Abstract:
|
A common modeling strategy for estimating treatment effects from non-experimental data is difference-in-difference (DiD) regression with fixed effects for the treated/untreated units over time. If treatment occurs at the group level, with many individual observations within each group, changes to the individuals within groups over time can make the fixed effects ineffective. Concerns about such within-unit changes, other than of treatment status, are typically addressed by including a linear additive specification of observed individual-level covariates in the fixed-effects regression. We propose a novel application of propensity/balancing score weights to supplement this strategy. In our application the units are large firms whose employee mixtures change over time. Within firm, we use boosted regression to create ATT weights for each person in each non-reference year so that the full weighted joint distribution of individual covariates matches that of the reference year. These weights are then applied in estimating the DiD regression to ensure that, on the observed covariates, the fixed effects are for units that stay fixed, reducing potential bias in treatment effects.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.