To determine the causal effect of a treatment using observational data, it is important to balance the covariate distributions between treated and control groups. However, achieving balance can be difficult when treated and control groups lack overlap. In the presence of limited overlap, researchers typically choose between two types of methods: 1) methods (e.g., inverse propensity score weighting) that imply traditional estimands (e.g., ATE) but whose estimators are at risk of variance inflation and considerable statistical bias; and 2) methods (e.g., overlap weighting) which imply a different estimand, thereby changing the target population to reduce variance. In this work, we introduce a framework for characterizing estimands by their target populations and the statistical performance of their estimators. We introduce a bias decomposition that encapsulates bias due to 1) the statistical bias of the estimator; and 2) estimand mismatch, i.e., deviation from the population of interest. We propose a design-based estimand selection procedure that helps navigate the tradeoff between these two sources of bias and variance of the resulting estimators. Our procedure allows the analyst to incorporate their domain-specific preference for preservation of the original population versus reduction of statistical bias. We demonstrate how to select an estimand based on these preferences by applying our framework to a right heart catheterization study.
翻译:暂无翻译