Comparison of multivariable-adjusted logistic regression model with propensity score techniques using pharmacy claims data

Abstract

Purpose To compare the multivariable-adjusted logistic regression model with the propensity score-matched, propensity score-stratified and propensity score-adjusted logistic regression models in estimating the effect of exposure to antidepressant agents in increasing the risk for type 2 diabetes mellitus. Methods A retrospective cohort study in the USA, using the Texas Medicaid prescription claims database was conducted from 1 January 1 2002 to 31 December 2009. Patients aged 18–64 years with new prescriptions for antidepressants (exposed group) or benzodiazepines (unexposed group) during the study period constituted the base population. Patients without diabetes at cohort entry were included in the study. Propensity scores, which predicted exposure to antidepressant agents, were used to create propensity score-matched, propensity score-stratified and propensity score-adjusted logistic regression models. Results A total of 44 715 patients formed the study sample. The risk estimates varied across different analytic methods. The propensity score-matched logistic regression model yielded the highest risk estimate (relative risk (RR) = 1.931; 95% confidence interval (CI): 1.705 to 2.187). The propensity score-stratified model (RR = 1.457; 95% CI: 1.127 to 1.884), the propensity score-adjusted regression model (RR = 1.491; 95% CI: 1.334 to 1.665) and the multivariable-adjusted logistic regression model (RR = 1.478; 95% CI: 1.323 to 1.653) yielded similar risk estimates. Conclusions Propensity score techniques using pharmacy claims data with a limited number of covariates yielded varied estimates of the treatment effect. The propensity score-matched model yielded a less biased treatment effect estimate than the multivariable-adjusted, the propensity score-stratified, and the propensity score-adjusted regression models.