# Cost-Overrun Hazards in Earth-Observing Missions: A Competing-Risks Model Separating Instrument-Driven from Launch-Driven Slip

**A Doctoral Dissertation**

**Candidate:** JPL_ASTRO_EARTH_08

**Program:** COLLEGIUM 1st Battalion

**NORTH STAR / JPL Category:** Earth Science Missions

**Methodological anchors (Hall of Shoulders):** Fine and Gray (competing-risks subdistribution-hazard apparatus); Fogel (structural decomposition against a bounded counterfactual); Callaway and Sant'Anna (heterogeneity discipline, refusal to pool); Flyvbjerg (optimism-bias / reference-class forecasting, the controlled rival)

**Stage:** Design-stage. The model specification, identification strategy, and analysis plan are complete; all numerical results are explicitly labeled illustrative and not yet executed on the full cohort.

**Date:** 2026-06-15


## Abstract

The stewardship of public resources entrusted to an Earth-observing mission rests, in the end, on the care with which a program office holds reserve against the delays it can foresee. Schedule slip is the proximate driver of cost growth in NASA Earth-observing missions, yet the standard treatment of slip aggregates all causes into a single delay variable. That aggregation collapses two mechanisms with different physics, different owners, and different policy levers: slip originating in instrument development and slip originating in launch-vehicle availability. This dissertation proposes a competing-risks survival model of cause-specific schedule-slip onset across an Earth-mission cohort assembled from NASA Cost Analysis Data Requirement (CADRe) records, the NASA Instrument Cost Model (NICM) dataset, the Government Accountability Office (GAO) annual assessments of major NASA projects, and TechPort sensor technology-readiness-level (TRL) records. The unit of analysis is the mission-development spell from Key Decision Point B; the two competing events are first instrument-driven slip and first launch-driven slip. The estimator is the Fine-Gray subdistribution hazard model, complemented by cause-specific Cox models and Gray's test, with the subdistribution form primary because the reserve-allocation question concerns the cumulative probability of each slip type rather than the instantaneous rate alone.

The falsifiable contribution is that the two slip sources are statistically distinct competing risks and that instrument-driven slip is the dominant subdistribution hazard for missions carrying first-of-kind active sensors but not for passive-radiometer heritage missions. The null is that the two sources are not separable competing risks. The work applies the structural-decomposition logic associated with Fogel, reading the cumulative incidence function of a removed risk as a bounded counterfactual, and the heterogeneity-aware separation logic associated with Callaway and Sant'Anna, treating sensor archetype as an explicit effect modifier rather than collapsing it into a pooled coefficient, while controlling the optimism-bias rival associated with Flyvbjerg through a reference-class proxy. This is a design-stage dissertation: the specification, identification strategy, and analysis plan are fully developed, and all reported numerical results are explicitly illustrative and not executed on the full cohort. If confirmed, the result would let NASA and JPL steer scarce schedule reserve to the dominant hazard for a given mission archetype rather than to an undifferentiated slip pool; if the null holds, the field learns that slip is one hazard after all, which is itself a result worth having.


## Table of Contents

- Front Matter: Title, Abstract, Table of Contents, List of Tables and Figures
- Chapter 1: Introduction
  - 1.0 The chapter thesis
  - 1.1 The problem in full
  - 1.2 Institutional and historical context
  - 1.3 The research questions, broken out explicitly
  - 1.4 Significance for NASA, JPL, and the named stakeholders
  - 1.5 Scope and delimitations
  - 1.6 Definitions of key terms
  - 1.7 The single falsifiable contribution
  - 1.8 The intellectual lineage that disciplines the design
  - 1.9 Roadmap of the dissertation
- Chapter 2: Theoretical Framework
  - 2.0 The chapter thesis
  - 2.1 The schedule-cost coupling as the conceptual starting point
  - 2.2 The competing-risks frame and the cumulative incidence function
  - 2.3 The Fogel lens: structural decomposition against a bounded counterfactual
  - 2.4 The Callaway and Sant'Anna lens: heterogeneity and the refusal to pool
  - 2.5 The optimism-bias rival as a theoretical contender
  - 2.6 The integrated conceptual model
  - 2.7 Summary
- Chapter 3: Literature Review
  - 3.0 The chapter thesis
  - 3.1 The schedule-cost coupling in space systems
  - 3.2 Technology readiness and schedule slip: the instrument side
  - 3.3 Complexity and cost-estimating relationships
  - 3.4 Instrument-development cost-time trends and instrument-schedule growth
  - 3.5 Contracting and policy effects on mission cost and schedule
  - 3.6 The launch-availability side
  - 3.7 The competing-risks methodological literature
  - 3.8 Project-overrun economics and the optimism-bias rival
  - 3.9 The synthesized gap and the propositions that follow
- Chapter 4: Data and Measurement
  - 4.0 The chapter thesis
  - 4.1 The four named datasets in depth
  - 4.2 Unit of analysis: the mission-development spell
  - 4.3 Variable construction and the measurement table
  - 4.4 Archetype construction from the NICM taxonomy
  - 4.5 Cause-coding by two-source reconciliation
  - 4.6 Data quality, validation, coverage, and ethics
- Chapter 5: Research Design and Identification
  - 5.0 The chapter's answer
  - 5.1 The estimator and why it is chosen
  - 5.2 The identification strategy
  - 5.3 Model specification
  - 5.4 Threats to validity
  - 5.5 The computational and software plan
  - 5.6 Chapter summary
- Chapter 6: Analysis Plan and Expected Results
  - 6.0 The chapter thesis
  - 6.1 The estimation procedure
  - 6.2 Expected, illustrative findings
  - 6.3 The falsification rule
  - 6.4 The pre-registered robustness battery
  - 6.5 Power and the minimum detectable effect
- Chapter 7: Discussion
  - 7.1 Implications under both outcomes
  - 7.2 Theoretical contribution back to each anchor framework
  - 7.3 Rival explanations engaged in full
  - 7.4 External validity
- Chapter 8: Conclusion
  - 8.0 The chapter thesis
  - 8.1 Restatement of the contribution: a specification and a falsification rule, not coefficients
  - 8.2 What stands even if the hypothesis is not confirmed
  - 8.3 Limitations, stated honestly
  - 8.4 A concrete future-research program
  - 8.5 Closing
- References
- Appendices
  - Appendix A: Variable and Data Dictionary
  - Appendix B: Slip-Event Cause-Coding Manual
  - Appendix C: NICM-Taxonomy-to-Archetype Crosswalk
  - Appendix D: Pre-Registration Block
  - Appendix E: Minimum-Detectable-Effect Power Tables (Specified, Not Executed)
  - Appendix F: Source-Provenance Log


## List of Tables and Figures

**Tables**

- Table 3.1. The TRL-to-schedule-slip evidence cluster (Chapter 3)
- Table 3.2. The competing-risks apparatus and its role in the design (Chapter 3)
- Table 3.3. The optimism-bias rival and the dissertation's response (Chapter 3)
- Table 4.1. Measurement table: construct, operational definition, source, scale (Chapter 4)
- Template T6.1. Cumulative incidence by cause and archetype (Chapter 6; specified, unpopulated by design)
- Template T6.2. Subdistribution and cause-specific hazard ratios (Chapter 6; specified, unpopulated by design)
- Template T6.3. Robustness summary (Chapter 6; specified, unpopulated by design)
- Appendix A variable and data dictionary table
- Appendix E minimum-detectable-effect power table (specified, unpopulated by design)

**Figures**

No figures are presented. Consistent with the design-stage guardrail, all cumulative-incidence profiles, hazard plots, and power curves are specified in form and deliberately left unexecuted; their result tables (Templates T6.1 through T6.3 and the Appendix E power table) stand in for figures and are populated only when the cohort is assembled.




# Chapter 1: Introduction

## 1.0 The chapter thesis

Few obligations in the conduct of a national science program are as quietly consequential as the duty to commit a mission's reserve wisely, since that commitment decides, long before launch, whether a mission keeps its promise to the public that funds it. This dissertation contributes the first application of the Fine-Gray competing-risks apparatus to NASA Earth-mission schedule slip, and it does so to settle a question that current acquisition practice cannot answer with the tools it presently uses: when an Earth-observing mission begins to slip, is the slip a single undifferentiated hazard, or is it the resolution of two structurally distinct competing risks, one originating in instrument development and one originating in launch-vehicle availability, whose relative dominance is conditioned by sensor archetype? The answer this chapter develops, and the rest of the dissertation tests, is that the two slip sources are separable competing risks, and that instrument-driven slip is the dominant subdistribution hazard for missions carrying first-of-kind active sensors but not for passive-radiometer heritage missions. What the dissertation delivers is the pre-registered design and the falsification conditions for that claim, not estimated coefficients, because this is a design-stage work whose cohort has not yet been assembled and whose model has not yet been executed. The chapter sets out to establish that the question is real, that it is material, that no prior work has answered it, and that the proposed design can answer it; and to fix the scope, the definitions, the single falsifiable contribution, and the road the remaining chapters travel.

I lead with the answer rather than with background because the argument of a dissertation chapter, like the argument of any mandatory-tier analytic artifact, should be auditable from its first line. The reader should know at the outset what is being claimed, what would refute it, and why it would matter if it were true. The sections that follow develop the claim in that order: the problem in full, the institutional and historical context in which it lives, the research questions broken out explicitly, the significance for the named stakeholders, the scope and delimitations, the definitions of the key terms, the formal statement of the contribution as a null and an alternative hypothesis, and a roadmap of the dissertation.
## 1.1 The problem in full

### 1.1.1 Current state: cost growth is a monetized image of schedule slip, and slip is treated as one thing

NASA Earth-observing missions overrun their committed budgets at rates that have proven stubborn across four decades of acquisition reform. The proximate mechanism connecting a budget commitment to its overrun is almost always schedule. A mission that takes longer than its baseline plan accrues standing-army labor costs, contractor fee adjustments, integration-and-test rework, and the carrying cost of facilities and reserves held open past their planned release. Cost growth is, in large part, a monetized image of schedule slip. The practitioner literature treats this coupling as nearly definitional. Lieber and Donor, examining complex NASA and defense projects through literature review, practitioner surveys, and direct program experience, document that cost overruns and schedule delays are tightly linked, and they catalogue the root causes that drive the coupling: unrealistic estimates, supply-chain difficulties, insufficient schedule margin, technical problems, scope changes, and the realization of risk events, alongside the more diffuse factors of complexity, over-optimism, and political pressure [\[86\]](#ref-86). Their central observation, that the phrase "directly related to" hides several distinct root causes operating at once, is the seed from which this dissertation grows.

The instrument-cost and instrument-schedule literature reinforces that the locus of much of this slip is the science payload rather than the bus. Kipp, Ringler, Chapman, and Freaner show, across recent NASA missions, that instrument schedule growth propagates into mission cost and schedule growth, establishing that the instrument is a primary and separable engine of overrun [\[77\]](#ref-77). Bearden's complexity-based cost-estimating relationships establish that mission complexity and design aggressiveness independently raise both cost and schedule risk, so the apparent severity of a mission's slip is partly a function of how hard the mission tried to do [\[17\]](#ref-17). Bitten and colleagues add a third layer: acquisition policy itself moves the cost-and-schedule-growth distribution, so the same technical mission attempted under different policy regimes carries a different slip profile [\[22\]](#ref-22). The current state of practice absorbs all of this into a single delay variable. A mission is recorded as having slipped by some number of months; that slip is then regressed on covariates, or rolled into a reserve estimate, as if it were one hazard with one cause.

### 1.1.2 The two structural origins of slip

The trouble with the single-delay treatment is that slip in an Earth-observing mission has at least two structurally different origins, and the difference is not cosmetic. The first origin is instrument development. A science instrument may fail environmental testing, may be unable to close its calibration or radiometric budget, or may carry a detector or laser technology below the maturity its schedule silently assumed. When this happens, instrument development becomes the binding path, and the mission's launch date moves because the payload is not ready. The instrument-side driver is well documented. Dubos, Saleh, and Braun show, on a cross-mission dataset, that the technology-readiness level of a spacecraft's least-mature technology at the time of authorization is a measurable driver of subsequent schedule slip, with a relationship that steepens nonlinearly at the low-maturity end [\[47\]](#ref-47) (a finding first developed in their conference antecedent of the previous year). Low entry maturity is, in this literature, a leading indicator of an instrument-driven slip.

The second origin is launch-vehicle availability. A mission may be ready, or nearly ready, and still slip because the launch service is delayed by manifest congestion, by an anomaly on a shared vehicle, or by the launch provider's own development slip, none of which has anything to do with the spacecraft's readiness. Earth-observing continuity missions are especially exposed to this second origin because they are frequently manifested on shared or transitioning launch services. The Landsat Data Continuity Mission account documents this kind of launch-driven schedule pressure as a structurally separate phenomenon from instrument maturation: a continuity mission carrying heritage instruments can be technically ready and still be governed by the dynamics of its launch slot [\[72\]](#ref-72).

These two origins have different owners, different physics, and different policy levers. The instrument-development origin is owned by the instrument provider and the project's payload organization; its physics is technology maturation and verification; its policy levers are technology-readiness gating and instrument reserve. The launch-availability origin is owned by the launch-services program and the launch provider; its physics is manifest dynamics and provider development; its policy levers are launch-vehicle diversity, manifest margin, and the timing of the launch-service commitment. Treating the two as one variable collapses structurally distinct causes into an undifferentiated pool and forecloses the single analysis a program office most needs at confirmation: the analysis that would tell it which lever to pull and where to hold reserve.

### 1.1.3 Desired state, gap, and consequence

The desired state is an archetype-specific basis for steering the scarce schedule and cost reserve that a mission holds at confirmation toward the dominant first-slip hazard for that mission's archetype: toward instrument maturation and technology-readiness gating for a first-of-kind active-sensor mission, and toward launch-manifest and provider dynamics for a heritage passive-radiometer continuity mission. The gap that stands between the current state and the desired state is not a missing dataset and not a missing estimator considered in isolation. It is a missing integration, and identifying it precisely is the burden of Section 1.2.

The consequence of leaving the gap unfilled is concrete and recurring. Reserve continues to flow against an undifferentiated pooled slip estimate. A first-of-kind active-sensor mission that is under-reserved against instrument maturation, or a heritage continuity mission that is under-reserved against manifest risk, overruns or descopes; and in either case the program office cannot say, after the fact, which lever it should have pulled, because the analytic apparatus it used never separated the levers in the first place. The problem is real because slip drives cost and has two structurally distinct origins that the standard treatment collapses; it is material because the reserve allocated against slip is a first-order cost driver and the wrong allocation is paid for in overrun or descope. These two assertions, that the problem is real and that the problem is material, anchor the argument this dissertation develops from beginning to end.

## 1.2 Institutional and historical context

The problem does not float free of institutions. It lives inside a specific acquisition apparatus with a specific history, and the design of this dissertation is shaped by where in that apparatus the decisions of interest are made.

NASA flight projects proceed through a phased life cycle punctuated by Key Decision Points. The decision this dissertation cares about is made at confirmation, around Key Decision Point B and into Key Decision Point C, when a project's cost and schedule baselines are committed and when reserve is sized and posted against the risks the project expects to face. Reserve at this stage is finite and is allocated under standing policy. The institutional question, stated in the project office's own terms, is how much reserve to hold and against what. If instrument-driven and launch-driven slip are one hazard, the answer is to hold reserve centrally against an undifferentiated pool. If they are two hazards with archetype-dependent dominance, the answer is to steer reserve toward the dominant hazard for the mission's archetype. The two answers prescribe materially different reserve postures, and the apparatus presently in use cannot choose between them on evidence.

The historical context supplies the cost-modeling lineage the dissertation inherits and the data it will use. The NASA cost-estimating community has built, over decades, a family of parametric cost-estimating relationships sensitive to mass, power, data rate, complexity, and technology maturity. The instrument-cost lineage runs through the NASA Instrument Cost Model and its Explorer-class extension, whose parameter records and instrument-type taxonomy supply both covariates and the basis for the archetype classification used here [\[65\]](#ref-65). The instrument cost-time-trend studies establish that instrument development is a primary locus of cost and schedule risk distinct from the bus and from launch, and that its trends can be measured over the long run [\[66\]](#ref-66). The contracting-and-policy lineage, exemplified by the firm-fixed-price-effectiveness analysis of Sobel and Tibor and by the policy-effect analysis of Bitten and colleagues, establishes that the institutional choices around how a mission is bought move its slip distribution and must therefore be controlled when slip is decomposed [\[123\]](#ref-123), [\[22\]](#ref-22). The electronics-cost lineage of Hahn and Sholder shows that even at the parts level, cost grows as design refines, and that reserve must be posted against that growth, a reminder that reserve allocation is the live decision the cost community is trying to inform [\[67\]](#ref-67).

The competing-risks apparatus this dissertation imports comes from an entirely different institutional tradition, biostatistics and epidemiology, where the structure of a first event that can occur for one of several mutually exclusive reasons has been formalized for decades. Fine and Gray's subdistribution-hazard model and the applied guidance built around it by Austin, Lee, and Fine, and the epidemiologic cautions of Andersen and colleagues, are mature and standard in that tradition [\[55\]](#ref-55), [\[9\]](#ref-9), [\[5\]](#ref-5). They have not crossed into NASA mission analysis. The institutional novelty of this dissertation is precisely this crossing: it takes an estimator built to answer whether a patient died of cardiovascular rather than non-cardiovascular causes, and it asks instead whether a mission first slipped for instrument rather than launch reasons. The structures are formally identical; the application is new.

## 1.3 The research questions, broken out explicitly

The dissertation is organized around one principal research question and three subsidiary questions that decompose it.

**Principal research question.** For NASA Earth-observing missions, is schedule slip a single undifferentiated hazard, or are slip that originates in instrument development and slip that originates in launch-vehicle availability two statistically distinct competing risks whose relative dominance depends on sensor archetype?

**Subsidiary question 1 (separability).** Can the first schedule-slip event of an Earth-observing mission be cause-coded reliably as instrument-driven or launch-driven by reconciling the CADRe Part A narrative with the GAO assessment narrative, and once so coded, are the two cause-specific hazard processes statistically distinguishable, so that the two slip sources behave as genuine competing risks rather than as one process with arbitrary labels?

**Subsidiary question 2 (archetype dependence).** Conditional on separability, does the dominance between the two slip hazards depend on sensor archetype, such that the instrument-slip subdistribution hazard exceeds the launch-slip subdistribution hazard for missions carrying first-of-kind active sensors but not for passive-radiometer heritage missions?

**Subsidiary question 3 (robustness against the rival).** Does any separation and archetype dependence found in answer to the first two questions survive a control for estimating optimism, the leading rival explanation in the project-overrun literature, so that the decomposition is not an artifact of which missions were optimistically baselined?

These three subsidiary questions map one-to-one onto the three conditions the falsification rule requires for confirmation, which Section 1.7 of the design (Chapter 6) fixes in advance: separability, archetype dependence, and robustness. Breaking the principal question into these three pieces is not a rhetorical convenience. It is the structure that makes the contribution falsifiable in three concrete and independent ways, so that a failure of any one piece is a clean, pre-specified refutation rather than a post-hoc reinterpretation.

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

The significance of the work follows directly from where the decision is made and who makes it.
For NASA as the funding and oversight institution, the result, if confirmed, converts an undifferentiated reserve posture into an archetype-specific one. Schedule and cost reserve at confirmation would be sized and steered according to the mission's dominant first-slip hazard rather than against a single pooled slip estimate. This is a sharper prescription than current practice because the decision-relevant quantity it rests on, the cumulative incidence of each slip cause by archetype, is exactly what the subdistribution model estimates, and the pooled approach cannot produce it. The causal mechanism that licenses the prescription is explicit. A mission that carries a first-of-kind active sensor below its assumed maturity at confirmation has, as its driver, an immature least-ready technology. The mechanism is that the technology fails environmental test, cannot close its calibration budget, or matures slowly, so instrument development becomes the binding path. The observable effect is an instrument-driven first slip that raises the instrument-slip cumulative incidence faster than the launch-slip cumulative incidence in the active-sensor stratum. The operational consequence is that the committed launch date moves and standing-army and rework cost overrun the baseline. The strategic implication is that reserve and technology-readiness gating should be steered to instrument maturation for this archetype. For the heritage passive-radiometer arm, the chain runs through a direct-rebuild instrument that carries little maturation risk, so the first slip is disproportionately launch-side: a shared-vehicle anomaly, a manifest reshuffle, or a provider's own development slip, and reserve should be steered to manifest and provider dynamics instead. This is a named mechanism, not a bare correlation. Where the dissertation can offer only correlation, as on the comparatively thin launch-availability side, it says so and downgrades its confidence accordingly.

For JPL specifically, the significance is direct because JPL disproportionately leads the first-of-kind active-sensor Earth missions, the lidars and radars with little or no flight heritage, that the dominance result is about. If instrument-driven slip is confirmed as the dominant first-slip hazard for that archetype, the result informs JPL's reserve posture and its technology-readiness gating at Key Decision Point B for exactly the class of mission JPL builds. The candidate's home category in the NORTH STAR and JPL framing, Earth Science Missions, is the category this dissertation studies, and the active-sensor archetype is JPL's institutional specialty.

For the named stakeholders in the program and cost-analysis communities, the significance is methodological and operational at once. The cost-analysis community that produces and consumes CADRe, NICM, and the GAO assessments gains a model that uses their own records, the milestone schedules, the instrument parameters, and the slip-cause narratives, to answer a question those records were never previously made to answer. The program office that posts reserve at confirmation gains a defensible, archetype-specific basis for an allocation it currently makes against a pooled estimate. The significance is bounded honestly: the result, if found, would be a better basis for one specific decision for one specific class of mission, not a universal theory of overrun.

## 1.5 Scope and delimitations

The scope is deliberately narrow, and narrowness is a methodological virtue here rather than a limitation to apologize for.

The population is NASA Earth-observing missions from roughly 1990 to the present that reach Key Decision Point B, an expected cohort on the order of thirty to sixty missions. The temporal window begins where the formal cost-and-schedule record becomes consistent and ends at the present. The life-cycle window of the analysis begins at Key Decision Point B, the start of preliminary design, and ends at the earlier of the first recorded slip event or launch readiness. The dissertation does not study pre-formulation slip, which is incurred before the cost-and-schedule record stabilizes, except to acknowledge it as left truncation in the spell-origin convention.

Three classes of mission are out of scope by construction. Commercial Earth-observing constellations are excluded because they face different launch economics, different procurement, and a different reserve logic. Planetary and astrophysics missions are excluded because their instrument and launch risk structures differ from the Earth-observing case. Non-US missions are excluded because their acquisition and cause-attribution records are not commensurable with the CADRe and GAO sources. Any result is therefore bounded to NASA Earth-observing missions of the studied era and does not transport beyond it; this boundary is restated explicitly in the discussion so that no reader mistakes a conditional, archetype-specific finding for a general law of cost overrun.

The contribution is also delimited in what it claims. It does not assert that instrument-driven slip is always the larger problem, nor that launch-driven slip is negligible, nor that sensor novelty is the only effect modifier that could matter. It asserts a specific conditional dominance: that instrument-driven slip is the leading first-slip hazard for one identifiable archetype and not for another. This deliberate narrowness is what makes the claim testable on a small cohort and what maps it onto a decision a program office actually makes. A broader claim would be harder to falsify and less useful to the people who would act on it.

One further delimitation governs the form of the dissertation itself. This is a design-stage work. The model specification, the identification strategy, the data sources, and the analysis plan are fully developed; but the cohort has not been assembled, the cause-coding has not been frozen, and the model has not been executed. Every numerical value that appears anywhere in the dissertation is labeled as illustrative, specifying the form of an expected output rather than reporting an empirical finding. No estimated coefficient, hazard ratio, or cumulative-incidence plateau is presented as a result. This honesty is a scope decision, not a hedge: the deliverable is the pre-registered design and its falsification conditions, and the dissertation is to be judged on the soundness of that design.

## 1.6 Definitions of key terms

Precision in the following terms is load-bearing, because the entire contribution turns on the separation they encode.

**Schedule slip event.** A slip event is a committed-baseline schedule movement exceeding a stated threshold, illustratively two months of net launch-date movement, attributable to a single dominant cause within a milestone period. The threshold and the single-dominant-cause requirement are part of the construct, not incidental settings, and the threshold is varied in the robustness battery so that no conclusion depends on a single cutoff.

**Instrument-driven slip and launch-driven slip (the competing events).** The two competing events are the first instrument-driven slip and the first launch-driven slip. A slip's cause is coded as instrument-driven or launch-driven by reconciling the CADRe Part A narrative with the GAO assessment narrative for the same project-year; where both name the same cause, the coding is high-confidence, and where they cannot be reconciled the event is flagged un-codable and handled in sensitivity analysis. The two cause-coded first-slip events are the competing risks, so named because the occurrence of one as the first event alters the at-risk set for the other and precludes its clean observation as the first cause.

**Mission-development spell (the unit of analysis).** The unit of analysis is the mission-development spell, the interval from the start of Phase B at Key Decision Point B to the earlier of the first recorded schedule-slip event of either cause or the launch readiness date. Time is measured in months from Key Decision Point B. Each mission contributes exactly one spell, and the spell ends in one of three states: instrument-driven first slip, launch-driven first slip, or administrative censoring at launch with no above-threshold slip.

**Sensor archetype (the effect modifier).** Each mission is classified from the NICM instrument taxonomy as carrying a first-of-kind active sensor, for example a lidar or radar with no direct flight heritage, or as a passive-radiometer heritage mission, a passive radiometer with direct heritage from a predecessor instrument. Missions whose sensors are mixed or ambiguous form a third category used only for robustness checks and excluded from the primary contrast. The archetype is the effect modifier whose interaction with the instrument-slip hazard operationalizes the dominance claim.

**Cumulative incidence function (the decision-relevant quantity).** The cumulative incidence function for cause k is the probability of experiencing a cause-k first slip by a given time, accounting for the competing event. It is the decision-relevant quantity because the reserve question is predictive: a program office needs the cumulative probability of each slip type by archetype, not only the instantaneous rate. The cumulative incidence function, and not the complement of a naive Kaplan-Meier estimate that treats the competing event as ordinary censoring, is the correct estimand, because the naive estimate is biased upward whether or not the competing events are independent [\[9\]](#ref-9), [\[5\]](#ref-5).

**Subdistribution hazard and cause-specific hazard.** The subdistribution hazard is the hazard that governs the cumulative incidence function and so answers the predictive question of how a covariate moves the absolute probability of a cause-k slip in the real population where the competing event occurs. The cause-specific hazard answers the etiologic question of how a covariate moves the instantaneous rate of a cause-k slip among missions still at risk. The dissertation estimates both, with the subdistribution hazard primary because the policy question is predictive, and reports divergence between them rather than suppressing it [\[9\]](#ref-9).

**Estimating optimism (the rival).** Estimating optimism is the systematic tendency for planning-stage baselines to be set more favorably than outcomes warrant, the mechanism that Flyvbjerg's reference-class-forecasting program identifies as the behavioral root of cost overrun [\[57\]](#ref-57). It is operationalized here as the ratio of the confirmation-baseline schedule to a reference-class median schedule for similar-class missions, and it is the control whose removal-and-reintroduction tests whether any apparent slip-cause separation is in fact an optimism artifact.

## 1.7 The single falsifiable contribution

The dissertation states one falsifiable contribution, given here verbatim as it is carried through every chapter.

**H1 (contribution).** For Earth-science missions, schedule slip from instrument development and schedule slip from launch-vehicle availability are statistically distinct competing risks, and instrument-driven slip is the dominant subdistribution hazard for missions carrying first-of-kind active sensors but not for passive-radiometer heritage missions.

**H0 (null).** The two slip sources are not separable competing risks; either their subdistribution hazards are statistically indistinguishable across the cohort, or the dominance of instrument-driven slip does not differ between first-of-kind active-sensor missions and passive-radiometer heritage missions.

The contribution is falsifiable in three concrete and independent ways, and I state them now so that the standard of refutation is fixed before any data are touched. First, if a nonparametric test of equality of the cumulative incidence functions across archetype strata, or a comparison of the cause-specific hazards, cannot distinguish the two slip processes, the events are not separable and H0 holds. Second, if the interaction between sensor archetype and the instrument-side covariates in the instrument-slip subdistribution model is not statistically distinguishable from zero, or carries the wrong sign, the dominance claim fails even if separability survives. Third, if any separation and dominance found disappear once estimating optimism is controlled, the contribution is not robust and is reported as an artifact of the estimating process rather than a real decomposition of slip. Confirmation requires all three of separability, archetype dependence, and robustness; failure of any one falsifies the contribution. Because the cohort is small, the design pre-commits to a power analysis run before estimation and to the rule that a non-rejection accompanied by wide confidence intervals is reported as inconclusive rather than as positive support for H0, which protects against over-claiming in either direction.

Three further commitments follow from this falsification standard, stated here in compressed form and carried in full through the design and discussion chapters. The proposed design addresses the causal mechanism, because it separates the two cause-specific hazards without contaminating one with the other. It improves on the alternatives, because a single pooled slip regression and a naive Kaplan-Meier that treats the competing event as censoring cannot decide separability or dominance, and the latter is biased upward. And its residual risk is acceptable, because results are stated as conditional, design-stage, and counterfactual, cause-coding error is bounded by two-source reconciliation and recoding sensitivity, small-sample power is stated in advance, and the estimating-optimism rival is explicitly controlled. The design and discussion chapters develop each of these commitments against the supporting evidence; they are named here so that the reader can track them as the argument unfolds.

I note explicitly that this dissertation does not construct an enterprise-architecture or systems-architecture traceability chain, and it deliberately omits the strategic-objective-to-capability-to-system-function vocabulary that such a chain would require. The contribution is an empirical-econometric survival study whose unit of analysis is the mission-development spell. There is no real capability, system, or data-service exchange in scope. The decision-relevance that an architecture chain would otherwise carry is carried instead, in prose, by the reserve-allocation policy implication: the objective is correct reserve allocation at confirmation, and the decision is the archetype-specific reserve posture and the technology-readiness gating at Key Decision Point B. Forcing architecture vocabulary onto a statistical contribution would be a category error, and the dissertation refuses it.
## 1.8 The intellectual lineage that disciplines the design

Three methodological traditions, beyond the competing-risks estimator itself, discipline the design, and naming them here prepares the reader for the theoretical framework that follows.

The first is the structural-decomposition tradition associated with Fogel, whose signature move is to take an aggregate outcome that everyone attributes to a single dominant cause and decompose it into separable channels whose individual magnitudes are measured against an explicitly constructed counterfactual, as in the social-savings method that separated the contribution of one transport mode from an economy-wide aggregate [\[82\]](#ref-82). Applied here, the lens forbids leaving schedule slip as an aggregate. The dissertation decomposes it into the instrument channel and the launch channel and reads the cumulative incidence function of a removed competing risk as the counterfactual-bearing quantity: what fraction of an archetype's missions would have slipped first for instrument reasons in the world where launch slip also competes for primacy. The same lens carries Fogel's own methodological warning, that a constructed counterfactual must be made explicit and bounded with sensitivity analysis, which is why the dissertation reports its decompositions with stated assumptions, stated bounds, and a sensitivity analysis on the cause-coding and the at-risk structure.

The second is the heterogeneity-aware tradition associated with Callaway and Sant'Anna, whose refusal to let a pooled regression coefficient stand in for a heterogeneous set of underlying comparisons is the discipline behind treating sensor archetype as an explicit effect modifier rather than collapsing it into a pooled coefficient [\[27\]](#ref-27). The analogous error this lens forbids is estimating a single instrument-slip hazard pooled across all Earth missions and reporting it as "the" instrument hazard, when the hazard differs sharply between first-of-kind active-sensor missions and heritage passive-radiometer missions. The design instead estimates archetype-specific hazards as separable building blocks and forms any aggregate as a transparent weighting of them, and the same authors' doubly-robust logic motivates the parallel use of penalized regression and reweighting as a robustness check [\[118\]](#ref-118).

The third is the optimism-bias tradition associated with Flyvbjerg, which is not a tool the dissertation adopts but the rival explanation it must defeat: the claim that overruns are driven by systematic planning-stage optimism rather than by project-specific technical cause [\[57\]](#ref-57). A credible separation of slip causes must show that the instrument-versus-launch distinction survives controls for this optimism, which is exactly what the third falsification route tests.

## 1.9 Roadmap of the dissertation

The dissertation proceeds in eight chapters. This first chapter has established the problem, its context, the research questions, the significance, the scope, the definitions, and the falsifiable contribution. Chapter 2 develops the theoretical framework: the schedule-cost coupling as the conceptual starting point, the competing-risks frame and why the cumulative incidence function rather than the Kaplan-Meier complement is the decision-relevant quantity, the Fogel structural-decomposition lens, the Callaway-Sant'Anna heterogeneity lens, the optimism-bias rival as a theoretical contender, and the integrated conceptual model the empirical work will test. Chapter 3, the longest, reviews the literature in full: the schedule-cost coupling in space systems, the technology-readiness-to-slip relationship on the instrument side, complexity and cost-estimating relationships, instrument cost-time trends, contracting and policy effects, the launch-availability side, the competing-risks methodological literature mature in biostatistics but never applied to mission slip, the project-overrun-economics rival, and the synthesized gap.

Chapter 4 describes the data and measurement in depth: the four named sources, CADRe through the ONCE database under data-use agreement, the NICM and NICM-E instrument-parameter records, the GAO Assessments of Major NASA Projects, and the NASA TechPort sensor technology-readiness histories; the mission-development spell as the unit of analysis with its left-truncation convention; the full measurement table for every variable; the archetype construction from the NICM taxonomy; the two-source cause-coding procedure; and the coverage and limitations. Chapter 5 specifies the research design and identification: the Fine-Gray subdistribution-hazard estimator with the cause-specific Cox model in parallel, the identification strategy resting on the three separability claims and the temporal ordering of covariates, the model specification with its ridge-penalized partial likelihood and events-per-variable cap, the threat catalogue across internal, external, construct, and statistical-conclusion validity, and the computational plan.

Chapter 6 lays out the analysis plan: the fixed estimation sequence, the expected and explicitly illustrative findings under H1 and under H0 and under partial support, the falsification rule fixed in advance, the pre-registered robustness battery, and the power and minimum-detectable-effect analysis run before estimation. Chapter 7 discusses the implications under each possible outcome, the theoretical contribution back to each anchor tradition, the five rival explanations and how the design distinguishes each, and the external-validity boundary; this is where the decision-relevance of the work lives, as policy rather than as architecture. Chapter 8 concludes by restating the contribution as the specification and falsification conditions rather than as coefficients, by setting out what stands even if H1 is not confirmed, by stating the limitations honestly, and by laying out the concrete future-research program that would execute the design. A backmatter supplies the full reference list with hyperlinked citations and a set of appendices: the variable and data dictionary, the cause-coding manual, the taxonomy-to-archetype crosswalk, the pre-registration block, the power tables, and the source-provenance log for the corpus.

The thread that runs through all eight chapters is the single claim stated in Section 1.7: that NASA Earth-mission schedule slip is two separable competing risks whose dominance is conditioned by sensor archetype, and that this can be tested, and falsified, by the design developed here. If the design is executed and the contribution confirmed, NASA and JPL gain a defensible, archetype-specific basis for allocating the scarce reserve that ultimately determines whether an Earth-observing mission overruns. If the design is executed and the null holds, the field learns that slip is one hazard after all, which is itself a result worth having and a cleaner foundation than the field currently stands on.


## Chapter 1 References

Citations in this chapter are numbered to the consolidated reference list in the Back Matter (Part I: References); each in-text marker links directly to its full entry there.


# Chapter 2: Theoretical Framework

## 2.0 The chapter thesis

This chapter argues that NASA Earth-mission schedule slip is correctly modeled as a competing-risks process with two cause-specific hazards, and that the conceptual apparatus needed to state and test that claim already exists, fully developed, in three literatures that have never been joined. The first is the competing-risks survival framework of Fine and Gray, which supplies the estimator and the decision-relevant quantity: the cumulative incidence function of a single cause in a world where other causes also occur. The second is the structural-decomposition tradition associated with Fogel, which insists that an aggregate everyone attributes to one dominant cause be split into separable channels and measured against an explicitly bounded counterfactual. The third is the heterogeneity discipline of Callaway and Sant'Anna, which refuses to let a single pooled coefficient stand in for a set of underlying comparisons that differ across subgroups. Each framework was built for a different empirical setting (biostatistical survival analysis, cliometric history, and panel-data econometrics), and each carries a primary literature with its own warnings. This chapter shows, anchor by anchor, how each transfers to the problem of separating instrument-driven from launch-driven slip, and then assembles the three into a single conceptual model that the empirical chapters will test.

The problem this chapter addresses can be stated as current state, desired state, gap, and consequence. The current state of the analytic practice is that schedule slip enters the spacecraft cost-and-schedule literature as a single continuous outcome regressed on covariates, so that a mission is recorded as having slipped and the slip is treated as one hazard with one cause. The desired state is a framework in which the two structurally distinct origins of slip, instrument development and launch availability, are carried as separate events with separate hazards, so that the cumulative probability of each can be estimated conditional on what kind of sensor the mission carries. The gap is that no existing framework does this for mission slip: the survival apparatus is mature in biostatistics but unapplied to acquisition outcomes, the decomposition logic is mature in economic history but has never been pointed at a hazard process, and the heterogeneity discipline is mature in policy evaluation but has not been used to keep an instrument hazard from being averaged across archetypes it does not describe. The consequence of leaving the gap open is that schedule reserve continues to be allocated against an undifferentiated slip pool, and a program office cannot learn from the data which of its two levers, technology gating or launch-manifest margin, the data say it should pull for the archetype in front of it.

The chapter proceeds by building each framework as a substantive section: its core construct, its primary sources, and its precise transfer to this problem. Section 2.1 establishes the schedule-cost coupling that makes slip the right object of study at all. Section 2.2 develops the competing-risks frame and defends the cumulative incidence function as the decision-relevant quantity. Section 2.3 develops the Fogel decomposition lens and its counterfactual discipline. Section 2.4 develops the Callaway and Sant'Anna heterogeneity lens and its doubly-robust motivation for the robustness design. Section 2.5 develops the optimism-bias rival as a theoretical contender that the framework must survive rather than ignore. Section 2.6 assembles the integrated conceptual model and states the causal mechanism the empirical work will test. Throughout, the design-stage discipline holds: every framework is developed as specification, no estimated quantity is reported, and any illustrative magnitude is labeled as such.


## 2.1 The schedule-cost coupling as the conceptual starting point

The premise on which the entire study rests is that cost growth in NASA Earth-observing missions is, to a first approximation, a monetized image of schedule slip. If this coupling did not hold, slip would be a secondary symptom and the right object of study would be cost directly. The coupling is what licenses treating the timing of slip as the primary outcome and cost as its downstream consequence.

The evidence for the coupling comes from the practitioner cost literature. Lieber and Donor, working from a literature review, surveys of project practitioners, and direct experience on NASA programs, document that cost overruns and schedule delays are tightly linked on complex projects, and that the everyday phrase "directly related to" conceals several distinct root causes operating at once [\[86\]](#ref-86). Their catalogue of causes is the useful part for this dissertation. They distinguish causes that are obvious in hindsight, such as unrealistic estimates, technical problems, supply-chain difficulties, and scope changes, from causes that are diffuse, such as complexity, optimism, and political factors. What connects this evidence to the present argument is the mechanism by which schedule converts to cost: a mission that runs long accrues standing-army labor, contractor fee adjustments, and rework, so that each month of slip carries a monetary cost that compounds. That mechanism is supported by the wider cost-modeling record. Kipp and colleagues isolate instrument schedule growth specifically and show it propagates into mission cost and schedule growth across recent NASA missions, establishing that the instrument is a distinct and measurable locus of the coupling rather than an undifferentiated part of the whole [\[77\]](#ref-77). Bitten and colleagues show that policy changes themselves move NASA science-mission cost and schedule growth, confirming that the coupling is sensitive to the acquisition environment and not a fixed technical constant [\[22\]](#ref-22).

One limit on the coupling matters and is carried forward into the design. Cost growth is largely, not entirely, a function of slip; a mission can overrun for reasons that are not schedule-mediated, such as a discrete materials price shock or a deliberate scope addition funded with new money. A critic might object that if some cost growth bypasses schedule, then a slip-centered framework misses part of the phenomenon. The response is that the dissertation does not claim to explain all cost growth; it claims that the slip channel is the dominant and policy-tractable one, and that the part of cost growth that bypasses slip is outside the scope deliberately, because it is not addressable by the reserve and gating levers the study is built to inform. Confidence in the coupling is high: it is supported by convergent practitioner, cost-model, and policy-effect evidence, and it is the least controversial premise in the chapter.

The transfer of this premise to the framework is direct. Because cost growth tracks slip, the object that must be decomposed is slip, not cost. And because slip has, by the Lieber and Donor cataloguing, multiple root causes that the aggregate hides, the decomposition is not a stylistic preference but a response to an explicitly documented aggregation problem. The remaining sections supply the three pieces of apparatus that turn that response into a testable model.


## 2.2 The competing-risks frame and the cumulative incidence function
### 2.2.1 The core construct

The competing-risks frame is the methodological backbone of the dissertation. Its central insight is that when a subject can fail from more than one cause, and the occurrence of one cause removes the subject from the risk set for the others, the failure causes are not independent processes to be analyzed one at a time but competing events that must be analyzed jointly. A mission's first schedule slip is exactly this kind of outcome: the first slip is either instrument-driven or launch-driven, and once a mission has slipped for one reason and that slip becomes the recorded first event, the clean observation of the other cause as the "first" event is foreclosed. This is the defining structure of competing risks, and it is why slip cannot be analyzed by running two independent survival models and ignoring the interference between them.

The frame distinguishes two hazard quantities, and that distinction is the conceptual heart of this section. The cause-specific hazard for event type k is the instantaneous rate of cause-k events among subjects still event-free at time t. It answers the etiologic question: among missions that have not yet slipped, how does a covariate change the rate at which instrument-driven slip occurs. The subdistribution hazard, introduced by Fine and Gray, governs the cumulative incidence function directly. It keeps subjects who have experienced a competing event in a modified risk set, so that the model regresses on the cumulative probability of cause-k slip by time t in the real population where the competing cause also operates [\[55\]](#ref-55). It answers the predictive question: for a mission of a given archetype, what is the probability it will have slipped for instrument reasons by month t, accounting for the fact that launch-driven slip is also competing to be its first slip.

### 2.2.2 The primary sources and what their convergence establishes

The primary literature on this frame is unusually settled, and reading it as a body rather than as a list yields a clear instruction for design. Fine and Gray provide the estimator and the formal object, the proportional-hazards model for the subdistribution of a competing risk, which allows direct regression modeling of the cumulative incidence of one event type in the presence of others [\[55\]](#ref-55). Gelber and Gray make the foundational normative argument that cumulative incidence functions should replace integrated cause-specific hazard functions when the goal is to describe the probability of competing events over time, because the integrated cause-specific hazard does not correspond to any real-world probability once competing events are present [\[63\]](#ref-63). Ray develops the class of K-sample tests for comparing the cumulative incidence of a competing risk across groups, the nonparametric instrument the dissertation uses to test whether the cumulative incidence functions differ across archetype strata [\[115\]](#ref-115).

The applied literature then converts these foundations into a usable decision rule. Austin, Lee, and Fine give the canonical guidance: the cause-specific hazard answers the etiologic question of how a covariate affects the instantaneous rate of an event among those at risk, while the subdistribution hazard answers the predictive question of how a covariate affects the cumulative probability of the event in the population where competing events occur, and the choice between them should follow the question being asked [\[9\]](#ref-9). Latouche, Allignol, Beyersmann, Labopin, and Fine sharpen this into a reporting standard: a competing-risks analysis should report results on all cause-specific hazards and the cumulative incidence functions, not one or the other, because each answers a different and complementary question [\[79\]](#ref-79). Lau, Cole, and Gange provide the epidemiologic primer that situates competing-risk regression models in observational data with the warnings that observational settings demand [\[80\]](#ref-80). Wolbers, Koller, Stel, Schaer, Jager, Leffondre, and Heinze summarize the objectives and approaches of competing-risks analysis for a cardiology readership and reinforce that the analyst must match the hazard quantity to the inferential objective [\[135\]](#ref-135).

