I am doing time to event analysis, measuring the time from first-ever stroke to first event of fracture. I am furthermore using cox-regression to analyze the associations between stroke severity and fractures, and civil status and fractures (adjusted for sex, age, AMI, diabetes and alcohol intake). I am doing a register-based retrospective cohort with a follow-up time of nearly 15 years (from 2003 to late 2017), including 106.082 patients.

I have tested the PH-assumption, using log-log plots and the ph-assumption statistical test. It is a bit more complicated because i have catergorical variables; stroke_severity (Very severe/severe/moderate/mild/Unknown) and civil_status (living with someome/living alone/other/unknown).
As you can see below, the hazards seem to overlap, so the PH-assumption does not seem to be met.

I have already recieved advice in this regard, but i would like to be absolutely certain about what the best approach would be. I would love some input, both on wether or not the ph-assumption is met, and on what the next optimal approach would be i.e. cox with time varying covariates, categorization of the follow-up period, stratified cox or focusing on other variables/associations.

Tabel ?. PH-assumption test for the association between stroke severity og fractures.
Stroke severity Rho Chi2 df Prob> Chi2
Very severe
Severe -0,00805 1,03 1 0,3107
Moderate 0,00698 0,77 1 0,3794
Mild 0,05719 51,78 1 0,0000
Unknown 0,01655 4,34 1 0,0372
Global test 285,34 4 0,0000








Tabel ?. PH-assumption test for the association between civil status og fractures.
Civil status Rho Chi2 df Prob> Chi2
Living with someone 1
Living alone -0,04920 38,09 1 0,0000
Other -0,06241 61,36 1 0,0000
Unknwon 0,02238 7,92 1 0,0049
Global test 88,18 3 0,0000







Log-log plots. Stroke severity Array



Log-log plots. Civil status Array