search

olilan / RCIC_matlab

Eclipse Public License 1.0
2 stars 0 forks source link

Averaging vs. linear model #12

Open rdotsch opened 11 years ago

rdotsch commented 11 years ago

Another interesting analysis alternative to the averaging procedure we use now is to use a linear model, where a binary (or continuous) response is predicted by the contrast values, and then taking the regression weights as the input for the contrast parameters to calculate the resulting CI.

This would take into account contrasts that in the stimulus set are covarying, instead of treating each contrast value as an independent correlation with the response criterion. Not sure how different the CIs would look, but I know that this can matter somewhat when doing reverse correlation using the FaceGen face space.

olilan commented 11 years ago

That is a cool idea. Without a really informed background: Would that lead to multiple-comparison problems? Is that maybe a step that should be dealt with by using functions from the Chauvin toolbox? If I remember correctly, they deal with the issue that multiple comparisons would be extremely conservative, but that one can use the fact of autocorrelations in the data. If one pixel is correlated with response, the adjacent one probably is also, so locally correlated. Using the Gaussian Random Field logic one can identify areas of response-related pixels without needing to correct for many independent tests.

As I said, only a first hunch...

rdotsch commented 11 years ago

The Chauvin toolbox indeed deals with the multiple comparisons problem with respect to finding significant pixels. However, the linear model I suggest is at the contrast level and besides would not be about significance testing, but just estimating the classification noise.