and the exogenous latency variation present, the fixed-effects and instrumented estimators applied to the synthetic data should return the imposed β within simulation error, and the instrumented estimator should recover it even when the program-difficulty-correlated component of latency is made large, while the naive fixed-effects estimator should be visibly biased in that case. A pipeline that fails this recovery test on synthetic data is debugged before it is ever applied to the real panel.
The second test is power. By imposing a sequence of β values ranging from zero upward and repeating the estimation across many synthetic draws, the design will trace the rejection rate of the decision rule as a function of true effect size, the realized cluster count, and the degree of panel imbalance. This power curve is the design's honest statement of what it can and cannot detect: it identifies the smallest latency effect the conjunctive decision rule could reliably reject the null against, given the data structure, and it shows how the wild-cluster bootstrap behaves at the modest cluster count so that the inference reported on the real data is calibrated rather than assumed. The simulation also stress-tests the heterogeneity-robust estimators by imposing heterogeneous and staggered regime effects in the synthetic data and confirming that the Callaway-Sant'Anna and de Chaisemartin-D'Haultfoeuille estimators recover the imposed effects where the naive two-way fixed-effects estimator does not, reproducing in a controlled setting the negative-weighting pathology that the Goodman-Bacon decomposition diagnoses [\[21\]](#ref-21), [\[22\]](#ref-22), [\[23\]](#ref-23), [\[29\]](#ref-29), [\[109\]](#ref-109), [\[135\]](#ref-135). This simulation supports the conclusion that the estimation machinery is correct and the design's detection limits are known before any real estimate is produced, resting on the standard practice of validating an estimator on data with a known answer before trusting it on data with an unknown one, the same logic that governs the published-figure validation of Step one, applied now to the estimator rather than the measure. Confidence that the simulation will be informative is high, because it is fully under the analyst's control. Even so, a simulation can only test the estimator against the data-generating process the analyst imposes, so a structural feature of the real data not built into the synthetic generator would not be caught here, which is why the simulation supplements, and does not replace, the published-figure validation of Step one and the robustness battery of Step five.

## 6.9 Interpreting the event-study and latency profile

The final interpretive component of the plan concerns how the event-study and latency-profile output will be read, since several of the heterogeneity-robust estimators produce event-study-style coefficient paths rather than single numbers. This section fixes the reading rules in advance so that a visually suggestive path is not over-read.

For any discrete administrative-regime change analyzed in Step five, the heterogeneity-robust estimators yield a profile of effects indexed by time relative to the regime change, and the design will display this profile as an event study with the pre-regime coefficients shown explicitly [\[99\]](#ref-99), [\[104\]](#ref-104), [\[133\]](#ref-133), [\[134\]](#ref-134). The pre-registered reading rule is that the pre-regime coefficients function as a placebo: if a regime that compressed authorization layers is associated with lower latency only after it took effect and not before, the pre-regime path should be flat and statistically indistinguishable from zero, and a non-flat pre-regime path is read as evidence of confounding or anticipation that undermines the causal reading, not as part of the effect. This is the event-study discipline that the modern staggered-difference-in-differences literature formalizes, in which the credibility of the post-treatment path rests on the flatness of the pre-treatment path [\[99\]](#ref-99), [\[104\]](#ref-104), [\[133\]](#ref-133). One caveat is that with a modest number of regime changes the pre-regime coefficients are themselves imprecisely estimated, so the flatness test is interpreted with the wild-cluster bootstrap inference rather than with naive standard errors, and a wide pre-regime confidence band is reported as a limit on the strength of the event-study reading rather than glossed over.

The latency profile itself, the descriptive trajectory of authorization latency across eras produced in Step two, is interpreted under a separate and explicitly weaker rule. The profile is a descriptive series, and the North reading of path dependence predicts that once a set of review and authorization rules is in place it generates increasing returns and locks in, so that latency should appear durable across eras rather than fluctuating freely [\[18\]](#ref-18). A latency profile that is high and persistent across multiple era regimes is consistent with this path-dependence reading, but the design pre-commits to treating that consistency as descriptive corroboration of the institutional theory, not as a causal claim about latency's effect on outcomes, which only the regression steps can address. This separation is the chapter's final guard: it keeps the descriptive contribution, which stands on its own and is valuable even under a retained null, cleanly distinct from the causal test, which is governed by the conjunctive decision rule and can fail.

## 6.10 The pre-registration record and the binding sequence

The plan acquires its force only if it is fixed before the data are seen and if the sequence in which the analyst confronts the data is itself constrained, and this final section states both. The credibility of the eventual result rests not on the analyst's good intentions but on a recorded, time-stamped commitment that removes the relevant discretion in advance. The recurring lesson of the optimism-bias and strategic-misrepresentation literature is that forecasts and evaluations in large public projects drift systematically toward the analyst's prior when discretion is available, regardless of the analyst's sincerity [\[13\]](#ref-13), [\[14\]](#ref-14), [\[30\]](#ref-30), [\[72\]](#ref-72). A pre-registration record converts the decision rule, the expected signs, the instrument set, and the robustness battery from claims the analyst could revise after seeing the data into commitments the analyst made before seeing it, so that any later deviation is visible as a deviation. The reference-class-forecasting tradition's central finding reinforces this: procedural pre-commitment, not exhortation, is what curbs the bias [\[30\]](#ref-30), [\[52\]](#ref-52), [\[54\]](#ref-54). Pre-registration constrains the confirmatory test and not the exploratory one, and the design preserves room for genuine discovery by labeling any analysis beyond the five steps as exploratory and excluding it from the rejection decision.

The pre-registration record is specified to contain four fixed elements, each committed before the panel is assembled. First, the hypotheses in their exact fixed form: the null that administrative decision-and-authorization latency has no association with cost growth, schedule slip, or mission cadence after program and era fixed effects and technical controls, against the alternative that longer latency is associated with greater cost growth, more schedule slip, and lower cadence. Second, the conjunctive decision rule of Section 6.6, including the requirement of consistent sign across the fixed-effects and instrumented specifications and survival of the enumerated robustness battery. Third, the expected signs of Section 6.7, recorded as directional predictions about the side of zero on which each bounded range will fall, with the weak-instrument first-stage threshold and the clustering and bootstrap choices fixed alongside them. Fourth, the instrument set and the enumerated robustness checks, so that neither the instruments nor the battery can be expanded or pruned after the data reveal which version is convenient. This record is the artifact that makes the falsification conditions binding rather than discretionary: the analyst commits, in advance, that the conditions under which the contribution fails are real, specified, and capable of occurring.

The binding sequence governs the order in which the analyst is permitted to see the data, and it is the operational complement to the record. The validation of Step one and the description of Step two are performed and inspected first, on the panel's construction and its marginal distributions, before the conditional associations of Steps three through five are estimated; the analyst sees that the constructed measures reproduce the published figures and sees the shape of every variable's distribution before seeing any latency coefficient. The illustrative simulation of Section 6.8 is run before the real estimation, so that the estimation machinery is validated against a known answer and the design's detection limits are calibrated before any real β is produced. Only then are the baseline, instrumented, and robustness estimates produced, in that fixed order, against the pre-registered rule. This sequencing is not bureaucratic ceremony; it is the mechanism by which the describe-before-estimate precept of the Maddison tradition [\[16\]](#ref-16), [\[26\]](#ref-26) and the bounded-counterfactual precept of the Fogel tradition [\[17\]](#ref-17) are enforced as an order of operations rather than left as aspirations. The objection that such sequencing is impossible to verify after the fact is met by the time-stamped pre-registration record, which fixes the commitments before the panel exists, so that the sequence is auditable against a document that predates the data.

The argument of the dissertation closes here in the design stage, and the binding record is the keystone that holds it. The problem is real and material because NASA programs systematically and consequentially grow in cost and slip in schedule [\[1\]](#ref-1), [\[2\]](#ref-2), [\[5\]](#ref-5), [\[6\]](#ref-6), [\[13\]](#ref-13); the design addresses the causal mechanism by operationalizing latency from documentary records and entering it net of program and era fixed effects with an instrument for the exogenous component [\[12\]](#ref-12), [\[18\]](#ref-18); it improves on the naive alternative by using heterogeneity-robust estimators with a Goodman-Bacon diagnostic rather than uncorrected two-way fixed effects [\[21\]](#ref-21), [\[22\]](#ref-22), [\[23\]](#ref-23), [\[29\]](#ref-29), [\[109\]](#ref-109), [\[135\]](#ref-135); and the residual risk is acceptable because endogeneity, baseline gaming, reverse causation, and construct fragility are each pre-named with an explicit check, and because the decision rule and the binding sequence are fixed in advance so that the answer, whatever it is, is one the analyst could not have manufactured. This chapter's contribution to that case is the operational guarantee that the test, once run, will be a test and not a confirmation.



## Chapter 7. Discussion

## 7.0 The chapter thesis

This chapter argues a single proposition: the research design in the preceding chapters is constructed so that both possible outcomes of the test are informative, and the interpretive burden of the dissertation therefore falls not on hoping for a particular result but on reading each result correctly. If the analysis rejects the null hypothesis in the predicted direction, the dissertation will have identified a portion of NASA cost growth and schedule slip that is attributable to a variable the agency controls directly, namely the elapsed time its own decision and authorization processes consume, and will have given the red-tape and administrative-burden constructs a dollar-and-month instantiation inside one agency over a sixty-eight-year horizon. If the analysis retains the null, the dissertation will have disciplined a widely repeated practitioner narrative by showing that it does not survive measurement, and will have redirected reform attention toward the technical and estimating factors the cost-growth literature already documents. The chapter develops this two-sided reading, weighs the rival explanations that must be defeated before any causal claim is warranted, states the limits of external validity, and fixes in advance the conditions under which the contribution fails. Throughout, the posture is design-stage: no estimates have been executed on the full panel, every expected sign is labeled as an expectation rather than a finding, and the illustrative coefficient table introduced in Chapter 5 remains unpopulated by design.

The problem this chapter addresses can be stated in the current-state, desired-state, gap, and consequence frame that organizes the dissertation. The current state is that the practitioner community attributes part of NASA's cost and schedule performance to slow internal decision-making, while the empirical literature that could test this attribution has never been assembled. The desired state is a reading protocol that converts either regression outcome into a defensible inference about whether process speed is a lever on program performance. The gap is that, absent such a protocol stated before estimation, a researcher confronting the eventual numbers would be free to narrate them in whichever direction the numbers happened to point, which is precisely the specification-search freedom the dissertation's pre-registration is designed to remove. The consequence of leaving that gap open is that the test, however carefully estimated, would carry no more evidentiary weight than the anecdotes it is meant to replace. This chapter closes the gap by committing the interpretation in advance.

The three commitments that structure the dissertation's method, drawn from Maddison, Fogel, and North, also govern its interpretation, and they decide which inferences the design licenses below. Maddison's requirement that comparison rest on a single, transparent measurement standard means that the latency series must be read as a constructed measure with known construction uncertainty, not as a natural quantity [\[16\]](#ref-16), [\[26\]](#ref-26). Fogel's insistence that any counterfactual claim name and cost its next-best alternative, and report a bounded range rather than a point, means that every statement of magnitude in this chapter is conditional and bounded [\[17\]](#ref-17). North's location of organizational performance in the transaction costs that institutions raise or lower, together with his account of adaptive efficiency and path dependence, supplies the causal vocabulary in which the latency mechanism is expressed and the reason the era fixed effects are required [\[18\]](#ref-18). Each commitment functions here as a working rule of inference, not as ornament.

## 7.1 Implications if H1 holds

The chapter thesis for this section is direct: if the latency coefficient is statistically distinguishable from zero and carries the predicted sign across the fixed-effects and instrumented specifications and survives the robustness battery, then the dissertation will have shown that a measurable share of NASA cost growth and schedule slip is a controllable process variable rather than an irreducible feature of spaceflight, and the practical consequence is that compressing authorization latency becomes a lever on program performance that is independent of the two levers the existing literature emphasizes, technology investment and contract structure.

Under H1, a portion of NASA cost and schedule performance is attributable to authorization latency, a variable internal to the agency, and the case for that attribution is set out below so that no major step travels without its justification. The evidence is the estimated coefficients: a positive latency coefficient for cost growth and for schedule slip, recovered net of program and era fixed effects, persisting under the instrumental-variable strategy and across both the coarse and fine latency resolutions. What licenses moving from those coefficients to the conclusion is the identification argument of Chapter 5: program fixed effects absorb time-invariant program difficulty, era fixed effects absorb common rule regimes and appropriations conditions, and the instruments isolate variation in latency that is plausibly unrelated to the phase-specific technical difficulty, so the surviving association is interpretable as the effect of process time rather than of difficulty or era. The precedent in Decarolis and colleagues supports that step, since they used contract-level United States federal procurement data and an instrumental-variable strategy to show that bureaucratic competence causally reduces delays and cost overruns, establishing that the administrative side of program execution can move cost and schedule outcomes and that an instrument-based identification of that effect is feasible in government data [\[12\]](#ref-12). One condition is essential and must be protected: the conclusion holds only to the extent that the instruments are valid, which is argued and tested rather than assumed, and the magnitude is reported as a bounded range conditioned on the stated assumptions, never as a single causal point, in keeping with Fogel's discipline [\[17\]](#ref-17). The section must also keep open the possibility that even a robust positive coefficient could reflect reverse causation or baseline gaming rather than a forward effect of latency on outcomes, and Section 7.3 weighs exactly those rivals before any causal language is permitted to stand.

The mechanism behind the claim is named, not merely asserted, because the dissertation forbids bare correlation as an explanation. The driver is fragmented decision authority combined with sequential, multi-layer authorization gates paced by an annual appropriations cycle. The mechanism is that each pending action accrues elapsed time, the authorization latency, during which standing program costs continue to be incurred and during which the technical baseline drifts as requirements are revisited. The observable effect is that measured latency per program-phase rises. The operational consequence is higher cost growth, because standing costs are a function of elapsed time and requirements drift invites rework, and more schedule slip, because authorization latency is itself a component of elapsed schedule. The strategic implication is that program execution management gains a lever distinct from technical risk reduction. This is the causal chain the design is built to test, and under H1 the data would be consistent with it. The dissertation's confidence in the chain, if H1 holds, would be **moderate to high** for schedule slip, where the mechanism is most direct because authorization latency is mechanically part of elapsed schedule, and **moderate** for cost growth, where the link runs through the two intermediate steps of standing-cost accrual and requirements drift and is therefore more exposed to confounding. Confidence would rise if the instrumented and fixed-effects estimates agreed closely and if the effect were stable across the two latency resolutions; it would fall if the effect appeared only at one resolution or only in the fixed-effects specification.

