Hi!

I have a question to statistics experts.

At the first level GLM after fitting the model and obtaining the parameter estimates (betas) we can perform a contrast estimate like c’betas and then we can obtain a voxel-specific t-value as (c’betas)/sqrt(var(e)c’(X’X)^-1 c)

The explanation and formula is from here: The General Linear Model (GLM)

I don’t understand where the variance of residuals (var(e)) is from. I thought we obtain only one error value per voxel. But if it’s only one, how can we calculate the variance of one number?

(Sorry for the dumb question, but I feel like I’m missing something and can’t find it with googling)

And another related question.

So we can get betas, contrasts (combinations of betas) or t values from the first level GLM. What should serve as an input to the second level.

And probably completely unrelated question: should we normalize (zcore) timeseries for GLM at the first level? Will it help to compare the results across subjects?

Thanks a lot in advance!