One further clarification; on the people that died of types 1-5, you would include the records that you have while they were still alive, but not the record indicating that they had died, and obviously, since you have no more data on them, no records after the point which they had died.

-Mary ----- Original Message ----- From: Mary To: SAS-L@LISTSERV.UGA.EDU ; coreklenelim@YAHOO.COM Sent: Tuesday, September 23, 2008 10:25 AM Subject: Re: Re: Competing Risks for Dummies

Darren,

I'm not an expert, but I did take the Survival Analysis using the Proportional Hazards Model course from SAS Institute. I would highly recommend taking the course; there is a 50% academic discount and it is offered via Live Web.

In the course, we learned about censored events, and I am wondering if since you are interested in modeling the probability of the 6th type of event, whether a death of event types 1-5 might be considered a "censored" event, sort of like someone dropping out of a clinical trial. Then you could get your data to be something like this:

days_since_study_start status 0=alive, 1=died of type 6 (you would drop records of anybody who died of types 1-5) independent_var1 independent_var2 and so forth, for all of your independent variables.

Then in the course we often used Proc PHREG, like this:

proc phreg data=set1 simple; model days_since_study_start*status(0)= independent_var1 independent_var2/ties=exact r1; run;

Hope that helps; but would definately recommend consulting with your statisticians and also taking the SAS course!

-Mary ----- Original Message ----- From: Shawn Haskell To: SAS-L@LISTSERV.UGA.EDU Sent: Tuesday, September 23, 2008 8:50 AM Subject: Re: Competing Risks for Dummies

On Sep 22, 12:13 pm, coreklene...@YAHOO.COM (Darren Oz) wrote: > Hello, > > A desperate newbie will greatly appreciate your help. > > First, let me try to identify my problem. I have a disease dataset where > the response variable has 7 levels: 0 means "still alive" and 1-6 stands > for 6 causes of death. A person might appear multiple times in the dataset > and the response variable will be a 0 if the end-of-month status of a > person is alive. The month the person dies is when the response variable > will be something between 1 and 6 I won't be seeing that person anymore. > Now, I want to focus on the probability of one cause of death, enumerated > as 6, for all the people in my holdout period. In other words, I want to > model the deaths using my historical data and "score" it on my holdout to > calculate the average deaths of one particular cause. > > My Google and SAS-L search gave me the idea of using Competing Risks > survival analysis for this problem. I have the "Survival Analysis Using > SAS: A Practical Guide" book, however, I am not a stats person and it's > impossible for me to do all the stuff myself. Could someone shed a light > on this subject? > > Darren

Darren, it sounds like your project may be important enough to seek professional biometrics help. FYI, the Kleinbaum and Klein text "Surival Analysis: a self-learning text" is a good introduction to survival analyses with a chapter on competing risks and some SAS code. i used the information in this text to run a Lunn-McNeil competing risks analysis for some deer fawn survival data (i.e., sickness-starvation vs. predation). I recall that it took considerable care to set up the input Excel file with just 2 competing risks; 6 competing risks would take greater care but should be doable. Cox proportional hazards (PH) is the standard (PROC PHREG or PROC TPHREG) when the question regards hazard risk of given covariates - not timing of the event (PROC LIFEREG used for questions of timing). Monthly analyses may not be necessary or appropriate. For Cox PH you need to test the time-independence PH assupmtion for covariates and extend the model if necessary - there are different ways to do this. In any style of competing risks analyses, the standard survival analysis assumption of noninformative censorships also means that competing risks are independent - for my analysis it meant that risk of predation was not influenced by risk of sickness- starvation - this may or may not not be true and requires careful thought and interpretation. Competing risks are tricky - good luck. SH