topic Re: Normality for Proc mixed in Statistical Procedures

Normality for Proc mixed

roshanm — Mon, 02 Feb 2015 19:24:43 GMT

I have a non-normal data. Do I need to make it normal (by appropriate transformation) prior to "Proc mixed" with "random" statement?

Re: Normality for Proc mixed

SteveDenham — Mon, 02 Feb 2015 20:58:00 GMT

If you know the distribution from which the data is generated, then PROC GLIMMIX will suit your needs. Unusual links (transformations) can be handled programmatically.

If you are determined to use PROC MIXED, then the key is making the RESIDUALS normal. For instance, a bimodal distribution due to sex can never be made "normal" and the analysis does not require that the response variable be normal--only that the residuals be normally and identically distributed.

Steve Denham

Re: Normality for Proc mixed

Ksharp — Tue, 03 Feb 2015 07:26:15 GMT

Fisher's Z transform could help you to normalize it .

z=1/2* log( (1+x)/(1-x) )

Re: Normality for Proc mixed

Babloo — Tue, 03 Feb 2015 12:28:01 GMT

Any possibilities to provide me an example of Fisher's Z transform?

Re: Normality for Proc mixed

Babloo — Tue, 03 Feb 2015 12:32:40 GMT

Seems Fisher's Z transform is part of correlation. Will it work for single variable?

Re: Normality for Proc mixed

Ksharp — Tue, 03 Feb 2015 15:32:56 GMT

Yeah. I think it should be worked on a single variable. It is just a transform function . But I don't how good it is, You can test it.

Of course. Don't forget try to use LOG() to make the variable normal too.

Re: Normality for Proc mixed

SteveDenham — Wed, 04 Feb 2015 13:55:15 GMT

Fisher's Z really only transforms values that are bounded below by -1 and above by 1 into normal deviates.

Steve Denham