Thoughts on modeling long duration blocks

I have an experiment in which conditions occur in 30 second blocks (TR = 2 sec). I have reason to believe, both from a theoretical perspective and from modeling the data using TENT() functions, that the shape of the response varies significantly from region to region. In some cases there may be a large transient followed by a sustained response. In other cases, there may be a sustained response with no transient or a transient with no sustained response. Some areas may show ramping behavior, increasing or decreasing over the course of the block. My primary interest is in understanding the late steady-state response (e.g., over the last 5 TRs). If I extract this effect as the sum of betas over the last 5 TRs as modeled with TENT functions, I am trying to understand the consequences of including TENTs for all of the TRs in the model, including those of no interest. I don’t have a good intuition for how the degrees of freedom associated with the TRs of no interest affect the stability of the TRs I care about. Would it make sense to use a 3dDeconvolve model that only had TENTs for those last 5 TRs? Alternatively, would it be defensible to use a BLOCK model to capture the steady state response but use a censor file to remove the first 10 TRs of each block from the analysis so that the fit or lack of fit of the early response to the model didn’t affect the overall estimates? --Mike

Bumping this in hopes of getting a response. Thanks!


In general I suggest using TENTzero to cover the event/block duration plus extra 12 seconds or so. It’s difficult to make recommendation without knowing the specifics of the experiment design. For example, would you have multicollinearity problem with so many regressors? You can also try BLOCK and compare the two. Don’t use TENT unless you expect to have BOLD response at the event onset. If you want to use TENT to capture the responses for the last 5 TRs only, I’m not so sure what you’re going to do with other time points.