Model-based standardization to adjust for confounding by cluster
Babette Brumback, PhD University of Florida
Model-based standardization uses a parametric statistical model to estimate an unconfounded population-average effect or association. With model-based standardization, one can compare expected group averages had the distribution of confounders been identical in each group to that of a standard population. Outcome-modeling and exposure-modeling are two commonly used approaches for model-based standardization. In addition, one can apply a doubly-robust approach that combines the outcome and exposure models and reduces the impact of model-misspecification of one or the other. We develop two outcome-modeling approaches based on generalized linear mixed models (GLMM) and generalized estimating equations (GEE). Our approaches are designed to handle a confounder, such as patient id or geographic area, that clusters the observations into a very large number of categories. We compare our outcome-modeling approaches to an exposure-modeling approach based on conditional logistic regression, and then compare both to a doubly robust approach. We illustrate the methods with 2014 Truven Health MarketScan Research Data to estimate proportions of acute respiratory tract infection (ARTI) diagnoses with an antibiotic prescription for emergency department versus outpatient visits, adjusting for confounding by unmeasured patient level variables and measured diagnosis-level variables.