Irregular measurement times in estimating time-varying treatment effects: Categorizing biases and comparing adjustment methods

Abstract

To estimate the causal effect of treatments that vary over time from observational data, one must adjust for time-varying confounding. A common procedure to address confounding is the use of inverse probability of treatment weighting methods. However, the timing of covariate measurements is often irregular, which may introduce additional confounding bias as well as selection bias into the causal effect estimate. Two reweighting methods have been proposed to adjust for these biases: time-as-confounder and reweighting by measurement time. However, it is currently not well understood in which situations these irregularly timed measurements induce bias, and how the available reweighting methods compare to each other in different situations. In this work, we provide a complete inventarization of all possible backdoor paths through which bias is induced. Based on these paths, we distinguish three categories of confounding bias by measurement time: direct confounding (DC), confounding through measured variables (CMV), and confounding through unmeasured variables (CUV). These categories differ in the assumptions and reweighting methods necessary to adjust for bias and may occur simultaneously with selection bias. Through simulation studies, we illustrate: 1. Reweighting by measurement time may be used to adjust for selection bias and confounding through measured variables; 2. Time-as-confounder may be used to adjust for all categories of confounding bias, but not selection bias; 3. In some cases, the use of a combination of both techniques may be used to adjust for both confounding and selection bias. We finally apply the categorization and reweighting methods on the pre-DIVA data set. Adjusting for measurement times is crucial in order to avoid bias, and the categorization of biases and techniques that we introduce may help researchers to choose the appropriate analysis.