The magnitude of AIC is not meaningful, just the difference in AIC between different models. 0 has no particular meaning. You want the smallest AIC. -75.8 is much smaller than 8.8 ... View more

Looks like a question for Rick Wicklin ... View more

SAS procedures have built in formats for the constants and variables that are printed. These carry over into the ods output files. But you can override these in PROC PRINT. For instance, if you saved an ods output file called test, and the p value was called p, then use: proc print data=test; format p 15.12; run; This gives you p to 12 decimal places (as an example). You can experiment with many other formats. Or to limit output just to p: proc print data=test; format p 15.12; var p; run; ... View more

No conflict at all, as indicated by Steve.Each of the first three time parameters is being tested versus 0 (which means, in this, case, with an over-parameterized model: Is the expected value for each time different from the last time?). The type 3 test consists of three contrasts simultaneously. Perhaps the mean for one time is different from a different time, or the mean of the first two times is different from the mean of the second two times, and so on and so forth. ... View more

Typically, you base distribution choice on properties of the response variable. Conditional residuals (histograms) can give hints, but keep the variable in mind. Time is a good example: it is bounded by zero, with no upper limit, and it is typically right skewed. This is a property of the gamma distribution. By the way, you don't choose a link function to stabilize variances (with non-normal distributions), you choose a link to provide a linear scale for the predictors. The unequal variances are automatically handled by the chosen (non-normal) distribution. For instance, with gamma, var(Y) = a.mu^2, where mu is the mean. So, one is automatically taking care of the fact that the variance is increasing quadratically with the mean. Once you move over into generalized linear models or generalized linear mixed models, lots of things change. ... View more

For your one question, the link is definitely not the same as a transformation. The distinction can be subtle, but important. The link is a function of the expected value (i.e., mean), not a function of the response variable; in contrast, a transformation is a function of the response variable (not the mean). Say, z = 1/time (the transformation you mention). We can also write f(time) = 1/time. Then in GLIMMIX (or MIXED), one is modeling (without random effects) E(z) = E(f(time)) = E(1/time) = X.beta Here, E(.) is the expectation. You are dealing with means of the transformation. Back-transforming does NOT give you the mean of time, but a shifted value away from the mean. A link can be written generically as g(mu), which in your case is g(E(time)), a function of the expected value. So, with link=inverse, you are modeling g(E(time)) = 1/E(time) = 1/mu =  X.beta Note that the expectation in the first case is of the function of time, but with the link, one has a function of the expectation. The back transformation of the link, known as the "inverse link" (a term unrelated to the use of 1/mu as the link), gives the actual expected value of time (your response variable). In general, use of link functions is preferred because one is working with means of the original response variable. With that said, inverses are tricky for modeling purposes, whether as data transformations or links. These can be especially problematical when the response varies greatly, with some points very close to 0. This is not the the situation with your data, apparently. There are lots of reasons why one can have problems with convergence. Can't tell based on your post. Maybe the algorithm is getting close or is stuck at a saddle point in the log-likelihood response surface. You could try some other optimization methods: NLOPTIONS TECH = ; You could try QUANEW or TRUREG or NRRIDG  for technique; there are others listed in the user's guide. If you are willing to move away from normal distributions, you have many possibilities. You mentioned that you have unequal variances. This is an essential property of variables with gamma or exponential distributions, or many other distributions. You could model the data with a gamma distribution, with time as your response variable (not 1/time). In this case the default link is log. "Inverse link" gives you expected time values. ... View more

Looks like you will want to use PROC GENMOD. You probably should read a few documents about the procedure, working through the examples in the documentation, before you are ready for your particular problem. We can help better after you try something (with your example code). Here are some good resources: http://www2.sas.com/proceedings/sugi26/p264-26.pdf SAS/STAT(R) 13.1 User's Guide https://support.sas.com/rnd/app/stat/papers/logistic.pdf Generalized Linear Model in SAS: PROC GENMOD There are many more. You can try the introductory example in the GENMOD User's Guide. ... View more

You can have this problem because of several issues, even with a small number of parameters. But the most likely culprit is the large number of parameters in your model (as suggested by Steve). Remember, if there are five levels of a factor, you can only have four parameters (when you have an intercept). Moreover, with many factors and an observational study (?), your remaining dummy variables may be linearly related (even when you drop the last level). ... View more

And a new reply to a different post reminded me that you can use PROC GLMMOD. With this PROC, you store the generated data set with the dummy variables and then use this in NLMIXED. You still need to write out the model with all the parameters. ... View more

This is a big and complex topic (doubly repeated measures). There are so many ways of tackling this, including with the approaches already listed. It is an area of active work by me, but I don't have any simple suggestions at this time. It would take too much writing. ... View more

You have to create dummy variables, either in a data step or within the PROC. This can be tedious. For instance, if WORK has five categories, you need four dummy variables. xw1=0; xw2=0; xw3=0; xw4=0; if (work eq 1) the xw1=1; if (work eq 2) then xw2=1; if (work eq 3) then xw3=1; if (work eq 4) then xw4=1; One then needs a parameter for each of these dummy covariates. ... bw1*xw1 + bw2*xw2 + bw3*xw3 + bw4*xw4 + ........ ; So, with so many categorical variables, you might have a LOT of parameters. This can get tricky to interpret. ... View more