Abstract:
|
Matching is increasingly applied to quasi-experimental studies to improve the comparability of treatment and control groups. However, matching on time-varying covariates at baseline results in pruned groups that are extreme relative to their group mean. Thus, on subsequent measurement, measures of change in these covariates and the treatment outcome may be biased by regression artifacts. This Monte Carlo simulation study estimates the bias introduced by regression artifacts when matching on serially correlated covariates in a widely used study design, difference-in-differences. Simulations vary the level of serial correlation in the matching variables and the strength of the confounding relationship between the matching variables, treatment assignment, and the outcome.
|