Out of Order: Specification Check Sequencing in Cox Models

• In Event: Models for TSCS, Time, and Duration Data

Abstract

The Cox duration model serves as the basis for more complex duration models like competing risks, repeated events, and multistate models. These models are characterized by multiple transitions that subjects can experience as they move through a process. The models also make a number of assumptions, many of which can be assessed empirically, often for substantive ends. However, we show that the order in which practitioners assess these assumptions can impact the model’s final specification, and ultimately, may produce misleading inferences if the corrections for violations of these assumptions are applied out of order. We focus on three assumptions that are simultaneously made for complex duration models: proportional hazards, stratified baseline hazards, and transition-specific covariate effects. We use Monte Carlo simulations to investigate the order in which practitioners should check these assumptions to obtain the most accurate inferences as often as possible. Our first round of simulations suggests that the final model specifications a practitioner obtains may differ starkly, depending on the order in which they test and correct for any assumption violations.

Authors

• Benjamin Thomas Jones, The University of Mississippi

• Shawna K. Metzger, Michigan State University