Primary Submission Category: Difference in Differences, Synthetic Control, Methods for Panel and Longitudinal Data
Parallel Trends in an Unparalleled Pandemic: Difference-in-differences for infectious disease policy evaluation
Authors: Alyssa Bilinski, Shuo Feng,
Presenting Author: Alyssa Bilinski*
Researchers frequently employ difference-in-differences (DiD) to study infectious disease policy. DiD assumes that treatment and comparison groups would have moved in parallel in expectation, absent the intervention (“parallel trends assumption”). Our work formalizes often unaddressed epidemiological assumptions required for common DiD specifications, assuming an underlying Susceptible-Infectious-Recovered (SIR) data-generating process, and proposes more robust specifications. We first demonstrate that popular specifications can encode strict assumptions: DiD modeling incident infections or rates will produce biased treatment effect estimates unless untreated potential outcomes for both groups come from a data-generating process with the same initial infection and transmission rates. Modeling log incidence or growth allows for different initial infection rates, but invokes conditions on transmission parameters. We propose alternative specifications based on epidemiological parameters — the effective reproduction number and the effective contact rate — that are both more robust to differences between groups and can be extended to more complex transmission dynamics. In power analyses, we highlight minimal differences between incidence and log incidence models; our alternative specifications have lower power than incidence or log incidence, but higher power than log growth. We illustrate practical implications re-analyzing published studies of COVID-19 mask policy.