When assessing associations of exposures with rare diseases based on case control studies designed within well-defined cohorts, individuals diagnosed prior to cohort entry are typically excluded to avoid the potential impact of survival bias on study findings.
We developed methods that in addition to data on controls and incident cases allow one to include information from prevalent cases to improve efficiency of case-control studies. We construct a constrained empirical likelihood assuming an exponential tilting model that leads to logistic regression and obtain efficient estimates of association parameters. We adjust for survival bias by modeling the backward time for prevalent cases using a parametric survival distribution. We develop an empirical likelihood ratio test for the association parameters in the logistic or survival model.
We quantify the efficiency gain when incident cases are supplemented with prevalent cases in simulations, and illustrate our methods by estimating the association of single nucleotide polymorphisms (SNPs) with breast cancer risk based on data from the U.S. Radiologic Technologists Health Study, a prospective cohort of radiologic technologists.
|