Methods/Statistics
Bias analysis adjustment for time-to-event outcomes Richard MacLehose* Richard MacLehose MacLehose MacLehose MacLehose Tulane University
Introduction
Misclassification is a threat to the validity of epidemiologic studies. Bayesian, frequentist, and probabilistic bias analysis (PBA) methods have been developed to adjust for misclassification, but little methodological development has focused on outcome misclassification in a time-to-event setting. We present methods to adjust for outcome misclassification in discrete-time follow-up analyses using both Bayesian and PBA approaches. We evaluate these approaches in simulations and apply them in a longitudinal cohort study of risk factors for time to dementia onset.
Methods
We assume that the true outcome status is unobserved at each discrete time point and once it occurs, it never resolves (an absorbing state). Outcome risk varies over time intervals and is a function of observed risk factors. The true disease state is latent and the observed outcome is misclassified. The influence of sensitivity and specificity is specified through prior distributions. We fit an absorbing-state hidden Markov model using Bayesian methods and approximated via a PBA algorithm. The methods are compared in a simulation study and illustrated in a longitudinal cohort study of dementia risk, following participants from adulthood into late life.
Results
Simulations showed both methods remove nearly all bias due to misclassification when the bias parameters were correctly specified (bias <0.01 and greater than 99% coverage). The Bayesian approach required more technical expertise and took 5 times the computational time. When these methods were applied to an existing cohort study, risk of dementia increased relative to methods that did not adjust for misclassification. Interval estimates were wider for the bias-adjusted risks because they incorporated uncertainty in the bias parameters.
Discussion
We presented two novel methods to adjust for outcome misclassification in time-to-event studies. Both methods had little bias and good coverage, with PBA maintaining an advantage in speed.