Autostrata: Improved Automatic Stratification for Coarsened Exact Matching

Abstract

We commonly adjust for confounding factors in analytical observational epidemiology to reduce biases that distort the results. Stratification and matching are standard methods for reducing confounder bias. Coarsened exact matching (CEM) is a recent method using stratification to coarsen variables into categorical variables to enable exact matching of exposed and nonexposed subjects. CEM’s standard approach to stratifying variables is histogram binning. However, histogram binning creates strata of uniform widths and does not distinguish between exposed and nonexposed. We present Autostrata, a novel algorithmic approach to stratification producing improved results in CEM and providing more control to the researcher.