It came up again today so a reminder:
No matter what you were taught,
don't bother to test for proportional hazards
in survival analyses.
Hazards are never proportional because hazard ratios vary over the follow-up if treatment has an effect on survival.
Bootstrap to get a valid 95% confidence interval for the average hazard ratio and avoid exclusive reliance on hazard ratios as effect measures.
Mats Stensrud and I explain it here 👇
https://jamanetwork.com/journals/jama/article-abstract/2763185
@MiguelHernan
Your view seems to be: proportional hazards never holds and, if it doesn't, the avg. HR is an avg. over the censoring distribution, so basically useless.
How does bootstrap help? Also, do you have a reference that shows it improves on model-based SE under non-PH? Thanks.
@Tim_P_Morris @MiguelHernan I think it is related to Lin & Wei 1989 (which I think you could also use their variance estimator instead of the bootstrap)
Lin DY & Wei LJ. (1989). The robust inference for the Cox proportional hazards model. JASA, 84(408), 1074-1078.
@PausalZ @MiguelHernan I’ll check it out, thanks Paul