The convergence of these sources establishes three things for the present argument. First, the subdistribution hazard and its cumulative incidence function are the decision-relevant quantities when the question is predictive, as the reserve-allocation question is: a program office wants to know the probability that a mission of a given archetype slips for a given cause, not only the instantaneous rate among the still-unslipped. Second, the cause-specific hazard must be reported alongside, not instead, because divergence between the two is itself informative about whether a covariate acts on the rate or on the cumulative burden, and the reporting standard forbids suppressing either [\[79\]](#ref-79). Third, the comparison across groups must be made on the cumulative incidence functions with a proper K-sample test [\[115\]](#ref-115), not on a quantity that ignores the competing structure. The dissertation adopts all three: the subdistribution model is primary, the cause-specific Cox model is estimated in parallel, and Gray's-type K-sample testing across archetype strata is the nonparametric check.

### 2.2.3 Why the cumulative incidence function and not Kaplan-Meier

One technical point in this literature carries the framework, and it deserves its own treatment, because the most common error in applied competing-risks work is exactly the error this dissertation must avoid. When a competing event is treated as ordinary right-censoring, the naive one-minus-Kaplan-Meier estimator of the cumulative probability of the event of interest is biased upward, because censoring assumes the censored subject remains at risk and could still experience the event, whereas a competing event means the subject can never experience the event of interest at all. Andersen, Geskus, de Witte, and Putter lay out this pitfall and its consequences in detail, and warn that the bias is not a small-sample curiosity but a structural feature of mishandling the competing event [\[5\]](#ref-5). The claim that the Kaplan-Meier approach overstates the probability of the cause of interest is therefore not a stylistic preference for the cumulative incidence function. It is a correction of a known bias whose direction and cause are understood.

The transfer of this point to mission slip is concrete. If launch-driven slip were treated as ordinary censoring in a model of instrument-driven slip, the estimated probability of instrument-driven slip would be biased upward, because missions that actually slipped first for launch reasons, and therefore could never be recorded as instrument-first, would be counted as if they remained at risk of an instrument-first slip. For an archetype contrast the bias is doubly dangerous, because the two archetypes face different mixtures of the two competing causes, so the upward bias would differ between them and could manufacture or mask an apparent archetype difference. The cumulative incidence function, estimated through the subdistribution hazard, is the quantity that handles the competing event correctly and so is the only defensible basis for the archetype comparison. Confidence in this commitment is high. It rests on a settled and convergent statistical literature with a known bias mechanism, and no part of the design depends on contradicting it.

The qualifier the same literature attaches must also be carried. Latouche and colleagues and the applied guidance all note that the subdistribution hazard's risk set, which retains subjects who have had a competing event, makes its coefficients harder to interpret etiologically. A covariate can raise the cumulative incidence of one cause partly by lowering the competing cause, so the subdistribution coefficient is a statement about cumulative burden, not about biological or physical rate [\[9\]](#ref-9), [\[79\]](#ref-79). The dissertation accepts this qualifier directly: it reports the subdistribution model for the predictive policy quantity and the cause-specific model for the rate interpretation, and it reads any divergence between them as a substantive finding about whether the archetype acts on the rate of instrument slip or on its cumulative share of first slips.


## 2.3 The Fogel lens: structural decomposition against a bounded counterfactual

### 2.3.1 The core construct

The second anchor supplies the interpretive discipline that turns a competing-risks estimate into an answer to the dissertation's real question. The competing-risks apparatus tells the analyst how to estimate the cumulative incidence of each cause without contaminating one with the other; it does not, by itself, tell the analyst to decompose the aggregate at all. The instruction to decompose comes from the structural-decomposition tradition associated with Robert Fogel. Its signature move is to take an aggregate outcome that the field attributes to a single dominant cause and split it into separable channels whose individual magnitudes are measured against an explicitly constructed counterfactual, rather than accepting the aggregate attribution on its face.

The construct has a precise analogue in the present problem. "Schedule slip" is the aggregate that the standard treatment attributes to a single delay process. The Fogel discipline demands that this aggregate not be left whole. It is decomposed into the instrument channel and the launch channel, and for each archetype the question is posed counterfactually: what would the cumulative incidence of slip be in the world where one channel were removed and only the other competed for primacy. The cumulative incidence function of a competing risk is, read through this lens, the counterfactual-bearing quantity. The instrument-slip cumulative incidence function answers what fraction of missions of a given archetype would experience instrument-driven slip first, in the world where launch-driven slip also competes; it is not a directly observed frequency but a constructed counterfactual quantity that depends on the modeled competing structure.

### 2.3.2 The primary sources and the warning they carry

The primary literature here is the social-savings tradition and its critics. Fogel's original demonstration, in the railroads-and-American-economic-growth program, decomposed an aggregate that contemporaries attributed almost entirely to one transport mode into separable channels and measured the marginal contribution of the railroad against a constructed counterfactual of alternative transport [\[113\]](#ref-113). The method was immediately contested precisely on the ground that the counterfactual was constructed rather than observed, which is the defining vulnerability of any decomposition that asks what would have happened in a world that did not occur. The subsequent literature is, in effect, a long argument about how to make such a construction credible. Leunig's reassessment of the passenger social savings from Victorian British railways is the most directly instructive primary source for this dissertation, because it re-executes the social-savings decomposition with explicit attention to the assumptions, shows how the headline number moves when the counterfactual valuation of time is changed, and thereby demonstrates that the decomposition is only as good as the transparency and sensitivity analysis attached to it [\[82\]](#ref-82). Leunig's later survey of the social-savings method generalizes the lesson: the method is powerful but its results are conditional on a stated counterfactual and must be bounded [\[124\]](#ref-124). The extension of the market-access decomposition logic to other settings, including the reallocation and market-access reworkings of the railroad question and parallel infrastructure-valuation work, confirms that the decomposition tradition is alive and that its discipline is the explicit, bounded counterfactual rather than any single estimate [\[112\]](#ref-112).

The convergent lesson from this body is not the magnitude of any railroad's contribution, which is irrelevant here, but the methodological warning that travels with the method. A decomposition's central number is a counterfactual quantity, and a counterfactual quantity must be reported with its assumptions stated, its construction made explicit, and its sensitivity to the construction shown. Fogel's own decompositions were defended exactly by making the counterfactual explicit and bounding it; the analogue for this dissertation is that the cumulative incidence function of a removed competing risk must be reported with the same explicitness, because it too is constructed rather than observed.

### 2.3.3 The transfer to mission slip

The transfer has two parts, one substantive and one disciplinary. Substantively, the Fogel lens is what authorizes the dissertation's framing of the question as a decomposition at all. Without it, a critic could ask why slip should be split into instrument and launch channels rather than analyzed as one process; the answer is that the aggregate hides causes with different owners and levers, exactly as the railroad aggregate hid the separable contribution of one mode, and the decision the program office faces is a decision about a channel, not about the aggregate. The cumulative incidence function of each cause within each archetype is the channel-specific, counterfactual-bearing quantity that the decision requires.

Disciplinarily, the lens imposes a reporting obligation that the design adopts in advance. Because the archetype-specific cumulative incidence functions are constructed counterfactuals, they must be accompanied by a sensitivity analysis that shows how the decomposition moves when the cause-coding or the at-risk structure is perturbed. This is not a generic robustness gesture; it is the specific defense the social-savings tradition developed for the specific vulnerability of constructed counterfactuals. The design therefore pre-commits to varying the slip threshold, to recoding ambiguous cause-attributions both ways, and to reporting the decomposition under each, so that a decomposition that flips under a plausible perturbation is reported as fragile rather than as a finding. What links the Fogel literature to this commitment is that a decomposition whose counterfactual is not bounded is, by the field's own hard-won standard, not yet a result, a principle that rests on the entire post-Fogel methodological argument about constructed counterfactuals [\[82\]](#ref-82), [\[124\]](#ref-124). Confidence that the decomposition framing is appropriate is high; confidence that any particular decomposed magnitude will be robust cannot be assessed at the design stage and is deliberately left open, the honest position the lens itself demands.


## 2.4 The Callaway and Sant'Anna lens: heterogeneity and the refusal to pool

### 2.4.1 The core construct

The third anchor supplies the discipline that protects the dissertation's central conditional claim from being averaged away. The competing-risks frame separates the two causes; the Fogel lens authorizes the decomposition; but neither, by itself, prevents the analyst from estimating a single instrument-slip hazard pooled across all Earth missions and reporting it as "the" instrument hazard. The Callaway and Sant'Anna tradition is the refusal to let a single pooled coefficient stand in for a heterogeneous set of underlying comparisons. Its core construct is that when an effect differs across subgroups or across timing, a pooled estimator returns a weighted average of the underlying group-specific effects, and that weighting can be opaque, can include negative weights, and need not correspond to any quantity the researcher intends to estimate.
The transfer of this construct to the dissertation is immediate and structural. The hypothesis is conditional: instrument-driven slip is the dominant first-slip hazard for missions carrying first-of-kind active sensors but not for passive-radiometer heritage missions. A pooled instrument-slip hazard estimated across both archetypes would average a stratum where the hazard is expected to be large against a stratum where it is expected to be small, returning a middling number that describes neither archetype. The Callaway and Sant'Anna discipline requires instead that sensor archetype enter as an explicit effect modifier, that the archetype-specific subdistribution hazards be estimated as separable building blocks, and that any aggregate be a transparent and defensible weighting of those blocks rather than a regression-imposed average. This is the interaction term that operationalizes the contribution: the parameter of interest is the coefficient on the archetype-by-instrument-side interaction, not a pooled main effect.

### 2.4.2 The primary sources and what they instruct

The primary literature is the recent reckoning in panel-data econometrics with heterogeneous treatment effects. Callaway and Sant'Anna provide the canonical statement that under heterogeneity the standard two-way fixed-effects estimator returns a contaminated weighted average of group-time effects, and they construct estimators that target well-defined, interpretable building-block parameters which can then be aggregated transparently [\[27\]](#ref-27). De Chaisemartin and d'Haultfoeuille establish, in parallel, that two-way fixed-effects and difference-in-differences estimators with heterogeneous effects can place negative weights on some underlying comparisons, so that the headline coefficient can even take a sign opposite to every underlying effect [\[33\]](#ref-33). Goodman-Bacon's decomposition of the difference-in-differences estimator with variation in treatment timing makes the weighting explicit and shows which comparisons drive the pooled number [\[41\]](#ref-41). The synthesis literature surveying these developments consolidates the field's converged position that heterogeneity must be modeled, not assumed away, and that aggregation must be a deliberate, weighted choice rather than a by-product of the estimator [\[134\]](#ref-134). The pre-test literature warns that conditioning the analysis on a prior test can itself distort the reported estimates, a caution the dissertation carries into its decision to fix the archetype contrast in advance rather than select it after looking at the data [\[108\]](#ref-108).

The convergent instruction from this body is that the building blocks come first and the aggregate comes second, transparently weighted. Applied to the dissertation, this means the archetype-stratified subdistribution hazards are estimated as the primary objects, the archetype contrast is the parameter of interest, and any cohort-wide summary is an explicit weighting of the two strata rather than a pooled coefficient that silently averages them. What connects this econometric literature to a survival model is the shared statistical structure: a pooled hazard across heterogeneous strata is subject to the same averaging pathology as a pooled treatment effect across heterogeneous groups, because in both cases a single coefficient is asked to summarize comparisons that differ in magnitude and possibly in sign. The explicit weighting decompositions make the pathology visible in the difference-in-differences setting [\[41\]](#ref-41), and the negative-weights results show how severe it can become [\[33\]](#ref-33).

### 2.4.3 The doubly-robust motivation for the robustness design

A second contribution of this tradition to the dissertation is methodological insurance rather than framing. Sant'Anna and Zhao develop doubly-robust difference-in-differences estimators that combine an outcome model and a propensity or weighting model so that the estimator remains consistent if either one is correctly specified, even when the other is misspecified [\[118\]](#ref-118). The construct is general: pairing a regression model with a reweighting model buys robustness to the failure of either component. The dissertation adopts the logic, not the specific estimator. Because the archetype-specific hazards are estimated on a small observational cohort where any single modeling choice could be wrong, the design pairs a penalized-regression specification with a reweighting robustness check, so that the archetype-specific hazard conclusions do not rest on one functional form. The principle is the doubly-robust insight that combining an outcome model and a weighting model protects against misspecification of either; the transfer is that the same pairing disciplines the small-sample hazard estimation against over-reliance on a single regularization or specification choice.

The limit the literature attaches is that double robustness is not a license to ignore specification entirely; it protects against the failure of one component, not both. The dissertation accepts this and treats the penalized-regression-plus-reweighting pairing as a robustness check that strengthens confidence when the two agree and flags fragility when they diverge, not as a guarantee of correctness. Confidence that the heterogeneity discipline is the right framing is high, because the conditional structure of the hypothesis makes pooling actively misleading; confidence in any specific archetype-specific magnitude is again deferred to execution.


## 2.5 The optimism-bias rival as a theoretical contender

### 2.5.1 Why the rival is theoretical in character, not only a control variable

A complete theoretical framework must include the leading rival account, because the dissertation's separation of slip into instrument and launch channels is a causal claim that a competing explanation could undercut at the level of theory, not only at the level of a control variable. The leading rival is the optimism-bias and reference-class-forecasting account associated with Flyvbjerg. Its core proposition is that cost and schedule overruns are driven less by project-specific technical causes than by systematic optimism in the estimating process itself: planners produce baselines that are predictably too aggressive, and the overrun is the gap between an optimistic plan and an ordinary outcome rather than the footprint of any particular technical event [\[57\]](#ref-57). If this account is correct in its strong form, then decomposing slip into instrument and launch channels could be decomposing an artifact, because the apparent technical causes would be downstream of a single estimating pathology that produced the over-aggressive baseline in the first place.

This rival is theoretical because it offers a different generative model of the data entirely: a single estimating-process bias, rather than two competing physical hazards. Under the instrument-launch decomposition model, the data are generated by two physical processes (technology maturation and manifest dynamics) whose hazards differ by archetype. Under the strong optimism model, the data are generated by one estimating process whose bias varies with how aggressive the baseline was, and the "cause" of slip recorded in the narrative is a post-hoc rationalization of a gap that optimism created. The two models can fit the same slip record while implying opposite policy levers: the decomposition model says steer reserve to the dominant physical channel by archetype, while the optimism model says fix the estimating process and reference-class the baseline, and the archetype channel is a distraction.

### 2.5.2 The primary sources and the procedure they imply

The primary literature is specific about both the claim and the remedy. Flyvbjerg's synthesis of what is known about cost overrun states the empirical regularities, including the persistence and predictability of overruns across project types, in a form that frames overrun as a property of the estimating process [\[57\]](#ref-57). The procedures-for-optimism-bias guidance operationalizes the remedy: adjust planning-stage estimates using the distribution of outcomes from a reference class of comparable completed projects, so that the optimistic point estimate is replaced by an empirically grounded distribution [\[60\]](#ref-60). The remedy is the source of the control the dissertation builds. An estimating-optimism proxy can be constructed as the ratio of a mission's confirmation-baseline schedule to a reference-class median schedule for missions of similar class, which is the reference-class-forecasting logic turned into a measurable covariate.

The procedure the rival implies for the dissertation is therefore not to dismiss optimism but to control for it and to test whether the channel separation survives the control. If the instrument-launch separation disappears once the optimism proxy is included, the rival wins and the decomposition was an artifact of which missions were optimistically estimated. If the separation persists with the optimism proxy in the model, the rival is engaged and defeated on its own terms, and the channels stand as more than re-descriptions of estimating bias. The design pre-commits to running the model with the optimism control included and then removing it, reporting both, so that the dependence of the separation on the control is visible rather than buried.

### 2.5.3 The qualifier and the calibrated confidence

A necessary qualification is that optimism bias and physical-channel slip are not mutually exclusive; the honest position is that both can operate, and the empirical question is whether the channel distinction adds explanatory and decision content beyond the optimism account. The objection a Flyvbjerg-aligned critic would press is that even a surviving separation could reflect archetype-correlated optimism, where first-of-kind active-sensor missions are systematically more optimistically estimated than heritage missions, so that the archetype interaction captures differential optimism rather than differential physical hazard. The design's answer is that the optimism proxy is archetype-specific in its reference class and that the complexity control after the cost-estimating-relationship literature absorbs the most obvious confound, so that the archetype interaction is identified net of both general optimism and general complexity. This answer is partial, not complete, and the dissertation states so: the separation of differential optimism from differential physical hazard is bounded by the quality of the optimism proxy, and the residual risk is acknowledged rather than assumed away. Confidence that the optimism rival must be controlled is very high; confidence that the control fully purges archetype-correlated optimism is moderate and is flagged as a limitation the executed analysis must probe.


## 2.6 The integrated conceptual model

### 2.6.1 Assembling the three anchors into one identification logic

The three anchors are not three alternative lenses to be chosen among; they are three complementary pieces of a single identification logic, and the integrated model is their assembly. The competing-risks frame supplies the estimator and the decision-relevant quantity: the cumulative incidence function of each cause, estimated through the subdistribution hazard, with the cause-specific hazard reported alongside and the cumulative incidence functions compared across strata by a proper K-sample test [\[55\]](#ref-55), [\[115\]](#ref-115), [\[9\]](#ref-9), [\[79\]](#ref-79), [\[5\]](#ref-5). The Fogel lens supplies the interpretive discipline: the aggregate slip is decomposed into instrument and launch channels, the cumulative incidence function of each is read as a constructed counterfactual, and the decomposition is bounded by sensitivity analysis on the cause-coding and the at-risk structure [\[113\]](#ref-113), [\[82\]](#ref-82), [\[124\]](#ref-124). The Callaway and Sant'Anna lens supplies the heterogeneity discipline: sensor archetype enters as an explicit effect modifier, the archetype-specific hazards are the building blocks, no pooled hazard is reported as the headline, and the penalized-regression-plus-reweighting pairing is the doubly-robust-motivated robustness check [\[27\]](#ref-27), [\[33\]](#ref-33), [\[118\]](#ref-118). The optimism account supplies the rival the model must defeat: the channel separation is estimated with and without a reference-class optimism control, and a separation that depends on the control is conceded to the rival [\[57\]](#ref-57), [\[60\]](#ref-60).

The fit among the three is tight rather than coincidental. The competing-risks cumulative incidence function is the counterfactual-bearing quantity the Fogel decomposition needs, so the estimator and the decomposition share an object. The archetype stratification the Callaway and Sant'Anna discipline demands is the stratification across which the competing-risks K-sample test compares the cumulative incidence functions, so the heterogeneity discipline and the survival comparison share a partition. The optimism control is the covariate that keeps the decomposition and the stratification from being artifacts of the estimating process, so the rival's remedy becomes the model's control. The integration is the contribution at the level of theory: no prior work joins the survival estimator, the decomposition discipline, the heterogeneity discipline, and the optimism control into one logic that can test separability and archetype dependence at once.

### 2.6.2 The notation the model carries forward

The model is stated in the notation fixed across the dissertation so that the design and analysis chapters inherit it without restatement. For event type k (k = instrument, launch), the Fine-Gray proportional subdistribution hazard model is

\[ \text{subhazard}_k(t \mid X) = \text{subhazard}_{k0}(t) \exp(X \, \boldsymbol{\beta}_k), \qquad\qquad (1) \]

where the subdistribution hazard governs the cumulative incidence function \( \text{CIF}_k(t \mid X) \), the probability of a cause-k first slip by time t accounting for the competing event. The coefficient on the archetype-by-instrument-side interaction is the parameter of interest for the contribution. The cause-specific Cox model is estimated in parallel for each event for the instantaneous-rate interpretation, and a Gray's-type test is the nonparametric check on whether the cumulative incidence functions differ across archetype strata. Estimation uses a ridge-penalized partial likelihood, with the penalty selected by cross-validation and the number of free covariates capped by an events-per-variable rule, given the small cohort. This notation is reproduced verbatim from the shared specification and is not altered here.
### 2.6.3 The causal mechanism the empirical work will test

The integrated model implies a specific causal chain, stated as driver, mechanism, observable effect, operational consequence, and strategic implication, so that the empirical work tests a mechanism and not a bare correlation. The driver is that a mission carries a first-of-kind active sensor, a lidar or a radar, below its assumed maturity at the start of preliminary design. The mechanism is that the least-mature sensor technology fails environmental test, cannot close its calibration budget, or runs long in maturation, so that instrument development becomes the binding path on the schedule. The observable effect is an instrument-driven first-slip event, cause-coded from the cost-and-technical narrative reconciled against the independent assessment narrative, which raises the instrument-slip cumulative incidence faster than the launch-slip cumulative incidence in the active-sensor stratum. The operational consequence is that the mission's committed launch date moves, accruing the standing-army and rework cost that overruns the confirmation baseline, which closes the schedule-cost coupling established in Section 2.1. The strategic implication is that schedule and cost reserve at confirmation should be steered to instrument maturation and to technology gating at the preliminary-design decision point for this archetype.

The mechanism runs the opposite way for the heritage archetype, and naming both is what makes the model a competing-risks model rather than a single-channel model with an interaction bolted on. For a passive-radiometer continuity mission whose instrument is a direct rebuild of a flown unit, the instrument carries little maturation risk, so the binding path is rarely instrument development. The first slip, when it occurs, is disproportionately launch-side: a shared-vehicle anomaly, a manifest reshuffle, or a launch provider's own development slip, each exogenous to the spacecraft's readiness. The launch-slip cumulative incidence therefore rises at least as fast as the instrument-slip cumulative incidence in the heritage stratum, the dominance reverses or vanishes, and the strategic implication inverts: reserve for the heritage archetype should be held against manifest and provider dynamics rather than against instrument maturation. The two mechanisms are the two competing hazards, and the archetype is the effect modifier that determines which one binds.

Where the framework offers a causal claim, it names the mechanism; where the available design supports only an association, it says so. The dominance claim is causal in the sense that it rests on a named physical mechanism with a documented direction, the instrument-maturation channel for the active archetype and the manifest channel for the heritage archetype, supported by the technology-readiness-to-slip evidence on the instrument side and the continuity-mission launch-pressure record on the launch side [\[47\]](#ref-47), [\[17\]](#ref-17), [\[72\]](#ref-72), [\[77\]](#ref-77). The strength of the causal reading is bounded by the observational nature of the cohort: sensor archetype is not randomized to missions, so the archetype contrast compares naturally occurring groups, and the controls for complexity and estimating optimism are the only defense against confounding of archetype with other risk drivers. The framework states this plainly and downgrades the causal language accordingly: the model identifies an archetype-conditional dominance under stated controls, not an experimentally established causal effect, and the empirical chapters carry that calibrated modality forward.

### 2.6.4 How this framework carries the dissertation's argument

The conceptual model is built to carry the argument that runs across the whole dissertation, and stating that argument here fixes what each later chapter must defend. The problem is real: cost growth is largely a monetized image of slip, and slip has structurally different instrument and launch origins that the single-hazard treatment collapses, a reading grounded in the schedule-cost coupling and in the documented distinctness of the instrument and launch channels [\[86\]](#ref-86), [\[77\]](#ref-77), [\[22\]](#ref-22), [\[47\]](#ref-47). The problem is material: slip and the reserve allocated against it are first-order cost drivers, and technology-readiness deficits and instrument-schedule growth measurably move mission cost and schedule, so getting the channel wrong has a price [\[77\]](#ref-77), [\[47\]](#ref-47), [\[22\]](#ref-22). The design addresses the causal mechanism: a competing-risks model with archetype as an explicit effect modifier separates the two cause-specific hazards without contaminating one with the other, which is what the integrated framework is constructed to do [\[55\]](#ref-55), [\[9\]](#ref-9), [\[27\]](#ref-27). The design improves on the alternatives: a single pooled slip regression cannot decide separability, and a naive Kaplan-Meier that treats a competing event as ordinary censoring is biased upward and cannot estimate the channel-specific cumulative incidence, while the subdistribution model can [\[5\]](#ref-5), [\[79\]](#ref-79). The residual risk is acceptable and stated: results are conditional, design-stage, and counterfactual in the cumulative incidence sense; cause-coding error is bounded by two-source reconciliation and recoding sensitivity in the Fogel discipline; small-sample power is committed to be reported in advance; and the estimating-optimism rival is controlled rather than ignored [\[82\]](#ref-82), [\[57\]](#ref-57), [\[60\]](#ref-60), [\[79\]](#ref-79). The later chapters develop each of these points in turn; this chapter's role is to show that the conceptual model can bear them.

This dissertation concerns a survival-analytic, policy-econometric contribution and not a real system, capability, or data-and-service exchange, so the architecture-traceability layer is deliberately out of scope. The decision relevance is carried in prose rather than in a capability-to-function chain: the objective is correct reserve allocation at confirmation, and the decision the model informs is the archetype-specific reserve posture and the technology-gating threshold at the preliminary-design decision point. No enterprise-architecture vocabulary is forced onto the econometric contribution, and the discussion chapter rather than an architecture table is where the objective-to-decision mapping is developed in full.


## 2.7 Summary

The framework this chapter has built is the joint of three settled literatures, assembled for a problem none of them was written for. From the competing-risks tradition it takes the subdistribution hazard and its cumulative incidence function as the decision-relevant, predictive quantity, the cause-specific hazard as the complementary etiologic quantity reported alongside, and the K-sample comparison of cumulative incidence functions as the test of separability, all disciplined by the known upward bias of treating a competing event as ordinary censoring [\[55\]](#ref-55), [\[115\]](#ref-115), [\[9\]](#ref-9), [\[79\]](#ref-79), [\[5\]](#ref-5). From the Fogel tradition it takes the instruction to decompose the slip aggregate into channels and to read each channel's cumulative incidence function as a constructed counterfactual bounded by explicit sensitivity analysis [\[113\]](#ref-113), [\[82\]](#ref-82), [\[124\]](#ref-124). From the Callaway and Sant'Anna tradition it takes the refusal to pool a heterogeneous hazard, the archetype-as-effect-modifier structure, and the doubly-robust-motivated pairing of penalized regression with reweighting [\[27\]](#ref-27), [\[33\]](#ref-33), [\[118\]](#ref-118). And from the optimism-bias account it takes the rival the model must survive and the reference-class control that lets it survive on the rival's own terms [\[57\]](#ref-57), [\[60\]](#ref-60). The integrated model states a named, archetype-conditional causal mechanism with calibrated, design-stage confidence: instrument-driven and launch-driven slip are separable competing risks whose dominance is conditioned by sensor archetype, a claim the empirical chapters will test under the fixed notation and the pre-committed falsification rule. No estimated quantity is reported here, and none is implied; the chapter delivers the conceptual model and its evidential spine, and leaves the numbers to an execution the design honestly defers.


## Chapter 2 References

Citations in this chapter are numbered to the consolidated reference list in the Back Matter (Part I: References); each in-text marker links directly to its full entry there.


# Chapter 3: Literature Review

## 3.0 The chapter thesis

The literature that bears on NASA Earth-mission schedule slip is broad, but it is not joined. This chapter argues a single claim and develops it across nine thematic sections: the three research traditions that touch the cost-overrun-through-slip problem each supply one necessary piece of an integrated identification strategy, yet no published study assembles all three, and that absence of integration, not the absence of any single piece, is the gap this dissertation fills. The spacecraft cost-and-schedule literature establishes that slip is real, costly, and partly traceable to technology immaturity, but it models slip as one continuous outcome and never separates its causes into competing events. The competing-risks survival literature supplies the estimator that would separate two cause-specific hazards without letting one contaminate the other, but it lives in biostatistics and has never been turned on a mission cohort. The project-overrun-economics literature supplies the rival explanation, estimating optimism, that any credible cause-separation must survive, but it offers no apparatus for decomposing slip by physical origin. The chapter funnels from the broad schedule-cost coupling down through the instrument-side drivers, the contracting and policy effects, and the launch-availability side, then crosses into the competing-risks toolkit and the overrun-economics rival, and closes by stating the synthesized gap and the propositions that follow from it.

The problem the chapter addresses can be framed in four moves. The current state of the literature is a body of work that treats schedule slip as a single hazard and regresses it on covariates, so that a program office reading the literature learns that slip is costly and that low technology readiness predicts it, but learns nothing about whether the slip it should reserve against is instrument-driven or launch-driven. The desired state is a literature that separates the two slip origins as distinct competing risks and tells a program office which dominant hazard a given mission archetype faces. The gap is that the three relevant literatures have never been integrated into one strategy that tests separability and archetype dependence at once. The consequence of leaving the gap open is that reserve keeps being allocated against an undifferentiated slip pool, and a first-of-kind active-sensor mission under-reserved against instrument maturation, or a heritage continuity mission under-reserved against manifest risk, overruns or descopes while the program office cannot say which lever it should have pulled.

Throughout, every cited source is interpreted rather than catalogued. For each work the chapter states what it found, by what method, with what limitation, and what its convergence (or divergence) with neighboring work means for the dissertation's argument. Causal claims name a mechanism; where the evidence supports only correlation, the chapter says so and downgrades its confidence accordingly. The chapter advances the dissertation's argument as it goes: the problem is real (Sections 3.1 through 3.6), the apparatus to address its causal structure exists but is unused on this problem (Section 3.7), the dominant rival explanation is identified and must be controlled (Section 3.8), and the residual gap is stated as a set of testable propositions (Section 3.9).


## 3.1 The schedule-cost coupling in space systems

The foundational empirical regularity for this dissertation is that cost growth in complex space and defense projects is, to first order, a monetized image of schedule slip. The clearest statement of this coupling comes from Lieber and Donor, who examined the relationship between schedule delays and cost overruns on complex NASA and defense projects through a combined approach of literature review, practitioner survey, and direct program experience [\[86\]](#ref-86). Their central finding is that practitioners broadly accept that cost overruns are "directly related to" schedule delays, but that the phrase conceals a heterogeneous set of root causes operating simultaneously: unrealistic estimates, supply-chain difficulties, insufficient schedule margin, technical problems, scope changes, and discrete risk events, alongside less tangible drivers such as project complexity, over-optimism, political pressure, and poor communication. Their method is qualitative and taxonomic rather than inferential; they catalogue and categorize the factors rather than estimate their relative weights on a cohort. That is the limitation this dissertation responds to. Lieber and Donor establish that slip has multiple structurally distinct causes and that the aggregate phrase "directly related to" hides them, but they stop at the catalogue. They do not build an estimator that separates the causes into competing events and measures which dominates for which kind of mission. The convergence of their catalogue with the rest of the cost-and-schedule literature is what licenses the dissertation's core move: if everyone agrees slip is multi-causal, then modeling it as a single hazard is a modeling choice of convenience, not a reflection of the underlying physics, and a competing-risks reframing is overdue.

The mechanism behind the coupling is concrete and worth naming, because the dissertation's entire decision-relevance rests on it. A mission that takes longer than its committed schedule continues to pay its standing army of engineers, technicians, and managers for the extra months; it incurs contractor fee adjustments and rework when delayed subsystems force re-test of already-completed work; and it absorbs the carrying cost of facilities and integration infrastructure held open past plan. The driver is elapsed time beyond baseline; the mechanism is the conversion of that time into recurring labor and overhead; the observable effect is committed-cost growth; the operational consequence is a breach of the confirmation cost baseline; the strategic implication is that controlling slip is the most direct way to control cost. This chain is supported across the literature reviewed below and is treated here as high-confidence, because it is corroborated by independent NASA cost-community studies (Sections 3.4 and 3.5) using different datasets and different methods. The dissertation does not need to re-establish the coupling; it takes the coupling as settled and asks the next question, which is what drives the slip.

A broad empirical anchor for the scale of the overrun problem outside the space domain is provided by the project-overrun literature synthesized under the heading of how common cost overrun and cost growth actually are [\[70\]](#ref-70). That synthesis documents that cost overrun is pervasive across project types and eras and that its distribution is not well-behaved; overruns are not symmetric errors around a correct estimate but are systematically biased upward and heavy-tailed. The relevance to the space domain is that NASA Earth missions are not an exception to a general pattern of well-estimated projects; they sit inside a population of large technical projects that overrun routinely and asymmetrically. This matters for the dissertation's framing because it means the question is not whether Earth missions slip, which is settled, but what the dominant first cause of their slip is, which is open. The limitation of importing the general-project evidence is external validity: construction and transport megaprojects differ from spaceflight missions in their technology content and their procurement structure, so the general distribution motivates the question without answering it for the space case.


## 3.2 Technology readiness and schedule slip: the instrument side

The most direct antecedent to the instrument arm of the dissertation's competing-risks structure is the body of work linking technology readiness level to schedule slippage. Dubos, Saleh, and Braun provide the canonical quantitative treatment, modeling schedule slippage as a function of the technology readiness level of a spacecraft's least-mature technology at the time of program authorization [\[47\]](#ref-47), with the conference antecedent developing the data analysis and modeling in fuller detail [\[46\]](#ref-46). Their method is a cross-mission statistical model relating starting TRL to subsequent realized schedule slip. Their central finding is twofold and both halves matter for this dissertation. First, a deficit in the entry TRL of the least-mature technology maps to a measurable increase in expected schedule slip, so technology immaturity is a quantitatively estimable driver of delay, with a measurable and nonlinear effect that goes well beyond a qualitative risk flag. Second, and more consequential, the relationship is nonlinear at the low-maturity end: the schedule penalty for starting a development at a very low TRL grows disproportionately as maturity falls, so the marginal cost of immaturity is highest precisely where missions carrying first-of-kind technologies live. This is the empirical foundation for treating entry TRL of the least-mature sensor as the leading instrument-side covariate and for expecting it to load most heavily on first-of-kind active-sensor missions.

The mechanism Dubos and colleagues' result implies is the instrument-arm causal chain the dissertation will test. A mission carries a sensor technology below its assumed maturity at the start of development; that technology must be matured during the development itself rather than before it; maturation activities such as detector qualification, calibration closure, and environmental test reveal problems that were latent at authorization; the resolution of those problems consumes calendar time that was not in the baseline; the schedule slips; and the slip is, in origin, instrument-driven. The limitation of the Dubos work for the dissertation's purpose is that it models slip as a single continuous outcome regressed on TRL. It does not distinguish slip that originates in the instrument from slip that originates elsewhere, and so it cannot say whether a given mission's realized slip was actually caused by the immature technology or merely co-occurred with it. The dissertation's competing-risks reframing is the direct response: it keeps the TRL covariate that Dubos validated but assigns it to a cause-specific instrument hazard rather than to a pooled slip outcome, so that the covariate's effect is estimated on the event it is theorized to drive.
Two further sources sharpen the TRL-schedule relationship and extend its credibility. Behdinan and Mishra develop a weighted technology readiness level framework that uses cardinal, multi-criteria weighting to ascertain the maturity, schedule, and trend of a set of NASA technologies, motivated by the observation that the Government Accountability Office has repeatedly reported cost overruns and schedule slippages tied to subjective and inconsistent maturity assessment [\[18\]](#ref-18). Their contribution is methodological. They show that a flat, ordinal TRL scale loses information because it treats the gap between adjacent levels as constant, when in fact the difficulty of advancing across levels is uneven, and they propose a weighting that recovers that unevenness. For this dissertation, the entry-TRL covariate must therefore be read as an ordinal proxy for an underlying continuous and nonlinear maturity construct. That reading reinforces the Dubos finding of nonlinearity and counsels caution in treating TRL deficits as additive. The limitation is that the Behdinan framework is a maturity-assessment method, not a slip-causation study; it improves the measurement of the covariate without separating slip causes. A complementary roadmapping methodology for future systems situates TRL-driven planning inside a structured technology-maturation process [\[130\]](#ref-130), and a flight-spares methodology for robotic life extension documents the heritage-build side of the same maturity question, showing that cost-effective spares strategies depend on the maturity and heritage of the design being spared [\[36\]](#ref-36). Together these establish that maturity is a structured, plannable, but uneven quantity, which is the construct the dissertation's archetype variable is built to capture at its extremes.

The official NASA position on technology readiness assessment is captured in the final report of the NASA Technology Readiness Assessment study team, which reviewed the agency's TRA and TRL processes, compared them with other government and international agencies, and recommended mitigations for the ambiguities the team found [\[69\]](#ref-69). The report's value here is that it documents, from inside the agency, that TRL assessment at the relevant milestones is detailed but ambiguity-prone, and that the GAO had an active interest in producing a TRA best-practices guide. This is institutional corroboration that entry TRL is both the right covariate to capture instrument maturity and a covariate measured with non-trivial error, which the dissertation's measurement and threats-to-validity discussion must carry forward. The limitation is that the study team's report is a process-improvement document, not an empirical slip study, so it informs the construct and its measurement error without contributing a finding about slip causation.

**Table 3.1. The TRL-to-schedule-slip evidence cluster.**

| Source | Method | Core finding | Limitation for this study |
|--------|--------|--------------|---------------------------|
| Dubos, Saleh & Braun [\[47\]](#ref-47), [\[46\]](#ref-46) | Cross-mission statistical model of slip on entry TRL | Entry-TRL deficit raises expected slip; relationship nonlinear at low maturity | Models slip as a single pooled outcome; does not separate slip causes |
| Behdinan & Mishra [\[18\]](#ref-18) | Multi-criteria weighted TRL framework | Adjacent TRL gaps are unequal; ordinal TRL loses information | Maturity-assessment method, not a slip-causation study |
| Tech-roadmapping methodology [\[130\]](#ref-130) | Roadmapping for future systems | TRL-driven maturation is a structured, plannable process | Process methodology, not an empirical slip estimate |
| Flight-spares methodology [\[36\]](#ref-36) | Cost-effective spares for life extension | Heritage and maturity govern spares strategy | Heritage-side illustration, not a competing-risks design |
| NASA TRA study team [\[69\]](#ref-69) | Agency process review | TRL assessment is detailed but ambiguity-prone | Process-improvement report, no slip finding |

Table 3.1 synthesizes to this: the instrument-side covariate is well-validated as a driver of slip in aggregate and is understood to be nonlinear and measured with error, but the entire cluster shares the limitation the dissertation exists to remove. None of it separates instrument-origin slip from other-origin slip. Confidence in the instrument-arm covariate is therefore high. Confidence that the existing literature has isolated the instrument hazard is, by construction, nil, because no source attempts the isolation.


## 3.3 Complexity and cost-estimating relationships

A precondition for the dissertation's archetype claim is that "first-of-kind active sensor" captures something distinct from general mission complexity rather than relabeling it. If sensor novelty and general complexity are the same thing, then any apparent instrument-slip dominance for novel-sensor missions would be an artifact of complexity, and the archetype contrast would carry no independent information. The complexity-based cost-estimating literature is what allows the dissertation to separate the two, so it must be reviewed not as background but as the source of an essential control.

Bearden's complexity-based cost-estimating relationships are the anchor [\[17\]](#ref-17). His method builds a complexity index for a space system by hypothesizing the technical elements that determine subsystem complexity, ranking each element against a historical database of comparable subsystems, and combining the ranks into an overall complexity measure that feeds a cost-estimating relationship. The central finding is that mission complexity and design aggressiveness independently raise both cost and schedule risk, and that complexity is estimable parametrically from technical parameters available early in design. For this dissertation the result is useful in two ways. It supplies a constructed complexity index that can enter the survival model as a control, so that the archetype effect is estimated net of general complexity; and by demonstrating that complexity is separately measurable, it establishes that complexity and sensor novelty are distinct constructs that can in principle be disentangled. The limitation is that Bearden's relationships are built to predict cost and schedule as outcomes, not to separate slip causes, and the complexity index is itself a compression of many technical parameters that may correlate with novelty. The control reduces but cannot fully eliminate the risk that archetype proxies for residual complexity, which the dissertation acknowledges as a construct-validity threat.

The instrument-cost-modeling tradition reinforces that instrument development is a distinct and parametrically modelable locus of risk. The JPL Advanced Projects Design Team's spacecraft instrument cost model develops an objective, multivariate approach to estimating instrument cost from physical and technical parameters [\[76\]](#ref-76), establishing early that instruments are estimated separately from the bus and that their cost is a function of mass, power, and technical characteristics. The literature already treats the instrument as a separable cost-and-schedule object with its own drivers, which is the conceptual permission the dissertation needs to treat instrument-origin slip as a separable event. The limitation of the instrument-cost-model tradition for the present purpose is that it estimates instrument cost, not the timing or cause of instrument-origin slip. It tells the dissertation which parameters describe an instrument's difficulty but not when or whether that difficulty becomes a schedule event. The dissertation uses these parameters as instrument-side covariates while supplying, through the competing-risks estimator, the timing-and-cause structure the cost models lack.


## 3.4 Instrument-development cost-time trends and instrument-schedule growth

The proposition that instrument development is specifically where Earth missions slip, as opposed to the bus or the ground system, is supported by a focused NASA cost-community literature that the dissertation treats as the strongest available direct evidence for its instrument arm. Kipp, Ringler, Chapman, and Freaner examine the impact of instrument schedule growth on mission cost and schedule growth for recent NASA missions [\[77\]](#ref-77). Their method relates instrument-level schedule growth to mission-level cost and schedule outcomes across a set of recent missions. Their finding, the load-bearing one for this dissertation, is that growth in the instrument's schedule propagates to growth in the mission's cost and schedule: the instrument is the primary driver, whose slip transmits to the whole mission. The mechanism is integration coupling. The instrument sits on the mission's critical path because the spacecraft cannot complete integration and test until the instrument is delivered and verified, so an instrument that runs long forces the whole integrated system to wait, accruing standing-army cost across the entire project rather than only on the instrument team. This is the clearest empirical statement in the corpus that an instrument-origin delay is a mission-level event, which is the claim the dissertation's instrument-driven competing event encodes. The limitation is that Kipp and colleagues measure instrument schedule growth as a continuous quantity correlated with mission outcomes. They do not frame the instrument as a competing risk against a launch-side alternative, and so they cannot say, for a given mission, whether the instrument or the launch was the first cause of slip. The dissertation's contribution is to take Kipp's validated instrument-to-mission propagation and embed it in a structure that lets it compete with the launch channel for first-cause primacy.

The cost-time-trend tradition provides the longitudinal backdrop. The NASA Instrument Cost Model and its Explorer-class extension estimate instrument cost from mass, power, data rate, and technology parameters, and the long-run instrument cost-time-trend studies document that instrument development is a persistent locus of cost and schedule risk distinct from the bus and from launch. These named sources are described in the data chapter as the origin of the instrument-parameter covariates and the active-versus-passive taxonomy. Their relevance here is that they establish, over decades and many instruments, that instrument development has its own cost-time signature, which is the empirical basis for treating it as a separable channel rather than folding it into a bus-and-instrument aggregate. The limitation, common to the cost-model tradition, is that the trends describe cost as a function of instrument parameters and do not isolate the cause or timing of slip; they justify the instrument channel's existence without estimating its hazard.

Supporting cost-estimating studies extend the instrument-side evidence into specific subsystems and contracting contexts. Hahn and Sholder develop a cost model for electronic, electrical, and electromechanical parts in NASA space missions and document, through multiple linear regression on actual mission and instrument costs, that parts costs grow as a design is refined and finalized, driven by parameters such as parts class, radiation environment, board area, and requirement complexity, and they connect this growth to the need for appropriate reserve allocation [\[67\]](#ref-67). Even at the component level, immaturity and complexity translate into cost growth that the program must reserve against, reinforcing the instrument-arm mechanism at a finer grain. Nilsen examines the effect of implementing data-driven uncertainty, through subsystem correlation matrices derived from cost-estimating-relationship residuals, on NASA cost models, and finds that correlation assumptions materially shift the probabilistic cost output and that residual-derived correlations reduce bias and error relative to blanket assumptions [\[98\]](#ref-98). The instrument channel's cost risk is correlated across subsystems in ways that naive models miss, which corroborates the dissertation's choice to model the instrument hazard with covariates rather than as an independent draw. Bart, Duda, and Hoffman estimate the cost to transition a space system from expendable to reusable and, in doing so, formalize technology-readiness-level growth factors and reliability-growth factors as multipliers on development cost [\[16\]](#ref-16), which independently confirms that TRL growth is treated by the cost community as a quantitative cost-and-schedule driver, consistent with the Dubos finding from the schedule side.

