www.cmm.bristol.ac.uk/forum

Hi,

I'm running cross-classified models with neighbourhoods, schools and children. This might be a basic question - but how can I best determine whether the between-group variances are statistically significant? Can I treat the standard deviation of the random effect estimates like a standard error?

Many thanks!
Anja

Hi Anja,

Sound like you are suggesting calculating a z-ratio for the between-group variance by dividing the point-estimate (mean of the MCMC chain for the between-group variance) by its standard error (standard deviation of the MCMC chain for the between-group variance) and then checking whether the test statistic is bigger than 1.96.

Don't do this!

This implicitly assumes that the between-group variance has a normal (i.e. symmetric) sampling distributions when variances typically have skewed sampling distributions. A better approach is to use the DIC statistic to compare the model with the between-group variance to the model without the between-group variance.

Best wishes

George

Many thanks George!