Hi,
On 11 Dec 2010, at 17:59, Christopher Benjamin wrote:
Dear FSL ninjas, I am writing with a question about model fit and be very grateful for any help.
I have three conditions of different durations (slow = long duration, comf = medium duration, fast = short duration). Not good, I know. I am trying
Yes, it couldn't distinguish strength of response due to stimulus variation v.s that caused by the differing durations.
to pull apart whether the difference we see for fast > slow (greater activation in short duration as compared to long duration condition) contrast is due to (a) a true difference in activation between conditions, or (b) differing model fit between the two conditions.
This seems a little circular to me - so I've probably misunderstood the question - the contrast is calculated directly from model fit so... (b) but, assumes the model is good so... (a).
Looking at this, it looks as though there is not a drastic difference in fit in the different conditions – esp. for slow and fast conditions, which is the key point I think. In each case, the raw data’s peak is about 60% of the value for the model’s peak.
Couple of comments: if you normalised your data to the model's range, as stated above, then why are these still so different; is looking at a single example of one of your best-fitting voxels likely to reveal very much about the results in general?
To investigate further, I took the same data and
- normalized all graphs by the maxima and minima for the slow model
- plotted this out (pdf, right side of page)
The "All raw data" plot is probably most useful here - even better if you include error-bars to show the subject variation - as it indicates a clear difference in strength of response to the different conditions.
I am trying to determine now what the exact effect on the beta weights will be, and how to determine the finding of significantly greater betas in the fast as compared to slow condition. Clarification would be very helpful.
Sorry, could you please clarify the question. In particular: "exact effect" of what on the beta weights?
Thanks people
Christopher
- Note: the data used was extracted from tsplot/tsplot_zstat1.txt columns one (data) and three (full model fit).