Abstract

Over the past few years, factor analysis (FA) has emerged as one of the most broadly used methods for estimating the impact of an intervention using observational time-series data on multiple units. This is due to the model’s ability to adjust for observed confounders and to allow for the presence of unmeasured confounding with a particular structure. In this talk, we present several variants of the standard causal FA model (including dynamic FA, multi-group FA and multi-study FA) and argue their usefulness for estimating causal effects in certain settings with limited data. Further, we discuss how FA can be extended to model the dependence of causal effects on a unit’s characteristics (modifiers). Finally, we demonstrate how fitting these models under the Bayesian paradigm leads to straightforward quantification of uncertainty for causal quantities of interest and can ensure data-driven model parsimony by exploiting regularizing priors on the model parameters.