Cause-specific cause of death

[Arnaud Lesegretain]

Hello,
I am working with a SEER data set for a hematological cancer; dataset also contains information about other causes of death (not the cancer diagnosis used to select my cases in SEER), such other cancer and non-cancer related death. For overall death, I am using the KM estimates and Cox PH regression methods to provide analysis about this outcome.
What is the most appropriate analytical approach to characterize outcomes for the other causes of death? Is it the Cumulating Incidence under a competing risk framework where all other causes of deaths are considered censored? Any other approach?
and what is the appropriate STATA command?
many thanks
Arnaud

[Clyde Schechter]

Yes, it would be most appropriate here to do a competing risks analysis where deaths from other causes compete with cancer deaths. (For that matter, if the focus is some particular type of cancer, death from any other type of cancer is also a competing risk.) Note that these deaths from competing risks are not censored events in this model. Unlike the simpler KM approach, in this model there are three ways the subject exits the study. 1. Failure from the event of primary interest (cancer death in your case), 2. failure from some other event that prevents the primary event from happening by preceding it (competing risks, such as death from other causes), and 3. censorship--which means that the subject is no longer being observed even though no failure event has yet occurred (such as end of study date or withdrawal).

The Stata command for estimating the cumulative incidence function is -stcompet-. It has a number of options as part of its syntax, so I recommend you read -help stcompet- for the details, and perhaps follow the link there to the PDF documentation as well. It's a fairly complicated command. And as you are not yet comfortable with the conceptual basis, I recommend you take a look at https://www.publichealth.columbia.ed...-risk-analysis, which has, in my opinion, an exceptionally clear presentation.

[Arnaud Lesegretain]

Many thanks Clyde. This is has been very useful and I have since read a lot and used the stcompet in my analysis. I was able to generate very nice ci curves which I think make sense overall in the context of the disease I am studying.
I would like to list the ci function to see individual values at various timepoints but cannot find how to do this ?

[Clyde Schechter]

OK, to do that, I'm afraid you'll have to use a different command, -stcrreg- to do the competing risks survival analysis. Then you can follow that up with -stcurve, cif outfile(filename, replace)-. This will graph the CIF but will also save the values used to plot the curve in filename as a Stata data file. Then you can -use- that file and displays those results are you see fit.

[Arnaud Lesegretain]

Hi Clyde, I am just trying to read the values of these 2 curves for example; is there not a more straight forward way?

[Clyde Schechter]

No, you don't have to read the values off the graphs. If you specified the -outfile()- option in -stcurve- Stata will have created a new data set that contains the values that are on the graphs. So -use- that file and -list- it or -browse- it to see the values. Here's an example using StataCorp's hypoxia.dta:

webuse hypoxia

stset dftime, failure(failtype = 1)
stcrreg ifp tumsize pelnode, compete(failtype == 2)

tempfile graph_results
stcurve, cif outfile(`graph_results')
use `graph_results', clear
browse

[Arnaud Lesegretain]

Many thanks Clyde. I will try this.
Maybe another more fundamental question: in my data set, i have ~ 8 different causes of death (COD) groupings (CVD, diabetes, renal, leukemia,....). Right now, I have structured specific COD codes:
For example: for CVD, cvd_cod=0 if alive, cvd_cod=1 if cvd is COD, cvd_cod=2 if COD is any other COD except CVD.
the stcompet command is :

stset survmonth, failure(cod_cvd==1)
stcompet cifcvd=ci, compet1(2)

Is this correct to group all other cause of death that are competing with CVD into a single code?

I have of course repeated this for each cause of death.

[Clyde Schechter]

Is this correct to group all other cause of death that are competing with CVD into a single code?

Yes.

[Andrea Discacciati]

Check out -stcomlist- from SSC.

 stcomlist is a wrapper for the SSC package stcompet to calculate
      and list  cumulative incidence functions and confidence intervals
      in the presence of competing risks.

[Arnaud Lesegretain]

thank you Andrea and Clyde

[Allen Buxton]

Check out -stcomlist- from SSC.

 stcomlist is a wrapper for the SSC package stcompet to calculate
and list cumulative incidence functions and confidence intervals
in the presence of competing risks.

As an option, I have the following that has allowed me the use of R package cmprsk to do Gray's test and get estimates and confidence intervals with a form to plot these estimates from Stata Re: (Gray R. J. (1988), A class of K-sample tests for comparing the cumulative incidence of a competing risk, Ann. Stat. 16:1141-1154). Stata does the Fine and Gray regression with stcrreg. https://www.statalist.org/forums/for...Here%20is%20an
stgrays.ado