Evaluate Propensity Score Matching Alternatives and Diagnostics
Overview
This question evaluates causal inference skills focused on propensity score methods, balance diagnostics, overlap/positivity assessment, estimand selection (ATE vs ATT), and variance estimation in observational studies.
Context
You are reviewing an observational study that used Propensity Score Matching (PSM) to estimate the causal impact of a UI change on user watch time. Randomized experimentation was not feasible, so historical logs and user covariates were leveraged to construct a matched sample and estimate an ATT (average treatment effect on the treated).
Task
Answer the following about best practices in PSM for product analytics:
- Why is standardized mean difference (SMD) ≤ 0.1 often used as a post-matching balance threshold?
- If logistic regression is not appropriate for the propensity model, what alternatives would you consider and why?
- How would you diagnose residual confounding after matching?
- Describe one method to estimate the variance of the treatment effect under PSM and when it is appropriate.
Hint: Discuss overlap/positivity, balance diagnostics (including higher moments and transformations), causal estimands (ATE vs ATT), and robust variance estimation.
Solution
Login required