You need to add a TEST statement:   proc mianalyze data=type3test; class effect; modeleffects effect; test effect; run; All of the examples in the documentation for the TEST statement look like they are for a regression on continuous variables though, so I am not sure this is the correct syntax for a categorical variable. If the code doesn't work (and I am guessing it may not which is why I am editing my response), then you will probably need to use an ODS output statement to get the parameter estimates and the covariance matrix, and then hand calculate (DATA step or PROC IML) the F test.     Using my friend Mr. Google, I found this link on Paul Allison's site: https://www.sas.upenn.edu/~allison/combchi.sas The COMBCHI macro requires PROC IML, so if that is not available to you, then this method wouldn't be an option.   BarbClothier presents a dummy variable approach in this thread: Proc Mianalyze: obtaining combined results for cat... - SAS Support Communities.  This approach will allow for the use of the TEST statement, but be prepared to edit the COVB dataset.   SteveDenham   ... View more

If ageA and AFF are completely confounded (as in @Rick_SAS 's example) where you can express one as a linear function of the other, then the easiest thing to do is to remove one of them as a model effect.  For example:   proc glm data=par1; class ageA wta; model FR1 = ageA wta ageA*wta; lsmeans ageA wta /pdiff; lsmeans ageA*wta / slice= (ageA wta); quit; run; This will give comparisons between the levels of ageA  as well as comparisons between the levels of wta.  The lsmeans statement for the interaction uses the SLICE= option to obtain F tests for the simple effects of wta at each level of ageA and of ageA at each level of wta.  If you want specific comparisons of the lsmeans at each level, you'll need to employ the STORE option to get the SLICE statement.  Note that now there is only one age related effect, so the overall F tests will not separate out a test for age at first calving.  However, that shouldn't be an issue if age at first calving is a simple additive function of age at arrival.   SteveDenham   ... View more

I am confused by the day to month correspondence you present (2 months = 168 days, 5 months = 532 days).  It is unlikely to affect your analysis, but I think interpretation of results might be difficult as it looks like the two cohorts are not commensurate.   SteveDenham ... View more

Take a look at the example https://documentation.sas.com/doc/en/pgmsascdc/9.4_3.4/statug/statug_glimmix_examples24.htm    The variance components needed to calculate an ICC appear in the Covariance Parameter Estimates table in the output, which has the ODS table name CovParms.   SteveDenham ... View more

Since you are using PROC PANEL (I think), you will probably get more specific results if you post this to the Analytics>SAS Forecasting and Econometrics community.   SteveDenham ... View more

I am a bit confused - with a single rater, what are classes in your design that led to the suggestion that an ICC would be useful.  I understand ICC's for multi-rater designs.  So, to be able to give some advice, knowing the design for your data would be really helpful.  Once the random effects are identified and output via ODS, calculation of the ICC is a proc transpose and a data step away.    icc = variance component for the classification variable of interest / sum of all variance components including residual   SteveDenham ... View more

Have you tried a generalized logit link?  A generalized logit link is available in PROC GLIMMIX.  My thought is that a generalized probit would give approximately the same results, so this may be an alternative.   If you stay with GENMOD, you will probably have to write your own link functions using the FWDLINK and INVLINK statements, to get the generalized types.     Since the probit distribution assumes an underlying normal distribution with "cutpoints", I am not sure that a generalized probit is particularly meaningful.  The alternative might be to set up various binary classifications.  That would probably require some (meaning, A LOT) of post processing programming.   SteveDenham ... View more

Look at the examples for PROC GENMOD - there is an example that fits a multinomial response with a cumulative logit distribution.  Changing to a cumulative probit is done with the LINK= option in the MODEL statement.  I would recommend running the example program with the cumulative probit to see how the program performs, and then adapting it to your particular design.   SteveDenham ... View more

Hi Will,   Not sure if this is the issue, but if you have a lot of zero's in your data, then a simple fit to a Poisson distribution may not be the best approach.  You may have a zero-inflated distribution.  Two ways to approach this: Mixture of Poisson and a mass at zero, or a hurdle model (also a mixture, but one of the distributions is binomial, with the probability there splitting the full population into "structured" zeros and everything else.  That might be interesting for your other predictors, as they may have a bigger influence on the split than on the actual level in the non-zero observations.   Of course, that answer tells you close to nothing about how to calculate residuals.  Try PROC FMM for the binomial-Poisson mixture and PROC GENMOD for a strictly zero inflated Poisson.   SteveDenham ... View more

Well, with only the information in the table, you won't be able to calculate by hand the SD and SE for each of the groups.  Well, first off, whatever analysis was done assumed homogeneity of variance - you can tell because the standard errors for the lsmeans are identical.  With different N's per group, that implies that the lsmeans are truly marginal means -> expected values of the distributions, rather than the averages for each group.   What you could try is to simulate data with these means and standard errors, having say 500 or 1000 samples of 298 and 294 observations each.  Calculate the mean, standard deviation and standard error for each sample. That would be enough to let the Central Limit Theorem give you an accurate approximation if you took the average of each set of samples.   SteveDenham ... View more

Everyone sees something in data based on what they do the most.  If I were faced with this dataset, I would consider some sort of repeated measures using a generalized linear model.  For instance, if you were looking for estimates for each player, as a conditional value over time, you might try:   data long; set topnfl; year=20;value=s20/100;output; year=19;value=s19/100;output; year=18;value=s18/100;output; year=17;value=s17/100;output; year=16;value=s16/100;output; year=15;value=s15/100;output; keep player year value; run; proc glimmix data=long; class player year; nloptions maxiter=1000; model value=player year/dist=bin ddf=32; random year/residual type=ar(1) group=player; lsmeans player/diff ilink; run; I set the denominator degrees of freedom to 32, which is what the 'skeleton' ANOVA table (no RANDOM statement) provides in PROC GLM. I used the binomial distribution because the measures are the result of an undetermined number of Bernoulli trials (either made or missed).   SteveDenham ... View more