| Although the propensity score method was developed 20 years ago, it has recently become a popular method in medical research. It is an analytic tool used to control for confounding resulting from non-random treatment assignment in observational studies. Researchers who use this method frequently estimate the propensity score from a multivariable logistic regression model, modeling treatment as a function of potential confounders. There is a large body of literature that provides guidelines for the development and assessment of logistic regression models. Yet it is unclear whether researchers adhere to these guidelines, and furthermore, whether these guidelines even apply to the propensity score.; In order to explore these questions, we conducted three studies. The first was a systematic review of the literature in which the propensity score method was employed to study a clinical question. The second study specifically focused on the usefulness of assessing model goodness of fit for evaluating propensity score models. In this study, we simulated data in order to test the sensitivity of summary measures of model fit when we omitted a confounder that was known to have a strong relationship with treatment assignment. In the final study, we employed the propensity score method to evaluate the effect of treatment with selective serotonin reuptake inhibitors compared to tricyclic antidepressants among elderly nursing home residents who were both depressed and had a cardiac condition. Our outcome of interest was cardiac events and death. Because of the nature of our sample and study design, it was important to control for many confounders.; Few studies that use the propensity score include information about how the propensity score model is developed and assessed. From the simulation study, summary measures of model goodness of fit are not useful in providing indications of when the propensity score is missing an important confounder. Additionally, using propensity scores missing important confounders will result in residual confounding in estimates of treatment effect. Finally, when we used a propensity score to control for confounding, we saw no difference in the rates of cardiac events or death by antidepressant class. |
