Associate Professor McGill University, Quebec, Canada
Background: Most administrative databases used in pharmacoepidemiologic research lack information on in-hospital drug use, leading to the misclassification of exposed patients as unexposed and introducing immeasurable time bias. Although this bias has been assessed in pharmacoepidemiologic studies analyzed using traditional modeling approaches, it has not been assessed when using marginal structural models (MSMs) to evaluate causal effects of time-varying exposures. In addition, little is known about the effectiveness of different methodological approaches to minimize immeasurable time bias in this setting.
Objectives: To quantify the magnitude of immeasurable time bias and assess the effectiveness of methodological approaches in addressing this bias when examining the association between metformin use and all-cause mortality among patients with type 2 diabetes (T2D) and nonalcoholic fatty liver disease (NAFLD) using MSMs.
Methods: We used nationwide administrative healthcare data from Korea to construct a cohort of patients with T2D and NAFLD. We used marginal structural models to estimate the effect of metformin use on all-cause mortality, controlling for sociodemographic factors, comorbidities, comedications, and diabetic complications. We estimated the ‘gold standard’ treatment effect using both inpatient and outpatient drug dispensing records and determined the amount of immeasurable time bias present by repeating the analyses using outpatient dispensing only. To compare the ability of different methodological approaches to minimize immeasurable time bias, we evaluated seven methodological approaches including adjusting for hospitalization as a binary variable, weighting by the proportion of measurable days, weighting by the probability of hospitalization, weighting by the inverse probability of hospitalization, assuming to be exposed during hospitalization if patients exposed prior to admission, conventional multiple imputation and Heckman's two-step imputation.
Results: The cohort included 12,545 patients and the gold standard odds ratio (OR) was 0.68 (95% CI 0.54-0.88) while the immeasurable time-biased OR was 0.51 (0.39-0.66). Among the approaches, adjustment for hospitalization (0.71; 0.54–0.95), assumed exposure during hospitalization (0.61; 0.48–0.79), two approaches to multiple imputations (0.81; 0.54–1.05, 0.70; 0.63–0.77, respectively) most effectively reduced bias.
Conclusions: Immeasurable time bias can result in substantial bias in pharmacoepidemiologic studies using marginal structural models. Adjustment for hospitalization, exposure continuity assumption, two multiple imputations appears to reduce the immeasurable time bias in the absence of inpatient medication data.
