Modeling Secondary Phenotypes Conditional on Genotypes in Case–Control Studies

Naomi C Brownstein,Shad Smith,Jianwen Cai,Eric Bair,Luda Diatchenko,Gary D Slade

doi:10.3390/stats5010014

Abstract

Traditional case–control genetic association studies examine relationships between case–control status and one or more covariates. It is becoming increasingly common to study secondary phenotypes and their association with the original covariates. The Orofacial Pain: Prospective Evaluation and Risk Assessment (OPPERA) project, a study of temporomandibular disorders (TMD), motivates this work. Numerous measures of interest are collected at enrollment, such as the number of comorbid pain conditions from which a participant suffers. Examining the potential genetic basis of these measures is of secondary interest. Assessing these associations is statistically challenging, as participants do not form a random sample from the population of interest. Standard methods may be biased and lack coverage and power. We propose a general method for the analysis of arbitrary phenotypes utilizing inverse probability weighting and bootstrapping for standard error estimation. The method may be applied to the complicated association tests used in next-generation sequencing studies, such as analyses of haplotypes with ambiguous phase. Simulation studies show that our method performs as well as competing methods when they are applicable and yield promising results for outcome types, such as time-to-event, to which other methods may not apply. The method is applied to the OPPERA baseline case–control genetic study.

Highlights

In addition to the cohort of initially temporomandibular disorders (TMD)-free adults enrolled in the prospective cohort study, people with examiner-verified chronic TMD were enrolled to create an unmatched case–control study
Study, Zi denotes the number of copies of the minor allele, Di is an indicator of whether participant i is a chronic case of TMD, and Yi is the ordinal number (0, 1, 2+) of comorbid pain conditions for participant i. (In general, Yi can take other forms, as described below.) If one were to ignore the case–control study design and consider the data as a random sample from the population, one would use standard methodology to study the relationship between Y = (Y1, . . . , Yn+m )0 and Z = ( Z1, . . . , Zn+m
The simulation results indicate that it is approximately unbiased, and has comparable coverage and confidence interval width to the method of inverse probability weighting (IPW) with generalized estimating equations (GEE) [2]

Summary

Introduction

Prospective studies are more straightforward and less prone to confounding than other study designs. They may require either extremely long follow-up periods or large sample sizes, and lack power. For rare diseases in particular, the sample sizes required in a prospective cohort study to have adequate statistical power to test hypotheses of interest may be prohibitively large. This can be especially problematic in genetic association studies, which may cost thousands of dollars per participant just to extract their genetic profiles. Retrospective case–control studies are more cost effective. The number of case–control studies focusing on the relationship between genetics and disease outcomes has grown astronomically in recent years

Methods

Results

Conclusion