We are getting strange results for one particular subject in a study, and I am stumped as to how this is possible. The study involvles three condtions, A, B, and C. The stimuli are all auditory sentences. Subjects listen to sentences while in the scanner in an event-related design. There are four runs, and each run has the conditions mixed randomly with varying intervals between them. TR=2 s. The entire sentence is presented within one TR.

Because these are auditory stimuli, you would expect stong auditory cortex activation when each condition is compared to silence. This is what you get in all subjects except one. This one subject shows normal looking strong auditory cortex activation for conditions B and C and some other scatterted activation in rest of the brain, like other subjects. Condition A, however, shows very weak auditory activation and very strong negative activation in large parts of the brain. There are many reason why the data can be corrupted or noisy (scanner aritifact, subject sleeping, etc.) but this cannot affect just one condition. At the very least, auditory cortex should be activated strongly when you are listening to sounds comapred to silence. Why this doesn't happen just for one condition is mysterious. The number of stimuli in each condition is about the same, so it is not a lack of statistical power for one condition. There are a number of other subjects that heard the stimuli in exactly the same order as this subject, and their activation is fine.

I have tried a number of things, but the results are essentially the same. Each run was analyzed separately, but each run shows the same pattern (weak auditory and strong -ve activation for A, normal for B and C).

An obvious thought was that there is an error in the regressors (they may be off by a TR or two). I double and triple checked the regressors, and they seem to be fine.

The original analysis was done by estimating the HRF with 3dDeconvolve. Instead of trying to estimate the lags, I also used a standard HRF (gamma variate and Cox special from waver) but results are still the same.

I also tried to simplyfy the analysis as much as possible, by removing all extra regressors like the motion parameters, but the results are still the same.

Any ideas about what could be wrong, and any tricks I could try?