quote:
Dear List,
I have been recently reading a lot about bootstrap test and come across a question comparing conventional bootstrap test with its counterpart implemented in PLS.Conventionally, people use bootstrap in order to get a proper distribution of the voxel saliencies and hence estimate the CIs. However, in the PLS approach, the bootstrap seems to be used as a method of eastimating the standard error of a normal distribution instead of using a theory to derive a formula for it.What is the reason to do so? Could it be attributed to the computationally expensive way of calculating the CI for voxel saliencies if the normality was not considered?
/Alireza
The reason for standard error estimation for the voxel salience is as you suspect - computational challenge. To do the confidence interval estimation based on percentiles (as we do for the behaviour correlatons for behav-PLS), you need to save the bootstrap distribution. This is a huge computational expense to do for each voxel and each latent variable, so we opted to go with standard error estimation. Its not perfect, but after examining the typical bootstrap distribution the assumption of Gaussianity is reasonable, thus proxy for the confidence interval estimation is good.