Section 3.4 synthesizes to this: the instrument channel is the best-evidenced of the two competing events. Multiple NASA-community studies, using different datasets and methods, converge on the conclusion that instrument schedule growth propagates to mission cost and schedule growth and that its drivers (TRL, complexity, parts immaturity, correlated subsystem risk) are estimable from parameters available at confirmation. Confidence in the instrument arm is therefore moderate-to-high at the design stage. What would raise it to high is what no existing study provides and what the dissertation proposes: an estimate of the instrument hazard as a competing risk against the launch channel on a defined cohort. What would lower it is a finding, on execution, that instrument slip rarely arrives first because anticipatory descoping suppresses it, a possibility the dissertation flags as a rival explanation.


## 3.5 Contracting and policy effects on mission cost and schedule

Slip and its cost consequence are not purely technical; they are shaped by the contract vehicle and the acquisition policy regime under which a mission is bought. This literature matters to the dissertation in two ways. It supplies covariates and controls that keep the cause-separation from being confounded by procurement structure, and it documents that the empirical base for the dissertation's question is the NASA cost community's own historical-cost analyses.

Sobel and Tibor study the effectiveness of firm-fixed-price spacecraft contracts in curbing cost growth, comparing the historical cost growth of cost-plus and firm-fixed-price spacecraft across NASA science missions launched over roughly two decades from contract start to delivery [\[123\]](#ref-123). Their finding is qualified and directly relevant: firm-fixed-price contracts do not uniformly curb cost growth, and a documented cause of cost growth on firm-fixed-price contracts is the mistaken assumption of high heritage to a previous spacecraft, alongside schedule delays, requirement changes, and added scope. The interpretation here is sharp. The heritage assumption is the very construct that defines the passive-radiometer heritage archetype, and Sobel and Tibor show that overstated heritage is itself a cost-growth driver. This validates heritage as a meaningful and consequential mission attribute, and it warns that heritage is sometimes nominal rather than real, which the dissertation's archetype coding must treat carefully: a mission claiming heritage that does not hold may carry more instrument risk than its archetype label suggests. The limitation is that Sobel and Tibor analyze cost growth as a continuous outcome by contract type, not slip as a competing risk by cause, so their result informs the controls and the heritage construct without separating the hazards.

Sholder and Whitley extend the contracting analysis with an explicit reserve-posture framing on which the dissertation's policy implications depend [\[122\]](#ref-122). Examining historical costs from past NASA missions, they find a high probability of cost growth from preliminary design review to launch and quantify how far missions exceed their budgets at the empirical fiftieth and seventieth percentiles, comparing in-house spacecraft builds against contracted builds. The decisive finding here is that NASA's reserve postures are calibrated against empirical cost-growth distributions and that those distributions differ between in-house and contracted builds. This is the institutional setting into which the dissertation's archetype-specific reserve recommendation would land: NASA already steers reserve by percentile against an empirical distribution, so a refinement that steers reserve by archetype against cause-specific cumulative-incidence distributions is a coherent extension of existing practice rather than a foreign imposition. The limitation is that Sholder and Whitley pool cost growth across causes. Their percentiles describe how much missions overrun, not which cause arrives first, which is the differentiation the dissertation adds.

Bitten and colleagues analyze the effect of policy changes on NASA science mission cost and schedule growth, relating shifts in acquisition policy to realized cost and schedule outcomes across missions [\[22\]](#ref-22). The finding is that policy regime measurably moves cost and schedule growth, which is the empirical basis for the dissertation's calendar-period fixed effects: era-specific acquisition policy is a confounder that must be absorbed so that an archetype effect is not mistaken for a policy effect, and vice versa. Without era controls, a cohort weighted toward a permissive or a stringent policy era could show a spurious cause-dominance pattern that reflects the era rather than the archetype. The limitation is that Bitten and colleagues model the aggregate effect of policy on pooled cost and schedule growth, so they establish that era matters without telling the dissertation how era interacts with cause; that interaction is left to the analysis. A probabilistic cost-schedule risk-analysis study using Monte Carlo simulation, drawn from the offshore-construction domain, reinforces the general point that deterministic plans systematically underestimate schedule risk and that probabilistic methods are needed to capture both aleatory and epistemic uncertainty [\[109\]](#ref-109). Its relevance is methodological and cross-domain rather than NASA-specific, and the dissertation cites it as corroboration that schedule risk is under-reserved when modeled deterministically, while noting that its offshore setting bounds its direct transfer.

Section 3.5 synthesizes to this: contracting and policy are first-order modifiers of the slip-to-cost coupling, and the NASA cost community has already built the empirical, percentile-based reserve machinery the dissertation seeks to refine. Confidence that contract vehicle and policy era must enter as controls is high. The limitation shared across this cluster is the same pooled-outcome limitation that runs through the whole domain literature: these studies measure how much and under what contract missions overrun, never which cause of slip arrives first.


## 3.6 The launch-availability side

The launch arm of the competing-risks structure is the thinnest in the open literature, and the dissertation is candid that this is the more evidence-constrained half of its design. There is no body of work that models launch-availability-driven schedule slip as a survival outcome with cause-specific hazards. What the literature provides instead is documentary evidence, primarily from the Landsat continuity program, that launch-side slip is real, structurally separate from instrument maturation, and driven by manifest and provider dynamics that are exogenous to the spacecraft's own readiness.

The Landsat Data Continuity Mission is the central case. Irons, Dwyer, and Barsi describe the next Landsat satellite and its development as a continuity mission carrying two Earth-viewing instruments, the Operational Land Imager and the Thermal Infrared Sensor, built to extend a multi-decadal land-observation record [\[72\]](#ref-72). The account documents a continuity mission whose scientific imperative is uninterrupted data and whose schedule is therefore acutely sensitive to launch timing. Jim Irons and colleagues' progress report toward the launch documents the mission tracking toward a specific launch date with the instruments in final integration and test [\[74\]](#ref-74), and Masek and colleagues' account of Landsat 9 describes a near-clone successor whose instruments are largely identical to its predecessor, the Operational Land Imager-2 being largely identical to the prior unit and the Thermal Infrared Sensor-2 improving on known issues [\[91\]](#ref-91). This cluster supplies the substantive basis for the launch arm. A continuity mission whose instrument is a direct or near-direct rebuild of a flown unit carries little instrument-maturation risk by design, so when such a mission slips, the slip is disproportionately likely to originate on the launch side, through a shared-vehicle anomaly, a manifest reshuffle, or a launch provider's own development slip, none of which the spacecraft controls. This is the heritage-archetype mechanism the dissertation tests: low instrument risk by construction shifts the dominant first-slip hazard toward the launch channel.
The exogeneity of much launch-side slip is what makes the competing-risks framing credible rather than circular, and the Landsat narrative supports it. Because a heritage continuity instrument is matured before development begins, its first slip, when it occurs, is structurally unlikely to be instrument-origin, which means the two archetypes face genuinely different mixtures of the two hazards. The limitation is severe and stated plainly: the Landsat sources are programmatic narratives, not quantitative slip studies, and they cannot be cited as DOI-bearing measurements of a launch hazard. The data chapter therefore leans on the primary GAO and CADRe records, which are restricted or non-DOI sources, for the launch-side cause coding, and the discussion frames the launch-side dominance claim for the heritage arm as the more evidence-thin half of the contribution, to be confirmed on the cohort rather than asserted from existing literature. Confidence in the launch arm is accordingly lower than in the instrument arm at the design stage. This asymmetry is acknowledged rather than hidden, and it is one reason the dissertation reports cause-specific hazards alongside subdistribution hazards, so a reviewer can judge the launch arm separately.

The broader Earth-observation-mission and launch-market context situates the launch arm in a changing environment. Miner and colleagues, surveying Earth-observation needs for a warming world ahead of the next decadal survey, document that the continuation of essential climate-variable time series is balanced against new measurement needs enabled by technological breakthroughs [\[93\]](#ref-93). Continuity missions and first-of-kind technology missions coexist in the Earth-science portfolio and face different risk structures, which is the dissertation's exact archetype distinction at the portfolio level. The governance-evolution literature on the transition from legacy-space to new-space delivery models documents that the launch and acquisition environment changed substantially over the studied era [\[142\]](#ref-142), the substantive justification for the dissertation's concern about era confounding of the launch market and for its calendar-period controls. The launch arm is conceptually well-motivated and documentarily supported but quantitatively under-served by the open literature, which is itself a finding: the absence of launch-side survival studies tells us where the literature has not looked, and it belongs in the statement of the gap.


## 3.7 The competing-risks methodological literature

The estimator the dissertation needs already exists, fully developed, in biostatistics and epidemiology. The competing-risks survival literature is mature, well-validated, and equipped with the machinery required to separate two cause-specific hazards and to report the cumulative incidence of each in the presence of the other. It has simply never been applied to NASA mission schedule slip. This section reviews that literature on its own terms, establishing what the apparatus does and why it is the right tool, so that Section 3.9 can state precisely what the unfilled gap is.

The foundational result is Fine and Gray's proportional-hazards model for the subdistribution of a competing risk [\[55\]](#ref-55). Their contribution is a semiparametric proportional-hazards model not for the cause-specific hazard but for the subdistribution hazard, which governs the cumulative incidence function, the marginal probability of a particular failure type over time. Their method uses the partial-likelihood principle with inverse-probability-of-censoring weighting to allow direct regression of covariates on the cumulative incidence of one competing event. The significance, which Fine and Gray themselves emphasize, is interpretive: the cause-specific hazard does not have a direct interpretation in terms of the survival probability for a particular failure type, whereas the cumulative incidence function does, and the latter is what a decision-maker needs when the question is the absolute probability of an event. For this dissertation that interpretive distinction is decisive. The policy question is the cumulative probability that a mission of a given archetype slips first for instrument reasons versus launch reasons, a cumulative-incidence question, so the subdistribution hazard is the primary estimator. The limitation the dissertation carries forward is that the subdistribution hazard's covariate effects are on the cumulative incidence and not on the instantaneous rate, so they answer the predictive and not the etiologic question, which requires the cause-specific model alongside.

The applied framing of the choice between the two hazard families is given by Austin, Lee, and Fine, whose introduction to survival analysis in the presence of competing risks is the clearest practical guide in the corpus [\[9\]](#ref-9). They establish two results the dissertation depends on. First, when estimating the crude incidence of outcomes, analysts must use the cumulative incidence function rather than the complement of the Kaplan-Meier survival function, because the Kaplan-Meier approach biases incidence estimates upward, regardless of whether the competing events are independent. This is the warning against treating a competing event as ordinary censoring, and it is the reason the dissertation cannot use a naive survival estimate: a naive Kaplan-Meier treatment of instrument slip that censors launch slip would overstate the instrument-slip probability. Second, the cause-specific hazard estimates the effect of covariates on the rate of the event among those still event-free (the etiologic question), while the subdistribution hazard estimates the effect on the absolute risk over time (the predictive question), and the analyst must choose based on which question is being asked. The dissertation's design follows this guidance exactly: subdistribution primary for the predictive policy question, cause-specific in parallel for the etiologic interpretation. Austin, Lee, and Fine carry no limitation of substance for the present purpose; the work is a guide, not a primary study, and its only constraint is that its examples are clinical, so the transfer to mission slip is the dissertation's own responsibility.

A refinement comes from the literature on the appropriate reporting of competing-risks analyses. Latouche and colleagues argue that a competing-risks analysis should report results on all cause-specific hazards and cumulative incidence functions, not a single chosen quantity, because the cause-specific and subdistribution views answer different questions and divergence between them is informative rather than a nuisance [\[79\]](#ref-79). For the dissertation this sets a reporting standard: it must present both the cause-specific Cox results and the Fine-Gray subdistribution results for both events, and treat any divergence between them as a substantive finding about the dependence structure of the two slip causes. Wolbers and colleagues provide a non-technical overview of competing-risks objectives and approaches for a clinical audience, stressing that competing-risks methods apply whenever competing events may preclude or alter the chance of observing the event of interest, and that the cumulative incidence function is the central descriptive tool [\[135\]](#ref-135). The dissertation's structure (a mission's first slip is either instrument-origin or launch-origin, and the occurrence of one precludes the clean observation of the other as "first") is the textbook competing-risks structure these authors describe, which is what licenses the borrowing.

The regression-modeling options are systematized by Lau, Cole, and Gange, who outline cause-specific and subdistribution regression approaches for epidemiologic data and contrast the structure of the risk sets and the interpretation of the parameters obtained from each [\[80\]](#ref-80). Their demonstration, on a cohort with competing events, that the cause-specific relative hazard and the subdistribution relative hazard can both be estimated and can differ in magnitude is the empirical illustration of why the dissertation estimates both. The limitation common to this methodological cluster is that it is developed and validated on clinical and epidemiologic data, where event coding is typically more objective (a death, a transplant, a diagnosis) than the narrative cause-attribution the dissertation must perform for mission slip. The transfer of the estimator is sound; the difficulty the dissertation inherits lies in the cause coding, not in the survival mathematics, a measurement problem the data and design chapters address rather than a flaw in the borrowed apparatus.

**Table 3.2. The competing-risks apparatus and its role in the design.**

| Source | Contribution | Role in this dissertation |
|--------|--------------|---------------------------|
| Fine & Gray [\[55\]](#ref-55) | Subdistribution-hazard model for the cumulative incidence of one competing risk | Primary estimator; CIF answers the predictive reserve-allocation question |
| Austin, Lee & Fine [\[9\]](#ref-9) | Cause-specific vs subdistribution choice; CIF not Kaplan-Meier | Licenses the dual estimation and rules out naive censoring |
| Latouche et al. [\[79\]](#ref-79) | Report all cause-specific hazards and CIFs | Reporting standard; divergence treated as a finding |
| Wolbers et al. [\[135\]](#ref-135) | Non-technical competing-risks overview; CIF as central tool | Confirms the slip structure is textbook competing-risks |
| Lau, Cole & Gange [\[80\]](#ref-80) | Cause-specific and subdistribution regression contrasted | Justifies estimating both relative hazards in parallel |

The synthesis of Table 3.2 and Section 3.7 is the central observation of the chapter. The apparatus is complete, validated, and interpretively well-understood, and it maps without strain onto the structure of mission first-slip. Confidence that the estimator is correct and available is very high, because it rests on decades of biostatistical development and replication. The single fact that makes the dissertation possible is that this complete apparatus has never been turned on NASA mission slip: no published study treats instrument-driven and launch-driven slip as competing risks and estimates their cause-specific and subdistribution hazards. The gap is therefore not in the tool and not in the domain evidence that slip is real and multi-causal; it is in the join.


## 3.8 Project-overrun economics and the optimism-bias rival

Any claim that mission slip separates into instrument and launch causes must survive a powerful rival explanation: that overruns are driven not by project-specific technical causes at all, but by systematic optimism in the planning-stage estimate. This rival is associated above all with Flyvbjerg, and the dissertation treats it not as background but as the explanation it must control to make its cause-separation credible.

Flyvbjerg's account of cost overrun holds that the root cause is behavioral bias rather than scope changes or complexity, that overrun is consistently fat-tailed rather than a symmetric estimating error, and that the remedy is reference-class forecasting that de-biases estimates by anchoring them to the actual outcomes of a class of comparable past projects [\[57\]](#ref-57). The empirical foundation is the demand-forecast-inaccuracy work of Flyvbjerg, Skamris Holm, and Buhl, whose study of traffic forecasts across 210 transport projects worth tens of billions of dollars in fourteen nations established with high statistical significance that forecasters systematically and persistently overestimate demand, that the inaccuracy did not improve over a thirty-year period, and that the resulting financial risks are large and routinely downplayed [\[58\]](#ref-58). For this dissertation the implication is that estimating optimism is a real, large, and durable phenomenon, and that if missions are optimistically estimated in a way correlated with their archetype, the apparent instrument-launch separation could be an artifact of which missions were optimistically baselined rather than of any difference in physical slip cause. The mechanism of the rival is concrete: an optimistic baseline schedule is short relative to the work required, so the mission slips against it regardless of cause, and if first-of-kind active-sensor missions are systematically more optimistically baselined than heritage missions, the instrument-slip dominance could reflect the baseline, not the instrument. The dissertation's defense, drawn directly from this literature, is the reference-class optimism proxy, constructed as the ratio of the confirmation-baseline schedule to a reference-class median schedule for similar-class missions, included as a control and removed in a sensitivity analysis to test whether the separation depends on it.

The reference-class-forecasting literature has matured into a substantial applied body that both strengthens the rival and equips the dissertation's control. Flyvbjerg, Hon, and Fok report the first application of reference-class forecasting in the Asia-Pacific region, benchmarking twenty-five Hong Kong roadwork projects against a sample of 863 similar projects and verifying the forecast-accuracy distribution at various development stages [\[61\]](#ref-61); Flyvbjerg, Glenting, and Ronnest produced the UK guidance document providing empirically based optimism-bias uplifts for reference classes of transport projects [\[60\]](#ref-60); and Flyvbjerg's earlier treatment details reference-class forecasting in planning practice as the operational remedy for optimism bias and strategic misrepresentation [\[56\]](#ref-56). The convergent finding across these is that reference-class forecasting measurably improves accuracy and that the size of the optimism uplift depends on how the reference class is defined. For the dissertation, this means its optimism proxy is methodologically grounded but sensitive to reference-class definition, a known limitation of the method that the dissertation inherits and must report. Park's before-and-after and with-and-without study of 107 major projects strengthens the empirical case, finding that average cost overrun declined sharply following the introduction of reference-class forecasting in the UK while the US, which did not adopt the practice, underperformed [\[103\]](#ref-103), which establishes that the optimism mechanism is causally consequential and remediable, a finding that reaches beyond correlation into demonstrated policy effect.

The broader optimism-bias-and-cost-overrun literature corroborates and qualifies the rival. Chen, Ahiaga-Dagbui, and colleagues provide a review of optimism bias and transport cost overruns and find a growing face-value acceptance of optimism bias as the primary cause, but identify gaps and unanswered questions about the relationship between optimism bias in cost appraisal and actual overruns, noting that the presence and nature of optimism bias in the complex institutional environment of cost appraisal are understudied [\[34\]](#ref-34). The dissertation takes this seriously: the optimism rival is widely accepted but not airtight, and treating it as the sole or proven cause would itself be an error, which is why the dissertation controls for optimism rather than assuming it away or assuming it dominant. Chadee, Hernandez, and Martin measure optimism bias directly among project participants and find moderate levels that influence time and cost overruns [\[32\]](#ref-32), and Chadee and colleagues' later work develops a simplified reference-class forecast for small island developing states, demonstrating practical reference-class construction on a sample of public housing projects [\[31\]](#ref-31); Awojobi and Jenkins apply reference-class forecasting to hydroelectric-dam cost-overrun risk [\[12\]](#ref-12), Ansar, Flyvbjerg, and colleagues find overwhelming evidence that large-dam budgets are systematically biased below actual costs using reference-class forecasting [\[6\]](#ref-6), and Baerenbold's review of forty-one reference-class-forecasting papers concludes that the method's effectiveness depends on the definition of the reference class and calls for an empirically based framework for reference-class formation [\[13\]](#ref-13). The cross-domain convergence (transport, dams, housing) raises confidence that the optimism mechanism is general; the persistent caveat about reference-class definition is the limitation the dissertation carries into its own proxy construction.

Two further strands round out the rival and its boundaries. Lovallo, Cristofaro, and Flyvbjerg's governance treatment organizes the remedies for behavioral bias into a forecasting-organizing-executing process and reviews the accuracy of reference-class forecasting alongside principal-agent solutions [\[88\]](#ref-88), which reads as evidence that the optimism rival has a governance as well as a cognitive dimension, relevant to the dissertation's discussion of anticipatory descoping as a reverse-causation threat. Natarajan combines reference-class forecasting with machine learning for offshore oil-and-gas megaproject planning, demonstrating that overrun distributions are non-normal and fat-tailed and that project-specific factors can be combined with reference-class uplifts in a predictive model [\[97\]](#ref-97), and Ahiaga-Dagbui and Smith use data mining on 1,600 completed projects to forecast final cost [\[1\]](#ref-1), and Brown, Lux, and Cowan apply reference-class forecasting to fusion-power-plant cost estimates in a setting where little data exist due to technological novelty [\[24\]](#ref-24). This last cluster shows that the most advanced overrun-economics work is moving toward project-specific predictive modeling that combines reference-class anchoring with covariates, methodologically adjacent to the dissertation's own covariate-driven hazard model; the difference is that none of it decomposes the outcome into cause-specific competing events, so the overrun-economics literature supplies the optimism control and the fat-tailed-distribution caution but not the cause decomposition.

**Table 3.3. The optimism-bias rival and the dissertation's response.**

| Strand | Representative sources | Finding the dissertation must address | Dissertation's response |
|--------|------------------------|----------------------------------------|--------------------------|
| Core optimism thesis | Flyvbjerg [\[57\]](#ref-57); Flyvbjerg, Skamris Holm & Buhl [\[58\]](#ref-58) | Overrun is behaviorally driven and fat-tailed, not technically caused | Optimism proxy as control; sensitivity removal |
| Reference-class forecasting practice | Flyvbjerg, Hon & Fok [\[61\]](#ref-61); Flyvbjerg et al. [\[60\]](#ref-60), [\[56\]](#ref-56); Park [\[103\]](#ref-103) | RCF improves accuracy; effect depends on class definition | Proxy is RCF-grounded; class-definition sensitivity reported |
| Optimism-bias reviews and measurement | Chen et al. [\[34\]](#ref-34); Chadee et al. [\[32\]](#ref-32), [\[31\]](#ref-31) | Optimism widely accepted but understudied and not airtight | Control, do not assume optimism proven or sole cause |
| Cross-domain RCF applications | Awojobi & Jenkins [\[12\]](#ref-12); Ansar et al. [\[6\]](#ref-6); Baerenbold [\[13\]](#ref-13) | Mechanism general across domains; class definition is the lever | General mechanism motivates the control; caveat carried |
| Predictive RCF-plus-covariates | Natarajan [\[97\]](#ref-97); Ahiaga-Dagbui & Smith [\[1\]](#ref-1); Brown et al. [\[24\]](#ref-24); Lovallo et al. [\[88\]](#ref-88) | Advanced work combines RCF with covariates but not cause decomposition | Adjacent method; dissertation adds the missing decomposition |

The synthesis of Table 3.3 is that the optimism rival is strong, general, and well-evidenced, and that the dissertation's credibility hinges on controlling it. Confidence that optimism is a real and material driver of overrun is high. Confidence that optimism, rather than physical slip cause, explains the instrument-launch separation is something only execution can settle, which is why the dissertation pre-commits to removing the optimism control and reporting whether the separation survives; if the separation depends on the control, the rival wins, and the dissertation says so in advance.


## 3.9 The synthesized gap and the propositions that follow

The three literatures reviewed above are each necessary and individually insufficient, and their non-integration is the gap. This closing section states the gap precisely and derives the propositions the empirical chapters test.

The spacecraft cost-and-schedule literature (Sections 3.1 through 3.6) establishes, at high confidence, that cost growth is largely a monetized image of schedule slip [\[86\]](#ref-86), [\[70\]](#ref-70); that slip is multi-causal with structurally distinct origins [\[86\]](#ref-86); that low entry technology readiness is a measurable and nonlinear driver of slip [\[47\]](#ref-47), [\[46\]](#ref-46), [\[18\]](#ref-18); that instrument schedule growth propagates to mission cost and schedule growth through critical-path coupling [\[77\]](#ref-77), [\[67\]](#ref-67), [\[98\]](#ref-98); that complexity is separately measurable so that novelty and complexity are demonstrably distinct constructs [\[17\]](#ref-17), [\[76\]](#ref-76); that contract vehicle and policy era materially move outcomes and must be controlled [\[123\]](#ref-123), [\[122\]](#ref-122), [\[22\]](#ref-22); and that the launch side is a real, documented, but quantitatively under-served slip origin driven by manifest and provider dynamics exogenous to the spacecraft [\[72\]](#ref-72), [\[74\]](#ref-74), [\[91\]](#ref-91), [\[93\]](#ref-93), [\[142\]](#ref-142). The decisive shared limitation of this entire literature is that it models slip as a single continuous or pooled outcome and never separates its causes into competing events. It tells a program office that slip is costly and partly TRL-driven; it does not tell the office whether the slip it should reserve against is instrument-origin or launch-origin for a given mission.
The competing-risks literature (Section 3.7) supplies, at very high confidence, the exact apparatus the domain literature lacks: a validated subdistribution-hazard model for the cumulative incidence of one competing event [\[55\]](#ref-55), a clear rule that the cumulative incidence function, not Kaplan-Meier, is the decision-relevant predictive quantity and that treating a competing event as ordinary censoring biases incidence upward [\[9\]](#ref-9), a reporting standard that both cause-specific and subdistribution results be presented because their divergence is informative [\[79\]](#ref-79), [\[135\]](#ref-135), and regression machinery for both hazard families [\[80\]](#ref-80). The limitation of this literature for the present problem is only that it has never been applied to mission slip. The mathematics transfers without strain, but the cause-coding it presupposes is harder for narrative mission records than for clinical endpoints, a measurement burden the dissertation accepts rather than a flaw in the tool.

The project-overrun-economics literature (Section 3.8) supplies, at high confidence, the rival explanation that the cause-separation must survive: that overrun is behaviorally driven, fat-tailed, and remediable by reference-class forecasting [\[57\]](#ref-57), [\[58\]](#ref-58), [\[61\]](#ref-61), [\[60\]](#ref-60), [\[56\]](#ref-56), [\[103\]](#ref-103), with cross-domain corroboration [\[12\]](#ref-12), [\[6\]](#ref-6), [\[13\]](#ref-13) and a caveat that optimism is accepted but not airtight [\[34\]](#ref-34). Its limitation is that it offers no apparatus for decomposing slip by physical origin. It explains why baselines are short, not which cause makes a mission slip first.

The gap, stated exactly, is the absence of a single identification strategy that (a) keeps the validated instrument-side covariates of the spacecraft literature, (b) embeds them in the competing-risks estimator so instrument-origin and launch-origin slip are separated rather than pooled, and (c) controls for the optimism rival so the separation is not an artifact of the estimating process, while testing both whether the two causes are separable and whether their relative dominance depends on sensor archetype. No prior study combines all three traditions. The gap is the join, not any one element, which is why the contribution is an integration rather than the discovery of a new fact.

From this synthesized gap, three propositions follow, mapping directly onto the dissertation's fixed hypotheses. The first is separability: instrument-driven and launch-driven schedule slip are statistically distinct competing risks, distinguishable by a test of whether their cause-specific hazards differ and whether their cumulative incidence functions differ across strata. This rests on the multi-causal slip evidence [\[86\]](#ref-86) and on the integration-coupling mechanism that makes instrument slip a mission-level event [\[77\]](#ref-77). The connecting logic is that structurally different owners, drivers, and levers (Section 3.1) imply statistically different hazards if the apparatus can detect them. The limit is that detection requires adequate cohort size and clean cause coding, so a non-rejection may reflect power rather than true non-separability. The second proposition is archetype dependence: instrument-driven slip is the dominant subdistribution hazard for first-of-kind active-sensor missions but not for passive-radiometer heritage missions. This rests on the nonlinear TRL-slip relationship that loads on low-maturity first-of-kind technologies [\[47\]](#ref-47), [\[46\]](#ref-46) and on the heritage mechanism by which a direct-rebuild instrument carries little maturation risk and shifts the dominant first-slip hazard to the launch channel [\[72\]](#ref-72), [\[91\]](#ref-91). The connecting logic is the Bearden-controlled distinction between novelty and general complexity [\[17\]](#ref-17). The limit is that heritage is sometimes nominal rather than real [\[123\]](#ref-123), so the archetype coding must guard against overstated heritage. The third proposition is robustness against the rival: the separation and the archetype dependence survive control for estimating optimism. Both the basis and the connecting logic come from the optimism literature itself [\[57\]](#ref-57), [\[58\]](#ref-58), which motivates the threat and supplies the reference-class control. The limit is the known sensitivity of reference-class methods to class definition [\[13\]](#ref-13), so the dissertation reports the result with and without the optimism control and treats dependence on the control as a refutation.

These three propositions are the separability, archetype-dependence, and robustness conditions of the dissertation's falsification rule, and they descend from the literature rather than from assertion. The remainder of the dissertation operationalizes them: the data and measurement chapter constructs the cohort, the cause coding, and the archetype variable from the named sources; the research-design chapter specifies the Fine-Gray estimator and the identification strategy; the analysis-plan chapter fixes the estimation sequence, the expected (illustrative, not executed) result patterns, and the pre-registered robustness battery that tests the third proposition. The literature has brought the argument to the edge of the empirical work. It has shown that the problem is real and material, that the apparatus to address its causal structure exists but is unused on this problem, that the dominant rival is identified and controllable, and that the residual risk of the design is bounded and stated. What remains is to execute, which the design-stage honesty of this dissertation defers while specifying it completely.


## Chapter 3 References

Citations in this chapter are numbered to the consolidated reference list in the Back Matter (Part I: References); each in-text marker links directly to its full entry there.


# Chapter 4: Data and Measurement

## 4.0 The chapter thesis

The data and measurement apparatus of this dissertation can be stated in one sentence before any of it is justified: a cohort of NASA Earth-observing missions, each contributing one development spell measured in months from Key Decision Point B (KDP-B), is assembled from four named sources (NASA Cost Analysis Data Requirement records accessed through the ONCE database, the NASA Instrument Cost Model parameter records, the Government Accountability Office annual assessments of major NASA projects, and NASA TechPort sensor technology-readiness-level histories), and every variable the model requires is operationalized against a specific field in one of those four sources, with the single non-trivial measurement act being the cause-coding of the first slip event by reconciling two independent narrative records. Everything that follows in this chapter defends that sentence: which source supplies which field, at what coverage and with what known bias, how each construct becomes a measurable quantity, and where the measurement is fragile enough that the analysis plan must carry a sensitivity check rather than a point estimate.

This framing matters because a competing-risks design is only as credible as its event-coding. The estimator described in Chapter 5 (the Fine-Gray subdistribution hazard model, with cause-specific Cox models and Gray's test in parallel) is a mature, well-understood machine. The risk to the contribution does not live in the estimator. It lives in whether the two competing events, instrument-driven first slip and launch-driven first slip, can be measured as distinct events in the historical record without one bleeding into the other. Current practice records NASA mission slip as a single delay variable and regresses it as if it had one cause; Lieber and Donor [\[86\]](#ref-86) document precisely this collapse, noting that the phrase "directly related to" hides several distinct root causes operating at once. The desired state is a dataset in which the first slip of every mission carries a defensible cause label drawn from two reconciled sources. The gap is that no such dataset has been assembled for NASA Earth-observing missions, because no prior study has needed the cause label as a competing-risks outcome. Leave that gap unfilled and the separability question of Chapter 1 cannot be tested at all: without cause-coded events there are no competing risks to separate. This chapter closes that gap at the level of measurement design, and it does so at the design stage, which means it specifies how the cohort would be built and coded and states the coverage and bias of each source, without claiming the cohort has yet been assembled or any event yet coded.

The chapter proceeds from sources to unit to variables to archetype to cause-coding to data-quality and ethics. Section 4.1 treats each of the four named datasets in depth: provenance, access path, coverage, unit of record, and known biases. Section 4.2 fixes the unit of analysis, the KDP-B-origin mission-development spell, and confronts the left-truncation problem that the spell-origin convention introduces. Section 4.3 operationalizes every variable in the bible against a source field and presents the full measurement table. Section 4.4 builds the archetype effect modifier from the NICM instrument taxonomy with active and passive exemplars grounded in the Earth-observation instrument literature. Section 4.5 specifies the two-source cause-coding procedure and the handling of un-codable events. Section 4.6 catalogues data quality, validation against known values, coverage limitations, and the ethics and access constraints that bound reproducibility.

## 4.1 The four named datasets in depth

The cohort is assembled from four sources, each contributing a distinct facet of the record. Two are restricted (CADRe and, for some fields, NICM), and two are public (the GAO assessments and TechPort). The division matters for reproducibility and is stated plainly here and again in Section 4.6.

### 4.1.1 NASA CADRe records, accessed through ONCE

**Provenance.** The Cost Analysis Data Requirement (CADRe) is the standardized cost-and-technical record that NASA projects above a cost threshold are required to produce at each major milestone. A CADRe document comes in three parts: Part A is a narrative and programmatic description of the project, Part B is a technical data dictionary of the spacecraft and its instruments, and Part C is the time-phased cost and schedule data. For this dissertation, CADRe is the spine of the cohort: it is the only source that carries, per mission and per milestone, the confirmation-baseline schedule, the current and as-flown schedule, the milestone dates, the mass and power parameters, and, in Part A, the narrative attribution of why a milestone moved. The schedule-cost coupling that motivates the entire dissertation is visible inside the CADRe record itself, because the cost growth recorded in Part C is, in large part, the monetized image of the schedule movement narrated in Part A.

**Access.** CADRe is not a public database. It is held by the NASA Cost Analysis Division and surfaced to analysts through ONCE, the One NASA Cost Engineering database. Access is granted to NASA personnel and to FFRDC cost analysts working under a data-use agreement. This access path is a hard constraint on reproducibility, discussed in Section 4.6.4, and it is the reason CADRe is described here by its provenance and field structure rather than cited as an open-literature source.

**Coverage.** CADRe coverage begins effectively in the mid-2000s as the requirement matured, with retrospective records of varying completeness for earlier missions. For the Earth-observing cohort targeted here (roughly 1990 to the present), CADRe is richest for the more recent missions and thinner for the oldest, a coverage gradient that interacts with the TechPort gradient described below and is treated as a known limitation. The unit of record in CADRe is the mission-milestone: one mission generates several CADRe snapshots across its life, at KDP-B, KDP-C (confirmation), and subsequent milestones.

**Known biases.** Two biases are material. First, the Part A cause narrative is written by the project, so it carries the project's own framing of why a milestone moved; a project office may have an incentive to attribute slip to exogenous causes (launch availability) rather than to its own instrument development. This is exactly the bias that the two-source reconciliation with GAO is designed to bound, because the GAO narrative is written by an external auditor with the opposite incentive. Second, CADRe completeness improved over time, so the richness of the cause narrative is correlated with calendar period, which the calendar-period fixed effects in the specification are intended to absorb.

### 4.1.2 The NICM and NICM-E instrument parameter records

**Provenance.** The NASA Instrument Cost Model (NICM) and its Explorer-class extension (NICM-E) are parametric cost models built on a database of flown NASA instruments. Habib-agahi documents NICM-E as a model that estimates instrument cost as a function of mass, power, data rate, and technology parameters [\[65\]](#ref-65), and the companion cost-time-trends work establishes that instrument development is a primary and distinct locus of cost and schedule risk, separate from the spacecraft bus and from the launch vehicle [\[66\]](#ref-66). For this dissertation, NICM contributes two things: the instrument-level parameter records (mass, power, data rate, instrument type, pointing) that serve as instrument-side covariates, and the instrument-type taxonomy that defines the archetype variable.

**Access.** The NICM tool and its underlying instrument database are NASA cost-community resources. The published NICM-E technical report is available through the NASA Technical Reports Server [\[65\]](#ref-65), but the full per-instrument parameter database that backs the model is a restricted cost-community asset, accessed under the same FFRDC arrangements as CADRe. The taxonomy and the modeling approach are public; the row-level parameter values for specific instruments are treated as restricted.

**Coverage.** NICM's instrument database spans NASA science instruments across mission classes, not only Earth-observing instruments. This is an advantage for taxonomy construction, since the active-versus-passive and radiometer-versus-other distinctions are well populated across the full instrument set, and a caution for parameter values, since Earth-observing instruments are a subset and the cohort uses only that subset. The unit of record is the instrument, not the mission; a multi-instrument mission contributes several NICM rows, which is why the count of distinct instruments per mission is itself a covariate.

**Known biases.** NICM is a cost model, so its database is curated toward instruments with reasonably complete cost actuals, which can under-represent instruments that were cancelled or radically descoped before flight. Because anticipatory descoping is one of the rival explanations the dissertation must confront (a mission that anticipates instrument risk may descope rather than slip), the under-representation of descoped instruments in NICM is a bias that pushes toward measuring the surviving, flown instruments and is flagged for the descope-history covariate discussed in Chapters 5 and 7.
### 4.1.3 The GAO Assessments of Major NASA Projects

**Provenance.** The Government Accountability Office publishes an annual assessment of major NASA projects. Each report tabulates, per project, the established cost and schedule baseline, the current estimate, and a narrative attribution of the sources of cost growth and schedule slip in the reporting year. The GAO assessment is the external, audit-side counterpart to the project-authored CADRe Part A narrative. Where CADRe Part A says, in the project's own words, why a milestone moved, the GAO narrative says, in the auditor's words, why the project's cost and schedule changed.

**Access.** The GAO assessments are public documents, published yearly and freely available. This is a decisive advantage for the cause-coding procedure. Because one of the two reconciled sources is open, the cause-coding rule can be partially audited by any reader with access to the GAO reports even if they lack CADRe access. The Bitten and Sobel cost-and-policy literature draws on exactly this public GAO record to study how policy changes and contract types move NASA science-mission cost and schedule [\[22\]](#ref-22), [\[123\]](#ref-123), which establishes the GAO assessments as a recognized, analyzable source for mission cost-schedule outcomes.

**Coverage.** The GAO assessment series covers major NASA projects above a reporting threshold. That threshold captures the Earth-observing missions of interest here consistently from roughly the late 2000s forward, and more sporadically before. The annual cadence means a single mission appears in several consecutive GAO reports across its development, so the GAO record, like CADRe, is naturally a mission-year panel from which the first cause-coded slip event is extracted.

**Known biases.** The GAO narrative is written to a Congressional-oversight audience, so it emphasizes accountability and can foreground management and contracting causes. This auditor framing is the deliberate counterweight to the project framing in CADRe Part A. Reconciling the two has value precisely because their biases run in opposite directions: an event coded the same way by both sources is robust to either bias alone.

### 4.1.4 NASA TechPort sensor TRL histories

**Provenance.** TechPort is NASA's technology-portfolio database, recording the technology-readiness-level (TRL) status and history of technologies across the agency's portfolio. For this dissertation, TechPort supplies the most important instrument-side covariate: the entry TRL of the least-mature sensor technology carried by each mission at KDP-B. The theoretical basis for this covariate is direct. Dubos, Saleh, and Braun show, on a cross-mission dataset, that the TRL of a spacecraft's least-mature technology at authorization drives subsequent schedule slip, with a nonlinear relationship that steepens at the low-maturity end [\[47\]](#ref-47), [\[46\]](#ref-46). The TRL construct itself rests on the NASA technology-readiness-assessment framework documented by Hirshorn and Jefferies, which also catalogues the ambiguities in TRL assignment that the measurement procedure here must respect [\[69\]](#ref-69).

**Access.** TechPort is publicly accessible through a documented API, which makes the entry-TRL covariate the most reproducible element of the instrument side. The API exposes technology records, their TRL histories, and their associations with missions and programs.

**Coverage.** TechPort coverage is uneven across the cohort's time span. It is well populated for technologies developed within the era of the database and progressively thinner for older missions whose technology histories predate systematic TRL tracking. This is the second of the two coverage gradients, the first being CADRe, and the two are correlated: the oldest missions in the cohort are the ones for which both the cause narrative and the TRL history are least complete. The unit of record in TechPort is the technology, which must be associated with the mission and with the specific sensor to yield the least-mature-sensor-technology TRL at KDP-B.

**Known biases.** TRL assignment is partly judgmental, as Hirshorn and Jefferies document [\[69\]](#ref-69), so two assessors can record different TRLs for the same technology at the same time. This introduces measurement error into the entry-TRL covariate that is not classical, since it can correlate with how aggressively a program reports its own maturity. It is one reason the analysis plan tests robustness to TRL coding rather than treating entry TRL as measured without error.

### 4.1.5 Why these four and not others

The four sources are chosen because together they close the measurement requirements of a competing-risks design, and individually none does. CADRe carries the schedule outcome and the project-side cause narrative but is restricted and project-framed. GAO carries an independent, public, auditor-side cause narrative but not the instrument parameters. NICM carries the instrument parameters and the taxonomy but not the schedule outcome or the cause. TechPort carries the maturity covariate but nothing about slip. The design rests on the join of all four on the mission key, and that join is what makes the cause-coding a two-source reconciliation rather than a single-source label. The broader Earth-observation literature that frames why these missions matter (the Landsat continuity record [\[72\]](#ref-72), [\[138\]](#ref-138), [\[139\]](#ref-139), [\[91\]](#ref-91), [\[149\]](#ref-149), the decadal-survey-driven mission pipeline [\[93\]](#ref-93), [\[102\]](#ref-102), [\[38\]](#ref-38), [\[101\]](#ref-101), [\[50\]](#ref-50), and the comparative Earth-observation context [\[92\]](#ref-92), [\[83\]](#ref-83), [\[105\]](#ref-105), [\[110\]](#ref-110), [\[44\]](#ref-44)) supplies the substantive grounding for the archetype examples in Section 4.4 but is not itself a data source for the cohort. It is the literature that tells us which missions are continuity missions and which carry first-of-kind active sensors.

## 4.2 Unit of analysis: the mission-development spell

### 4.2.1 Definition

The unit of analysis is the mission-development spell, defined exactly as in the prospectus and bible and carried forward here without alteration. The spell is the interval from the start of Phase B, beginning at Key Decision Point B (KDP-B), to the earlier of (a) the first recorded schedule-slip event of either cause or (b) the launch readiness date. Time is measured in months from KDP-B. Each mission contributes exactly one spell. The spell ends in one of three mutually exclusive states: instrument-driven first slip, launch-driven first slip, or administrative censoring at launch with no above-threshold slip recorded. The two slip states are the competing events; the censoring state is the third, non-event outcome.

This is a single-spell, first-event survival design, not a recurrent-event design. The choice is deliberate and follows the policy question. A program office allocates reserve at confirmation against the risk of the first slip that breaks the committed baseline, so the decision-relevant quantity is the cumulative incidence of each cause of first slip, not the full sequence of subsequent slips. Modeling only the first event keeps the unit aligned with the decision and keeps the cause-coding tractable, because attributing a single dominant cause to the first baseline-breaking movement is more defensible than disentangling causes across a long sequence of later movements where instrument and launch effects increasingly entangle.

### 4.2.2 The spell-origin convention and left truncation

The choice of KDP-B as the spell origin is a convention, and the chapter states it as such rather than as a true zero of risk. Schedule risk is incurred before KDP-B: missions enter the formal cost-and-schedule record only after a period of pre-formulation and Phase A study during which technology maturation and concept risk have already accrued. The instrument cost-time-trends literature makes this concrete, showing that instrument development risk builds across the full development arc and is not switched on at a single milestone [\[66\]](#ref-66). KDP-B is chosen as the origin because it is the first milestone at which the four sources reliably carry a consistent baseline schedule, a recorded entry TRL, and an instrument taxonomy, so it is the earliest point at which the spell can be measured uniformly across missions. The convention is therefore a measurement choice driven by source availability, and it is part of the construct definition for the spell, not a claim about the physics of risk onset.

