Survival Curve following Propensity Score Kernel Matching

[Christian Connell]

I am trying to obtain accurate survival and/or smoothed hazard function graphs following analysis of a cox regression model run using a propensity score matched sample (with kernel matching using psmatch2). The psmatch2 process created a p-weight variable, which I used in the stcox model (see below) and then also tried to obtain survival curves comparing treatment and control conditions. The resulting curves did not match the survival output in terms of 25th and 50th percentile survival times, which I believe may be due to the kernel-weighted propensity matching. I see that kernel-smoothed cumulative hazard curves can be generated, but am concerned that is a separate issue from incorporating the p-weights. Is there a way to generate either survival curves or smoothed hazard curves that incorporate the kernel-weighted propensity score results?

*Survival.
stset timetormvl18 [pweight = _weight], failure(flrmvl==1)
stsum, by(Completers2)
stcox Completers2, vce(cluster Provider2)
stcurve, survival at1(Completers2=0) at2(Completers2=1)
stcurve, kernel(epan) hazard at1(Completers2=0) at2(Completers2=1)

Thank you for any assistance you can provide,
Christian Connell
Yale University School of Medicine

[Steve Samuels]

Welcome to Statalist, Connell!
Please look FAQ 12, which asks that you show us all commands and output between CODE delimiters. In this case, reporting the results of stsum and the percentiles you see on the curves would help (but see code below). I am guessing that you are looking at the survival curve from stcurve and comparing percentiles to those reported by stsum. There is no reason to expect them to be the same. The curves from stcox are based on a model that forces the curves to have proportional hazards, whereas stsum shows percentiles from Kaplan-Meier curves that have no such assumption. That they are different (how different?) suggests to me that the hazards for the two curves are not proportional. To investigate non-proportionality, see the Manual entry for "stcox PH assumption tests — Tests of proportional-hazards assumption".

You can get a closer look at the estimated percentiles for the survival curve from stcurve with the following code:

tempfile t1
stcurve, survival at1(Completers2=0) at2(Completers2=1), outfile(`t1', replace)
use `t1', clear
list _t surv1 if abs(surv1-.5)<.01  
list _t surv2 if abs(surv2-.5)<.01