First question, is the effect of any of your repeated predictors linked solely to the current response (this will have an easy response), or does the predictor at a previous timepoint affect the response as well (this gets into timeseries analysis)?   If it is the first case, then get your data in "long" format with each record having an entry for the subject_id, the timepoint, the response variable value at that timepoint, and each predictor variable value at that timepoint.   From there, questions like "Is this predictor continuous or categorical?" and "Is there a possibility of predictors interacting that could affect my interpretation?" and "What do I do if all my predictors are continuous, but have very different ranges of values?" become important.   In any case any of the linear modeling PROCs are capable of handling data of this sort.   SteveDenham ... View more

Save your code and data. then restart SAS, or disconnect and reconnect SAS Studio. At some point, it looks like you have an unmatched single or double quote.   From what you posted, it also looks to me like you called a macro named logistic_aic_sbc_score. The unbalanced quote may be in that macro..   SteveDenham ... View more

Question (actually a trick question) - how do you know that the distribution for residuals is not normal? Did you do some sort of test? There are well-known issues with almost every hypothesis test for normality (overpowered with N greater than about 40, underpowered for N less than about 15), and the linear mixed model is remarkably robust to the assumption of normality of the residuals, so long as the empirical distribution is mono-modal, not truncated, and lacks extremely large absolute values. The mono-modal basically boils down to sex differences.   So here are some ways to attack the issue, from simple to complex: bin your responses to four or five categories and consider using Cochran-Mantel-Haenszel methods where you stratify by sex. Plot your data and see what the shape looks like. From that, use a generalized linear model, assuming the distribution you have a picture of. If you have what might be considered random effects, use a generalized linear mixed model. Bootstrap your data. Simulate a lot of datasets that could possibly occur based on your current data. Use a Bayesian analysis with noninformative priors. This does a lot better job of simulating the data needed to construct credible intervals as you can include correlations over time or clusters. I don't think you have any random effects, so a good start on this can be found by looking through the documentation for PROC BGLIMM. Given what you have done so far, I would recommend #4. You can use most of your PROC MIXED code, and you can examine each distribution/link to see which best fits your data.   SteveDenham ... View more

Two part answer here. First a reply to @Ksharp : After fitting a model, the residuals may or may not be normal (Gaussian). For example, if you fit binomial data without accounting for the distribution with a link function, the residuals will not look Gaussian (it might take a lot of data). Second a reply to @SAS-questioner : If you only have 6 data points, why are you bothering to fit a model? The mixed model or GEE model parameters will have such large standard errors you probably won't be able to correctly infer from them.    SteveDenham ... View more

Just be sure that you use the monotonically decreasing part of these curves to estimate the half-life if you implement the method in @Rick_SAS 's blog. Including the increasing phase will lead to a best fit line that has a much, much smaller slope than the part of the curve dominated by elimination from the pool sampled. As a result, the half life will be greatly over-estimated.   SteveDenham    PS I am working through the paper by Kurada & Chen to use compartment modeling on this data. Here is a link: https://support.sas.com/resources/papers/proceedings18/1883-2018.pdf . Then on to  the paper by Hannion and Boulanger https://www.lexjansen.com/phuse/2020/as/PRE_AS06.pdf . From these, all of the data can be used, not just the monotonically decreasing data. ... View more

