Statistical analysis of neural organization

Abstract

1. The responses of 32 taste neurons in the solitary nucleus of the rat to 12 stimuli were analyzed with multidimensional scaling (MDS) and cluster analysis (CA) procedures. These analyses of empirical taste data were compared with similar analyses of two model data sets of known configuration to help clarify the implications of these methods commonly used in forming conclusions about the organization of the taste system. 2. To relate to possible conclusions about groupings in taste, both model data sets were chosen as the best possible examples of ungrouped data, the first being completely regular (in the form of a checkerboard) across the taste space, the second randomly arranged. The analysis of the present empirical data appear to be similar to the present ungrouped models, more so the random than the regular model, in the sense that all are amenable to grouping. 3. Because of the similarity of these model MDS and CA solutions to the present empirical solutions and to most published analyses of this sort, the idea is suggested that the appearance of the plots per se for empirical data does not support the conclusion of grouping. And, technically, MDS and CA do not have the statistical power to provide conclusions about issues of neural organization. 4. MDS and CA analyses have two very powerful roles relating to their ability to disclose the hidden organization of complex data sets; they may lend support for or refute theories about the data sets developed from other considerations, and may help generate theories for further consideration. The question of groupings is only one of many such issues. 5. Because data in the present and other reports are quite adequately accounted by MDS solutions of low dimensionality, it is suggested that their organization is characterized as continuous (i.e., rather than belonging to several disjoint spaces). 6. The use of correlations as distance measures in MDS and CA procedures distorts the spatial solutions, making analysis by visual inspection misleading. For example, using correlations, the true or natural spatial arrangements of data sets are probably less circular or spherical than shown in published MDS solutions. Also they are probably more evenly distributed across the space in the sense that the points are actually more concentrated toward the centers of the spaces; this may have strong influences on interpretations of the general form of the solutions. CA solutions can be influenced in analogous fashion. These problems of distortion of the solutions can be avoided with use of direct, linear estimates of distances. (ABSTRACT TRUNCATED AT 400 WORDS)