Hello Robin,

thank you for your comment. Yes, in these classes there are only failures (event NE 'T'). So the standard errors should be large. The problem is that the parameter estimates for the other levels change (only slightly) between model 1 and model 2. So, for simulation purposes which parameter estimate should I take?

Sebastian

-----Ursprüngliche Nachricht----- Von: Robin High [mailto:robinh@unlserve.unl.edu] Gesendet: Donnerstag, 11. Januar 2007 17:51 An: Hein.Sebastian (FORST) Cc: SAS-L@LISTSERV.UGA.EDU Betreff: Re: PROC LOGISTIC

> > I am running a logistic model with one categorial variable "bo" coded in > the GLM way. The first two intercepts bo=1 and bo=2 are not significant, > whereas the intercepts for "bo" >= 3 are significant. Contrast statements > (not in code here) allow to test for differences between parameter > estimates of "bo". (Model1) So far it's ok. > > But now, for SIMULATION purposes (not for interferences on differences > between parameter estimate of (bo") of "bo" = 3, should I use the parameter > estimates of the code below OR run separate models by individual "bo". > (Model2) > > Of course the parameter estimate of the intercept of "bo" will vary > (slightly). Any help is greatly appreciated. > > Sebastian > > /*Model 1*/ > PROC LOGISTIC data=test; > class bo /param=GLM ; > model LT (event='T')= bo X1 X3 / noint lackfit(2) ; > RUN;QUIT;/ > > Type 3 Analysis of Effects > Wald > Effect DF Chi-Square Pr > ChiSq > bo 5 403.4486 <.0001 > X1 1 140.0823 <.0001 > X3 1 226.4185 <.0001 > > Analysis of Maximum Likelihood Estimates > Standard Wald > Parameter DF Estimate Error Chi-Square Pr > ChiSq > bo 1 1 -18.1066 399.3 0.0021 0.9638 > bo 2 1 -17.6244 2432.6 0.0001 0.9942 ^^^^^^^^ ^^^^^^ > bo 3 1 -4.0945 0.2184 351.5006 <.0001 > bo 4 1 -1.2139 0.1668 52.9599 <.0001 > bo 5 1 -2.5070 0.1570 255.1079 <.0001 > X1 1 -0.0693 0.00586 140.0823 <.0001 > X3 1 0.0631 0.00420 226.4185 <.0001

Hi Sebastian,

Before you run anything else, you should check why you are getting such large estimates and standard errors for bo=1 and bo=2 -- a sign that you have "complete" separation, which probably means you have all successes (event='T') or all failures (event NE 'T') for those levels.

Robin High