Performence of statistical methods to control for unmeasured confounding in pharmacoepidemiology: Focus on Instrumental Variable Analysis

Abstract

Randomized controlled trials are not always feasible to measure the effect of a treatment, for instance, to detect rare adverse drug reactions. In that case, observational studies are essential. However, in observational studies, the treatment effect can be biased due to possible confounding by patient characteristics. Several statistical methods, e.g., multivariable regression, can be used to control for confounding by measured confounders, but confounding by unmeasured patient characteristics remains a potential problem of observational studies. Alternative methods have therefore been proposed to control for unmeasured confounding. In this thesis, we focus on three of those methods: instrumental variable (IV) analysis, prior event rate ratio (PERR) adjustment, and the self-controlled case series (SCCS) design. First, we discuss the conceptual framework of IV analysis. Moreover, we reviewed different methods for IV estimation and provide guidance on when different methods may be applicable. Second, we focus on the performance of IV estimates in the cohort and the nested case-control (NCC) designs and the performance of balance measures (i.e., standardized difference) to falsify one of the three assumptions of IV, the assumption that the IV is independent of confounders. Our analyses show that IV estimates are very unstable and biased when the IV is weakly associated with the exposure. The variability of IV estimates is larger in the NCC study compared to a cohort study, which can be partly remedied by increasing the number of cases, or by increasing the number of controls per case. The standardized difference can be a useful tool to falsify the above mentioned IV assumption. Nevertheless, an observed balance of measured confounders between IV categories does not guarantee balance of unmeasured confounders. Third, two empirical studies of IV analysis based on general practice databases are described. The performance of IV analyses varied between time-fixed and time-varying exposures, across the databases, and strongly depends on the definition of IVs and the sample size of the study. Finally, the performance of the PERR adjustment method and the impact of violations of the assumptions of the SCCS design are studied. The PERR method can be applied to reduce bias due to unmeasured confounding; however, in particular situations (e.g., when prior events influence the probability of subsequent exposure) it may produce biased exposure effects. The SCCS analyses show that the incorrect definitions of observation periods (e.g., if subjects are censored at the event) may bias exposure effect estimates. In conclusion, although IV analysis appears a powerful statistical tool to control for unmeasured confounding, its validity and applicability in pharmacoepidemiology still have to be established particularly for studies of time-varying exposures. The PERR method and the SCCS design can be applied in case of time-varying exposure, yet these may provide biased estimates if confounders change over time and these methods may not be suitable for all possible clinical outcomes. Therefore, the assumptions of these methods should be routinely evaluated and statistical evidence as well as clinical knowledge is essential when interpreting the study results.