The contribution to the public-administration literature under H1 deserves separate statement because it is where the dissertation most enlarges existing theory. Rainey, Pandey, and Bozeman established that the procedural rules an organization imposes are measurable and that managers in public and private settings perceive them differently, and the subsequent measurement program built validated perceptual scales for red tape [\[7\]](#ref-7). Brewer and Walker provided an empirical analysis of red tape's impact on governmental performance and found effects that are real but more nuanced than conventional wisdom assumes, a caution this dissertation has carried from the outset [\[8\]](#ref-8). The limitation common to that tradition is that its central construct is typically operationalized perceptually, through survey instruments that ask managers how burdensome they find the rules, and its outcomes are typically organizational performance ratings rather than dollars and months. Under H1, this dissertation would supply something the perceptual tradition has not had for a single agency over the long run: a behaviorally grounded, documentary measure of one specific administrative delay, denominated in months, linked to outcomes denominated in constant fiscal-year dollars and in schedule months. In North's vocabulary, latency is the internal transaction cost of measuring a proposed action against the rules and obtaining authorization to proceed, and an H1 result would estimate the size of that transaction cost in the currency of program execution [\[18\]](#ref-18). This is the theoretical pay-back the section promises: not a new construct, but a hard, long-run measurement of an old one, returned to the literatures that named it.

The mission and stakeholder implications under H1 are where the dissertation serves its NORTH STAR and JPL category, mission program execution management. The implication is that investment in faster, lower-layer authorization may return cost and schedule savings comparable to investment in technical risk reduction, and that the return is largest where authorization events accumulate over many years. The Jet Propulsion Laboratory is the clearest instance, because its programs concentrate in deep-space and planetary missions with long development cycles, and a long cycle is precisely a structure in which authorization latency compounds: every additional review gate, every appropriations cycle that pages a phase transition, adds elapsed months that under H1 translate into cost and schedule. The qualifier here is again Fogelian. The dissertation would not assert that any specific program would have cost a specific amount less under faster authorization, because that is an unrun counterfactual; it would assert a bounded, within-program and within-era estimate of the average association, conditioned on the identification assumptions, and would leave the program-specific counterfactual explicitly uncomputed [\[17\]](#ref-17). The single permitted institutional-design implication, developed at greater length in Section 7.6, is that the lever is real and is expressed in program-execution-management terms, not in the architecture-layer vocabulary that the dissertation's scope decision excludes.

A second-order implication under H1 deserves attention because it sharpens what the result would and would not authorize. A positive latency coefficient is an average association recovered net of the fixed effects, and an average is not a uniform effect. The dissertation's robustness battery, by re-estimating at both the coarse and fine latency resolutions and across alternative cadence-family definitions, would expose whether the effect is concentrated in particular eras, particular mission classes, or particular phases of the lifecycle. If, under H1, the latency effect on schedule slip were strongest in the formulation and early-development phases, where authorization gates cluster, and weaker in operations, where few authorization events occur, that heterogeneity would itself be a finding with operational content: it would localize the lever to the phases where decision-cycle compression is most actionable. The dissertation does not pre-commit to a particular heterogeneity pattern, because doing so would reintroduce the specification-search freedom the pre-registration removes, but it does commit to reporting the heterogeneity it finds, and an H1 reading is incomplete until that reporting is done. Confidence in any heterogeneity claim would be **lower** than confidence in the average effect, because subsetting the panel reduces the variation available within each cell and widens the cluster-robust intervals, and the dissertation would calibrate its language to that reduced precision rather than reading a noisy subgroup difference as a structural feature.

There is also a measurement contribution under H1 that is logically prior to the regression and survives independent of it, and this section would be incomplete without stating it, because it is the part of the dissertation that stands even if every causal qualifier is contested. To estimate the latency coefficient at all, the dissertation must first construct a constant-unit, single-documentary-rule series of NASA authorization latency spanning 1958 to 2026, a series that does not currently exist. That construction is the Maddison contribution in its purest form: a transparent, replicable measure expressed in common units so that one era can be compared with another on the same footing [\[16\]](#ref-16), [\[26\]](#ref-26). Under H1 the series is the input to a causal finding; but the series is valuable to the cost-estimating and public-administration communities whether or not the coefficient is nonzero, because it makes a previously unmeasured institutional variable visible and comparable across the agency's history. The dissertation's confidence in the measurement contribution is **high and largely independent of the regression result**, because it rests on documentary construction rather than on identification assumptions, and the only thing that would lower it is the construction uncertainty in the earliest decades, which the two-resolution design is built to bound.

## 7.2 Implications if H0 holds
The chapter thesis for this section is its mirror: if the analysis retains the null, the result is not a failure of the study but a substantive finding, namely that once program difficulty and era are accounted for, authorization latency does not move cost, schedule, or cadence, and the practical consequence is that the agency should concentrate reform on the technical and estimating factors the cost-growth literature documents rather than on process speed.

The symmetry between this section and the last is deliberate, and it is itself a methodological commitment. A test that can only be interesting in one direction is not a test; it is an advocacy exercise dressed in regression notation. The dissertation has been designed, per the prospectus, so that the null is a real possibility and not a straw figure, because latency may be endogenous to program difficulty and any raw correlation could reflect difficulty rather than process. Retaining H0 would therefore carry genuine information, and the inference must be set out carefully to keep it honest. Under H0, authorization latency is not a lever on cost, schedule, or cadence once difficulty and era are removed. The evidence would be a latency coefficient statistically indistinguishable from zero in the preferred specification, or one that reverses sign between the fixed-effects and instrumented specifications, or one that vanishes under the heterogeneity-robust estimators or the alternative latency resolution. The same identification machinery that licenses the H1 reading licenses this one, now read in the opposite direction: if the design is credible enough to interpret a nonzero coefficient as an effect, it is credible enough to interpret a zero coefficient as the absence of one. The cost-growth literature's own finding supports the plausibility of the null world, since technical and estimating parameters, instrument mass and power, technology readiness, contract type, and the relationship between schedule and cost carry substantial explanatory weight for NASA outcomes, which makes a technical-dominance world theoretically plausible rather than merely a fallback [\[1\]](#ref-1), [\[4\]](#ref-4), [\[6\]](#ref-6). A retained null is a statement about average association net of the controls, not proof that latency never matters for any program, and the section must keep open the possibility that a null could be a power problem, an artifact of the unbalanced panel and the modest number of program clusters, rather than a true absence of effect, which is why the analysis plan specifies wild-cluster bootstrap inference and reports the precision of the zero.

The reform implication under H0 is concrete and useful. The technical-estimating tradition is mature: Emmons, Bitten, and Freaner set reserve guidelines from historical growth distributions and attempted to separate growth due to external programmatic reasons from growth due to internal technical reasons [\[1\]](#ref-1); Bitten and colleagues showed that policy changes move NASA science-mission cost and schedule growth [\[4\]](#ref-4); the National Research Council's consensus study catalogued the management, technical, and budgetary causes of cost growth in Earth and space science missions [\[6\]](#ref-6). If H0 holds, these are the levers that survive, and the dissertation would direct the agency's reform energy toward them rather than toward process-speed initiatives. The mechanism distinction is the point. Under H0, the elapsed time of authorization is not the binding driver of outcomes; the binding drivers are the technical parameters at commitment and the estimating practices that set the baselines against which growth is measured. A reform program that compressed review cycles while leaving optimistic baselines and technical risk untouched would, under H0, be expected to yield little, and the dissertation would have produced the evidence to say so.

A retained null disciplines a narrative that the agency has long repeated without testing, and that is the section's strongest contribution. The practitioner claim that overruns flow from slow decision-making is widely held and rarely measured, and the vivid examples that animate it, the roughly nine billion dollars spent over about five years by Constellation with no operational flights, the long interval between the first and second crewed Space Launch System flights, are anecdotes rather than measurements, as the prospectus has been careful to mark them. A retained null would show that the narrative does not survive a properly identified test, and that is a service to the agency's reasoning even though, and indeed because, it disappoints the narrative. The confidence the dissertation could place in a null is itself bounded: a precisely estimated zero, with tight cluster-robust intervals that exclude managerially meaningful effect sizes, would warrant **moderate to high** confidence that latency is not a first-order lever; a noisily estimated zero, with intervals wide enough to contain both null and substantial effects, would warrant only **low** confidence and would be reported as an inconclusive test rather than as evidence for H0. This is the McCloskey caution operationalized: statistical insignificance is not substantive nullity unless the estimate is precise enough to exclude substantively important magnitudes, and the dissertation reports the precision so the reader can judge.

A retained null would also carry a theoretical implication for the institutional account that motivates the study, and the section must address it rather than treat H0 as the mere absence of an interpretation. North's framework predicts that authorization regimes, once established, generate increasing returns and lock in, which is the reason latency might be both large and durable; but durability of the variable is not the same as efficacy of the variable [\[18\]](#ref-18). A retained null would be consistent with a world in which authorization regimes are indeed durable and path-dependent, exactly as the institutional theory expects, yet in which the elapsed time they generate is absorbed by other program slack rather than transmitted to cost and schedule outcomes. In that world the transaction cost named by North is real but is not the binding one for execution performance; some other transaction cost, perhaps the cost of measuring and re-specifying technical requirements, dominates. The dissertation would not be able to adjudicate among such alternative transaction costs from a single null coefficient, and it would not pretend to; it would report the null as a finding about latency specifically and would flag the identification of the binding transaction cost as future work. This is the honest boundary of a null: it falsifies one mechanism without certifying another, and the dissertation states the boundary rather than filling it with speculation.

A retained null also bears on the cadence outcome in a way that differs from its bearing on cost and schedule, and the asymmetry must be carried into the reading. Cadence is the most construct-fragile of the three outcomes, dependent on program-family definition, and the analysis reports it under multiple family definitions for exactly that reason. A null on cadence is therefore harder to interpret than a null on cost or schedule, because it could reflect either a true absence of a latency-cadence link or the dilution of a real link across inconsistent family definitions. The dissertation's decision rule treats a cadence null with more caution than a cost or schedule null, requiring concordance across family definitions before reading the cadence result as informative either way. The confidence the dissertation places in any cadence inference, null or otherwise, is accordingly **lower** than in the cost and schedule inferences, and the chapter is explicit that the cadence channel is the one most likely to remain inconclusive after estimation.

## 7.3 Rival explanations weighed even if H1's pattern appears

The chapter thesis for this section is that finding the predicted association is necessary but not sufficient for the causal claim, because three rival explanations could each generate the H1 pattern without latency causing the outcomes, and the contribution stands only if each rival is confronted with a specific, pre-named check rather than waved away. This is the rival-explanation discipline that Yin places at the center of credible case-based and quasi-experimental inference: a candidate explanation earns belief only after the plausible alternatives have been articulated and tested against the same evidence [\[33\]](#ref-33).

The first rival is reverse causation. The concern is that programs already overrunning may generate more authorization events and longer latency as a consequence of trouble rather than as a cause of it: a program in difficulty triggers additional reviews, re-baselining actions, and authorization cycles, so high latency and high cost growth co-occur because the overrun produces the latency. The rival's mechanism, named precisely, is what makes it testable. The driver under the rival is emergent program trouble; the mechanism is that trouble summons additional oversight; the observable effect is elevated latency contemporaneous with elevated growth; the implication is a positive latency coefficient with the causal arrow running backward. The dissertation's responses are two, both specified in advance. First, the instrumental-variable strategy isolates variation in latency driven by the authorizing office's contemporaneous workload and by the action's timing relative to the appropriations calendar, sources of latency that are plausibly unrelated to the specific program's emerging trouble; if the instrumented coefficient retains the predicted sign and magnitude, the reverse-causation reading is weakened because the instrument does not respond to program trouble [\[12\]](#ref-12). Second, latency is measured early in each phase, at the phase's opening authorization events, so that the latency variable predates the bulk of the phase's cost and schedule realization and cannot be a downstream product of it. The confidence the design can defeat reverse causation is therefore **conditional on instrument strength**: strong first-stage relationships and a stable instrumented coefficient raise it; a weak first stage leaves the reverse-causation objection alive and would be reported as such.

The second rival is common cause. The concern is that an underlying era condition, most plausibly a period of budget instability, produces both high latency and high cost growth without latency causing the growth: when appropriations are unstable, authorizations stall and programs are simultaneously starved and re-planned, so a third factor drives both variables. The mechanism of the rival is that fiscal turbulence is a confounder common to the explanatory and outcome variables. The dissertation's responses are again two and again pre-specified. First, era fixed effects absorb common shocks and rule regimes that affect all programs in a given period, including agency-wide reorganizations and appropriations-cycle conditions, so a common-cause story operating at the era level is removed by construction. Second, a funding-instability index, built from year-over-year deviations between requested and appropriated funds, enters the specification as a control, so that variation in fiscal turbulence finer than the era band is also conditioned out. The interpretive consequence is that the surviving latency coefficient is identified from within-program and within-era variation net of measured funding instability, which is the configuration under which the common-cause rival is least able to masquerade as a latency effect.

The third rival is optimistic baselines, and it is the subtlest because it implicates the outcome measures themselves. The concern, which the megaproject literature emphasizes, is that agencies set optimistic cost and schedule baselines, so measured growth reflects the optimism of the baseline rather than anything about process [\[13\]](#ref-13), [\[30\]](#ref-30). Flyvbjerg and colleagues document that large public projects overrun systematically and locate part of the cause in optimism bias and strategic misrepresentation at the decision stage, and reference-class forecasting is the corrective they propose precisely because baselines are chronically optimistic [\[13\]](#ref-13), [\[30\]](#ref-30). If baselines are set optimistically and independently of latency, then measured growth is partly an artifact of baseline choice, and a latency coefficient could be picking up a baseline-setting pattern correlated with latency rather than a process effect. The more dangerous version of the rival is joint determination: agencies that anticipate slow authorization may set optimistic baselines for that very reason, so baseline choice and latency are co-determined and the coefficient confounds the two. The dissertation's response is the baseline-conservatism test specified in Chapter 5: the study examines whether baseline conservatism, proxied by the ratio of held reserves to baseline, correlates with latency, and if the correlation is absent the joint-determination version is weakened, while if it is present the analysis conditions on the conservatism proxy and reports how much the latency coefficient moves. The honest statement is that this rival cannot be fully eliminated by the design, only bounded, and the dissertation accordingly reports the latency coefficient both with and without the baseline-conservatism conditioning so the reader can see its sensitivity. Confidence that the optimistic-baseline rival has been defeated is therefore **moderate at best**, and the dissertation says so rather than overclaiming.

