Miller (2023) sugggests that

  • Check regression output to ensure that no variables are being unexpected dropped.
  • Choice pre-event reference period
    • When a more extended reference period is used for normalizing to zero, the standard errors can be noticeably smaller.
    • When there is a dip in outcomes before the event
      • Do not use the period of the dip as counterfactual baseline
      • Use a period prior to the beginning of the dip
  • Report a combination of actual and counterfactual average outcomes separately for each unit type
    • (1) estimate the event study model; (2) “zero out” the event-time dummies and Make predictions; (3) average these predictions within calendar time for the treated units; and (4) plot out this counterfactual.
  • When the never-treated units are problematic comparisons for the treated units
    • Using a re-weighting or matching procedure prior to estimation of the event study
  • Choice event window
    • Limit the event window to bring more visual attention to the period of interest, even you have more data
    • Offer a robustness check
  • End-cap variables are common
    • The endpoint sometimes appear to be offset a bit from the rest of the graph
    • Plot the endpoint coefficients to indicate these are differently estimated
  • If there exist pre-trends, we can control for unit-specific trends, by including a (continuous) time variable interacted with unit dummies.

Statistical Inference Link to heading

  • Standard cluster-robust methods provide accurate standard errors only if the number of clusters is large enough (42 or 50 clusters by rules of thumb)
  • Too few clusters lead over-rejection

Reference Link to heading

Miller, D. L. (2023). An Introductory Guide to Event Study Models. Journal of Economic Perspectives, 37(2), 203–230.