Luiz,

My understanding about the censoring is this: If you want to exclude the signal at a timepoint 'k' in a time series of Z(1), Z(2), ..., Z(n), basically 3dDeconvolve would estimate the baseline coefficients and regressor parameters based on Z(1), Z(2), ..., Z(k-1), Z(k+1), ..., Z(n). The reason you want to keep that regressor and the corresponding design matrix intact is because you still want to model this condition in those valid trials.

Since regression analysis is done with least squares approach, ignoring the signal at timepoint 'k' in the process would be equivalent to fitting the value at time 'k' in the regression model with the original 'invalid' signal at timepoint 'k', which leads to no contribution from timepoint 'k' in the sum of squares error (SSE) in the least squares estimation:

SSE = (Z - Z^)'(Z - Z^) (Z^ is the vector of fitting values)

This is what I meant by 'being ignored for modeling'. Hope this makes sense.

Gang