Prespecified Adjusted Estimation Plan¶
Protocol — instantiates Baseline Covariate Balance Verification
A pre-registered rule that fixes, before any outcome is seen, which baseline covariates the effect estimate will adjust for and how — so adjustment corrects imbalance without becoming a fishing license.
A Prespecified Adjusted Estimation Plan is the ex-ante response to the possibility of imbalance. Before any outcome is unblinded, it commits the analysis to a covariate-adjusted estimate — naming the covariate set, the model form (commonly ANCOVA on strong baseline predictors), and a hard boundary excluding anything measured after exposure — so that if imbalance turns up, the correction is a rule executed, not a choice improvised. The distinction is everything: adjustment chosen after seeing which covariates move the result is indistinguishable from data-dredging, while the same adjustment fixed in advance is a credible remedy. This plan is the mechanism that makes covariate adjustment honest by locking the response before the data can whisper which response would be convenient.
Example¶
A behavioral-nutrition trial tests whether a meal-planning app reduces six-month weight change, and its pre-registration — filed before unblinding — specifies the primary analysis as an ANCOVA adjusting for exactly two covariates: baseline weight and a baseline physical-activity score, chosen because they are the strongest known predictors of the outcome. The plan draws an explicit line: no post-randomization variable (mid-study adherence, for instance) may enter the adjustment set. When the data arrive, baseline weight turns out mildly imbalanced between arms. The pre-set ANCOVA absorbs the imbalance as a matter of course, and — crucially — because the adjustment was fixed months earlier, no reviewer can allege the team went looking for the covariate set that flattered the app. The remedy is trusted precisely because it was blind to the result.
How it works¶
- Rank prognostic covariates before outcomes, and name the adjustment set by predictive strength rather than by what improves the estimate.
- Specify the model — the adjustment form (ANCOVA, covariate-adjusted regression, or prespecified stratification) and how the covariates enter it.
- Draw the post-treatment exclusion boundary — a hard rule that nothing measured after exposure may join the adjustment set, since conditioning on a post-treatment variable (a mediator or collider) reintroduces bias.
- Register and lock the plan, then, on observed imbalance, execute it rather than negotiating a new analysis.
Tuning parameters¶
- Covariate-set size — a parsimonious set of top predictors vs. a broad set. Broad adjustment covers more imbalance but spends degrees of freedom and courts overfitting.
- Model form — ANCOVA vs. regression adjustment vs. prespecified stratification. Each trades interpretability against flexibility.
- Prognosis basis — how covariate strength is judged (prior evidence, blinded baseline data, domain theory). Better priors sharpen the set; weak ones misprioritize it.
- Degree of pre-commitment — a fully locked plan vs. prespecified guidance with narrow discretion. Locking maximizes credibility but cannot adapt to genuine surprises.
When it helps, and when it misleads¶
Its strength is that it converts covariate adjustment from a suspicious post-hoc move into a credible prespecified one, protecting the estimand and precision at once: imbalance on a strong predictor is handled by a rule everyone agreed to before the data existed.
Its failure modes are sharp. Adjusting for a variable measured after exposure silently reintroduces the very bias adjustment was meant to remove, which is why the post-treatment exclusion boundary is load-bearing rather than decorative. And a plan that is prespecified in name but loose in substance still leaves room for the garden of forking paths — the many defensible analytic choices that, taken after seeing the data, quietly bias the result.[n1] The classic misuse is quietly adding a covariate to the model after peeking at the outcome. The guarding discipline is to lock the plan, register it, and honor the post-treatment boundary without exception, so the adjustment answers to a commitment rather than to the result.
How it implements the components¶
imbalance_response_pathway— it is the pre-chosen response: covariate-adjusted estimation invoked when imbalance appears, rather than an improvised reaction.prognostic_variable_priority_weighting— the adjustment set is selected by prognostic strength, so the covariates that matter most for the outcome are the ones corrected for.post_treatment_exclusion_boundary— a hard rule bars any variable measured after exposure from the adjustment set, protecting the estimand.
It computes no balance metric or figure — those are Standardized Mean Difference Table and Covariate Balance Plot — and it does not itself audit assignment integrity, which is Randomization Integrity Audit.
Related¶
- Instantiates: Baseline Covariate Balance Verification — it is the prespecified remedial pathway that handles imbalance without compromising the causal contrast.
- Consumes: Standardized Mean Difference Table — the imbalance signal whose breaches trigger the prespecified adjustment.
- Sibling mechanisms: Standardized Mean Difference Table · Baseline Characteristics Table · Covariate Balance Plot · Randomization Integrity Audit · Stratified Balance Check · Automated A/B Balance Dashboard · Balance Exception Report
Editorial Notes¶
Form Classification¶
Form family: Rule, Policy & Commitment
Rationale: The mechanism precommits covariates, model form, exclusion boundaries, and deviations before outcomes, constraining future estimation conduct.
Nearest alternative: Representation, Specification & Plan — The plan is written, but its operative force is the standing rule preventing post-outcome fishing.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Specialized
Rationale: Prespecified Adjusted Estimation Plan is most plausibly rooted in the statistics_experimental_design tradition because its characteristic form depends on probability, calibrated inference, experimental design, and uncertainty analysis. The assignment tracks that formative lineage, not the many settings in which the mechanism can now be applied.
Related originating lineages:
- Medicine & Healthcare — Randomized clinical trials materially institutionalized prespecified adjusted analyses in protocols and analysis plans.
Review resolution: Both blind reviewers agree that statistics experimental design is the primary origin. Explicit reconciliation resolves alternate origin disagreement. Formative alternate lineages are retained as medicine_healthcare; later breadth of use is recorded separately as domain_reach=specialized, while origin_mode=cross_disciplinary_synthesis describes the relationship among origin lineages.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
A prespecified adjustment corrects for imbalance; it does not certify that the original comparison was clean. The two must not be conflated: adjustment is a remedy and an estimand choice, not proof that randomization worked. This is why the plan is downstream of the metric mechanisms that detect imbalance, not a substitute for them.
[n1] The garden of forking paths, a phrase from Andrew Gelman and Eric Loken, names the way many individually reasonable analytic decisions — which covariates, which model, which subset — can, when made after seeing the data, produce a biased result even without deliberate p-hacking. Full prespecification is the standard defense, which is why a merely nominal plan does not close the loophole. ↩