I need help to figure out how to implement a marginal effects model in sas PHREG

[marcel]

Hello SAS Experts,

I am running a cox proportional hazards model. I am running PHREG because I want to test if there is a significant interaction between TTO and a continuous covariate.

My TTO has two levels (0 and 1)
Let the continuous covariate be "B"

I run the following:

proc phreg data = data;
class TTO;
model survivaltime*censored(0) = TTO B TTO*B  / ties=Efron;
id id;
run;

As can be seen in the table above, the interaction TTO*B is not significant. Of course, some people are of the opinion that I should just drop the interaction because it is not significant and be done with it.

However, I read a recommendation that when the interaction is not significant, then "interpret your model using marginal effects in the same way as if the interaction were significant".   How do you this? Or better yet, how do you implement a marginal effects model in sas PHREG? Can anyone direct me to a good source with examples?

Best regards,
Marcel

[SteveDenham · Accepted solution]

Hi @marcel ,

I would interpret that article as saying - look at the main effects least squares means (if both of the variables in the interaction are categorical).

The simple way of getting marginal means is to drop the interaction.

However, what you have here is a bit trickier, as B is a continuous covariate.  The test of the interaction is whether the beta's for the two levels of TTO are identical.  Since you have no evidence that they are different, the parsimonious model would drop out the interaction.  If you keep the interaction (because a: it is significant or b; you have additional knowledge that the slopes are not the same from prior work), then when you compare lsmeans, you should do it at 3 levels of B, a low level, the mean level and a high level.  See Milliken and Johnson's Analysis of Messy Data, vol. 3: Analysis of Covariance for examples.

SteveDenham

[marcel]

Thank you, Steve. That is a point of start for me. I will check the reference you recommend.

Regards,

Marcel