quote:
When running task PLS analyses, I have always understood the condition LV weights to be a linear contrast between the conditions. I recently ran a behavioural PLS analysis on some EEG data, and the resulting condition weights were ALL positive. For e.g., the weights across 4 conditions in LV1 were: 0.5329 0.5580 0.5494 0.3207. Obviously, these cannot describe a linear contrast across the conditions, so I'm confused about how to interpret them. They are not direct multiples of the condition/task correlations either, though they seem to follow the pattern of the correlations reasonably closely (the correlations, in this case were: 0.7913 0.7437 0.7278 0.4633).
Does it make sense to obtain all positive condition weights in a behavioural PLS, or is this an indication that I've run the analysis incorrectly?
Hi Jesse,
Your results are correct for behavior PLS. In behavior PLS, unlike task PLS, the grand mean is left in as a possible dimension, so you can get a latent variable where all the weights are in the same direction. This is essentially the mean and identifies where brain-behavior correlations are similar across tasks. The other LVs will be variations around the mean. In task PLS, since we mean centre the data, the LV that would be the grand mean is zeroed out.
Randy