According to the PROC MIXED documentation, cov(i,j) = s^2*p^(d(i,j)) where d(i,j) is the Euclidean distance between i and j d(i,j) = sqrt( sum [over m from 1 to k](c(m,i) - c(m,j)^2). So for one dimensional data, like time series, your expression should be correct. Steve Denham ... View more

Dale, Did you mean GLIMMIX, rather than GLM? I got sensible answers from GLIMMIX, but only error messages from GLM. Maybe I missed something in the syntax. Steve Denham ... View more

I suppose that you currently have something like: proc glm data=one; model y=x; run; if so, then try: data two; set one; diff=y-x; run; proc glm data=two; model diff=; run; The standard error of the estimate will be slightly biased, but the point estimate should be OK. Steve Denham ... View more

Anna, If the descending option is kept, then at least try: proc logistic data=total descending noprint; model Y(event='0')=X / outroc=rocdata_sptot; output out=sptot1 p=predprob; run; Good luck, Steve Denham ... View more

Hi Anna, Try this: proc logistic data=total descending noprint; model Y(event='1')=X / outroc=rocdata_sptot; output out=sptot1 p=predprob; run; The way I read the documentation, this should give what you need. Steve Denham ... View more

You can get these estimates in a couple ways. What might be the most efficient is to include "dummy" data for the next year for each store in the input data set, with the response variable set to . Then, use the OUTP= option on the MODEL statement to generate a dataset that will have the predicted values for all observations, including those not used in fitting the model (i.e., those with the response variable set to missing for the next year). Steve Denham ... View more

Which procedure did you use to run the Poisson regression? With that info, we could probably come up with a way to calculate lsmeans and standard errors. Steve Denham ... View more

Peat, How about PROC NLIN? Try this: data one; input x y; cards; 1 28 2 40 3 45 4 48 5 52 6 54 ; proc nlin data=one method=marquardt hougaard; parms a=3 b=2; model y=a*x**b; run; I think the results will be straightforward. Now if variances are not gaussian, or there is clustering or hierarchical issues, you may wish to consider PROC NLMIXED. Steve Denham ... View more

I did a quick search of the PROC TTEST documentation, and learned something new (as I almost always do when I read the manual). For the lognormal distribution, the formula for the CV is NOT the standard deviation divided by the mean. It is instead sqrt(exp(variance) - 1). Given that, it doesn't surprise me so much that the CV of a lognormal variable is not the standard deviation divided by the mean. See also the documentation for PROC GLIMMIX, under the MODEL statement for the DIST= option, where the estimators for the lognormal distribution are given. Steve Denham ... View more

Check out the PARMSDATA= option. This should allow you to load in the values you want. Steve Denham ... View more

The statements you have will give marginal means averaging over blocks. This seems logical to me, as the realization of blocks in this particular study is a sample from all possible blocks. I think your measurements are at equal intervals, and if there are more than three or four, you may wish to use a time-dependent covariate structure, rather than the unstructured matrix, simply because the number of parameters to be estimated is smaller. You could check the AIC for this. Finally, you may want BLUPs for heartrate for the three blocks. These differe from the marginal means (lsmeans), but can be obtained from estimate statements. See the documentation for examples. Good luck. Steve Denham ... View more

The following code is untested. I assume that dose is categorical and fixed across drugs. If not, then see part 2. proc mixed data = mydata; class drug patient hour dose gen; model heartrate = gen|drug|dose|hour; repeated hour / subject=patient(drug) type=un; run; If dose is continuous and varies by type of drug then: proc mixed data = mydata; class drug patient hour gen; model heartrate = gen|drug|hour dose drug*dose; repeated hour / subject=patient(drug) type=un; lsmeans gen|drug|hour/at dose={insert value of interest}; run; These should provide a good start, but it may be that the covariance structure contains too many parameters, in which case an autoregressive or spatial power structure might be more usable. Steve Denham ... View more

Some pointers at what I was attmepting: proc glimmix data=multinom method=laplace ; I selected the method=laplace so that the residual variance would be included in the optimization. I can't give a good reason other than for small samples it is supposed to be better than the pseudo-likelihood methods. class animal day time group; t=time; nloptions maxiter=500; model score=group/dist=mult oddsratio(diff=all) solution; All this seems standard so far. random int day/subject=animal s cl; This fits random effects for animal and animal by day, in the most efficient manner. Be sure the data are properly sorted. random time/ subject=animal*day type=sp(pow)(t) s cl; This is my rather feeble attempt at handling the repeated nature of the data. I know I can't use _residual_ with a multinomial, so I tried to construct something on the G side that would mimic the true R side design. Looking at it, it shows some sort of relationship within animal-days, but assumes independence of the time points. That would be a best case scenario, and it is as close to the design as I could get. covtest glm; This is a handy thing that is currently only available in GLIMMIX, and provides a test of the model against a null model of full independence. I look at the ratio of the chi-squared value presented to the number of parameters excluded in the test as a possible indicator of model rankings for goodness of fit. WARNING WARNING WARNING: I have no real theoretical references for this. WARNING WARNING WARNING. run; As far as the output, it made sense to me. The odds ratios look about right, and the tests meet the eyeball tests of the plots. Steve ... View more

More: Perhaps the following will give what you are looking for: proc mixed data=data covtest nobound; class patientid hospital trt sex; model outcome=time trt time*trt sex age/solution; random hospital: random intercept time/ type=UN sub=patientid(hospital); run; This would give a variance component due to hospitals, and would fit the random regression to the patients within each hospital. However, I cannot guarantee that the linear dependency is removed. Steve Denham ... View more