Propensity Score Matching in Nonrandomized Studies: A Concept Simply Explained Using Antidepressant Treatment During Pregnancy as an Example

Abstract

In prospective and retrospective observational studies, such as those that examine the effects of antidepressant drugs for the treatment of depression during pregnancy, patients are not randomized to whatever treatments they do or do not receive. As a result, treatment groups may be unbalanced for a wide range of sociodemographic and clinical variables. In such studies, because the illness (and its correlates) for which the treatment is indicated may itself influence the study outcomes, the use of the treatment becomes indirectly linked to these outcomes (this is known as confounding by indication), and the effects of treatment and illness on the study outcomes are difficult if not impossible to separate. Case-control research designs, regression analyses that adjust for independent variables, and propensity score matching are ways in which baseline differences between groups and confounding by indication are statistically addressed. This article examines the concepts involved, explains what is done in propensity score matching procedures, and discusses the advantages and limitations of propensity score matching. Whereas propensity score matching can substantially reduce baseline differences between groups in observational studies, it can never correct for unmeasured confounds; therefore, cause-effect relationships can never be deduced from such studies. However, in situations in which gold standard randomized controlled trials are impractical, researchers must make do with statistical approaches such as propensity score matching.