Innovation in the technical means of empirical inquiry is a necessary and perhaps inevitable component of scientific progress. New tools not only allow investigators to study phenomena that had previously been inaccessible to them, but also permit them to look at existing phenomena in novel ways, and occasionally provide metaphors that serve as building blocks of original theory.^1,2 Yet technical innovation also poses dangers. Among these is the possibility that people lacking the requisite training effectively will be excluded from meaningful participation in scientific discourse. As the methods of empirical research grow more specialised, those who have mastered their use may become increasingly insulated from the criticism of their peers. And science without criticism is bound to go badly. The question, then, is not whether technical innovation is good or bad, but rather how scientific disciplines can capitalise on such advances while simultaneously mitigating the attendant dangers.

Questions of this sort now face our field as social epidemiologists collectively seek to incorporate into our practices a comparatively new class of statistical models, referred to variously as multilevel, hierarchical, random effects, and mixed models. Certainly these models have a great deal to offer us, given our shared interest in exposures that occur at the community level but whose ultimate effects on health are mediated through physiological processes within individual bodies. In fact, multilevel models provide a highly flexible framework for studying a broad range of phenomena, and social epidemiology has just begun to scratch the surface of their potential applications. For many of us, however, statistical methods comprise some of the most difficult and least interesting aspects of our work. And some of us may find the investment of time and energy that would be required truly to master multilevel models to be simply prohibitive. As the use of multilevel models becomes more widespread, then, there is a danger that some of us will be left out of our field’s evolving discourse, while others are tempted to use the technical complexity of their models to shield their work from critical evaluation.

In this issue of the journal, Merlo and his colleagues³ provide a tutorial—the first in a series—on multilevel models in social epidemiology. Theirs is far from the only tutorial on this topic to appear in the epidemiological literature. Indeed, the publications listed in their references section represent the tip of a large and growing iceberg of books and articles that have been written on multilevel models. This tutorial differs from its predecessors in being written on a more conceptual level. The authors present the material in the context of a simple simulated dataset, and provide several figures that may help readers achieve some visual intuition about the interpretation of multilevel models. Of course, no one will become an expert simply by reading this tutorial or even the whole series. But if these tutorials help to demystify multilevel models for some readers, then Merlo and his colleagues will have made a valuable contribution to preserving the inclusiveness of discourse in our field.

A second and perhaps more fundamental danger accompanying the incorporation of multilevel models into the practice of social epidemiology relates to the dominance of empirical work over theory development in our discipline. A glance at the table of contents of any epidemiology textbook shows, in essence, a list of study designs and statistical methods.^4–7 What little theory we have is largely focused on the physiological mechanisms that transform exposures, be they at the individual or community level, into health outcomes.⁸ Virtually non-existent in our field is any systematic theory of the social and economic processes that generate health hazards and shape the distribution of exposures among people and populations. These processes are undoubtedly complex and dynamic, and will require theory that goes far beyond a few paragraphs in the background sections of empirical papers. This, in essence, was the message of Susser and Susser^9,10 when they issued their eloquent call, nearly 10 years ago, for a new paradigm to replace the “black box” approach to chronic disease epidemiology with a multilevel ecological perspective.

Formidable institutional barriers stand in the way of rigorous theory development in social epidemiology. Governmental and other funding sources are interested almost exclusively in supporting empirical research targeting specific health outcomes. Promotion, tenure, and other forms of professional recognition follow along the same lines. Yet somehow we social epidemiologists must find ways of rewarding serious theoretical work on a par with empirical work. For without strong theory we have little basis for knowing what parameters are truly worth measuring—what data to collect, what statistical models (multilevel or otherwise) to fit to them, and how to interpret the results. Continued progress of our field may depend more critically at this point upon our ability to promote theoretical innovation than upon our proper use of the latest statistical methodology.