This convention creates a left-truncation problem that the design handles explicitly. A mission whose risk clock effectively started in Phase A but whose spell is measured from KDP-B has had some of its risk-bearing time truncated from the record. The Fine-Gray apparatus accommodates left truncation [\[55\]](#ref-55), and the analysis uses the left-truncated form where a mission's entry time into the risk set differs from the common origin. Where a mission entered the formal record later than KDP-B (for example, a mission whose CADRe baseline was first established at confirmation rather than at KDP-B), its delayed entry is modeled as left truncation rather than treated as if the mission were at risk from the common origin. Ignoring left truncation would bias the at-risk set in the early months of the spell, inflating the apparent early hazard. The correction has a limit: it depends on accurately dating each mission's entry into the record, which is itself a CADRe-completeness-dependent quantity, so the left-truncation handling is reported with the same coverage caveat as the rest of the CADRe-derived data. Confidence in the spell-origin convention is moderate. It is the defensible choice given source availability, and the left-truncation form protects against the main bias, but the convention would be strengthened by a sensitivity analysis (specified in Chapter 6) that re-origins the spell at confirmation for the subset of missions with complete pre-confirmation records.

### 4.2.3 Why one spell per mission and not the mission-year panel

The four sources are naturally mission-year panels: CADRe snapshots, GAO annual assessments, and TechPort histories all record multiple observations per mission across time. The design collapses this panel to one spell per mission by extracting the first cause-coded slip event, or the censoring at launch. The collapse is what defines the competing-risks structure: the occurrence of one cause of first slip removes the mission from the risk set for the other cause as a "first" event, which is the defining feature of competing risks rather than two independent processes [\[9\]](#ref-9), [\[135\]](#ref-135). Retaining the full panel would change the question from "what is the cumulative incidence of each cause of first slip" to "what is the rate of all slips," which is not the decision-relevant quantity and would not map onto the reserve-allocation decision the dissertation exists to inform.

## 4.3 Variable construction and the measurement table

This section operationalizes every variable named in the bible against a specific source field. The construct, operational definition, source, and scale for each variable are summarized in the measurement table (Table 4.1), and the prose below it elaborates the constructs that require interpretation rather than a one-line definition. Each variable is carried forward verbatim from the prospectus data section and from the bible; this section adds the operational detail (the exact field, the scale, the coding rule) that turns the bible definition into a measurable quantity.

### Table 4.1: Measurement table

| Construct | Operational definition | Source | Scale |
|---|---|---|---|
| Spell time | Months from KDP-B to first slip event or launch readiness | CADRe Part A/C milestone dates | Continuous (months), positive |
| Competing event (outcome) | First above-threshold committed-baseline slip, cause-coded instrument-driven or launch-driven; else censored at launch | CADRe Part A reconciled with GAO assessment narrative | Categorical, 3 states (instrument slip / launch slip / censored) |
| Slip threshold | Net launch-date movement exceeding threshold (default 2 months; varied 1-4 in sensitivity) | CADRe Part C committed vs current launch date | Continuous trigger; binary above/below |
| Sensor archetype (effect modifier) | First-of-kind active sensor vs passive-radiometer heritage; mixed/ambiguous as third category | NICM instrument taxonomy | Categorical, 2 primary + 1 robustness |
| Entry TRL | TRL of least-mature sensor technology at KDP-B | TechPort technology history (API) | Ordinal, TRL 1-9 (lower = less mature) |
| Instrument count | Number of distinct science instruments on the mission | NICM instrument records / CADRe Part B | Count, integer >= 1 |
| Instrument mass | Total or per-instrument mass | NICM parameter record | Continuous (kg), positive |
| Instrument power | Instrument power draw | NICM parameter record | Continuous (W), positive |
| Instrument data rate | Instrument data rate | NICM parameter record | Continuous (Mbps), positive |
| Launch-vehicle class | Vehicle performance class | CADRe Part B / GAO / public manifest | Categorical |
| Launch provider | Launch services provider | CADRe / GAO / public manifest | Categorical |
| Shared-manifest indicator | Shared (rideshare/co-manifest) vs dedicated launch | CADRe / public manifest | Binary |
| Provider-in-development indicator | Launch service in its own development at manifest assignment | Public launch record / GAO | Binary |
| Mission complexity index | Bearden-style complexity score from subsystem parameters | NICM/CADRe parameters per Bearden method | Continuous index |
| Estimating-optimism proxy | Ratio of confirmation-baseline schedule to reference-class median schedule for similar-class missions | CADRe schedule vs constructed reference class | Continuous ratio (>1 = optimistic) |
| Calendar period | Era of KDP-B (acquisition-policy regime) | CADRe / GAO milestone dates | Categorical (period fixed effects) |
| Descope history (auxiliary) | Recorded instrument descope before slip window | CADRe Part A / GAO narrative | Binary/categorical, where recorded |
### 4.3.1 The outcome and the slip threshold

The outcome is the cause-coded first slip event, and its measurement has two parts: detecting that a slip occurred and was large enough to count, and coding its cause. The detection part is treated here. Cause-coding is the subject of Section 4.5, because it is the one genuinely difficult measurement act in the design.

A slip event is a committed-baseline schedule movement exceeding a threshold, defined illustratively as two months of net launch-date movement, attributable to a single dominant cause in the milestone period. The threshold is operationalized from CADRe Part C as the net movement of the committed launch readiness date between consecutive baselines. The two-month default is a starting value, not a finding; the analysis plan varies it from one month to four months (Chapter 6) and requires the dominance pattern to be stable across that range to count as confirmed. A threshold exists at all because not every baseline adjustment is a slip event in the sense relevant to reserve allocation. Small, routine replans should not trigger an event, and the threshold filters them. The design must respect one objection: the threshold is a researcher choice that could be tuned to produce a result. For that reason it is pre-registered as a sensitivity range rather than fixed at a single convenient value. Confidence in the threshold construct is moderate. It is defensible and pre-registered, and the stability requirement across the one-to-four-month range is the evidence that would raise or lower confidence in any specific result derived under it.

### 4.3.2 Instrument-side covariates

The instrument-side covariates are entry TRL, instrument count, mass, power, and data rate. Entry TRL is the headline covariate, measured as the TRL of the least-mature sensor technology at KDP-B and drawn from TechPort. "Least-mature sensor technology" is a deliberate choice that follows Dubos, Saleh, and Braun, whose result concerns the maturity of the least-ready technology, not the average maturity [\[47\]](#ref-47). The operational rule is to identify, for each mission, the set of sensor technologies carried, retrieve each technology's TRL history from TechPort, take the TRL recorded at or immediately before KDP-B, and assign the minimum across the sensor set as the mission's entry TRL. The scale is ordinal, TRL 1 through 9, with lower values indicating less maturity and, by the Dubos result, higher expected instrument-side slip hazard. Instrument count, mass, power, and data rate come from the NICM parameter records. Instrument count is the integer number of distinct science instruments, and the continuous parameters are taken at the per-instrument or mission-total level as the specification requires. The Kipp instrument-schedule-growth result supplies the mechanism that licenses these as instrument-side covariates: instrument schedule growth propagates to mission cost and schedule growth, so instrument parameters that predict instrument schedule growth are the right covariates for the instrument-driven slip hazard [\[77\]](#ref-77).

### 4.3.3 Launch-side covariates

The launch-side covariates are launch-vehicle class, launch provider, a shared-versus-dedicated manifest indicator, and an indicator for a launch service in its own development at the time of manifest assignment. Vehicle class and provider are categorical, taken from CADRe Part B, the GAO record, or the public launch manifest, cross-checked where the sources overlap. The shared-manifest indicator is binary, distinguishing missions co-manifested or riding shared with another payload from missions on a dedicated vehicle. These are separated because shared manifests expose a mission to a second payload's schedule and to manifest-reshuffle dynamics that a dedicated mission does not face. The provider-in-development indicator is binary and flags missions whose assigned launch service was itself still in development at manifest assignment, exposing the mission to the provider's own development slip. The exogeneity of much launch-side slip, a shared-vehicle anomaly, a manifest reshuffle, or a provider's development slip, is what makes the launch event a credible competing risk distinct from instrument maturation, as the Landsat continuity record documents [\[72\]](#ref-72), [\[74\]](#ref-74). The launch-side covariates are the measurable correlates of that exogenous exposure.

### 4.3.4 Controls

Three controls guard against the rival explanations. The mission complexity index is constructed after Bearden's complexity-based cost-estimating method, which ranks subsystem complexity from technical parameters and combines the ranks into an overall index [\[17\]](#ref-17); its role is to ensure that "first-of-kind active sensor" carries independent explanatory content beyond general mission complexity, the second rival explanation the discussion must defeat. The estimating-optimism proxy is the ratio of the confirmation-baseline schedule to a reference-class median schedule for missions of similar class, constructed following the reference-class-forecasting logic that Flyvbjerg establishes as the discipline for detecting planning-stage optimism [\[57\]](#ref-57). A ratio below one indicates a schedule baseline shorter than the reference class would predict, the signature of optimism, and the proxy is the control that prevents the instrument-launch separation from being an artifact of which missions were optimistically estimated. The reference class for each mission is constructed from the cohort and from the broader NICM instrument set, so that "similar-class missions" is an explicit, auditable set rather than a judgment. Calendar-period fixed effects absorb era-specific acquisition policy. The policy-effects literature shows that policy changes and contract types measurably move NASA science-mission cost and schedule [\[22\]](#ref-22), [\[123\]](#ref-123), so the era a mission was baselined in is a confound that the period effects remove. The cost-modeling context for the electronic-parts and instrument cost drivers that feed both the complexity index and the optimism reference class is documented in the NASA cost-estimating literature [\[67\]](#ref-67).

## 4.4 Archetype construction from the NICM taxonomy

### 4.4.1 The construct and its source

The archetype variable is the effect modifier on which the entire H1 contribution turns, so its construction is given its own section. The construct is sensor novelty, operationalized as a categorical contrast between two archetypes: a first-of-kind active sensor (for example, a lidar or a radar with no direct flight heritage) and a passive-radiometer heritage mission (a passive radiometer with direct heritage from a predecessor instrument). Missions that are mixed or ambiguous, carrying both archetypes or a sensor that fits neither cleanly, form a third category used only for robustness and excluded from the primary contrast. The source for the classification is the NICM instrument taxonomy, which carries the active-versus-passive and radiometer-versus-other distinctions across the full NASA instrument set [\[65\]](#ref-65).

The operational rule is a two-step crosswalk. First, each mission's instruments are mapped to their NICM instrument-type categories. Second, the mission is classified: if its defining science instrument is an active sensor (lidar, radar, or other active illumination) without a direct flight-heritage predecessor, the mission is first-of-kind active; if its defining instrument is a passive radiometer with a direct heritage rebuild or near-rebuild of a flown unit, the mission is passive-radiometer heritage; otherwise it is mixed or ambiguous. "Direct flight heritage" is the decisive sub-construct, coded from whether the instrument is a documented rebuild, recurring build, or close derivative of a previously flown instrument, which the NICM record and the mission documentation jointly establish.

### 4.4.2 Active-sensor exemplars

The first-of-kind active-sensor archetype is grounded in the spaceborne active-instrument literature. A spaceborne lidar mission carries an active instrument whose detector, laser, and optics technology maturity drives the development schedule; the Acta Astronautica spacecraft-and-optics design study for a spaceborne lidar mission illustrates this archetype, where technology maturity is the binding constraint on the development schedule [\[127\]](#ref-127). Active Earth-observation instruments of this kind, lidar and radar, are the class for which entry TRL is most likely to be low and most likely to drive instrument-side slip, which is the substantive reason the dissertation predicts instrument-driven slip to dominate in this stratum. The hyperspectral-instrument development history [\[110\]](#ref-110) supplies a parallel case of a relatively novel sensing technology whose maturation drives schedule, reinforcing that the active and novel-sensor end of the taxonomy is where instrument maturation risk concentrates.

### 4.4.3 Passive-radiometer heritage exemplars

The passive-radiometer heritage archetype is grounded in the continuity-mission and passive-instrument literature. The Landsat program is the canonical Earth-observation continuity case: a long-running series in which later missions rebuild or closely derive their instruments from flown predecessors, carrying little instrument-maturation risk and facing their dominant schedule pressure on the launch side [\[72\]](#ref-72), [\[138\]](#ref-138), [\[139\]](#ref-139), [\[91\]](#ref-91). A passive thermal-infrared or multi-channel radiometer with direct heritage, of the kind the Acta Astronautica radiometer design study describes [\[45\]](#ref-45), is the archetypal passive-radiometer heritage instrument: its technology is mature, its development risk is low, and when its mission slips, the slip is disproportionately launch-driven (a shared-vehicle anomaly, a manifest reshuffle, or a provider's development slip), as the Landsat Data Continuity Mission launch record documents [\[74\]](#ref-74). The decadal-survey-driven passive-instrument missions (for example, the soil-moisture and atmospheric-aerosol continuity missions [\[102\]](#ref-102), [\[38\]](#ref-38)) extend the set of candidate heritage missions in the cohort.

### 4.4.4 Why the binary contrast and its known cost

The archetype is a binary simplification of a continuous novelty spectrum, and the chapter states this as a known construct-validity cost rather than hiding it. The binary is justified by decision relevance: a program office allocates reserve against a discrete archetype judgment, not against a continuous novelty score, so the binary maps onto the decision. The objection is that forcing a continuum into two bins could manufacture a contrast where the underlying reality is graded. For that reason the mixed-or-ambiguous third category is retained and brought into the model as a separate stratum in a robustness check (Chapter 6), to confirm that the binary is not imposing a spurious dichotomy. The Callaway-Sant'Anna discipline that the bible imports requires the archetype to enter as an explicit effect modifier and the archetype-specific hazards to be estimated as separable building blocks rather than collapsed into a pooled coefficient [\[27\]](#ref-27); the binary archetype is the operational form of that effect modifier, and the third category is the guard against the binary becoming a pooling of its own. Confidence in the archetype construct is moderate to high for the clear exemplars (a flagship lidar mission, a Landsat continuity mission) and lower for the boundary cases, which is precisely why the boundary cases are routed to the robustness stratum rather than forced into the primary contrast.

## 4.5 Cause-coding by two-source reconciliation

### 4.5.1 The procedure

The one measurement act that carries the design is coding the cause of the first slip event as instrument-driven or launch-driven. The procedure is a two-source reconciliation. For each mission and the milestone period in which its first above-threshold slip occurs, two independent narratives are read: the CADRe Part A narrative, written by the project, and the GAO assessment narrative for the same project-year, written by the external auditor. Each narrative is coded for the dominant attributed cause of the slip: instrument development (the science instrument failed environmental test, could not close its calibration budget, or carried a technology below its assumed maturity) or launch-vehicle availability (manifest congestion, a shared-vehicle anomaly, or the provider's own development slip). The slip event's cause is coded high-confidence when both sources name the same dominant cause. The cause-coding is frozen before any modeling, so that the modeling cannot feed back into the coding.

### 4.5.2 Handling un-codable events

Events whose cause cannot be coded consistently across the two sources, where CADRe names one cause and GAO names another, or where neither names a single dominant cause, are flagged as un-codable rather than forced into a bin. The flagging is itself a measurement output, not a failure: it quantifies how much of the cohort's first-slip experience is multi-cause or contested. Un-codable events are handled in sensitivity analysis (Chapter 6), recoded both ways (once as instrument-driven, once as launch-driven), with the result reported under both codings; a conclusion that flips between codings is reported as fragile. This is the operational form of the Fogel methodological warning that the bible imports: the cause decomposition is a constructed quantity that must be reported with explicit bounds and a sensitivity analysis, not asserted as observed [\[82\]](#ref-82).

### 4.5.3 Why two sources and what the reconciliation buys
The reason for using two sources is that each single source carries a directional bias, and the two biases run opposite. CADRe Part A is project-authored and may under-attribute slip to the project's own instrument development; the GAO narrative is auditor-authored and may over-attribute it to management and contracting. An event coded the same way by both is resistant to either bias alone, because a project incentive to blame the launch side and an auditor incentive to blame management would both have to fail in the same direction to manufacture a false instrument-driven code. The basis for treating GAO as an independent, analyzable cause source is the policy-effects studies that use the GAO record to attribute NASA science-mission cost and schedule growth [\[22\]](#ref-22), [\[123\]](#ref-123). The objection the design respects is that both sources remain narratives, and narrative attribution is subjective even when two narratives agree; the reconciliation reduces this problem without eliminating it, which is why construct validity for "instrument-driven slip" is reported as defensible but imperfect and why the un-codable flag and the recoding sensitivity are mandatory rather than optional. Confidence in the cause-coding is moderate. The two-source rule improves on single-source coding, the directional-bias argument is sound, and the recoding sensitivity bounds the residual error, but the construct remains narrative-based, and the design does not claim to measure the physical cause without error.

## 4.6 Data quality, validation, coverage, and ethics

### 4.6.1 Data-quality checks

Three data-quality checks are specified at the cohort-assembly stage. The first is cross-source key reconciliation: the mission key that joins CADRe, NICM, GAO, and TechPort must resolve to the same mission across all four, and missions that fail to join cleanly (a renamed mission, a re-baselined mission that appears under two identifiers) are reconciled by hand against the mission documentation before entering the cohort. The second is schedule-internal consistency: the milestone dates extracted from CADRe Part A and Part C must be internally consistent (KDP-B before confirmation before launch readiness), and inconsistencies are resolved against the GAO baseline record. The third is covariate range checks: instrument mass, power, and data rate from NICM, and entry TRL from TechPort, are range-checked against the plausible envelope for Earth-observing instruments, with out-of-range values flagged for source re-checking rather than silently retained. These checks are mechanical and are run before the cause-coding is frozen.

### 4.6.2 Validation against known values

The cohort admits a partial validation against known values, specified here even though it is not yet executed. The validation has two arms. The first arm validates the schedule outcome. For the subset of cohort missions whose launch dates and major slips are a matter of public record (the Landsat continuity missions are the clearest case, with a well-documented launch history [\[72\]](#ref-72), [\[74\]](#ref-74), [\[91\]](#ref-91)), the cohort's recorded first-slip timing and the public launch record must agree; a cohort mission whose coded slip history contradicts the public launch record is a data-quality failure to be investigated. The second arm validates the instrument-side covariate against the established TRL-slip relationship. Because Dubos, Saleh, and Braun established a quantitative entry-TRL-to-slip relationship on an independent cross-mission dataset [\[47\]](#ref-47), the cohort's own entry-TRL-to-instrument-slip association can be checked for sign and rough magnitude consistency with that prior result. Consistency is reassurance, not confirmation of H1 (which concerns the archetype-conditioned competing-risks structure, not the marginal TRL-slip association); a gross inconsistency would signal a measurement problem in the entry-TRL covariate. Both validation arms are design-stage specifications: they state what the validation would check and against which known values, and they report no executed validation result.

### 4.6.3 Coverage and limitations

The coverage limitations are real and are stated plainly, carrying forward and elaborating the prospectus limitations. The cohort targets NASA Earth-observing missions from roughly 1990 to the present that reach KDP-B, an expected sample on the order of thirty to sixty missions. This is small for a survival model with two competing events and an archetype interaction, which is the central statistical-conclusion-validity concern and the reason the specification pre-commits to ridge-penalized estimation and an events-per-variable cap (Chapter 5). Cause coding depends on narrative attribution, which is subjective and can mask multi-cause slip; the two-source reconciliation and the un-codable flag bound this but do not remove it. The two coverage gradients, CADRe completeness and TechPort completeness, are correlated and both thin for the oldest missions, so the oldest era of the cohort carries the most measurement uncertainty in both the cause narrative and the entry-TRL covariate. The launch-availability side is the thinnest domain-empirical facet: launch-manifest-driven slip as a survival outcome is under-documented in the open literature, so the launch-side construct leans on the Landsat continuity narrative [\[72\]](#ref-72), [\[74\]](#ref-74) and on the GAO and CADRe primary records rather than on a quantified launch-slip literature, and the launch-side dominance claim for the heritage arm is, by construction, the more evidence-thin half of H1, to be confirmed on the cohort. The cohort is observational: sensor archetype is not randomized to missions, so the archetype contrast compares naturally occurring groups, and the complexity index and optimism proxy are the only defenses against confounding of archetype with other risk drivers. These limitations bound external validity to NASA Earth-observing missions of the studied era and do not extend to commercial constellations (which face different launch economics [\[44\]](#ref-44), [\[92\]](#ref-92)), to non-Earth science, or to non-US programs (the ESA Earth-observation pipeline [\[90\]](#ref-90) is outside the cohort).

### 4.6.4 Ethics, access, and reproducibility

The data are programmatic records of public missions, not human-subjects data, so the ethical surface is access and reproducibility rather than consent or privacy. The binding ethical and practical constraint is that CADRe (through ONCE) and the restricted NICM parameter database are available only to analysts under a NASA or FFRDC data-use agreement. This has two consequences that the dissertation states openly. First, the full analysis is reproducible only by analysts with the data-use agreement; a reader without CADRe access can reproduce the public arm (the GAO cause narratives, the TechPort entry-TRL covariate, the public launch records) but not the project-side narrative or the restricted instrument parameters. Second, because one of the two cause-coding sources (GAO) is public, the cause-coding rule is partially auditable even by readers without CADRe access, a deliberate design choice to maximize transparency under the access constraint. The dissertation handles this by describing CADRe and the restricted NICM database by their provenance, field structure, access path, and known biases (Sections 4.1.1, 4.1.2) rather than reproducing restricted values, and by committing in the analysis plan to report the cause-coding manual, the NICM-to-archetype crosswalk, and the variable dictionary in full (the appendices), so that an analyst who later secures the data-use agreement can reproduce the cohort and the coding exactly. No restricted values are reported in this dissertation; the restricted sources are used to define the measurement procedure, and the procedure is what is published.

### 4.6.5 Summary of the measurement design

The measurement design supports the dissertation's argument at the level of data. The problem is real: cost growth is a monetized image of slip [\[86\]](#ref-86), and slip has structurally distinct instrument and launch origins visible in the CADRe and GAO records. The problem is material: TRL deficit and instrument schedule growth measurably move mission cost and schedule [\[47\]](#ref-47), [\[77\]](#ref-77), and launch-side slip is a separate, documented continuity-mission pressure [\[72\]](#ref-72). The measurement addresses the causal mechanism: the cohort carries the entry-TRL covariate that drives the instrument hazard and the manifest covariates that drive the launch hazard, joined to a two-source cause-coded outcome that keeps the two hazards from contaminating each other. The measurement improves on the single-source alternative: a one-source cause label would inherit one source's directional bias, while the reconciliation bounds it. The residual risk is acceptable and bounded. The cause-coding is narrative-based and the cohort is small, but the un-codable flag, the recoding sensitivity, the threshold sensitivity, the two validation arms, and the explicit access and coverage limitations are the design's honest statement of where the measurement is fragile and what would raise or lower confidence in any result built on it. The chapter has specified the data and the measurement; it has executed neither, and it claims neither, which is the correct posture for a design-stage dissertation whose deliverable is the pre-registered measurement design and not an empirical finding.



# Chapter 5: Research Design and Identification

## 5.0 The chapter's answer

The research design that licenses this dissertation's contribution is a cause-specific competing-risks survival design built on the Fine-Gray proportional subdistribution hazard model, estimated separately for the instrument-driven and the launch-driven first-slip events, with sensor archetype entering as an explicit effect modifier and with the whole apparatus disciplined by penalized partial-likelihood estimation under an events-per-variable cap. This is the chapter's thesis, and the rest of the chapter develops it. The design is chosen because the policy question that motivates the work is predictive rather than purely etiologic: a program office at confirmation wants to know the cumulative probability that a mission of a given archetype will slip first for instrument reasons, in the real world where launch-driven slip also competes to be the first cause, and that cumulative-probability question is exactly the object that the subdistribution hazard governs and the object that an ordinary single-hazard regression cannot deliver [\[55\]](#ref-55), [\[9\]](#ref-9). The identification of the contribution does not rest on an instrument or an exclusion restriction in the econometric sense, because there is no endogenous treatment to be purged; it rests instead on three temporal-ordering and separability claims that the design makes verifiable, on the explicit refusal to pool a heterogeneous hazard into a single coefficient, and on a counterfactual reading of the cumulative incidence function that is stated and bounded rather than assumed. The residual risk in the design is concentrated in cause-coding error and small-sample power, and the chapter ends by showing that each of these residual risks is bounded by a pre-committed mitigation rather than left open.

The problem this chapter addresses, stated in the current-state and desired-state frame that the dissertation carries throughout, is methodological rather than substantive. The current state of practice models schedule slip as a single undifferentiated outcome and regresses it on covariates, which means the analyst has no estimator that can keep instrument-caused slip and launch-caused slip from contaminating each other and no estimator that returns the archetype-conditional cumulative incidence a reserve decision actually needs. The desired state is a fully specified, pre-registered estimator and identification strategy that test separability and archetype dependence at once and that report a counterfactual decomposition of slip into its two channels. The gap is that the competing-risks apparatus that could supply this estimator is mature in biostatistics but has never been transferred to NASA mission schedule slip [\[55\]](#ref-55), [\[5\]](#ref-5), [\[146\]](#ref-146). Leaving the gap open means reserve continues to be allocated against a pooled slip estimate with no design capable of telling a program office which of the two hazards dominates for the mission in front of it. This chapter closes the design half of that gap; Chapter 6 closes the analysis-plan half.

A note on scope is owed at the outset. This is a design-stage dissertation. Every specification, every identification claim, and every threat-mitigation in this chapter is fully developed, but no coefficient, hazard ratio, or cumulative-incidence plateau is reported as an executed empirical result. Where the chapter states an expected sign or an illustrative magnitude, it labels that statement as illustrative and not yet estimated on the full cohort, in keeping with the dissertation's standing guardrail. The deliverable of this chapter is the design and its defense, not findings.

## 5.1 The estimator and why it is chosen

### 5.1.1 The object to be estimated

The decision the design serves is the allocation of scarce schedule and cost reserve at the confirmation milestone. That decision turns on a probability, not a rate: the program office wants to know, for a mission of a given sensor archetype, the probability that its first schedule slip will be instrument-driven by some horizon in development, and the matching probability that its first slip will be launch-driven, so that reserve can be steered toward the dominant channel. The quantity that carries this probability is the cause-specific cumulative incidence function, the probability of experiencing a cause-k first slip by time t in the presence of the competing cause. The estimator must therefore be one whose covariate effects map directly onto the cumulative incidence function rather than onto a hazard whose translation into a probability is mediated by the other cause's hazard in a way that has no closed interpretation. This requirement is the reason the design takes the subdistribution hazard as its primary estimator rather than the cause-specific hazard alone.

The claim that the cumulative incidence function, and not the naive Kaplan-Meier complement, is the correct object is not a stylistic preference; it is a correctness requirement with a documented failure mode for the alternative. Treating the competing event as ordinary right-censoring and applying the Kaplan-Meier estimator to one cause biases the estimated event probability upward, because the Kaplan-Meier construction implicitly assumes that a censored subject remains at risk of the event of interest, whereas a mission whose first slip was launch-driven is no longer at risk of an instrument-driven first slip [\[5\]](#ref-5), [\[135\]](#ref-135). The magnitude of this bias is not negligible and grows with the incidence of the competing event; an applied demonstration in a competing-risks setting shows the Kaplan-Meier complement overstating absolute event risk precisely because it censors the competing event rather than accounting for it [\[140\]](#ref-140). Because the dissertation's whole purpose is to compare the cumulative incidence of two competing causes, an estimator that systematically inflates each cause's incidence in proportion to the other's frequency would corrupt the very comparison the work exists to make. The cumulative incidence function, estimated non-parametrically by the Aalen-Johansen form and modeled covariate-wise by the subdistribution hazard, is the estimator that avoids this failure [\[146\]](#ref-146), [\[37\]](#ref-37).

### 5.1.2 The Fine-Gray subdistribution model, stated

The primary estimator is the Fine-Gray proportional subdistribution hazard model [\[55\]](#ref-55), estimated separately for the instrument-driven event and for the launch-driven event. Carrying the dissertation's fixed notation verbatim, for event type k (k = instrument, launch) the model is

\[ \text{subhazard}_k(t \mid X) = \text{subhazard}_{k0}(t) \exp(X \, \boldsymbol{\beta}_k), \qquad\qquad (1) \]
where \( \text{subhazard}_{k0}(t) \) is the baseline subdistribution hazard for cause k, X is the covariate vector, \( \boldsymbol{\beta}_k \) is the cause-k coefficient vector, and the subdistribution hazard governs the cumulative incidence function \( \text{CIF}_k(t \mid X) \), the probability of a cause-k first slip by time t accounting for the competing event. The coefficient on the archetype-by-instrument-side interaction is the parameter of interest for the contribution H1. The defining feature of the subdistribution construction, the one that separates it from the cause-specific hazard, is its risk-set bookkeeping. A subject who experiences the competing event is retained in the risk set for the event of interest with a decaying weight rather than removed, so that the modeled hazard maps monotonically onto the cumulative incidence function and a positive coefficient translates into a higher cumulative probability of that cause [\[55\]](#ref-55), [\[146\]](#ref-146). This makes \( \beta_{\text{instrument}} \) directly interpretable as the effect of a covariate, for example an entry-TRL deficit, on the cumulative probability that a mission slips first for instrument reasons, which is the decision-relevant statement.

The proportional-subdistribution-hazards assumption that the model imposes is testable, and it will be tested. The design checks proportionality of the subdistribution hazards through Schoenfeld-type residual diagnostics adapted to the Fine-Gray weighting and through time-interacted covariate terms. Where proportionality fails for a covariate, the design admits a time-varying coefficient for that covariate rather than forcing a constant effect, following the standard remedy in the competing-risks literature [\[9\]](#ref-9), [\[10\]](#ref-10). The model is also estimated in its left-truncated form. Some missions enter the formal cost-and-schedule record after a period of pre-formulation study, so the spell origin at Key Decision Point B (KDP-B) is a construct convention rather than a true zero of risk. The design uses the subdistribution model formulation that accommodates right-censored and left-truncated competing-risks data, so that a mission entering the at-risk record later than its KDP-B is handled correctly [\[143\]](#ref-143). This left-truncation-capable form is essential rather than ornamental: mishandling delayed entry would bias the baseline subdistribution hazard and therefore the cumulative incidence comparison at the heart of the contribution.

### 5.1.3 Why the cause-specific Cox model runs in parallel

The subdistribution hazard answers the predictive question, but it is not the only question worth asking, and a design that reported only the subdistribution model would be incomplete by the field's own reporting standards. The cause-specific Cox proportional-hazards model is therefore estimated in parallel for each event, giving the instantaneous-rate interpretation: how a covariate affects the rate of an instrument-driven first slip among missions still at risk and not yet slipped for either cause. The applied competing-risks literature is explicit that the two models answer complementary questions, the cause-specific hazard the etiologic question and the subdistribution hazard the predictive question, and that a complete analysis reports both because divergence between them is itself informative about the mechanism [\[9\]](#ref-9), [\[146\]](#ref-146). The methodological guidance is unambiguous that a competing-risks analysis should report results on all cause-specific hazards and on the cumulative incidence functions, not on one to the exclusion of the other [\[79\]](#ref-79). The design honors this directly. Where the subdistribution coefficient for an instrument-side covariate is large but the cause-specific coefficient is modest, the design reads this as the covariate operating partly through the competing-event structure rather than purely through the instrument rate, and reports the divergence rather than suppressing it.

Gray's test is the nonparametric backbone of the separability claim. Gray's K-sample test compares cumulative incidence functions across groups without imposing the proportional-subdistribution-hazards assumption, and the design uses it as the model-free check on whether the instrument-slip cumulative incidence function differs across the two archetype strata [\[37\]](#ref-37). Because the test is nonparametric, a rejection by Gray's test that agrees with a significant archetype interaction in the parametric subdistribution model provides convergent evidence for separability that does not depend on the parametric form. A disagreement between the two would warn that the parametric result is an artifact of the proportional-hazards assumption, and the design pre-commits to reading such a disagreement as a reason to downgrade confidence rather than to report the parametric result alone.

### 5.1.4 Confidence in the estimator choice

Confidence in the estimator choice is high, and the basis for that grade is the convergence of an entire methodological literature on the same recommendation for the same kind of question. The original derivation of the subdistribution model [\[55\]](#ref-55), the introductory and recommendation papers that map the predictive question onto the subdistribution hazard and the etiologic question onto the cause-specific hazard [\[9\]](#ref-9), [\[146\]](#ref-146), the prognostic-modeling literature that uses the subdistribution model precisely when the target is an absolute risk for decision-making [\[136\]](#ref-136), and the textbook epidemiologic warning against the Kaplan-Meier failure mode [\[5\]](#ref-5), [\[135\]](#ref-135) all point to the same design. What would lower this confidence is a demonstration that the proportional-subdistribution-hazards assumption fails pervasively across the covariates of interest and cannot be repaired by time-varying coefficients, in which case the design would migrate to a flexible or additive subdistribution form [\[84\]](#ref-84), [\[20\]](#ref-20); the design pre-commits to that contingency rather than treating the proportional form as inviolable. What would raise confidence further is an executed proportionality check on the assembled cohort showing the assumption holds for the archetype and TRL covariates, which is a Chapter 6 deliverable, not a Chapter 5 one.

## 5.2 The identification strategy

### 5.2.1 What "identification" means in this design

Identification in this dissertation is not the identification of a structural treatment effect through an instrument or a discontinuity, and it is worth saying so plainly so that the design is judged against the right standard. There is no randomized or quasi-randomized assignment of sensor archetype to missions, and the design does not pretend otherwise; the archetype contrast is a comparison of naturally occurring groups, and the controls for complexity and estimating optimism are the defense against confounding rather than a substitute for randomization. What the design must identify is something more specific and more tractable: that the two slip causes are genuinely separable competing risks in the data, that the archetype variable is a pre-outcome quantity rather than a consequence of the slip it is meant to predict, and that the covariates respect the temporal ordering required for their coefficients to carry the intended interpretation. These are the three identification claims, and each is argued formally below, stating the basis for the claim, the principle that licenses it, the methodological authority that backs that principle, the limits within which it holds, and the objection it must answer.

### 5.2.2 First identification claim: the events are genuinely competing

The instrument-driven and launch-driven first-slip events are genuine competing risks in the data, not two independent processes artificially yoked together. The basis for treating them so is structural. Each mission contributes a single development spell from KDP-B, and the spell ends in exactly one of three mutually exclusive states: instrument-driven first slip, launch-driven first slip, or administrative censoring at launch with no above-threshold slip. By construction, the occurrence of one cause-coded first slip alters the at-risk set for the other, because a mission that has already slipped first for instrument reasons is no longer a candidate to slip first for launch reasons. This mutual exclusivity at the level of the first event is the defining structure of competing risks [\[5\]](#ref-5), [\[146\]](#ref-146).

That structure is what licenses the choice of model. When the occurrence of one event precludes the clean observation of the other as the first cause, the events are competing by definition, and the correct apparatus is the competing-risks model rather than two separate single-event survival models, because separate single-event models would treat each competing event as ordinary censoring and inherit the upward Kaplan-Meier bias [\[5\]](#ref-5), [\[135\]](#ref-135), [\[140\]](#ref-140). The definition is canonical in the methodological literature: a competing risk is an event whose occurrence precludes the occurrence of the primary event of interest [\[9\]](#ref-9), [\[146\]](#ref-146), and the epidemiologic guidance is explicit that treating a competing event as ordinary censoring is the standard error the framework exists to avoid [\[5\]](#ref-5).

Two limits bound the claim. The events are competing for the first-slip outcome only. The design does not claim that instrument and launch problems cannot co-occur over a mission's life; it claims that the first cause-coded slip is one or the other, which is the outcome the spell measures. A mission can and often will experience both kinds of trouble; the competing-risks structure concerns which kind reaches the threshold first. The objection that the design must answer is that if the cause-coding cannot reliably assign a single dominant cause to the first slip, the competing structure is contaminated, and two causes that actually co-occur would be split arbitrarily. That objection is real, and it is the dominant internal-validity threat addressed in Section 5.4; the design answers it with two-source reconciliation and recoding sensitivity rather than by assuming clean separation.

A fourth state that the design must confront is whole-mission cancellation before launch. Under the current three-state specification, a mission that is formally cancelled during development simply exits the at-risk set without recording a first-slip event, and the analysis proceeds as though that exit were uninformative. That assumption is not defensible. Missions are most commonly cancelled after accumulating enough cost and schedule trouble to exhaust political or budgetary support, which means the probability of cancellation is positively correlated with the latent severity of both instrument-driven and launch-driven slip risk. Treating cancellation as independent right-censoring therefore understates the true slip hazard in both arms of the model: the missions most likely to have slipped first for instrument reasons, and those most likely to have slipped first for launch reasons, are precisely the ones at elevated cancellation risk and hence most likely to exit before the slip event is observed [\[150\]](#ref-150). The bias runs in the same direction as the Kaplan-Meier competing-event error the design already corrects, and it would compress the estimated cumulative incidence functions downward, most severely for the active-sensor stratum where instrument-driven slip pressure is hypothesized to be highest.

The design's response is to treat whole-mission cancellation as a third competing event rather than as independent right-censoring. Cancelled missions receive a cause code of "cancellation" rather than a censoring flag, and the Aalen-Johansen nonparametric estimator is applied to the three-state system, instrument-driven first slip, launch-driven first slip, and cancellation, so that the cumulative incidence of each slip cause is estimated in the real population where cancellation also competes. This reclassification has a direct consequence for the instrument-driven versus launch-driven slip separation that is the heart of the contribution. If cancellation is informatively correlated with instrument trouble, as the TRL and complexity evidence suggests it would be for first-of-kind active-sensor missions that run out of margin before launch, then reclassifying it shifts probability mass from the censoring set back into the competing-event structure and raises the instrument-slip cumulative incidence in the active-sensor stratum relative to the naive censoring treatment. The heritage passive stratum, where the instrument risk is lower and cancellations are rarer, is affected less. The net effect sharpens rather than blurs the archetype contrast the contribution predicts: treating cancellation as a competing event makes the hypothesis harder to confirm by chance, because it no longer benefits from inflated censoring-induced survival in the high-risk stratum. The pre-registered robustness battery carries a sensitivity analysis in which cancelled missions are returned to the censoring treatment, so that the dependence of the primary result on the informative-censoring reclassification is directly assessable rather than assumed away.

### 5.2.3 Second identification claim: the archetype is pre-outcome

The sensor-archetype effect modifier is determined before the outcome window and is therefore not a post-treatment quantity contaminated by the slip it is meant to predict. The basis for this is the timing of the classification. The archetype, first-of-kind active sensor versus passive-radiometer heritage mission, is a function of the instrument design fixed at KDP-B and classified from the NASA Instrument Cost Model taxonomy. The spell, and therefore the outcome window, begins at KDP-B, so the classifying information predates the first byte of outcome.

A moderator measured strictly before the outcome cannot be a consequence of the outcome, so an interaction between such a moderator and a covariate is interpretable as effect modification rather than as reverse causation [\[27\]](#ref-27), [\[19\]](#ref-19). This is the central discipline the heterogeneity-aware methodologists impose: the moderator is a fixed pre-period characteristic, and the effect is built up from moderator-specific comparisons rather than imposed by a pooled regression. The refusal to let a pooled coefficient stand in for a heterogeneous set of underlying comparisons is the explicit lesson of the modern difference-in-differences literature, which shows that a single pooled estimator under heterogeneity returns a contaminated weighted average that need not correspond to any quantity the researcher intends [\[27\]](#ref-27), and which therefore insists that heterogeneity enter through explicit, pre-determined group structure [\[19\]](#ref-19), [\[43\]](#ref-43).

The claim holds within a stated limit. The archetype is a binary simplification of a continuous novelty spectrum, and missions that are genuinely mixed are coded into a third category used for robustness rather than forced into the primary contrast. The pre-outcome status of the classification is not weakened by this simplification, but the binary cut is a construct choice that Section 5.4 flags as a construct-validity limitation. The objection that the claim must answer is that a mission which anticipates instrument risk might descope its instrument before KDP-B, which would change its archetype classification in a way correlated with its latent slip risk. This is the anticipatory-descoping channel, and it is the most serious threat to the pre-outcome claim. The design's answer, developed in Section 5.4 and revisited in the discussion chapter, is to measure entry TRL at KDP-B before the slip window opens and to carry descope history as a covariate where the record supports it, so that anticipatory descoping is observed rather than hidden.

### 5.2.4 Third identification claim: covariate temporal ordering

The instrument-side and launch-side covariates are measured at or before KDP-B, preserving the temporal ordering required for their coefficients to be read as effects on subsequent slip. Each covariate is read at confirmation: entry TRL of the least-mature sensor technology is read from TechPort as of KDP-B; instrument mass, power, and data rate are the confirmation-baseline values from the instrument cost-model record; launch-vehicle class, the shared-versus-dedicated manifest indicator, and the provider-in-development indicator are taken from the manifest assignment in force at confirmation. The complexity index and the estimating-optimism proxy are both functions of the confirmation-baseline plan.

A covariate measured before the outcome cannot have been caused by the outcome, so its coefficient is free of the simplest form of reverse-causation contamination, and the temporal ordering is a necessary condition for the regression coefficient to carry a forward-looking, decision-relevant interpretation [\[136\]](#ref-136), [\[27\]](#ref-27). The prognostic-modeling literature that motivates the predictive use of the subdistribution model requires precisely this: predictors fixed at the prediction time so that the absolute-risk estimate is one a decision-maker could actually have formed at that time [\[136\]](#ref-136).

Temporal ordering rules out reverse causation but not omitted-variable confounding. A covariate measured at KDP-B can still be correlated with an unobserved driver of slip, which is why the design carries the complexity and optimism controls and why the archetype contrast remains observational. The objection here is that entry TRL could itself be chosen in anticipation of slip risk, an endogenous-gating concern. The design addresses this by using the TRL recorded at KDP-B before slip is observed, which fixes the temporal ordering, while acknowledging in Section 5.4 that anticipatory gating remains a residual concern that the descope covariate and the cause-specific-versus-subdistribution divergence are designed to surface.

### 5.2.5 The Callaway-Sant'Anna discipline on the interaction
The dominance claim in the contribution is identified by the archetype interaction, and the design imposes a specific methodological discipline on how that interaction is estimated and reported. The discipline is the refusal, inherited from the heterogeneity-aware difference-in-differences literature, to let a single pooled instrument-slip hazard stand in for what is in fact a set of archetype-specific hazards [\[27\]](#ref-27). The analogous error in this setting would be to estimate one instrument-slip subdistribution hazard pooled across all Earth missions and to report it as "the" instrument hazard, when the hazard differs sharply between first-of-kind active-sensor missions and heritage passive-radiometer missions. The pooled coefficient under that heterogeneity is a weighting of the two archetype-specific hazards whose weights are imposed by the estimator rather than chosen by the analyst, and it need not correspond to either archetype's true hazard or to any quantity a reserve decision would use.

The design therefore estimates the archetype-stratified subdistribution hazards as separable building blocks. It fits the model within the first-of-kind active stratum and within the heritage passive stratum, and it tests whether the instrument-slip subdistribution hazard exceeds the launch-slip subdistribution hazard in the active stratum but not in the heritage stratum. Any aggregate across strata that the dissertation reports is a transparent, defensible weighting of those archetype-specific blocks rather than a regression-imposed average, as the heterogeneity literature requires [\[27\]](#ref-27), [\[43\]](#ref-43). The pooled specification with the archetype-by-TRL interaction is retained, but it is read as a compact test of effect modification, not as the source of the headline magnitude; the magnitudes come from the stratified blocks. This is the operational meaning of the contribution's parameter of interest, the archetype-by-instrument-side interaction: the formal hypothesis test that the stratified blocks differ in the predicted direction.

The same authors' doubly robust logic, which combines an outcome model and a weighting model so that consistency survives misspecification of either one [\[118\]](#ref-118), motivates a parallel robustness construction in the design. The archetype-specific hazards are estimated both by the penalized partial-likelihood subdistribution model (an outcome-model route) and by an inverse-probability-of-archetype reweighting that balances the complexity and optimism covariates across strata (a weighting route), so that a conclusion holding under both routes is robust to misspecification of either. This parallel estimation is not a separate study. It is a robustness check on the archetype-specific blocks, and Section 5.5 specifies its computational form.

