Non-Collapsibility and Selection Bias of Hazard Ratio

• Activity Number: 67
• Type: Topic Contributed
• Date/Time: Sunday, August 9, 2015 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Menglan Pang* and Michal Abrahamowicz and Robert Platt
• Companies: and McGill University and McGill University
• Keywords: Non-collapsibility ; marginal hazard ratio ; selection bias ; inverse probability weighting

Abstract

Non-collapsibility of the hazard ratio (HR) can be induced by a nonlinear relationship between the covariate and partial likelihood estimation when omitting a covariate in a Cox proportional hazard model. Meanwhile, selection bias may also occur. We used a longitudinal simulation study to simultaneously investigate the impact of (1) the non-collapsibility effect and (2) the selection bias of the HR when omitting a time-fixed risk factor. The non-collapsibility effect and selection bias were evaluated in different simulated scenarios, assuming a variety of treatment effects and covariate effects. We also compared the performance of two inverse probability weighted Cox models that used (i) the pooled propensity score and (ii) the time-specific propensity score. Our results suggested that omitting an important risk factor resulted in substantial changes in the Cox model estimates. This change was predominantly explained by the non-collapsibility effect, whereas selection bias was barely present in our simulation settings. On the other hand, both inverse probability weighted models provided unbiased estimate of the true marginal HR with comparable precision in our simulation scenarios.