topic Re: distribution option in Proc Glimmix in Statistical Procedures

distribution option in Proc Glimmix

nlpurumi — Sat, 17 Feb 2018 19:27:42 GMT

Hi. I chose to run proc glimmix instead of proc mixed because when I test shapiro wilk's test in proc mixed, the normality assumption was violated even after tranformation.

However, when I choose distribution of gamma instead of default distribution (gaussian), the model does not converge.

Is there any solution for this?

Should I still seek a model to converge when the distribution is gamma not gaussian?

Or can I go with gaussian distribution even the normality assumption was violated?

Thanks for your advice in advance.

Re: distribution option in Proc Glimmix

Ksharp — Sun, 18 Feb 2018 01:57:50 GMT

You could try different initial value. Check PARAM statement.

Re: distribution option in Proc Glimmix

PGStats — Sun, 18 Feb 2018 06:53:10 GMT

Given enough observations, the normality checking tests will always give significant results. First, make sure you don't have an outlier problem. Then, if your residuals distribution has a single mode and not too much skewness, go with the gaussian analysis.

If you really have to model your data as gamma, then post your code for more help.

Re: distribution option in Proc Glimmix

nlpurumi — Sun, 18 Feb 2018 13:10:06 GMT

So do you mean it is not a problem to do gaussian distribution even when the purpose of running proc glimmix is because I thought my data is non-normally distributed?

Thanks.

Re: distribution option in Proc Glimmix

Rick_SAS — Wed, 21 Feb 2018 02:08:53 GMT

> So do you mean it is not a problem to do gaussian distribution even when the purpose of running proc glimmix is because I thought my data is non-normally distributed?

It is not a problem, but I think you are misquoting the purpose of glimmix.

The assumptions for general linear regression models are that the residuals (not the data) are normally (and independently) distributed with constant variance. Generalized linear models (eg, PROC GENMOD) support alternative residual structures such as Poisson regression where the variance is assumed to vary, as well as non-normal error distributions. PROC GLIMMIX supports generalized linear models with random effects.