Dear Adam,

In fast event-related designs it is by the very nature of the design that subsequent stimuli are temporally close together, leading to large overlaps in activations. Although activations tend to sum up linearly, thus allowing to distangle activations from different conditions (assuming randomized presentation and such), and although there have been studies with an interval as low as 500 ms, deviations from the classical BOLD response can be detected up to intervals of about 6 s. Scott Huettel has published a series of papers on that topic, summarized in Huettel (2012, Neuroimage, "Event-related fMRI in Cognition").

In your case the stimulus period is always followed by a decision period, which results in a causal relationship between the two. I'd say the interval in between should thus be at least several seconds if you really want to look at the different components of the trial. However, you should still be unable to properly discern working memory from perceptually driven and decision-related activations, as it's "floating in between".

Of course there are also studies that simply go with a series of stick functions, e.g. one "component condition" every one or two TRs, forwarding only a few of the conditions into analyses (e.g. the first and the last one). This was done in some of the early working memory studies, and I think Ollinger et al. (2001, Neuroimage, there are two associated papers) were the first ones to come up with "half"/partial trials to solve the issue with correlated component regressors. Other studies go with separate models, in the first one just modelling the stimulus, in the second modelling just the response period. Although this seems to work to some extent the regressors can be assumed to highly correlate if they were thrown into a single model. By going with two separate models you avoid the correlations at first sight, but the regressors are contaminated by preceding/subsequent components/activations to a large extent of course.

So I'd say, either go with a relatively long interval, which doesn't really solve the issue with the working memory component, or go with half trials (which would be useful as groups might differ with regard to working memory and corresponding activations, this could be tested by comparing partial trials consisting of stimulus + mask across groups).

Best,

Helmut