Anonymous posted on Monday, April 11, 2005 - 1:46 pm
I'm trying to fit univariate 2-level-models. When I just use the variables (without usage of the CATEGORICAL ARE - option), I can obtain ICCs. But as soon as I specify the variables as being ordinal with the CATEGORICAL ARE - option, there are no ICCs computed at all. What can I do about that? Calculating the ICCs by hand or are there other ways to obtain the ICCs for ordinal variables?
Thank you very much.
BMuthen posted on Thursday, April 14, 2005 - 5:30 am
For categorical outcomes there is no within-evel variance estimated, so the usual formula of between divided by within plus between does not apply. For example, in a logit regression of a binary u on x, the residual variance for u given x does not exist unless you consider an underlying u* variable which is regressed on x with a residual that has a logistic density variance. You could use that logistic density variance as the within-level variance.
Anonymous posted on Thursday, April 14, 2005 - 7:13 am
Would another solution of the problem be to compute an asymptotic covariance matrix (with PRELIS for example) and to calculate the ICCs from that ACM independently from the other analysis (with MPlus)?
BMuthen posted on Friday, April 15, 2005 - 7:59 am
The asymptotic covariance matrix refers to the precision with which the polytomous correlations are computed. So these variances are not useful for this purpose.
BMuthen posted on Friday, April 15, 2005 - 8:01 am
You should look at the Snijders and Bosker book which I think covers ICC's with ordinal variables in line with the u* approach I mentioned above.