A fourth rival, less central than the three above but admitted because the case against confounding is only as strong as its enumeration is complete, is construct slippage in the latency measure itself. The concern is that the documentary operationalization of latency, elapsed months between a documented trigger and the documented authorization that resolves it, may capture something other than administrative decision time: it may capture, for example, the time programs spend gathering technical information before an authorization can responsibly be granted, in which case the latency variable is partly a measure of technical work rather than of process delay. If so, a positive coefficient would partly reflect the very engineering difficulty the design means to net out. The dissertation's responses are the two-resolution construction, which checks whether results are stable when latency is measured from milestone-to-milestone intervals versus from individual key-decision-point records, and the construct-validity sensitivity analyses specified in Chapter 5. The honest statement is that latency is a proxy for the administrative-time concept and may not capture it fully, and the dissertation reports the sensitivity of its conclusion to the operationalization rather than asserting that the proxy is the concept. Confidence that latency measures process delay rather than technical work is **moderate**, raised by stability across resolutions and lowered by any divergence between them.

A fifth rival, survivorship and cancellation selection, is less prominent in the existing literature's treatment of panel data but is structurally important for the dissertation's cost-and-schedule panel and must be addressed before any causal reading is permitted. The concern is that programs which were cancelled drop out of the panel entirely, while programs that survived to completion are, by construction, the programs that obtained authorization and secured resources, however slowly or costlily. If cancellation is informative, meaning that the unobserved outcomes of cancelled programs would have been systematically different from the observed outcomes of surviving programs, the cost-and-schedule panel is a selected sample rather than a random draw from the population of NASA programs, and the latency coefficient estimated on the surviving panel may not equal the coefficient that would be obtained on the full, including-cancelled population. The mechanism of the selection threat is specific: a program that encounters high latency early in its development may be cancelled before its full cost and schedule growth is realized, so the panel observes only the surviving programs, which are precisely those that persisted through high-latency episodes and eventually overran. This selection could bias the latency coefficient in either direction. If cancelled programs were disproportionately those in which high latency preceded early termination before overrun could accumulate, the surviving sample would understate the latency-cost association, attenuating the coefficient toward zero. If, conversely, cancellation selectively removed lower-overrun programs, the surviving sample would overstate it. The net direction of the bias is an empirical question the design must address rather than assume away.

The design's responses to this rival are two. First, the panel construction codebook, specified in Appendix A, records cancelled programs explicitly and tracks their available cost and schedule data through the phase at which cancellation occurred; any program with a documented baseline and at least one subsequent actual is included in the panel for the phases completed before cancellation, rather than excluded entirely. This partial-inclusion rule means that cancellation reduces the number of observations from a cancelled program but does not drop it from the analysis, limiting the selection to programs terminated so early that no subsequent data exist, which are programs whose contribution to the panel would in any case be marginal. Second, the analysis plan pre-specifies an inverse-probability-of-completion weighting step as part of the robustness battery: a completion propensity score is estimated from program characteristics observed at formulation, and observations are weighted by the inverse of that estimated propensity before the baseline regression is run, so that programs resembling those that were cancelled receive higher weight and the coefficient is adjusted toward what it would be in the full population. A latency coefficient that is stable across the unweighted and inverse-probability-weighted specifications provides evidence that the survivorship selection is not driving the result; a coefficient that shifts substantially between the two flags the selection as a material threat and is reported as such. Confidence that the survivorship rival can be fully neutralized is low, because propensity-score weighting corrects only for observed characteristics and not for unobserved reasons for cancellation; the dissertation therefore downgrades its causal confidence by one step when the survivorship concern applies and reports the sensitivity explicitly rather than treating the weighted estimate as a definitive resolution.

The cumulative logic of this section is that the design improves on the alternatives, which is one of the points the dissertation carries across its chapters. The predicted pattern, if observed, is consistent with H1, but it is also consistent in part with each rival, and only after each rival has been confronted with its specific check does the causal reading earn the qualified belief the dissertation is willing to assert. The rivals, reverse causation, common cause, optimistic baselines, construct slippage, and survivorship selection, are not defeated to the same degree: reverse causation and common cause are addressed by the strongest instruments the design has, the workload and appropriations-timing instruments and the era fixed effects with the funding-instability control; optimistic baselines and construct slippage are bounded rather than eliminated; and survivorship selection is partially addressed by partial inclusion of cancelled programs and by inverse-probability weighting, with a residual that is reported rather than concealed. Where a rival cannot be fully defeated, the dissertation downgrades its confidence rather than concealing the residual, which is the correlation-versus-causation discipline applied honestly at the interpretive stage. The residual risk that survives this confrontation is acceptable and managed rather than zero, and the dissertation's claim is calibrated to that residual, the final element of the case the dissertation carries through every chapter.

## 7.4 External validity

The chapter thesis for this section is that the dissertation is a single-agency long-run case study whose external validity rests on the consistency of its mechanism with broader procurement evidence rather than on statistical generalization to a population of agencies, and that the appropriate external claim is therefore bounded and mechanism-based, not extrapolative.

The findings, whichever way they fall, pertain to NASA, and the limits of what the design supports follow from that. The panel is one agency's programs over its history, and the identifying variation is within-program and within-era variation inside that single institution. There is no random sample of agencies from which NASA was drawn, and no basis for a frequentist generalization to the population of public technical agencies. To claim otherwise would violate the measurement discipline the dissertation has adopted, because the constant-unit, single-rule construction of the latency series is calibrated to NASA's documentary record and would have to be rebuilt, with its own construction uncertainty, for any other agency [\[16\]](#ref-16). The correct framing is the case-study framing: the dissertation offers analytic, not statistical, generalization, in which the basis for extending the finding beyond the case is the credibility of the mechanism rather than the representativeness of the sample [\[33\]](#ref-33).

The mechanism-based external case runs through the nearest external benchmark, Decarolis and colleagues. That study, using contract-level United States federal procurement data across agencies, finds that bureaucratic competence causally reduces delays and cost overruns, which is direct evidence that the administrative side of program execution affects cost and schedule outcomes in government generally, not only at NASA [\[12\]](#ref-12). The interpretive move the dissertation makes is precise and limited. If this study's NASA-specific finding under H1 agrees in sign with the Decarolis federal-procurement result, the agreement strengthens the case that the underlying mechanism, administrative process consuming time and money, is general, while leaving the magnitude NASA-specific. The Decarolis result is used as corroboration of the mechanism, not as a basis for transferring NASA's estimated coefficients to other settings. The transaction-cost theory supplies the reason the mechanism should be expected to recur: North's account locates organizational performance in the transaction costs institutions raise or lower, and there is no feature of that account special to NASA, so the mechanism's portability is theoretically grounded even where its magnitude is not [\[18\]](#ref-18). This is the logic by which the problem is general while the estimate is local.

The boundary conditions on external validity must be stated because they are where naive transfer would go wrong. NASA's authorization structure, with its key decision points, program commitment reviews, standing review boards, and joint-confidence-level policy, paced by a federal annual appropriations cycle, is the structure in which the latency variable is defined. Agencies or commercial programs with materially different authorization structures, fewer review layers, different funding mechanisms, or different oversight regimes, would generate latency from a different process, and the same documentary rule would not apply unchanged. Commercial programs in particular operate under authorization structures so different that the dissertation explicitly declines to extend its claims to them and frames its results as NASA-specific. The honest external statement is therefore narrow and defensible: the dissertation tests a mechanism in one well-documented agency, reports whether that mechanism operates there, and points to the procurement-competence literature as evidence that the mechanism is not unique to the agency, while disclaiming any transfer of magnitude. Confidence in the external claim is **moderate for the mechanism and low for the magnitude**, and the dissertation calibrates its language accordingly.

## 7.5 What would falsify the contribution

The chapter thesis for this section is that the contribution is falsifiable in the strict sense, that the conditions of falsification are fixed before estimation, and that holding to those conditions rather than narrating whatever the data deliver is what distinguishes this study from the anecdotes it replaces.

The pre-registered decision rule is the spine of the falsification logic, and it is stated here in the same terms it carries throughout the dissertation so that the interpretation cannot drift. H0 is rejected only if the latency coefficient is statistically distinguishable from zero at conventional levels, carries the predicted sign in the same direction across the fixed-effects and instrumented specifications, and survives the robustness battery. The sign convention is restated to keep it unambiguous: under H1 the latency coefficient is positive for cost growth and for schedule slip and negative for mission cadence. If the coefficient is statistically indistinguishable from zero in the preferred specification, the contribution's causal claim fails and H0 is retained. If the coefficient reverses sign between the fixed-effects and instrumented specifications, it fails, because a sign that depends on whether endogeneity is addressed is not a stable effect. If the coefficient vanishes under the heterogeneity-robust estimators of Callaway and Sant'Anna or de Chaisemartin and D'Haultfoeuille, or under the alternative latency resolution, it fails, because an effect that exists only under a naive estimator or only at one documentary resolution is an artifact rather than a finding. And if the apparent association is fully explained by baseline gaming or by reverse causation, as tested by the baseline-conservatism check and the instrumental-variable strategy, it fails, because the association would then belong to a rival rather than to latency.

The reason these conditions are binding rather than discretionary is the central methodological commitment of the dissertation, and it is worth making the reasoning explicit. The falsification conditions cannot be relaxed after seeing the data, because the conditions are pre-registered, time-stamped before estimation in the pre-registration record that the dissertation's backmatter preserves. Pre-registration removes the specification-search freedom that would otherwise let a researcher choose, after the fact, the specification under which the desired result appears; once the decision rule is fixed in advance, the result is whatever the fixed rule returns. The modern econometric literature documents how flexible estimator choice can manufacture significance, which is why the dissertation commits to the heterogeneity-robust estimators and the Goodman-Bacon diagnostic before estimation rather than selecting among them afterward. Pre-registration disciplines inference but does not guarantee truth; a pre-registered test can still be underpowered or confounded, which is why the falsification conditions include the precision and rival-explanation checks and not only the sign and significance. The section keeps open the possibility that an unforeseen data pathology could require a departure from the pre-registered plan, and the dissertation's response is that any such departure would be reported as a departure, with the original pre-registered result shown alongside, so the reader can see what the fixed rule would have returned.

The deliberate refusal to populate the illustrative coefficient table is the most visible expression of this falsification posture, and it belongs in the discussion because it is an interpretive choice, not a gap. The expected signs stated in Chapter 5 and recalled in Section 7.1 are directional expectations under H1, offered so the test is interpretable; they are not findings, and the table that would hold the estimated coefficients is left as placeholders by design. Reporting fabricated coefficients as if real would violate the falsifiability standard on which the dissertation is built, because it would present as evidence numbers that no estimation produced. The dissertation's honesty about its design stage is therefore not a limitation to be apologized for but the very thing that makes its eventual result, when the panel is assembled and the procedure executed, defensible.
Falsification of the contribution and failure of the dissertation are distinct outcomes, and conflating them would misread the study's structure. The causal contribution, that longer latency raises cost growth and schedule slip and lowers cadence, is falsifiable and may be falsified; if the latency coefficient is an indistinguishable or unstable zero, that contribution fails. But the dissertation's first contribution, the constant-unit long-run latency series, is not falsifiable in the same sense, because it is a measurement rather than a hypothesis, and it stands whether the coefficient is zero or nonzero. The second contribution, joining the NASA cost-and-schedule literature to the public-administration literature on administrative process in a single long-run study of one agency, likewise stands as a synthesis regardless of the coefficient's value. The falsification conditions of this section therefore bite on the causal claim alone, and a reader who saw a retained null should conclude that the causal lever was not demonstrated, not that the dissertation produced nothing. This layered structure, a falsifiable causal claim resting on a non-falsifiable measurement and a non-falsifiable synthesis, is what allows the study to be both genuinely at risk of disconfirmation and productive under either outcome.

## 7.6 The institutional-design implication

The chapter thesis for this final section is that the dissertation permits exactly one institutional-design implication, that this implication is expressed in program-execution-management terms rather than in enterprise-architecture vocabulary, and that this restraint is a deliberate scope decision rather than an oversight.

One scope decision governs what this section may and may not say. The dissertation is a cliometric econometric study whose objects are programs, phases, documentary events, and panel estimates. It does not design or deliver a capability, a system, a data or service exchange, or an enterprise architecture to be fielded. The architecture-traceability layer that would map a strategic objective through a capability, an operational activity, a system function, and a data exchange to a measure and a decision is therefore out of scope, and the dissertation neither builds that table nor borrows its vocabulary. To force DoDAF or BEA terminology onto a panel regression about authorization latency would be a category error, attaching an architecture grammar to an object that is not an architecture. The discipline here is the same one that governs the citation practice and the unpopulated table: say only what the design supports, in the vocabulary the design warrants.

Within that boundary, the single permitted implication is the management lever, and it is conditional on H1. If longer authorization latency is associated with greater cost growth and schedule slip net of difficulty and era, then faster, lower-layer authorization is a candidate management intervention, because it would compress the elapsed time that, under H1, drives the outcomes. The mechanism is the one named in Section 7.1: lower-layer authorization shortens the interval between an action becoming due and its resolution, reducing the standing-cost accrual and requirements drift that the latency mechanism converts into cost and schedule. The implication is expressed in execution-management terms, that a program might invest in decision-cycle compression as it invests in technical risk reduction, and weighed against technical-risk levers on the same cost-and-schedule footing. It is not expressed as an architecture-layer prescription, because the dissertation has no architecture to prescribe. The Jet Propulsion Laboratory is again the salient stakeholder, because its long-cycle deep-space programs accumulate the most authorization events and would, under H1, see the largest return from decision-cycle compression; but even this is stated as a conditional management implication, not as a system design.

The qualifier that closes the chapter returns to the institutional theory that opened it. North's account of adaptive efficiency and path dependence is the reason the management lever, even under H1, would be difficult to pull: once a set of review and authorization rules is in place it generates increasing returns, organizations adapt to it, and the cost of changing it rises, so an agency can lock onto a persistent, even inefficient, decision trajectory [\[18\]](#ref-18). The era fixed effects in the design exist precisely because rule regimes are durable in this way. The strategic implication, then, is doubly conditional. It is conditional on H1, because absent a latency effect there is no lever to pull, and it is conditional on the agency's capacity to overcome the path dependence that makes authorization regimes self-reinforcing. The dissertation's contribution under H1 is to measure the size of the prize; whether and how the agency could claim it is an institutional-change question the path-dependence literature warns is harder than the measurement, and the dissertation is careful not to promise the change on the strength of the measurement alone. That restraint, holding the claim to exactly what the design supports, is the posture in which the chapter, and the case for the dissertation's contribution, rests.



