# SAS Code for Survival Analysis

In PROC PHREG, how do you set a continuous variable at a certain value as the reference level? For example, suppose x = 3.5, 3.6, 4.3, 5.4 and we want x = 6 to be the reference level. How would you do this in SAS?

Would this be correct:

      data new;
set old;
x1 = (x=3.5);
x2 = (x= 3.6);
x3 = (x = 4.3);
x4 = (x= 5.4);

proc phreg data = new;
class x ( ref = '6');
model time*censor(0) = x1 x2 x3 x4 /r1;
run;


Edit. Actually, could one just use the point estimate of the coefficient of $x$? In other words suppose the estimated coefficient for $x$ is $0.512$. Then the hazard ratio between $x = 3.5$ and $x = 6$ would be $\exp(0.512(3.5-6))$, the hazards ratio between $x = 3.6$ and $x = 6$ would be $\exp(0.512(3.6-6))$ etc...?

So there is no need to even form these groups? Just do the following:

    proc phreg data = new;
model time*censor(0) = x /r1;
run;

• Doesn't SO have a SAS section? I thought SE was for theory. – DWin Dec 11 '11 at 18:34

If X is a continuous variable, there is no reference level; or, I guess, in a way, the reference level is 0. Then you could, sort of, change the reference level to 6 by forming a new variable x2 = x - 6. Then you could use your second code.

If you want to treat x as a class variable, then (unfortunately) there's no easy way to set the reference level in SAS. One thing you can do is either format the variable so thath that the reference level comes last, or create a new variable something like this:

data new;
set old;
if x = 3.5 then x2 = 'A:3.5';
else if x = 3.6 then x2 = 'B:3.6';

....
run;


then

   proc phreg data = new;
class x2
model time*censor(0) = x2 /r1;
run;


That loses the continuity and ordinality of x, but that may be desirable (to capture nonlinearity and because your x are not very evenly spaced). However, I suggest exploring polynomial terms and/or splines.

you interpret the HR in terms of a unit change in x eg if your covariate is weight then it is 1kg. You could even then plot the HR against the covariate ie use "ods output parameterestimates=...." and then the estimate from this ods: val=exp(estimate * x), and then "plot val * x" to see the HR as a function of the change in the variable. Incidentally, i'd recommend david collett's book on survival data