Causal Discovery & Inference — Quick Reference

Why this matters

Observational data are full of useful signals—but without a causal frame they can mislead. This guide helps you move from “what changed?” to “what caused it?” so you can design interventions, choose reliable levers, and avoid costly mistakes. It’s aimed at product, operations, healthcare, policy, and service teams who need practical, testable causal reasoning rather than formal proofs.

Start with a clear causal question

Good causal work begins with a crisp question framed in terms of a counterfactual. Example: “What would average customer retention have been last quarter if we hadn’t launched feature X?” Make the outcome, treatment, population, and time window concrete. If you can’t state the counterfactual, you don’t have a causal question yet.

Core concept: directed acyclic graphs (DAGs)

DAGs are a compact way to surface assumptions about which variables cause which. Draw the key variables and arrows for causation, not correlation. Use DAGs to identify confounders you must adjust for, instruments you might exploit, or variables you must avoid conditioning on (colliders).

Quick tips:

• Label variables: treatment, outcome, measured confounders, unmeasured confounders, instruments, mediators.
• Ask: is there a backdoor path (a noncausal path) from treatment to outcome? If yes, you must close it by design or adjustment.
• Watch for colliders—conditioning on them can create spurious associations.

Common practical methods (intuition, assumptions, and what to check)

Quick decision guide: which approach fits common constraints?

1. If you have a credible external shock or cutoff that assigns treatment as‑if randomly → consider RDD (cutoff) or natural experiment approaches.
2. If treatment occurs for a group at a specific time and you can observe a comparable control group → Difference‑in‑Differences (with pre‑trend checks).
3. If unobserved confounding is likely and you can find a variable that shifts treatment but not outcomes directly → Instrumental Variables.
4. If confounding is likely but you have rich covariates → Matching/propensity methods plus sensitivity analysis.
5. If one aggregate unit is treated (region, state) → Synthetic control, with placebo tests and donor pool checks.
6. If none of the above are available → be transparent: estimate associations, state assumptions, and treat findings as hypothesis‑generating rather than definitive causal claims.

Robustness & validation checklist (use before acting)

• State causal question and counterfactual explicitly.
• Draw a DAG to document assumptions and potential confounders.
• Run pre‑trend and placebo tests where applicable.
• Check balance on pre‑treatment covariates and composition stability.
• Test sensitivity to model specification, bandwidths, and functional forms.
• If using IV, report first‑stage strength and discuss the exclusion restriction transparently.
• Report heterogeneous effects and consider external validity limits.
• Document all choices, diagnostics, and plausible alternative explanations in a short reproducible notebook.

Short examples

Product: A/B tests are ideal; when randomization breaks, an event that changes only some users (a rollout threshold) might permit DiD or RDD.

Operations: A supply disruption that affects some lines but not others can be a natural experiment; use synthetic controls if treated unit is unique.

Healthcare: Policy changes by region often invite DiD or synthetic control; physician prescribing thresholds can be exploited through RDD.

Next practical steps

1. Draft your causal question and draw a simple DAG. Share it with a teammate for critique.
2. Inventory available data: units, time granularity, potential instruments, and pre‑periods.
3. Select a candidate method using the decision guide. Run the basic diagnostics in the robustness checklist.
4. Document results, assumptions, and a short risk memo describing threats to validity before acting.

Recommended resources

• Angrist & Pischke — Mostly Harmless Econometrics (practical intuition)
• Imbens & Rubin — Causal Inference for Statistics (strong foundations)
• DAGitty (online DAG tool) for drawing and testing DAG‑based adjustment sets

Parting guidance

Causal answers require assumptions. Be explicit, run sensible diagnostics, and treat surprising causal claims as hypotheses to validate with new data or a designed experiment where feasible. When experiments are possible, prefer them; when not, rigorous quasi‑experimental methods plus transparent reporting are the next best path.