Hi Rick That does not solve the problem. There are still 26 leverage points and the warning is still given. Peter ... View more

Thanks Steve.  Hmm.  Maybe the developer of QUANTREG will chime in. ... View more

You are dealing with the question of whether you want a stratified analysis (one for the low BMI group, one for the high BMI group) or a moderated model with an interaction term. If you stratify the analysis, you could use a BY statement: proc sort data = inputdata; by bmigroup; run; proc glm data=inputdata;        class trgrp ;        model outcome = trgrp;       by bmigroup; run; if you want a moderated model you could run proc glm data=inputdata;        class trgrp bmigrp ;        model outcome = trgrp bmigrp trgrp*bmigrp;       lsmeans trgrp bmigrp trgrp*bmigrp; run; and you could have an estimate statement as well - although I find LSMEANS easer to write and to have easier output to read. ... View more

If you want to estimate the effect of trgrp in one BMI group then do this: proc glm data=inputdata;        class trgrp ;        model outcome = trgrp;       where bmigroup = 0; run; ... View more

Thanks again! ... View more

Thanks Steve, Helpful as always.  That ran.  One thing that worries me is that the df is the same for the city level effects and the person level effects.  Is that correct?  I know estimating the df in these models is tricky and full of options Peter ... View more

Nothing in NLMIXED is as simple as it is in GLIMMIX.  However, the NLMIXED documentation provides examples of models with covariates. See e.g. SAS/STAT(R) 9.2 User's Guide, Second Edition  where they give this code for a model with several covariates: proc nlmixed data=inhaler corr ecorr;   parms b0=0 b1=0 b2=0 b3=0 sd=1 i1=1 i2=1;   bounds i1 > 0, i2 > 0;   eta = b0 + b1*group + b2*time + b3*gt + u;   if (clarity=1) then p = probnorm(-eta);   else if (clarity=2) then   p = probnorm(i1-eta) - probnorm(-eta);   else if (clarity=3) then   p = probnorm(i1+i2-eta) - probnorm(i1-eta);   else p = 1 - probnorm(i1+i2-eta);   if (p > 1e-8) then ll = log(p);   else ll = -1e100;   model clarity ~ general(ll);   random u ~ normal(0,sd*sd) subject=sub;   replicate freq;   estimate 'thresh2' i1;   estimate 'thresh3' i1 + i2;   estimate 'icc' sd*sd/(1+sd*sd);   run; ... View more

A preliminary warning: These are complex models that are easy to misuse and misinterpret.  A substantial amount of study will be necessary to use these models well. First, I'd ask if you are using multilevel models to deal with longitudinal data or clustered data. If it's longitudinal data then I recommend two books by the "two Geerts": Mollenberghs and Verbeke, Linear Mixed Models for Longitudinal Data - for MIXED and Mollenberghs and Verbeke, Models for Discrete Longitudinal Data - for GLIMMIX I  find these considerably clearer than Littel et al (which is also a good book, but demands more math) An even easier book is Hedeker and Gibbons Longitudinal Data Analysis but it is less specific to SAS. If you have data that is clustered, then one good book (especially if you have educational data) is Hierarchical Linear Models by Raudenbush and Bryk; it's not SAS specific, but it will provide the background you need. ... View more

Yes, you're right in your interpretation. ... View more

No, I don't think it's better to delete some levels of a categorical variable.  That winds up being an uninterpretable model. E.g. suppose the variable is race and you have White, Black, Asian, Other.  Suppose only White and Asian are significant.  Then if you delete the other levels, you are comparing Whites to Asians without controlling for Black or Other.  Keep all levels. Parsimony is often the enemy ... View more

Looks OK to me ... View more

Everything in STEPWISE is, at best, highly questionable and, at worst, outright wrong. However, here, you have shown that you can make stepwise behave in either of two ways: Treat the categorical variable as a single variable or treat each level as a single variable.  I recommend the first. Perhaps you want to exclude any variable that is insignificant at any level?  I think that would be an (added) mistake, but you could certainly do it by  hand (e.g. by removing that variable from the list). ... View more

It depends on where the separation is.  When there is quasi-complete separation, usually one variable will have very odd confidence intervals.  Those shouldn't be used. But the other ones can.  See this paper by Paul Allison ... View more