John M. Neuhaus

An assessment of bias in longitudinal studies with outcome-dependent visits

The timing and frequency of the measurement of longitudinal outcomes in electronic health records may be associated with the value of the outcome. Such visit times are called outcome-dependent and previous work has indicated that ignoring outcome-dependent visit times can produce biased estimates of the associations of covariates with outcomes. In this talk we present theory and use simulation studies to further assess estimation bias in a range of settings that vary from those where all the visits are scheduled and regularly spaced in time to those where where all the visits are unscheduled, irregularly spaced in time and associated with the value of the outcome.  Our results show that while ignoring outcome-dependent visit times can yield biased estimates of some covariate effects, other covariate effects can be estimated with little bias.  We also show that the presence of a small number of regularly scheduled visits in the data set protects mixed model analyses from bias.