search

0todd0000 / spm1d

One-Dimensional Statistical Parametric Mapping in Python
GNU General Public License v3.0
61 stars 21 forks source link

Normality check #65

Closed zof1985 closed 7 years ago

zof1985 commented 7 years ago

Hi Todd,

A reviewer asked me about the check of the normality of my data for anova3rm analysis. Briefly, he would like to see a single p value rather than a p value for each point of my interpolated (101 points) timeseries. Therefore I would know your opinion about using some multivariate normality tests like the Henze-Zirkler’s test or similar (ref.).

Many thanks, Luca.

0todd0000 commented 7 years ago

Hi Luca,

Mark, Jos and I are coincidentally preparing a paper right now precisely on that issue, and we've got three different multivariate normality tests implemented including the Henze-Zirkler test. All three will be available in spm1d in the next few months. However, one problem is that there is really no good test of multivariate normality; the tests work generally well on 0D data but their small imperfections become magnified when extending the tests to 1D.

So instead of explicit normality testing I'd recommend running non-parametric analyses, which are available in the current version of spm1d. Insofar as the parametric and non-parametric results qualitatively agree one can conclude that the parametric approach's assumption of normality is a reasonable one. In this case explicit normality testing is redundant. If, however, the results qualitatively disagree then the non-parametric results should be used because they are valid for all distributions, normal or otherwise.

Todd

zof1985 commented 7 years ago

Hi Todd,

As always. Your answers are very helpful.

Many thanks, Luca