So I started to work on this with NLMIXED when I noticed something critical - these data do not describe a single pool decaying process of the type  c = c0 * exp (-k * time). It appears that at time=0 the response value is at or near zero, then increases to about time = 43, followed by a decrease. There are two logical models that fit this - a time independent two-pool model with sampling from the second pool, so that there is an entry rate constant, or a time-dependent model with sampling from a single pool.     Neither of those models fit the linearized (log transform) approach you have. One thing that might help would be to estimate the removal coefficient and estimate t1/2 from that. That would be easiest by removing time points less than 43 from the analysis and rescaling the time axis. However, that estimate of t1/2 will be biased by not having an accurate initial estimate of the maximum concentration and when it occurs.  However, in many pharmacokinetic studies, this is the standard approach, and it gives a good estimate of the removal coefficient and the resulting half-life in the pool.  Here is code to accomplish that:   proc nlmixed data=have2(where=(atptn>40)) maxiter=5000 ; parms c0 = 1.5790 d1 = 0 k0 = -0.5291 d2 = 0 s2 = 0.2069 ; pred = c0*trt1*exp(k0*trt1*(ATPTN-43)) + (c0 + d1)*trt2*exp((k0 + d2)*trt2*(ATPTN-43)); model aval ~ normal(pred,s2); predict pred out=pred; estimate 'initial for trt1' c0 ; estimate 'removal coefficient for trt1' k0; estimate 'initial for trt2' c0 + d1; estimate 'removal coefficient for trt2' k0 + d2; estimate 'half-life for trt1' -log(2)/k0; estimate 'half-life for trt 2' -log(2)/(k0 + d2); ods output parameterestimates=parms additionalestimates=addl; run; I think the use of the CMPTMODEL statement would enable fitting all of the data, but I will have to learn more about it before I post anything.   SteveDenham   ... View more

@Ksharp 's code has a residual option in one of the random statements. METHOD=QUAD doesn't accept R-side (residual side) effects. I would add week to the CLASS statement and try this:   proc glimmix data=counts method=quad; class sub week ; model count = week / link=log s dist=poisson; random int / subject=sub; random week/ subject=sub type=ar(1) ; /* You may want to consider other structures */ run; I am going to call in @StatDave because I think this question might be addressed better using PROC GEE (and he is much more familiar with PROC GEE than I am), as I believe you are interested in the marginal effects of time, rather than subject-specific effects.   SteveDenham     ... View more

Just some miscellaneous comments - it doesn't make any difference what base log transform is used, the estimate for t1/2 should be the same. This is because the various log values differ only by a multiplicative constant, which factors out.    Next, log transforming the data makes some assumptions regarding the variance (that the variance is proportional to the mean, for instance). Consequently, this could be done on untransformed data using PROC NLIN or PROC NLMIXED. By including an indicator variable and some clever programming steps, you could get an estimate of the difference in t1/2 between treatments using an ESTIMATE statement in NLMIXED. If you are interested in this approach, let us know.   Finally, I had no trouble running your code. I assume that the method is historically what has been done in your field, so I would stay with it.   SteveDenham ... View more

Would a two step method be a possibility? Step 1: Use LAPLACE or QUAD (if you have enough data) to fit the RANDOM effects, and output the variance/covariance parameter estimates to a dataset. This would enable selection of an error structure with the smallest corrected AIC. Step 2.Fit your current model using the pseudolikelihood method and a residual R side effect for the repeated factor. You could use the values obtained in the first step as starting values in a PARMS statement.   NOTE WELL: THIS IS UNTESTED AND THERE IS NO GUARANTEE THAT IT WILL SOLVE THE PROBLEM     Additionally, you should consider that since this is a GLMM with a binary distribution the best approach may be to do this all as a G side analysis.   SteveDenham ... View more

Put enum in a CLASS statement. If the only issue before was PHREG trying to fit enum as continuous, that should address the issue. The issue then becomes whether you get an appropriate analysis when enum is still considered as a stratification variable.   A shout out to @StatDave @JacobSimonsen or @FreelanceReinh for additional help in this area, since the above paragraph exhausts what I know about PHREG.   SteveDenham ... View more

For comparison of multiple groups to a single control for a nonparametric analysis (Kruskal-Wallis), Dunn's test is often recommended. While also influenced by ties (as is any rank based analysis), it is not as affected as Dunnett's test. SAS does not currently offer Dunn's as a multiplicity adjustment, so you would probably have to search for it on the internet.   However, SAS does offer the method of Edwards and Berry (ADJUST=SIMULATE), which is truly distribution-free, and provides superior family-wise error control to any of the other methods available for the LSMEANS or LSMESTIMATE statement.   SteveDenham ... View more

And if you still have non-estimable means, you may have to resort to a "means model" where you fit only the three way interaction in the MODEL statement, and use LSMESTIMATE statements to generate the means and standard errors.   SteveDenham ... View more

That is fairly specialized. Try searching LexJansen.com for papers on the subject, at least as a start.   SteveDenham ... View more