Are multilevel (hierarchical) models the coolest thing in the world or just a fancy way to fool yourself? Jan de Leeuw (statistics, UCLA) thinks both!

Skepticism about multilevel models

Jan is the author (with Ita Kreft) of an introductory text on multilevel models but also wrote this attack on a recent paper by Iain Pardoe and Robert Weidner.

Jan’s general point is that multilevel modeling can be used to estimate varying coefficients that can then be given spurious causal interpretations. I agree with the general point (as I wrote in this article) but I think he’s a bit over the top in critizing Iain and his colleagues, who actually do know a bit about their application area.

Response and discussion

Here’s Iain’s response and also the original Pardoe and Weidner paper.

The paper was also discussed by others. Here’s the discussion by Bill Browne and here’s the discussion by Alan Zaslavsky. These two discussants use mlm’s extensively in educational and health research, and their discussions are more technical and are thus more useful for practitioners of these methods (but less inflammatory for the casual reader).

Pardoe and Weidner’s article, along with the discussions by de Leeuw, Browne, and Zaslavsky, and Pardoe’s rejoinder, will appear in the Journal of Statistical Planning and Inference.