www.cmm.bristol.ac.uk/forum

Hi,

I am running a 2 level logistic multilevel model. I would like to add an interaction term between two level 2 variables to my current model (they are not collinear). However, the level 2 residual variance is now non-significant (0.004 s.e. 0.003). I understand the Wald test is a pretty rough approximation. Is it justifiable to add an extra explanatory variable in to my model at this stage for exploratory purposes?

I am estimating using PQL2 as I have multiple models to run, but checking with MCMC shows little difference in coefficient values and residual variance.

thanks

Shane

Is anyone able to offer an opinion on this?

I would usually consider the justification for including an explanatory variable as a theoretical issue - whether there is good reason to believe that the variable in question may be related to the outcome. But you probably mean something different by justification here. If you mean 'Is there a reasonable chance of obtaining significance on the new L2 coefficient when the L2 residual variance is non-significant?', I would say 'No chance'. I'd consider whether your data are sufficient in quantity and quality to obtain the result you're after.