Dear Colleagues

How do you approach the baseline levels in a design with two measurements? I have two randomly assigned groups, drug and placebo (subject effects) and two resting states pre and post intervention (conditions). If I am interested in the interaction “drug>placebo, post>pre” [1 -1] [1 -1], is there any way to set the contrasts to consider the significant interactions, which are not driven by baseline differences?

All your comments are highly appreciated,
Lucas

Hi, Lucas,
 
You can accomplish what you want by using an exclusion mask from a simple baseline comparison of groups in the SPM GUI.  Run a simple CONN baseline comparison of groups (i.e., [1 -1][1 0]) and set your threshold a bit more liberal (i.e., p<0.005) and then save any resulting blobs as a mask image.  Then, run your desired group x time comparison (i.e., [1 -1][1 -1]) in CONN and threshold appropriately.  In the results window you can click on SPM view and then when the SPM GUI results pop-up it will ask you if you wish to mask the results.  It is here where you would select your ROI/mask from the [1 -1][1 0] group x baseline contrast and select exclusion.  This will then show you which of your results from the group x time contrast were not a result of reasonably significant baseline differences.
 
Hope this helps,
Jeff

Hi Jeff,

Thank you kindly for the great tip! Can something similar be done for ROI-to-ROI or even voxel-to-voxel (MVPA)? Being rather beginner, I may be wrong, but I also believe to have seen a publication combining vectors like -1 -1 3 -1 or similar.

Best,
Lucas

Hi Lucas/Jeff,

While I totally agree with Jeff on how you would do these analyses, I would like to double-check why exactly you would want to perform that masking. To give some context to this, when looking at a Group-by-Condition interaction (e.g. [1 -1] contrast across drug/placebo groups, and [1 -1] contrast across pre/post scans) those differences are always orthogonal to the corresponding main group and condition effects (i.e. differences between the drug/placebo groups during both the pre-intervention and post-intervention scans, or differences between the pre/post scans in both the drug and placebo groups), so an interaction effect cannot arise solely from a difference between the drug/placebo groups at baseline (those differences need either to not appear in the post-intervention scans, or need to appear in different magnitude in the post-intervention scans, for an interaction to be present). The conjunction analysis that Jeff describes allows you to exclude any regions where the connectivity might be different between your two groups in the pre-intervention scans, but just note that those differences alone are not "causing" or "explaining" the interaction effects, since those interactions indicate that those between-group differences are being modulated by the intervention, whether they were present at baseline or not. This raises the question of why exactly would you want to mask-out those regions that show pre-treatment differences in connectivity between your drug/placebo groups when analyzing the somewhat-orthogonal question of which regions show differences in the post- vs. pre-treatment changes in connectivity between your drug/placebo groups. 

As a only-partially-related comment (feel free to skip/disregard this paragraph), it is true that in pre/post drug/placebo designs where the placebo/drug group assignment is random there are no reasons to expect a pre-treatment difference between the groups beyond sampling/random effects. In this case it is often argued that an ANCOVA analysis, where pre-treatment measures are used as a covariate when analyzing between-group differences in post-treatment measures, might be preferable over a repeated-measures/MANOVA analysis (both analyses will typically lead to similar results and interpretation, but the power in the former analysis is expected to be larger than in the latter, particularly in the case of low correlations between pre- and post- treatment measures; see for example Rausch et al. 2003, Analytic Methods for Questions Pertaining to a Randomized Pretest, Posttest, Follow-Up Design). If, on the other hand, the group assignment is not random then the two analyses can potentially lead to different results, and the differences arise from the different interpretation of these analyses: repeated-measures ANOVA/MANOVA allows you to look at the intervention "change" effects (differences in connectivity between post- and pre-intervention scans) and determine whether those changes are different in the control vs. placebo groups, while ANCOVA, on the other hand, allows you to look at the post-intervention connectivity values, and determine whether those are different in the control vs. placebo groups when comparing subjects at the same level of pre-intervention connectivity. In this case (non-random assignment) repeated-measures/MANOVA analyses are often recommended instead, though this may depend on the specifics of the research questions that one is interested in, and the source of pre-treatment differences between the groups. In practice, in fMRI analyses I have not really seen ANCOVA models used much in this context (perhaps mostly due to this sort of models requiring voxelwise regressors which are not that widely implemented/known?; just for reference using FSL's FEAT would allow you to do this, and we are working on an equivalent implementation in CONN) so most people seem to be using the (perhaps less powerful in some context but perhaps also more generally valid) repeated-measures/MANOVA approach instead irrespective of the potential source of pre-treatment differences. 

Hope this helps
Alfonso

Originally posted by Jeff Browndyke:

Thanks for the methodological correction and guidance, Alfonso!  As usual...a wealth of information and mentoring beyond what (I suspect) many of us may get from our respective institutions and in-house colleagues!  I suspect that Lucas may be in the same boat as me in having to explain to non-imaging MD-types why pre-treatment functional connectivity differences may not be fatal confounds in any analysis results.

Warm regards,
Jeff

Hi Alfonso
Hi Jeff

I can only join Jeff's thanks for your invaluable guidance in the world of connectivity analysis! It is highly appreciated. After reading your explanation and the reference paper, I am convinced by the value of ANCOVA for this type of analysis. I am unsure whether I understood you right about the practical implementation. Is there no easy way to do it in Conn? If not, would you simply stick to the [1,-1] [1,-1] contrast? Strange enough, at least visually, almost all the found interactions seem to have very unreasonably different baselines in my sample.

Warm regards,
Lucas

Hi Alfonso,

I was reading through the forum and came across your post below describing the use of a one-way ANCOVA where pre-treatment measures are used as a covariate to account for baseline differences (instead of a 2-way mixed ANOVA). Very interesting! Could you please let me know whether it is possible to implement this in CONN for seed-to-voxel and MVPA analyses, and how easy it is to implement? Also, do you think this approach is now commonly known/understood/used in the imaging field (I have never seen it in any paper)?

Many thanks

Hi Alfonso, 

Thank you for the very detailed information in your response to the original post. Like msc_22, I am wondering if it is now possible to analyze post-intervention connectivity while controlling for pre-intervention connectivity, but in an ROI-ROI analysis?

For context, I am working on a double-blind RCT. One of my analyses looks at the effect of a vitamin intervention vs placebo. Pre-intervention, my control group demonstrates significantly greater between-network connectivity. Post-intervention, there are no longer any significant between group differences. 

Regards, 

Nick


Originally posted by msc_22: