Impact of Drug Exposure Definitions on Observed Associations in Pharmacoepidemiology Research.

BACKGROUND - A variety of methods are used to define exposure in pharmacoepidemiologic studies. Although each method has known biases, the relative effect of these biases on an observed association has not been fully examined. OBJECTIVE - To explore the influence of different exposure definitions on estimates, using the association between metformin and all-cause mortality as a proto-typical model. METHODS - New users of oral anti-hyperglycemic drugs were identified using administrative health databases from Alberta, Canada between 1998 and 2010. Drug exposure was described using definitions that are commonly used in observational studies. All analyses included the same covariates of age, gender, and a comorbidity score, and subjects not exposed to metformin served as the reference group. The measure of association was assessed using a Cox Proportional Hazards model for cohort studies and conditional logistic regression for case-control studies. RESULTS - We identified 64,293 new oral anti-hyperglycemic drugs users; mean age 68.9 years, 33,131 (52%) males, and 24,745 (39%) deaths during a mean follow-up of 6 years.  In adjusted models, the association between metformin and mortality ranged from 0.23 (95% CI 0.22-0.25) to 0.92 (95% CI 0.88-0.95) reduction. Most metformin exposure definitions, however, provided estimates in the 0.6-0.8 reduction range, aligning with the results of previous observational studies. CONCLUSIONS - The variety of exposure definitions tested in this analysis produced a wide range of associations between metformin and mortality risk. Therefore, pharmacoepidemiological studies should implement sensitivity analyses including at least two exposure definitions to provide more robust and potentially valid study estimates.