I am not sure what is going to happen with this code, but it might trigger something on your side. What happens if you change the RANDOM _residual_ to RANDOM _residual_/group=sex ? I would expect to see two estimates for residual error in that case. To actually see them and their confidence bounds you need to a COVTEST statement. In this case it would look like:   covtest homogeneity /cl; I don't see a good method in GLIMMIX to get the SE for the residual variance, but that doesn't mean there isn't one. GENMOD does give an SE for the scale parameter in a Poisson regression, that may or may not have an SE>0, depending on the data.   SteveDenham ... View more

For GLM: To test for homogeneity of variance across CLASS levels in a one-way ANOVA setting, use the MEANS statement with the HOVTEST option. There are a variety of tests available. I prefer the Brown-Forsythe method.   For mixed models (GLIMMIX being preferred); Include a RANDOM statement with the following syntax:   RANDOM _residual_/group = <class variable>; And add the following COVTEST statement:   COVTEST 'common variance' homogeneity This conducts a likelihood ratio test. One nice thing about this approach is that even for moderate sized datasets, you can ignore the testing part, and rely on convergence in distribution of the variance components if, in fact, there is no heteroscedasticity. You do this at the cost of degrees of freedom, so if you are in a situation where power of the test is at a premium, you can remove the statements if the test is non-significant.   For a situation where you have a main linear regression and you are looking for the influence of fixed categorical covariates, the GLIMMIX approach is quite handy.   SteveDenham     ... View more

Consider moving your post over to the SAS Forecasting and Econometrics forum, as SYSLIN is part of SAS/ETS   SteveDenham ... View more

Could it be as simple as the variable is called "new_surgery1*timetrying" in the MODEL statement for PHREG, but is called "TIMETRYING*NEW_SURGERY1" in the MODELEFFECTS statement for MIANALYZE? MIANALYZE can be picky about things like the variables not having identical names (including case on occastion).   SteveDenham ... View more

What do you see when you add the inverse option to the model statement:   proc glm data=outmi; by _imputation_; class TRT02PN; model W24CHG=TRT02PN/solution inverse; means TRT02PN /hovtest=leven(type=abs); ods output ParameterEstimates=coeff; run; And sometimes it is easier to step away from GLM and try something like this:   proc glimmix data=outmi; by _imputation_; class TRT02PN; model W24CHG=TRT02PN/solution covb; ods output ParameterEstimates=coeff covb=covb; run; This is solely to get the covariance matrix for the fixed effects, and not to combine the imputations.   SteveDenham     ... View more

Consider that the two dependent variables are correlated measures on each subject, i.e. two manifestations of a latent variable. PROC MIXED supports Kronecker product covariance structures to handle this. See this paper by Tao, Kiernan and Gibbs  https://support.sas.com/resources/papers/proceedings15/SAS1919-2015.pdf    To make this work you will need to move from the long form of your data to an "extra-long" format, adding a new variable (call it new_DV that takes on a value of 1 when the response variable is DV1 and 2 when the response variable is DV2. In the extra-long format, each record will have a unique combination of id, condition, school, class, place, sex, time and new_DV   Code for the analysis could look like this: proc mixed data=long; class id Condition school class place sex time new_dv; model response = new_DV|condition|time|sex|age/solution ; Random intercept /subject=class(school); Random intercept /subject=school(place); Random intercept /subject=place; Repeated new_DV time/type=un@un subject=id r;  This assumes that ID is unique across all schools, classes and places. If not, you'll have to nest ID in the subject= option for the REPEATED statement.   I hope this is along the lines of what you are looking for.  I fear that fitting a five-way interaction and all lower level interactions might be a bit dicey, unless you have thousands of records.   SteveDenham ... View more

PROC MEANS and PROC UNIVARIATE are likely candidates. Either will estimate medians, quartiles and the interquartile range. Those would be the descriptive non-parametric parameters for a single arm.   If you have access to SAS 15.2, PROC TTEST has a bootstrap option. In my opinion, this is a more robust nonparametric approach. It does require you to do some exercises to make it applicable to a single arm, but that is not a stopper.   SteveDenham ... View more

This is a "not sure if it will work" answer, but the degrees of freedom look a bit wonky, and it is probably because you are using the Kenward-Roger denominator degrees of freedom with a very sparse dataset due to the missing values. Try, in order, ddfm=satterthwaite, ddfm=bw, or using ddf= with fixed values based on a skeleton ANOVA table. You could then add in an EMPIRICAL option to get the shrinkage estimates.   SteveDenham ... View more

LLOD = can of worms. There is a lot of statistical literature out there on how to handle values that might be left-censored at a lower limit of quantitation, but few that apply to LLOD, as that can be defined as the value below which measurement noise is greater than the signal. I would impute values below LLOD as coming from a uniform distribution on [0, LLOD], which happens to have an expected value of LLOD/2. So, your imputation is probably not the issue. What I suspect is that the plot of the data has a hockey-stick appearance, with a long flat area out to the left, where most of the values are either very small or imputed. If you fit a line out there, it is going to have a slope approaching zero, with a relatively large residual SD. When you plug that into the back calculation for t1/2 , you get a really large value.   But think about it, the concentrations out in that area are due to non-single compartment behavior, either slow release from a precursor compartment or some sort of flip-flop kinetics.  What my colleagues in the pharmacokinetics group do is to examine the time course curve, and only use the data from the descending portion prior to entering LLOD territory. They set values <LLOD to zero, and any values after the first zero are also set to zero. That more or less restricts the analysis to a single compartment with a single elimination rate constant that is time-independent. I can see their point - I would rather estimate the rate constant with 4 points descending than 104 points where the last 100 are effectively zero.   Now if your time-concentration curve isn't nicely behaved, such as having two (or more) peaks, long flat periods, or increasing concentrations well after initial dose, then noncompartmental analysis may not give you a believable half-life.   SteveDenham ... View more

Add the NOINT option to your MODEL statement.   SteveDenham ... View more

For the first question, I have a hunch that it might be due to a change in the denominator degrees of freedom as you have moved farm from a G side effect to an R side effect. Also, consider that the results for the F value do change when you go from an all G side parameterization to a G and R parameterization. Walt Stroup's text covers what could be happening with your situation.   The second part about why the number of events and trials differ so greatly between the two approaches is harder to answer. I am not sure how your data are structured, but I think that may be part of what is happening. For things like this, I have always trusted what @jiltao has said, and @Mike_N  has made some great comments recently.   SteveDenham ... View more

Mr. Google was my friend today. Searching on "Kenward roger approximation" led to this: https://documentation.sas.com/doc/en/statug/15.2/statug_glimmix_details40.htm    From this you will see that the denominator degrees of freedom is a solution to an equation, rather than the (n-1) or (n-1)(m-1) you think of as the design denominator degrees of freedom. If they happen to come out to the same value, it is because the data are balanced and the covariance matrices associated do not result in any shrinkage of the standard errors.   So if the KR ddfm values are not completely inappropriate (such as ddf =1 in a repeated measures design), they represent a less biased approach to the analysis. Even better for controlling bias would be KR2.   SteveDenham ... View more

Do you have estimates of the variances and the covariance? You could then try to calculate a correlation that way by using PROC FCMP or macro code to get a value for the correlation. No guarantees, though.   SteveDenham ... View more