### 5.2.6 The Fogel counterfactual reading of the cumulative incidence function

The identification of the contribution carries a counterfactual interpretation that the design states explicitly rather than smuggling in. The structural-decomposition tradition demands that an aggregate outcome attributed to a single dominant cause be decomposed into separable channels whose individual magnitudes are measured against a constructed counterfactual [\[82\]](#ref-82). Applied here, schedule slip is the aggregate, the instrument channel and the launch channel are the separable channels, and the archetype-specific cumulative incidence function is the counterfactual-bearing quantity. The cumulative incidence of instrument-driven slip in the active-sensor stratum answers a counterfactual question: what fraction of active-sensor missions would experience instrument-driven slip first, in the world where launch-driven slip also competes to be the first cause. The cumulative incidence function is the right object for this reading because, unlike the cause-specific hazard, it is defined in the real population where both causes operate, so it carries the "in the world where the other cause competes" clause that the decomposition needs.

The decomposition tradition also carries a methodological warning that the design takes seriously and converts into a reporting rule. Social-savings decompositions were attacked on the ground that the counterfactual was constructed rather than observed, and the defense was always to make the construction explicit and to bound it with sensitivity analysis [\[82\]](#ref-82). The analogue here is exact: the cumulative incidence function of a removed competing risk is a counterfactual quantity, not a directly observed one, and the design pre-commits to reporting it with the same explicitness, namely stated assumptions, stated bounds, and a sensitivity analysis that shows how the decomposition moves when the cause-coding or the at-risk structure is perturbed. Section 5.4 and the analysis plan in Chapter 6 carry this commitment into the robustness battery. The confidence attached to the counterfactual reading is therefore moderate rather than high by deliberate calibration: moderate because the counterfactual is constructed, and raised toward high only to the extent the sensitivity analysis shows the decomposition stable under perturbation.

## 5.3 Model specification

### 5.3.1 The primary archetype-stratified specification

The primary specification estimates, for each event k and each archetype stratum separately, the subdistribution hazard as a function of the instrument-side covariates (entry TRL of the least-mature sensor technology, instrument count, instrument mass, instrument power), the launch-side covariates (launch-vehicle class, the shared-versus-dedicated manifest indicator, and the provider-in-development indicator), and the controls (the complexity index after the complexity-based cost-estimating tradition [\[17\]](#ref-17), the estimating-optimism proxy constructed as the ratio of the confirmation-baseline schedule to a reference-class median schedule for similar-class missions [\[57\]](#ref-57), and calendar-period fixed effects to absorb era-specific acquisition policy). Estimating the model within each stratum is the operational form of the Callaway-Sant'Anna discipline: it produces the archetype-specific subdistribution hazards as separable blocks rather than as a pooled average [\[27\]](#ref-27). The instrument-side covariate set rests on the established finding that the maturity of a spacecraft's least-ready technology at authorization drives subsequent schedule slip [\[47\]](#ref-47), which is the substantive reason entry TRL is the leading instrument-side covariate rather than an arbitrary inclusion.

The choice to lead with entry TRL among the instrument-side covariates is a mechanism claim, and the design names the mechanism rather than asserting a bare correlation. The driver is a mission carrying a first-of-kind active sensor below its assumed maturity at KDP-B. The mechanism is that the least-mature sensor technology fails environmental test, cannot close its calibration budget, or runs long in maturation, so that instrument development becomes the binding path. The observable effect is an instrument-driven first-slip event that raises the instrument-slip cumulative incidence faster than the launch-slip cumulative incidence in the active-sensor stratum. The operational consequence is a committed-launch-date movement that accrues standing-army and rework cost against the confirmation baseline, and the strategic implication is that reserve and TRL gating should be steered to instrument maturation for this archetype. The TRL-schedule literature supplies the empirical backing for the driver-to-effect link, finding a measurable and nonlinear relationship between starting-TRL deficit and expected slip [\[47\]](#ref-47); the design's contribution is to carry that link into a cause-specific cumulative-incidence comparison rather than into a single pooled slip regression.

### 5.3.2 The pooled specification with interactions

The pooled specification is estimated alongside the stratified blocks and adds the archetype-by-TRL interaction and the archetype-by-shared-manifest interaction to a model fit across both primary strata. The archetype-by-TRL interaction is the compact test of the instrument-side half of the contribution: it asks whether the effect of an entry-TRL deficit on the instrument-slip subdistribution hazard is larger in the first-of-kind active stratum than in the heritage passive stratum. The archetype-by-shared-manifest interaction is the matching test of the launch-side half: it asks whether the effect of a shared-manifest assignment on the launch-slip subdistribution hazard is larger in the heritage passive stratum, consistent with the mechanism by which a direct-rebuild instrument carrying little maturation risk leaves the mission's first slip disproportionately launch-side. The pooled model is a test, not a magnitude source; its coefficients confirm or deny effect modification, while the stratified blocks supply the cumulative incidence functions a reserve decision reads.

The two specifications are complementary and the design reports both, because each guards against a failure of the other. The stratified blocks could mislead if a stratum is so small that its block is noisy; the pooled interaction borrows strength across strata and is more stable. The pooled interaction could mislead if the proportional-subdistribution-hazards assumption holds within strata but not across them; the stratified blocks do not impose cross-stratum proportionality. Reporting both, and reading a divergence between them as a reason to lower confidence, is the design's defense against either single specification carrying the whole claim.

### 5.3.3 Penalization and the events-per-variable cap

The cohort is small, on the order of thirty to sixty missions, with two competing events and an interaction, and a naive maximum-partial-likelihood fit on a covariate set of this size would be unstable and prone to separation. The design's default is therefore a ridge-penalized partial likelihood for the subdistribution model, with the ridge penalty selected by cross-validation, and with the number of free covariates capped by an events-per-variable rule so that no model is fit with more covariates than the realized event count can support. Penalized estimation in the competing-risks setting is well established for this small-sample, many-covariate situation, and the additive and flexible subdistribution-model literature provides the estimation machinery the design draws on where the proportional form needs relaxing [\[84\]](#ref-84), [\[20\]](#ref-20). The events-per-variable cap is the quantitative discipline that prevents the interaction-bearing model from overfitting: the cap is set in advance, the covariate set is trimmed to satisfy it, and the trimming order is pre-registered so that it cannot be tuned to produce a desired result.

The penalization choice interacts with the doubly robust robustness construction from Section 5.2.5. The penalized subdistribution model is the outcome-model route; the inverse-probability-of-archetype reweighting is the weighting route; and the design reports the archetype-specific hazards under both, treating agreement across routes as evidence of robustness to misspecification [\[118\]](#ref-118). The reweighting route is itself stabilized against small-sample instability by trimming extreme weights, a standard precaution that the design pre-commits to so that a handful of missions with extreme archetype-propensity scores cannot dominate the reweighted estimate.

A measurement-induced specification subtlety is worth recording because it bears on covariate handling. Entry TRL for older missions is unevenly recorded in TechPort, and some instrument-side covariates may be subject to detection-limit or below-threshold recording rather than clean measurement. The competing-risks literature has developed estimators for covariates subject to detection limits in exactly this regression setting [\[137\]](#ref-137), and the design adopts that handling for any instrument-side covariate whose record is censored at a measurement floor rather than dropping such missions, because dropping them would select the cohort on data completeness in a way correlated with mission era and therefore with the launch-market conditions the design is trying to control.

## 5.4 Threats to validity

### 5.4.1 Internal validity

The dominant internal-validity threat is cause-coding error. The two competing events are distinguished by attributing each first slip to a single dominant cause, and that attribution rests on reconciling the narrative cause statement in the cost-and-technical record with the independent narrative in the GAO assessment for the same project-year. If instrument and launch problems co-occur and the dominant cause is misattributed, the two events bleed into each other, and separability is either spuriously rejected, because true instrument slips are scattered into the launch category and vice versa, or spuriously accepted, because the coding rule itself imposes a separation the data do not support. This is the most consequential threat in the design because it strikes at the first identification claim directly. The mitigation has three parts, all pre-committed: cause-coding requires agreement across the two independent narrative sources, so that a high-confidence code is one both sources support; events whose cause cannot be reconciled across the two sources are flagged as un-codable rather than forced; and the analysis plan recodes the ambiguous events both ways and reports the result under both codings, so that a conclusion which flips between codings is reported as fragile rather than as confirmed. The confidence that cause-coding error is controlled is moderate, not high, because narrative attribution is irreducibly subjective; the two-source reconciliation reduces the error but cannot eliminate it, and the design says so.

The second internal-validity threat is endogenous TRL gating. NASA gates technology maturity at milestones, so the entry TRL recorded at KDP-B is not a passive observation but a quantity the program has already acted on, and a mission that anticipates instrument risk may have raised its instrument's maturity, delayed its confirmation, or descoped before KDP-B. Any of these would suppress the very instrument-driven slip the model is trying to measure, biasing the instrument-slip hazard downward and potentially masking a true dominance. The defense is to measure entry TRL at the moment of KDP-B, before the slip window opens, which fixes the temporal ordering, and to include descope history as a covariate where the record supports it, so that anticipatory descoping is observed rather than hidden. This is a partial defense and the design grades it as such: it converts an unobserved selection into a partly observed covariate, but it cannot recover the slip that a descope prevented from ever being recorded. The divergence between the cause-specific and subdistribution results is the additional diagnostic here, because a gating-induced suppression would show a characteristic pattern of a modest cause-specific rate alongside a cumulative incidence depressed by the competing structure.

### 5.4.2 External validity

The cohort is NASA Earth-observing missions of a specific era, roughly 1990 to the present, that reach KDP-B. The results, if found, do not transport to commercial Earth-observing constellations, which face different launch economics and a different manifest structure; they do not transport to planetary or astrophysics missions, whose instrument and launch risk structures differ; and they do not transport to non-US programs. The archetype contrast is specific to the active-versus-passive sensor distinction and should not be read as a general novelty result that would apply to novelty in the spacecraft bus, the ground system, or the operations concept. The design states this boundary as a delimitation rather than a defect: a narrow, well-bounded claim is more falsifiable and more useful to the program offices that would act on it than a broad claim that strains its evidence. The confidence in the external-validity boundary is high, because it follows directly from the cohort definition rather than from any estimated quantity; the design knows the boundary of its population with certainty even before estimation.

A further external-validity concern is era confounding. Launch-market conditions changed substantially across the studied period, with the entry of new providers and the retirement of established vehicles, so a cohort weighted toward one era could show a launch-side dominance that reflects the era rather than the archetype. The calendar-period fixed effects absorb common era shocks, and the archetype contrast is estimated within era to the extent the sample permits, but the small cohort limits how finely era can be partitioned, and the design acknowledges that a perfectly clean era-versus-archetype separation is not achievable on a cohort of this size. This is one reason the launch-side dominance claim for the heritage arm is graded as the more evidence-thin half of the contribution, to be confirmed on the cohort rather than asserted.

### 5.4.3 Construct validity
Two construct-validity concerns are central. The first is that "instrument-driven slip" is a constructed category resting on narrative attribution; it is a defensible but imperfect proxy for the underlying physical cause, and a milestone slip with genuinely mixed causes is forced into a single dominant-cause bin by the coding rule. The design reports the coding rule in full so that a reviewer can audit it, flags the un-codable events rather than hiding them, and treats the recoding sensitivity as a direct test of how much the constructed category drives the result. The second concern is that the archetype variable is a binary simplification of a continuous novelty spectrum: a sensor that is partly novel and partly heritage is not cleanly active-first-of-kind or passive-heritage, and the binary cut imposes a dichotomy on a continuum. The design's answer is the third archetype category for mixed or ambiguous sensors, which is excluded from the primary contrast but brought into the model as a separate stratum in robustness to confirm that the binary contrast is not forcing a spurious dichotomy. Confidence that the constructs are valid is moderate. The constructs are transparent and auditable, which is the strongest defense available for a narrative-derived category, but transparency is not physical correctness, and the design does not overclaim.

### 5.4.4 Statistical-conclusion validity

The sample is small and the model carries an interaction, so statistical-conclusion validity, specifically the risk of a Type II error that reports a non-rejection of the null when the design simply lacked power, is the central statistical concern. The design pre-commits to three defenses. It uses penalized estimation and the events-per-variable cap to keep the model from overfitting the small cohort. It reports confidence intervals and the realized power rather than point estimates alone, so that a non-rejection accompanied by wide intervals is distinguishable from a non-rejection accompanied by tight intervals around a null effect. It runs a power and minimum-detectable-effect analysis before estimation, not after, so that the design states in advance what effect it can and cannot detect. The decisive pre-commitment is interpretive: a non-rejection of the null accompanied by wide intervals is reported as inconclusive rather than as support for the null, which protects against over-claiming a null result from a small sample. This commitment is what makes the falsification rule honest, because it forecloses the move of treating a power failure as evidence for non-separability. The minimum-detectable-effect analysis is specified in Chapter 6 as a deliverable of the analysis plan; Chapter 5 records only the design commitment that it is run in advance and that its result conditions the interpretation of any non-rejection.

### 5.4.5 How the threat catalogue bears on the contribution

The threat catalogue is not a list of weaknesses but a statement of the residual risk that survives the design, and reading it that way clarifies what the design has and has not secured. The dissertation's argument holds that the problem is real, that it is material, that the design addresses the causal mechanism, that it improves on the alternatives, and that the residual risk is acceptable. The first four of these are carried by the earlier sections and the sibling chapters: the problem of undifferentiated slip is real and documented in the schedule-cost-coupling literature; it is material because reserve allocated against slip is a first-order cost driver; the competing-risks design addresses the mechanism by separating the two cause-specific hazards without contaminating one with the other; and it improves on the alternatives because a pooled slip regression and a naive Kaplan-Meier cannot decide separability or dominance and the Kaplan-Meier alternative is biased upward by construction [\[5\]](#ref-5), [\[140\]](#ref-140). This Section 5.4 addresses the fifth point, residual risk. The residual risk is concentrated in cause-coding error and small-sample power, and the design's claim is not that these risks are absent but that each is bounded by a pre-committed mitigation: two-source reconciliation and recoding sensitivity for cause-coding, and the in-advance power analysis with the inconclusive-non-rejection rule for power. The residual risk is therefore acceptable in the specific sense that it is bounded, stated, and tested, not in the sense that it is zero.

## 5.5 The computational and software plan

### 5.5.1 Estimation software

The competing-risks estimation is implemented in a reproducible scripted environment rather than by point-and-click, so that every cohort, coding, and model decision is captured in version-controlled code and the entire analysis can be re-run from the frozen cohort by an analyst with the data-use agreement. The subdistribution model is estimated by the Fine-Gray method as implemented in mature, peer-reviewed survival libraries; the original method and its standard implementations are the reference point [\[55\]](#ref-55), and the cumulative-incidence and Gray-test machinery is available in well-documented form, including the implementation that brought the cumulative-incidence estimator and the K-sample comparison into general analytic use [\[37\]](#ref-37). The cause-specific Cox models are estimated by standard partial-likelihood routines. The left-truncated subdistribution form is estimated using the method developed for right-censored and left-truncated competing-risks data, which is the form required by the spell-origin convention [\[143\]](#ref-143). The design does not commit to a single language at the level of dogma, because the methods are available in more than one ecosystem, but it does commit to an implementation whose competing-risks routines are validated against the published estimators rather than a hand-rolled fit, because a custom implementation of the subdistribution weighting would introduce an avoidable verification burden.

### 5.5.2 Penalized and reweighted estimation

The ridge-penalized partial likelihood is implemented through the penalized-regression extensions of the survival ecosystem, with the penalty path computed and the penalty selected by cross-validation on the partial-likelihood deviance. The cross-validation is stratified by event type so that the penalty is not selected on a fold that happens to contain few events of one cause, which would destabilize the rarer-event model. The inverse-probability-of-archetype reweighting route is implemented by first estimating an archetype-propensity model on the complexity and optimism covariates, trimming extreme weights at a pre-registered threshold, and then fitting the subdistribution model with the stabilized weights, so that the doubly robust comparison of the outcome-model and weighting-model routes is a direct re-run with the same outcome specification under two estimation strategies [\[118\]](#ref-118). The additive and flexible subdistribution-model implementations are held in reserve for the contingency, flagged in Section 5.1.4, that the proportional-subdistribution-hazards assumption fails and cannot be repaired by time-varying coefficients [\[84\]](#ref-84), [\[20\]](#ref-20).

### 5.5.3 Reproducibility, provenance, and pre-registration

The computational plan is built around three reproducibility commitments that bind the analysis before any model is fit. First, the cohort and the cause-coding are frozen before modeling: the assembled spells, the reconciled cause codes, the un-codable flags, and the archetype classifications are written to a versioned, read-only dataset, and the modeling code reads only that frozen dataset, so that no model decision can feed back into the coding. Second, the specification, the threshold settings, the events-per-variable cap, the penalty-selection procedure, and the falsification rule are pre-registered as a frozen block before estimation, so that the robustness battery cannot be tuned after the result is seen; the pre-registration block is the design's primary defense against the optimism-bias critique turning inward on the analysis itself, because a pre-registered analysis cannot be optimistically re-specified to rescue a desired finding [\[57\]](#ref-57). Third, the full provenance of the cohort, the data-source versions, the API pull dates for TechPort, and the software and package versions are logged alongside the frozen dataset, so that a later analyst can reconstruct not only the analysis but the exact data state it ran on. A frozen cohort, a frozen specification, and a logged provenance, taken together, are what make a design-stage dissertation honest: they commit the analysis in advance and in public, so that the eventual execution tests a pre-stated hypothesis rather than searches for a publishable pattern.

The pre-registration commitment also discharges the design's obligation to the reference-class-forecasting critique. The rival explanation that overruns are driven by systematic planning-stage optimism rather than by project-specific technical cause [\[57\]](#ref-57) applies not only to the missions under study but, as a methodological hazard, to the analyst. A pre-registered specification with a fixed falsification rule is the analytic analogue of reference-class discipline: it fixes the decision rule before the outcome is seen, just as reference-class forecasting fixes the estimate against an outside view before the inside view can bias it. The design therefore treats pre-registration not as a bureaucratic formality but as the structural feature that lets the dissertation make a falsifiable claim, because a claim that could be re-specified after the fact to avoid falsification is not falsifiable at all. Confidence that the computational plan secures reproducibility is high, because reproducibility is a property of the plan itself rather than of an estimated quantity, and the plan's frozen-cohort, frozen-specification, logged-provenance structure is fully specified in this chapter.

### 5.5.4 What the computational plan does not claim

In keeping with the dissertation's guardrail, the computational plan stops short of any executed result. It specifies the software, the penalization, the reweighting, and the reproducibility commitments, and it leaves the result tables specified but unpopulated for Chapter 6 to describe in their expected, illustrative form. No coefficient, hazard ratio, cross-validated penalty value, or cumulative-incidence plateau appears in this chapter, because none has been estimated on the full cohort, and reporting one would violate the design-stage honesty the dissertation is built on. The chapter's deliverable is a design that an analyst with the data-use agreement could execute exactly as written, returning a falsification decision against the pre-stated rule, and it is the completeness and pre-commitment of that design, not any number, that this chapter offers as its contribution.

## 5.6 Chapter summary

This chapter specified the full empirical design that licenses the dissertation's contribution and argued its identification formally. The estimator is the Fine-Gray proportional subdistribution hazard model, carried in the fixed notation \( \text{subhazard}_k(t \mid X) = \text{subhazard}_{k0}(t) \exp(X \, \boldsymbol{\beta}_k) \), estimated separately for the instrument-driven and launch-driven first-slip events, with the cause-specific Cox model and Gray's test run in parallel, and with the subdistribution form chosen as primary because the reserve-allocation decision turns on the cumulative incidence of each cause rather than on its instantaneous rate, and because the naive Kaplan-Meier alternative is biased upward by construction in the presence of a competing event [\[55\]](#ref-55), [\[5\]](#ref-5), [\[140\]](#ref-140). Identification rests on three argued claims, that the events are genuinely competing, that the archetype is a pre-outcome moderator, and that the covariates respect temporal ordering, each defended by stating its basis, the principle that licenses it, the limits within which it holds, and the objection it must answer, and disciplined by the Callaway-Sant'Anna refusal to pool heterogeneous hazards [\[27\]](#ref-27), [\[43\]](#ref-43) and by the Fogel counterfactual reading of the cumulative incidence function with its explicit-and-bounded-construction warning [\[82\]](#ref-82). The specification estimates archetype-stratified subdistribution hazards as separable blocks plus a pooled interaction test, under ridge penalization and an events-per-variable cap with a parallel doubly robust reweighting route [\[84\]](#ref-84), [\[118\]](#ref-118). The threat catalogue treats internal validity (cause-coding error and endogenous TRL gating), external validity (the era and program boundary and era confounding), construct validity (narrative attribution and the binary archetype), and statistical-conclusion validity (small-sample power with the inconclusive-non-rejection rule) as the residual risk that survives the design, each bounded by a pre-committed mitigation. The computational plan binds the analysis through a frozen cohort, a frozen and pre-registered specification, and a logged provenance, and it stops deliberately short of any executed result, in keeping with the design-stage guardrail. Chapter 6 takes this design and specifies the analysis plan: the fixed estimation sequence, the expected illustrative findings under the contribution and under the null, the pre-registered robustness battery, and the power analysis that conditions the interpretation of any non-rejection.


## Chapter 5 References

Citations in this chapter are numbered to the consolidated reference list in the Back Matter (Part I: References); each in-text marker links directly to its full entry there.


# Chapter 6: Analysis Plan and Expected Results

## 6.0 The chapter thesis

The analysis plan of this dissertation can be stated in one sentence before any of it is justified: the cohort will be run through a fixed six-step estimation sequence whose decision rule on the hypothesis is written down in advance, whose expected results are stated only as signed directions with named mechanisms rather than as numbers, and whose result tables are specified now and left deliberately empty until the cohort is assembled and the cause-coding is frozen. Everything in this chapter develops that sentence. It sets out the order of operations so that no analytic choice can be made after seeing an estimate; it states what each estimate would have to look like for the contribution to be confirmed and what it would look like for the null to stand; it reasons through why each expected sign should appear, tracing a driver to a mechanism to an observable effect rather than asserting a correlation; it designs the illustrative simulation that will demonstrate the estimator behaves as intended before it touches real data; and it interprets the cumulative incidence profile and the Gray's-test contrast as the decision-relevant objects they are. The single discipline that governs the entire chapter is that it is written at the design stage. No estimate has been produced. Every quantity named below is labeled as expected or as illustrative, and every result table is presented as a template with its cells unpopulated by design.

This framing matters because the credibility of a pre-registered design rests on whether the analyst has tied his own hands before the data can talk back. In the current state of NASA mission-slip analysis, slip is regressed as a single delayed outcome after the fact, with the analyst free to choose specifications until the result is congenial; Flyvbjerg's critique of cost-overrun practice is precisely that the latitude to sample data, remove outliers, and select baselines after the fact is what produces the field's century-long record of biased estimates [\[57\]](#ref-57). The desired state is an analysis whose every consequential choice, the threshold that defines a slip event, the events-per-variable cap, the penalty, the cause-coding, and above all the rule that decides between H0 and H1, is fixed in writing before the cohort is run, so that the eventual estimate can only confirm or falsify and cannot be tuned. The gap is that no such pre-registered competing-risks plan exists for mission schedule slip, because the competing-risks framing of slip is itself novel to this dissertation. Leaving that gap unfilled means that any future estimate, however carefully computed, would inherit the same suspicion of specification search that haunts the overrun literature. This chapter closes the gap by writing the plan down in full and in advance, and it does so honestly at the design stage, which means it specifies the analysis without executing it and reports no findings dressed as results.

The chapter proceeds from procedure to expectation to decision rule to diagnostics to power. Section 6.1 fixes the six-step estimation sequence and the illustrative simulation that validates the estimator. Section 6.2 states the expected, illustrative findings under H1, under H0, and under the partial-support pattern, with every number labeled illustrative and every result table left empty by design. Section 6.3 reproduces the falsification rule exactly as it is fixed in the dissertation bible. Section 6.4 specifies the pre-registered robustness battery. Section 6.5 sets out the minimum-detectable-effect analysis that is run before estimation and the in-advance commitment that a wide-interval non-rejection is reported as inconclusive rather than as support for the null.
## 6.1 The estimation procedure

The estimation procedure is a fixed sequence of six steps, ordered so that each consequential decision is made and frozen before the next step can see its consequences. The sequence is reproduced here verbatim from the dissertation's analysis plan, and the order is itself part of the pre-registration: a reviewer can check not only what was done but that it was done in the committed order.

### 6.1.1 The six-step sequence

**Step one. Assemble the cohort and freeze the cause-coding.** The cohort of NASA Earth-observing missions reaching Key Decision Point B (KDP-B) is assembled from the four named sources fixed in Chapter 4, and the two competing events, instrument-driven first slip and launch-driven first slip, are coded by reconciling the CADRe Part A narrative with the Government Accountability Office assessment narrative for the same project-year. The defining act of this step is that the cause-coding is frozen before any modeling begins. This ordering is not a convenience. It is the firewall that prevents the most dangerous form of specification search in a competing-risks design, the silent re-attribution of an ambiguous event to whichever cause makes the eventual hazard contrast cleaner. Once the coding is frozen, the events-per-variable accounting that constrains the rest of the design is known, because the number of cause-coded events of each type is fixed.

**Step two. Construct the archetype effect modifier and the entry-TRL covariate.** Each mission is classified from the NASA Instrument Cost Model taxonomy as carrying a first-of-kind active sensor or as a passive-radiometer heritage mission, with mixed or ambiguous sensors held in a third category for robustness only. The entry technology-readiness level of the least-mature sensor technology at KDP-B is taken from TechPort. Both constructs are built from records that predate the slip window, preserving the temporal ordering on which the identification argument of Chapter 5 depends. This step is separated from the modeling steps so that the effect modifier and the key instrument-side covariate are fixed as data, not chosen as the model is fit.

**Step three. Estimate the nonparametric cumulative incidence functions and run Gray's test.** For each cause within each archetype stratum, the cumulative incidence function (CIF) is estimated nonparametrically, and Gray's K-sample test is run for equality of the CIFs across the archetype strata. This step is placed before any regression. The nonparametric CIF is the assumption-light object: it requires only the competing-risks accounting, not the proportional-subdistribution-hazards assumption that the regression step will impose. Gelber and Gray established the normative reason for leading with the CIF: the cumulative incidence function, not the integrated cause-specific hazard, is the quantity that corresponds to a real-world probability once competing events are present [\[63\]](#ref-63). Ray's class of K-sample tests for comparing the cumulative incidence of a competing risk across groups is the instrument used for the formal contrast [\[115\]](#ref-115), and the Stata implementation documented by Coviello and Boggess supplies the practical estimator for the CIF and the K-sample comparison [\[37\]](#ref-37). Running this step first gives the separability question a first, model-free answer before any parametric structure is layered on.

**Step four. Estimate the cause-specific Cox and Fine-Gray models, first without and then with the archetype interaction.** For each event, the cause-specific Cox model and the Fine-Gray subdistribution hazard model are estimated, in that paired form, first in a specification without the archetype interaction and then in the pooled specification with the archetype-by-instrument-side interactions. The paired estimation is not redundancy. Latouche, Allignol, Beyersmann, Labopin, and Fine make the reporting standard explicit: a competing-risks analysis should report results on all cause-specific hazards and the cumulative incidence functions, because the cause-specific hazard answers the etiologic question of the instantaneous rate among those still at risk while the subdistribution hazard answers the predictive question of the cumulative probability in the population where competing events occur [\[79\]](#ref-79). Wolbers, Koller, Stel, Schaer, Jager, Leffondre, and Heinze reinforce that the analyst must match the hazard quantity to the inferential objective rather than report one as if it answered both questions [\[135\]](#ref-135). For this dissertation the policy question is predictive, so the Fine-Gray subdistribution hazard is the primary object, but the cause-specific Cox model is estimated alongside it because divergence between the two is itself informative and is reported rather than suppressed.

**Step five. Test H1 by the interaction and the dominant-hazard contrast.** The contribution hypothesis is tested by the significance and sign of the archetype-by-instrument-side interaction in the instrument-slip subdistribution model, and by the contrast of which cause is the dominant subdistribution hazard within each archetype stratum. The test is not a single p-value on a single coefficient. It is a conjunction: the interaction must carry the predicted sign, and the within-stratum dominance must reverse between the first-of-kind active stratum and the heritage passive stratum. This conjunctive structure is what makes the test demanding, and it is set out in full in Section 6.3.

**Step six. Run the pre-registered sensitivity analyses.** The robustness battery fixed in Section 6.4 is run last: the slip-threshold variation, the ambiguous-event recoding both ways, the penalty and events-per-variable variation, the removal and reintroduction of the estimating-optimism control, and the introduction of the mixed-sensor third stratum. These are run after the primary estimate, not interleaved with it, so that the primary result is committed before its robustness is probed and the sensitivity analyses cannot be used to select the primary specification.

### 6.1.2 Why the ordering is itself a method claim

The argument that the ordering matters deserves to be made fully, because it is the spine of the chapter: fixing the six steps in this order materially improves the credibility of any eventual estimate. The evidence is that the overrun-economics literature documents, repeatedly and across domains, that latitude in the estimating process is the proximate source of biased forecasts. Flyvbjerg's account names inconsistent baselines, idiosyncratic sampling, and outlier removal as the bad practices that the discipline must foreclose [\[57\]](#ref-57), and the reference-class-forecasting program that Flyvbjerg and colleagues built for Hong Kong's roadworks projects [\[61\]](#ref-61) and that Natarajan adapted for offshore oil and gas megaprojects [\[97\]](#ref-97) is, at bottom, an apparatus for removing analyst latitude by committing to an outside view before the project-specific estimate is formed. This evidence bears on the present design because a survival design with a small cohort, two competing events, and an interaction term is exactly the kind of analysis where latitude is most dangerous: the number of defensible specifications is large relative to the number of events, so the space of congenial results an unconstrained analyst could reach is correspondingly large. That danger is the events-per-variable concern that runs through the competing-risks methodology literature and that Chapter 5 made the basis for penalized estimation. One limit is protected here: pre-registration improves credibility, it does not manufacture power, and a small cohort that has been honestly pre-registered can still fail to detect a real effect. The ordering protects against false confidence, not against the genuine limits of the sample. The objection that must be acknowledged is that an overly rigid pre-registration can prevent the analyst from correcting a genuine error discovered mid-analysis, for instance a data-entry mistake in the cause-coding. The design answers this by distinguishing, in advance, between specification changes (forbidden after the freeze) and error corrections (permitted but logged), so that the firewall blocks tuning without blocking honesty. Confidence in this conclusion is high, because the value of pre-committing analytic choices is one of the better-evidenced propositions in the overrun-economics literature the dissertation draws on, and because the cost of the discipline is low at the design stage where nothing has yet been estimated.

### 6.1.3 The illustrative validation simulation

Before the estimator is applied to the real cohort, the analysis plan calls for an illustrative simulation whose sole purpose is to confirm that the estimating machinery recovers a known structure and that the Gray's-test and interaction diagnostics behave as the design assumes. This simulation is illustrative by construction. It uses synthetic data with parameters chosen by the analyst, not the real cohort, and it produces no finding about NASA missions; its output is a statement about the estimator, not about the world.

The simulation is designed as follows. A synthetic cohort of mission-development spells is generated with a sample size and an event split matched to the expected real cohort (an order of 30 to 60 missions, with the event counts that the power analysis of Section 6.5 treats as the planning values). Two competing events are generated from cause-specific hazards whose forms are set by the analyst so that, in the first-of-kind active stratum, the instrument-slip cause-specific hazard rises with a synthetic entry-TRL deficit while the launch-slip hazard does not, and, in the heritage passive stratum, the two hazards are set close together. The data-generating process is specified at the level of the cause-specific hazards rather than the subdistribution hazards, following the warning of Bonneville, de Wreede, and Putter that proportional subdistribution hazards generally cannot hold simultaneously for more than one cause, so that a simulation which imposes proportional subdistribution hazards on both competing events at once would be generating data from an internally inconsistent model [\[23\]](#ref-23). Generating from cause-specific hazards and then deriving the implied cumulative incidence functions is the internally consistent route, and it is the route this simulation takes. Tanaka's result on the rare-event approximation between the subdistribution hazard ratio and the cause-specific hazard ratio [\[129\]](#ref-129) serves as a sanity anchor: in the low-event-rate region the two ratios should approximately coincide, and the simulation checks that the estimated quantities respect this approximation where it is expected to hold, a cheap and informative test that the estimation code is wired correctly.

The simulation then runs the full six-step sequence on the synthetic data and checks three things. First, that the nonparametric CIFs recover the planted shape, with the instrument-slip CIF rising faster in the active stratum and the two CIFs converging in the passive stratum. Second, that Gray's test rejects equality of the instrument-slip CIFs across strata at the planted separation and fails to reject when the separation is set to zero, which establishes that the test has the size and power the design assumes at the planning sample size. Third, that the archetype-by-TRL interaction in the pooled Fine-Gray model recovers the planted sign and magnitude within simulation error, and that the ridge penalty selected by cross-validation does not shrink the interaction to undetectability when the planted effect is at the practically meaningful size. The simulation is run across a grid of planted effect sizes spanning zero to a value the spacecraft literature would consider large, so that the realized rejection rate as a function of planted effect is the simulated power curve that Section 6.5 reports as the design's minimum-detectable-effect basis.

The reasoning behind insisting on this simulation is direct. A competing-risks estimator on a small cohort with an interaction is a piece of machinery with many ways to be silently miswired, from an incorrect at-risk set in the subdistribution construction to a penalty that over-shrinks the parameter of interest. The mechanism by which a miswiring would damage the dissertation is that a miswired estimator could return a clean-looking null on the real cohort that is an artifact of the code rather than a property of the world, and at the design stage there would be no way to tell the two apart. Running the simulation first documents the estimator's behavior on a known structure before the real data are touched, so that a null on the real cohort can be read as a statement about missions rather than about software. The dissertation can then distinguish, in advance, between three readings of an eventual non-rejection: a true null, an underpowered cohort, and a broken estimator. The simulation is therefore not a flourish but a precondition for the honest interpretation of the real result, and the analysis plan treats it as a mandatory step that precedes any contact with the cohort. Confidence in the necessity of the simulation is high; confidence in any particular planted parameter is not claimed, because the planted parameters are illustrative analyst choices and are labeled as such.

## 6.2 Expected, illustrative findings

This section states the expected form of the results. It is the section where the design-stage discipline is most load-bearing, and so it carries the guardrail in the strongest form: no result below has been executed on the cohort, every numerical statement is an illustrative placeholder that specifies the form of the output rather than a finding, and every result table is presented as a template whose cells are deliberately left unpopulated. The purpose of the section is to commit, in advance, to what each pattern would look like, so that the eventual estimate can be matched against a pre-stated expectation rather than interpreted post hoc.

### 6.2.1 The expected pattern if H1 holds

If the contribution hypothesis holds, the expected pattern has four linked features, each stated as a direction with a named mechanism.

The first feature is that, in the first-of-kind active-sensor stratum, the instrument-slip cumulative incidence function would rise faster and reach a higher plateau than the launch-slip cumulative incidence function. The driver is a mission carrying a first-of-kind active sensor, a lidar or radar with no direct flight heritage, below its assumed maturity at KDP-B. The mechanism is that the least-mature sensor technology becomes the binding development path: it fails environmental test, cannot close its calibration budget, or its maturation simply runs longer than the schedule assumed, so that instrument development is the activity most likely to generate the first above-threshold schedule movement. The observable effect is that the instrument-slip CIF in this stratum accumulates probability earlier and to a higher level than the launch-slip CIF. The operational consequence is that the mission's committed launch date moves for instrument reasons, accruing the standing-army and rework cost that overruns the confirmation baseline. The strategic implication is that schedule and cost reserve for this archetype should be steered toward instrument maturation and toward TRL gating at KDP-B. This chain rests on the antecedent established in the spacecraft literature, that the technology readiness of a spacecraft's least-mature technology at authorization is a measurable driver of subsequent schedule slip, which Dubos, Saleh, and Braun document on a cross-mission dataset [\[47\]](#ref-47), and on the complexity-based cost-estimating relationships of Bearden that make complexity and design aggressiveness independent drivers of cost and schedule risk [\[17\]](#ref-17).

The second feature is the illustrative magnitude. If H1 holds, the instrument-slip subdistribution hazard ratio on the entry-TRL deficit would be the dominant and statistically distinguishable term in the active stratum, with a subdistribution hazard ratio above one for each level of TRL deficit. The value "above one" is labeled here as illustrative: it specifies the direction the hazard ratio would have to take, not a number the dissertation claims to have estimated. The form of the expected output is a subdistribution hazard ratio with a confidence interval that excludes one for the TRL-deficit term in the active stratum, and the cell where that number would appear is left empty in the result template of Section 6.2.4.

The third feature is the reversal in the heritage passive stratum. In the passive-radiometer heritage stratum, the two cumulative incidence functions would be closer together, with the launch-slip subdistribution hazard at least as large as the instrument-slip hazard, so that the dominance would reverse or vanish. The driver here is a direct-rebuild passive radiometer carrying little maturation risk because it has flown before. The mechanism is that, with the instrument path largely de-risked, the first above-threshold slip is disproportionately likely to originate on the launch side: a shared-vehicle anomaly, a manifest reshuffle, or a launch provider's own development slip, all exogenous to the spacecraft's readiness. The observable effect is that the launch-slip CIF in this stratum is at least as high as the instrument-slip CIF. The operational consequence is that reserve for this archetype should be steered toward manifest and provider dynamics rather than toward instrument maturation. The strategic implication is the archetype-conditional reserve posture that is the dissertation's policy payoff. The Landsat Data Continuity Mission account documented by Irons, Dwyer, and Barsi is the substantive grounding for the launch-driven pressure on heritage continuity missions [\[72\]](#ref-72); the dissertation reads that narrative as the qualitative analogue of the quantitative reversal it expects to find.

The fourth feature is the formal separability signal. If H1 holds, Gray's test would reject equality of the instrument-slip cumulative incidence functions across the two archetype strata, and the archetype-by-TRL interaction in the pooled subdistribution model would be statistically distinguishable from zero with the predicted sign. These are the two formal objects that convert the visual pattern of the CIFs into a testable contrast, and their expected behavior is the rejection of equality and a nonzero interaction, respectively.

### 6.2.2 The expected pattern if H0 holds
If the null holds, the expected pattern is the mirror image, stated with equal specificity so that the null is falsifiable in both directions. The two cumulative incidence functions would track each other within both archetype strata, so that neither cause dominates in either archetype. Gray's test would fail to reject equality of the instrument-slip CIFs across the strata, and the archetype-by-TRL interaction would be statistically indistinguishable from zero. On the mechanism reading of the null, slip behaves in the data as a single hazard whose cause attribution carries no archetype-conditional structure, despite its two nominal origins: either the instrument and launch causes are not separable in the historical record, or their separation does not depend on sensor archetype. The dissertation treats this as a reportable result, not a failure. A credible demonstration that slip is one hazard after all would change how reserve is modeled, since it would license the central, undifferentiated reserve pool that current practice already uses.

### 6.2.3 The partial-support pattern

A third pattern is anticipated in advance because it is genuinely possible and carries its own implication. The two slip causes could be separable while the archetype dependence fails. Gray's test could reject equality of the cause-specific or cumulative-incidence structures, establishing that instrument-driven and launch-driven slip are distinct competing risks, while the archetype-by-TRL interaction remains indistinguishable from zero, so that the dominance of instrument-driven slip does not differ between the first-of-kind active and heritage passive strata. This pattern rejects part of H1, the archetype-dependence half, while confirming the separability half. The dissertation pre-commits to reporting it as partial support rather than as either confirmation or null. The implication survives: if slip is two distinct hazards even without archetype dependence, then slip should be modeled as two hazards rather than one, which still moves the reserve-modeling architecture away from a single pooled delay variable, even if it does not license the archetype-conditional steering that full H1 would license. Stating this pattern in advance prevents the analyst from being forced, after the fact, to call a mixed result either a confirmation or a null.

### 6.2.4 The result-table templates, specified and left empty by design

The analysis plan specifies the result tables now and leaves their cells empty, so that the form of the eventual output is fixed before the numbers exist. Three templates are specified.

**Template T6.1, cumulative incidence by cause and archetype.** Rows are the two archetype strata (first-of-kind active; heritage passive). Columns are the estimated CIF plateau for instrument-slip and for launch-slip, each with a confidence band, and the Gray's-test statistic and p-value for the across-strata equality of the instrument-slip CIF. Every cell is left empty by design at the design stage.

| Archetype stratum | Instrument-slip CIF plateau (95% band) | Launch-slip CIF plateau (95% band) | Gray's test (statistic, p) |
|---|---|---|---|
| First-of-kind active | (to be estimated) | (to be estimated) | (to be estimated) |
| Heritage passive | (to be estimated) | (to be estimated) | (to be estimated) |

**Template T6.2, subdistribution and cause-specific hazard ratios.** Rows are the covariates of the primary specification (entry-TRL deficit, instrument count, mass, power, vehicle class, shared-manifest indicator, provider-in-development indicator, complexity index, optimism proxy). Columns are the Fine-Gray subdistribution hazard ratio and the cause-specific Cox hazard ratio, each for the instrument event and the launch event, with confidence intervals. The interaction rows (archetype-by-TRL, archetype-by-shared-manifest) are flagged as the parameters of interest. Every cell is left empty by design.

| Covariate | Instrument event: subdistribution HR (95% CI) | Instrument event: cause-specific HR (95% CI) | Launch event: subdistribution HR (95% CI) | Launch event: cause-specific HR (95% CI) |
|---|---|---|---|---|
| Entry-TRL deficit | (to be estimated) | (to be estimated) | (to be estimated) | (to be estimated) |
| Archetype-by-TRL interaction (parameter of interest) | (to be estimated) | (to be estimated) | (to be estimated) | (to be estimated) |
| Archetype-by-shared-manifest interaction | (to be estimated) | (to be estimated) | (to be estimated) | (to be estimated) |
| (remaining covariates) | (to be estimated) | (to be estimated) | (to be estimated) | (to be estimated) |

**Template T6.3, robustness summary.** Rows are the five pre-registered sensitivity analyses of Section 6.4. Columns record whether the H1 pattern (separability, archetype dependence, dominance reversal) is preserved under each perturbation. Every cell is left empty by design.

| Sensitivity analysis | Separability preserved? | Archetype dependence preserved? | Dominance reversal preserved? |
|---|---|---|---|
| Threshold 1 to 4 months | (to be filled) | (to be filled) | (to be filled) |
| Ambiguous-event recoding both ways | (to be filled) | (to be filled) | (to be filled) |
| Penalty and EPV variation | (to be filled) | (to be filled) | (to be filled) |
| Optimism-control removal | (to be filled) | (to be filled) | (to be filled) |
| Mixed-sensor third stratum | (to be filled) | (to be filled) | (to be filled) |

The discipline of specifying and emptying these tables is itself a substantive commitment. Committing the output format in advance protects against the form of reporting bias in which the table that is shown is the one whose cells came out favorably. The reporting-standards literature supports this. Austin and Fine give practical recommendations for reporting Fine-Gray analyses that center on reporting both hazard families and the absolute-risk quantities rather than selecting the more favorable [\[10\]](#ref-10), and Latouche and colleagues make the same point as a standard [\[79\]](#ref-79). The logic is straightforward: a result table whose structure is fixed before estimation cannot be the product of selecting which result to show. The limit is that fixing the table format does not fix the conclusion; the cells can still come out either way, and the design intends exactly that. The practice of pre-specifying output tables is a direct and well-understood guard against selective reporting.

