# Survival Analysis

#### CharmSchool

##### New Member
Has anyone ever done a survival analysis (in medical research) where you first include the treatment variable as the only covariate and then do a sort-of manual forward stepwise selection process, adding one covariate at a time and
keeping only those that change the parameter estimate of the treatment
variable by 20%?

If so, could you please let me know if I am going about this correctly? In short, the analyses has proceed as follows:

Model 1:
Status*Time = TxGroup
(TxGroup has a significant p-value)

Model 2:
Status*Time = TxGroup Covar3
(the addition of Covar3 changes the TxGroup beta by more than 20% but now neither is significant)

Model 3:
Status*Time = TxGroup Covar8
(the addition of Covar8 changes the TxGroup beta by more than 20% but now neither is significant)

...let's say that of the 10 covariates tested as above, Covar3 and Covar8 are the only covariates that change the TxGroup parameter estimate (beta) by more than 20%. Then the "full" model (I think) would be:

Full Model:
Status*Time = TxGroup Covar3 Covar8
(again, none of the p-values are significant)

Is this the correct way to perform this type of analysis? Which model should be used for interpretation?

Any guidance would be very much appreciated!

C

#### vinux

##### Dark Knight
Which s/w you are using?
which model you are using? i mean cox regression or survival regression?
finally did you check the multicollinearity?

#### CharmSchool

##### New Member
Hi Vinux,

Thanks for helping me out with this! I am using SAS and running a Cox regression (proc phreg). I checked for multicollinearity and everything checked out as fine. So, based on my above post (in blue), my code essentially looks like (in red):

Model 1:
Status*Time = TxGroup
(TxGroup has a significant p-value)
proc phreg data=x;
model Time*Status(0) = TxGroup / ties=exact rl;
run;

Model 2:
Status*Time = TxGroup Covar3
(the addition of Covar3 changes the TxGroup beta by more than 20% but now neither is significant)
proc phreg data=x;
model Time*Status(0) = TxGroup Covar3/ ties=exact rl;
run;

Model 3:
Status*Time = TxGroup Covar8
(the addition of Covar8 changes the TxGroup beta by more than 20% but now neither is significant)
proc phreg data=x;
model Time*Status(0) = TxGroup Covar8/ ties=exact rl;
run;

...let's say that I did this for 10 covariates, and Covar3 and Covar8 are the only covariates that change the TxGroup parameter estimate (beta) by more than 20%. Then the "full" model (I think) would be:

Full Model:
Status*Time = TxGroup Covar3 Covar8
(again, none of the p-values are significant)

proc phreg data=x;
model Time*Status(0) = TxGroup Covar3 Covar8/ ties=exact rl;
run;

Is this the correct way to perform this type of analysis? Which model should be used for interpretation?

Any guidance would be very much appreciated!

C

#### vinux

##### Dark Knight
You could have directly used stepwise ( in PHREG).

Following cases, we may not be able to come up with statistically significant model

1. Small sample
2. Small % non censored population ( Status =1(i guess), having complete data).