psm-causal-inference

name

psm-causal-inference

description

Use when the user needs propensity score methods for causal inference, including propensity score matching (PSM), inverse probability weighting (IPW), overlap/common support checks, caliper or nearest-neighbor matching, covariate balance diagnostics using standardized mean differences, ATT/ATE estimation, sensitivity analysis, and clear reporting for observational treatment effect studies in Python, R, Stata, or SQL-backed analytics workflows.

PSM Causal Inference

Use this skill for observational treatment effect questions where the user wants to compare treated and untreated units after adjusting for observed confounding with propensity scores.

Workflow

1. Define the estimand before code:

   • Treatment, outcome, unit of analysis, time window, and eligible population.
   • Target estimand: ATT for effect on treated units, ATE for the whole population, or ATC for untreated units.
   • Confounders must be pre-treatment variables only.

2. Build the propensity model:

   • Estimate P(T = 1 | X) with logistic regression, regularized logistic regression, gradient boosting, random forest, or another calibrated classifier.
   • Do not include post-treatment variables, mediators, colliders, or outcome-derived features.
   • Inspect overlap/common support before matching or weighting.

3. Choose adjustment strategy:

   • Nearest-neighbor matching for interpretable matched cohorts.
   • Caliper matching when poor matches would bias estimates; common default is 0.2 * SD(logit(propensity_score)).
   • Matching with replacement when control pool is small.
   • IPW/stabilized weights when the goal is ATE and overlap is adequate.
   • Trimming when propensity scores are near 0 or 1 and positivity is violated.

4. Validate balance:

   • Report standardized mean differences (SMD) before and after adjustment.
   • Aim for absolute SMD below 0.1; stricter studies may use 0.05.
   • Check variance ratios for continuous covariates and category balance for categorical covariates.
   • Produce a love plot or balance table when useful.

5. Estimate treatment effect:

   • Use matched-pair outcome differences for 1:1 ATT matching.
   • Use regression on the matched/weighted sample for adjusted estimates when appropriate.
   • Use robust, bootstrap, or cluster-aware standard errors depending on the design.
   • Report uncertainty: confidence intervals, p-values if needed, and practical effect size.

6. Stress test the conclusion:

   • Try alternative propensity specifications.
   • Vary caliper width and matching ratio.
   • Compare matching with IPW or doubly robust estimation when feasible.
   • State residual risks from unmeasured confounding.

Reporting Checklist

• Estimand, inclusion criteria, treatment timing, and outcome window.
• Propensity model covariates and why they are pre-treatment confounders.
• Overlap/common support diagnostics.
• Matching or weighting choices, including caliper, replacement, ratio, and discarded units.
• Balance before and after adjustment.
• Treatment effect estimate with uncertainty.
• Sensitivity analyses and unmeasured-confounding caveats.

Guardrails

• PSM adjusts only for observed confounders.
• Good predictive propensity scores are not enough; balance is the goal.
• Avoid matching on variables caused by treatment.
• Do not present matched sample results as population-wide ATE unless the design supports it.
• If overlap is weak, say the causal question is not well supported by the data.