That is how I would view them.  As compared to method 2, in any case. ... View more

Noise. The additional parameters probably fit the residual noise after fitting all of the other variables. ... View more

I think I can help with some of the questions, but may miss some things. First, if you want pairwise odds ratios for genotype, try changing your lsmeans statement to: lsmeans genotype/at sweek=-2 diff oddsratio; For the odds ratio of the covariate, I think you have what you need.  The last four lines of the "Odds Ratio Estimates" table should give the results comparing sweek=-2 to sweek=6, as you move out 8 units.  If you want to compare sweek=-2 to sweek=8, you must change the portion in parenthesis to: (diff=all at sweek=-2 unit sweek=10) The next part I am really unsure about, but since you are testing odds ratios in the lsmeans statement, you may be able to change the parenthesis clause in the MODEL statement to: (at sweek=-2 unit sweek=8) I think the documentation says that the DIFF option applies to classification main effects.  Removing it here may make things easier to follow. As far as the quadratic effect, sweek*sweek is the correct syntax.  However, I'm not sure about the use of the AT statement under these conditions--check the documentation as for how interactions of continuous effects are calculated.  You may have to use the AT, to get around the use of mean values, I'm just not sure. Residuals for binomials?  Would plots=residualpanel get you started? I hope this helps. ... View more

Warning.  Untested code will follow. I think you will need lsmeans statements that set the value of pre.  Something like: proc mixed;   class group;   model post = group pre group*pre/ddfm=kr;   lsmeans group/diff at pre=val1 cl adjdfe=row;   lsmeans group/diff at pre=val2 cl adjdfe=row;   lsmeans group/diff at pre=val3 cl adjdfe=row;   lsmeans group/diff at pre=val4 cl adjdfe=row;   lsmeans group/diff at pre=val5 cl adjdfe=row;   ods output diffs=diffs; run; The dataset diffs should include the differences between all least squares means of group at various values (val1 through val5) of pre. Here I have variable names, but you really need the values at which you want to estimate the least squares means.  I include an ODS output statement, in case you want to plot the differences and their confidence limits. Good luck. ... View more

This sure looks like a repeated measures analysis, with random effects thrown in.  In this code, trial is NOT a repeated factor, but rather a random effect.  The repeated nature is found in the three response variables measured on each subject.  The documentation for PROC MIXED offers a different approach, that considers both the measure and trial as repeated factors.  Look for the Kronecker product structures for multivariate repeated measures.  I think that this would provide a better analysis in the long run, if in fact trial is repeated. The code would look something like: proc mixed data=long;    class Subject Diameter Height Slope Trial Var; /* The variable Var identifies the three response variables*/     model response=                Diameter|Height|Slope|Trial|Var ; repeated Var Trial/ type=un@ar(1) subject=subject; /* This assumes that the time between trials is equally spaced.  If not, then type=un@un would probably be a first choice*/                LSMEANS Diameter|Height|Slope|Var / ADJUST=TUKEY DIFF ;                        run; This requires re-arranging the input dataset. Suppose you have the input set now, call it work.wide. data long; length var $12.; set wide; response=Shld_P1_Abd, var='Shld_P1_Abd';output; response=Shld_P1_Flex; var='Shld_P1_Flex';output; response=Wrist_P1;var='Wrist_P1';output; drop Shld_P1_Abd Shld_P1_Flex Wrist_P1; run; This sets up the dataset for the doubly repeated measures. ... View more

data have; input ID othervar $6.; datalines; 1001 a 1001 b 1001 c 1004 d 1004 dd 1005 e 1006 f 1006 g 1006 h 1006 i ; /*strip down to unique ids*/ proc sort data=have out=temp1 nodupkey;      by ID; run; /*generate uniform random variable for each unique ID*/ /*I'm sure PROC SURVEYSELECT would be a good choice here if you want a fixed number of IDs, but this is a little more general, and a lot less precise*/ data temp2;      set temp1;      ranno=ranuni(1); /*so that this is repeatable, I specified a fixed seed*/      if ranno>0.3; /*You'll need to fiddle with this so that you get approximately the right number in your sample*/      keep ID ranno; run; proc sql;      create table want as select a.* from have a, temp2 b where a.ID=b.ID; quit; This should provide a good start... Edited to improve readability.  I hope this works.  Cutting and pasting from the SAS editor doesn't work so nicely ... View more

I think you will need a subject= specification in the random statement. Try: PROC GLIMMIX data=test; CLASS subjid; MODEL disease= grass200m / dist=binary solution oddsratio; random _residual_ / type=sp(exp) (x_cor y_cor) subject=subjid; run; Each subject should have its own set of x and y coordinates. SteveDenham ... View more

This still returns the x value in the OP. I believe he is looking to return lambda (or m in this example). The problem is that m is the solution for a cumulative sum of the following type: probability = sum (i from 0 to n) {exp (-m) * (m**i) / i!} I can only guess that some sort of grid search/optimization scenario would be able to do this on a general basis. SteveDenham ... View more

The value for a-b is ., which is less than any value. I would expect c=2 as the only possible result. Steve Denham ... View more

Just a guess. PROC REG handles missing values such that if any variable needed for any regression is missing, the observation is excluded from all estimates. If you had missing values for some of the independent variables that were NOT included in the final model, then the sample size, and hence Cp, would be different. Steve Denham ... View more

I really don't know if you can extend the methods to TOST but I would look at PROC NPAR1WAY for Hodges-Lehman estimates of location shift, and perhaps to PROC MULTTEST for bootstrap resampling of the t-test. Good luck with this. Steve Denham ... View more

How many independent variables are you dealing with? Are there a priori values that you will be interested in? A cartesian join of all the values generates a new data set, to be appended to the source data, with missing dependent values. If you are only interested in particular values that arise after gathering more data, then you are stuck with re-running everything. Steve Denham ... View more

Do you know the values of the independent variables that you will want predictions for prior to running the regression? If so, simply append them as observations, with the dependent variable set to missing (=.). The output data set should contain predicted values, standard errors, confidence limits, etc. for all observations in the dataset, so these can quickly be extracted. Steve Denham ... View more

You should read Milliken and Johnson's Analysis of Messy Data. Probably the best approach is to fit a means model, and address the tests of effects through the use of contrast and estimate statements. Another approach, find out what combinations of the independent variables are missing. If you can combine levels so that there are no empty cells in the highest level interaction in your model, and it still addresses your objectives, then you are home free. Also, check that the covariates age and wbvcalc are not completely confounded with region. Good luck, Steve Denham Added thought: Last, you say that you are fitting a "reduced model". I would guess that you have dropped some terms that were "non-significant." Is that really a good idea? Were those terms part of the study design, so that they have meaning, even if not significant? And most importantly, are the least squares means estimable in the "full model"? Message was edited by: SteveDenham ... View more