Our general approach has been to analyze fMRI data without making any shape assumptions. This is done by modeling each time point of the BOLD response with a delta function (Buckner et al. 1996). The BOLD timecourse is then estimated with the general linear model (Friston et al. 1995). Statistical inferences are performed across a group of subjects by analyzing the timecourses with a repeated measures analysis of variance.

To give an example, consider a match-to-sample paradigm: a subject is shown a sample object, a delay is introduced, and then a test object is presented. The subject's task is to determine whether the test object is the same as the sample object. During a trial the subject must encode the sample information, maintain that information during the delay interval, and match that information to the test stimulus. Different brain regions may emphasize each process. Regions that encode the sample stimulus and regions that match the sample stimulus to the test stimulus should show a relatively transient response, while regions that maintain that information over the duration of the delay period should show a more sustained response. Thus information is carried not only by the existence of a hemodynamic response, but also by its shape. If a shape for the BOLD response is assumed that is representative of only the transient response, then regions with a non-transient response will suffer a reduction in estimated response magnitude and thus be detected with decreased sensitivity. With a shape-free analysis, the timecourses of transient and sustained responses can be estimated unambiguously with equal sensitivity.