Causal Inference and Scientific Paradigms in Epidemiology

Empiricist philosophers such as Francis Bacon, John Locke, and David Hume believed that knowledge is gained through observations of natural phenomena. In contrast to deductive logic, inductive logic is not self-contained and therefore is open to error. On the other hand, deductive logic cannot by itself establish a theory of prediction since it has no connection to the natural world. Hume observed that inductive inference does not carry a logical necessity. He challenged the notion that causality could be proved, while highlighting the subjectivity of knowledge and the fallibility of inductive reasoning. In the nineteenth century, Robert Koch provided a framework for identifying acute diseases associated with microorganisms. In the twentieth century, following World War II, efforts were made by Sir Austin Bradford Hill and others to systematize and justify causal inference in observational research. More recent authors, including Mervyn Susser, have offered refined accounts of causal criteria. The Bradford Hill criteria for causal inference or subsets of the criteria are still widely used as a heuristic aid for assessing whether associations observed in epidemiologic research are causal. The model of sufficient component causes proposed by Kenneth Rothman is widely used in epidemiology as a framework for teaching and understanding multicausality.

There have been many important developments in quantitative models for assessing causality in the last two decades as well as important related developments in epidemiology, statistics, philosophy, and computer science. Writers have conceptualized causality using deterministic models, quasi-deterministic models, and probabilistic models. In epidemiology, a probabilistic model of causation holds that a cause increases the probability that a disease or other adverse health condition will occur. There has been extensive discussion of probabilistic causation in the literature on the philosophy of science. Under a probabilistic model of causation, a cause may be neither necessary nor sufficient for the disease to occur. In contrast, a deterministic model holds that diseases have causes and if these causes are present then diseases will follow. Both probabilistic and deterministic models of disease causation can be linked to sufficient-component models of disease causation. Sufficient-component models of disease causation are especially suitable for understanding the complex causation of many chronic diseases. Approaches for warranting causal claims in epidemiology include quantitative models such as the counterfactual model and structural models.

Much of the literature on causal inference in epidemiologic research has dealt with causal inference in the field of epidemiology as a whole, or has focused on causal inference in certain areas such as chronic diseases and environmental causes of diseases and adverse health outcomes. In recent years, however, several authors have dealt with causal inference within a variety of epidemiologic subdisciplines including nutritional epidemiology, genetic epidemiology, infectious disease epidemiology, and social epidemiology. Although criteriabased approaches are still widely cited and used, enthusiasm for the Bradford Hill criteria or subsets of the criteria appears to be waning in some areas of epidemiology (or among some groups of epidemiologists). An increasing number of authors have argued that traditional criteria for causal inference in observational research do not apply to particular epidemiology subdisciplines or that certain criteria should be modified. A large and growing literature has dealt with quantitative models for estimating causal parameters using data from observational studies (for example, counterfactual models and structural equation models).

A notable feature of the debate about scientific paradigms and theories in epidemiology has been the juxtaposition of scientific, social, and ethical arguments. In some articles on epidemiologic theory and scientific paradigms, ethical issues and professional values have been discussed in order to bolster or clarify the authors' scientific arguments or to point to future directions. However, the introduction of statements laden with social concerns, ethics, and values into a scientific discussion of epidemiologic theory tends to obscure the scientific issues at the heart of the debate. Analyses of professional values and ethics in epidemiology are vitally important, but they should be thorough and well balanced and clearly distinguished as considerations of ethical and social concerns. Judgments about the relative scientific merits of alternative scientific paradigms (or proposed refinements in existing paradigms) should be based upon scientific considerations.

Metaphors and visual images have been used in epidemiology in conjunction with frameworks and theories proposed to explain disease causation and the processes that give rise to social patterning of disease risks and the ways in which risk behaviors arise and become maintained in social groups. The metaphor of a running stream has recently been extended by Glass and McAttee to stimulate creative thinking about the determinants of risk behaviors in populations and emerging paradigms of disease causation. The frameworks and theories that have been proposed to illustrate or explain the processes that give rise to social patterning of disease have been paralleled by the development and refinement of methods of multilevel statistical analysis and complex systems modeling in public health. The metaphors and conceptual models employed in epidemiology have moved beyond short, causal chains and reductionist world-views to include more holistic, dynamic perspectives that are more consistent with the complexity of real world situations.

Over the past decade, leading epidemiologists have noted the importance of social factors in studying and understanding the distribution and determinants of disease in human populations; but to what extent are epidemiologic studies integrating genetic information and other biologic variables with information about individual-level risk factors and group-level or societal factors related to the broader residential, behavioral, or cultural context? There remains a need to consider ways to integrate genetic information with social and contextual information in epidemiologic studies, partly to combat the overemphasis on the importance of genetic factors as determinants of disease in human populations. Even in genome-wide association studies of coronary heart disease and other common complex diseases, only a small proportion of heritability is explained by the genetic variants identified to date. It is possible that familial clustering due to genetic factors has been overestimated and that important environmental or social influences (acting alone or in combination with genetic variants) have been overlooked.

Changes in research paradigms and theories about disease causation have frequently led to refinements in frameworks for causal inference. Among the most promising paradigm shifts in contemporary epidemiology has been an increasing willingness to examine disease etiology using a multilevel or systems approach and a parallel trend towards the mathematization of causal inference. However, these two important developments (adoption of a multilevel or systems approach for epidemiologic research and use of quantitative models for causal inference) have not been adequately linked. One important impediment to reaching the full potential of multilevel studies is the need for further refinement of quantitative and graphical models for causal inference that are suitable for emerging research paradigms. More efforts should be made to clarify conceptually what approaches work best for identifying causal relationships in the context of complex systems. Optimal approaches to causal inference are not necessarily identical across epidemiology subdisciplines and researchers should not assume that there is one correct or optimal theory of causality that accounts for causal relationships identified in epidemiologic research.