Yes there is something wrong going on. A possible explanation is that there is no variation associated with your grouping level.

Check this out.

http://support.sas.com/kb/22/614.html

It sounds like you might be overfitting your model.   

Well in SAS notation, the G matrix refers to the variance-covariance matrix of random effects. In this case (a single random effect associated with the intercecpt), this is a scalar. So there is no way to simplify it. Instead, either you remove the random effect, or you indicate a broader level of grouping, if any.

I concur with the earlier correspondents - one of the things that can happen is that during the estimation the level 2 variance has gone negative at an iteration - clearly such a negative value is problematic.  But such values are possible implying that is negative dependence between your groups - they are more un-alike than a random sample eg if pigs in a litter are under competition for food , some may grow at the expense of others.

It would be interesting to know the substance of your data - how many groups ; and what is the typical number of lower level units in a higher level unit ; and what is the degree of imbalance, Also do you have a very non- normal distribution for the response and are there outliers.

I have 17 classrooms from three 3 schools. Classroom sizes range from 12 to 16. I don't have very non-normal distribution of outcomes. 

Thanks 

Ibraihm 

Just 17 is pushing it a bit - 3 is far too few. Remember that in even in a random intercepts model you are estimating a mean and a variance with at the school level only 3 "observations"

Try a two level model - pupils in classes (the class variance will have any school effect incorporated) and then put in a fixed effect for the schools eg two dummy variables.

see also

https://www.researchgate.net/post/What_is_the_explanation_of_negative_variance_in_the_statistical_models