PURPOSE:  Since randomized controlled trials are not always feasible, observational studies are often relied upon to evaluate the comparative effectiveness of treatments in real-world settings. The ultimate goal of comparative effectiveness research is causal analysis to assess the relationship between a treatment and outcome. Appropriate control of confounding is essential to causal analysis, though correct model specification with respect to covariate selection can be a difficult task. The validity of causal inference based on observational data hinges on this fundamental yet untestable assumption which cannot be verified using statistical tests. Doubly robust (DR) estimation combines a model for the outcome (i.e., outcome regression) and a model for the treatment exposure (i.e., propensity score), offering investigators with two chances to fulfill the correct model specification assumption. Provided one of the models is correctly specified, the DR estimator will be robust to misspecification of the other model. That is, if either the outcome regression model or the propensity score model for the treatment is correctly specified, one can still get unbiased estimates of the association between treatment and outcome. DESCRIPTION:  The workshop will consist of four topics. First, confounding bias, model misspecification, general causal assumptions, and different adjustment techniques (including outcome regression and inverse probability of treatment weighting) will be discussed by way of introduction. Second, a conceptual overview of DR estimation will be presented, and estimators with the DR property, including the augmented inverse probability weighted estimator, will be explored. Third, a study using DR estimation to compare the effect of oral versus BRACE therapies on relapse in multiple sclerosis will be presented and discussed. Fourth, audience members will be invited to participate in building DR models (e.g., by suggesting covariates for control) for the case example as workshop moderators run live sensitivity analyses to demonstrate the implementation of DR estimation.