## Chapter 8. Conclusion

## 8.0 The chapter's answer, stated first

The central claim of this dissertation survives the worst case. Even if the panel test returns a latency coefficient that cannot be distinguished from zero, this work will have produced a thing that did not exist before it: a consistent, documentary-rule-based series of NASA decision-and-authorization latency spanning 1958 to 2026, built to a single measurement standard so that one era can be set against another on the same footing. That measurement is the load-bearing contribution, and it does not depend on the regression result. On top of it sits a pre-registered, falsifiable test of whether longer authorization latency is associated with greater cost growth, more schedule slip, and lower mission cadence, with a fixed decision rule and an explicit treatment of the counterfactual through within-program and within-era comparison. The honest posture of the closing is therefore unembarrassed: the test has not been executed on the full dataset, the expected signs are stated only so the test is interpretable, and the illustrative coefficient table is left unpopulated by design. What this chapter does is restate that threefold contribution precisely, set out what stands and what does not under each possible outcome, name the limitations without softening them, and lay down a concrete research program for executing the design and extending it.

This is the closing of a design-stage dissertation, and it mirrors the opening deliberately. Chapter 1 led with the answer; this chapter returns to the same answer so the document closes on the claim it opened with. The discipline that governs the return is the discipline of quantitative economic history as practiced by Maddison, Fogel, and North: state the proposition quantitatively, build the measures from primary records, bound the estimate rather than asserting a point, and let the data decide [\[16\]](#ref-16), [\[17\]](#ref-17), [\[18\]](#ref-18).

## 8.1 The threefold contribution, restated and weighted

The contribution of this dissertation is threefold, and the three parts are not of equal epistemic standing. They are ordered here from the part that stands regardless of the empirical result to the part that depends entirely on it.

### 8.1.1 First contribution: a consistent long-run latency series

The construction of a constant-unit, single-rule series of NASA decision-and-authorization latency across 1958 to 2026 is a standalone scientific contribution whose value does not depend on the sign or significance of the latency coefficient. No such series presently exists. The NASA cost-and-schedule literature is mature and quantitative, but it treats schedule as an outcome and, at most, as a driver of cost; it does not isolate administrative authorization time as a measured explanatory variable built to a uniform documentary rule across the agency's history [\[1\]](#ref-1), [\[6\]](#ref-6). The public-administration literature has measured the rules, clearances, and procedural delays that organizations impose, and it has related them to performance, but it has done so largely through perceptual scales and cross-sectional designs, not through a long-run documentary time series for a single agency [\[12\]](#ref-12). The latency series this dissertation specifies fills the empty cell: it operationalizes latency as the elapsed months between a documented trigger event and the documented authorization event that resolves it, takes the per-phase median across that phase's authorization events, and reports it at two resolutions, a coarse milestone-to-milestone measure for the full span and a fine key-decision-point measure for the modern subperiod.

A measurement built to a single transparent rule and expressed in constant units is the precondition of any comparison across periods or units; this is the Maddison standard, that no two eras can be compared until a replicable measurement standard exists [\[16\]](#ref-16). The latency series satisfies that standard by construction, because the same documentary rule is applied identically across all eras and because cost is deflated to constant fiscal-year dollars using the NASA New Start Inflation Index before any cross-era comparison. A series built this way is informative whether or not it correlates with anything, in the same way that the Maddison Project's long-run national accounts are valuable as a measurement even before they are used in any growth regression [\[16\]](#ref-16). The cliometric tradition treats the consistent series as the first deliverable and the regression as the second; the canonical economic-history projects are remembered for their measurements at least as much as for their estimates [\[16\]](#ref-16), [\[17\]](#ref-17). The descriptive step in the analysis plan, reporting the univariate distributions of latency, cost growth, schedule slip, and cadence by era under constant-dollar deflation, is itself a contribution and is sequenced before any estimation precisely so that the series can stand on its own.

This contribution holds with **high** confidence as a matter of design, conditional on the panel being assembled to the specification. One failure mode bounds that confidence: if the documentary record proves too sparse in the earliest decades to construct even the coarse measure for a usable share of early programs, the series would be truncated rather than full-span, and the contribution would be a modern-era latency series rather than a 1958-to-2026 one. That is a degradation of scope, not a falsification of the contribution. One might object that a latency series with no demonstrated association to outcomes is a curiosity, not a contribution. The reply is that measurement precedes inference in this tradition by design, and that a null association is itself a finding the series makes possible; a series that could only ever be reported if it confirmed a hypothesis would be the product of specification search, which the pre-registered decision rule exists to prevent.

### 8.1.2 Second contribution: joining two literatures for one agency over the long run

This dissertation joins the NASA cost-and-schedule literature to the public-administration literature on administrative process, two mature bodies of work that have not previously been combined in a long-run quantitative study of a single agency. The first literature models cost growth as a function of technical and contractual parameters, instrument mass and power, technology readiness, contract type, and relates schedule growth to cost growth [\[1\]](#ref-1), [\[6\]](#ref-6). It acknowledges administrative and programmatic causes in narrative but does not construct them as measured regressors over the full history. The second literature has built validated constructs for red tape and administrative burden and has shown, in the procurement setting most adjacent to NASA, that administrative competence causally reduces delays and cost overruns [\[12\]](#ref-12). The dissertation supplies the missing connective tissue: it imports the administrative-time construct from the second literature, operationalizes it from NASA's own documentary record, and enters it into the cost-and-schedule estimating framework of the first.

A bridge between two literatures is a contribution when each contains the half of the answer the other lacks and neither has reached across. Here the NASA literature has the outcome measures and the technical controls but not the administrative explanatory variable; the public-administration literature has the explanatory construct and the precedent that it matters but not the long-run single-agency panel. Joining them is not a mechanical merge; it requires translating a perceptual administrative-burden construct into a documentary, dated, program-phase measure, which is the methodological work this dissertation performs. The institutional reading that makes the join coherent is North's: decision-and-authorization latency is, in his vocabulary, a transaction cost internal to a public organization, the time cost of measuring a proposed action against the rules and obtaining authorization to proceed [\[18\]](#ref-18). His concepts of adaptive efficiency and path dependence explain why such latency would be both large and durable across the agency's history, and they supply the theoretical reason era fixed effects are necessary to separate rule regimes from program-specific variation [\[18\]](#ref-18). The Decarolis procurement result provides the empirical precedent that the administrative side of program execution moves cost and schedule outcomes [\[12\]](#ref-12).

This contribution holds with **high** confidence as a conceptual and design matter, because the join is performed in the framing, the variable construction, and the specification rather than being contingent on the estimate. Its empirical payoff, a dollar-and-month-denominated instantiation of the red-tape construct inside one agency, is contingent on the test (Section 8.1.3). A skeptic might say the two literatures were never joined because the join is not worth making, that administrative latency is already subsumed under the cost literature's schedule variable. The reply is that schedule slip is an outcome that bundles engineering difficulty, supply problems, funding instability, and authorization delay; isolating authorization latency as a separately measured input is exactly what neither literature has done, and the bundling is the reason the practitioner attribution to slow decisions has never been tested.

### 8.1.3 Third contribution: a pre-registered falsifiable test

This dissertation states and pre-registers a single falsifiable hypothesis with a fixed decision rule, an explicit counterfactual treated through within-program and within-era comparison, and a refusal to report invented estimates as real. The contribution is stated as competing hypotheses. H0 holds that administrative decision-and-authorization latency has no association with cost growth, schedule slip, or mission cadence after accounting for program and era fixed effects and technical controls. H1 holds that longer latency is associated with greater cost growth, more schedule slip, and lower cadence after the same accounting. The decision rule is fixed in advance: H0 is rejected only if the latency coefficient is statistically distinguishable from zero, carries the predicted sign in the same direction across the fixed-effects and instrumented specifications, and survives the robustness battery. If the coefficient is indistinguishable from zero, reverses sign between specifications, or vanishes under the heterogeneity-robust estimators or the alternative latency resolution, H0 is retained and the contribution fails.

A hypothesis is a scientific contribution only if it can fail; pre-registration of the decision rule before estimation is what makes the falsification binding rather than discretionary, and it is what prevents the specification search that would otherwise let a researcher report whichever specification confirmed the prior. The counterfactual discipline follows Fogel: the claim that a program would have cost less or flown sooner under faster authorization is an unmeasured counterfactual until the next-best alternative is named, and the feasible approximation is within-program and within-era comparison, with the central estimate reported as a bounded range conditioned on stated assumptions rather than as a single causal point [\[17\]](#ref-17). The notation and the estimator are fixed: \(\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t)\), where \(\alpha_i\) are program fixed effects, \(\delta_t\) are era fixed effects, and \(\varepsilon\) is clustered by program. Under H1, \(\beta\) is positive for cost growth and schedule slip and negative for cadence. The choice of heterogeneity-robust estimators over a naive two-way fixed-effects specification for the discrete-regime analyses, and the use of an instrumental-variable strategy in the spirit of Decarolis for the endogeneity of latency to phase-specific difficulty, are what make the test well posed rather than rigged [\[12\]](#ref-12).

This is the contribution that depends entirely on execution; it is design-stage and carries **no** empirical confidence because no estimate has been produced. The expected signs are stated for interpretability only and are not findings. One could argue that a pre-registered test that has not been run is not a contribution at all, only a promise. The reply distinguishes the design from the result: a falsifiable design with a fixed decision rule, a defended identification strategy, and a refusal to fabricate is a methodological contribution that other researchers can execute, criticize, or improve, and it is the part of this dissertation most directly reusable by the program-execution-management community it serves.
## 8.2 What stands if the hypothesis is not confirmed

The argument that runs through the dissertation is the right frame for stating what survives a null. It rests on five propositions: the problem is real, the problem is material, the design addresses the causal mechanism, the design improves on the alternatives, and the residual risk is acceptable. Four of these five stand whether the data reject or retain H0.

**The problem is real, and a null does not unmake it.** NASA programs grow in cost and slip in schedule; this is documented across the estimating literature, which set reserve guidelines explicitly by reference to historical growth distributions [\[1\]](#ref-1), [\[6\]](#ref-6). A null on latency would mean that authorization time is not the driver, not that the growth is imaginary.

**The problem is material, and a null does not shrink it.** The overruns are large and consequential. A null would redirect the search for causes toward the technical and estimating factors the cost-growth literature already emphasizes [\[1\]](#ref-1), [\[6\]](#ref-6). That redirection is itself a useful result, because it would discipline the widely shared practitioner narrative that attributes overruns to slow decision-making, showing that the narrative does not survive measurement.

**The design addresses a real mechanism, whether or not the mechanism fires.** The latency construct is operationalized from documentary records, key decision points, program commitment reviews, and the budget actions that authorize each phase, and entered net of program and era fixed effects with an instrument for authorizing-office workload and appropriations-calendar timing. A null tells the field that this carefully specified channel is empirically quiet, a stronger and more publishable statement than an untested assertion either way.

**The design improves on the naive alternative regardless of result.** The use of heterogeneity-robust staggered estimators with a Goodman-Bacon diagnostic, instead of a naive two-way fixed-effects difference-in-differences, is a methodological improvement that holds whatever the coefficient turns out to be.

Only the fifth element, the strategic implication that compressing authorization latency is a controllable lever on cost and schedule, depends on rejecting H0 in the predicted direction. That asymmetry is the design working as intended: a well-posed test is one in which either outcome is informative, and four of the five propositions are insulated from the empirical result by construction. The confidence attached to the surviving four is **high**; the confidence attached to the fifth is, at this stage, **not assessable**, because it awaits the estimate.

## 8.3 Limitations, stated without softening

Honesty about limitations is part of the contribution, not a concession against it. Six limitations bound this work, and each is paired with the design response already built in, so the reader can judge how far the response reaches.

**The test has not been run.** This is the governing limitation. Every coefficient in the dissertation is expected or illustrative; the illustrative table is left unpopulated by design, and the deliberate refusal to fill it with invented numbers is a methodological choice. No claim of a realized association is made anywhere in this document. The response is the full pre-registered procedure of Chapter 6, which can be executed once the panel is assembled.

**Baseline-definition drift.** The documentary definition of a program baseline changed over the agency's history, so the earliest observations carry more construction uncertainty than the modern ones. The two-resolution latency measure is the response: a coarse measure for the full span and a fine measure for the modern subperiod, with results reported separately so that no finding is an artifact of changing documentary density.

**Undated events and bounded latency.** Decision and authorization events are not always recorded with dates, so latency for some program-phases must be bounded rather than point-identified. This reduces precision and is consistent with, rather than a violation of, the bounded-estimate discipline; ranges are reported in the Fogel manner rather than false points [\[17\]](#ref-17).

**Optimistic baselines and baseline gaming.** Cost and schedule baselines are chosen by the agency and may be set optimistically, a selection problem the megaproject literature emphasizes. If agencies set optimistic baselines when they anticipate slow authorization, baseline choice and latency are jointly determined and measured growth reflects baseline choice rather than process. The response is the baseline-conservatism test, which examines whether the ratio of held reserves to baseline correlates with latency.

**Endogeneity to phase-specific difficulty.** Harder phases may both take longer to authorize and overrun more, producing a positive correlation that reflects difficulty rather than process. Program fixed effects remove the part of difficulty fixed across a program's life but not the part that varies by phase. The response is the instrumental-variable strategy using authorizing-office workload and appropriations-calendar timing as sources of latency variation plausibly unrelated to the technical difficulty of the specific phase, with both fixed-effects and instrumented estimates reported so the reader can see how much the conclusion depends on the instrument [\[12\]](#ref-12). The honest qualifier here is that the instruments' validity is argued and tested, not assumed, and a finding that no valid instrument can be constructed would itself bound the causal interpretation to a correlational one, with confidence downgraded accordingly.

**External validity.** The findings, when produced, will pertain to NASA and may not generalize to other agencies or to commercial programs with different authorization structures. The study frames its claims as NASA-specific and uses the Decarolis federal-procurement evidence only as corroboration that the mechanism is general, not as a basis for generalizing the magnitude [\[12\]](#ref-12). Cadence is additionally fragile because it depends on program-family definition, with no single definition uniquely correct, which is why cadence results are reported under multiple family definitions.

