Tree models for assessing covariate-dependent method agreement with an application to physical activity measurements

Abstract

Method comparison studies assess agreement between different measurement methods. In the present work, we are interested in comparing physical activity measurements using two different accelerometers. However, a potential issue arises with the popular Bland–Altman analysis, as it assumes that differences between measurements are identically distributed across all observational units. In the case of the physical activity measurements, agreement might depend on sex, height, weight, or age of the person wearing the accelerometers, among others. To capture this potential dependency, we introduce the concept of conditional method agreement, which defines subgroups with heterogeneous agreement in dependence of covariates. We propose several tree-based models that can detect such a dependency and incorporate it into the model by splitting the data into subgroups, showing that the agreement of the activity measurements is conditional on the participant’s age. Simulation studies also showed that all models were able to detect subgroups with high accuracy as the sample size increased. We call the proposed modelling approach conditional method agreement trees and make them publicly available through the R package coat.