Zero level 2 variance: Why?

[dcoul07]

In the context of an analysis involving two levels (50 Groups & 1106 Individuals) using a likert scale (0-8) that I consider as a continuous DV, I get the puzzling results: although the averages vary across groups, fitting the Variance Component model (random intercepts) gives a 0.0000000 estimate of the level 2 variance. I simulated sets of data (binomial, p=.5 & N=8) and obtained the same results. This suspect variance is just the variance of the averages of the 50 groups, and I know there is variation there. How come MlwiN estimates it as being equal to 0? I am using IGLS for the estimation... I used the MIXED MODELS procedure in SPSS and got the same results, so this is not a bug... But something that I am overlooking, most likely very simple, that produces that result. Any help would be appreciated... i posted a similar message in the LEMMA forum, including a sample dataset, if somebody needs to see the problem... Unfortunately, it is not possible to attach files here.

Thanks for any help

[dcoul07]

Addendum: I found another post on this forum, where a similar problem was encountered. Somebody suggested to 'Allow for negative Level 2 variances', which I did. But the final estimate was indeed negative... How is that possible for such a simple model: 2 levels, a single variable? What can numerically cause this?