Your review meeting
A teammate says,“The adjusted result is positive, so we should launch.” Prepare a one-page assumption-led review using Elm 1–3. No real launch, customer intervention or automated decision is made. Deliver estimand, design, observed contrast, identification argument, threats, sensitivity and a bounded recommendation.
Write an assumption ledger
For each assumption, state why it is needed, what observed evidence can inform it, what cannot be checked, and what happens if it fails. Include intervention consistency, measurement/follow-up, interference, exchangeability or parallel trends as appropriate, and positivity where relevant. Do not merge randomized and observational identification arguments into one generic checklist.
Reference solution
The stratified table illustrates composition-driven association reversal; its adjusted contrast needs substantive identification assumptions. The randomization fixture gives an assignment-based estimate under a declared complete-randomization design, not an automatically identified receipt effect. The DiD fixture gives a contrast requiring an untreated parallel-trend argument; a zero pre gap change is not proof. Each result answers a different question.
Communicate evidence and limits
Keep estimate, units, target population, time, design and assumptions together. A p-value concerns a specified null/test/design; sensitivity values concern assumed departures. Neither proves a practically useful effect, transportability to another population or operational safety.
analysis | numerical result | interpretive limit
stratified association | crude +0.3667, standardised -0.10 | adequate adjustment unproven
random assignment | ITT +2; sharp-null p .2 | hypothetical design, tiny fixture
DiD | +4 | parallel untreated trends needed
DiD delta4 scenario | 0 | assumption, not confidence boundA useful next-evidence plan
Specify which additional pre-treatment causes, historical periods, assignment logs, exposure records, outcome completeness checks or independent review could strengthen the design. Avoid collecting unnecessary personal information. If a design cannot support the desired causal estimand, describe the association honestly and propose a more credible prospective evaluation with qualified oversight.
Failure rehearsal
A colleague filters the randomized fixture to people who attended, or drops incomplete panels without reporting attrition. Explain how this changes the estimand or observed composition. Preserve the original dataset, record exclusions and reasons, and rebuild the stated analysis. Do not make a clean dashboard substitute for a design audit.
Honest portfolio statement
“Implemented local causal-reasoning exercises on fictional fixtures, with estimand definitions, assignment-preserving randomization enumeration, DiD arithmetic and trend sensitivity.” Do not claim a real randomized study, measured policy benefit, removal of all confounding, certification or employment outcomes.
Reviewer rubric and interview prompts
Score 0–2 for estimand, design-specific assumptions, numerical correctness, sensitivity/limitations and honest communication. Explain why more covariates can hurt; why post-treatment selection matters; why a pretrend check is not a counterfactual; and why effect-of-assignment differs from effect-of-receipt. Record an unresolved assumption and the next evidence that could challenge it.
Authoritative reference
Miguel A Hernán and James M Robins, Causal Inference: What If. Public author resources remain public. Supporter value is the original offline workflow and review materials; no public book or author endorsement is presented as exclusive.
Try the activity before reading its reference. Keep your sources, commands and assumptions with your work. Companion activities are unsaved self-practice and do not award a certificate or change lesson progress.