External Measures How-to
From BVwiki
Contents |
Introduction
It is possible to modify the GLM design matrix in a way that behavioral data which are available on a trial-by-trial basis are incorporated into the analysis. This will augment your design matrix in such a way that additional predictors are included and, thus, additional regressors can be statistically evaluated. Common measures would be reaction times or, for more complex reactions, accuracy measures.
This article has been prepared from a documentation written and kindly provided by Fabrizio Esposito.
Step-by step description
Preparation
However, it is necessary that the behavioral data undergo some specific transformation outside of BV before being available as new predictors. Namely it is necessary to standardize these measures. Let's say you want to use the reaction times on a per-trial basis (but any other discrete or integrate continuous data would do): the series of reaction times must be z-trasformed, i.e. de-meaned and re-scaled with respect to its standard deviation.
BrainVoyager GUI
Then, you can create a new predictor on a per-trial basis (coding different trials in different sequential conditions might be helpful for this purpose) by weighting each interval with the corresponding z-transformed value before applying HRF convolution. This is easily done by Ctrl-right-clicking on each trial interval and specifying the weighting value. In the end this predictor will express the reaction time modulation free of the main effect of the corresponding task. Of course the task must be modelled itself as well. Thus an additional predictor withouth weighting must be added to the model. This has to be done, in principle, for all different conditions of the protocol. Especially with multiple conditions it might convenient to create a new protocol version where the conditions are splitted condition_trial, e.g.:
- Rest
- Task1_Trial1
- Task1_Trial2
- ...
- Task2_Trial1
- ...
This is not strictly necessary but it is probably the more error-free procedure for building such a model (a little bit of extra time might be necessary in the protocol editing part but this can be done a reasonably fast way if you work on the text PRT file using any text editor and starting from the original version of the protocol).
Contrasts
After estimation, the best contrast to try is probably a conjunction of the main effect of a task (is there task-related activity?) and the reaction-time modulated predictor (is there a reaction-time related modulation of the activation?).
Further Implications
Please note the resulting design, although orthogonal thanks to the z-transformation, is not balanced. The variance of the modulation predictor depends much on the actual dispersion of the reaction time measures across trials. To have an idea of that, if a person would have always the same reaction, the resulting predictor would be all zeros and no modulation can be estimated. Of course, in general people ensure a sufficient amount of, e.g., difficulty modulation to the task in order to ensure some dispersion in the behavioral measures. In any case the efficiency of the design cannot be fully stated in advance.
Improvements
In low-power cases a second-level ROI-based analysis might incorporate some spatial averaging of the data and be more effective for post-hoc evaluation of the model. But this depends on the background and target of the application.