Two evidence gaps in the supporting corpus are acknowledged for the same reason. The Government Accountability Office's recurring NASA major-project assessment series is a named dataset that anchors the modern portion of the panel but is cited by report number and agency URL rather than by digital object identifier, because those reports do not carry one; the NASA New Start Inflation Index used for constant-dollar deflation is likewise an operational source rather than a peer-reviewed citation. Naming these explicitly is preferable to obscuring them, and both must be entered as discrete citable records before the panel is finalized.

## 8.4 The discipline that governs the work

The method this dissertation follows is the method of quantitative economic history, and stating it plainly is the right way to close, because the discipline is what guarantees the answer will be defensible whenever it arrives. Four commitments structure the work, each tied to one of the anchoring practitioners.

State the proposition quantitatively before reasoning about cause. The research question is posed as a measurable association between a constructed latency variable and three measured outcomes, not as a narrative about institutional dysfunction. Maddison's rule that proximate sources be quantified before ultimate causes are inferred governs the sequence: measure latency and its proximate correlates first, draw institutional conclusions only after [\[16\]](#ref-16).

Build the measures from primary records to a single rule. Latency, cost growth, schedule slip, and cadence are all constructed from documentary sources, NASA budget and program records, the NASA Technical Reports Server, and GAO assessments, by rules applied identically across eras, with constant-dollar deflation wherever cost is compared across time [\[16\]](#ref-16). This is the Maddison standard operationalized for one agency.

Bound the estimate rather than asserting a point. Following Fogel, the counterfactual claim that faster authorization would have lowered cost or accelerated cadence is approximated through within-program and within-era comparison and reported as a bounded range conditioned on stated assumptions, not as a single causal number [\[17\]](#ref-17). The undated-event problem and the endogeneity problem both push toward bounds rather than points, and the design accepts that honestly.

Let the data decide, under a rule fixed in advance. The pre-registered decision rule makes the falsification conditions binding. The institutional reading that explains why latency would be durable enough to matter, North's path dependence and adaptive efficiency, is what justifies the era fixed effects that make the test fair [\[18\]](#ref-18). Whether the data reject or retain the null, the procedure that produces the answer is fixed before the answer is seen, which is the only way a single-agency observational study can credibly claim to have tested rather than illustrated its hypothesis.

One point of scope closes the methodological statement. This is a cliometric econometric study; its objects are programs, phases, documentary events, and panel estimates, not a capability, system, or data exchange to be fielded. No architecture is being designed or delivered, and the dissertation therefore carries no architecture-traceability layer and no capability-to-system vocabulary. The single architecture-adjacent statement permitted is the institutional-design implication, expressed in program-execution-management terms: if the alternative holds, faster and lower-layer authorization is a management lever on cost and schedule independent of technology investment. That statement is conditional, and it is the only place the dissertation reaches toward design.

## 8.5 A concrete future-research program

The future work is not aspirational; it is a sequenced program with a defined path from the present design-stage document to an executed study and beyond. It is organized as three horizons.
### 8.5.1 Near horizon: executing the design on the full data

The immediate program is the five-step procedure of the analysis plan, executed in order. First, assemble and validate the program-phase panel from NASA budget and program records, NTRS documentation, and GAO assessments, validating constructed cost and schedule growth against published Aerospace Conference and GAO figures for overlapping programs so that the measures reproduce known values where known values exist [\[1\]](#ref-1), [\[6\]](#ref-6). Second, describe before estimating: report the univariate distributions of latency and the three outcomes by era under constant-dollar deflation, which delivers the standalone latency series as the first product. Third, estimate the fixed-effects baseline for all three outcomes with program-clustered standard errors. Fourth, address endogeneity with the instrumental-variable strategy, reporting first-stage strength and comparing instrumented and fixed-effects estimates [\[12\]](#ref-12). Fifth, run the robustness and heterogeneity battery at both latency resolutions, under alternative cadence definitions, with the heterogeneity-robust estimators for the discrete-regime analyses and the Goodman-Bacon decomposition as a diagnostic, reporting central estimates as bounded ranges in the Fogel manner [\[17\]](#ref-17).

Two corpus actions precede this horizon, and they are named as such. At least two specific GAO NASA major-project assessment reports must be entered as discrete citable records by report number and agency URL, and the NASA New Start Inflation Index source document must be added as the deflator citation. These are the highest-priority gaps, and they gate finalization of the data chapter.

The deliverable of the near horizon is the populated version of the illustrative table, with the bracketed placeholders replaced by estimated coefficients and cluster-robust intervals, accompanied by the descriptive latency series. Only at that point does the dissertation move from design-stage to executed, and only then is the strategic implication of Section 8.2 either earned or foreclosed.

### 8.5.2 Middle horizon: extending to comparator agencies

The appropriations-calendar instrument is, by construction, generalizable. The timing of an authorization action relative to the federal appropriations cycle paces authorization for budget-dependent reasons that are independent of the technical difficulty of any specific program, and that logic is not unique to NASA. The natural extension is a comparator-agency study that applies the same documentary-rule latency construction and the same identification strategy to other federal organizations that execute large technical programs under the same appropriations cycle. The Decarolis procurement result, which finds that administrative competence causally reduces delays and overruns across United States federal contracting, is the nearest external benchmark and the bridge to this horizon: agreement between a NASA-specific latency finding and the broader procurement evidence would strengthen the case that the mechanism is general while keeping the magnitude agency-specific [\[12\]](#ref-12).

This extension converts the single-agency external-validity limitation into a research design. A multi-agency panel with agency fixed effects would let the latency coefficient be estimated across organizations that share the appropriations instrument but differ in authorization architecture, the cleanest available test of whether the mechanism travels. Confidence in this horizon is presently **low to moderate**, because it depends on comparator agencies maintaining documentary records of decision events at a resolution comparable to NASA's; the evidence that would raise it is a feasibility audit of decision-event documentation in at least one comparator agency.

### 8.5.3 Far horizon: from measurement to management instrument

If the near-horizon execution rejects H0 in the predicted direction, a third research program becomes available: turning the latency series from a descriptive measurement into a forward-looking management instrument. The mechanism the dissertation names runs from fragmented decision authority and sequential, multi-layer authorization gates, through the accrual of elapsed time during which standing program costs continue and requirements drift, to higher cost growth and schedule slip and lower cadence, and finally to the strategic implication that a portion of NASA program performance is a controllable process variable [\[18\]](#ref-18). A management instrument would operationalize the front of that chain: a real-time latency monitor that tracks authorization intervals against program-family historical distributions and flags accruing delay before it converts into measured growth. This is explicitly a far horizon and explicitly conditional on a non-null result; it is named here so the line from this dissertation's measurement to a future operational use is visible, not because any such instrument is proposed now. The path dependence that North identifies is the caution that travels with this horizon: authorization regimes generate increasing returns and lock in, so any instrument that aims to compress latency must contend with the same adaptive inertia that made latency durable in the first place [\[18\]](#ref-18).

## 8.6 Closing

This dissertation set out to ask whether the time NASA takes to decide and authorize program actions is a measurable variable across the agency's history, and whether longer decision-and-authorization latency is associated with worse program performance measured as cost growth, schedule slip, and mission cadence. It answers the first question with a design that constructs the measure to a single documentary rule across 1958 to 2026, and it poses the second as a pre-registered, falsifiable test with a fixed decision rule, a defended identification strategy, and a refusal to report fabricated estimates as real. The first answer stands now; the second awaits execution on the full data.

The closing claim is the same one the document opened with, and that symmetry is intentional. A consistent long-run series of NASA authorization latency can be built, and it is worth building whether or not it ultimately predicts cost, schedule, or cadence, because the measurement is the precondition of the inference and the inference is honestly deferred until the panel is assembled. Whether the data reject or retain the null, the answer will be defensible, because the measures are built from primary records to a transparent rule, the estimate is bounded rather than asserted, and the decision rule is fixed before the answer is seen. That is the discipline of quantitative economic history, and it is the discipline this dissertation has tried to honor: state the proposition quantitatively, build the measures from primary records, bound the estimate, and let the data decide [\[16\]](#ref-16), [\[17\]](#ref-17), [\[18\]](#ref-18). The contribution is the design and the measurement; the verdict belongs to the data, when they are made to speak. There is a quiet service in measuring carefully what others have only asserted, and in leaving to those who come after a record they can trust and build upon. That service, more than any single result, is what this work hopes to render to the institution it studies and to the stewardship of the resources placed in its care.
## References

