Print

Print
Sterling Johnson wrote :
[log in to unmask]" type="cite"> 
Dear SPMers:
I need some anova help,
  We have a between groups one-way anova with one con image per subject taken from a two-condition fMRI paradigm. The hypothesis we wish to test is that that there will be a linear trend in group mean signal change in a specific region:  g1<g2<g3<g4.
Questions:
1. In basic models should i be using the straight ANOVA option, or ANOVA (with constant) to test this trend effect?
2. the contrast -3 -1 1 3 should tell me if there is a linear trend, but using equally spaced positive integers (1 3 5 7) in the contrast yields a better result when using the ANOVA no constant model, but is this contrast valid?
Thanks
Sterling
Well, it seems to me a little bit complicated.

The use of the constant term in ANOVAs reflects the fact that you want to test that the mean if each condition differ. It means that it is mainly use to contrast conditions (and contrasts sum needs to equal 0). When working without constant term, we usually test that one or more of the conditions are different from 0 (contrasts like 1000, 0100 ...). ANOVAs without cst are also use when working on differences (if I'm wrong please correct me! :-) ). In your case, it seems that you set, at the first level, a contrast C1>C2 and then use this result at the second level to compare groups (that why I suppose your ANOVA without constant is better).

For the second part of the question, I'm a little bit lost. To be sure that g1<g2, I think one need to test it explicitly with a contrast 1 -1 0 0 and not only with a contrast like -3 -1 1 3 (maybe someone would comment that?). In addition, I don't understand why you are using contrasts like -3 -1 1 3? Maybe there is here, a theoretical reason? Can't you use -2 1 1 2? The contrast 1 3 5 7 seems to be invalid (at least for me) because I think it tests for differences from 0 and not between groups.

I hope it helps you, (and that I'm right ;-) )

Best,
Cyril