Adjusting for confounding in population administrative data when confounders are only measured in a linked cohort.

Introduction: Analyses of population administrative data can often only be minimally adjusted due to a lack of control variables, potentially leading to bias due to residual confounding. Objectives: We aimed to use linked cohort data to help address residual confounding in analyses of population administrative data. One particular aim was to explore strategies for when linked cohort and population administrative data cannot be accessed together in the same environment (are "siloed"). Methods: We propose a multiple imputation-based approach, introduced through application to simulated data in three different scenarios related to the structure of the datasets. We then apply this approach to a real-world example - examining the association between pupil mobility (changing schools at non-standard times) and Key Stage 2 (age 11) attainment using data from the UK National Pupil Database (NPD). The limited control variables available in the NPD are supplemented by multiple measures of socioeconomic deprivation captured in linked Millennium Cohort Study (MCS) data. Results: In our real-world example, we included 509,670 individuals in the population NPD data, of whom 7,768 (1.5%) were MCS cohort members. The unadjusted estimate of -1.86 (95% CI -1.92, -1.81) for the association between pupil mobility and Key Stage 2 attainment was attenuated to -0.92 (95% CI -0.97, -0.88) through adjustment for the NPD control variables, and further attenuated to -0.76 (95% CI -0.86, -0.67) through adjustment for the MCS control variables. Conclusions: Linked cohort data can be used to address residual confounding in analyses of population administrative data, and our proposed approach performed well across a range of simulated and real-world scenarios. The underlying principles are widely applicable: any analysis of administrative data could potentially be strengthened by linking a subset of individuals into richer cohort data. More research is required to understand how these methods can be applied more broadly.