Which residuals should be evaluated for normality: Studentized, Pearson, or raw residuals?

[Dicarlis]

Hello everyone,

I have been analyzing performance data from livestock. When dealing with such data—even in blocked experimental designs—it can be challenging to determine whether an extreme laboratory result (e.g., blood urea) is a genuine analytical error (a true outlier) or simply a biological variation that causes a specific animal to stand out despite local control. This variation could even represent a direct or indirect effect of the applied treatments.

Given this scenario, would the most appropriate approach for assessing residual normality be to use Studentized residuals rather than Pearson or raw residuals? And why?

[sbxkoenk]

What model do you intend to build?

• ANOVA
• mixed model
• a  nonparametric alternative
• ... ??

Which SAS procedure did you use?

Is it a crossover design?

BR, Koen

[Dicarlis]

Hello Koen
How're you doing?
Thanks for you awnser,
Usually I have used the mixed models procedure, in a randomized block design,
Also, in my research team, we have done a batch of trails using Latin sqare,
My best regards,

[sbxkoenk]

I understand you want to check the typical normality assumptions.

I think PROC MIXED (to fit models with fixed and random effects) has "conditional residuals".

Conditional residuals are the residuals accounting for the random effects, so it helps you to evaluate the normal distribution given the random effects.

BR, Koen

[Dicarlis]

I understand. So ..., What is the correct way to assess the normality of the residuals: using the Pearson transformation, the Student's transformation, or the residuals without transformations?

Att.

[sbxkoenk]

Statisticians look at all of the residual plots, but tend to focus more on the studentized, Pearson, and conditional plots (especially if there are random effects in the model).  

BR,

Koen