### 6.2.5 Interpreting the cumulative incidence profile as the decision object

The cumulative incidence profile is the decision-relevant object, and the chapter interprets it as such from the outset. The reason the CIF rather than the hazard ratio is the decision object is that a program office allocating reserve at confirmation needs the absolute probability that a mission of a given archetype will experience an instrument-driven first slip by a given time, not the instantaneous rate among missions that have not yet slipped. Austin, Lee, and Fine make this the canonical reason to prefer the cumulative incidence function for predictive questions: the subdistribution hazard and its CIF answer the absolute-risk question that decision-makers actually pose, while the cause-specific hazard answers the etiologic question of the rate [\[9\]](#ref-9). The Fogel structural-decomposition reading, carried from Chapter 2, gives the CIF its counterfactual content. The archetype-specific cumulative incidence function answers what fraction of a given archetype's missions would experience instrument-driven slip first, in the world where launch-driven slip also competes for primacy, which is the counterfactual-bearing quantity that a reserve decision implicitly requires. Leunig's re-assessment of the Victorian social-savings method [\[82\]](#ref-82) is the methodological reminder that this counterfactual is constructed rather than observed and must therefore be reported with its assumptions and bounds made explicit, which is why the CIF is always reported with its confidence band and why the robustness battery perturbs the cause-coding that the CIF rests on.

The profile interpretation deserves one further clarification, because the time axis of the CIF is the development spell measured in months from KDP-B, and the choice of that time scale has consequences that the analysis plan must own. Vilsmeier, Buchele, Rehm, Unseld, Rothenbacher, and Beyersmann show that the choice between calendar time and time-since-origin can produce contradicting hazard estimates and even a sign reversal when the two scales are confused [\[132\]](#ref-132). The dissertation's time scale is time-since-KDP-B, the development-spell clock, not calendar time, and the analysis plan fixes this so that the CIF profile is read as the accumulation of slip probability over a mission's own development clock rather than over the launch-market calendar. The calendar-period fixed effects absorb era shocks as covariates; they do not become the time axis. Holding this distinction keeps the profile interpretation clean and prevents the launch-market era from masquerading as the development clock.

## 6.3 The falsification rule

The falsification rule is fixed in advance and is reproduced here exactly as it stands in the dissertation bible, because the credibility of the design rests on the rule being unchangeable after estimation. The rule is conjunctive, which is what makes it demanding.

Confirmation of the contribution requires all three of the following.

**Separability.** Gray's test rejects equality of the cumulative incidence functions across the archetype strata, or the cause-specific hazards differ. This is the requirement that instrument-driven and launch-driven slip are distinct competing risks rather than one hazard with a cosmetic cause label.

**Archetype dependence.** The archetype-by-instrument-side interaction is statistically distinguishable from zero with the predicted sign, meaning the instrument-slip subdistribution hazard exceeds the launch-slip subdistribution hazard in the first-of-kind active stratum but not in the heritage passive stratum. This is the requirement that the dominance is conditional on sensor archetype, which is the specific, decision-relevant content of H1.

**Robustness.** The pattern survives the optimism control and the ambiguous-event recoding. This is the requirement that the separation is not an artifact of the estimating process or of the cause-coding.

Failure of any one of the three falsifies the contribution. The conjunction is the point. A design that allowed any single one of the three to count as confirmation would be far easier to satisfy by chance or by specification search, and the dissertation deliberately raises the bar by requiring all three together.

A single, explicit small-sample caveat is built into the rule and is itself pre-registered. A non-rejection accompanied by wide confidence intervals is reported as inconclusive rather than as support for H0. This caveat is the hinge between Section 6.3 and Section 6.5, because whether a non-rejection counts as inconclusive or as a genuine null depends on the width of the intervals relative to the minimum detectable effect, which is the quantity the power analysis fixes in advance. The reason to pre-register this caveat rather than to decide it after seeing the intervals is that the temptation to read an underpowered null as a confirmation of H0 is strong and is precisely the kind of after-the-fact latitude the whole design is built to remove.

The logic of the falsification rule is worth making explicit because the rule is the chapter's central claim: this three-part conjunction, fixed in advance, is the correct decision rule for the contribution. Each of the three parts maps onto a distinct way the contribution could be wrong: the events could fail to be separable, the separation could fail to depend on archetype, or the apparent result could be an artifact of optimism or coding. A contribution that is true in the sense H1 intends must pass all three, while each of the three failure modes is individually sufficient to make the contribution false, so the conjunction is both necessary and sufficient for the claim as stated. This rests on the identification argument of Chapter 5, which established that archetype is assigned before the outcome, that the covariates are measured at or before KDP-B, and that the two events genuinely compete, so that the three tests are testing what they purport to test. The small-sample caveat is protected: the rule decides between confirmation and falsification only when the cohort has the power to distinguish them, and otherwise returns inconclusive, a third verdict the rule explicitly admits. The objection is that a conjunctive rule could in principle reject a contribution that is true on two of three counts; the dissertation accepts this as the deliberate cost of a demanding rule and notes that the partial-support pattern of Section 6.2.3 is exactly the reporting category for a contribution that is true on the separability count but not on the archetype-dependence count, so that a partial truth is recorded as partial rather than discarded. Confidence in the rule is high, because it is derived directly from the structure of the hypothesis rather than chosen for convenience, and because it was fixed before any data contact.

## 6.4 The pre-registered robustness battery
Five robustness analyses are fixed in advance so that the result cannot be tuned after the fact. They run as step six of the estimation sequence, after the primary estimate is committed. Each is specified here with the perturbation it applies and the mechanism by which that perturbation probes a specific threat.

**First, threshold variation.** The slip threshold is varied from one month to four months of net launch-date movement, and the dominance pattern must be stable across this range to count as confirmed. The threat this probes is that the two-month threshold fixed in the variable definition is an arbitrary cut that could be manufacturing the pattern. A threshold that is too low admits noise events whose cause-coding is least reliable, while a threshold that is too high discards genuine early slips. A pattern that is real should be visible across a band of reasonable thresholds; a pattern that appears only at one threshold is an artifact of the cut. A conclusion that holds only at two months and collapses at one or four is reported as fragile.

**Second, ambiguous-event recoding both ways.** Events whose cause cannot be reconciled across the CADRe Part A and GAO narratives are recoded both ways, once as instrument-driven and once as launch-driven, and the result is reported under both codings. The threat this probes is cause-coding error, which Chapter 5 named the dominant internal-validity threat. The un-codable events are exactly the ones most likely to bleed the two competing risks into each other, so recoding them to both extremes brackets the range of conclusions the coding uncertainty can support. A conclusion that flips between the two codings is reported as fragile; a conclusion that survives both is robust to the worst case of coding uncertainty.

**Third, penalty and events-per-variable variation.** The ridge penalty is varied and the events-per-variable cap is tightened and loosened to show the estimate is not an artifact of one regularization choice. The threat this probes is that the small cohort forces a regularization decision, and a result that depends on the exact penalty is a result driven by the analyst's regularization rather than by the data. Varying the penalty traces how much of the interaction estimate is signal that survives shrinkage and how much is an artifact of a particular penalty value. The discrete-time subdistribution-hazard formulation of Berger, Schmid, Welchowski, Schmitz-Valckenberg, and Beyersmann [\[20\]](#ref-20) is noted in the plan as an alternative estimation route available if the continuous-time penalized fit proves unstable on the small cohort, and its availability is part of why the penalty variation is a meaningful rather than a cosmetic check.

**Fourth, optimism-control removal and reintroduction.** The estimating-optimism control is removed and reintroduced to isolate whether the separation depends on it. The threat this probes is the leading rival explanation, that the apparent instrument-launch separation is an artifact of which missions were optimistically estimated rather than of genuine cause structure. If the separation survives the removal of the optimism proxy and is essentially unchanged by its reintroduction, then the separation is not being carried by estimating optimism. If the separation appears only when the optimism control is in the model, the rival explanation is doing the work and the contribution is not robust. This is the single most important robustness check for adjudicating the Flyvbjerg rival [\[57\]](#ref-57), [\[61\]](#ref-61), and its result feeds directly into the third clause of the falsification rule.

**Fifth, the mixed-sensor third stratum.** The third archetype category, missions carrying mixed or ambiguous sensors, is brought into the model as a separate stratum to confirm that the binary contrast is not forcing a spurious dichotomy onto a continuum. The threat this probes is construct-validity: the archetype variable is a binary simplification of a continuous novelty spectrum, and a result that appears clean only because the messy middle was excluded would be an artifact of the dichotomy. If the mixed stratum sits, as expected, between the two poles in its instrument-slip CIF, the continuum interpretation is supported and the binary contrast is a defensible simplification rather than a fabrication; if the mixed stratum behaves erratically, the dichotomy is suspect.

The choice to pre-register these five and to run them last is itself disciplined by the methodological anchors. Bonneville, de Wreede, and Putter caution that fitting multiple Fine-Gray models, one per competing cause, can be internally inconsistent because proportional subdistribution hazards generally cannot hold for more than one cause at once [\[23\]](#ref-23), and Austin, Putter, Lee, and Steyerberg show concretely that summing the subject-specific risks from two Fine-Gray models can exceed one [\[8\]](#ref-8). The analysis plan responds to both warnings by treating the cause-specific Cox models as the internally consistent backbone for any quantity that must cohere across both causes, and by reserving the Fine-Gray subdistribution model for the single-cause predictive statements where it is the right tool. The robustness battery is where this division is enforced: the penalty and recoding checks are run on both hazard families, and any place where the two families disagree is reported rather than reconciled by fiat, in keeping with the Latouche reporting standard [\[79\]](#ref-79). The Callaway and Sant'Anna discipline against letting a pooled coefficient stand for heterogeneous effects [\[27\]](#ref-27), and the doubly robust logic that motivates pairing an outcome model with a reweighting model [\[118\]](#ref-118), are realized in the third and fourth checks: the penalty variation guards the archetype-stratified estimates as separable building blocks rather than a regression-imposed average, and the optimism-control check is the reweighting-versus-outcome-model robustness that doubly robust estimation calls for. The synthesis literature on difference-in-differences methods that the same authors anchor [\[33\]](#ref-33), [\[134\]](#ref-134) is the source of the broader instruction the battery follows: that heterogeneity must be probed by re-estimation under alternative weightings rather than asserted away by a single specification.

## 6.5 Power and the minimum detectable effect

The power analysis is run before estimation, not after, and this ordering is the final piece of the pre-registration. The reason to compute power in advance is that the meaning of a non-rejection depends entirely on whether the cohort could have detected the effect it failed to find, and that question must be answered before the result is seen, or it becomes a post-hoc rationalization of whatever the cohort returned.

The procedure is as follows. Given the expected sample size of 30 to 60 missions and the expected event split between the two competing causes, the design computes the minimum detectable subdistribution hazard ratio for the archetype-by-instrument-side interaction at conventional significance and power. This minimum detectable effect is the smallest interaction the cohort could reliably distinguish from zero, and it is computed using the same simulated machinery as the validation simulation of Section 6.1.3: the synthetic cohort is generated at the planning sample size and event split, the interaction is planted at a grid of sizes, and the realized rejection rate as a function of planted size is the simulated power curve from which the minimum detectable effect is read. This simulation-based route is preferred to a closed-form power formula because the small cohort, the penalization, and the competing-risks structure together make the closed-form approximations unreliable, and because the rare-event regime in which Tanaka's approximation between the two hazard ratios holds [\[129\]](#ref-129) is exactly the regime a small cohort may sit in, so the power curve must be computed where the estimator will actually operate rather than read off an asymptotic formula.

The design then makes a commitment in advance that is the operational form of the small-sample caveat in the falsification rule. If the minimum detectable effect is larger than the effect the spacecraft literature would consider practically meaningful, the dissertation states in advance that a non-rejection of H0 will be reported as inconclusive rather than as confirmatory. The benchmark for "practically meaningful" is drawn from the magnitudes in the TRL-schedule literature: Dubos, Saleh, and Braun report that a one-level deficit in starting TRL maps to a measurable and nonlinear increase in expected schedule slip [\[47\]](#ref-47), and an interaction smaller than the effect implied by that relationship would be one the cohort cannot be expected to resolve and whose non-detection therefore carries no information about H1. Stating this benchmark in advance is what converts the small-sample caveat from a vague hedge into an operational rule: the dissertation knows, before estimation, what size of interaction it can and cannot detect, and it commits to reading a non-rejection in the undetectable range as silence rather than as evidence.

The mechanism reasoning for placing power before estimation is the same firewall logic that orders the rest of the plan. The driver is that a small observational cohort has limited power for an interaction term. The mechanism by which this would damage the dissertation is that an underpowered null could be misread as a confirmation of H0, manufacturing a false negative result that looks like a finding. The observable effect of computing power first is that the minimum detectable effect is fixed as a number before the result is seen, so the interval width of an eventual non-rejection can be compared against a pre-stated threshold rather than against the analyst's after-the-fact intuition. The operational consequence is that the three verdicts the design admits, confirmation, falsification, and inconclusive, are separated by a line drawn in advance. The strategic implication is that the dissertation cannot over-claim from its small sample in either direction: it cannot claim H1 from a chance rejection, because the falsification rule is conjunctive and robustness-gated, and it cannot claim H0 from an underpowered null, because the power analysis pre-commits to calling that null inconclusive. Confidence in the necessity of the pre-estimation power analysis is high; confidence in any particular minimum-detectable-effect value is not asserted, because that value is an output of an illustrative simulation at planning parameters and is labeled accordingly.

A final honesty about power closes the chapter. The launch-availability side of the structure is the thinner half of the evidence base, because launch-manifest-driven slip as a survival outcome is under-documented in the open literature relative to the instrument-side TRL-schedule relationship. The launch-slip arm of the cohort may therefore carry fewer well-coded events than the instrument-slip arm, which would lower the power for any statement about the heritage passive stratum specifically, where the launch-slip dominance is expected. The analysis plan flags this asymmetry in advance: the heritage-arm dominance claim is the more evidence-thin half of H1 and is the half most likely to return an inconclusive verdict on a small cohort. Park, Bakoyannis, Zhang, and Yiannoutsos provide the semiparametric machinery for cumulative incidence regression under interval-censored competing risks with missing event types [\[104\]](#ref-104), which is the relevant tool if the launch-side events prove to be more often interval-censored or missing a clean cause label than the instrument-side events, and the plan names it as the available extension for that contingency rather than discovering the need for it mid-analysis. Mozumder, Booth, Riley, Rutherford, and Lambert supply the calibration logic for cause-specific absolute risk under competing events [\[95\]](#ref-95) that would govern any external validation of the fitted CIFs, and Rufibach, Beyersmann, Friede, Schmoor, and Stegherr summarize the findings of a large methodological program on survival analysis for events with varying follow-up times [\[117\]](#ref-117), which is the regime a mission cohort with staggered KDP-B dates and varying launch readiness dates inhabits; the plan cites both as the standards any eventual execution will be held to. Carvalho's treatment of forecasting and financial predictability in complex infrastructure and energy projects [\[28\]](#ref-28) is noted as the cross-domain reminder that the predictability of cost and schedule in large technical projects is itself limited, so that even a fully powered and confirmed result describes a tendency rather than a deterministic law, and the reserve-allocation implications of Chapter 7 are framed as probabilistic steering rather than as certainty. The minimum-detectable-effect tables that operationalize all of this are specified in Appendix E and, consistent with the design-stage discipline of the whole dissertation, are presented as templates to be populated when the cohort is assembled, not as executed power computations.


## Chapter 6 References

Citations in this chapter are numbered to the consolidated reference list in the Back Matter (Part I: References); each in-text marker links directly to its full entry there.


# Chapter 7: Discussion

The central answer of this chapter is that the dissertation earns a decision-relevant payoff under every outcome the analysis plan can return, and that the design has been built so that no single result, including the null, leaves the program office without a usable lesson. If the contribution holds, NASA and the Jet Propulsion Laboratory gain an archetype-specific rule for steering scarce schedule and cost reserve: hold reserve and gate technology readiness against instrument maturation for first-of-kind active-sensor missions, and hold reserve against launch-manifest and provider dynamics for passive-radiometer heritage continuity missions. If the null holds, the field learns that schedule slip is one hazard after all, which retires a plausible and expensive hypothesis and tells reserve managers to keep holding a single undifferentiated pool. If only part of the contribution holds, the field still learns that slip is two separable hazards even where sensor archetype does not modify their dominance, which by itself is enough to change how slip is modeled. This chapter develops that answer along four axes: the operational and theoretical implications under both outcomes; the contribution returned to each anchor framework; the full engagement with rival explanations; and the explicit statement of external validity. None of the numbers discussed below are executed findings. Consistent with the design-stage guardrail that governs the whole dissertation, every quantitative pattern named here is the illustrative form of an expected result, not an estimate produced on the cohort.

The problem this chapter must close deserves to be named at the outset, because the discussion is where a dissertation either converts a method into a decision or fails to. The current state of NASA Earth-mission reserve practice is that schedule slip, the proximate driver of cost growth, is treated as a single delay variable and reserve is held centrally against an undifferentiated pool at confirmation. The desired state is a reserve posture steered to the dominant first-slip hazard for the mission archetype actually in front of the program office. The gap between the two is not the absence of any one ingredient but the absence of an interpretation that tells a manager what to do once the model has run. The consequence of leaving that gap open is concrete: a first-of-kind active-sensor mission that is under-reserved against instrument maturation overruns or descopes its science, and a heritage continuity mission that is under-reserved against manifest risk slips on a shared vehicle, and in each case the program office cannot say afterward which lever it should have pulled. The remainder of this chapter is the argument that the design closes that gap whatever the data return.

## 7.1 Implications under both outcomes

The discipline this section imposes on itself is that a credible design-stage dissertation must specify its implications before it sees its results, and must specify them for the outcome it hopes for, the outcome it fears, and the outcome in between. Stating the policy payoff only for the favorable outcome is a known failure mode of applied work; it converts a falsifiable claim into an advertisement. The analysis plan in Chapter 6 fixes three exhaustive outcome states, and each carries a distinct, pre-committed interpretation.

### 7.1.1 If H1 holds: archetype-steered reserve and gated technology readiness

If the contribution is confirmed, the operational implication is that schedule and cost reserve at confirmation should be allocated by sensor archetype rather than against a pooled slip estimate, with first-of-kind active-sensor missions reserving and gating primarily against instrument maturation and heritage passive-radiometer continuity missions reserving primarily against launch-manifest and provider dynamics.

The evidence for this implication is the archetype-specific cumulative incidence functions that the Fine-Gray estimator produces. Recall the notation carried from the prospectus: for event type k, the subdistribution hazard \( \text{subhazard}_k(t \mid X) = \text{subhazard}_{k0}(t) \exp(X \, \boldsymbol{\beta}_k) \) governs the cumulative incidence function \( \text{CIF}_k(t \mid X) \), the probability of a cause-k first slip by time t accounting for the competing event. Under H1 the expected, illustrative form of the result is that in the first-of-kind active stratum the instrument-slip CIF rises faster and reaches a higher plateau than the launch-slip CIF, while in the heritage passive stratum the two CIFs converge and the dominance reverses or vanishes. The CIF is the decision-relevant quantity because reserve is held against the cumulative probability of a slip by the launch date, not against an instantaneous rate; a program office does not allocate against \( \text{subhazard}_k(t) \) directly, it allocates against the chance that a slip of cause k will have happened by the time it needs the vehicle.

What connects this evidence to the prescription is that reserve, being finite and allocated once at confirmation under standing policy, is most efficiently placed against the cause with the largest cumulative incidence for the archetype at hand. If two causes of slip have measurably different cumulative incidence functions within an archetype, then a reserve dollar or a reserve month buys more risk retirement when placed against the dominant cause than when spread across both. This is the same logic that justifies any targeted contingency over a flat contingency: targeting beats spreading when the target is both identifiable in advance and materially larger.
That logic rests on the documented coupling of cost to schedule and on the documented sensitivity of that coupling to its drivers. Bitten and colleagues show that NASA science-mission cost and schedule growth respond to policy and programmatic changes, which establishes that the growth is steerable rather than fixed [\[22\]](#ref-22). Kipp and colleagues show that instrument schedule growth propagates into mission cost and schedule growth for the class of missions this dissertation studies. That finding is the empirical anchor for treating instrument maturation as a lever worth reserving against [\[77\]](#ref-77). Sobel and Tibor show that the contract instrument itself, firm-fixed-price versus other forms, measurably affects cost growth, which confirms that the apparatus of confirmation, reserve and gating included, is a real and consequential set of levers rather than a formality [\[123\]](#ref-123). Bearden's complexity-based cost-estimating relationships add a further point: the design must hold complexity constant, so that an archetype-steered reserve rule carries content that a complexity-steered rule alone cannot supply [\[17\]](#ref-17).

One condition is essential, and it is protected here rather than buried. The prescription is conditional on the archetype contrast surviving the controls for complexity and estimating optimism, and on the cumulative incidence separation remaining stable across the slip-threshold range fixed in Chapter 6. Confidence in the prescription is moderate at the design stage, not high, because no coefficient has been estimated. The prescription is a conditional plan, and its confidence will be set by the width of the estimated intervals and by whether the pattern survives the pre-registered robustness battery. The evidence that would raise confidence is a stable, sign-correct archetype interaction with intervals narrow enough to exclude the null across the threshold range. The evidence that would lower it is an interaction whose sign or significance flips when the optimism control is removed or when ambiguous slip events are recoded.

The standing objection is that even a confirmed archetype rule could be operationally inert. The population of future missions might be too small or too archetype-homogeneous for differential reserve to matter, or reserve might prove fungible across causes once held. The design partly answers this by noting that the Jet Propulsion Laboratory disproportionately leads first-of-kind active-sensor Earth missions, so the active-stratum prescription maps onto a recurring slice of the portfolio rather than a rare edge case. The fuller answer is that the prescription concerns where to gate technology readiness at Key Decision Point B, not only where to hold dollars, and gating is not fungible: a TRL gate placed on the least-mature sensor technology is a different and non-substitutable control from a manifest-margin policy on a launch service.

The operational content of the H1 prescription is therefore specific and falsifiable in its own right. For a JPL-led first-of-kind active-sensor mission, the leading indicator named by the design is the entry TRL of the least-mature sensor technology at KDP-B, and the implied management action is to gate that maturity and to hold instrument reserve against it. This is sharper than the current practice of holding undifferentiated reserve against a pooled slip estimate, and it follows from the archetype-specific cumulative incidence functions rather than from a general intuition that novel instruments are risky.

### 7.1.2 If H0 holds: slip is one hazard, a result worth having

If the null holds, the correct interpretation is not that the dissertation failed but that NASA Earth-mission schedule slip is, for reserve purposes, a single hazard. Reserve should then continue to be held against an undifferentiated pool, because the two candidate causes cannot be told apart in a way that would change the allocation.

The basis for an H0 reading would be the expected null pattern fixed in Chapter 6: the two cumulative incidence functions track each other within both strata, Gray's test fails to reject equality of the CIFs across archetype strata, and the archetype-by-instrument-side interaction is statistically indistinguishable from zero. The competing-risks literature gives the precise reading of this pattern. Latouche and colleagues insist that a competing-risks analysis report all cause-specific hazards and the cumulative incidence together, so that a non-separation is visible and is not mistaken for an artifact of looking at one quantity alone [\[79\]](#ref-79). Reporting both is what makes an H0 reading interpretable rather than a silence.

What makes this reading worth stating is that a falsifiable design must treat its null as informative, and a null here is informative because it retires a specific, costly hypothesis. If the data say the two slip causes are not separable, then the analytic and data-curation effort that a differential reserve regime would require is not justified, and the program office is told, on evidence, to keep its simpler pooled posture. A null that closes off a tempting but unsupported complication is a contribution to practice, not a void.

This rests on the warning, central to the competing-risks apparatus, that treating a competing event as ordinary censoring biases the naive estimate. Andersen and colleagues catalogue the possibilities and pitfalls of competing risks in exactly this spirit, showing that the wrong handling of a competing event manufactures apparent structure where none exists [\[5\]](#ref-5). The relevance is that an H0 result obtained from the correctly specified subdistribution model is more trustworthy as a null than a non-finding from a naive single-hazard regression, because the correct apparatus rules out the artifactual separation that a careless one could produce. An H0 from this design is therefore a stronger negative than an H0 from the methods it replaces.

The H0 reading holds within a small-sample caveat that the design pre-commits to in Chapter 6: a non-rejection accompanied by wide confidence intervals is reported as inconclusive rather than as positive support for H0. Confidence in a clean H0 reading is therefore conditional on the realized power, which is why the minimum-detectable-effect analysis is run before estimation rather than after. The evidence that would let H0 be read as a genuine non-separation rather than as a power failure is intervals tight enough to exclude a practically meaningful subdistribution hazard ratio for the archetype interaction; absent that, the honest report is silence, not confirmation of the null.

The objection an H0 reading must answer is that the null could be an artifact of cause-coding that smears the two causes together, so that a real separation is hidden by measurement rather than absent in the world. The two-source reconciliation of CADRe Part A and GAO narratives, and the ambiguous-event recoding sensitivity, are the design's answer: if the null survives both codings, it is more credibly a true non-separation; if it appears only under one coding, it is reported as fragile. The null, in other words, is itself subjected to the robustness battery, not accepted at face value.

A null is so often the harder result to report honestly, and this design makes it reportable. The temptation under a null runs two ways. The first is to over-interpret a wide-interval non-rejection as confirmation of H0, which would assert non-separation on the strength of insufficient power. The second is to keep searching for the specification under which a separation emerges, which would convert a null into a false positive by way of unrestrained model search. Both failure modes are foreclosed in advance. The minimum-detectable-effect analysis, run before estimation, fixes the threshold below which a non-rejection is to be called inconclusive rather than confirmatory, so the first failure mode cannot occur without violating a pre-commitment. The frozen cause-coding and the pre-registered specification foreclose the second, because the model that returns the null is the model fixed before the data were seen, not the survivor of a search across many. The value of an H0 reading is therefore proportional to the discipline of the pre-registration, and the dissertation's claim is that under that discipline a null is genuine knowledge: the reserve community is told, on correctly specified evidence rather than on the absence of a finding from a careless one, that the simpler pooled posture is the right one and that the cost of building and maintaining a differential reserve regime is not warranted by separable hazards that the data cannot resolve.

### 7.1.3 If only part holds: two hazards without archetype dependence

The intermediate outcome, separable events but no archetype dependence, is not a muddle to be explained away but a distinct and useful result: slip should be modeled as two hazards even where sensor archetype does not modify which one dominates.

The basis for this reading is the partial-support pattern named in Chapter 6, in which Gray's test or the cause-specific hazards establish that the instrument and launch events are genuinely distinct competing risks, while the archetype-by-instrument-side interaction is indistinguishable from zero, so the dominance of instrument-driven slip does not differ between the first-of-kind active and heritage passive strata.

What makes this a result rather than a failure is that separability and archetype dependence are logically independent claims, and confirming the first while rejecting the second is a coherent and informative state of the world, not a failed version of H1. H1 is a conjunction; its first conjunct can hold while its second fails. The decision-relevance of the first conjunct alone is that reserve modeling should carry two hazards rather than one, even if the same split applies to every archetype, because a two-hazard model forecasts the cumulative incidence of each slip cause and a one-hazard model cannot.

This is the same structural-decomposition logic that motivates the whole design, articulated through the Fogel lens and its modern illustration. Leunig's reassessment of the passenger social savings from Victorian railways demonstrates that decomposing an aggregate into separable channels yields knowledge even when the channels do not interact with some third factor in the way one hypothesized; the value of the decomposition is the channel-level measurement, which survives the failure of an interaction hypothesis [\[82\]](#ref-82). Transferred here, a confirmed separation of instrument-channel from launch-channel slip is valuable on its own terms, because it lets a forecaster estimate CIF_instrument and CIF_launch separately, regardless of whether archetype moves their ratio.

This reading holds within a stated limit: it still requires the separation to survive the robustness battery; a separation that depends on the optimism control or that flips under recoding is fragile under the partial-support outcome exactly as it would be under full H1. Confidence in the two-hazards-without-archetype reading is therefore moderate and contingent on the same diagnostics.

The objection is that two hazards with identical archetype loadings might be operationally equivalent to one hazard for reserve purposes, since if the split is the same everywhere a manager could hold a single pool sized to the combined incidence. The answer is that even an archetype-invariant split is forecastable as two series, and a program that wants to know not just how much it might slip but why it might slip, in order to act on the cause during development rather than only to hold money against it, gains from the decomposition. Knowing the cause early enables intervention; knowing only the pooled magnitude enables only reservation.

## 7.2 Theoretical contribution back to each anchor framework

The chapter thesis for this section is that the dissertation borrows three frameworks and returns something to each: it extends the Fine-Gray apparatus to a domain it has never touched, it realizes the Fogel structural-decomposition program on a problem where the counterfactual is sharply defined, and it enforces the Callaway-Sant'Anna refusal to pool heterogeneous effects in a setting where the temptation to pool is strong and the cost of pooling is a wrong reserve rule.

### 7.2.1 To the Fine-Gray competing-risks apparatus

The contribution to the competing-risks literature is the first application of the subdistribution-hazard apparatus to NASA mission schedule slip. This is not a trivial transplant. The apparatus is mature in biostatistics, where death from one cause competes with death from another, but its assumptions and its reporting standards have to be re-earned in a setting where the events are slip causes, the at-risk population is missions in development, and the competing structure is that a mission's first slip is coded as one cause or the other. The dissertation's theoretical move is to show that this setting satisfies the conditions the apparatus requires: the events genuinely compete, because a mission's first slip is one cause or the other and the occurrence of one alters the at-risk set for the other; the subdistribution form is the right primary estimator, because the policy question is predictive (what is the cumulative probability of each slip cause) rather than purely etiologic; and the cause-specific hazards must be reported alongside, per the standing guidance, so that the etiologic and predictive answers are both visible. Latouche and colleagues argue that reporting only one of the two families of hazard is a recurrent error in applied competing-risks work [\[79\]](#ref-79); the dissertation honors that by carrying both. Andersen and colleagues warn against the upward bias of treating a competing event as ordinary censoring [\[5\]](#ref-5); the dissertation's contribution is to make that warning concrete in a mission-engineering context where the naive Kaplan-Meier of "time to slip" would do exactly the prohibited thing and overstate the incidence of either cause. The return to the apparatus is therefore an existence proof that the competing-risks frame is well posed for mission schedule slip, together with a worked specification that future mission-acquisition researchers can reuse.

### 7.2.2 To the Fogel structural-decomposition lens

The contribution to the Fogel lens is to realize structural decomposition with an explicit, bounded counterfactual on a problem where the counterfactual quantity is unusually well defined. Fogel's signature move, illustrated in the social-savings tradition that Leunig extends and re-measures, is to refuse to leave an aggregate as a single cause and instead to decompose it into separable channels measured against a constructed counterfactual, while making the construction explicit and bounding it with sensitivity analysis [\[82\]](#ref-82). The dissertation's realization of this move is that the cumulative incidence function of a removed competing risk is the counterfactual-bearing quantity: CIF_instrument answers what fraction of an archetype's missions would slip first for instrument reasons in the world where launch slip also competes for primacy. This is a cleaner counterfactual than the transport social savings ever had, because the competing-risks formalism defines the removed-risk cumulative incidence exactly, rather than requiring a constructed counterfactual transport network. But the dissertation also imports Fogel's discipline along with his method. The Fogel social-savings decompositions were attacked precisely because the counterfactual was constructed rather than observed, and the defense was always explicitness and bounding. The dissertation takes the warning seriously by reporting the cumulative incidence of a removed risk as a counterfactual quantity, not an observed one, with stated assumptions, the slip-threshold sensitivity that bounds it, and the ambiguous-event recoding that perturbs the cause structure. The return to the Fogel lens is a demonstration that the structural-decomposition program transfers to mission acquisition and that, in this setting, the formal estimator supplies the bounded counterfactual the program demands.
### 7.2.3 To the Callaway-Sant'Anna heterogeneity discipline

The contribution to the Callaway-Sant'Anna lens is to enforce the refusal to let a pooled coefficient stand in for a heterogeneous set of underlying effects in a setting where the pooling temptation is acute. Their critique of two-way fixed effects is that under heterogeneity a single coefficient returns a contaminated weighted average of distinct effects, so the headline number may correspond to no causal quantity the researcher intended. The analogous error here would be to estimate a single instrument-slip hazard pooled across all Earth missions and report it as the instrument hazard, when the hazard differs sharply between first-of-kind active-sensor missions and heritage passive-radiometer missions. The dissertation enforces the discipline by making sensor archetype an explicit effect modifier, estimating the archetype-specific subdistribution hazards as separable building blocks, and requiring that any aggregate be a transparent, defensible weighting of those blocks rather than a regression-imposed average. The interaction term that operationalizes H1 is this enforcement made testable. The doubly robust logic of the same research program, which combines an outcome model and a weighting model so that consistency survives misspecification of either, motivates the parallel use of penalized regression and reweighting as a robustness check on the archetype-specific hazards. The return to the lens is an applied case where pooling would produce a materially wrong reserve rule, holding undifferentiated reserve against a pooled hazard for a portfolio whose archetypes face different dominant causes. This sharpens the stakes of the heterogeneity discipline beyond the difference-in-differences setting where it was first articulated.

A note on what the dissertation does not claim to return to the anchors. It does not extend the Fine-Gray estimator mathematically, propose a new social-savings formalism, or modify the doubly robust difference-in-differences estimator. The contribution is applied and integrative. It is the first place these three lenses are combined into a single identification strategy for one problem, and the theoretical payoff is the demonstration that they cohere: the competing-risks estimator supplies the decomposition the Fogel lens demands and the archetype stratification the Callaway-Sant'Anna lens demands, and the optimism-bias literature supplies the control that keeps the decomposition from being an artifact of the estimating process.

## 7.3 Rival explanations engaged in full

The thesis for this section is that the contribution survives only if it beats the alternatives, and that the design has been built to let each rival win if it is true. A separation of slip causes that cannot distinguish itself from estimating optimism, from general complexity, from a cause-coding artifact, from anticipatory descoping, or from a shifting launch market would not be a contribution. It would be a re-description of one of those rivals. Each rival is therefore named with its mechanism, the design feature that adjudicates it, and the confidence that adjudication warrants.

### 7.3.1 Estimating optimism as the true driver

The leading rival is that overruns, and therefore the slip that proxies them, are driven by systematic optimism in the planning-stage estimate rather than by project-specific technical causes. Flyvbjerg's account holds that optimism bias and strategic misrepresentation, not idiosyncratic technical trouble, explain the persistence of cost overrun across project types [\[57\]](#ref-57), and the demand-forecast-inaccuracy evidence reinforces that estimates are biased at the planning stage in a way that is systematic rather than random [\[58\]](#ref-58). The rival's mechanism is that a mission optimistically estimated at confirmation will appear to slip regardless of its instrument or launch particulars, so the apparent instrument-launch separation could be an artifact of which missions were optimistically estimated rather than of any real difference in slip cause. If this rival is correct, the separation should weaken or vanish once estimating optimism is held constant.

The design adjudicates this rival directly. The reference-class optimism proxy, constructed as the ratio of the confirmation-baseline schedule to a reference-class median schedule for similar-class missions, enters as a control, following the reference-class-forecasting logic that Flyvbjerg's program advances and that Baerenbold develops as a risk-reduction tool for large projects [\[13\]](#ref-13). The pre-registered robustness battery removes and reintroduces this control specifically to isolate whether the separation depends on it. The decision rule is stated in advance and is not negotiable after the fact: if the archetype separation disappears once estimating optimism is controlled, the rival wins and the contribution is not robust. Chen and colleagues deepen the optimism-bias account by distinguishing its mechanisms in transport cost overrun [\[34\]](#ref-34), and Lovallo, Cristofaro, and Flyvbjerg situate the de-biasing remedy in a three-stage governance process [\[88\]](#ref-88). The dissertation's posture toward this literature is not to dismiss it but to treat it as the rival most likely to be partly right and to control for it as carefully as the data permit. Confidence that the contribution beats this rival is, at the design stage, moderate and contingent on the removal-and-reintroduction test. The dissertation does not assert that optimism is irrelevant, only that the design can tell whether the separation is more than an optimism artifact.

### 7.3.2 Archetype as a mere complexity proxy

The second rival is that first-of-kind active sensor is not a distinct construct at all but a proxy for general mission complexity, so that any apparent archetype effect is really a complexity effect wearing a sensor label. The mechanism is that complex, aggressive missions slip more for many reasons at once, and active-sensor missions tend to be complex, so a correlation between the active archetype and instrument slip could be entirely mediated by complexity. Bearden's complexity-based cost-estimating relationships establish that complexity and design aggressiveness independently raise cost and schedule risk [\[17\]](#ref-17), which is why complexity must be held constant before any archetype claim can stand.

The design adjudicates this rival by including the Bearden complexity index as a control in every specification, so that the archetype contrast is estimated within complexity rather than across it. If the complexity index absorbs the archetype effect, the rival wins: the dissertation would then report that what looked like a sensor-archetype hazard is a complexity hazard, and the reserve prescription would revert to a complexity-steered rule rather than an archetype-steered one. This is a clean, pre-committed adjudication, and its honesty is that it can take the contribution away. Confidence that archetype carries information beyond complexity is moderate at the design stage. It will be set by whether the archetype interaction remains distinguishable from zero with the complexity index in the model, and the dissertation states in advance that an archetype effect that vanishes under the complexity control is to be reported as a complexity effect, not defended.

### 7.3.3 Cause-coding manufacturing the separation

The third rival is that the separation is manufactured by the cause-coding procedure itself, so that the analyst, by deciding whether each slip is instrument-driven or launch-driven from a narrative, imposes a distinction that the underlying physics does not support. The mechanism is that slip is often multi-causal, an instrument problem and a manifest problem can co-occur, and forcing each first slip into a single dominant cause could create two clean events out of one messy reality, generating an apparent competing-risks structure as a coding artifact.

The design adjudicates this rival with the two-source reconciliation and the ambiguous-event sensitivity. Each slip cause is coded by reconciling the CADRe Part A narrative with the GAO assessment narrative for the same project-year, so that a cause counts as high-confidence only where two independent records name it; events whose cause cannot be reconciled across the two sources are flagged and recoded both ways in sensitivity analysis. The decision rule is that a separation that appears only under one coding of the ambiguous events is reported as fragile, and a separation that survives both codings is more credibly real. This rival cannot be eliminated, because narrative attribution is irreducibly subjective, but it can be bounded, and the dissertation's posture is to report the bound rather than to claim the construct is clean. Confidence that the separation is more than a coding artifact is moderate and tied to the recoding sensitivity. The construct-validity limitation, that instrument-driven slip is a constructed category resting on narrative attribution, is acknowledged as a standing weakness rather than argued away.

### 7.3.4 Reverse causation through anticipatory descoping and gating

The fourth rival is the subtlest and deserves the most careful mechanism statement, because it threatens the contribution in the direction of a false null rather than a false positive. NASA gates technology maturity at milestones, so a mission that anticipates instrument risk may delay its own confirmation, lower its committed schedule ambition, or descope the troublesome instrument. Any of these anticipatory moves would suppress the very instrument-driven slip the model is trying to measure: the mission that would have slipped on its instrument instead never commits to the schedule that the slip would have violated, or removes the instrument that would have slipped. This is selection on the dependent variable through the gating process, and its effect is to bias the instrument-slip hazard downward and potentially to mask a true dominance, producing a spurious null or a spurious archetype-invariance.

The design adjudicates this rival in two ways. First, entry TRL is measured at the moment of KDP-B, before the slip window opens, so the covariate that drives instrument slip is recorded prior to the anticipatory behavior rather than after it. Second, the descope history is included as a covariate where the record supports it, so that anticipatory descoping is observed and modeled rather than silently absorbed into the at-risk structure. The honest limitation is that descope decisions are not always documented in a way the design can recover, so this rival is bounded rather than eliminated. The residual risk is that an undocumented anticipatory descope removes an instrument-slip event from the data, and the dissertation flags this as the rival most likely to bias toward H0 if it is operating. Confidence that the contribution is not undone by anticipatory descoping is moderate, and the design states that a null result should be read with this rival in mind: a non-rejection could reflect a true non-separation or a gating process that has already removed the instrument-slip events the model would have counted.

### 7.3.5 Era confounding of the launch market

The fifth rival is that the launch market changed substantially across the studied period, so a cohort weighted toward one era could show a launch-side dominance that reflects the era rather than the archetype. The mechanism is that manifest congestion, provider diversity, and launch-service reliability all shifted across the decades the cohort spans, so a heritage continuity mission flown in a congested-manifest era and an active-sensor mission flown in a different era would differ in launch-side slip for reasons that have nothing to do with their sensors. Zancan and colleagues document the evolution of space-sector governance from legacy to new-space models, which is the institutional backdrop for exactly this kind of era shift in launch availability [\[142\]](#ref-142); the launch market a 1990s continuity mission faced is not the launch market a recent mission faces.

The design adjudicates this rival with calendar-period fixed effects that absorb common era shocks, and by estimating the archetype contrast within era to the extent the sample permits. The limitation, stated plainly, is that the small cohort constrains how finely the era can be partitioned, so within-era estimation is feasible only coarsely; the dissertation acknowledges that era and archetype are partially confounded by the historical fact that mission archetypes were not evenly distributed across launch-market eras. The launch-availability side is, as the evidence-gap register notes, the thinnest domain-empirical facet of the whole design, because launch-manifest-driven slip as a survival outcome is under-documented in the open literature. The corpus carries the continuity-mission narrative through Irons and colleagues on the Landsat Data Continuity Mission [\[72\]](#ref-72) and the Landsat-science synthesis of Wulder and colleagues [\[139\]](#ref-139), but few sources quantify launch-side slip directly. The dissertation therefore frames the launch-side dominance claim for the heritage arm as the more evidence-thin half of the contribution, to be confirmed on the cohort rather than asserted from the literature, and assigns it lower confidence than the instrument-side claim accordingly.

Taken together, the five rivals are not a defensive afterthought but the core of the case that the contribution improves on the alternatives. Each rival is given a mechanism, a pre-committed adjudication, and a confidence grade, and in three of the five the adjudication can take the contribution away. That asymmetry, a design that lets its rivals win if they are true, is what distinguishes a falsifiable contribution from a plausible story.

## 7.4 External validity

The thesis for this section is that the contribution, if found, is real but bounded, and that stating the boundary precisely is part of the contribution rather than a concession. A result whose scope is overclaimed is worse than a narrow result honestly bounded, because the overclaim invites misapplication to missions the design never studied.

The population to which any result transports is NASA Earth-observing missions that reached Key Decision Point B in roughly the 1990-to-present era, the cohort defined in the data chapter. The boundary has three faces. First, the result does not transport to commercial Earth constellations, which face different launch economics, different manifest dynamics, and different instrument-heritage structures than the NASA missions in the cohort; the new-space governance shift that Zancan and colleagues document means that the launch-availability hazard in particular is era- and sector-specific and cannot be carried across the public-commercial boundary [\[142\]](#ref-142). Second, the result does not transport to planetary or astrophysics missions, whose instrument and launch risk structures differ from Earth observation in ways the design does not model; an active-sensor lidar on an Earth mission and an instrument on a planetary mission do not share the same maturation or manifest physics, and the archetype taxonomy is calibrated to Earth-observing sensors. Third, the archetype contrast is specific to the active-versus-passive sensor axis and should not be read as a general novelty result; novelty in a bus subsystem, an avionics suite, or a ground system is a different effect modifier that would require re-estimation, and the dissertation claims nothing about it.

