Hierarchical Data Modeling in the Social Sciences

Abstract

The last 10 years of active research in the area of hierarchical, multilevel data modeling has brought problems as well as benefits. The three conference papers reflect well both the potentialities of the new procedures and some of the dangers we need to guard against. As in all statistical modeling of the real world, our inferences are no better than the data upon which they are based and the adequacy of the assumptions we are prepared to make. The paper by de Leeuw and Kreft sounds some useful warnings, and I will discuss that one first. The paper by Rogosa and Saner focuses in detail on a repeated measures application and one software package, and asks questions about the usefulness of the available analysis procedures. I shall have some general remarks about ways of handling repeated measures data, but leave comments about the HLM software to Professor Raudenbush to respond to. The paper by Draper is concerned with causal inference and ways in which this can be strengthened by using multilevel models. He also places these models in their historical context, and his discussion of competing estimation procedures raises some interesting topics for future research.