Violation of Homogeneity: A Methodologic Issue in the Use of Data Mining Tools

Excerpt

The recent publication by Bate et al.[1] suggests that use of the ‘method Bayesian confidence propagation neural network (BCPNN)’ facilitates the identification of signals in the WHO database. As described by Lindquist et al.,[2] the technique is premised on the appropriate use of proportionate reporting rates. Such use makes an implicit assumption regarding the homogeneity of the population to which it is applied. Bate et al.[1] describe the WHO database as one in which 67 countries provide data. We suggest that the assumption of homogeneity is violated by the use of data from so many countries representing diverse medical care delivery systems, regulatory environments, pharmaceutical utilisation, and pharmacovigilance systems. It seems reasonable to conclude that use of BCPNN may result in the identification of potential signals but also create considerable noise. …