Model testing for proportionality - competing risk regression (stcrreg)

[LarsFolkestad]

Dear listers

I am running an competing risk regression and want to test my model to see if the proportionality assumptions holds and the goodness-of-fit.

i asked a statistician for help and he sent me good advise saying:
The competing risk models can be tested the same way as the other proportional hazards models. The only difference is that instead of the all-cause hazard, you would be interested in the cause-specific hazards (event of interest, competing event 1, competing event 2,…). You can check the proportionality assumption by making a log-log plot (log-cumulative hazard against log of time) as in the stphplot (which is only defined for cox regression as stata default).

he gave the following code:

streg gender2, dist(gompertz)
predict s, surv
gen ls0 = log(-log(s)) if gender2==0
gen ls1 = log(-log(s)) if gender2==1
gen lt0 = log(time) if gender2==0
gen lt1 = log(time) if gender2==1
twoway lowess ls0 lt0 || lowess ls1 lt1

but seeng that i use stcrreg and then when i try to

predict s, surv

i get an error saying survival option not allowed.

My question: How do i model test for the proportionality assumptions using stcrreg in Stata?

lars

[LarsFolkestad]

reading up on the stcrreg in the book by Cleves et al "An introduction to survival analysis Using Stata".
They say to test for the proportionality assumptions by using the tvc and texp options - arguing that if the proportional-subhazardsassumptions holds than coefficients should be constant with time and time interactions should not be statistically significant.

Thats what i'll do than.

lastly they use

stcurve, cif at1(drug = 0) at2(group=1)

to show the cumulative incidence of the event.

My problem - the cause of the use of stcrreg - is that i am looking at the incidence of a non-leathal event in a group of patients (with increased all-cause-mortality) compared to a reference population.

comments on that?

Lars