Confounder Misclassification in Claims-Based Pharmacoepidemiology: Methods and Application

Abstract

Pharmacoepidemiologists often rely on administrative insurance claims data when conducting observational studies of prescription drug safety and effectiveness. These data are advantageous for efficiently assembling large populations of patients treated in routine care, and when linked to pharmacy data provide accurate information on filled prescriptions. However, because they are collected for billing purposes, information useful to the conduct of observational studies may be omitted or recorded inaccurately. The objective of this thesis was to quantify the bias associated with confounder misclassification arising from both study design as well as inherent limitations of claims data, propose strategies to ameliorate bias due to misclassification, and to apply those strategies in a drug safety study. In Chapter 1, we compared fixed-duration versus all-available approaches to confounder measurement in settings where there was differential confounder misclassification due to differential data availability between exposure groups using simulations. We demonstrated that, contrary to prior simulation studies, using all available pre-cohort entry data to measure confounders is not always preferable to a fixed-duration approach, especially when there is a large discrepancy in the amount of available data. In Chapter 2, we considered that investigators working with claims often have multiple valid definitions for the same confounder which vary in their measurement characteristics (i.e. sensitivity and specificity); however, it is generally unclear which definition would remove more bias. We proposed two novel measurement methods combining a more sensitive and more specific confounder definition and evaluated each using simulations. We recommend one of our proposed approaches which reliably returned an effect estimate at least as unbiased as the better of the two component definitions. In Chapter 3, we compared new users of non-vitamin K antagonist oral anticoagulant drugs and new users of warfarin with respect to angioedema risk using a cohort study. We also compared warfarin to non-use using a case-crossover design. In the cohort study, we utilized an all-available approach to confounder measurement after finding the amount of available data was similar in each exposure group. Our findings were inconsistent with large increases in the relative rate of angioedema associated with oral anticoagulants.