Extreme-value sampling design is cost-beneficial only with a valid statistical approach for exposure–secondary outcome association analyses

Abstract

In epidemiology cohort studies, exposure data are collected in sub-studies based on a primary outcome (PO) of interest, as with the extreme-value sampling design (EVSD), to investigate their correlation. Secondary outcomes (SOs) data are also readily available, enabling researchers to assess the correlations between the exposure and the SOs. However, when the EVSD is used, the data for SOs are not representative samples of a general population; thus, many commonly used statistical methods, such as the generalized linear model (GLM), are not valid. A prospective likelihood method has been developed to associate SOs with single-nucleotide polymorphisms under an extreme phenotype sequencing design. In this paper, we describe the application of the prospective likelihood method (STEVSD) to exposure–SO association analysis under an EVSD. We undertook extensive simulations to assess the performance of the STEVSD method in associating binary and continuous exposures with SOs, comparing it to the simple GLM method that ignores the EVSD. To demonstrate the cost-benefit of the STEVSD method, we also mimicked the design of two new retrospective studies, as would be done in actual practice, based on the PO of interest, which was the same as the SO in the EVSD study. We then analyzed these data by using the GLM method and compared its power to that of the STEVSD method. We demonstrated the usefulness of the STEVSD method by applying it to a benign ethnic neutropenia dataset. Our results indicate that the STEVSD method can control type I error well, whereas the GLM method cannot do so owing to its ignorance of EVSD, and that the STEVSD method is cost-effective because it has statistical power similar to that of two new retrospective studies that require collecting new exposure data for selected individuals.