Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

ANOVAs of individual MaxDiff scores between different subgroups?

I want to compare means of individual MaxDiff scores between subgroups of respondents for particular items - is it OK to use an ANOVA on the rescaled scores? Does anyone know of any useful documentation on analysing heterogenity of individual scores within an item of interest?

Thank you.
asked Aug 30, 2012 by anonymous

1 Answer

0 votes
I'm pretty sure that you can. Anova is robust enough for rescaled MD scores, even thou:
1) rescaled MD scores have upper and lower bounds so are by definition non-normal
2) are often extremely skewed
You should be more cautious when you try to check differences between MD variables, because they are not independent for rescaled scores.

If you are not sure if anova is aplicable due to extreme skewness or kurtosis, you can always try Kruskal-Wallis.
answered Aug 30, 2012 by lkomenda Bronze (2,810 points)
Keep in mind it's not the raw scores that need to be normal. It's the residuals.  So u should try an ANOVA first and check the residuals are approximately normal.  Keeping in mnd Linear models are quite robust to this assumption.

It's more of a problem  if  the variances differ between groups, which may happen if some groups are very close to 0 or 100, while others are in the middle.
Of course resids should be normal, not the whole distribution of rescaled scores.
In the subject of homogenity of variances you should especially check for correlation between means and variances which can be a case with rescaled md scores.
But you could always go for kruskal-wallis with almost the same power as anova.
Thank you both for your help. Although variances appear to be equal between groups the residual distributions are far from normal - I'll try Kruskal Wallis instead. Your comment about correlation between means and variances is interesting - I'd already noticed a strong positive correlation between the mean score of items and their standard deviation - why is this often the case with rescaled scores?

Thanks again,