Violation to Normality
If there's an outlier, ANOVA will not be robust to normality, so what do you do?
posted by Karlin @ 3:34 PM
Previous Posts
- Robustness of Normality
- re. transformation
- Latin Square Design
- Write-up information
- Two or more constructs and assumptions paragraph?
- Reporting of Transformations and Categorization
- re. qq plot and histogram
- Post Hocs
- Parental Bonding and Parent Child Relationship var...
- re. independence of observations
3 Comments:
At 3:53 PM,
Karlin said…
(and outliers in homogeneity is taken care of because levene's test shows to be non-significant anyway)
At 8:24 PM,
Unknown said…
Sorry for not getting to this post sooner - I thought it had been answered, but I didn't realize that the comment was another comment from you.
ANOVA actually IS robust to MINOR violations of normality. When we say that it is robust if you are using 2-tailed tests, cell sizes are equal, df error > 20, and there are no outliers - this does not mean that there cannot be a single outlier. It means that with 2-tailed tests, df error > 20, RELATIVELY equal cell sizes, and FEW EXTREME outliers, you are USUALLY fine.
I would ALWAYS check normality with a histogram or QQ plot. These will show you whether normality is violated or not. If it is (or looks like it might be), you can transform your DV. Start with a square root transformation and don't forget to reverse score the DV first if it is negatively skewed. (You don't need to reverse score it before transforming it if it is positively skeweed.)
At 8:56 PM,
Mari said…
Alternatively, if the data are normally distributed, with just a few outliers, you can (a) drop those participants or (b) recode just those outliers. (See your regression notes for options, but consider recoding to X SD units above mean--keeping relative positioning of the variable as most extreme, but pulling in the score.)
As always, whatever you do, report it.
Post a Comment
<< Home