Causal Inference

Skill — Causal Inference

When this skill activates

This skill activates when analyzing treatment effects, designing A/B tests with statistical rigor, building uplift models, or performing counterfactual reasoning. Use when you need to establish causation (not just correlation) between interventions and outcomes.

Mandatory actions when this skill is active

Before writing any code

1. Define the causal estimand precisely (ATE, ATT, CATE) and document all causal assumptions (SUTVA, ignorability, positivity)
2. Draw a Directed Acyclic Graph (DAG) identifying treatment, outcome, confounders, mediators, colliders, and instrumental variables
3. Select the appropriate causal inference method (RCT analysis, propensity scoring, instrumental variables, diff-in-diff, RDD, synthetic control) based on data structure and identification strategy
4. Calculate minimum detectable effect size and required sample size with power analysis (typically 80% power, 5% alpha)

During implementation

• Validate the overlap/common support assumption by examining propensity score distributions across treatment groups
• Test for balance on all confounders after matching/weighting (standardized mean differences <0.1, variance ratios 0.5-2.0)
• Implement sensitivity analysis to assess robustness to unmeasured confounding (e.g., Rosenbaum bounds, E-values)
• Use doubly robust estimators when possible (combining outcome regression and propensity scoring for consistent estimates)
• Check for heterogeneous treatment effects across subgroups and report conditional average treatment effects (CATE)
• Validate parallel trends assumption for difference-in-differences designs with pre-treatment period analysis
• Report confidence intervals using robust standard errors (clustered at appropriate level) or bootstrap resampling

After implementation

• Present results with effect sizes in original units (not just p-values) and practical significance interpretation
• Document all assumptions made, sensitivity analysis results, and potential violations of causal identification
• Create visualization showing treatment effect heterogeneity across covariates and subgroups
• Generate reproducible analysis pipeline with versioned data, code, and randomization logs

Self-check before task completion

• Causal DAG documented with clear identification strategy and testable implications
• Balance diagnostics show adequate covariate balance (SMD <0.1) after adjustment
• Sensitivity analysis demonstrates results are robust to reasonable violations of assumptions
• Confidence intervals and effect sizes reported in interpretable units with business context
• All code is reproducible with seed setting for randomization and resampling procedures