Testing for covariate balance in comparative studies.

摘要

In comparative studies, causal inference necessitates effective adjustments for important covariates. This becomes particularly relevant in observational studies, where covariates are rarely jointly balanced between treatment and control groups, and this lack of covariate balance can generate misleading results. Such adjustments as propensity score matching and stratification are frequently used to align dissimilar groups. The credibility of the analysis may be bolstered by demonstrating that such an adjustment has improved balance on observed covariates. It is not automatic that these measures achieve this objective. To appraise whether they have, practitioners use a variety of techniques, many of them common in hypothesis testing. However, some of these 'balance tests" lack formal motivation, may give contradictory messages about balance, and have in some cases been shown to have undesirable statistical properties.;We begin by identifying goals of balance testing in comparative studies and evaluating arguments used in the literature to support and oppose using significance tests to appraise covariate balance. We survey the literature for existing measures and appraisals of balance, with an interest in their advantages and shortcomings in relation to the goals we identify. We study the performance of some existing ways to assess covariate balance through an examination of contemporary research and identify that a permutation version of the balance-testing procedure originally suggested in Dehejia and Wahba (2002) has been shown to outperform some of the other approaches. We supplement our findings from the existing literature with a thorough simulation study based on real observational data. We use this simulation study to evaluate the impact of using the various balance diagnostics on the validity of statistical inferences about treatment effects.;Our literature review and simulation results suggest an important role for a new formal balance diagnostic. We develop several ideas aimed at this end and test them in various simulation settings that resemble authentic analysis conditions. Using the results of the literature survey and our simulation study, we are able to recommend new and existing techniques for testing covariate balance using randomization-based inference. We also propose dependable combinations of procedures for inference in comparative studies and provide some examples for application of our recommended techniques.