The only way to know is to try. You may run into problems because of little between variation. But it depends on how many clusters you have, how many observations per cluster, etc. The need for taking clustering into account is not determined solely by the size of the intraclass correlation. It is also determined by the average cluster size. An approximation to the design effect (DEFF) is 1 + (s - 1) r where s is the average cluster size and r is the intraclass correlation. Values of DEFF over 2 suggest that taking clustering into account is necessary.
Daan Stam posted on Thursday, October 16, 2003 - 12:15 am
Thank you for your swift reaction,
I noticed you using the same formula in another question thread. However my need for using a multilevel approach lies for the utmost part in the group variables that have no variation on the individual level. Without multilevel I would have to drop these variables.
The problem now lies in the fact that most of my other variables have design effects of below 2. Only 7 of my 25 variables have design effects of a little above 2. I have tried some multilevel models, but have founs no convergence yet.
Your convergence problems are most likely related to your low between-level variance. No amount of trying can solve that. It doesn't sound like your grouping variable is a cluster variable. What is it actually? And also, have you tried to convergence suggestions in the Mplus User's Guide on pages 160-162. If you have started with a really complicated model, you could try starting with one part, getting that going, and then adding the rest a bit at a time.
daan stam posted on Friday, October 17, 2003 - 4:32 am
Thanks you for your respons,
My grouping variable is a the score of a certain group on a specific domain. The entire group gets 1 score, so all the members in the group get that score.
Without multilevel things could get sticky. However things are looking up. I tried your suggestion of using a simple model first and that seems to work.