Analytical Approaches to Reduce Selection Bias in As-Treated Analyses with Missing In-Hospital Drug Information.

Abstract

While much attention has focused on immeasurable time bias as a potential exposure misclassification bias, it may also result in potential selection bias in cohort studies using an as-treated (or per protocol) exposure definition in which patients are censored upon treatment discontinuation. We examined analytical approaches to minimise informative censoring due to the absence of in-hospital drug data using a case study of β-blocker use and mortality in heart failure. We conducted a cohort study using Korea's healthcare database, including inpatient and outpatient drug data. Using an as-treated exposure definition, patients were followed up until death, β-blocker discontinuation (in the exposed), β-blocker initiation (in the unexposed), or end of study period. In 'complete prescription' analysis using inpatient and outpatient drug data, we estimated hazard ratios (HR) and 95% confidence intervals (CI) using a Cox proportional hazard model. In outpatient drug-based analyses, we attempted to reduce the bias using stabilised inverse probability weighting (IPW) for treatment crossovers, hospitalisation, and all artificial censorings. An HR of 0.89 (95% CI 0.74-1.07) for β-blocker use versus non-use for all-cause mortality was found in 'complete prescription' analysis. Benefits were exaggerated when follow-up was assessed using outpatient drug data only (HR 0.71; 95% CI 0.57-0.89). Weighting by stabilised IPW for treatment crossovers and hospitalisation reduced the bias. When using an as-treated exposure definition, missing in-hospital drug data induced selection bias in our case study. Using IPW for censoring mitigated bias from the hospitalisation-induced censorings.