<span id="ref-1"></span>1. D. Emmons, R. Bitten, and C. W. Freaner (2007). Using Historical NASA Cost and Schedule Growth to Set Future Program and Project Reserve Guidelines. *IEEE Aerospace Conference*. doi: [10.1109/aero.2007.353027](https://doi.org/10.1109/aero.2007.353027).
<span id="ref-2"></span>2. R. Bitten, and et al. (2014). Historical mass, power, schedule, and cost growth for NASA science instruments. *IEEE Aerospace Conference*. doi: [10.1109/aero.2014.6836219](https://doi.org/10.1109/aero.2014.6836219).
<span id="ref-3"></span>3. R. Bitten, and et al. (2016). Historical mass, power, schedule, and cost growth for NASA spacecraft. *IEEE Aerospace Conference*. doi: [10.1109/aero.2016.7500553](https://doi.org/10.1109/aero.2016.7500553).
<span id="ref-4"></span>4. R. Bitten, et al. (2018). The effect of policy changes on NASA science mission cost and schedule growth. *IEEE Aerospace Conference*. doi: [10.1109/aero.2018.8396408](https://doi.org/10.1109/aero.2018.8396408).
<span id="ref-5"></span>5. W. Majerowicz, and S. A. Shinn (2016). Schedule matters: Understanding the relationship between schedule delays and costs on overruns. *IEEE Aerospace Conference*. doi: [10.1109/aero.2016.7500722](https://doi.org/10.1109/aero.2016.7500722).
<span id="ref-6"></span>6. National Research Council (2010). Controlling Cost Growth of NASA Earth and Space Science Missions. *National Academies Press*. doi: [10.17226/12946](https://doi.org/10.17226/12946).
<span id="ref-7"></span>7. H. G. Rainey, S. K. Pandey, and B. Bozeman (1995). Research Note: Public and Private Managers' Perceptions of Red Tape. *Public Administration Review*. doi: [10.2307/3110348](https://doi.org/10.2307/3110348).
<span id="ref-8"></span>8. G. A. Brewer, and R. M. Walker (2009). The Impact of Red Tape on Governmental Performance: An Empirical Analysis. *Journal of Public Administration Research and Theory*. doi: [10.1093/jopart/mun040](https://doi.org/10.1093/jopart/mun040).
<span id="ref-9"></span>9. D. Moynihan, P. Herd, and H. Harvey (2010). Administrative Exclusion: Organizations and the Hidden Costs of Welfare Claiming. *Journal of Public Administration Research and Theory*. doi: [10.1093/jopart/mup046](https://doi.org/10.1093/jopart/mup046).
<span id="ref-10"></span>10. L. DeHart-Davis (2008). Green Tape: A Theory of Effective Organizational Rules. *Journal of Public Administration Research and Theory*. doi: [10.1093/jopart/mun004](https://doi.org/10.1093/jopart/mun004).
<span id="ref-11"></span>11. E. L. Borry (2016). A New Measure of Red Tape: Introducing the Three-Item Red Tape (TIRT) Scale. *International Public Management Journal*. doi: [10.1080/10967494.2016.1143421](https://doi.org/10.1080/10967494.2016.1143421).
<span id="ref-12"></span>12. F. Decarolis, et al. (2018). Bureaucratic Competence and Procurement Outcomes. *NBER Working Paper 24201*. doi: [10.3386/w24201](https://doi.org/10.3386/w24201).
<span id="ref-13"></span>13. B. Flyvbjerg, et al. (2018). Five things you should know about cost overrun. *Transportation Research Part A*. doi: [10.1016/j.tra.2018.07.013](https://doi.org/10.1016/j.tra.2018.07.013).
<span id="ref-14"></span>14. B. Flyvbjerg (2018). Cost Overruns on Infrastructure Projects: Patterns, Causes, and Cures. *(book chapter)*. doi: [10.1515/9781553394570-010](https://doi.org/10.1515/9781553394570-010).
<span id="ref-15"></span>15. C. C. Cantarelli, and et al. (2010). Cost Overruns in Large-scale Transportation Infrastructure Projects: Explanations and Their Theoretical Embeddedness. *European Journal of Transport and Infrastructure Research*. doi: [10.18757/ejtir.2010.10.1.2864](https://doi.org/10.18757/ejtir.2010.10.1.2864).
<span id="ref-16"></span>16. J. Bolt, and J. L. van Zanden (2017). The Maddison Project: Historical GDP estimates worldwide. *Journal of World-Historical Information / Maddison Project Database*. doi: [10.5195/jwhi.2017.46](https://doi.org/10.5195/jwhi.2017.46).
<span id="ref-17"></span>17. A. Herranz-Loncán (2006). Railroad Impact in Backward Economies: Spain, 1850-1913. *The Journal of Economic History*. doi: [10.1017/s0022050706000350](https://doi.org/10.1017/s0022050706000350).
<span id="ref-18"></span>18. D. C. North (1990). Institutions, Institutional Change and Economic Performance. *Cambridge University Press*. doi: [10.1017/CBO9780511808678](https://doi.org/10.1017/CBO9780511808678).
<span id="ref-19"></span>19. H. P. Stahl (2010). Survey of cost models for space telescopes. *Optical Engineering*. doi: [10.1117/1.3430603](https://doi.org/10.1117/1.3430603).
<span id="ref-20"></span>20. (NASA cost community) (2021). Effectiveness of Firm-Fixed Price Spacecraft Contracts to Curb Cost Growth. *IEEE Aerospace Conference*. doi: [10.1109/AERO50100.2021.9438356](https://doi.org/10.1109/AERO50100.2021.9438356).
<span id="ref-21"></span>21. A. Goodman-Bacon (2018). Difference-in-Differences with Variation in Treatment Timing. *NBER Working Paper 25018*. doi: [10.3386/w25018](https://doi.org/10.3386/w25018).
<span id="ref-22"></span>22. B. Callaway, and P. H. C. Sant'Anna (2020). Difference-in-Differences with multiple time periods. *Journal of Econometrics*. doi: [10.1016/j.jeconom.2020.12.001](https://doi.org/10.1016/j.jeconom.2020.12.001).
<span id="ref-23"></span>23. C. de Chaisemartin, and X. D'Haultfoeuille (2021). Difference-in-Differences Estimators of Intertemporal Treatment Effects. *The Review of Economics and Statistics*. doi: [10.1162/rest_a_01414](https://doi.org/10.1162/rest_a_01414).
<span id="ref-24"></span>24. K. Imai, and I. S. Kim (2020). On the Use of Two-Way Fixed Effects Regression Models for Causal Inference with Panel Data. *Political Analysis*. doi: [10.1017/pan.2020.33](https://doi.org/10.1017/pan.2020.33).
<span id="ref-25"></span>25. B. P. Tucker, and H. C. Alewine (2022). The Roles of Management Control: Lessons from the Apollo Program. *Contemporary Accounting Research*. doi: [10.1111/1911-3846.12833](https://doi.org/10.1111/1911-3846.12833).
<span id="ref-26"></span>26. A. Maddison (2001). The World Economy: A Millennial Perspective. *OECD Development Centre*. doi: [10.1787/9789264189980-en](https://doi.org/10.1787/9789264189980-en).
<span id="ref-27"></span>27. O. E. Williamson (1996). The Mechanisms of Governance. *Oxford University Press*. doi: [10.1093/oso/9780195078244.001.0001](https://doi.org/10.1093/oso/9780195078244.001.0001).
<span id="ref-28"></span>28. O. E. Williamson (1981). The Economics of Organization: The Transaction Cost Approach. *American Journal of Sociology*. doi: [10.1086/227496](https://doi.org/10.1086/227496).
<span id="ref-29"></span>29. A. Goodman-Bacon (2021). Difference-in-differences with variation in treatment timing (journal version). *Journal of Econometrics*. doi: [10.1016/j.jeconom.2021.03.014](https://doi.org/10.1016/j.jeconom.2021.03.014).
<span id="ref-30"></span>30. B. Flyvbjerg (2008). Curbing Optimism Bias and Strategic Misrepresentation in Planning: Reference Class Forecasting in Practice. *European Planning Studies*. doi: [10.1080/09654310701747936](https://doi.org/10.1080/09654310701747936).
<span id="ref-31"></span>31. E. Ostrom (1990). Governing the Commons: The Evolution of Institutions for Collective Action. *Cambridge University Press*. doi: [10.1017/CBO9780511807763](https://doi.org/10.1017/CBO9780511807763).
<span id="ref-32"></span>32. M. Weinzierl (2018). Space, the Final Economic Frontier. *Journal of Economic Perspectives*. doi: [10.1257/jep.32.2.173](https://doi.org/10.1257/jep.32.2.173).
<span id="ref-33"></span>33. R. K. Yin (2010). Rival Explanations (Encyclopedia of Case Study Research). *SAGE Encyclopedia of Case Study Research*. doi: [10.4135/9781412957397.n306](https://doi.org/10.4135/9781412957397.n306).
<span id="ref-34"></span>34. Kevin T. Rich, et al. (2025). Houston...We Have a Cost Problem: A Financial Analysis of NASA Contractors. *54th International Conference on Environmental Systems*. doi: [10.32865/2346/102676](https://doi.org/10.32865/2346/102676).
<span id="ref-35"></span>35. Boyd Malali Makuyu, and Lynn Kazembe (2026). Examining of Project Cost Overrun in Project Management: A Case Study of the Road Development Agency in Lusaka. *International Journal of Advanced Multidisciplinary Research and Studies*. doi: [10.62225/2583049x.2026.6.2.6033](https://doi.org/10.62225/2583049x.2026.6.2.6033).
<span id="ref-36"></span>36. Dominic D. Ahiaga-Dagbui, and Simon Smith (2014). Dealing with construction cost overruns using data mining. *Construction Management and Economics*. doi: [10.1080/01446193.2014.933854](https://doi.org/10.1080/01446193.2014.933854).
<span id="ref-37"></span>37. Eduardo Monterisi Almeida de Carvalho (2026). FORECASTING AND FINANCIAL PREDICTABILITY IN COMPLEX INFRASTRUCTURE AND ENERGY PROJECTS. *Revista de Estudos Interdisciplinares*. doi: [10.23900/artefactum.v25i2.2931](https://doi.org/10.23900/artefactum.v25i2.2931).
<span id="ref-38"></span>38. David Ackah, and Kwasi Opoku Boadu (2025). The Impact of Strategic Procurement on Cost Overruns and Delays in Petroleum Exploration and Production Projects. *African Journal of Procurement, Logistics & Supply Chain Management*. doi: [10.4314/ajplscm.v8i5.3](https://doi.org/10.4314/ajplscm.v8i5.3).
<span id="ref-39"></span>39. David Ackah, and Kwasi Opoku Boadu (2025). The Theoretical Foundation of Strategic Procurement: Mitigating Cost Overruns and Delays in Petroleum Exploration and Production Projects. *African Journal of Procurement, Logistics & Supply Chain Management*. doi: [10.4314/ajplscm.v8i5.4](https://doi.org/10.4314/ajplscm.v8i5.4).
<span id="ref-40"></span>40. Muhammad Saiful Islam, Madhav Nepal, and Martin Skitmore (2018). Modified Fuzzy Group Decision-Making Approach to Cost Overrun Risk Assessment of Power Plant Projects. *Journal of Construction Engineering and Management*. doi: [10.1061/(asce)co.1943-7862.0001593](https://doi.org/10.1061/(asce)co.1943-7862.0001593).
<span id="ref-41"></span>41. Hui-min Liu, et al. (2018). Optimism Bias Evaluation and Decision-Making Risk Forecast on Bridge Project Cost Based on Reference Class Forecasting: Evidence from China. *Sustainability*. doi: [10.3390/su10113981](https://doi.org/10.3390/su10113981).
<span id="ref-42"></span>42. Bryan Barley, Allen S. Bacskay, and Marilyn Newhouse (2010). Heritage and Advanced Technology Systems Engineering Lessons Learned from NASA Deep Space Missions. doi: [10.2514/6.2010-8622](https://doi.org/10.2514/6.2010-8622).
<span id="ref-43"></span>43. Maan A. Shafaay, et al. (2025). Modeling construction cost overrun risks at the FEED stage for mining projects using PLS-SEM. *Journal of Asian Architecture and Building Engineering*. doi: [10.1080/13467581.2025.2481242](https://doi.org/10.1080/13467581.2025.2481242).
<span id="ref-44"></span>44. Mehmet Sahinoglu, and Julia Petty (2023). Quantitative Risk Assessment and Management of National Defense Acquisition with a Game-Theoretic Security Risk Meter Tool. *International Journal of Computer Theory and Engineering*. doi: [10.7763/ijcte.2023.v15.1344](https://doi.org/10.7763/ijcte.2023.v15.1344).
<span id="ref-45"></span>45. David Ackah, and Joseph Sekyi-Ansah (2026). The Effect of Strategic Procurement Management on Cost Overruns and Schedule Delays in Petroleum Engineering Projects in Ghana. *Project Management and Scientific Journal*. doi: [10.4314/pmsj.v9i1.3](https://doi.org/10.4314/pmsj.v9i1.3).
<span id="ref-46"></span>46. David Ackah, and Kwasi Opoku Boadu (2025). Optimising Procurement Strategies for Cost Efficiency and Risk Mitigation in Petroleum Engineering Projects: The Study of Offshore Drilling Operations. *African Journal of Procurement, Logistics & Supply Chain Management*. doi: [10.4314/ajplscm.v8i5.2](https://doi.org/10.4314/ajplscm.v8i5.2).
<span id="ref-47"></span>47. Aaron Chadee, et al. (2023). Reducing Cost Overrun in Public Housing Projects: A Simplified Reference Class Forecast for Small Island Developing States. *Buildings*. doi: [10.3390/buildings13040998](https://doi.org/10.3390/buildings13040998).
<span id="ref-48"></span>48. Hani Al Sadi, and Mahmoud Dawood (2021). Oil and Gas Projects in Sultanate of Oman: Construction Schedule and Cost Overrun. *Journal of student-scientists' research*. doi: [10.47611/jsr.v10i3.1286](https://doi.org/10.47611/jsr.v10i3.1286).
<span id="ref-49"></span>49. Francis T. Hoban, William M. Lawbaugh, and Edward J. Hoffman (1994). Readings in program control. *NASA STI Repository (National Aeronautics and Space Administration)*. [https://openalex.org/W778002722](https://openalex.org/W778002722).
<span id="ref-50"></span>50. D. Debnath (2025). From Delay to Delivery: Tackling Time and Cost Risks in Hydropower Projects. *International Journal for Research in Applied Science and Engineering Technology*. doi: [10.22214/ijraset.2025.76218](https://doi.org/10.22214/ijraset.2025.76218).
<span id="ref-51"></span>51. A. Saif (2024). Real-Time Delivery Control for ICT Programs: A Telemetry-Driven Framework for Cost, Schedule, and Risk Optimization in Enterprise Project Execution. *Journal of Information Systems Engineering & Management*. doi: [10.52783/jisem.v9i4s.14949](https://doi.org/10.52783/jisem.v9i4s.14949).
<span id="ref-52"></span>52. C. Brown, Hanni Lux, and James R. Cowan (2024). Reference Class Forecasting and Its Application to Fusion Power Plant Cost Estimates. *IEEE Transactions on Plasma Science*. doi: [10.1109/tps.2024.3405631](https://doi.org/10.1109/tps.2024.3405631).
<span id="ref-53"></span>53. Jörg L. Spenkuch, Edoardo Teso, and Guo Xu (2023). Ideology and Performance in Public Organizations. *Econometrica*. doi: [10.3982/ecta20355](https://doi.org/10.3982/ecta20355).
<span id="ref-54"></span>54. Rebekka Baerenbold (2023). Reducing risks in megaprojects: The potential of reference class forecasting. *Project Leadership and Society*. doi: [10.1016/j.plas.2023.100103](https://doi.org/10.1016/j.plas.2023.100103).
<span id="ref-55"></span>55. Dan Lovallo, Matteo Cristofaro, and Bent Flyvbjerg (2023). Governing Large Projects: A Three-Stage Process to Get It Right. *Academy of Management Perspectives*. doi: [10.5465/amp.2021.0129](https://doi.org/10.5465/amp.2021.0129).
<span id="ref-56"></span>56. Aaron Anil Chadee, Salisha R. Hernandez, and Héctor Martín (2021). The Influence of Optimism Bias on Time and Cost on Construction Projects. *Emerging Science Journal*. doi: [10.28991/esj-2021-01287](https://doi.org/10.28991/esj-2021-01287).
<span id="ref-57"></span>57. Jelena M. Živković, et al. (2019). THE COST PERFORMANCE AND CAUSES OF OVERRUNS IN INFRASTRUCTURE DEVELOPMENT PROJECTS IN ASIA. *Journal of Civil Engineering and Management*. doi: [10.3846/jcem.2019.8646](https://doi.org/10.3846/jcem.2019.8646).
<span id="ref-58"></span>58. Ali Shehadeh, and M. Abuaddous (2025). XAI-enabled probabilistic pipeline for predicting delay and cost overrun risk in construction. *Engineering Construction and Architectural Management*. doi: [10.1108/ecam-06-2025-1024](https://doi.org/10.1108/ecam-06-2025-1024).
<span id="ref-59"></span>59. Atif Ansar, and Bent Flyvbjerg (2022). How to solve big problems: bespoke versus platform strategies. *Oxford Review of Economic Policy*. doi: [10.1093/oxrep/grac009](https://doi.org/10.1093/oxrep/grac009).
<span id="ref-60"></span>60. Solomon Melaku Belay, et al. (2021). Analysis of Cost Overrun and Schedule Delays of Infrastructure Projects in Low Income Economies: Case Studies in Ethiopia. *Advance in Civil Engineering*. doi: [10.1155/2021/4991204](https://doi.org/10.1155/2021/4991204).
<span id="ref-61"></span>61. P. E. Love, et al. (2021). Risk and Uncertainty in the Cost Contingency of Transport Projects: Accommodating Bias or Heuristics, or Both?. *IEEE transactions on engineering management*. doi: [10.1109/tem.2021.3119064](https://doi.org/10.1109/tem.2021.3119064).
<span id="ref-62"></span>62. Zane Scott, and Antony J. Williams (2020). Applying MBSE 2.0 with Intent Aforethought. *Day 2 Tue, May 05, 2020*. doi: [10.4043/30507-ms](https://doi.org/10.4043/30507-ms).
<span id="ref-63"></span>63. R. Susanti (2020). Cost overrun and time delay of construction project in Indonesia. *Journal of Physics: Conference Series*. doi: [10.1088/1742-6596/1444/1/012050](https://doi.org/10.1088/1742-6596/1444/1/012050).
<span id="ref-64"></span>64. Bent Flyvbjerg, Chi-keung Hon, and Wing Huen Fok (2016). Reference class forecasting for Hong Kong's major roadworks projects. *Proceedings of the Institution of Civil Engineers - Civil Engineering*. doi: [10.1680/jcien.15.00075](https://doi.org/10.1680/jcien.15.00075).
<span id="ref-65"></span>65. Anon. (n.d.). Planning and Estimation of Operations Support Requirements. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20110006923](https://ntrs.nasa.gov/citations/20110006923).
<span id="ref-66"></span>66. Guangyi Wang, Rita Hamad, and Justin S. White (2024). Advances in Difference-in-differences Methods for Policy Evaluation Research. *Epidemiology*. doi: [10.1097/ede.0000000000001755](https://doi.org/10.1097/ede.0000000000001755).
<span id="ref-67"></span>67. Jung Eun Park (2021). Curbing cost overruns in infrastructure investment. *European journal of transport and infrastructure research*. doi: [10.18757/ejtir.2021.21.2.5504](https://doi.org/10.18757/ejtir.2021.21.2.5504).
<span id="ref-68"></span>68. B. Flyvbjerg (2018). Planning Fallacy or Hiding Hand: Which Is the Better Explanation?. doi: [10.1016/j.worlddev.2017.10.002](https://doi.org/10.1016/j.worlddev.2017.10.002).
<span id="ref-69"></span>69. Munawir, R. Arifuddin, and M. A. Abdurrahman (2026). Analysis of the Delay Risk of the Engineering, Procurement, Construction, and Commissioning (EPCC) Phases in Oil and Gas Industry Projects. *Engineering, Technology & Applied Science Research*. doi: [10.48084/etasr.13969](https://doi.org/10.48084/etasr.13969).
