George,

I can think of 4 different methods to handle the data if you have multiple runs.

(1) Average

This probably applies to simple experiments such as block designs. Other than this restrictive condition, it is also less kosher in terms of trend modeling.

(2) Concatenate all runs and treat cross-run events the same

This approach assumes no cross-run variation for the same event class. In the end you have one beta for each event class.

(3) Concatenate all runs but treat cross-run event class as separate regressor

Now you can model cross-run variation of the same event type, and you would get multiple beta's for each event class. But you still assume the residual errors (noise) share the same structure across runs since they are pooled and averaged across runs in the model. Also you need to have enough sample size for each event class in each run. The mutliple beta's will be averaged in group analysis.

(4) Analyze each run separately (no concatenation)

You get multiple beta's for each event type without most of the above assumptions except that you need to have enough sample size for each event class in each run. The mutliple beta's will be averaged in group analysis.

Gang