Everything should be made as simple as possible but not simpler.

Can complexity theory throw some different light on the aetiology of complex diseases, 1 currently explored mostly by the probe of molecular genetics? 2 Epidemiology, particularly as developing in the last half a century, has been dealing with disease aetiology—‘a priori’ not known whether simple or complex—with quite simple tools from an epistemological viewpoint. Observational epidemiology studies of aetiological factors are conceived, and whenever feasible carried out, as association studies at the individual level. An association and its nature, causal or non-causal, is researched within the same individuals of one or more exposures of interest with an outcome adjusting for other exposures, which may distort (confound) the association. This basic and invariant study concept has been developed into a vast and sound array of methods of study design and statistical analysis aimed at (i) making it applicable within a variety of purely observational circumstances and (ii) approaching the same study validity attainable by an experiment in which exposures are selected and assigned at will by the investigator and randomization is employed to control confounding and biasing factors unknown or known but not controlled by systematic arrangement. This approach is in principle universally applicable and robust, as it does not depend on the specific nature of the exposures and of the disease under study or on any formal and quantitative model of disease causation. Only a very limited number of such models have in fact been produced, the best known being the multi-stage models of carcinogenesis initially proposed by Armitage and Doll, 3 useful in refining but not crucial for the interpretation of the data from the association studies. Ironically for a discipline defined as ‘The study of the distribution and determinants of health-related states and events in specified populations. . .’ the research approach operates in fact at the level of individuals, whose disease risk produced by an exposure is estimated using the collection of conceptually identical and independent subjects called a ‘population’ (often a sub-group of the real population of interest). The distributions of the exposures and outcomes in the population and their evolution in time are regarded as just descriptive data providing only suggestive or ancillary evidence on causation, always infiltrated by the possibility of an ecological fallacy. In a paradigmatic study the real population is stripped of all its history to extract, as closely as possible to a controlled experiment, selected exposure values for individuals and to find out what frequencies of outcome correspond to them. Two domains escape this restriction: infectious disease epidemiology and population genetics, in which models of transmission of putative causal agents contribute substantially to their identification as actual agents. Still in both fields the final word goes to the association method at the individual level, as shown by the studies on HPV viruses and cervical cancer 5 or Helicobacter pylori 6 and gastric cancer, and by the recent major shift in emphasis towards association studies, both exploratory and confirmatory, in genetic epidemiology. A second stripping of possibly available information may occur when moving from the population level, above the individual, to the levels below the individual as a unit, namely when biological markers are the exposures under investigation. Common models of data analysis, as logistic, Poisson or proportional hazard models, do not take into account (or only partially in Cox’s model with time-dependent covariates) the time sequence, the interconnections and the hierarchical level within the body physiology of the various markers potentially involved in disease pathogenesis. Basically they are ‘omnibus’ models, in which the interrelationships between exposures collapse into first order (two factors) or at the maximum second order (three factors) interaction terms in a linear model (higher-order interactions are very hard to assess even in balanced randomized factorial designs). 7