A method making fewer assumptions gave the most reliable estimates of exposure–outcome associations in stratified case–cohort studies

Abstract

ObjectiveA case–cohort study is an efficient epidemiological study design for estimating exposure–outcome associations. When sampling of the subcohort is stratified, several methods of analysis are possible, but it is unclear how they compare. Our objective was to compare five analysis methods using Cox regression for this type of data, ranging from a crude model that ignores the stratification to a flexible one that allows nonproportional hazards and varying covariate effects across the strata. Study Design and SettingWe applied the five methods to estimate the association between physical activity and incident type 2 diabetes using data from a stratified case–cohort study and also used artificial data sets to exemplify circumstances in which they can give different results. ResultsIn the diabetes study, all methods except the method that ignores the stratification gave similar results for the hazard ratio associated with physical activity. In the artificial data sets, the more flexible methods were shown to be necessary when certain assumptions of the simpler models failed. The most flexible method gave reliable results for all the artificial data sets. ConclusionThe most flexible method is computationally straightforward, and appropriate whether or not key assumptions made by the simpler models are valid.