Spatio-Temporal Partition Modelling: An Example from Neurophysiology
Neurophysiology tries to understand the function of the brain. Frequently activation experiments using positron emission tomography (PET) or magnetic resonance imaging (MRI) are undertaken to map the activation of certain areas of the brain following an experimental stimulus. Such functional mapping experiments produce data consisting of three-dimensional images where voxel values are indicative of regional neuronal activity. Usually no prior knowledge is available and thus the analysis proceeds at the voxel level. In this setting images of statistics are formed where each voxel has an
associated value of a simple statistic. These statistics express the activation at the voxel level. Many eﬀorts have been undertaken in order to provide methods for the classiﬁcation of individual voxels in neuro-imaging. More precisely this attempt deals with two problems at a time: For one to assess if there are any functional diﬀerences in an activation experiment and secondly to ﬁnd methods which allow to address the activity of a voxel at a certain location. One of the earliest attempts was the approach by Duﬀy et al. (1981)
who apply z-scores or t-scores to EEG scalp data. Friston et al. (1990) use an ANOVA model and Holmes et al. (1996) use a randomization test in order to investigate these two questions. In the following we describe the application of methods from spatial epidemiology (Schlattmann and Bo¨hning 1993) and meta analysis (Bo¨hning et al. 1998, Bo¨hning 1999) to the neuro-imaging problem. A review of spatial partition modelling is found in Chapter 7 in this volume.