Currently convolve_hrf simply downsamples the event time-course to give event weights for each volume acquisition. However, this means that for very brief events (less than the duration of an acquisition) then this event will be ignored. This behavior may or (most likely) may not be desired but currently there is little explanation of this. Possible solutions include downsampling after the convolution (but this increases computational demands) or taking the event value the mean of each acquisition epoch.