Using lsmeans in phreg

[SASUser64]

I have a competing risks model using phreg and am trying to find the group means through lsmeans. I'm not sure I understand the results; the estimates are incredibly small (or negative if I include control variables in the model) compared to the actual group means when I run the same model through proc glimmix as a linear regression model. The outcome is hospital length of stay. Here is some of the code:

proc phreg data=cif_icp_crani;
class group / param=glm order=internal ref=last;
model los_adj*outcome(0) = group / rl eventcode=1;
format group groupf.;
hazardratio group / diff=pairwise;;
lsmeans group / means exp cl;
run;

proc glimmix data=cif_icp_crani;
class group;
model los_adj = group;
lsmeans group;
run;

Can someone please explain to me what the lsmeans output in a phreg model actually means and why the estimates are so different than what the actually group means are? Thank you.

[SteveDenham]

A comparison of the lsmeans output from the two PROCs would be useful to check whether the difference can be explained by the scale on which the models are operating.  That is what I think is happening, but I could be WAAAAAY off.

SteveDenham

[SASUser64]

Thanks, Steve! Here is the output from the phreg model:

Group Least Squares Means
Group	Estimate	Standard Error	z Value	Pr > |z|	Alpha	Lower	Upper	Exponentiated	Exponentiated Lower	Exponentiated Upper
Medicaid MCO	0.2717	0.06714	4.05	<.0001	0.05	0.1401	0.4033	1.3122	1.1504	1.4968
Medicaid FFS	0.2227	0.06518	3.42	0.0006	0.05	0.09491	0.3504	1.2494	1.0996	1.4197
Private Ins.	0	0	.	.	0.05	0	0	1	1	1

Here is the output from the regression model:

Group Least Squares Means
Group	Estimate	Standard	DF	t Value	Pr > |t|
		Error
Medicaid FFS	36.3212	1.3628	2074	26.65	<.0001
Medicaid MCO	29.9707	1.3407	2074	22.36	<.0001
Private Ins.	21.9483	1.2914	2074	17	<.0001

[SteveDenham]

And I am really off - it's not a scale thing. Maybe @SAS_Rob would be a good resource, or could at least point you to someone who is.

SteveDenham

[SAS_Rob]

I would say that there is really no fair way to compare the two models as they are now.  PHREG is fitting a competing risks model with a very different likelihood function than GLIMMIX would fit (treating the response as having a normal LL by default).

This section of the documentation might be helpful in understanding some of the differences.

SAS Help Center: Analysis of Competing-Risks Data