Proc Phreg Hazardratio statement

[Wafflecakes]

Howdy yall,

Does anybody know alternatives to the hazardratio statement that allow the statistician to specify different units for continuous variables?  This statement does not offer p-values.

[OsoGris]

You can use the  E option of the HAZARDRATIO statement to display the coefficient vector of the linear combination that gives rise to the hazard ratio.  Note that the E goes after a forward slash.  Suppose you desire a hazard ratio for a covariate named 'treatement'; for example,

HAZARDRATIO treatment / DIFF=REF E;

Then you can use those coefficients in a CONTRAST statement with the ESTMATE=EXP option which should replicate the hazard ratio value and also give a p-value.  The ESTIMATE=EXP also goes after a forward slash.  For example,

CONTRAST 'label' treatment . . . list of coefficients . . .  / ESTIMATE=EXP;

For a good example with categorical variables see Example 86.3 Modeling with Categorical Predictors which is informative.  In particular, the 3rd PHREG step there has HAZARDRATIO statements 'H1', 'H2' and 'H3' which correspond with CONTRAST statements 'C1', 'C2' and 'C3', respectively (and which naturally give the same results except that the CONTRAST statement also gives a p-value).    The entire example is informative but the material including and above Output 86.3.6 & Output 86.3.7 illustrates these corresponding HAZARDRATIO and CONTRAST statement results.

[Wafflecakes]

How would I figure out the coefficients for two-way interactions between categorical variables? I dont know how to come up with the design matrix.

[OsoGris]

Norte: the above answer was geared towards a categorical variable.  I haven't tried the / E option in conjunction with the UNITS= option on the HAZARDRATIO statement with a continuous covariate.  If it doesn't work seamlessly you can always do the HAZARDRATIO/CONTRAST statement construct for UNITS=1 and then just manually exponentiate the resulting 1-unit hazard ratio (say HR_1) for UNITS=k as HR_k = (HR_1)**k.

[OsoGris]

When you have an interaction term you cannot interpret the main effects in isolation.  You must involve the levels/value of the interacting variable(s).  So, regarding HAZARDRATIOS for interaction models you can still use the E option and the CONTRAST statement construct.  However, you will likely also need the AT= option to compute the hazard ratio  for a one variable at a particular level of the interacting covariate.   For instance, if continuous variables x1 and x2 appear in an interaction term then you can do something like:

HAZARDRATIO x1 / AT(x2=3) E;

Then you can take the coefficients to a CONTRAST statement with ESTIMATE=EXP as explained above.

[OsoGris]

Sorry, again my reply was geared towards continuous variables which was your initial question.  Regarding interactions of categorical variables say coviartes c1 and c2 you can do something like

HAZARDRATIO c1 /  AT(c2=ALL) E;

Then, again, take the coefficients to a CONTRAST statement.  See the Syntax => HAZARDRATIO Statement section of the PHREG documentation for more details on the HAZARDRATIO Statement

[Wafflecakes]

The E option does not work I don't think when used in conjunction with the 'UNITS' option. how does this:

"If it doesn't work seamlessly you can always do the HAZARDRATIO/CONTRAST statement construct for UNITS=1 and then just manually exponentiate the resulting 1-unit hazard ratio (say HR_1) for UNITS=k as HR_k = (HR_1)**k."

Get you the p-value?

[OsoGris]

I think the best thing to do at this point is to open up a ticket (Tracking Number) with SAS Technical Support.  You can do so by emailing to support@sas.com and make sure you reference PHREG in the subject.   However, the most efficient way to open a ticket is to use our online Submit A Problem form and select SAS/STAT from the  "Product" list (you'll have to scroll down in this list pretty far).   In so doing it is a good idea to provide your site number, SAS release and OS.   Please run the following SAS statement and include the results of which appear in the SAS log: 　　　　　　

　

　 　　　　　　　%put Site: &syssite 　　　Release: 　&sysver 　(&sysvlong) 　　　System: &sysscp (&sysscpl);

　

[Wafflecakes]

I do get the coefficients in these tables at the bottom but it is equal to the units that I set for the continuous variable and the other variables are listed as well with coefficients that equal 0.  Is that correct?  I don't think so?