"Radford Neal" <[EMAIL PROTECTED]> wrote: > Yes, but are the distributions of symptoms given the diseases > adequately described by Gaussian distributions? If not, then the > number of Gaussian components needed to described the overall > distribution of symptoms will not equal the number of diseases.
A very valid concern -- the problem is in using only Gaussian distributions, and would be remedied by using other types in modelling. An axis-aligned Gaussian with a diagonal covariance matrix is nothing else than a mean with a central tendency and can be described as follows: age = 40 +- 10 temperature = 37 +- 2 glucose = 130 +- 20 urine_test = positive (p=0.8) (binomial treated as Gaussian) A state of health can be nicely described this way, because health is about maintaining homeostasis. People use these descriptions very often and without realizing that they correspond to Gaussians. However, in practice, there are situations where a distinguishing characteristic is not a central tendency but a breakdown in it. For example temperature +- 10 > 40 For describing states like that, Gaussians are no longer particularly suitable one would need a different kind of a model, and I haven't seen good alternatives, other than the usual practice of ditching interval scales and adopting ordinal/nominal ones. Perhaps someone reading this post has? -- mag. Aleks Jakulin http://ai.fri.uni-lj.si/aleks/ Artificial Intelligence Laboratory, Faculty of Computer and Information Science, University of Ljubljana. . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
