3dDeconvolve has a linear model for data in each voxel:
v[n] = stim[0]*iresp[n] + stim[1]*iresp[n-1] + stim[2]*iresp[n-2] + ... + baseline[n] + noise
The stim[] time series you input with the
-stim_file options. The program computes the iresp[] time series (with iresp[j] nonzero only for j between min lag and max lag).
The fitts output is the fitted model, which is the equation above without the noise. This is good for plotting on top of the data (v[]) time series so that you can see if the model captures most of the interesting features of the data.
The iresp[] time series, which is usually very short compared to the data and fitts time series, is useful for determining the amplitude and shape of the response. The statistics of the iresp[] time series (the partial and/or overall F scores) are what you use "to figure out how well stimuli in a experiment are causing an effect".
bob cox