Abstract:
|
Correlated binary outcomes arise frequently in public health research. When an outcome of interest is rare, sampling the data via a case-control design is preferred for its higher power, but when there are multiple outcomes, it is more convenient to sample based on a composite outcome. Ignoring this sampling in the analysis produces biased estimates, which we correct with inverse probability weighting. An additional layer of complication arises when gold standard outcome information is not available due to practical difficulties in its ascertainment. Under such settings, the outcome status is often estimated based on predicted probabilities derived from fitting a risk prediction model in a validation set. Traditionally, such probabilities are thresholded to classify the true outcome status, resulting in potential misclassification. We discuss estimation and testing procedures to improve power by directly modeling the predicted probabilities, and show via simulations that the proposed methods perform well in finite samples. The methodology is illustrated on data from the Army STARRS study, where the outcomes are psychiatric disorders, and the predictors of interest are genetics.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.