Including an infrequently measured time-varying error-prone covariate in survival analyses: a simulation-based comparison of methods.

Abstract

Epidemiologic studies often evaluate the association between an exposure and an event risk. When time-varying, exposures are usually updated at discrete visits although changes are in continuous time and survival models require values to be constantly known. Moreover, exposures are likely measured with error, and their observation truncated at the event time. We aimed to quantify in a Cox regression the bias in the association resulting from intermittent measurements of an error-prone exposure. Using simulations under various scenarios, we compared five methods: last observation carried-forward (LOCF), classical two-stage regression calibration using measurements up to the event (RC) or also after (PE-RC), multiple imputation (MI) and joint modeling of the exposure and the event (JM). The LOCF, and to a lesser extent the classical RC, showed substantial bias in almost all 45 scenarios. The RC bias was avoided when considering post-event information. The MI performed relatively well, as did the JM. Illustrations exploring the association of Body Mass Index and Executive Functioning with dementia risk showed consistent conclusions. Accounting for measurement error and discrete updates is critical when studying time-varying exposures. MI and JM techniques may be applied in this context, while classical RC should be avoided due to the informative truncation.