Dear Paloma, dear Donald,
> Third, you specify and FIR for your task and then include the save
> regressors from the first design as "user regressors".
Note that "regressor" predictors are mean-centered automatically, while "condition" regressors are not. This should affect the model then, as the FIR regressors obtained from model 1 and entered as regressors into model 2 are not identical to the original FIR regressors due to the centering.
> Visscher (2003) and some other authors recommend modelling blocks with canonical HRF but events with FIR.
Concerning the study by Visscher et al. (2003), the main limitation is the predicor with a fixed shape for the sustained activation (convolved gamma function) and the FIR / series of delta functions for the transient effect. They also discuss this and conclude:
If the transient response shape does not fit the assumed response
shape, activity corresponding to the difference between
the true activity and the modeled shape could be
misapplied to the sustained regressor.
In the following passage they focus on misapplied sustained activation, based on simulated data reflecting transient canonical shape regressors with different durations. As far as I understand the predictor remains the same though ("assumed the basic SPM canonical shape for this activity" - no indication of a convolution). Well yes, if the actual time course is based on a stimulus duration of e.g. 1.25 s but the predictor is based on a stimulus duration of 6 s then there's some discrepancy, and this will affect the model negatively. As the transient trials are placed randomly within the block, but are packed relatively densely, this will result in some level of positive activation throughout the block, which might be caught up by the transient block regressor. It's not/less a problem for the FIR model because the FIR model doesn't have any assumptions about the time course, accordingly, the FIR regressor set can't be "wrong", it can only be "too short" (this is why the 10 regressor FIR set has less misapplied activation compared to the 7 regressor FIR set for longer stimulus durations).
Now, we can also argue that if the sustained response shape does not fit the assumed response shape then the unmodeled activity might be misapplied to the transient regressor. They discuss this on page 1706. Based on their observations in the acquired data there seems to be no major problem, but it might well be different *with different data*. Even if it doesn't affect estimation of the transient activation it leaves open the question about the misestimated sustained regressor. Thus, their statement "transient responses should not be modeled in the GLM with assumed shapes" has to be interpreted within the context (regressors constructed with a certain duration predicting simulated data constucted with other durations, or more generally, if BOLD response doesn't agree well with the predictor; unsurprisingly, the midmodeled activation becomes larger for larger duration discrepancies (Fig. 11 B)).
In summary, I'd say:
- With real data, the deviation from the predictor is hopefully within a limited range. Otherwise, and/or if the response shape is unknown, then we should never go with an assumed shape, be it a block, event, mixed design, be it a transient or sustained effect, but instead turn to more flexible predictors, e.g. FIR sets.
- Accordingly, depending on expectations it might well make sense to go with canonical HRF regressors for both effects in a mixed design, and there are various papers that do so. Canonical HRF regressors might also be preferable in some instances, as Donald has already pointed out, that is if you want to take into account trial-specific durations (which you can't do with FIR; except if you go with an extra FIR set for each of the possible durations) or if stimulus onsets are not fixed relative to TR onset, as in FIR, the onsets are rounded to the next TR.
- There's going to be a problem with shared variance due to non-orthogonal sustained and transient regressors, which should hold for any of the predictor combinations (FIR & FIR, FIR & canonical, canonical & canoncial). One common approach is to orthogonalize one (set of) regressor(s) onto the other, thus testing for sustained effects plus transient effects that cannot be explained by sustained effects. The shared variance would be explained by the sustained regressor then (whether this is what you want to do is another issue though).
> we found bizarre results.
When working with FIR there are two common error sources:
1) The duration has to be specified as 0, as the length of the "on" periods of the stick functions and the overall length / no. of stick functions is determined via the FIR settings
2) in your design, different trials of a transient condition might be relatively close together, leading to an overlap of the FIR responses. Per default SPM orthogonalizes within conditions, resulting in an orthogonalization of the FIR regressors. To obtain a design matrix similar to the one in Henson et al. (http://www.fil.ion.ucl.ac.uk/spm/doc/papers/rnah_choice.pdf ) the ortho. should be disabled, see https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;f525092b.1508 .