UCLA

In multisite observational studies where level-one units are nested within level-two groups and treatment assignment occurs within (as opposed to between) groups, treatment assignment processes may vary between groups. A possible consequence of such group-varying treatment assignment processes is that the conditional exchangeability assumption may hold for some groups in a given sample while other groups are susceptible to omitted variable bias. This paper employs a simulation study to explore the potential for leveraging information about group-specific treatment assignment processes in order to mitigate omitted variable bias in hierarchical linear models with random intercepts and treatment effect slopes. The simulation demonstrates that an “augmented” model that incorporates information about group-varying treatment assignment can substantially reduce bias in treatment main effect estimates and also reduce the mean squared error of treatment random effect variance estimates.