reconstruction tests

[jeremymanning]

For each pyFR subject, we have one correlation matrix (at the observed locations for that subject). For each RAM subject, we have up to two correlation matrices (at the observed locations for that subject)-- (up to) one for RAMFR and (up to) one for CatFR, depending on which tasks the subject ran in.

Within vs. across subject:

• Include data from whatever tasks are being included in the given iteration of the analysis. E.g. for pyFR everything is from the "same" task, so there's no task info to take into account. For RAM data, a given subject's data (or part of it) may or may not be included.
• For within subject analyses: take the correlation matrix (averaged across whatever sessions we have, and then averaged across whatever tasks are being included). Remove the one electrode we're trying to predict. Blur the subject correlation matrix out to that missing location. Now reconstruct the "missing" location, treating the remaining locations as "observed."
• For across subject analyses: use the union of all locations (across all subjects/tasks) as the "full set of locations." Blur every subject's correlation matrix out to the full set of locations. For each subject, compute an average (blurred out) correlation matrix by averaging over all of the recording sessions (within each task) and then averaging across tasks. To do the reconstructions:
  • Build a model by combining correlation matrix over all other subjects
  • Treat all locations (except the to-be-reconstructed location) as observed, and use the average model (from other subjects) to reconstruct the missing data

Within vs. across vs. combined tasks:

• Within task: only include data from the same task. Some subjects may not have done that task, in which case they're excluded from whatever conditions they would otherwise have been included in.
• Across task: only include data from the other task. Some subjects may not have done the other task, in which case they're excluded from whatever conditions they would otherwise have been included in.
• Combined: include the average model from every subject/task, even if they've only run in one of them.
• Note: for all of these conditions, still exclude the to-be-reconstructed subject if we're in the "across subjects" condition, and include only the to-be-reconstructed subject (if they've run in the appropriate/necessary tasks) if we're in the "within subjects" condition"

For each of these combinations of conditions (pyFR: within/across subjects -- 2 conds total; RAM: within/across subjects * within/across/combined tasks -- 6 conds total) report (for every electrode) the average correlation between the true vs. reconstructed voltage trace. (The average is taken across recording sessions.)