Dear SPMers,
@Donald: thanks for the quick & helpful suggestions! However, I’m still a bit confused how to set up my 2nd level analysis. The original questions with answers & a full description of the event-related fmri design can be found below. Briefly, I ran a memory task with 3 levels of memory load, trials were separated by a fixation cross. The aim of the study is to compare two groups of participants on the 2nd level.
1. On advice of Donald I will revise my 1st level analysis excluding the explicit regressor for the fixation cross (as a baseline condition). Now the first level design includes the three regressors:
Memory-load 1
Memory-load 2
Memory-load 3
2. Which images should be passed to the second level analysis?
a) Beta-images for load 1 / 2 /3?
b) Contrast-images? However, given the 1st level design does not explicitly model the fixation cross (baseline) should I use the t-
contrasts:
1 0 0 -> load 1
0 1 0 -> load 2
0 0 1 –> load 3 ?
c) I think in this case betas & cons are model the same activity – or do they differ because of the implicit baseline?
3. Which analysis should be used to reveal the most robust analysis of between-group & interaction effects (load x group) on the 2nd level?
a) A full factorial design with the factors group (2levels) & load (3levels)
This design would reveal results for the main effect of task & the interaction effect
To asses the main effect of group I would put the 1st level contrast for the mean effect of load (t contrast: 1/3 load 1 + 1/3 load 2 +
1/3 load 3) in a 2-sample t-test.
b) An analysis with extracted beta-values. Therefore I would compute a 1-sample t-test for the whole sample using the 1st level
contrast for the mean effect of task/load (t contrast: 1/3 load 1 + 1/3 load 2 + 1/3 load 3). Significant clusters from this analysis
would represent the memory-networks. In a next step I would extract individual beta-values for these clusters using individual 1st
level beta-images (wm load 1, wm load 2, wm load3). Individual data from these ROIs would be passed forward to a statistics
program to analyse main effects & interaction effects.
Thanks again for the help & best regards,
Ben
Original post with answers:
I’m analyzing an event-related working memory task with three levels of wm-load and two groups (placebo vs. verum), the conditions are:
1. Baseline (fixation-cross)
2. wm-load 1
3. wm-load 2
4. wm-load 3
I would only model your conditions, then the betas are implicitly with relation to your fixation cross. Otherwise your beta are relative to an unknown value -- not to mention that you probably begin your study with a fixation cross and have a BOLD response that is predicted to increase right away even though the cross was there before some of the data.
To compare groups on the second level I used a full-factorial design (factors: group (2-levels) x load (3-levels).
My questions:
1. I ran the analysis with con-images (wm-load 1-Baseline / wm-load2-Baseline / wm-load3-Baseline) and beta-images (from load1 / load2 / load3). The analyses reveal different results. Which analysis reveals the most reliable results?
Subtractions.
2. Both analyses reveal a main effect of group, however no group x load interaction. How can I be sure, that the effects are working-memory specific? Would it be valid the mask the contrast of the main group effect with the main effect of task?
In a mixed-design the main effect of group is invalid (see below for correct method). The main effect of task and interaction are valid.
3. For a more sensitive analysis I’m planning to compute the contrast for:
1/3 wm-load 1 + 1/3 wm-load2 + 1/3 wm-load3 vs baseline for the whole sample. Than compute a 1-sample t-test with these contrast images. Significant clusters from this analysis are thought to represent the working memory networks. In a next step individual beta values will be extracted using these significant clusters as ROIs. Does this analysis lead to valid results?
This contrast should also be used to compute the main effect of group in a two-sample t-test or a one-way ANOVA.
You want to extract the contrasts for each part, not the betas. Also, you might want to revise your first-level models to exclude the fixation condition.
See www.martinos.org/~mclaren/ftp/Utilities_DGM/peak_nii for a way to extract from all clusters and select which images to use for the extraction. A previous email provides more detail on its usage.
Thanks in advance & kind regards,
Ben
|