Cox Regression tvc command interpretation

[Ben Morton]

We are looking at the "Sabadell score" as a predictor of long term outcome at the point of discharge from critical care in hospital.

This is an ordinal score of 0,1,2 or 3.

Most patients fall in categories 0 and 1, fewer in 2 and 3 as the "sickest" groups of patients.

I have constructed a Cox-Regression model and found that this variable violates PH assumptions (this makes sense as the effect of this variable is likely to diminish with time)

I've read with interest the following conversation on: https://protect-eu.mimecast.com/s/fB...=statalist.org

Based on this I've drafted the following code to explore the effect at 90 days (we have five year survival data):

stcox i.Sabadell, tvc(i.Sabadell) texp(exp(_t>90))

I wanted to check that my interpretation is correct:

So at 90 days the hazard ratio observed in Sabadell score 3 is reduced by a factor of 0.363 - this would mean 120.32*0.363 = 43.68 - is this right???

Would it be reasonable to re-run the analysis at one, two and three years to examine how the effect changes over time?

If so what would be the optimal way to describe/present this data?

Advice gratefully received

Ben

[Paul Dickman]

You can reproduce your results by using -stsplit- (before -stcox-) to split person-time. The advantage of that approach, as opposed to using tvc() and texp() is that you can very easily split follow-up into more than two categories so you get, for example, estimated hazard ratios for each and every year of follow-up (or whatever intervals you choose).

That gives you hazard ratios that are step functions (easy to interpret but not clinically plausible). One alternative is to fine split with -stsplit- and fit, for example, a natural spline. Alternatively, you could go back to using tvc() and texp() but choose a functional form other than a step function (e.g., texp(_t) or texp(ln(t))).

I'm a big fan of flexible parametric models (the user-written -stpm2- command) where it's very easy to model time-varying effects.