The decadal context sharpens why the bounded result still matters. Miner and colleagues argue that the Earth-observation needs of a warming world constitute a decadal survey without analogs, meaning the future Earth-mission portfolio will carry first-of-kind active sensors at a rate the historical cohort only begins to sample [\[93\]](#ref-93). This cuts two ways for external validity. It strengthens the relevance of the active-stratum prescription, because the archetype the dissertation flags as instrument-slip-dominant is precisely the archetype the coming portfolio will field more of. But it also warns that the cohort underrepresents the future, so a hazard estimated on historical active-sensor missions may not transport cleanly to a future in which such missions are more numerous, more ambitious, and built under different maturation and launch conditions. The honest statement is that the design measures the hazard structure of the missions that have flown, offers it as the best available evidence for the missions about to fly, and flags the extrapolation as an extrapolation rather than a guarantee.
Two further validity limits, carried from the design chapters, bound the result internally as well as externally. The construct-validity limit is that instrument-driven slip and the binary sensor archetype are constructed categories resting on narrative attribution and on a dichotomization of a continuous novelty spectrum; they are defensible but imperfect proxies, reported with their coding rules so a reviewer can audit them. The statistical-conclusion-validity limit is that the cohort is small and the model carries an interaction, so a non-rejection of the null must be read against the pre-computed minimum detectable effect, and a wide-interval non-rejection is reported as inconclusive rather than as support for H0. These two limits are why the external-validity statement is paired with a confidence statement: even within the studied population, the dissertation's claims are conditional, design-stage, and counterfactual in the cumulative-incidence sense, and their confidence is set by the robustness battery and the realized power rather than asserted in advance.

One question the external-validity section owes an answer to is whether the contribution, even if confirmed, would move a reserve board to change its KDP-B posture. A finding that the instrument-slip cumulative incidence plateau is higher than the launch-slip plateau for first-of-kind active-sensor missions is a statistical result; converting it into a reserve decision requires a translation rule that the contribution does not supply but that the existing NASA cost-and-schedule guidance comes close to supplying already. Emmons, Bitten, and Freaner show, from the historical NASA cost and schedule growth record, that reserve guidelines calibrated to realized growth distributions are achievable and that the relevant quantity is the growth at a target cost-confidence level, typically the 50th or 70th percentile of the realized cost-growth distribution conditional on mission class [\[151\]](#ref-151). The archetype-specific cumulative incidence functions that the dissertation produces are the survival-analytic analogue of that conditional growth distribution: they give the probability that an instrument-driven slip will have occurred by a given development horizon, conditional on archetype and entry TRL. A program office that already uses a cost-confidence-level framework to set reserve could read the archetype-specific CIF plateau directly as an input to that framework, replacing the current class-level growth percentile with an archetype-and-TRL-conditional slip probability. The dissertation does not claim to redesign the reserve governance process; it claims to supply the conditional probability estimate the governance process would need to make its undifferentiated pool archetype-conditional. Whether NASA and JPL program offices would adopt that input given their existing acquisition gates and organizational incentives is an implementation question that sits outside the scope of this design-stage study but is named here so that a future study can close it.

The residual risk that remains after all of this is, by the design's own standard, acceptable, and naming why completes the chapter's argument. The problem is real, because cost growth is largely a monetized image of schedule slip and slip has structurally distinct instrument and launch origins. The problem is material, because reserve and the slip it guards against are first-order cost drivers and the instrument and launch levers are documented and steerable. The design addresses the causal mechanism, because the Fine-Gray apparatus with archetype as an explicit effect modifier separates the two cause-specific hazards without contaminating one with the other. The design improves on the alternatives, because a pooled slip regression or a naive Kaplan-Meier that treats a competing event as ordinary censoring cannot decide separability or dominance, and the rivals to the substantive claim each have a pre-committed adjudication that can take the contribution away. The residual risk is acceptable, because the results are stated as conditional and counterfactual, the cause-coding error is bounded by two-source reconciliation and recoding sensitivity, the small-sample power is stated in advance, and the estimating-optimism rival is explicitly controlled. What remains uncertain is not whether the design is sound but what the cohort will say when the pre-registered specification is finally run, and that uncertainty is the proper and intended state of a design-stage dissertation that has done its work honestly.

The discussion returns to where it began. The dissertation is built so that each of its three possible outcomes carries a usable lesson for the people who allocate the scarce schedule reserve that ultimately decides whether an Earth-observing mission overruns. If instrument-driven and launch-driven slip are separable competing risks whose dominance depends on sensor archetype, NASA and JPL gain an archetype-specific reserve and gating rule. If they are not separable, the field learns that slip is one hazard and the simpler posture is the right one. If they are separable but archetype-invariant, slip should still be modeled as two hazards. In every case the contribution is the honest design and the falsification conditions it fixes in advance, not a coefficient, and the value of the work does not depend on the data returning the answer the candidate hopes for.


## Chapter 7 References

Citations in this chapter are numbered to the consolidated reference list in the Back Matter (Part I: References); each in-text marker links directly to its full entry there.


# Chapter 8: Conclusion

## 8.0 The chapter thesis

The deliverable of this dissertation is a pre-registered design and a falsification rule, not a set of estimated coefficients, and that distinction is the whole point of the concluding argument. Stated in one sentence before it is defended: this dissertation contributes the first application of the Fine-Gray competing-risks apparatus to NASA Earth-mission schedule slip, reframing slip as two separable hazards (instrument-driven and launch-driven) whose relative dominance is conditioned by sensor archetype, and the thing handed over to a future analyst is the complete specification of how to test that reframing and the explicit conditions under which it would be confirmed or falsified [\[55\]](#ref-55), [\[27\]](#ref-27). Everything that follows restates that contribution with the precision a concluding chapter owes its reader, separates what the dissertation has actually established from what it has only proposed, and lays out the concrete path from the present design-stage artifact to an executed result on the full cohort.

A conclusion to a design-stage dissertation carries an unusual burden. It must resist two opposite temptations at once. The first is to quietly promote the illustrative expected findings of Chapter 6 into something that sounds like a result, borrowing the rhetorical confidence of an empirical conclusion without the evidence that would earn it. The second is to under-claim, to treat a design that has not yet been executed as if it had contributed nothing, when in fact a correctly specified and falsifiable design is itself a contribution that stands independent of which way the eventual estimation breaks. This chapter holds the line between the two. It claims exactly what a design has the standing to claim: that the question is well posed, that the estimator fits the question, that the identification strategy licenses the causal reading the question demands, that the falsification conditions are fixed in advance so the result cannot be tuned after the fact, and that the limitations are stated honestly enough that a reader can judge what an executed result would and would not be allowed to conclude. The current state is a literature that models Earth-mission slip as one undifferentiated hazard and a reserve practice that allocates against a single pooled slip pool at confirmation; the desired state is an archetype-specific basis for steering scarce reserve to the dominant first-slip hazard of a given mission archetype; the gap this dissertation closes is the integration of three previously disjoint literatures into one identification strategy; and the consequence of leaving that gap open is that program offices keep reserving against an aggregate they cannot decompose, unable to tell which lever they should have pulled when a mission overruns.

The chapter proceeds in five movements. Section 8.1 restates the contribution as a specification and a falsification rule. Section 8.2 isolates what stands even if the headline hypothesis is not confirmed, which is the part of the work most easily lost in a result-focused reading. Section 8.3 states the limitations honestly and without hedging. Section 8.4 lays out a concrete future-research program, the path from the data-use agreement to an executed and reported falsification decision and then to extensions. Section 8.5 closes.

## 8.1 Restatement of the contribution: a specification and a falsification rule, not coefficients

The contribution of this dissertation is a single falsifiable claim and the apparatus built to test it. The claim, carried verbatim from the prospectus and unaltered through every chapter, is H1: for Earth-science missions, schedule slip from instrument development and schedule slip from launch-vehicle availability are statistically distinct competing risks, and instrument-driven slip is the dominant subdistribution hazard for missions carrying first-of-kind active sensors but not for passive-radiometer heritage missions. The null, H0, is that the two slip sources are not separable competing risks: either their subdistribution hazards are statistically indistinguishable across the cohort, or the dominance of instrument-driven slip does not differ between the two archetypes. The contribution is not the assertion that H1 is true. It is the demonstration that H1 is the right question, that it is answerable with a named estimator on a nameable cohort, and that the answer is constrained in advance by a rule that cannot be relaxed once the estimation runs.

The methodological core of the contribution is the importation of the Fine-Gray subdistribution-hazard machinery into a domain that has never used it. The spacecraft cost-and-schedule literature has modeled slip as a single continuous outcome regressed on covariates such as the technology readiness level of a mission's least-mature technology at authorization [\[47\]](#ref-47), and that literature is genuinely informative about the marginal relationship between maturity and slip. What it cannot do, by construction, is separate the cause of slip into competing events with distinct hazards, because a single continuous outcome has no place to put the distinction between an instrument cause and a launch cause. The competing-risks apparatus supplies exactly that place. The subdistribution hazard for event type k governs the cumulative incidence function, the probability of a cause-k first slip by a given time accounting for the competing event, and it is the cumulative incidence function, not the Kaplan-Meier estimate that treats a competing event as ordinary censoring, that answers the predictive policy question a program office actually faces [\[55\]](#ref-55), [\[5\]](#ref-5). The claim that this is the first such application is a claim about a gap, and the gap is real: the corpus assembled for this dissertation is strong on the spacecraft TRL-schedule literature and strong on the competing-risks toolkit, and it contains no source that joins them. That absence is not a weakness of the literature search; it is the reason the dissertation exists.

The contribution is disciplined by two methodological lenses that keep the apparatus from being misused. The first, in the spirit of Fogel's structural-decomposition method, refuses to leave schedule slip as an aggregate attributed to a single dominant cause and instead decomposes it into separable channels measured against an explicitly constructed and bounded counterfactual [\[82\]](#ref-82). The cumulative incidence function of a removed competing risk is the counterfactual-bearing quantity: it answers what fraction of a given archetype's missions would have slipped first for instrument reasons in the world where launch slip also competes for primacy. Fogel's method also carries a warning the dissertation takes seriously, that a constructed counterfactual must be reported with stated assumptions, stated bounds, and a sensitivity analysis rather than asserted as observed, which is why the pre-registered robustness battery is part of the contribution and not an afterthought. The second lens, in the spirit of the Callaway and Sant'Anna refusal to let a pooled coefficient stand for a heterogeneous set of underlying comparisons [\[27\]](#ref-27), requires sensor archetype to enter as an explicit effect modifier rather than be collapsed into a single pooled instrument-slip hazard. The dominance claim in H1 is precisely an interaction: the instrument-slip subdistribution hazard exceeds the launch-slip subdistribution hazard in the first-of-kind active stratum but not in the heritage passive stratum. Reporting a single pooled instrument hazard would average across that interaction and report a number corresponding to no archetype any program office actually manages.

The falsification rule is the second half of the contribution and is fixed in advance with the same standing as the hypothesis. Confirmation requires all three of: separability, meaning Gray's test rejects equality of the cumulative incidence functions across archetype strata or the cause-specific hazards differ; archetype dependence, meaning the archetype-by-instrument-side interaction is statistically distinguishable from zero with the predicted sign; and robustness, meaning the pattern survives the optimism control and the ambiguous-event recoding. Failure of any one of the three falsifies the contribution, with the small-sample caveat that a non-rejection accompanied by wide intervals is reported as inconclusive rather than as support for H0. This rule is what converts the dissertation from a plausible story into a falsifiable design. Treating the rule as a contribution in its own right is warranted because it is what a future analyst inherits: it tells them not only what to estimate but what conclusion each estimation outcome is permitted to license, before they see the data. The principle is the standard one for pre-registration, that a decision rule fixed before estimation cannot be reverse-engineered from a result the analyst wishes to report, and the limit is that the rule binds only an analyst who chooses to honor it, so its value depends on the future execution being conducted in the same pre-committed spirit in which the design was written. Confidence that the design is correctly specified is high; confidence in any particular outcome of executing it is, by the design's own honesty, withheld until the estimation is run.

## 8.2 What stands even if the hypothesis is not confirmed

The most important thing a design-stage conclusion can establish is what survives a null result, because a design whose entire value is contingent on its preferred hypothesis being true is a weak design. Three things stand regardless of which way the eventual estimation breaks, and they are worth stating in order of increasing strength.

The first thing that stands is the reframing of slip as a competing-risks process, independent of the archetype-dominance claim. There is a partial-support pattern, anticipated in Chapter 6 and in the prospectus, in which the two slip causes turn out to be separable competing risks but the archetype does not modify the dominance in the predicted direction. Under that pattern, part of H1 fails (the dominance conditioning) while the more fundamental half is confirmed (the two slip causes are distinct hazards rather than one). This is not a failure of the dissertation; it is a result with its own policy consequence. If slip is two hazards even without archetype dependence, then reserve modeling should still be built on two cause-specific hazards rather than on a single pooled slip estimate, because a pooled estimate that mixes two distinct hazards is a worse predictor than two estimated separately even when their relative magnitudes do not depend on archetype. The mechanism is concrete: a pooled hazard regressed on covariates assigns a single coefficient to a covariate that may drive the instrument channel and the launch channel in opposite directions, so the pooled coefficient is a contaminated weighted average in exactly the sense the heterogeneity lens warns against [\[27\]](#ref-27). Separating the two hazards removes that contamination whether or not the separation is archetype-conditioned. The reframing therefore carries value down to the weaker of the two partial-support patterns, and only the full H0, in which the two causes cannot be separated at all, removes it.

The second thing that stands is the honest design itself, considered as a contribution to how the field studies mission slip. The dissertation specifies a two-source cause-coding procedure that reconciles the CADRe Part A narrative with the GAO assessment narrative, flags un-codable events rather than forcing them into a bin, and bounds the residual coding error with a recoding sensitivity that reports the conclusion under both codings of every ambiguous event [\[82\]](#ref-82). It specifies an events-per-variable cap and a ridge-penalized partial likelihood so that a small cohort does not produce an over-fit hazard model presented with false precision. It pre-commits to a power analysis run before estimation, so that a non-rejection can be diagnosed as either true non-separability or insufficient power by the width of the reported intervals rather than being silently read as support for the null. None of these design elements depends on H1 being true. They are reusable: an analyst studying a different decomposition of mission slip, or the same decomposition on a different mission class, inherits a coding manual, a measurement table, and a falsification rule that they can adapt. The design is, in this sense, a methodological template for cause-decomposed survival analysis of acquisition outcomes, and the template stands whether or not the first application of it confirms its motivating hypothesis.

The third thing that stands is the value of a clean negative result, which a design-stage dissertation is unusually well positioned to deliver because it has pre-committed to interpreting one. If the executed estimation returns the full H0, that the two slip causes cannot be statistically separated, with intervals tight enough that the non-rejection is well-powered and informative, then the field has learned that slip is one hazard after all, and that reserve held centrally against an undifferentiated pool is the correct posture rather than a lazy one. That is a result worth having, and it is worth having precisely because the present practice of central pooled reserve would then be vindicated rather than merely assumed. The case for valuing the negative result is that the current practice is currently un-tested: nobody has shown that pooling is wrong, but nobody has shown it is right either, and a falsifiable design that returns a tight non-rejection converts an untested convention into a tested one. The principle is the basic asymmetry of pre-registered inference, that a result interpreted under a rule fixed in advance is informative whichever way it breaks, while a result interpreted after the fact is informative in only the direction the analyst hoped. The limit, carried from the falsification rule, is that the negative result is only clean if the intervals are tight; a wide-interval non-rejection on a thirty-to-sixty-mission cohort is inconclusive, not confirmatory of H0, and the dissertation says so in advance precisely so that it cannot later be read as a confirmation it has not earned. Confidence that a tight non-rejection would be informative is high; confidence that the cohort is large enough to produce a tight non-rejection is moderate, which is itself a limitation and is treated as one below.

## 8.3 Limitations, stated honestly

The limitations of this dissertation are not incidental qualifications to be minimized; they are the boundary of what an executed result would be allowed to claim, and stating them plainly is part of the design's honesty. Four limitations are first-order.

The first is cohort size. The target is NASA Earth-observing missions from roughly 1990 to the present that reach Key Decision Point B, an expected sample on the order of thirty to sixty missions. This is small for a survival model carrying two competing events and an archetype interaction, and it is the central statistical-conclusion-validity concern. The consequence is direct: the number of free covariates is capped by an events-per-variable rule, the estimation is ridge-penalized to stabilize it, and the design pre-commits to reporting realized power and interval width so that a non-rejection of H0 is interpretable rather than ambiguous. The mechanism by which small samples threaten the contribution is that the dominance claim lives in an interaction term, and interaction terms are the least powered quantities in any small-sample model, so the design's most decision-relevant parameter is also its most fragile. The honest statement is that the cohort may simply be too small to resolve the archetype interaction with confidence, in which case the dissertation's own falsification rule directs that the result be reported as inconclusive rather than dressed as either confirmation or refutation. No amount of estimator sophistication manufactures power that the sample does not contain.
The second limitation is the narrative basis of the cause-coding. "Instrument-driven slip" and "launch-driven slip" are constructed categories resting on the reconciliation of two narrative attributions, the project-authored CADRe Part A and the auditor-authored GAO assessment. Narrative attribution is subjective even when two narratives agree, and a slip with genuinely multiple simultaneous causes can be coded to a single dominant cause that the record names but the physics does not cleanly support. The two-source reconciliation improves on single-source coding because the two sources carry opposite directional biases: a project may under-attribute slip to its own instrument, an auditor may over-attribute to management and contracting. An event coded the same way by both is therefore robust to either bias alone [\[22\]](#ref-22). The reconciliation reduces the subjectivity rather than removing it, which is why construct validity for the slip-cause categories is reported as defensible-but-imperfect, why the un-codable flag is mandatory, and why the recoding sensitivity that reports the conclusion under both codings of every ambiguous event is a required diagnostic rather than an optional one. The category is a proxy for the underlying physical cause. It is a good proxy, but it is a proxy, and any executed result inherits that gap between the coded cause and the true cause.

The third limitation is restricted and uneven data. The CADRe records and the restricted NICM instrument parameters are available only to analysts under a NASA or FFRDC data-use agreement, so the full analysis is reproducible only by such analysts; a reader without the agreement can reproduce the public arm (the GAO narratives, the TechPort entry-TRL covariate, the public launch records) but not the project-side narrative or the restricted parameters. TechPort technology-readiness histories are uneven in completeness for older missions, and that incompleteness is correlated with CADRe completeness, so the oldest era of the cohort carries the most measurement uncertainty in both the cause narrative and the entry-TRL covariate at once. A further data limitation is that the launch-availability side is the thinnest domain-empirical facet: launch-manifest-driven slip as a survival outcome is under-documented in the open literature, so the launch-side construct leans on the Landsat continuity narrative and on the GAO and CADRe primary records rather than on a quantified launch-slip literature [\[72\]](#ref-72). By construction, then, the launch-side dominance claim for the heritage passive arm is the more evidence-thin half of H1, to be confirmed on the cohort rather than supported by prior quantified work.

The fourth limitation is that the cohort is observational and the design is therefore exposed to confounding it can control but not eliminate. Sensor archetype is not randomized to missions; first-of-kind active sensors are assigned to missions for reasons (scientific ambition, programmatic priority, institutional capability) that may themselves correlate with slip risk through channels other than the instrument maturation the dissertation means to isolate. The complexity index after Bearden and the reference-class estimating-optimism proxy are the design's defenses against the two most plausible confounders, complexity and planning-stage optimism, but they are controls, not exogeneity, and a confounder unmeasured by either would bias the archetype contrast. The left-truncation convention of dating the spell origin at KDP-B is a related construct limitation: some missions incur schedule risk during pre-formulation study before the formal record opens, so the KDP-B origin is a defensible convention for consistency rather than a true zero of risk, and the archetype contrast is a comparison of naturally occurring groups whose pre-KDP-B histories the design observes only partially. These limitations bound external validity to NASA Earth-observing missions of the studied era and explicitly do not extend the result to commercial constellations, to non-Earth science missions, or to non-US programs.

## 8.4 A concrete future-research program

The dissertation's design-stage status is not a permanent condition. It is the first phase of a research program whose remaining phases are specified concretely enough to be executed. The program has five steps, in fixed order, and the order matters because each step gates the next.

The first step is to secure the CADRe data-use agreement and the access to the ONCE database and the restricted NICM parameter records. This is a data-access dependency rather than a scientific one, but it is the binding constraint: until an analyst has the agreement, the project-side cause narrative and the restricted instrument parameters cannot be read, and the cohort cannot be fully assembled. The public arm of the data (the GAO assessment narratives, the TechPort entry-TRL histories, the public launch records) can be assembled in advance of the agreement, so the program can begin the cohort scaffolding while the access is pending, but the cause-coding cannot be completed until both the project-authored and auditor-authored narratives are in hand. The step is named first because it is the longest-lead item and because mis-sequencing it would strand the rest of the program.

The second step is to assemble the cohort and freeze the cause-coding before any modeling is done. The freezing is the methodological heart of the execution. The two competing events are coded by reconciling CADRe Part A with the GAO narrative for each project-year, un-codable events are flagged, and the entire coding is locked before a single hazard model is estimated, so that no modeling result can feed back into the cause labels. The archetype variable is constructed from the NICM instrument taxonomy with the active-versus-passive and first-of-kind-versus-heritage rules applied as specified in the data chapter, and the entry-TRL covariate is constructed from TechPort. The output of this step is a frozen analysis dataset and a written cause-coding record auditable against the public GAO arm, the deliberate design choice that keeps the coding partially reproducible even by analysts without CADRe access.

The third step is to execute the pre-registered specification exactly as written, with no additions discovered in the data. The fixed sequence is: estimate the nonparametric cumulative incidence functions for each cause within each archetype stratum and run Gray's test for equality across strata; estimate the cause-specific Cox models and the Fine-Gray subdistribution models for each event, first without and then with the archetype interaction; test H1 by the significance and sign of the archetype-by-instrument-side interaction and by the contrast of dominant hazards across strata [\[55\]](#ref-55). The subdistribution model is primary because the policy question is predictive, but the cause-specific results are estimated and reported alongside it because the two answer complementary etiologic and predictive questions, and any divergence between them is itself informative rather than something to suppress. The power analysis is run before this step, not after, so that the interpretation of a non-rejection is fixed in advance.

The fourth step is to report the falsification decision exactly as the rule defines it, whichever way it breaks. If all three conditions (separability, archetype dependence, robustness) hold, the contribution is confirmed and the reserve-allocation implication follows: first-of-kind active-sensor missions, which JPL disproportionately leads, should steer schedule and cost reserve toward instrument maturation and treat entry TRL at KDP-B as the leading gating indicator, while heritage passive-radiometer continuity missions should reserve against launch-manifest and provider dynamics. If any condition fails, the contribution is falsified in the specific way the failed condition names, and the report states which condition failed and what that implies, including the partial-support pattern in which the two slip causes are separable but archetype does not modify the dominance. If the non-rejection comes with wide intervals, the report states that the result is inconclusive and that a larger cohort, not a different estimator, is the remedy. The commitment to report all of these outcomes symmetrically is what protects the eventual result from the optimism bias that the dissertation spends its design controlling for in others [\[57\]](#ref-57).

The fifth step is extension, conditional on the first four being executed. Two extensions are natural and were flagged in the design as out of the present scope. The first extends the archetype axis beyond the active-versus-passive sensor distinction to other novelty dimensions (a novel bus, a novel ground system, a first-use launch vehicle), re-estimating the competing-risks structure with the new effect modifier rather than assuming the sensor result transports. The second extends the mission class beyond Earth observation to planetary or astrophysics missions, whose instrument and launch risk structures differ and which therefore require re-estimation rather than generalization. Both extensions inherit the same falsification discipline: each is a new pre-registered test of a new conditional dominance claim, not a borrowing of the Earth-observation result. A third, more distant extension would relax the binary archetype into a continuous novelty measure and estimate the dominance as a function of position on that continuum, which would require a larger cohort than the present design assumes and is therefore deferred until the cohort can be enlarged across eras or across agencies under comparable cost-and-schedule recording.

## 8.5 Closing

This dissertation began from a coupling that the space-systems literature treats as nearly definitional: cost growth in NASA Earth-observing missions is in large part a monetized image of schedule slip [\[86\]](#ref-86). It then made one move that the literature had not made, which was to refuse to leave slip as a single aggregated hazard and instead to model it as two competing risks with different owners, different physical drivers, and different policy levers, asking whether the dominance between them depends on whether a mission carries a first-of-kind active sensor or a heritage passive radiometer. The estimator that fits that question is the Fine-Gray subdistribution-hazard model. The discipline that keeps it honest is the structural-decomposition logic that reads the cumulative incidence function of a removed risk as a bounded counterfactual [\[82\]](#ref-82) and the heterogeneity logic that refuses to pool a coefficient across archetypes that do not share it [\[27\]](#ref-27), and the identification strategy that licenses the causal reading rests on covariates and an archetype assignment all fixed at or before Key Decision Point B [\[47\]](#ref-47).

What is offered here is the complete design and not the executed result, and the dissertation has been explicit about that boundary at every step, labeling every numerical pattern illustrative and claiming no estimated coefficient as a finding. The honesty is not a hedge; it is the contribution's form. A falsifiable design with a pre-committed decision rule is worth more to the field than a tuned result, because it can be executed by anyone who secures the data and it will return an informative answer whichever way it breaks. If the plan is executed and H1 is confirmed, NASA and JPL gain a defensible, archetype-specific basis for allocating the scarce schedule reserve that ultimately determines whether an Earth-observing mission overruns, allowing reserve to be steered to instrument maturation for the missions that face that hazard and to launch-manifest dynamics for the missions that face the other [\[72\]](#ref-72). If the plan is executed and the null holds with tight intervals, the field learns that slip is one hazard after all and that central pooled reserve is correct rather than merely customary, a result worth having and one that the present convention has never actually been tested against. Either way, the question moves from assumed to answered. The design is finished; the execution awaits the data; and the rule for reading whatever the data say is fixed in advance, which is the most a design-stage dissertation can offer and exactly what this one set out to deliver. If the work that follows helps those who steward these missions to hold the right reserve against the right hazard, and so to keep faith with the public purpose the missions serve, it will have answered the duty that prompted it.



## Back Matter: References and Appendices

This chapter discharges two obligations that the body of the dissertation incurred but deferred: it makes every claim in Chapters 1 through 8 auditable by compiling the complete, real reference apparatus in a single consistent style, and it makes the design executable by a second analyst by recording, in operational detail, the measurement rules, the cause-coding manual, the archetype crosswalk, the frozen pre-registration block, the specified-but-unexecuted power tables, and the full source-provenance log. The problem these appendices solve is concrete. A design-stage dissertation asks to be believed on the strength of its specification rather than its results, and a specification that lives only in prose is not yet a protocol; the current state is a complete argument with its operational backbone implicit, the desired state is a protocol an analyst holding the CADRe data-use agreement could run without reconstructing a single coding decision, and the gap between them is exactly the material assembled below. The consequence of leaving that gap open is that the pre-registration discipline the dissertation claims for itself would be unenforceable, because there would be no frozen artifact against which to check whether the executed analysis matched the promised one. Every reference in the list that follows exists in the project corpus and resolves to a live DOI or a resolvable repository URL; no entry is invented, and the numbering is the canonical key for the in-text citations throughout the dissertation.


## Part I: References

The reference list is compiled from the project corpus (149 entries) in a single IEEE-derived style, ordered alphabetically by first-author surname, with corporate-authored and repository-only sources interleaved alphabetically by title. Each entry carries a clickable DOI (rendered as a `doi.org` link) or, for NASA Technical Reports Server, OpenAlex, arXiv, and NIST records that have no assigned DOI, a clickable resolvable URL. Anchor spans (`<span id="ref-N"></span>`) at the head of each entry match the in-text citation markers used in the preceding chapters, each rendered as a clickable link to its entry, so that every numbered citation resolves to its full entry here. The dissertation does not cite all 149 entries in its running text; the corpus is the vetted population from which each chapter drew, and the full list is reproduced so that the evidentiary base is inspectable in its entirety rather than only through the subset that surfaced as inline marks.

<span id="ref-1"></span>[1] Dominic D. Ahiaga-Dagbui, Simon Smith, "Dealing with construction cost overruns using data mining," *Construction Management and Economics*, 2014. doi: [10.1080/01446193.2014.933854](https://doi.org/10.1080/01446193.2014.933854).

<span id="ref-2"></span>[2] Dominic D. Ahiaga-Dagbui, Simon Smith, "Rethinking construction cost overruns: cognition, learning and estimation," *Journal of Financial Management of Property and Construction*, 2014. doi: [10.1108/jfmpc-06-2013-0027](https://doi.org/10.1108/jfmpc-06-2013-0027).

<span id="ref-3"></span>[3] Olusina Temidayo Akinyokun, Onifade. Morakinyo Kehinde, Adegoke Michael Abejide, "A Review of the Application of Decision Tree Analysis and Artificial Neural Networks in Project Management," *Jurnal Teknik Industri: Jurnal Hasil Penelitian dan Karya Ilmiah dalam Bidang Teknik Industri*, 2025. doi: [10.24014/jti.v11i1.16480](https://doi.org/10.24014/jti.v11i1.16480).

<span id="ref-4"></span>[4] A. Aliyev, "The AI and Quantum Era: Transforming Project Management Practices," *Journal of Future Artificial Intelligence and Technologies*, 2025. doi: [10.62411/faith.3048-3719-59](https://doi.org/10.62411/faith.3048-3719-59).

<span id="ref-5"></span>[5] P. K. Andersen, R. B. Geskus, T. de Witte, H. Putter, "Competing risks in epidemiology: possibilities and pitfalls," *International Journal of Epidemiology*, 2012. doi: [10.1093/ije/dyr213](https://doi.org/10.1093/ije/dyr213).

<span id="ref-6"></span>[6] A. Ansar, B. Flyvbjerg, Alexander Budzier, Daniel Lunn, "Should We Build More Large Dams? The Actual Costs of Hydropower Megaproject Development," 2014. doi: [10.1016/j.enpol.2013.10.069](https://doi.org/10.1016/j.enpol.2013.10.069).

<span id="ref-7"></span>[7] Peter C. Austin, Aurélien Latouche, Jason P. Fine, "A review of the use of time‐varying covariates in the Fine‐Gray subdistribution hazard competing risk regression model," *Statistics in Medicine*, 2019. doi: [10.1002/sim.8399](https://doi.org/10.1002/sim.8399).
<span id="ref-8"></span>[8] P. Austin, H. Putter, Douglas S. Lee, E. Steyerberg, "Estimation of the Absolute Risk of Cardiovascular Disease and Other Events: Issues With the Use of Multiple Fine-Gray Subdistribution Hazard Models," *Circulation. Cardiovascular Quality and Outcomes*, 2022. doi: [10.1161/circoutcomes.121.008368](https://doi.org/10.1161/circoutcomes.121.008368).

<span id="ref-9"></span>[9] P. C. Austin, D. S. Lee, J. P. Fine, "Introduction to the Analysis of Survival Data in the Presence of Competing Risks," *Circulation*, 2016. doi: [10.1161/circulationaha.115.017719](https://doi.org/10.1161/circulationaha.115.017719).

<span id="ref-10"></span>[10] P. C. Austin, J. P. Fine, "Practical recommendations for reporting Fine-Gray model analyses for competing risk data," *Statistics in Medicine*, 2017. doi: [10.1002/sim.7501](https://doi.org/10.1002/sim.7501).

<span id="ref-11"></span>[11] P. Austin, A. Allignol, J. Fine, "The number of primary events per variable affects estimation of the subdistribution hazard competing risks model.," *Journal of Clinical Epidemiology*, 2017. doi: [10.1016/j.jclinepi.2016.11.017](https://doi.org/10.1016/j.jclinepi.2016.11.017).

<span id="ref-12"></span>[12] Omotola Awojobi, Glenn P. Jenkins, "Managing the cost overrun risks of hydroelectric dams: An application of reference class forecasting techniques," *Renewable and Sustainable Energy Reviews*, 2016. doi: [10.1016/j.rser.2016.05.006](https://doi.org/10.1016/j.rser.2016.05.006).

<span id="ref-13"></span>[13] Rebekka Baerenbold, "Reducing risks in megaprojects: The potential of reference class forecasting," *Project Leadership and Society*, 2023. doi: [10.1016/j.plas.2023.100103](https://doi.org/10.1016/j.plas.2023.100103).

<span id="ref-14"></span>[14] G. Bakoyannis, Menggang Yu, C. Yiannoutsos, "Semiparametric regression on cumulative incidence function with interval-censored competing risks data," *Statistics in Medicine*, 2017. doi: [10.1002/sim.7350](https://doi.org/10.1002/sim.7350).

<span id="ref-15"></span>[15] Bryan Barley, Allen S. Bacskay, Marilyn Newhouse, "Heritage and Advanced Technology Systems Engineering Lessons Learned from NASA Deep Space Missions," 2010. doi: [10.2514/6.2010-8622](https://doi.org/10.2514/6.2010-8622).

<span id="ref-16"></span>[16] Ryan de Freitas Bart, K. Duda, J. Hoffman, "Estimating the Cost to Transition a Space System from Expendable to Reusable," *IEEE Aerospace Conference*, 2023. doi: [10.1109/aero55745.2023.10115848](https://doi.org/10.1109/aero55745.2023.10115848).

<span id="ref-17"></span>[17] D. A. Bearden, "Complexity based cost estimating relationships for space systems," *2004 IEEE Aerospace Conference*, 2004. doi: [10.1109/aero.2004.1368229](https://doi.org/10.1109/aero.2004.1368229).

<span id="ref-18"></span>[18] K. Behdinan, S. Mishra, "CARDINAL WTRL: TECHNOLOGY MATURITY, SCHEDULE SLIPPAGE AND TREND FORECASTING.," *Proceedings of the Design Society*, 2021. doi: [10.1017/pds.2021.60](https://doi.org/10.1017/pds.2021.60).

<span id="ref-19"></span>[19] Christophe Bell'ego, David Benatia, Vincent Dortet-Bernardet, "The chained difference-in-differences," *Journal of Econometrics*, 2023. doi: [10.1016/j.jeconom.2024.105783](https://doi.org/10.1016/j.jeconom.2024.105783).

<span id="ref-20"></span>[20] M. Berger, M. Schmid, T. Welchowski, S. Schmitz-Valckenberg, J. Beyersmann, "Subdistribution hazard models for competing risks in discrete time.," *Biostatistics*, 2018. doi: [10.1093/biostatistics/kxy069](https://doi.org/10.1093/biostatistics/kxy069).

<span id="ref-21"></span>[21] Jan Beyersmann, Martin Schumacher, "Time-dependent covariates in the proportional subdistribution hazards model for competing risks," *Biostatistics*, 2008. doi: [10.1093/biostatistics/kxn009](https://doi.org/10.1093/biostatistics/kxn009).

<span id="ref-22"></span>[22] Robert Bitten, Charles Hunt, Debra Emmons, Robert Kellogg, Eric Mahr, Sarah Lang, "The effect of policy changes on NASA science mission cost &amp; schedule growth," *2018 IEEE Aerospace Conference*, 2018. doi: [10.1109/aero.2018.8396408](https://doi.org/10.1109/aero.2018.8396408).

<span id="ref-23"></span>[23] E. Bonneville, L. D. de Wreede, H. Putter, "Why you should avoid using multiple Fine–Gray models: insights from (attempts at) simulating proportional subdistribution hazards data," *Journal of the Royal Statistical Society: Series A (Statistics in Society)*, 2024. doi: [10.1093/jrsssa/qnae056](https://doi.org/10.1093/jrsssa/qnae056).

<span id="ref-24"></span>[24] C. Brown, Hanni Lux, James R. Cowan, "Reference Class Forecasting and Its Application to Fusion Power Plant Cost Estimates," *IEEE Transactions on Plasma Science*, 2024. doi: [10.1109/tps.2024.3405631](https://doi.org/10.1109/tps.2024.3405631).

<span id="ref-25"></span>[25] Alexander Budzier, Bent Flyvbjerg, "The Oxford Olympics Study 2024: Are Cost and Cost Overrun at the Games Coming Down?," *Social Science Research Network*, 2024. doi: [10.2139/ssrn.4849892](https://doi.org/10.2139/ssrn.4849892).

<span id="ref-26"></span>[26] Ruel, C., Larouche, M., Donato, M., "Monitored background radiometer," *NASA Technical Reports Server*, 1986. Available: [https://ntrs.nasa.gov/citations/19880001449](https://ntrs.nasa.gov/citations/19880001449).

<span id="ref-27"></span>[27] B. Callaway, P. H. C. Sant'Anna, "Difference-in-Differences with multiple time periods," *Journal of Econometrics*, 2021. doi: [10.1016/j.jeconom.2020.12.001](https://doi.org/10.1016/j.jeconom.2020.12.001).

<span id="ref-28"></span>[28] Eduardo Monterisi Almeida de Carvalho, "FORECASTING AND FINANCIAL PREDICTABILITY IN COMPLEX INFRASTRUCTURE AND ENERGY PROJECTS," *Revista de Estudos Interdisciplinares*, 2026. doi: [10.23900/artefactum.v25i2.2931](https://doi.org/10.23900/artefactum.v25i2.2931).

<span id="ref-29"></span>[29] P. Carydias, "Engineering Visual Presentation E08: Tackling the 65% megaproject failure epidemic: six key levers amplified by digital and AI," *Australian Energy Producers Journal*, 2025. doi: [10.1071/ep24445](https://doi.org/10.1071/ep24445).

<span id="ref-30"></span>[30] P. Carydias, R. Kennedy, "Tackling the 65% megaproject failure epidemic: six key levers amplified by digital and AI," *Australian Energy Producers Journal*, 2025. doi: [10.1071/ep24270](https://doi.org/10.1071/ep24270).

<span id="ref-31"></span>[31] Aaron Chadee, Hector Martin, Sihara Gallage, Upaka Rathnayake, "Reducing Cost Overrun in Public Housing Projects: A Simplified Reference Class Forecast for Small Island Developing States," *Buildings*, 2023. doi: [10.3390/buildings13040998](https://doi.org/10.3390/buildings13040998).

<span id="ref-32"></span>[32] Aaron Anil Chadee, Salisha R. Hernandez, Héctor Martín, "The Influence of Optimism Bias on Time and Cost on Construction Projects," *Emerging Science Journal*, 2021. doi: [10.28991/esj-2021-01287](https://doi.org/10.28991/esj-2021-01287).
<span id="ref-33"></span>[33] Clément de Chaisemartin, Xavier d'Haultfoeuille, "Two-Way Fixed Effects and Difference-in-Differences Estimators with Heterogeneous Treatment Effects and Imperfect Parallel Trends," *SSRN Electronic Journal*, 2023. doi: [10.2139/ssrn.4487202](https://doi.org/10.2139/ssrn.4487202).

<span id="ref-34"></span>[34] Yizi Chen, Dominic D. Ahiaga-Dagbui, Muhammad Jamaluddin Thaheem, Asheem Shrestha, "Toward a Deeper Understanding of Optimism Bias and Transport Project Cost Overrun," *Project Management Journal*, 2023. doi: [10.1177/87569728231180268](https://doi.org/10.1177/87569728231180268).

<span id="ref-35"></span>[35] Mario A. Cleves, William Gould, Roberto G. Gutierrez, Yulia Marchenko, "An Introduction to Survival Analysis Using Stata," *RePEc: Research Papers in Economics*, 2003. Available: [https://openalex.org/W1529206929](https://openalex.org/W1529206929).

<span id="ref-36"></span>[36] "A cost effective methodology for building flight spares for robotic life extension," *Acta Astronautica*, 2019. doi: [10.1016/j.actaastro.2019.06.025](https://doi.org/10.1016/j.actaastro.2019.06.025).

<span id="ref-37"></span>[37] V. Coviello, M. Boggess, "Cumulative Incidence Estimation in the Presence of Competing Risks," *The Stata Journal*, 2004. doi: [10.1177/1536867x0400400201](https://doi.org/10.1177/1536867x0400400201).

<span id="ref-38"></span>[38] Robert Damadeo, Charles Hill, Michael Obland, John Leckey, "The Stratospheric Aerosol and Gas Experiment (SAGE) IV Pathfinder," *NASA Technical Reports Server*, 2020. Available: [https://ntrs.nasa.gov/citations/20205011684](https://ntrs.nasa.gov/citations/20205011684).

<span id="ref-39"></span>[39] Guoqing Diao, Joseph G Ibrahim, "Quantifying time-varying cause-specific hazard and subdistribution hazard ratios with competing risks data," *Clinical Trials*, 2019. doi: [10.1177/1740774519852708](https://doi.org/10.1177/1740774519852708).

<span id="ref-40"></span>[40] "Difference-in-Differences with a Continuous Treatment," doi: [10.2139/ssrn.4716682](https://doi.org/10.2139/ssrn.4716682).

<span id="ref-41"></span>[41] "Difference-in-differences with variation in treatment timing," doi: [10.1016/j.jeconom.2021.03.014](https://doi.org/10.1016/j.jeconom.2021.03.014).

<span id="ref-42"></span>[42] "DISE: A Dynamic Integrated Space-Economy Model for Orbital Debris Mitigation Policy Evaluation (2025)," doi: [10.1007/s10640-025-01003-y](https://doi.org/10.1007/s10640-025-01003-y).

<span id="ref-43"></span>[43] Stephanie N. Dixon, Gerarda A. Darlington, Anthony F. Desmond, "A competing risks model for correlated data based on the subdistribution hazard," *Lifetime Data Analysis*, 2011. doi: [10.1007/s10985-011-9198-9](https://doi.org/10.1007/s10985-011-9198-9).

<span id="ref-44"></span>[44] Dave Donaldson, Adam Storeygard, "The View from Above: Applications of Satellite Data in Economics," doi: [10.1257/jep.30.4.171](https://doi.org/10.1257/jep.30.4.171).

<span id="ref-45"></span>[45] "Dual-telescope multi-channel thermal-infrared radiometer for outer planet fly-by," *Acta Astronautica*, 2016. doi: [10.1016/j.actaastro.2016.08.009](https://doi.org/10.1016/j.actaastro.2016.08.009).

<span id="ref-46"></span>[46] G. F. Dubos, J. H. Saleh, R. Braun, "Technology Readiness Level, Schedule Risk and Slippage in Spacecraft Design: Data Analysis and Modeling (conference antecedent)," *AIAA SPACE Conference*, 2007. doi: [10.2514/6.2007-6020](https://doi.org/10.2514/6.2007-6020).

<span id="ref-47"></span>[47] G. F. Dubos, J. H. Saleh, R. Braun, "Technology Readiness Level, Schedule Risk, and Slippage in Spacecraft Design," *Journal of Spacecraft and Rockets*, 2008. doi: [10.2514/1.34947](https://doi.org/10.2514/1.34947).

<span id="ref-48"></span>[48] "Economic Benefits of the Global Positioning System (GPS) to the U.S. Private Sector (RTI International for NIST, 2019)," doi: [10.6028/nist.gcr.25-062](https://doi.org/10.6028/nist.gcr.25-062).

<span id="ref-49"></span>[49] "The environmental impact of emissions from space launches: A comprehensive review," doi: [10.1016/j.jclepro.2020.120209](https://doi.org/10.1016/j.jclepro.2020.120209).

<span id="ref-50"></span>[50] Angela M Erb, Peter B Boucher, Miguel O Roman, Ian L Paynter, "1.02  NASA Earth Observations," *NASA Technical Reports Server*, 2025. Available: [https://ntrs.nasa.gov/citations/20250010859](https://ntrs.nasa.gov/citations/20250010859).

<span id="ref-51"></span>[51] "Evaluating the environmental impact of space missions," doi: [10.1016/j.actaastro.2025.06.024](https://doi.org/10.1016/j.actaastro.2025.06.024).

<span id="ref-52"></span>[52] "Evaluating the impact of space activities in low earth orbit," doi: [10.1016/j.actaastro.2021.03.030](https://doi.org/10.1016/j.actaastro.2021.03.030).

<span id="ref-53"></span>[53] "An evaluation of dependencies of critical infrastructure timing systems on the Global Positioning System (GPS) (NIST 2021)," doi: [10.6028/nist.tn.2189](https://doi.org/10.6028/nist.tn.2189).

