6

I'm using xtmixed in Stata to test a Hierarchical Linear Model. My problem is that variance at level 2 is about 4% of the total variance. So most of the variance is at level 1 and the intraclass correlation coefficient is 4%.

In other studies (different fields) I see much higher intraclass correlation coefficients (30-50%). Is there a threshold for a multilevel model to make sense? Or can I publish my results just showing the 4% (and its reduction when adding predictors to the empty model)?

Forinstance
  • 755
  • 2
  • 8
  • 29

1 Answers1

5

John B. Nezlek argues that ICC should not be a ground for justifying decisions on multilevel models, because it's values could be misleading. In his article he gives a synthetic example of varying within-group relationships when intraclass correlations are 0 (attached below). So some, like Nezlek, would say that this is not a problem.

enter image description here

See: Nezlek, J.B. (2008). An Introduction to Multilevel Modeling for Social and Personality Psychology. Social and Personality Psychology Compass, 2(2): 842–860.

Tim
  • 108,699
  • 20
  • 212
  • 390
  • 3
    I have never seen this paper before but I totally agree with this advice. – Jake Westfall Nov 26 '14 at 17:41
  • Thanks! So is it ok to publish a model where I start from a 0.5% ICC. Then I add some predictors (lvl1) reducing of 25% the variance at level 1 and just one predictor at lvl2 reducing the variance at level 2 of 100%. Predictors are significant. But does it make sense to add the one predictor at level 2 in this case (reducing the variance of 100%)? Should I use the Cohen's f2 to estimate the effect size? – Forinstance Dec 02 '14 at 10:25