Abstract
Studies using health administrative databases (HAD) may lead to biased results since information on potential confounders is often missing. Methods that integrate confounder data from cohort studies, such as multivariate imputation by chained equations (MICE) and two-stage calibration (TSC), aim to reduce confounding bias. We provide new insights into their behavior under different deviations from representativeness of the cohort. We conducted an extensive simulation study to assess the performance of these two methods under different deviations from representativeness of the cohort. We illustrate these approaches by studying the association between benzodiazepine use and fractures in the elderly using the general sample of French health insurance beneficiaries (EGB) as main database and two French cohorts (Paquid and 3C) as validation samples. When the cohort was representative from the same population as the HAD, the two methods are unbiased. TSC was more efficient and faster but its variance could be slightly underestimated when confounders were non-Gaussian. If the cohort was a subsample of the HAD (internal validation) with the probability of the subject being included in the cohort depending on both exposure and outcome, MICE was unbiased while TSC was biased. The two methods appeared biased when the inclusion probability in the cohort depended on unobserved confounders. When choosing the most appropriate method, epidemiologists should consider the origin of the cohort (internal or external validation) as well as the (anticipated or observed) selection biases of the validation sample.