<span id="ref-54"></span>[54] "Feature Detection Systems Enhance Satellite Imagery," *NASA Technical Reports Server*, 2009. Available: [https://ntrs.nasa.gov/citations/20090039420](https://ntrs.nasa.gov/citations/20090039420).

<span id="ref-55"></span>[55] J. P. Fine, R. J. Gray, "A Proportional Hazards Model for the Subdistribution of a Competing Risk," *Journal of the American Statistical Association*, 1999. doi: [10.1080/01621459.1999.10474144](https://doi.org/10.1080/01621459.1999.10474144).

<span id="ref-56"></span>[56] Bent Flyvbjerg, "Curbing Optimism Bias and Strategic Misrepresentation in Planning: Reference Class Forecasting in Practice," *European Planning Studies*, 2007. doi: [10.1080/09654310701747936](https://doi.org/10.1080/09654310701747936).

<span id="ref-57"></span>[57] B. Flyvbjerg, "Five things you should know about cost overrun," *Transportation Research Part A: Policy and Practice*, 2018. doi: [10.1016/j.tra.2018.07.013](https://doi.org/10.1016/j.tra.2018.07.013).
<span id="ref-58"></span>[58] B. Flyvbjerg, M. K. Skamris Holm, S. L. Buhl, "How (In)accurate Are Demand Forecasts in Public Works Projects? The Case of Transportation," *Journal of the American Planning Association*, 2005. doi: [10.1080/01944360508976688](https://doi.org/10.1080/01944360508976688).

<span id="ref-59"></span>[59] Bent Flyvbjerg, "Megaproject policy and planning: Problems, causes, cures," *VBN Forskningsportal (Aalborg Universitet)*, 2007. Available: [https://openalex.org/W1832657426](https://openalex.org/W1832657426).

<span id="ref-60"></span>[60] Bent Flyvbjerg, Carsten Glenting, Arne Rønnest, "Procedures for Dealing with Optimism Bias in Transport Planning: Guidance Document," *VBN Forskningsportal (Aalborg Universitet)*, 2004. Available: [https://openalex.org/W2739390026](https://openalex.org/W2739390026).

<span id="ref-61"></span>[61] Bent Flyvbjerg, Chi-keung Hon, Wing Huen Fok, "Reference class forecasting for Hong Kong’s major roadworks projects," *Proceedings of the Institution of Civil Engineers - Civil Engineering*, 2016. doi: [10.1680/jcien.15.00075](https://doi.org/10.1680/jcien.15.00075).

<span id="ref-62"></span>[62] Tamrat Endebu Gebre, Girma Taye, W. Deressa, "Time-dependent predictors of loss to follow-up in HIV care in low-resource settings: A competing risks approach," *PLoS ONE*, 2025. doi: [10.1371/journal.pone.0329132](https://doi.org/10.1371/journal.pone.0329132).

<span id="ref-63"></span>[63] Richard D. Gelber, Robert J. Gray, "Cumulative incidence functions should replace integrated cause-specific hazard functions to describe probabilities of competing events over time," *Controlled Clinical Trials*, 1991. doi: [10.1016/0197-2456(91)90089-5](https://doi.org/10.1016/0197-2456(91)90089-5).

<span id="ref-64"></span>[64] Ronald B. Geskus, "Cause-Specific Cumulative Incidence Estimation and the Fine and Gray Model Under Both Left Truncation and Right Censoring," *Biometrics*, 2010. doi: [10.1111/j.1541-0420.2010.01420.x](https://doi.org/10.1111/j.1541-0420.2010.01420.x).

<span id="ref-65"></span>[65] H. Habib-agahi, "NASA Instrument Cost Model for Explorer-Like Mission Instruments (NICM-E)," *NASA Technical Reports Server*, 2015. Available: [https://ntrs.nasa.gov/citations/20150008233](https://ntrs.nasa.gov/citations/20150008233).

<span id="ref-66"></span>[66] H. Habib-agahi, "NASA Space Flight Instruments: Cost-Time Trends," *NASA Technical Reports Server*, 2019. Available: [https://ntrs.nasa.gov/citations/20190026811](https://ntrs.nasa.gov/citations/20190026811).

<span id="ref-67"></span>[67] Meagan Hahn, Rachel Sholder, "Cost Estimation of Electronic Parts in NASA Space Missions," *IEEE Aerospace Conference*, 2020. doi: [10.1109/aero47225.2020.9172574](https://doi.org/10.1109/aero47225.2020.9172574).

<span id="ref-68"></span>[68] Sally R. Hinchliffe, Paul C. Lambert, "Flexible parametric modelling of cause-specific hazards to estimate cumulative incidence functions," *BMC Medical Research Methodology*, 2013. doi: [10.1186/1471-2288-13-13](https://doi.org/10.1186/1471-2288-13-13).

<span id="ref-69"></span>[69] Steven R. Hirshorn, Sharon Jefferies, "Final Report of the NASA Technology Readiness Assessment (TRA) Study Team," *NASA STI Repository (National Aeronautics and Space Administration)*, 2016. Available: [https://openalex.org/W2739239624](https://openalex.org/W2739239624).

<span id="ref-70"></span>[70] "How Common Is Cost Overrun and Cost Growth?," *Project Cost Overrun*, doi: [10.1017/9781316779675.007](https://doi.org/10.1017/9781316779675.007).

<span id="ref-71"></span>[71] "Impact of Rocket Launch and Space Debris Air Pollutant Emissions on Stratospheric Ozone and Global Climate," doi: [10.1029/2021ef002612](https://doi.org/10.1029/2021ef002612).

<span id="ref-72"></span>[72] J. R. Irons, J. L. Dwyer, J. A. Barsi, "The next Landsat satellite: The Landsat Data Continuity Mission," *Remote Sensing of Environment*, 2012. doi: [10.1016/j.rse.2011.08.026](https://doi.org/10.1016/j.rse.2011.08.026).

<span id="ref-73"></span>[73] Perez G.J., "PHILIPPINE EARTH OBSERVATION SATELLITE MISSIONS AND APPLICATIONS: A DECADAL SURVEY," *International Archives of the Photogrammetry Remote Sensing and Spatial Information Sciences ISPRS Archives*, 2023. doi: [10.5194/isprs-archives-xlviii-1-w2-2023-1507-2023](https://doi.org/10.5194/isprs-archives-xlviii-1-w2-2023-1507-2023).

<span id="ref-74"></span>[74] Irons, Jim, Sabelhaus, Phil, Masek, Jeff, Cook, Bruce, Dabney, Phil, Loveland, Tom, "Progress Towards a 2012 Landsat Launch," *NASA Technical Reports Server*, 2012. Available: [https://ntrs.nasa.gov/citations/20120009180](https://ntrs.nasa.gov/citations/20120009180).

<span id="ref-75"></span>[75] Ferebee, Melvin J., Jr., Troutman, Patrick A., Monell, Donald W., "Satellite Systems Design/Simulation Environment: A Systems Approach to Pre-Phase A Design," *NASA Technical Reports Server*, 1997. Available: [https://ntrs.nasa.gov/citations/20040105652](https://ntrs.nasa.gov/citations/20040105652).

<span id="ref-76"></span>[76] Warfield, K., Roust, K., "The JPL Advanced Projects Design Team's Spacecraft Instrument Cost Model: an Objective, Multivariate Approach," *NASA Technical Reports Server*, 1998. Available: [https://ntrs.nasa.gov/citations/20060034648](https://ntrs.nasa.gov/citations/20060034648).

<span id="ref-77"></span>[77] K. A. Kipp, S. C. Ringler, E. L. Chapman, C. W. Freaner, "Impact of instrument schedule growth on mission cost and schedule growth for recent NASA missions," *2012 IEEE Aerospace Conference*, 2012. doi: [10.1109/aero.2012.6187407](https://doi.org/10.1109/aero.2012.6187407).

<span id="ref-78"></span>[78] Rasheed K. Lamidi, Bello Ishola Sanni, Saheed Kunle Ajibade, B. I. Doroh, Aishat Olaosebikan, "Cumulative Incidence Function in Competing Risks: A Case Study of Primary Biliary Cirrhosis in Liver Disease," *International Journal of Science for Global Sustainability*, 2025. doi: [10.57233/ijsgs.v11i4.991](https://doi.org/10.57233/ijsgs.v11i4.991).

<span id="ref-79"></span>[79] Aurélien Latouche, Arthur Allignol, Jan Beyersmann, Myriam Labopin, Jason P. Fine, "A competing risks analysis should report results on all cause-specific hazards and cumulative incidence functions," *Journal of Clinical Epidemiology*, 2013. doi: [10.1016/j.jclinepi.2012.09.017](https://doi.org/10.1016/j.jclinepi.2012.09.017).

<span id="ref-80"></span>[80] Bonnie Lau, Stephen R. Cole, Stephen J. Gange, "Competing Risk Regression Models for Epidemiologic Data," *American Journal of Epidemiology*, 2009. doi: [10.1093/aje/kwp107](https://doi.org/10.1093/aje/kwp107).

<span id="ref-81"></span>[81] Jalal Meftah Mohamed Lekraik, Opeoluwaseun Ojekemi, "Advancing Digital Project Management Through AI: An Interpretable POA-LightGBM Framework for Cost Overrun Prediction," *Syst.*, 2025. doi: [10.3390/systems13121047](https://doi.org/10.3390/systems13121047).

<span id="ref-82"></span>[82] T. Leunig, "Time is Money: A Re-Assessment of the Passenger Social Savings from Victorian British Railways," *Journal of Economic History*, 2006. doi: [10.1017/s0022050706000283](https://doi.org/10.1017/s0022050706000283).
<span id="ref-83"></span>[83] Jian Li, David P. Roy, "A Global Analysis of Sentinel-2A, Sentinel-2B and Landsat-8 Data Revisit Intervals and Implications for Terrestrial Monitoring," *Remote Sensing*, 2017. doi: [10.3390/rs9090902](https://doi.org/10.3390/rs9090902).

<span id="ref-84"></span>[84] Wanxing Li, Xiaoming Xue, Yonghong Long, "An additive subdistribution hazard model for competing risks data," *Communications in Statistics - Theory and Methods*, 2017. doi: [10.1080/03610926.2016.1277759](https://doi.org/10.1080/03610926.2016.1277759).

<span id="ref-85"></span>[85] Wanxing Li, Yonghong Long, "An Additive-Multiplicative Cox-Aalen Subdistribution Hazard Model for Competing Risks Data," *Journal of Systems Science and Complexity*, 2019. doi: [10.1007/s11424-019-7281-6](https://doi.org/10.1007/s11424-019-7281-6).

<span id="ref-86"></span>[86] M. Lieber, B. Donor, "Schedule matters: Understanding the relationship between schedule delays and costs on overruns," *2016 IEEE Aerospace Conference*, 2016. doi: [10.1109/aero.2016.7500722](https://doi.org/10.1109/aero.2016.7500722).

<span id="ref-87"></span>[87] Huimin Liu, Canhui Jiang, Yan Liu, Marcel Hertogh, Xue Lyu, "Optimism Bias Evaluation and Decision-Making Risk Forecast on Bridge Project Cost Based on Reference Class Forecasting: Evidence from China," *Sustainability*, 2018. doi: [10.3390/su10113981](https://doi.org/10.3390/su10113981).

<span id="ref-88"></span>[88] Dan Lovallo, Matteo Cristofaro, Bent Flyvbjerg, "Governing Large Projects: A Three-Stage Process to Get It Right," *Academy of Management Perspectives*, 2023. doi: [10.5465/amp.2021.0129](https://doi.org/10.5465/amp.2021.0129).

<span id="ref-89"></span>[89] Dr. Muhammad Imran Majeed, Dr. Abdullah Hammad, Dr Sahar Munir, Rida Maqsood, Isha Safdar, "Evaluating Cost Estimation and Budgeting Challenges in Public Health and Education Infrastructure Projects," *Social Science Review Archives*, 2025. doi: [10.70670/sra.v3i4.1385](https://doi.org/10.70670/sra.v3i4.1385).

<span id="ref-90"></span>[90] P. Martimort, B. C. Domínguez, Arnaud Hélière, Josep Roselló, Martin Suess, J. Angevain, Adrien Bardou, F. Bräuer, "On-Going and Planned Mission Concept Studies for the Preparation of Future ESA Earth Observation Satellites," *IEEE International Geoscience and Remote Sensing Symposium*, 2023. doi: [10.1109/igarss52108.2023.10283393](https://doi.org/10.1109/igarss52108.2023.10283393).

<span id="ref-91"></span>[91] Jeffrey G. Masek, Michael A. Wulder, Brian Markham, Joel Mccorkel,  Christopher J. Crawford, James Storey, Del T Jenstrom, "Landsat 9: Empowering Open Science and Applications Through Continuity," *NASA Technical Reports Server*, 2020. doi: [10.1016/j.rse.2020.111968](https://doi.org/10.1016/j.rse.2020.111968).

<span id="ref-92"></span>[92] Matthew F. McCabe, Matthew Rodell, Douglas Alsdorf, Diego G. Miralles, R. Uijlenhoet, Wolfgang Wagner, Arko Lucieer, Rasmus Houborg, "The future of Earth observation in hydrology," *Hydrology and earth system sciences*, 2017. doi: [10.5194/hess-21-3879-2017](https://doi.org/10.5194/hess-21-3879-2017).

<span id="ref-93"></span>[93] K. R. Miner, R. K. Braghiere, C. E. Miller, N. Schlegel, D. Schimel, "A Decadal Survey Without Analogs: Earth Observation Needs for a Warming World," *AGU Advances*, 2024. doi: [10.1029/2023av001148](https://doi.org/10.1029/2023av001148).

<span id="ref-94"></span>[94] Chantrice Moss, Suman Chakrabarti, David W. Scott, "Parts quality management: Direct part marking of data matrix symbol for mission assurance," 2013. doi: [10.1109/aero.2013.6497372](https://doi.org/10.1109/aero.2013.6497372).

<span id="ref-95"></span>[95] S. I. Mozumder, Sarah Booth, Richard D Riley, M. Rutherford, Paul C Lambert, "Calibration of cause-specific absolute risk for external validation using each cause-specific hazards model in the presence of competing events," *Diagnostic and Prognostic Research*, 2025. doi: [10.1186/s41512-025-00197-5](https://doi.org/10.1186/s41512-025-00197-5).

<span id="ref-96"></span>[96] S. I. Mozumder, M. Rutherford, P. Lambert, "stpm2cr: A flexible parametric competing risks model using a direct likelihood approach for the cause-specific cumulative incidence function," *The Stata Journal*, 2017. doi: [10.1177/1536867x1701700212](https://doi.org/10.1177/1536867x1701700212).

<span id="ref-97"></span>[97] Ananth Natarajan, "Reference Class Forecasting and Machine Learning for Improved Offshore Oil and Gas Megaproject Planning: Methods and Application," *Project Management Journal*, 2022. doi: [10.1177/87569728211045889](https://doi.org/10.1177/87569728211045889).

<span id="ref-98"></span>[98] Victoria Nilsen, "Examining the Effects of Implementing Data-Driven Uncertainty in Cost Estimating Models," *IEEE Aerospace Conference*, 2023. doi: [10.1109/aero55745.2023.10116026](https://doi.org/10.1109/aero55745.2023.10116026).

<span id="ref-99"></span>[99] "OPUS: An Integrated Assessment Model for Satellites and Orbital Debris (2023)," doi: [10.48550/arxiv.2309.10252](https://doi.org/10.48550/arxiv.2309.10252).

<span id="ref-100"></span>[100] "Orbital debris and the market for satellites (Rao 2023)," doi: [10.1016/j.ecolecon.2023.107831](https://doi.org/10.1016/j.ecolecon.2023.107831).

<span id="ref-101"></span>[101] Neeck, Steven P., "The NASA Earth Science Program and Small Satellites," *NASA Technical Reports Server*, 2015. Available: [https://ntrs.nasa.gov/citations/20150022481](https://ntrs.nasa.gov/citations/20150022481).

<span id="ref-102"></span>[102] O'Neill, P., Podest, E., "Utilization of Ancillary Data Sets for Conceptual SMAP Mission Algorithm Development and Product Generation," *NASA Technical Reports Server*, 2011. Available: [https://ntrs.nasa.gov/citations/20150005750](https://ntrs.nasa.gov/citations/20150005750).

<span id="ref-103"></span>[103] Jung Eun Park, "Curbing cost overruns in infrastructure investment," *European journal of transport and infrastructure research*, 2021. doi: [10.18757/ejtir.2021.21.2.5504](https://doi.org/10.18757/ejtir.2021.21.2.5504).

<span id="ref-104"></span>[104] Jun Park, G. Bakoyannis, Ying Zhang, C. Yiannoutsos, "Semiparametric regression on cumulative incidence function with interval-censored competing risks data and missing event types," *Biostatistics*, 2021. doi: [10.1093/biostatistics/kxaa052](https://doi.org/10.1093/biostatistics/kxaa052).

<span id="ref-105"></span>[105] Darius Phiri, Justin Morgenroth, "Developments in Landsat Land Cover Classification Methods: A Review," *Remote Sensing*, 2017. doi: [10.3390/rs9090967](https://doi.org/10.3390/rs9090967).

<span id="ref-106"></span>[106] Melania Pintilie, "Competing Risks: A Practical Perspective," 2006. Available: [https://openalex.org/W179935622](https://openalex.org/W179935622).

<span id="ref-107"></span>[107] Doe Precious, "Infrastructure project cost overrun and schedule delay in Ghana: Is it an issue of resource misallocation or financial constraints?," *Project Leadership and Society*, 2025. doi: [10.1016/j.plas.2025.100188](https://doi.org/10.1016/j.plas.2025.100188).
<span id="ref-108"></span>[108] "Pretest with Caution: Event-Study Estimates after Testing for Parallel Trends," doi: [10.1257/aeri.20210236](https://doi.org/10.1257/aeri.20210236).

<span id="ref-109"></span>[109] "Probabilistic Cost-Schedule Risk Analysis in Offshore Construction Projects Using Monte Carlo Simulation and Quantitative Assessment," *International Journal of Business and Technology Management*, 2025. doi: [10.55057/ijbtm.2025.7.9.34](https://doi.org/10.55057/ijbtm.2025.7.9.34).

<span id="ref-110"></span>[110] Shen‐En Qian, "Hyperspectral Satellites, Evolution, and Development History," *IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing*, 2021. doi: [10.1109/jstars.2021.3090256](https://doi.org/10.1109/jstars.2021.3090256).

<span id="ref-111"></span>[111] "Railroad Impact in Backward Economies: Spain, 1850-1913 (Herranz-Loncan 2006)," doi: [10.1017/s0022050706000350](https://doi.org/10.1017/s0022050706000350).

<span id="ref-112"></span>[112] "Railroads and American Economic Growth: A Market Access Approach (Donaldson & Hornbeck 2013)," doi: [10.3386/w19213](https://doi.org/10.3386/w19213).

<span id="ref-113"></span>[113] "Railroads and American Economic Growth: Essays in Econometric History (Fogel 1964)," doi: [10.2307/2552284](https://doi.org/10.2307/2552284).

<span id="ref-114"></span>[114] "Railroads, Reallocation, and the Rise of American Manufacturing (2019)," doi: [10.3386/w26594](https://doi.org/10.3386/w26594).

<span id="ref-115"></span>[115] Malcolm H. Ray, "A Class of $K$-Sample Tests for Comparing the Cumulative Incidence of a Competing Risk," *The Annals of Statistics*, 1988. doi: [10.1214/aos/1176350951](https://doi.org/10.1214/aos/1176350951).

<span id="ref-116"></span>[116] Ping K. Ruan, Robert J. Gray, "Analyses of cumulative incidence functions via non‐parametric multiple imputation," *Statistics in Medicine*, 2008. doi: [10.1002/sim.3402](https://doi.org/10.1002/sim.3402).

<span id="ref-117"></span>[117] K. Rufibach, J. Beyersmann, T. Friede, Claudia Schmoor, Regina Stegherr, "Survival analysis for AdVerse events with VarYing follow-up times (SAVVY): summary of findings and assessment of existing guidelines," *Trials*, 2024. doi: [10.1186/s13063-024-08186-7](https://doi.org/10.1186/s13063-024-08186-7).

<span id="ref-118"></span>[118] P. H. C. Sant'Anna, J. Zhao, "Doubly robust difference-in-differences estimators," *Journal of Econometrics*, 2020. doi: [10.1016/j.jeconom.2020.06.003](https://doi.org/10.1016/j.jeconom.2020.06.003).

<span id="ref-119"></span>[119] "Satellite launch traffic forecast based on the global Gross Domestic Product and constellation plans (Retagne & Colombo 2026)," doi: [10.1016/j.actaastro.2026.04.046](https://doi.org/10.1016/j.actaastro.2026.04.046).

<span id="ref-120"></span>[120] Sezer Savaş, "Artificial intelligence in construction project management: Trends, challenges and future directions," *Journal of Design for Resilience in Architecture and Planning*, 2025. doi: [10.47818/drarch.2025.v6i2165](https://doi.org/10.47818/drarch.2025.v6i2165).

<span id="ref-121"></span>[121] Emil Scosyrev, "Improved confidence intervals for a difference of two cause‐specific cumulative incidence functions estimated in the presence of competing risks and random censoring," *Biometrical Journal*, 2020. doi: [10.1002/bimj.201900060](https://doi.org/10.1002/bimj.201900060).

<span id="ref-122"></span>[122] Rachel Sholder, S. Whitley, "Math is EZIE: How Contracts Help Control Cost," *IEEE Aerospace Conference*, 2023. doi: [10.1109/aero55745.2023.10115565](https://doi.org/10.1109/aero55745.2023.10115565).

<span id="ref-123"></span>[123] L. Sobel, Elliott B. Tibor, "Effectiveness of Firm-Fixed Price Spacecraft Contracts to Curb Cost Growth," *IEEE Aerospace Conference*, 2021. doi: [10.1109/aero50100.2021.9438356](https://doi.org/10.1109/aero50100.2021.9438356).

<span id="ref-124"></span>[124] "Social Savings (Leunig 2010)," doi: [10.1111/j.1467-6419.2010.00636.x](https://doi.org/10.1111/j.1467-6419.2010.00636.x).

<span id="ref-125"></span>[125] "Space sustainability rating: Designing a composite indicator to incentivize satellite operators to pursue long-term sustainability of the space environment (Rathnasabapathy et al. 2025)," doi: [10.1016/j.actaastro.2025.03.034](https://doi.org/10.1016/j.actaastro.2025.03.034).

<span id="ref-126"></span>[126] "Space, the Final Economic Frontier (Weinzierl 2018)," doi: [10.1257/jep.32.2.173](https://doi.org/10.1257/jep.32.2.173).

<span id="ref-127"></span>[127] "Spacecraft and optics design considerations for a spaceborne lidar mission," *Acta Astronautica*, 2023. doi: [10.1016/j.actaastro.2023.10.042](https://doi.org/10.1016/j.actaastro.2023.10.042).

<span id="ref-128"></span>[128] BC Tai, ZJ Chen, D Machin, "Estimating sample size in the presence of competing risks – Cause-specific hazard or cumulative incidence approach?," *Statistical Methods in Medical Research*, 2018. doi: [10.1177/0962280215623107](https://doi.org/10.1177/0962280215623107).

<span id="ref-129"></span>[129] Shiro Tanaka, "Rare Event Approximation Between Subdistribution Hazard Ratio and Cause-specific Hazard Ratio in Survival Analysis With Competing Risks," *Journal of Epidemiology*, 2024. doi: [10.2188/jea.je20240063](https://doi.org/10.2188/jea.je20240063).

<span id="ref-130"></span>[130] "Technology roadmapping methodology for future transportation systems," *Acta Astronautica*, 2022. doi: [10.1016/j.actaastro.2022.03.038](https://doi.org/10.1016/j.actaastro.2022.03.038).

<span id="ref-131"></span>[131] "Tipping Points of Space Debris in Low Earth Orbit," doi: [10.5334/ijc.1275](https://doi.org/10.5334/ijc.1275).

<span id="ref-132"></span>[132] J. Vilsmeier, Gisela Büchele, M. Rehm, Theresa Unseld, D. Rothenbacher, J. Beyersmann, "The choice of time scales in survival analysis has implications: calendar time versus patients’ time-to-event," *BMC Medical Research Methodology*, 2026. doi: [10.1186/s12874-026-02884-3](https://doi.org/10.1186/s12874-026-02884-3).
<span id="ref-133"></span>[133] "What is value in the New Space Economy? The end-users' perspective on satellite data and solutions (Paravano, Locatelli & Trucco 2023)," doi: [10.1016/j.actaastro.2023.05.001](https://doi.org/10.1016/j.actaastro.2023.05.001).

<span id="ref-134"></span>[134] "What's trending in difference-in-differences? A synthesis of the recent econometrics literature," doi: [10.1016/j.jeconom.2023.03.008](https://doi.org/10.1016/j.jeconom.2023.03.008).

<span id="ref-135"></span>[135] Marcel Wolbers, Michael Koller, Vianda S Stel, Beat Schaer, Kitty J. Jager, Karen Leffondré, Georg Heinze, "Competing risks analyses: objectives and approaches," *European Heart Journal*, 2014. doi: [10.1093/eurheartj/ehu131](https://doi.org/10.1093/eurheartj/ehu131).

<span id="ref-136"></span>[136] Marcel Wolbers, Michael Koller, Jacqueline C.M. Witteman, Ewout W. Steyerberg, "Prognostic Models With Competing Risks," *Epidemiology*, 2009. doi: [10.1097/ede.0b013e3181a39056](https://doi.org/10.1097/ede.0b013e3181a39056).

<span id="ref-137"></span>[137] Yilin Wu, Rui Huang, Huixia Judy Wang, Liming Xiang, "Analysis of Competing Risks Data with Covariates Subject to Detection Limits," *Journal of Computational And Graphical Statistics*, 2025. doi: [10.1080/10618600.2025.2526420](https://doi.org/10.1080/10618600.2025.2526420).

<span id="ref-138"></span>[138] Michael A. Wulder, Thomas R. Loveland, David P. Roy, Christopher J. Crawford, Jeffrey G. Masek, Curtis E. Woodcock, Richard G. Allen, Martha C. Anderson, "Current status of Landsat program, science, and applications," *Remote Sensing of Environment*, 2019. doi: [10.1016/j.rse.2019.02.015](https://doi.org/10.1016/j.rse.2019.02.015).

<span id="ref-139"></span>[139] Michael A. Wulder, David P. Roy, Volker C. Radeloff, Thomas R. Loveland, Martha C. Anderson, David M. Johnson, Sean P. Healey, Zhe Zhu, "Fifty years of Landsat science and impacts," *Remote Sensing of Environment*, 2022. doi: [10.1016/j.rse.2022.113195](https://doi.org/10.1016/j.rse.2022.113195).

<span id="ref-140"></span>[140] Mohammad Zaino, "Overestimation of Cardiovascular Mortality Risk by Kaplan-Meier in Competing Risks Settings: A Web-Based Calculator and NHANES Analysis," *International Journal of Statistics in Medical Research*, 2025. doi: [10.6000/1929-6029.2025.14.31](https://doi.org/10.6000/1929-6029.2025.14.31).

<span id="ref-141"></span>[141] Ammar Zaki, "Factors Affecting Schedule Delay and Cost Overrun in Egyptian Construction Projects," 2026. doi: [10.2139/ssrn.6103586](https://doi.org/10.2139/ssrn.6103586).

<span id="ref-142"></span>[142] E. Zancan, A. Paravano, G. Locatelli, P. Trucco, "Evolving governance in the space sector: From Legacy Space to New Space models (Zancan, Paravano, Locatelli & Trucco 2024)," doi: [10.1016/j.actaastro.2024.09.005](https://doi.org/10.1016/j.actaastro.2024.09.005).

<span id="ref-143"></span>[143] Xu Zhang, Mei‐Jie Zhang, Jason P. Fine, "A proportional hazards regression model for the subdistribution with right‐censored and left‐truncated competing risks data," *Statistics in Medicine*, 2011. doi: [10.1002/sim.4264](https://doi.org/10.1002/sim.4264).

<span id="ref-144"></span>[144] Feipeng Zhang, Heng Peng, Yong Zhou, "Fine–Gray proportional subdistribution hazards model for competing risks data under length-biased sampling," *Statistics and Its Interface*, 2019. doi: [10.4310/sii.2019.v12.n1.a10](https://doi.org/10.4310/sii.2019.v12.n1.a10).

<span id="ref-145"></span>[145] Xu Zhang, Mei‐Jie Zhang, "SAS macros for estimation of direct adjusted cumulative incidence curves under proportional subdistribution hazards models," *Computer Methods and Programs in Biomedicine*, 2010. doi: [10.1016/j.cmpb.2010.07.005](https://doi.org/10.1016/j.cmpb.2010.07.005).

<span id="ref-146"></span>[146] Zhongheng Zhang, "Survival analysis in the presence of competing risks," *Annals of Translational Medicine*, 2017. doi: [10.21037/atm.2016.08.62](https://doi.org/10.21037/atm.2016.08.62).

<span id="ref-147"></span>[147] Bo Zhou, Jason P. Fine, Aurélien Latouche, Myriam Labopin, "Competing risks regression for clustered data," *Biostatistics*, 2011. doi: [10.1093/biostatistics/kxr032](https://doi.org/10.1093/biostatistics/kxr032).

<span id="ref-148"></span>[148] Mouqing Zhou, Evangelos I. Kritsotakis, Baohua Xu, Zhusheng Guo, Yongfeng Zeng, Bin Zhou, R. Brinks, Jiancong Wang, "In-hospital mortality from healthcare-associated infection by multidrug-resistant Pseudomonas aeruginosa: a competing risks analysis of a 4-year propensity-matched cohort study in southern China," *GMS Hygiene and Infection Control*, 2025. doi: [10.3205/dgkh000597](https://doi.org/10.3205/dgkh000597).

<span id="ref-149"></span>[149] Zhe Zhu, Michael A. Wulder, David P. Roy, Curtis E. Woodcock, Matthew C. Hansen, Volker C. Radeloff, Sean P. Healey, Crystal Schaaf, "Benefits of the free and open Landsat data policy," *Remote Sensing of Environment*, 2019. doi: [10.1016/j.rse.2019.02.016](https://doi.org/10.1016/j.rse.2019.02.016).

<span id="ref-150"></span>[150] L. E. Thomas, M. P. Turakhia, and M. J. Pencina, "Competing Risks, Treatment Switching, and Informative Censoring," *JAMA Cardiology*, vol. 6, 2021. doi: [10.1001/jamacardio.2021.1239](https://doi.org/10.1001/jamacardio.2021.1239).

<span id="ref-151"></span>[151] D. L. Emmons, R. E. Bitten, and C. W. Freaner, "Using Historical NASA Cost and Schedule Growth to Set Future Program and Project Reserve Guidelines," *2007 IEEE Aerospace Conference*, 2007. doi: [10.1109/aero.2007.353027](https://doi.org/10.1109/aero.2007.353027).


## Part II: Appendices

### Appendix A: Variable and Data Dictionary

Appendix A fixes the operational definition of every construct named in the shared bible, so that two analysts coding the same mission would record the same value. The dictionary is the measurement contract for the dissertation. Where a construct rests on narrative attribution rather than a recorded number, the entry says so plainly, because the credibility of the cause-specific separation depends on the reader knowing which variables are measured and which are coded.

| Construct | Operational definition | Source | Scale / type | Notes |
|---|---|---|---|---|
| Mission-development spell | Interval from the start of Phase B (KDP-B) to the earlier of first cause-coded slip or launch readiness | CADRe milestone dates [\[86\]](#ref-86) | Duration, months from KDP-B | One spell per mission; left-truncated form used where record entry post-dates KDP-B |
| Instrument-driven first slip | First above-threshold committed-baseline schedule movement whose dominant cause is instrument development | CADRe Part A reconciled with GAO narrative | Binary event + event time | Competing event 1; see Appendix B |
| Launch-driven first slip | First above-threshold committed-baseline schedule movement whose dominant cause is launch-vehicle availability | CADRe Part A reconciled with GAO narrative | Binary event + event time | Competing event 2; see Appendix B |
| Administrative censoring | Reaching launch readiness with no above-threshold slip of either cause | CADRe / GAO | Censoring indicator | Third terminal state of the spell |
| Slip threshold | Net launch-date movement above which an event is recorded (illustratively two months) | Derived from CADRe schedule baselines | Months | Varied 1 to 4 months in the robustness battery (Appendix D) |
| Sensor archetype | First-of-kind active sensor (lidar/radar, no direct flight heritage) vs passive-radiometer heritage mission | NICM instrument taxonomy [\[65\]](#ref-65) | Binary (mixed/ambiguous = third stratum) | Effect modifier; crosswalk in Appendix C |
| Entry TRL | Technology-readiness level of the least-mature sensor technology at KDP-B | TechPort technology records | Ordinal, 1 to 9 | Key instrument-side covariate; measured before the slip window opens |
| Instrument mass / power / data rate | Recorded instrument parameters at confirmation | NICM / NICM-E [\[65\]](#ref-65) | Continuous | Instrument-side covariates |
| Number of distinct instruments | Count of separately developed science instruments on the mission | NICM / CADRe | Integer | Instrument-side covariate |
| Launch-vehicle class and provider | Vehicle family and launch-services provider at manifest assignment | CADRe / GAO / public manifest records | Categorical | Launch-side covariate |
| Shared-manifest indicator | Whether the mission shares a vehicle with another payload | CADRe / public manifest | Binary | Launch-side covariate |
| Provider-in-development indicator | Whether the assigned launch service is itself in active development at manifest assignment | Public launch-program records | Binary | Launch-side covariate; the exogenous-slip channel |
| Mission complexity index | Composite complexity score after the complexity-based cost-estimating-relationship method | Constructed after Bearden [\[17\]](#ref-17) | Continuous | Control; rules out archetype as a mere complexity proxy |
| Estimating-optimism proxy | Ratio of confirmation-baseline schedule to a reference-class median schedule for similar-class missions | Constructed after the reference-class-forecasting logic [\[57\]](#ref-57), [\[58\]](#ref-58) | Ratio | Control for the rival explanation |
| Calendar-period effect | Fixed effect for the acquisition-policy era of confirmation | Derived from KDP-B date | Categorical | Absorbs era-specific launch-market and policy shocks |

The instrument-side covariates are measured at or before KDP-B by construction, and this is what preserves the temporal ordering the identification strategy requires: a covariate fixed before the slip window cannot be an effect of the slip it is used to predict. The two constructed controls, the complexity index and the optimism proxy, defend against the two principal rival explanations. Their definitions are stated here in full so that a reviewer can judge whether the separation the dissertation reports is genuine or an artifact of how those controls were built.
### Appendix B: Slip-Event Cause-Coding Manual

The cause-coding manual is the most consequential operational artifact in the dissertation, because the competing-risks separation is only as trustworthy as the rule that assigns a slip to the instrument channel or the launch channel. The driver-to-consequence chain is explicit. A slip is recorded in the milestone period. The analyst reads the CADRe Part A narrative that attributes its proximate cause, then independently reads the GAO assessment narrative for the same project-year, and the slip is cause-coded only when the two sources agree on a single dominant cause. Where they disagree, or where neither names a dominant cause, the event is flagged as un-codable and held out of the primary contrast, entering only the recoding sensitivity analysis. This two-source reconciliation is the mechanism by which narrative subjectivity is bounded rather than eliminated. It cannot manufacture certainty where the record is silent, but it prevents a single analyst's reading of a single document from determining a competing-event assignment.

The coding rules are fixed in advance. First, instrument-driven slip is coded when the dominant narrative cause is instrument development: environmental-test failure, calibration-budget closure problems, detector or focal-plane maturation, or instrument delivery slip. Second, launch-driven slip is coded when the dominant narrative cause is launch-vehicle availability: manifest congestion, a shared-vehicle anomaly, or the launch provider's own development slip, in each case independent of spacecraft readiness. Third, a slip whose narrative names a spacecraft-bus, ground-system, or programmatic-funding cause is neither instrument-driven nor launch-driven and does not constitute a competing event of interest; the spell continues at risk. Fourth, a slip whose dominant cause cannot be reconciled across the two sources is flagged un-codable. The manual does not attempt to apportion a multi-cause slip across channels, because the competing-risks structure requires a single dominant cause per first-slip event. Multi-cause slips are exactly the un-codable population, and their handling is reported transparently rather than resolved by assumption.

The confidence calibration follows from the reconciliation design. Where both sources name the same cause, the coding is high-confidence; where only one source names a cause and the other is silent, the event is low-confidence and is recoded both ways in the sensitivity battery. The dissertation commits in advance to reporting any conclusion that flips between codings as fragile rather than confirmed, the operational expression of the design-stage honesty the work claims throughout.

### Appendix C: NICM-Taxonomy-to-Archetype Crosswalk

Appendix C makes the archetype assignment auditable by mapping the NICM instrument taxonomy onto the binary effect modifier. The crosswalk matters because the archetype is the variable that operationalizes the entire H1 contrast; if its assignment were opaque, the dominance claim would be untestable. The mapping rule is twofold. An instrument is classified as a first-of-kind active sensor when it is an active instrument (it transmits and senses its own signal, as a lidar or a radar does) and carries no direct flight heritage from a previously flown unit of the same design. An instrument is classified as passive-radiometer heritage when it is a passive radiometer (it senses naturally emitted or reflected radiation) and traces direct design heritage to a predecessor flight instrument. Missions whose primary instrument satisfies neither rule cleanly, including active instruments with substantial heritage and passive instruments that are themselves first-of-kind, are assigned to the mixed-or-ambiguous third stratum and excluded from the primary contrast, entering only the robustness check that confirms the binary contrast is not forcing a dichotomy onto a continuum.

The crosswalk is grounded in the archetype exemplars carried in the corpus: the spaceborne-lidar design literature [\[127\]](#ref-127) anchors the first-of-kind active-sensor pole, and the thermal-infrared-radiometer literature [\[45\]](#ref-45) together with the Landsat continuity record [\[72\]](#ref-72), [\[74\]](#ref-74), [\[91\]](#ref-91) anchors the passive-radiometer heritage pole. The crosswalk is applied to the mission's archetype-defining primary instrument, not to its full payload, because the effect modifier is a property of the dominant instrument risk the mission carries, not an average over secondary sensors. Where a mission carries more than one science instrument, the archetype is assigned from the instrument whose maturation most plausibly governs the critical path, and that assignment is recorded with its justification so that a second analyst can audit it.

### Appendix D: Pre-Registration Block

The pre-registration block is the frozen specification against which the executed analysis will be checked. It is reproduced here verbatim from the analysis plan so that the protocol is fixed before any cohort data are touched, the precondition for the falsification rule to have teeth. The estimator is the Fine-Gray proportional subdistribution hazard model, estimated separately for the instrument-driven event and the launch-driven event, with the cause-specific Cox model estimated in parallel for each event and Gray's test used as the nonparametric check on whether the cumulative incidence functions differ across archetype strata. For event type k, the subdistribution hazard is \( \text{subhazard}_k(t \mid X) = \text{subhazard}_{k0}(t) \exp(X \, \boldsymbol{\beta}_k) \), governing the cumulative incidence function \( \text{CIF}_k(t \mid X) \); the coefficient on the archetype-by-instrument-side interaction is the parameter of interest for H1.

The frozen settings are as follows. The primary slip threshold is two months of net launch-date movement, with the robustness range fixed at one to four months. Estimation uses a ridge-penalized partial likelihood with the penalty selected by cross-validation and an events-per-variable cap that bounds the number of free covariates given the small cohort (an expected 30 to 60 missions). The pooled specification adds the archetype-by-TRL and archetype-by-shared-manifest interactions. The robustness battery is fixed at five analyses: threshold variation across one to four months; both-way recoding of un-codable events; ridge-penalty and events-per-variable variation; removal and reintroduction of the estimating-optimism control; and introduction of the mixed-sensor third stratum as a separate group. The falsification rule is fixed. Confirmation requires separability (Gray's test rejects equality of cumulative incidence functions across archetype strata, or the cause-specific hazards differ), archetype dependence (the archetype-by-instrument-side interaction is distinguishable from zero with the predicted sign, namely the instrument-slip subdistribution hazard exceeds the launch-slip subdistribution hazard in the first-of-kind active stratum but not in the heritage passive stratum), and robustness (the pattern survives the optimism control and the ambiguous-event recoding). Failure of any one of the three falsifies the contribution, with the standing caveat that a non-rejection accompanied by wide intervals is reported as inconclusive rather than as support for the null.

### Appendix E: Minimum-Detectable-Effect Power Tables (Specified, Not Executed)

The power tables are specified in form and left unpopulated in value, because the cohort has not been assembled and no estimation has been run. This is a deliberate design-stage commitment, not an omission. The dissertation pre-registers that the power analysis is run before estimation rather than after, so that a non-rejection of the null can be interpreted honestly as either true non-separability or insufficient power, distinguished by the realized interval width. The table below fixes the cells that will be filled once the cohort size and the instrument-versus-launch event split are known; every numeric entry is shown as a placeholder to be computed, never as a result.

| Cohort size (n) | Event split (instr : launch) | Minimum detectable subdistribution HR (archetype interaction), alpha = 0.05, power = 0.80 |
|---|---|---|
| 30 | to be computed on cohort | to be computed (illustrative placeholder) |
| 40 | to be computed on cohort | to be computed (illustrative placeholder) |
| 50 | to be computed on cohort | to be computed (illustrative placeholder) |
| 60 | to be computed on cohort | to be computed (illustrative placeholder) |

The interpretive rule attached to the table is fixed in advance. If the minimum detectable subdistribution hazard ratio for the archetype interaction exceeds the effect size the spacecraft cost-and-schedule literature would regard as practically meaningful, the dissertation states ahead of estimation that a non-rejection of the null will be inconclusive rather than confirmatory. This pre-commitment is the protection against over-claiming from a small cohort, and it is the reason the table is presented empty: populating it with invented numbers would defeat the honesty the design-stage posture is meant to preserve.

### Appendix F: Source-Provenance Log

Appendix F records where the evidence base came from, in service of the transparency rule that a research deliverable must disclose its discovery path. The corpus of 149 entries was assembled from two sweeps. The brain sweep queried the Hall of Shoulders anchor brains for the structural-decomposition lens [\[82\]](#ref-82), [\[124\]](#ref-124), [\[111\]](#ref-111), [\[112\]](#ref-112), [\[113\]](#ref-113), [\[114\]](#ref-114) and the heterogeneity-aware separation lens [\[27\]](#ref-27), [\[118\]](#ref-118), [\[41\]](#ref-41), [\[134\]](#ref-134), [\[108\]](#ref-108), [\[33\]](#ref-33), and queried the ACTA_Papers full-text brain for the archetype exemplars in the spaceborne-lidar [\[127\]](#ref-127) and thermal-infrared-radiometer [\[45\]](#ref-45) literatures together with the flight-spares [\[36\]](#ref-36) and technology-roadmapping [\[130\]](#ref-130) papers; the AMOS_Papers and Space_Economy_Papers brains were queried and retained no net entries after relevance screening and de-duplication. The vault API sweep queried OpenAlex, Crossref, NASA Technical Reports Server, and arXiv among the free sources, and Semantic Scholar and Scopus among the keyed sources. IEEE Xplore was not entitled in this run and Springer Metadata was not added; the engineering-systems facet was recovered through the other sources, and the gap is flagged for a future entitled pull rather than concealed. Of the 149 retained entries, 133 carry a DOI and 16 are URL-only NTRS, OpenAlex, arXiv, or NIST records that all resolve; a 14-DOI random sample was verified against the Crossref registry with all fourteen confirmed, and one off-topic false positive that scored on generic method words was removed before the corpus was frozen. The restricted primary records that the design depends on, the CADRe data accessed through the ONCE database under a data-use agreement and the linked GAO and TechPort records, are not open-literature DOIs and are therefore described by their access path in the data chapter rather than enumerated in this reference list; their absence from the bibliography is a data-access dependency, not a citation gap.