Proc PHREG: final model with different variables when different model selection methods were used

[xinjian]

Dear All,

I used PROC PHREG to do Cox proportional hazards regression. I got 5 variables in the model with back ward elimination. I got 3 variables of the model when I use forward or stepwise model selection methods. Is it possible with different variables in the final model because using different model selection methods? Which method I should use?

Thanks,

Xinjian

[Reeza]

xinjian wrote:
Is it possible with different variables in the final model because using different model selection methods? Which method I should use?

Yes, its possible.

There is no one method that should be used, in fact using the automated selection methods without applying business logic isn't a good method in general.

[xinjian]

Hi Reeza,

Thanks for your quick response. I used univariate cox regression (proc phreg) to select variables (if p<0.25). I got 9 variables. Then I run proc phreg with different model selection methods to find final model if p<0.05. So what is the best way to select final COX model?

Thanks,

Xinjian