Hazards, Causality, and Practical Relevance of Collider Effects – Comment on Beyersmann et al. “Hazards Constitute Key Quantities for Analyzing, Interpreting and Understanding Time-to-Event Data”

Abstract

Hazards constitute key quantities for analyzing, interpreting, and understanding time-to-event data. Hazards and corresponding effect measures, such as the hazard ratio from the Cox proportional hazards model, have a valid causal interpretation if the hazard function is considered as a function in time rather than hazards at specific time points. In this comment, we would like to add two points: (1) The hazard ratio is also a useful population-level estimand with a valid causal interpretation. (2) Empirical evidence shows that problematic situations, which could occur in theory due to strong heterogeneity, are usually avoided in typical randomized controlled trials.