Causal Inference Cheat Sheet for Practitioners

Turn observations into defensible causal insight — quick

Use this cheat sheet when you have observational data and a decision to make, but you cannot run a randomized trial. It helps you pick a practical quasi‑experimental strategy, document assumptions, run simple diagnostics, and prepare the results so leaders can act with appropriate caution.

Quick answer — which design when

• Randomized experiment (gold standard) — use if you can randomize assignment or rollout.
• Difference‑in‑Differences (DiD) — use for staggered or before/after changes across treated & control groups when parallel trends are plausible.
• Regression Discontinuity (RD) — use when treatment assignment sharply changes at a known cutoff (and manipulation around cutoff is unlikely).
• Instrumental Variables (IV) — use when you have a plausible instrument that shifts treatment but not outcome directly (exclusion holds).
• Matching / Propensity Scores — use to make treated and comparison groups more similar on observables; beware unobserved confounding.
• Synthetic Control — use for one treated unit with rich pre‑treatment data to construct a weighted control.

Key terminology (first time use explained)

• Counterfactual — what would have happened to the same unit had the treatment not occurred.
• Confounder — a factor that influences both treatment assignment and outcome and can bias naive comparisons.
• Instrument — a variable that causes variation in treatment but does not directly affect the outcome except through treatment (used in IV).
• Mediator — a variable on the causal path from treatment to outcome (not a confounder).
• External validity — whether an estimated effect generalizes beyond the studied sample or context.

Decision checklist — quick workflow

1. Write a clear causal question: who, what, when, and the policy or decision it informs. (Template below.)
2. Sketch a simple causal diagram (DAG): show treatment, outcome, key confounders, mediators, and instruments.
3. List assumptions required for candidate designs (parallel trends, no manipulation, instrument exogeneity).
4. Pick the simplest design that plausibly meets assumptions given your context and data quality.
5. Pre‑specify the primary test, covariates, window, and robustness checks (pre‑trends, placebo, sensitivity).
6. Run the primary analysis and then robustness checks. Document any fragile results or violations.
7. Translate effect sizes into practical impact and decision thresholds (costs, scale, risk tolerances).

Simple diagnostics & common threats (and what to do)

• Confounding — check covariate balance; if imbalance exists, control, match, or collect more data. Recognize limits if unobserved confounding is plausible.
• Selection bias — look for systematic differences in who receives treatment; consider IV or RD if selection is based on observables or a running variable.
• Violation of parallel trends (DiD) — plot pre‑treatment trends visually and test for pre‑trends; if failing, restrict window, reweight, or choose another design.
• Manipulation at cutoff (RD) — run density tests at cutoff and examine baseline covariates for discontinuities.
• Weak instrument (IV) — check first‑stage strength (F‑statistic); weak instruments produce unreliable estimates.
• P‑hacking & specification search — pre‑register or record analysis choices, and present robustness checks transparently.

Practical robustness checks (quick list)

• Pre‑trend / placebo (DiD): estimate fake treatment dates or earlier windows.
• Falsification outcomes: test outcomes that should not be affected.
• Covariate balance: compare observables before treatment; if imbalance persists, adjust and report.
• Sensitivity analysis: report how large an unobserved confounder would need to be to overturn conclusions (qualitative if formal methods not available).
• Alternative specifications: vary windows, functional forms, and control sets; report stability of effects.

Short templates you can copy

Causal question template: For population X, does treatment T during period P cause outcome Y within timeframe F, compared to alternative A, in context C?

Assumptions record (one line each): Design: _______, Key assumption: _______, Why plausible: _______, How to test: _______

Reporting checklist (share with decision‑makers): design, data window, sample, primary estimate (+ CI), pre‑trends, 2 robustness checks, key caveats, decision implications and uncertainty.

From estimate to decision — short guide

1. Translate effect into operational terms (e.g., additional events per 1,000 users; revenue per month).
2. Estimate uncertainty and worst‑case scenarios; use them for downside risk assessment.
3. Map alternative actions: pilot, scale, or withhold — and the evidence threshold for each.
4. Recommend a follow‑up plan: further data collection, randomized trial if feasible, or staged rollout with monitoring.

When not to attempt causal inference

• When key confounders are unobserved and no credible instrument or discontinuity exists.
• When data lack pre‑treatment observations needed for trend checks.
• When effect sizes are tiny relative to measurement error and practical decision thresholds.

Further reading (practical accessible works)

• Angrist & Pischke — Mostly Harmless Econometrics (practical methods)
• Imbens & Rubin — Causal Inference for observational studies (reference)
• Hernán & Robins — Causal Inference: What If (clear applied perspective)
• Short tutorials and DAG tools: Judea Pearl primers and DAGitty (for drawing DAGs)

Keep this cheat sheet handy

Before acting on observational evidence, use the checklist and the templates above to document exactly which assumptions your decision depends on, run at least two robustness checks, and present effect sizes in operational terms that matter to decision‑makers.

Image suggestion: causal inference diagram