Back to PLS HelpI'm Online
Posted on 06/04/12 08:21:25
Number of posts: 1
Hi,
I'm using PLS to analyse an fMRI Blocked design with six conditions, and four groups of subjects. In SPM I ran it as a 2x2 between subjects ANOVA design, and I'm wanting to redo that in PLS, to see if using the multivariate approach gives me more power.
I'm trying to figure out how to set up the contrasts file to get the equivalent of the ANOVA. I see that Rajah & McIntosh (2008) did something similar, and I think I've figured out how to follow that. However I would love some reassurance from a PLS guru that I'm doing it right!!
My design has four groups of subjects: Control, MD (maths disabilities), RD (reading disabilities), and MDRD (both). The two between subject factors are MD (present, absent) and RD (present, absent).
I've got all the session datamats made, and no problem entering groups and conditions into PLS.
I have a contrast matrix of LVs specified, and this same matrix repeats for each of the four groups:
| LV1 | LV2 | LV3 | LV4 | LV5 | LV6 | |
| Words | 1 | 0 | 0 | 0 | 0 | 0 |
| Lines | 0 | 1 | 0 | 0 | 0 | 0 |
| Number | 0 | 0 | 1 | 0 | 0 | 0 |
| Colours | 0 | 0 | 0 | 1 | 0 | 0 |
| Mult | 0 | 0 | 0 | 0 | 1 | 0 |
| Digits | 0 | 0 | 0 | 0 | 0 | 1 |
| Rest | -1 | -1 | -1 | -1 | -1 | -1 |
And I know what the ANOVA weights should be for my four groups to give two main effects and an interaction:
| Effects | |||
| Group | MainMD | MainRD | MDxRD |
| Control | -1 | -1 | -1 |
| MD | 1 | -1 | 1 |
| RD | -1 | 1 | 1 |
| MDRD | 1 | 1 | -1 |
So to run the ANOVA, I've made three separate contrast matrix files, one for each effect, and in each one all I did was multiply the LV x condition submatrix for each group by the corresponding multiplier according to the contrast (so e.g. for the main effect of MD, I multipled the control section by -1, the MD section by 1, the RD section by -1 and the MDRD section by 1).
Then I just ran all three analyses - et voila! It seems so easy... but would love to know I haven't done it the wrong way!
Thanks,
Anna Wilson, University of Canterbury, New Zealand
Replies:
Posted on 06/07/12 08:07:21
Number of posts: 394
I am not sure I completely follow you, but I think you are close. If I understand, you run the main effects contrasts (the LVs) for all groups without reweighting by group to get the main effects and then another set of contrasts multiplying your LV matrix by the Effect matrix to give you the groupXtask interactions?
Let me provide this example and see if it works
Say we have a design with two groups and two conditions, so we code group as:
G1: 1
G2: -1
and condition as:
C1: 1
C2: -1
so a fully parameterized design matrix would be:
G1C1: 1 1 1
G1C2: 1 -1 -1
G2C1: -1 1 -1
G2C2: -1 -1 1
where column one is group main effect, column 2 is the condition main effect and column 3 is the groupXcondition interaction
is this similar to what you did?
BTW, for the non-rotated PLS (i.e., contrast-based), the contrasts can be entered all at once in a single analysis or separately, since there is no adjustment for correlations between contrasts in the analysis.
Login to reply to this topic.
