Anita,
This is an interesting question!
Most investigators in the field are usually obsessed with the statistical values, and pay little attention to the effect estimates. The range of effect estiamtes (e.g., percent signal change) varies across experimental types (event-related vs. block), experimental conditions, brain regions, subjects and groups. If you just use one effect estimate (e.g., beta value) for each condition, it might be hard to establish a criterion or cutoff for outliers. I can think of two possible methods. One possible solution is to downgrade those potential outliers with, for example, option -model_outliers in 3dMEMA. Another approach is to model the hemodynamic response with a model-free method (e..g, TENT or CSPLIN in 3dDeconvolve), and use the signature HDR shape as an identifier for anomalies or head motion effects.
Gang