Department Seminars

In Epidemiologic studies, measurement error in the exposure variable can have large effects on the power of hypothesis testing for detecting the impact of exposure in the development of a disease. As it distorts the structure of data, more uncertainty is associated with the inferential procedure involving such exposure variables. The underlying theme of this thesis is the adjustment for misclassification in the hypothesis testing procedure. We consider problems involving a correctly measured binary response and a misclassified binary exposure variable in a retrospective case-control scenario. We account for misclassification error via validation data under the assumption of non-differential misclassification. The objective here is to develop a test to check whether the exposure prevalence rates of cases and controls are same or not, under frequentist and Bayesian point of view. To evaluate the test developed under Bayesian approach, we compare that with an equivalent test developed under frequentist approaches. Both these approaches were developed under the two further assumptions: in presence or absence of validation data, and to evaluate whether there is any gain in hypothesis testing for having such validation data or not. The frequentist approach involves likelihood ratio test, while the Bayesian test is developed from posterior distribution generated by a mixed MCMC algorithm and a normal prior under realistic assumptions. The comparison between these two approaches is conducted using different simulated scenarios, as well as two real case-control studies having partial validation (internal) data. Different scenarios include the settings with varying sensitivity and specificity, sample sizes, exposure prevalence proportion of unvalidated and validated data and under fixed budgetary constraint. In the scenarios under consideration, we reach the same conclusion from the two hypothesis testing procedures.

In many research areas, measurement error frequently occurs when investigators are studying the association between exposure variables and response variables in observational study. Several results can be caused by mismeasured exposure variables, such as loss of power, biased estimators and mis-leading conclusions. The underlying theme of this thesis is to evaluate a proposed "formal" approach that adjusts the measurement error under the Bayesian analysis. The approach is applied when the response variable is precisely measured but the exposure variable is either misclassified or mis-measured under the non-differential assumption. Gibbs sampler and Metropolis - Hasting algorithms are used to generate the posterior distributions for unknown parameters in the model.  In both binary exposure and continuous exposure situations, three cases are studied to evaluate the performance of our formal approach as: when the measurement error is known (from previous study); when the measurement error is unknown but has some prior information available; when both prior information of the measurement error and some validation data are ready to use.  Meanwhile, our formal approach is also compared with informal or naive approaches by studying the sampling distributions of the log odds ratio (in the binary exposure case) and the estimated coefficients (in the continuous exposure case). Finally, the proposed formal approach is applied on a real world dataset and similar conclusions (as from the simulated datasets) are able to reach when prior information is properly adjusted.