Cox proportional hazards models have more statistical power than logistic regression models in cross-sectional genetic association studies

Abstract

Cross-sectional genetic association studies can be analyzed using Cox proportional hazards models with age as time scale, if age at onset of disease is known for the cases and age at data collection is known for the controls. We assessed to what degree and under what conditions Cox proportional hazards models have more statistical power than logistic regression models in cross-sectional genetic association analyses. Analyses were conducted in an empirical study on the association of 65 polymorphisms and risk of coronary heart disease among 2400 familial hypercholesterolemia patients, and in a simulation study that considered various combinations of sample size, genotype frequency, and strength of association between the genotype and coronary heart disease. We applied Cox proportional hazards models and logistic regression models, and compared effect estimates (hazard ratios and odds ratios) and statistical power. In the empirical study, Cox proportional hazards models generally showed lower P-values for polymorphisms than logistic regression models. In the simulation study, Cox proportional hazards models had higher statistical power in all scenarios. Absolute differences in power did depend on the effect estimate, genotype frequency and sample size, and were most prominent for genotypes with minor effects. For example, when the genotype frequency was 30% in a sample with size n=2000 individuals, the absolute differences were the largest for effect estimates between 1.1 and 1.5. In conclusion, Cox proportional hazards models can increase statistical power in cross-sectional genetic association studies, especially in the range of effect estimates that are expected for genetic associations in common diseases.