Use DiD for staggered treatment adoption
Overview
This question evaluates proficiency in causal inference for panel data, specifically difference-in-differences with staggered adoption, heterogeneous and dynamic treatment effects, event-study design, pre-trend diagnostics, cluster-robust inference, and comparison with propensity-score matching.
Staggered DiD for a Weekly RPU Rollout (50 Regions, 2025-06-01 to 2025-08-15)
Context and assumptions:
- You have panel data at the region-week level with weekly revenue per user (RPU) as the outcome.
- A product/feature is rolled out once per region during 2025-06-01 to 2025-08-15 and remains active thereafter (no reversals). Assume no never-treated regions; at any week t, the valid controls are regions not yet treated (G_i > t). You also have a pre-period (e.g., Jan–May 2025) to test pre-trends.
- Goal: estimate the causal effect of the rollout on RPU in the presence of potentially heterogeneous and dynamic effects across cohorts and over time since adoption.
Tasks: (a) State when two-way fixed effects (TWFE) is biased under staggered adoption and what it identifies. (b) Implement a heterogeneous-effects-robust estimator (Callaway–Sant’Anna or Sun–Abraham), define group-time ATT ATT(g, t), and specify aggregation weights to obtain overall and dynamic effects. (c) Specify an event-study setup with a clear reference period and how you would visualize it. (d) Describe pre-trend testing (joint F-test on leads) and remedies if violated. (e) Explain inference with few clusters, including wild cluster bootstrap and handling serial correlation. (f) Explain how you would reconcile DiD estimates with a parallel propensity score matching (PSM) analysis.
Solution
Login required