Adjusting for non-confounding covariates in case-control association studies

Considerable debate has been generated in recent literature on whether non-confounding covariates should be adjusted for in the analysis of case-control data through logistic regression, and limited theoretical results are available regarding this problem. Zhang et al. (2018) proposed a constrained maximum likelihood approach that is seemingly more powerful than the approaches with or without adjusting for non-confounding covariates in logistic regression, but no theoretical justification was provided regarding this empirical finding. We provide rigorous justification for the relative performances of the above three approaches through Pitman's asymptotic relative efficiencies. Specifically, the constrained maximum likelihood approach is proved to be uniformly most powerful. On the other hand, the relative performance of the other two approaches heavily depends on disease prevalence, that is, adjust for non-confounding covariates can lead to power loss when the disease prevalence is low, but this is not the case otherwise.