<span id="ref-70"></span>70. Norenna S. Sarahadil (2025). The concept of red tape and efficiency among corporate and government manager: Analysis on the effects to their management performances. *Environment and Social Psychology*. doi: [10.59429/esp.v9i11.3168](https://doi.org/10.59429/esp.v9i11.3168).
<span id="ref-71"></span>71. Yizi Chen, et al. (2023). Toward a Deeper Understanding of Optimism Bias and Transport Project Cost Overrun. *Project Management Journal*. doi: [10.1177/87569728231180268](https://doi.org/10.1177/87569728231180268).
<span id="ref-72"></span>72. David Christensen (2023). Cost Overrun Optimism: Fact of Fiction?. *Defense Acquisition Research Journal*. doi: [10.22594/dau.23-910.30.03](https://doi.org/10.22594/dau.23-910.30.03).
<span id="ref-73"></span>73. Dharmesh Oza (2023). Quality-Induced Impacts on Time Overrun, Cost Overrun, Dispute, and Litigation in EPC Contracts. *International Journal for Research in Applied Science and Engineering Technology*. doi: [10.22214/ijraset.2023.56126](https://doi.org/10.22214/ijraset.2023.56126).
<span id="ref-74"></span>74. Saket Patni, and Venkata Krishna Kumar Sadhu (2023). Investigating Cost Overrun Factors in High-Rise Housing Projects in India. *The Asian Review of Civil Engineering*. doi: [10.51983/tarce-2023.12.1.3666](https://doi.org/10.51983/tarce-2023.12.1.3666).
<span id="ref-75"></span>75. Ananth Natarajan (2022). Reference Class Forecasting and Machine Learning for Improved Offshore Oil and Gas Megaproject Planning: Methods and Application. *Project Management Journal*. doi: [10.1177/87569728211045889](https://doi.org/10.1177/87569728211045889).
<span id="ref-76"></span>76. Tim Neerup Themsen (2019). The processes of public megaproject cost estimation: The inaccuracy of reference class forecasting. *Financial Accountability & Management*. doi: [10.1111/faam.12210](https://doi.org/10.1111/faam.12210).
<span id="ref-77"></span>77. Debra Emmons, et al. (2010). Affordability assessments to support strategic planning and decisions at NASA. doi: [10.1109/aero.2010.5446893](https://doi.org/10.1109/aero.2010.5446893).
<span id="ref-78"></span>78. Ram Singh (2009). Delays and Cost Overruns in Infrastructure Projects -- An Enquiry into Extents, Causes and Remedies. *RePEc: Research Papers in Economics*. [https://openalex.org/W1513149336](https://openalex.org/W1513149336).
<span id="ref-79"></span>79. Bent Flyvbjerg (2007). Megaproject policy and planning: Problems, causes, cures. *VBN Forskningsportal (Aalborg Universitet)*. [https://openalex.org/W1832657426](https://openalex.org/W1832657426).
<span id="ref-80"></span>80. Anon. (n.d.). NASA Standing Review Board Handbook. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20230001306](https://ntrs.nasa.gov/citations/20230001306).
<span id="ref-81"></span>81. Timothy Landucci, and Frank W. Ciarallo (2026). Operationalizing Transaction Cost Economics for Contract Governance: Evidence from Procurement Data. doi: [10.2139/ssrn.6776392](https://doi.org/10.2139/ssrn.6776392).
<span id="ref-82"></span>82. Baraka Israel (2023). A study of stakeholders' procurement deficiencies, delays and cost overruns in Tanzania's construction projects. *International Journal of Procurement Management*. doi: [10.1504/ijpm.2023.134629](https://doi.org/10.1504/ijpm.2023.134629).
<span id="ref-83"></span>83. Ashem E. Egila, O. Balogun, and S. Yusuf (2020). Assessment of delay and cost-overrun in federal road construction project in Abuja. doi: [10.14807/ijmp.v11i4.1065](https://doi.org/10.14807/ijmp.v11i4.1065).
<span id="ref-84"></span>84. Oliver E. Williamson (2008). Outsourcing: Transaction Cost Economics and Supply Chain Management. *Journal of Supply Chain Management*. doi: [10.1111/j.1745-493x.2008.00051.x](https://doi.org/10.1111/j.1745-493x.2008.00051.x).
<span id="ref-85"></span>85. Rachel Sholder, and Sally Whitley (2023). Math is EZIE: How Contracts Help Control Cost. doi: [10.1109/aero55745.2023.10115565](https://doi.org/10.1109/aero55745.2023.10115565).
<span id="ref-86"></span>86. Anon. (n.d.). Joint Confidence Level Requirement: Policy and Issues. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20110013438](https://ntrs.nasa.gov/citations/20110013438).
<span id="ref-87"></span>87. Takunda Gumbu (2024). To collaborate or not to collaborate: A transaction cost economics approach to construction contracts in public-private partnerships. *Journal of Sustainable Development Law and Policy*. doi: [10.4314/jsdlp.v15i1.6](https://doi.org/10.4314/jsdlp.v15i1.6).
<span id="ref-88"></span>88. Arvin Karl Demillo Capiral (2023). The Dynamics of Congressional Committees in Budget Legislation and Its Impact to Philippine Economic Development. *Journal of Contemporary Sociological Issues*. doi: [10.19184/csi.v3i1.31362](https://doi.org/10.19184/csi.v3i1.31362).
<span id="ref-89"></span>89. Muhammad Waheed (2023). Elements of Cost and Schedule Overrun in Construction Projects. *Indonesian Journal of Engineering Research*. doi: [10.11594/ijer.v4i2.46](https://doi.org/10.11594/ijer.v4i2.46).
<span id="ref-90"></span>90. J. Crooke, M. Bolcar, and J. Hylan (2019). Evolving management strategies to improve NASA flagship's cost and schedule performance: LUVOIR as a case study. *Optical Engineering + Applications*. doi: [10.1117/12.2529294](https://doi.org/10.1117/12.2529294).
<span id="ref-91"></span>91. Douglass C. North (2016). Institutions and Economic Theory. *The American Economist*. doi: [10.1177/0569434516630194](https://doi.org/10.1177/0569434516630194).
<span id="ref-92"></span>92. Ram Singh (2010). Delays and Cost Overruns in Infrastructure Projects: Extent, Causes and Remedies. [https://openalex.org/W2340068089](https://openalex.org/W2340068089).
<span id="ref-93"></span>93. Charles N. Halaby (2004). Panel Models in Sociological Research: Theory into Practice. *Annual Review of Sociology*. doi: [10.1146/annurev.soc.30.012703.110629](https://doi.org/10.1146/annurev.soc.30.012703.110629).
<span id="ref-94"></span>94. Anon. (n.d.). The NPG 7120.5A Electronic Review Process. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20000028283](https://ntrs.nasa.gov/citations/20000028283).
<span id="ref-95"></span>95. Anon. (n.d.). Reducing NPR 7120.5D to Practice: Preparing for a Life-Cycle Review. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20100032895](https://ntrs.nasa.gov/citations/20100032895).
<span id="ref-96"></span>96. Anon. (n.d.). Composite Crew Module: Primary Structure. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20110020665](https://ntrs.nasa.gov/citations/20110020665).
<span id="ref-97"></span>97. Z. Neal, et al. (2018). Making or buying evidence: Using transaction cost economics to understand decision-making in public school districts. *Evidence & Policy: A Journal of Research, Debate and Practice*. doi: [10.1332/174426416x14778277473701](https://doi.org/10.1332/174426416x14778277473701).
<span id="ref-98"></span>98. S. Shinn, L. Wolfarth, and Meagan Hahn (2010). Estimating incremental cost and schedule growth for systems engineering and project management. *IEEE Aerospace Conference*. doi: [10.1109/aero.2010.5446866](https://doi.org/10.1109/aero.2010.5446866).
<span id="ref-99"></span>99. Roberto Esposti (2026). Evaluating policy impact under sparse and staggered adoption. A synthetic difference-in-differences application to EU rural development measures. *Evaluation and Program Planning*. doi: [10.1016/j.evalprogplan.2026.102751](https://doi.org/10.1016/j.evalprogplan.2026.102751).
<span id="ref-100"></span>100. Brantly Callaway, Andrew Goodman-Bacon, and Pedro H. C. Sant'Anna (2021). Difference-in-Differences with a Continuous Treatment. *arXiv (Cornell University)*. doi: [10.48550/arxiv.2107.02637](https://doi.org/10.48550/arxiv.2107.02637).
<span id="ref-101"></span>101. Christina F. Rusnock (2012). Predicting Cost and Schedule Growth for Military and Civil Space Systems. *Defense Technical Information Center (DTIC)*. [https://openalex.org/W223598845](https://openalex.org/W223598845).
<span id="ref-102"></span>102. Anon. (n.d.). Constellation Program Life-cycle Cost Analysis Model (LCAM). *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20080031533](https://ntrs.nasa.gov/citations/20080031533).
<span id="ref-103"></span>103. Yasin Kalafatoglu (2026). Deterministic Decision Authority: A Governance-First Architecture for Human-Final AI Systems. doi: [10.2139/ssrn.6193898](https://doi.org/10.2139/ssrn.6193898).
<span id="ref-104"></span>104. Xuliang Wang (2025). Modern Staggered Difference-in-Differences: From the Pitfalls of Two-Way Fixed Effects (TWFE) to Robust Estimation. doi: [10.2139/ssrn.5456874](https://doi.org/10.2139/ssrn.5456874).
<span id="ref-105"></span>105. Saba Pourreza, R. Scott, and Brian Sauser (2024). Lifecycle Cost Affordability and Performance-Based Contracting – A Managerial Decision Framework Based on Literature Review. *Operations and Supply Chain Management An International Journal*. doi: [10.31387/oscm0560420](https://doi.org/10.31387/oscm0560420).
<span id="ref-106"></span>106. Anon. (2024). Implementing an Objectives-Driven, Risk-Informed, and Case-Assured Approach to Safety and Mission Success at NASA. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20240006733](https://ntrs.nasa.gov/citations/20240006733).
<span id="ref-107"></span>107. Rong Ran, et al. (2024). Red tape reform, transaction costs, and corporate social performance: A natural quasi-experiment in China. *International Public Management Journal*. doi: [10.1080/10967494.2024.2332702](https://doi.org/10.1080/10967494.2024.2332702).
<span id="ref-108"></span>108. Ianire Taboada, et al. (2023). Artificial Intelligence Enabled Project Management: A Systematic Literature Review. *Applied Sciences*. doi: [10.3390/app13085014](https://doi.org/10.3390/app13085014).
<span id="ref-109"></span>109. Clément de Chaisemartin, and Xavier d'Haultfoeuille (2023). Two-Way Fixed Effects and Difference-in-Differences Estimators with Heterogeneous Treatment Effects and Imperfect Parallel Trends. *SSRN Electronic Journal*. doi: [10.2139/ssrn.4487202](https://doi.org/10.2139/ssrn.4487202).
<span id="ref-110"></span>110. Zachary Porreca (2022). Synthetic Difference-In-Differences Estimation With Staggered Treatment Timing. *SSRN Electronic Journal*. doi: [10.2139/ssrn.4015931](https://doi.org/10.2139/ssrn.4015931).
<span id="ref-111"></span>111. Nikolai Miklin, et al. (2022). Causal inference with imperfect instrumental variables. *Journal of Causal Inference*. doi: [10.1515/jci-2021-0065](https://doi.org/10.1515/jci-2021-0065).
<span id="ref-112"></span>112. Francis T. Hoban (2019). Issues in NASA program and project management. *NASA STI Repository (National Aeronautics and Space Administration)*. [https://openalex.org/W1595991887](https://openalex.org/W1595991887).
<span id="ref-113"></span>113. Fatih Canıtez, and Dilay Çelebi (2018). Transaction cost economics of procurement models in public transport: An institutional perspective. *Research on Transport Economics*. doi: [10.1016/j.retrec.2018.03.002](https://doi.org/10.1016/j.retrec.2018.03.002).
<span id="ref-114"></span>114. Edward Ralph Carroll, and Robert Malins (2016). Systematic Literature Review: How is Model-Based Systems Engineering Justified?. doi: [10.2172/1561164](https://doi.org/10.2172/1561164).
<span id="ref-115"></span>115. Morgan Dwyer, et al. (2014). The Cost Impacts of Jointness: Insights From the NPOESS Program. doi: [10.21236/ada624606](https://doi.org/10.21236/ada624606).
<span id="ref-116"></span>116. Andrew Bell, and Kelvyn Jones (2014). Explaining Fixed Effects: Random Effects Modeling of Time-Series Cross-Sectional and Panel Data. *Political Science Research and Methods*. doi: [10.1017/psrm.2014.7](https://doi.org/10.1017/psrm.2014.7).
<span id="ref-117"></span>117. Henning Best, and Christof Wolf (2014). The SAGE Handbook of Regression Analysis and Causal Inference. doi: [10.4135/9781446288146](https://doi.org/10.4135/9781446288146).
<span id="ref-118"></span>118. Stephen Howard Chadwick (2007). Defense Acquisition: Overview, Issues, and Options for Congress. *University of North Texas Digital Library (University of North Texas)*. [https://openalex.org/W1529725485](https://openalex.org/W1529725485).
<span id="ref-119"></span>119. Laura Poppo, and Todd Zenger (1998). Testing alternative theories of the firm: transaction cost, knowledge-based, and measurement explanations for make-or-buy decisions in information services. *Strategic Management Journal*. doi: [10.1002/(sici)1097-0266(199809)19:9<853::aid-smj977>3.0.co;2-b](https://doi.org/10.1002/(sici)1097-0266(199809)19:9<853::aid-smj977>3.0.co;2-b).
<span id="ref-120"></span>120. John F. Muratore, et al. (1990). Real-time data acquisition at mission control. *Communications of the ACM*. doi: [10.1145/96267.96277](https://doi.org/10.1145/96267.96277).
<span id="ref-121"></span>121. Oliver E. Williamson (1987). The economic institutions of capitalism firms, markets, relational contracting. [https://openalex.org/W3122093892](https://openalex.org/W3122093892).
<span id="ref-122"></span>122. Anon. (n.d.). Cost Estimation and Control for Flight Systems. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20020050339](https://ntrs.nasa.gov/citations/20020050339).
<span id="ref-123"></span>123. Anon. (n.d.). NASA Human Spaceflight Scenarios - Do All Our Models Still Say No?. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20170008892](https://ntrs.nasa.gov/citations/20170008892).
<span id="ref-124"></span>124. Anon. (n.d.). Process-Based Cost Estimation for Ramjet/Scramjet Engines. *NASA Technical Reports Server*. [https://ntrs.nasa.gov/citations/20030112778](https://ntrs.nasa.gov/citations/20030112778).
<span id="ref-125"></span>125. Jangsu Yoon (2024). Inference on Nonparametric Panel Data Models with Fixed Effects and Censored Dependent Variables. doi: [10.2139/ssrn.4953073](https://doi.org/10.2139/ssrn.4953073).
<span id="ref-126"></span>126. Vikram Dayal, and Anand Murugesan (2023). Panel Data and Fixed Effects. *Demystifying Causal Inference*. doi: [10.1007/978-981-99-3905-3_10](https://doi.org/10.1007/978-981-99-3905-3_10).
<span id="ref-127"></span>127. David Carl, et al. (2022). TSCI: Tools for Causal Inference with Possibly Invalid Instrumental Variables. *CRAN: Contributed Packages*. doi: [10.32614/cran.package.tsci](https://doi.org/10.32614/cran.package.tsci).
<span id="ref-128"></span>128. Andrew Bell, Malcolm Fairbrother, and Kelvyn Jones (2018). Fixed and random effects models: making an informed choice. *Quality & Quantity*. doi: [10.1007/s11135-018-0802-x](https://doi.org/10.1007/s11135-018-0802-x).
<span id="ref-129"></span>129. Anon. (2017). The Budget and Appropriations Cycle. *Working the Federal Budget*. doi: [10.4324/9781315181202-7](https://doi.org/10.4324/9781315181202-7).
<span id="ref-130"></span>130. K. A. Kipp, et al. (2012). Impact of instrument schedule growth on mission cost and schedule growth for recent NASA missions. *2012 IEEE Aerospace Conference*. doi: [10.1109/aero.2012.6187407](https://doi.org/10.1109/aero.2012.6187407).
<span id="ref-131"></span>131. O. Williamson (1999). Public and Private Bureaucracies: A Transaction Cost Economics Perspective. doi: [10.1093/jleo/15.1.306](https://doi.org/10.1093/jleo/15.1.306).
<span id="ref-132"></span>132. Jeffrey M. Wooldridge (2025). Two-way fixed effects, the two-way mundlak regression, and difference-in-differences estimators. *Empirical Economics*. doi: [10.1007/s00181-025-02807-z](https://doi.org/10.1007/s00181-025-02807-z).
<span id="ref-133"></span>133. Kirill Borusyak, Xavier Jaravel, and Jann Spiess (2024). Revisiting Event-Study Designs: Robust and Efficient Estimation. *The Review of Economic Studies*. doi: [10.1093/restud/rdae007](https://doi.org/10.1093/restud/rdae007).
<span id="ref-134"></span>134. Kurt Schmidheiny, and Sebastian Siegloch (2023). On event studies and distributed-lags in two-way fixed effects models: Identification, equivalence, and generalization. *Journal of Applied Econometrics*. doi: [10.1002/jae.2971](https://doi.org/10.1002/jae.2971).
<span id="ref-135"></span>135. Dana E. Goin, and C. Riddell (2023). Comparing two-way fixed effects and new estimators for differences-in-differences: A simulation study and empirical example. *Epidemiology*. doi: [10.1097/ede.0000000000001611](https://doi.org/10.1097/ede.0000000000001611).
<span id="ref-136"></span>136. Clément de Chaisemartin, and Xavier D'Haultfœuille (2022). Two-Way Fixed Effects and Differences-in-Differences with Heterogeneous Treatment Effects: A Survey. *National Bureau of Economic Research*. doi: [10.3386/w29734](https://doi.org/10.3386/w29734).
<span id="ref-137"></span>137. Claude Diebolt, and Michael Haupert (2021). Cliometrics: Past, Present, and Future. *Oxford Research Encyclopedia of Economics and Finance*. doi: [10.1093/acrefore/9780190625979.013.552](https://doi.org/10.1093/acrefore/9780190625979.013.552).
<span id="ref-138"></span>138. Douglass C. North, et al. (2016). Structure and Performance: The Task of Economic History. *Journal of Economic Literature*. [https://openalex.org/W1510710415](https://openalex.org/W1510710415).
<span id="ref-139"></span>139. Gareth Austin (2015). African Economic History in Africa. *Economic History of Developing Regions*. doi: [10.1080/20780389.2015.1033686](https://doi.org/10.1080/20780389.2015.1033686).
<span id="ref-140"></span>140. M. Dwyer, Z. Szajnfarber, B. Cameron, and E. Crawley (2018). A model for understanding and managing cost growth on joint programs. *Acta Astronautica*. doi: [10.1016/j.actaastro.2018.07.004](https://doi.org/10.1016/j.actaastro.2018.07.004).



