Abstract
Self selection into cohort studies and selective loss to follow up potentially undermine the representativeness of exposure outcome associations. Previous analyses using large cohort studies reported few significant differences between sample and population estimates for some exposure outcome associations, and suggested that selection bias might not be a problem. The aim of the current article is to examine presence and magnitude of bias due to self selection and loss to follow up in a large population based cohort study. We employ the structural approach to selection bias to (a) highlight why previous results do not generalize beyond the examined exposure-outcome associations, (b) show that selection bias is not unlikely in studies involving mental health related exposures and outcomes, and (c) explain why inverse probability weighting (IPW) but not adjustment for selection variables can more generally correct bias. Further, we propose to estimate the magnitude of selection bias by comparing results of IPW weighted and non-weighted analyses and show how evidence for the absence of bias can be obtained in the Bayesian framework using the concept of the region of practical equivalence. The example of the association of risk factors for ADHD and preschoolers' ADHD symptoms assessed in the Norwegian Mother and Child Cohort Study provides clear evidence for the presence of selection bias in a population based large cohort study. More generally, consideration of the likely causal relationships between mental health, alcohol use, smoking, and education suggests that analyses involving these variables require IPW to control bias. Importantly, assessment of bias for entire multi-exposure multi-outcome cohort studies is not possible. Instead, assessing the potential for bias and its correction needs to be done on the level of individual analyses.