Commentary and opinion: III. Some nonontological and functionally unconnected views on current issues in the analysis of PET datasets.
Strother et al. (1995) and Friston (1995) both raise important issues and provide useful reviews of various aspects of PET data analysis. Statisticians would not assume that any single piece of methodology would answer all questions about a type of data in a variety of experimental and observational contexts. The fundamental importance of hypothesis-driven inference, based on well designed experiments, cannot be overestimated for its ability to progress scientific understanding in an orderly manner. However, hypothesis-generating experiments are also vital in their own right. In practice, we generally do not have the luxury of both types of experiment, and we should note Strother et al.'s comment on the importance of extracting as much information as possible from each dataset. Friston (1995) also sees formal testing methods and exploratory methods such as principal components analysis as complementary. The correct approach would therefore seem to be (a) to select methods for formal and exploratory data analysis from the rich existing tool kit of statistical procedures, (b) to modify these as necessary to deal with special PET problems such as multiplicity, (c) to be aware of the assumptions underlying the methods being used and to investigate the problems that can arise if these assumptions fail to hold, (d) to appreciate the complexity of both PET data and of the potential questions that can be asked of it, and (e) to be aware of the limitations of any statistical analysis and the need for caution in interpreting conclusions not based on any predefined hypothesis.