Models and meanings: Therapist effects and the stories we tell

Click to increase image sizeClick to decrease image size

Notes

1. Truth-in-packaging warning: I was a reviewer for one of these studies.

2. I personally prefer the term multilevel model, which also seems to be the dominant one in the statistical literature. The term hierarchical linear model and its acronym HLM are less desirable for two reasons. First, HLM is the name of a particular computer program, leading to confusion between software and analytic technique (similar to that in structural equation modeling, which is sometimes referred to as LISREL modeling, after the computer program). Second, these models are being generalized to nonlinear models, including those for nominal and ordinal dependent variables (Fitzmaurice, Laird, & Ware, 2004; Skrondal & Rabe-Hesketh, 2004). We thus have the odd situation in which the HLM (hierarchical linear models) program now contains procedures for nonlinear modeling!

3. At the risk of offending, I should say that error, although unlikely with such experienced research teams, cannot be completely ruled out. Multilevel models are relatively complicated, and the software to analyze them can be difficult to use. Although both of these studies are by research teams with considerable experience with these models, I am concerned about the tendency for routine use of multilevel models by those less experienced in their use than these authors. Perhaps we are entering the stage at which studies should be required to include computer code and data (when possible), perhaps on the World Wide Web, along with the published study. It would be useful for the authors of these two studies to exchange computer code and even to explore each other's models to locate more precisely the exact reasons for the discrepant findings.

4. Although not the authors of these studies, I suspect.

5. Fortunately, Elkin et al. are well aware of this issue and deal with it successfully.