I think this is a data dependent problem, more than a syntax problem.  I will bet that most of your information is at a limited number of flowpos values.  Reparameterizing these to something where the responses are more evenly spaced could help if that is the situation. I also think you would have better luck not having two versions of the flowpos variable.  Without data to try this on, what about sorting the data set by plot treatment poll_supp ind flowpos? Then trying to execute: proc glimmix data=sorteddata initglm scoring=5;       class treatment plot poll_supp flowpos ind;       model nseeds = treatment poll_supp flowpos treatment*poll_supp treatment*flowpos poll_supp*flowpos treatment*flowpos*poll_supp / dist=negbin offset=lntovules             ddfm=kenwardroger solution;       random flowpos/ subject=ind type=ar(1) residual;       random plot; run; I included ind as a class variable.  The subject=ind assumes that all individuals are uniquely identified, so that an infinite likelihood is avoided. The initglm option should generate starting values based on excluding the random effects, and the scoring=5 will give five iterations of Fisher scoring to get things started in the right direction (hopefully). Good luck on this one. Steve Denham ... View more

You are definitely on the right track.  The last two group= statements are equivalent, when it comes to constructing the R matrix.  It will be block diagonal, with each block being a CS submatrix referring to the subject-trial combination.  I am curious about the full design--it may be that a doubly repeated measures using type=UN@CS might be an approach to consider.  However, the best specification is all dependent on the design.  More info would probably clear this up pretty quickly. Steve Denham ... View more

First, a t-test on the proportions may be the only test you can get from SURVEYMEANS.  I could be wrong and if others know better, I hope they will contribute. To get the ANOVA with weighted data you are probably going to have to shift to PROC SURVEYREG.  For that to work, you will need classification variables and a response variable.  I presume vary and varx are the classification variables.  More info on this part will make responding easier. To get mean score values from SURVEYMEANS, try dropping the class statement and adding 'mean' to the proc statement: proc surveymeans data=XXXX t mean; weight weightvar; var vary varx; run; I do wonder what meaning you can attach to a mean for categorical variables, however.  It will be dependent on the number of categories and response patterns.  Perhaps the median and various quantiles would be more informative.  I don't know the field you are working in, so I can't say what might be standard practice. With more info on the design and sampling schema, I am sure that SURVEYREG can be adapted to do the ANOVA. Steve Denham ... View more

You may want to look at PROC NLMIXED.  You would specify the model similarly, but include a term for the measurement error.  Take a look at exaomple 63.1 in the NLMIXED documentation, where a one-compartment model for PK data is fit.  I think you could easily gin up your NLIN code to accommodate the measurement error. Steve Denham ... View more

Once upon a time, when mastodons roamed the earth, and I was a first year graduate student in a field that wasn't statistics, but used statistics a lot, I took a course that used Sokal and Rohlf's Biometry as a text.  In it was an example of how to analyze count data, using the square root transformation.  At the bottom of page 384, it said, "For reporting of means the transformed means are squared again and confidence limits are reported in lieu of standard errors." So, I supposed that you had transformed the data using a log transform, calculated a mean, and then exponentiated that value to get the expected value of your counts, for the "raw" data. It is important to get one thing out of that old book--Confidence limits are reported in lieu of standard errors.  Backtransforming the estimate of the standared error obtained gets you into trouble.  That's why lvm suggests reporting both.  One of the great things about PROC GLIMMIX is its use of the delta method to get back-transformed standard errors.  Still, I always expect asymmetric confidence bounds on non-normal data. Good luck. Steve Denham ... View more

Would this not simply be a simple IF-THEN-ELSE statement? if increase>=0.2 and aval>uln then critfn=1;      else critfn=0; Where increase, aval, and uln are previously defined on the data set. Steve Denham ... View more

Hi Serena, When you calculate means and standard errors of the original data, do you transform before calculating and then back transform?  If not, the values are somewhat misleading to begin with.  I truly believe that best estimators are those obtained from the analysis, rather than the raw data. Given all of that, I would suggest the following version for your lsmeans statement: lsmeans trt/ilink diff lines; This will give the mean and standard error on the original scale (ILINK) and letters for the bar graph (LINES). Good luck, Steve Denham ... View more

You may want to look into PROC TTEST.  It can compare data that comes from summary statistics.  The work will be in getting your coefficient datasets into shape, with the necessary labeled variables (mean, standard deviation, number of observations). Good luck Steve Denham ... View more

Difficulties I see, without having a good dataset to test this on: 1. The multinomial distribution option in GLIMMIX does not support calculation of BLUEs (LSMEANS), so residual calculation may be similarly not supported.  Maybe someone else has better information, but this is my best guess as to why you aren't getting residuals. 2. You are not modeling the repeated nature of the data in the current code.  The RANDOM statement should look something like: random day/residual sub=id type=cs (for example only, you may have a better idea of what covariance structure fits your data).  You will also have to include day in the CLASS statement. 3. If you include R side effects like this, then method=quad will not work--the only available methods are the pseudo-likelihood methods. 4. To get odds ratios, add oddsratio to your model statement.  Perhaps this will give you the info you need to check the proportionality assumption. Good luck, and keep posting on this.  I am facing a similar challenge, and think I can learn from what happens here. Steve Denham ... View more

Following on to Ksharp's response that PROC REG is not a good tool for forecasting, what you might do is run PROC REG and get the Durbin-Watson autocorrelation statistic.  If it is not extreme, then extrapolation may be good for at least short projections ahead. But to limit the response to a specified range is difficult.  You could create the CI on the predicted values, and when it exceeds your range, state that the model is good only out to that point. Another approach could be to apply a differencing schema to your data in an attempt to introduce stationarity.  You still have the problem of extrapolation, but autocorrelation should be reduced.  Produce the confidence interval at each extrapolated time point.  Back-calculate the values from the differenced predicted values, and see how far out you can still meet your pre-specified range. One more thought:  I really don't know of any methodology that allows you to limit the range of predicted values for extrapolated independent variables a priori.  If someone does, I would be very, very interested in learning about it. Steve Denham ... View more

I think you might have to sneak up on this one.  Use PROC MIXED to calculate non-central F values for a range of sample sizes and expected effects, and see which one best approximates the power desired.  See Chapter 12 in Littel et al.'s SAS for Mixed Models, 2nd. ed., and expand from there. Steve Denham ... View more

There are at least two curses to watch out for, in my opinion.  Overfitting is one--the more variables you have the more likely you can fit even the most random process results.  Remember, if I have ten data points and nine variables, there is at most one solution satisfying all constraints.  To gain any information on the variables that are in the model, we need substantially more data points.  The other is, loosely speaking, collinearity.  Some variables are just too closely related in some n-space to provided independent information. And then there comes the problem of prediction.  Simply fitting the existing data becomes easier and easier as we add variables.  However, that does not guarantee the ability to predict.  A validation data set is needed, or at least a good cross-validation plan.  For example, minimizing the mean squared prediction error is a different task than minimizing the residual variability in a training set. Steve Denham ... View more

That is not an easy answer that can be dropped in this space.  I have attached some code that does a five parameter logistic regression for a cell based bioassay that might be adaptable, but beware, it is not documented, it does get convoluted, and it was written for v9.1.3. I hope this will help. Steve Denham ... View more