Causal Discovery & Inference — Practitioner Primer

Welcome — why this matters

If you need to know whether a change will cause a difference (not just be correlated with one), this primer helps teams choose practical, responsible paths from observational data to decision‑ready causal insight. The goal is not technical perfection but defensible evidence you can act on: clear causal questions, explicit assumptions, appropriate analytic design, robustness checks, and a reproducible decision memo.

Quick decision checklist: do you need causal methods?

1. Is the question about the effect of a change, policy, or intervention? (If yes, causal methods are relevant.)
2. Can you run a randomized experiment within a reasonable timeframe, budget, and ethical constraints? (If yes, prefer an experiment.)
3. If an experiment is infeasible, can you find a plausible quasi‑experimental or natural experiment design? (If yes, proceed with causal inference; if no, limit conclusions to associations.)
4. Is there sufficient pre/post variation, good covariate data, or an instrument/cutoff that plausibly isolates exogenous variation? (If yes, a stronger causal design is possible.)

Core approaches — what they are, when to use them, and a simple workflow

Difference‑in‑Differences (DiD)

Use when a subset of units receives a change at a known time and others do not. Intuition: compare outcome changes over time between treated and control groups.

Simple workflow:

1. Define treated and control groups and pre/post windows.
2. Check pre‑trend: are trends parallel before treatment? (Visualize and test.)
3. Estimate DiD regression: outcome ~ treated + post + treated*post + covariates.
4. Run robustness checks: varying windows, weighted regressions, clustered SEs, event‑study.

Key assumption: parallel trends (absent treatment, groups would have followed the same trend).

Instrumental Variables (IV)

Use when a variable (instrument) causes variation in treatment but plausibly affects the outcome only through treatment. Intuition: use the instrument to isolate exogenous variation in the treatment.

Simple workflow:

1. Identify candidate instrument and justify its plausibility (relevance and exclusion).
2. First stage: regress treatment on instrument (+ covariates). Check F‑stat > 10 for weak instruments.
3. Second stage: regress outcome on predicted treatment from first stage.
4. Run overidentification (if multiple instruments), sensitivity, and heterogeneity checks.

Key assumptions: instrument relevance, exclusion restriction (no direct path from instrument to outcome), and monotonicity (for LATE interpretation).

Regression Discontinuity (RD) / Natural Cutoffs

Use when assignment to treatment changes sharply at a known cutoff (age, score, date). Compare units just above and below the cutoff.

Workflow:

1. Plot outcome vs running variable near cutoff; look for discontinuity.
2. Estimate local regression on both sides with appropriate bandwidth and polynomial order.
3. Check for manipulation of the running variable and balance of covariates near the cutoff.

Synthetic Control & Natural Experiments

When you have a single treated unit (e.g., one city or product line) and many potential controls, synthetic control builds a weighted combination of controls to approximate the counterfactual. Natural experiments include external shocks (weather, policy rollouts) that create plausibly exogenous variation.

Matching, Propensity Scores, and Regression Adjustment

Matching and regression can help reduce bias when treatment assignment depends on observables. Use them carefully: they rely on the assumption of no unobserved confounding (selection on observables). Combine with sensitivity analysis.

Compact assumptions checklist (explicitly document these)

• Clear causal estimand: Who/what is treated, what outcome, and what estimand (ATE, ATT, LATE)?
• Exchangeability / no unobserved confounders: Justify why treated and control are comparable or why the design isolates exogenous variation.
• No interference: One unit's treatment does not affect another's outcome (or document spillovers).
• Positivity / overlap: There is sufficient variation in treatment across relevant covariates.
• Design‑specific: parallel trends (DiD), valid instrument (IV), no running‑variable manipulation (RD).
• Model specification: functional form, clustering, and heteroskedasticity handling are appropriate.

Minimum evidence checklist before an internal decision

• Pre‑analysis plan or documented analysis choices (even short) to reduce researcher degrees of freedom.
• Visual diagnostics: time series, event studies, balance tables, density of running variable.
• Key statistical estimates with confidence intervals and effect sizes in business units.
• Robustness checks: alternative windows, controls, functional forms, placebo tests.
• Falsification tests: outcomes or groups that should not be affected.
• Sensitivity analysis for unobserved confounding (e.g., Rosenbaum bounds, or simple bounding exercises).
• Clear statement of external validity limits: who and where results likely generalize.

Practical integration with experiments

Always prefer randomized experiments when feasible. Use quasi‑experimental methods to:

• Inform experiment design (sample sizes, expected effect size ranges).
• Provide interim evidence when experiments are slow or infeasible.
• Triangulate with experiments: if both point to similar effects, confidence rises.

Operational workflow:

1. Pose the causal question and preferred estimand.
2. Decide: run an experiment, or identify the best quasi‑experimental design available.
3. Pre‑register analytic choices where possible and record reasons for deviations.
4. Run analyses, produce diagnostic plots, and document robustness checks.
5. Write a short decision memo summarizing assumptions, evidence, remaining uncertainties, and recommended action with risk notes.

Simple pseudo‑code (conceptual steps)

DiD (conceptual):

1. Compute mean outcome by group/time and plot trends.
2. Fit: Y_it = alpha + beta*Treated_i*Post_t + gamma_i + delta_t + epsilon_it
3. Report beta, clustered SEs, and event‑study coefficients for pre/trends.

IV (conceptual):

1. First stage: Treatment_i = a + b*Instrument_i + controls + e1
2. Second stage: Outcome_i = c + d*PredictedTreatment_i + controls + e2
3. Check first‑stage F, overid tests, and interpret LATE.

What to record in a decision memo

• Question and estimand clearly stated.
• Data sources, time windows, and sample construction.
• Chosen design and why (DiD, IV, RD, synthetic control, etc.).
• Key diagnostics and robustness checks (with figures attached).
• Effect estimate(s) with confidence intervals and practical interpretation.
• Assumptions, risks, and suggested next steps or experiments.

Common pitfalls to avoid

• Claiming causality from simple correlations without design or assumptions.
• P‑hacking and overfitting by trying many specifications without documentation.
• Ignoring external validity: local effects may not generalize.
• Misusing weak instruments or failing to check parallel trends.

Next steps and resources

If you want to go deeper, standard accessible references include:

• Angrist & Pischke — Mostly Harmless Econometrics (practical, applied focus)
• Imbens & Rubin — Causal Inference for more formal foundations
• Hernán & Robins — For epidemiologic perspectives

This guide aims to help teams decide whether a causal method is appropriate and to run a compact, defensible analysis. Use the checklist items to build a short pre‑analysis note and a reproducible script. When the next step is an operational template (pre‑registered plan, interactive checklist, or decision memo form), consider turning this guide into an interactive workflow so teams capture assumptions and build institutional memory.