On Wed, 26 Mar 2003 13:18:17 +0100, "david" <[EMAIL PROTECTED]> wrote:
> Hello > I would like to know a measure for testing the binormality of a data set. > thanks in advanced I don't use the term, binormal. And Google turns up something concerning topology, which is not a usual topic in these groups. Are you referring to bivariate normality? If so -- the straight, simple approach, to start, is to test the variables separately, and combine tests, preserving power or alpha, depending on which matters to you. Bivariate normality implies that the linear combinations of the two should also be normal. The obvious combinations are the principal components: form them and test them; and combine tests. There are various tests for 'normality.' How the data are generated could suggest the fashion in which non-normality is apt to arise. If the variables are two versions of the same measurement, there might be special salience to simple difference score. If you have a particular need -- symmetry, continuity (lack of ties), absence of outliers, presence of exponential tails, infinite theoretical range -- you should select or design a test to meet that need. If you have a particular reason to suspect deviations that you can characterize -- and they matter -- you should test for those deviations. - "goodness of fit" depends on what aspects you want to fit, so there is never a single test for something like "normality." Hope this helps. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
