Estimate genmod

[Enas]

 I have this model and I try to estimate the difference between the high volume (1 2 5 ml ) compare to lower volume (.1 and .5)

proc genmod data= germ; *where Primary_Response_EDR_SP_NR in('EDR','SP') and Infusion_Volume ne 5;
class MRN DBI_Category (ref='NORMAL') Infusion_Media Infusion_Volume (ref='0.1') ;
model Secondary_Response_EDR_SP_NR= DBI Infusion_Media Infusion_Volume DBI*Infusion_volume DBI*Infusion_Media /dist=multinomial
link=cumlogit type3 wald;
repeated subject=MRN/ sorted;

*lsmeans DBI_Category /pdiff=all adjust=bon exp cl;

estimate "Upper in Air " intercept 6 dbi 6 Infusion_Media 6 0 0 Infusion_Volume 2 2 2 0 0 DBI*Infusion_volume 2 2 2 0 0 DBI*Infusion_Media 6 0 0/ divisor=6 ;
estimate "lower in Air " intercept 6 dbi 6 Infusion_Media 6 0 0 Infusion_Volume 0 0 0 3 3 DBI*Infusion_volume 0 0 0 3 3 DBI*Infusion_Media 6 0 0/ divisor=6 ;
estimate " upper -lower in Air" Infusion_Volume 2 2 2 -3 -3 DBI*Infusion_volume 2 2 2 -3 -3 / divisor=6;
run;

I have this outcome and I don't why it is not estimable ???

Contrast Estimate ResultsLabel Mean Estimate Mean L'Beta Estimate StandardError Alpha L'Beta Chi-Square Pr > ChiSqConfidence Limits Confidence LimitsUpper in Airlower in Airupper -lower in AIr

0.2944	0.2022	0.4071	-0.8743	0.2542	0.05	-1.3726	-0.3761	11.83	0.0006
0.2347	0.1249	0.3972	-1.1821	0.3904	0.05	-1.9472	-0.4170	9.17	0.0025
Non-est	.	.	Non-est	.	.	.	.	.	.

I

[SteveDenham]

I think it is because DBI is a continuous variable in your model, which leads to issues with the continuous by categorical interactions.  Also, with a cumulative multinomial, you should have multiple intercepts, some of which need to be included in the estimate statement. 

Have you tried using LSMESTIMATE statements? It appears to me that you want to test for a difference in the lsmeans (log odds), so the advantage that LSMESTIMATE has is that all of the parameter matching, etc is handled.

SteveDenham

[StatDave]

I strongly advise using LSMEANS or SLICE instead of the ESTIMATE statement which requires determining the correct coefficients on the model parameters to estimate the quantity of interest and is a very error prone task. In a binomial or multinomial model, differences like you want are differences of log odds and exponentiating those differences yields odds ratios. You can easily do this with the LSMEANS statement with the DIFF and ODDSRATIO options. You might also want to add the E option to see a table showing the coefficients on the model parameters used to estimate each log odds. Note that in a multinomial model with k response levels, there are k-1 cumulative logits so you will get k-1 log odds estimates for each level of the predictor in the LSMEANS statement. The difference, however, is the same for each logit, so there is only one difference (and odds ratio).

lsmeans Infusion_volume / e diff oddsratio;