On the Correction for Misclassification Bias in Drug Safety Data Using Validation Sample Approaches

02/23/2016 - 15:30
02/23/2016 - 16:30
Chris Gravel, McGill University
Purvis Hall, 1020 Pine Ave. West, Room 24

 Outcome misclassification in patient health records can bias estimation of adverse drug reaction risk. In this discussion, we will first consider a binary setting and demonstrate the use of internal validation sampling to offset misclassification bias in estimation of the odds-ratio. Investigation of the relative efficiency of odds-ratio estimators arising from the use of conditional versus random validation sampling will be investigated in simulation studies, focusing on differences in the selection of the categorical composition underlying the validation data. A Monte Carlo approximation to validation sample size determination will be recommended. To address the additional influence of confounding, we will introduce an inverse probability weighted approach to rebalance covariates across treatment groups while continuing to mitigate the impact of misclassification bias.

Next, for right censored continuous time survival data, failing to observe the event of interest can introduce misclassification bias in risk estimation. Incorrectly observing cause-specific event types at correctly recorded event times can also introduce bias. An internal validation sampling approach is used to update a set of parametric likelihoods to produce unbiased estimates in scenarios with the presence of either or both of these errors. These approaches are validated through large simulation studies.

Last edited by on Thu, 02/18/2016 - 14:40