Hi, I'm working on a regression analysis, but the results make me wonder whether I have a correct understanding of how it's working.
First, I am trying to see how well brain activity in various regions predicts a variable, A. It seems the procedure is to do the reverse (see how well A predicts brain activity in various regions) and use the fact that a correlation between A and Y is the same as the correlation between Y and A.
Next, I have at least 1 other variable, B which theoretically should also account for some of the variance. Also, B and A may theoretically interact, so I have a 3rd variable A*B, the product of A and B, which serves as an interaction term.
In any multiple regression analysis I have done on behavioural data using the Enter method (which seems to be the only available option), variance is ascribed to one of the predictor variables. So if 2 correlated variables can account for the variance to various degrees, the variance gets assigned to the variables like slices of pie until the whole pie has been doled out.
Here's the problem: when I look at the activation map for A, B and A*B (which is clearly correlated with A and B), the maps are overlapping. In fact, there is very little difference between the map for A and the map for A*B. I would have thought that if A*B was a better predictor of activity than A (or vice versa) voxels significant in one map should not appear for the other predictor, and not for both.
Can anyone set me straight? There may be more than one place where my understanding has gone wrong.
Many thanks,
Chris
|