Hi everyone

I am working a project looking at the clustering of suicide behaviours among secondary school students. We have a data set of over 9000 students from about 90 schools. My question is around using the latent logistic distribution to calculate the intra-class correlation for dichotomised indicators versus using the student-level and school-level variances for continuous indicators. My problem is that the results as quite different for arguably the same underlying behaviour.

When I a use a scale for suicide behaviour (combining 3 likert questions with indicators) the school level variance is 0.004 and individual level variance is 1.837, giving an ICC of 0.002 or 0.26%.

However if I just use I dichotomised indicator for suicide risk, then the school level variance is 0.547 and using (pi^2)/3 as per Snijders & Bosker. Which gives me an ICC of 0.14 or 14.2%! (I am using Glimmix in SAS with the cumulative logit link function.)

I am having difficulty reconciling these differences, as the two measures use the same underlying information to derive the indicators.

Regards