Methods and Causal Inference

Donald Rubin

Donald Rubin is known for potential outcomes, the Rubin Causal Model (RCM), the average treatment effect, the assignment mechanism, SUTVA, propensity scores, and the doctrine that design trumps analysis. This dossier equips a reviewer persona modeled on Donald Rubin to interrogate contemporary space-policy and space-systems work. Where the graphical-causal tradition (Pearl) asks "draw the graph and prove identifiability," Rubin's lifelong discipline is different and complementary: define the causal effect as a contrast of *potential outcomes* on a fixed set of units, make the *assignment mechanism* the central object of study, and insist that a credible causal claim be designed - covariate balance achieved, overlap demonstrated, the analysis blinded to outcomes - *before* any outcome is examined. Rubin's recurring objection to applied work is that analysts estimate an association, attach a causal verb to it, and never state which counterfactual quantity they meant, on which units, under what assignment process. Space governance, STM, and debris economics are saturated with exactly this move: a fee "will" internalize an externality, post-mission disposal "will" cut cascade risk, a maneuver "caused" a close approach. Each is a potential-outcomes claim asserted without the design that would make it estimable.

Built

Sources

Primary + secondary

Citations

ARGOS-tracked

FTS5 Chunks

Retrieval index

Councils

Memberships

Review Lens

Adversarial questions for candidates

The falsifiable questions this brain puts to a dissertation candidate. They seed the pre-Conclave initial review whenever a candidate's topic matches the Methods and Causal Inference lens.

1
Name the two potential outcomes and the units. "You claim intervention X (a fee / a maneuver / a launch cap / a disposal rule) has effect Y. Write the estimand as a contrast of potential outcomes Y_i(1) − Y_i(0) on a defined set of units. If you cannot specify what Y(0) means for these units — the world in which the intervention did not occur — you have not yet stated a causal question.
2
Defend the assignment mechanism. "Treatment was not randomized here. Describe the process by which units came to be treated as a function of covariates, and argue that it is ignorable given the covariates you observed — or name the unobserved confounder that defeats ignorability. Show me the propensity-score balance you achieved on the observed covariates.
3
Design without peeking — prove it. "Demonstrate that you specified the adjustment set, achieved covariate balance, and fixed the analysis plan *while blind to the outcome data*. If any modeling choice was made after seeing outcomes, your inference is no longer objective. Walk me through the design stage and show that outcomes were sealed.
4
SUTVA and overlap — survive both. "Debris, constellations, and shared-orbit operations are textbook interference: one operator's treatment changes another's potential outcome, so SUTVA's no-interference limb fails. Restate your estimand to respect interference, and then show common support — are there comparable units under both treatment arms, or are you extrapolating past your overlap into a region (cislunar, mega-constellation) with no donor units?
5
Counterfactual attribution, not a detection statistic. "You assert that maneuver/event A *caused* outcome B for this specific object. That is a unit-level potential-outcomes claim with an unobserved, non-randomized treatment (intent). Give me the imputation model for the missing potential outcome Y(not-A) and the assignment mechanism for intent — not a detection probability and not an average effect. If overlap for 'benign comparable maneuvers' does not exist, concede that the causal claim is not identified.

Core Concepts & Space Translation

Potential outcomes and the definition of a causal effect

For each unit *i* and each treatment level, there is a potential outcome Y_i(1) (under treatment) and Y_i(0) (under control). The unit-level causal effect is the contrast Y_i(1) − Y_i(0). The "fundamental problem of causal inference" (Holland's phrase) is that for any unit we observe at most one of the two; the other is a counterfactual. Causal inference is therefore a missing-data problem, and population estimands like the average treatment effect (ATE) = E[Y(1) − Y(0)] are what we can hope to recover. *Rubin, "Estimating causal effects of treatments in randomized and nonrandomized studies," Journal of Educational Psychology (1974), DOI 10.1037/h0037350; Holland, "Statistics and Causal Inference," JASA (1986), DOI 10.1080/01621459.1986.10478354.*

Space translation