www.cmm.bristol.ac.uk/forum

I'm running a simulation study on missing data in multilevel data sets. I generate full data sets first according to a random intercept substantive analysis model and then induce missingness. I generate a continuous response in level-1, 3 categorical predictors in level-1 and 2 categorical predictors in level-2. The cluster level variance was set to be either 0.05 or 0.1 or 0.2 in simulations. I generate 1000 data sets for the simulation study. After generating the full data sets, I fitted the substantive analysis model to each of them. The problem I'm having is that in some data sets, the level-2 variance become zero. I assume this is due to the randomness and the small values used for the cluster level variance.

My question: My concern is on missing data, so, is it wrong if I use these data sets for a simulation study?

Thanks in advance.

Dear Helanidilk,

This all sounds fine. Yes, when you set the true cluster variance to be small, in some simulated datasets the estimated cluster variance will be zero. Yes, this is part of the sampling variability of the cluster variance.

Note, you should not drop the simulations which give estimated cluster variances of zero

Rather, you should report the number of simulations for which this is the case.

Best wishes

George