Neuroskeptic has a great post on this topic, listing some of the researcher degrees of freedom in analyzing a hypothetical fMRI experiment:
"Let's assume a very simple fMRI experiment. The task is a facial emotion visual response. Volunteers are shown 30 second blocks of Neutral, Fearful and Happy faces during a standard functional EPI scanning. We also collect a standard structural MRI as required to analyze that data."What are some of the options for analyzing this with MVPA? This is not an exhaustive list by any stretch, just the first few that came to mind.
- Average the volumes to one per block. Which volumes to include in the average (i.e. to account for the hemodynamic lag)?
- Create parameter estimate images (PEIs) (i.e. fit a linear model and do MVPA on the beta weights), one per block. The linear model could be canonical or individualized.
- Average the volumes to one per run. Calculate the averages from the block files or all at once from the raw images.
- Create one PEI for each run.
- Analyze individual volumes (first volume in each block, second volume in each block, etc).
- the "default": linear svm, c=1.
- a linear svm, but fit the c.
- a nonlinear svm (which type?).
- a different classifier (random forest, naive bayes, ....).
- linear discriminants (multiple options)
- on the runs
- on a combination of runs (first two runs out, next two out, etc)
- ignoring the runs (ten-fold, leave-three-examples-out, etc)
- on the subjects (leave-one-subject-out)
- on the runs, but including multiple subjects
- ROI (anatomical, functional, hybrid)
- searchlight (which radius? which shape? how to combine across subjects?)
- resize the voxels?
- scale (normalize) the data? (across voxels within an example, across examples?). Center, normalize the variance, take out linear trends, take out nonlinear trends?
Simmons JP, Nelson LD, & Simonsohn U (2011). False-positive psychology: undisclosed flexibility in data collection and analysis allows presenting anything as significant. Psychological science, 22 (11), 1359-66 PMID: 22006061