## Appendix A. Variable-Construction Codebook

This codebook is the single source of truth for how each variable enters the panel. It is written so that a second researcher could reconstruct every column from the named primary records without consulting the author. The governing principle is the Maddison single-rule standard: one documentary rule, applied identically across all eras, is the precondition for any cross-era comparison.

**Table A.1. Variable dictionary.**

| Variable | Symbol | Role | Definition | Unit | Primary source |
|---|---|---|---|---|---|
| Authorization latency | Latency(i,p,t) | Explanatory | Median elapsed time across the authorization events in program-phase (i,p,t), from documented trigger event to documented resolving authorization | Months | NASA budget and program records; NTRS process records [80, 86, 94, 95] |
| Cost growth | Outcome (1) | Dependent | (actual phase cost − baseline phase cost) / baseline phase cost, both in constant FY dollars | Fraction of baseline | Agency records; GAO assessments; deflated by NASA New Start Inflation Index |
| Schedule slip | Outcome (2) | Dependent | (actual phase duration − baseline phase duration) / baseline duration | Fraction of baseline | Agency records; GAO assessments |
| Mission cadence | Outcome (3) | Dependent | Era level: operational mission events per unit time within a program family; program level: interval between successive flight/delivery events | Events per period; months between events | Flight/delivery records, multiple family definitions |
| Spacecraft/instrument mass | X | Control | Mass at commitment | kg | Cost-estimating datasets [1, 2, 3, 130] |
| Power | X | Control | Power at commitment | W | Cost-estimating datasets [2, 3] |
| Technology readiness level | X | Control | TRL at commitment | Ordinal 1–9 | Program records |
| Mission class | X | Control | Mission class at commitment | Categorical | Program records |
| Contract type | X | Control | Primary contract structure | Categorical | Contract records [20, 85] |
| Number of external partners | X | Control | Count of distinct partner organizations | Count | Program records |
| Funding-instability index | X | Control | Year-over-year deviation between requested and appropriated funds for the program | Index | Appropriations records [88, 129] |
| Program fixed effect | \(\alpha_i\) | Nuisance | Absorbs all time-invariant program characteristics | n/a | Construction |
| Era fixed effect | \(\delta_t\) | Nuisance | Absorbs common shocks and rule regimes | n/a | Appendix C |

**A.1. The authorization-latency rule.** Latency is defined as the elapsed time between a documented trigger event (the point at which a decision or authorization becomes due) and the documented authorization event that resolves it. For each program-phase, the constructed value is the median elapsed time, in months, across that phase's authorization events. The median is preferred to the mean to reduce sensitivity to a single unusually long or short interval.

**A.2. Trigger and resolution event taxonomy.** Trigger events include the scheduling of a key decision point, the convening of a program commitment review, the opening of a life-cycle review under NPR 7120.5 [94, 95], the chartering of a Standing Review Board assessment [80], and the establishment of a joint cost-and-schedule confidence-level requirement [86]. Resolution events are the corresponding decision-authority sign-offs, confirmation decisions, and budget actions that authorize the phase to proceed. The taxonomy is applied identically across eras; where the modern record names an event the early record does not, the coarse resolution (Appendix A.3) governs.

**A.3. Coarse and fine resolutions.** The coarse measure, available for the full 1958–2026 span, is built from milestone-to-milestone intervals. The fine measure, available only for the modern subperiod, is built from individual key-decision-point records. Results are reported separately at each resolution. This two-resolution design is the response to the changing documentary density of the record and prevents a measurement artifact from being read as a historical trend.

**A.4. Deflation.** Wherever cost is compared across eras, nominal dollars are converted to constant fiscal-year dollars using the NASA New Start Inflation Index before any ratio is formed. This step is mandatory and precedes the cost-growth calculation in Table A.1.


## Appendix B. Program-Phase Inclusion List and Coverage Table

The unit of analysis is the program-phase observation. A program contributes one observation per documented lifecycle phase for which both a baseline and an actual exist, so a program with formulation, development, and operations phases contributes several observations. The panel is unbalanced by construction.

**Table B.1. Coverage by era and resolution (structure; entries populated at panel assembly).**

| Era band | Latency resolution available | Coverage density | Baseline definition | Primary cross-check |
|---|---|---|---|---|
| 1958–1969 | Coarse only | Sparse | Informal/early | Agency records |
| 1970–1989 | Coarse only | Moderate | Maturing | Agency records |
| 1990–2009 | Coarse; partial fine | Dense | Formal commitment | Aerospace Conference datasets [1, 98] |
| 2010–2026 | Coarse and fine | Dense | NPR 7120.5 commitment | GAO major-project assessments |

The earliest band carries the most construction uncertainty, which is why its observations enter only at coarse resolution and why the two-resolution reporting (Appendix A.3) exists. The inclusion list itself, naming each program-phase and its resolution flag, is generated at panel assembly and validated against published Aerospace Conference and GAO figures for overlapping programs so that the constructed measures reproduce known values where known values exist.


## Appendix C. Era-Regime Definition Table

Era fixed effects \(\delta_t\) absorb the rule regimes that, by North's logic of adaptive efficiency and path dependence, generate a persistent latency component within a regime. The table maps each regime to its \(\delta_t\) so that identification rests on within-program and within-era variation net of these regimes.

**Table C.1. Era-regime mapping (illustrative bands; exact boundaries fixed at assembly from documentary regime changes).**

| \(\delta_t\) band | Rule-regime character | Documentary marker |
|---|---|---|
| \(\delta_1\) | Pre-formal-baseline era | Absence of standardized commitment review |
| \(\delta_2\) | Early program-management formalization | Introduction of program-control practice [49, 112] |
| \(\delta_3\) | Life-cycle-review regime | NPG/NPR 7120.5 review process [94, 95] |
| \(\delta_4\) | Joint-confidence-level regime | JCL policy adoption [86] |
| \(\delta_5\) | Standing-Review-Board regime | SRB Handbook practice [80] |

Regime boundaries are set from documented rule changes, not from calendar convenience, and a program spanning two regimes contributes observations to each, with \(\delta_t\) absorbing the shared regime component in each.


## Appendix D. Robustness-Battery Specification List

The battery is fixed before estimation so that survival of it is informative rather than the product of post-hoc selection. Each item targets a named threat.

**Table D.1. Pre-specified robustness checks.**

| Check | Variation | Threat addressed |
|---|---|---|
| Clustering variants | Program-clustered; alternative clusters | Statistical-conclusion validity |
| Wild-cluster bootstrap | Given modest program-cluster count | Few-clusters inference [22, 24] |
| Alternative fixed-effects structures | Program-only; era-only; two-way | Specification dependence [116, 128] |
| Two latency resolutions | Coarse vs. fine | Construct/measurement artifact |
| Cadence-family definitions | Multiple program-family definitions | Cadence construct fragility |
| Heterogeneity-robust estimators | Callaway–Sant'Anna; de Chaisemartin–D'Haultfoeuille | Staggered-treatment heterogeneity [22, 23, 136] |
| Goodman-Bacon decomposition | Diagnostic for discrete-regime analyses | TWFE negative weighting [21, 29] |
| Instrumental-variable estimates | Office workload; appropriations timing | Endogeneity to phase difficulty [12, 111, 127] |
| Baseline-conservatism test | Reserves-to-baseline vs. latency | Baseline gaming [13, 30] |

The continuous-latency framing is the first defense against the staggered-binary pathology; the heterogeneity-robust estimators and the Goodman-Bacon diagnostic are reserved for any analysis of a discrete administrative-regime change. Central estimates are to be reported as bounded ranges conditioned on stated assumptions, in the Fogel manner, not as single causal points.


## Appendix E. Pre-Registration Record

This record is time-stamped before any estimation on the full panel and is reproduced here verbatim from the approved prospectus so that the decision rule is binding rather than discretionary.

**E.1. Hypotheses.**
- **H0 (null).** Administrative decision-and-authorization latency in NASA programs has no association with program cost growth, schedule slip, or mission cadence, after accounting for program and era fixed effects and technical controls.
- **H1 (alternative).** Longer administrative decision-and-authorization latency is associated with greater program cost growth, more schedule slip, and lower mission cadence, after accounting for program and era fixed effects and technical controls.

**E.2. Specification (fixed notation).** For program i, phase p, era t:

\[
\text{Outcome}(i,p,t) = \beta \cdot \text{Latency}(i,p,t) + \gamma \cdot X(i,p,t) + \alpha_i + \delta_t + \varepsilon(i,p,t) \qquad\qquad (1)
\]

with \(\alpha_i\) program fixed effects, \(\delta_t\) era fixed effects, \(\varepsilon\) clustered by program. Under H1: \(\beta > 0\) for cost growth and schedule slip; \(\beta < 0\) for cadence.

**E.3. Decision rule.** H0 is rejected only if the latency coefficient is statistically distinguishable from zero at conventional levels, carries the predicted sign in the same direction across the fixed-effects and instrumented specifications, and survives the robustness battery of Appendix D. If the coefficient is indistinguishable from zero, reverses sign between specifications, or vanishes under the heterogeneity-robust estimators or the alternative latency resolution, H0 is retained and the contribution fails.

**E.4. Expected signs (illustrative, NOT findings).** Cost growth: \(\beta\) expected positive. Schedule slip: \(\beta\) expected positive and largest of the three, because latency is itself a component of elapsed schedule. Cadence: \(\beta\) expected negative. These are directional expectations stated for interpretability; they are not results.

**E.5. Unpopulated illustrative table.** The illustrative coefficient table from the analysis plan remains unpopulated by design. The latency-coefficient cells read "sign expected +" for cost growth and schedule slip and "sign expected −" for cadence, with the interval and point estimate marked "to be estimated." Reporting fabricated coefficients as if real would violate the falsifiability standard on which this dissertation rests, so the cells are left empty until the panel is assembled and the pre-registered procedure is executed.