Date: Tue, 19 May 2009 17:13:58 -0500 Reply-To: Robin R High Sender: "SAS(r) Discussion" From: Robin R High Subject: Re: glimmix question Comments: To: elodie In-Reply-To: Content-Type: text/plain; charset="US-ASCII" elodie Sent by: "SAS(r) Discussion" 05/19/2009 10:44 AM Please respond to elodie To SAS-L@LISTSERV.UGA.EDU cc Subject Re: glimmix question On May 19, 9:37 am, rh...@UNMC.EDU (Robin R High) wrote: > For the record, I hastily typed in the incorrect formula for the > cumulative probabilities yesterday for the cumulative logit model, so the > "modified" version based on the printed intercepts below (if I copied them > correctly) should be: > > Cumulative Probs Individual probs add to > 1 > row 1: EXP(-3.6743)/(1+EXP(-3.6743)) = 0.02474 prob A = .02474 = 0.02474 > row 2: EXP(-2.3303)/(1+EXP(-2.3303)) = 0.08864 prob B = .08864 - .02474 > = 0.06390 > row 3 1 prob C = 1 - .08864 > = 0.91136 > > The ILINK option gives the cumulative probs for all but the last level -- > the individual ones can be computed as shown above. > > Robin High > UNMC > > elodie > Sent by: "SAS(r) Discussion" > 05/18/2009 09:00 AM > Please respond to > elodie > > To > SA...@LISTSERV.UGA.EDU > cc > > Subject > glimmix question > > Hi all, > > I am running a generalized regression. The dependent variable has 3 > categories. > > The syntax I am using is > > proc glimmix data=; > class timeclass ; > model dependent var categorized = / solution ; > random intercept / subject=id; > run; > > The output reads: > The GLIMMIX procedure is modeling the probabilities of levels of > the dep variable having lower Ordered Values. > > Effect dep var categorized Estimate > > Intercept 1 -3.6743 > Intercept 2 -2.3303 > > I am confused about what those two intercepts mean. I think that if I > take the exponential of intercept 1, I get an odds ratio. But I do not > know whether it measures the odds ratio of going from category 2 to > category 1, or from category 3 to category 1, or some other > transition. The manual of proc glimmix leaves much to be desired. Any > ideas? > > Thanks a lot for the help. Thank you so much for your messages. I very much appreciate your help. You are right, glimmix is a difficult procedure to use, and I need more practice with genmod, logistic and the more basic procs. The ILINK option is useful for getting the estimated probabillities, thanks. I have another question. When I include a (0-1) binary covariate to the model of my original post, I get a (small but significant) negative coefficient. I will denote the coefficient beta. For some reason, I do not get an odds ratio when I ask for it. This is confusing. So I do not know how to interpret that negative coefficient. Should I interpret beta using the following equation exp(\beta) = \frac{P(Y=B|X=1)/P(Y=A|X=1)}{P(Y=B|X=0)/P(Y=A|X=0)}? or maybe exp(\beta) = \frac{P(Y=C|X=1)/P(Y=A|X=1)}{P(Y=C|X=0)/P(Y=A|X=0)} Any ideas? Thanks a lot for your help You didn't specify how you asked for it, though the following example may be of some help: PROC glimmix DATA=phy; FREQ count; MODEL resp = x / dist=multinomial link=clogit solution oddsratio ; * link=glogit; OUTPUT out=prd predicted(ilink)=pred; RUN; The cumulative logit works with data placed in a table like this: ----------------------------------- |Counts | Resp | | | |-----------------| | | | A | B | C |Total| |----------+-----+-----+-----+-----| | x | | | | | |1 | 10| 4| 12| 26| |0 | 8| 8| 33| 49| ------------------------------------ The SAS code above gives one odds ratio that is the same for groupings of adjacent columns {P(Y=A|X=1)/P(Y=B or C|X=1)} EXP (beta) = ----------------------------- {P(Y=A|X=0)/P(Y=B or C|X=0)} and {P(Y=A or B|X=1)/P(Y=C|X=1)} EXP (beta) = ----------------------------- {P(Y=A or B|X=0)/P(Y=C|X=0)} predictions in a table form: --------------------------------------- |Est Probs | prob1 | prob2 | prob3 | |-------------+-------+-------+-------| |x | | | | |1 | 0.3640| 0.1925| 0.4435| |0 | 0.1774| 0.1436| 0.6790| --------------------------------------- if you proceed with the calculations, something like: prb23 = prob2 + prob3 ------------------------------- |odds ratio 1 | prob1 | prb23 | odds = prob1/prb23 |-------------+-------+-------| |x | | | |1 | 0.3640| 0.6360| 0.5724 Odds Ratio = 0.5724/0.2157 = 2.65 |0 | 0.1774| 0.8226| 0.2157 ------------------------------- prb12 = prob1 + prob2 ------------------------------- |odds ratio 2 | prb12 | prob3 | odds = prb12/prob3 |-------------+-------+-------| |x | | | |1 | 0.5565| 0.4435| 1.2547 Odds Ratio = 1.2547/0.4728 = 2.65 |0 | 0.3210| 0.6790| 0.4728 ------------------------------- It's been a while since I checked all the details, but it works much like this with clogit. The glogit option produces separate odds ratios. Robin High UNMC 

Back to: Top of message | Previous page | Main SAS-L page