GLIMMIX for continuous response whith non-normal residuals?

[marcel]

I am analyzing a continuous response and fixed and random factors. I run a Proc MIXED analysis (code below = "Main Analysis"). The diagnostic indicates that the residuals are non-normal ("Residual Diagnostic", below). The best transformation to normalize the residuals is the Johnson transformation. I read opinions from many statisticians on normalizing the residuals through transformation, and many are in favor but many others are against using transformations.

Is it advisable to use GLIMMIX in this situation?

Below are my original codes:

MAIN ANALYSIS

proc mixed data = WORK.FED28 order=internal covtest
class Status IDs InSta OrIs MotherSt
model Fed28 = InSta/ solution ddfm=sat residual outpred=pdat128Fec28;
random OrIs OrIs*InfSta MotherSt(InSta OrIs);
format InSta InfStafmt.;
ods output solutionR = eblupsdatf28 influence = inff28;
run;

RESIDUAL DIAGNOSTIC

proc univariate data=pdat128Fec28 normal;
class Status ;
var studentresid;
qqplot / normal(mu=est sigma=est) nrow=1;
format Status Statusfmt.;
run;

 

Thank you.

Marcel

[StatsMan]

The tests for Normality produced by PROC UNIVARIATE are very strict and will detect very small departures from normality.  Most statisticians rely on looking at a plot of the residuals.  If the plot of the residuals is not too far out of whack, then accept the assumption of the normality of the residuals and move on.  

In your case, the residuals are a little peaked according to the QQ plot.  This assessment comes down to an opinion rather than a strict statistical test, but again most statisticians just look at this plot and make a decision from that look.

You asked in the title about using GLIMMIX if the response has non-normal residuals.  You do have different distributions you can assume for the response, but nothing that will adjust the residuals if you want to use a normal distribution. 

[marcel]

Sir StatsMan,

It looks that the residuals are not really "out of whack". It seems reasonable to decided that the residuals do not violate the assumption of normality. Thank you for your advise.

Regards,

Marcel