On 7 Mar 2002 05:49:13 -0800, [EMAIL PROTECTED] (Mobile Survey) wrote: > Rich Ulrich <[EMAIL PROTECTED]> wrote in message >news:<[EMAIL PROTECTED]>... > > > If one item has a terrible distribution, it is going to > > correlate poorly. Okay, that means, Drop it. > > For the statistics, you drop both: bad content and > > bad wording. (Wording is usually easier to mend.)
> No. the univariate distributions are normal. But what I meant was that > when I plot one item against the other in a scatter, then since the > variables are measured discretely, I do not get a continuous scatter. > can I continue using the Pearson correlations? Should inter-item and > item to total correlations be reported in terms of the Pearsopn > correlation coefficient? For reporting, use Pearson. People are going to expect it, and people know what it says. Doing the rank-transform (Spearman) is mainly useful when there are outliers on continuous variables with funny scoring. Rank- transforming is more apt to hurt the spacing than to help it, when you start out with an item with a few categories, where the categories were established on purpose. The lack of continuous scatter means that there is some slight 'attenuation' of the correlation when there are just two or three values; so that it is smaller than it ought to be - especially if a dichotomy is near either extreme. For most purposes, you are stuck with the Pearsons; Spearman's, especially, serve no end that I think of. For factor analysis, and other 'latent structures', some people have very large Ns and compute tetrachoric r for their 2x2 contrasts. There's also a 'polychoric r' (more than 2 categories) which is mentioned even less. -- 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/ . =================================================================
