Law, GR orcid.org/0000-0001-7904-0264, Green, R and Ellison, GTH orcid.org/0000-0001-8914-6812 (2012) Confounding and causal path diagrams. In: Tu, Y-K and Greenwood, D, (eds.) Modern Methods for Epidemiology. Springer , Dordrecht, Netherlands , pp. 1-13. ISBN 9789400730243
Abstract
The issue of confounding, and the bias it can induce, is a key concern in epidemiology, and yet there remains substantial debate about the definition and identification of confounding variables. Although there is less disagreement concerning the definition of variables acting as competing exposures, these can still prove difficult to identify and/or differentiate from confounders. In this Chapter we describe the use of causal path diagrams, in the form of Directed Acyclic Graphs (DAGs), to define and identify variables thought to act as potential confounders and competing exposures. DAGs have become increasingly popular in the epidemiological literature for modelling hypothesised causal relationships between risk factors and disease. In these diagrams, each variable is represented as a ‘node’. Known, likely, and speculative causal relationships between variables are indicated using unidirectional arrows connecting the two nodes concerned. The arrows are unidirectional because DAGs assume that no variable can be both the cause and the effect of any other variable. In addition a complete circuit from one variable through one or more other variables and back to the original variable is not allowed – hence the diagrams being described as ‘acyclic’. Developing these diagrams as models that accurately represent causal relationships between different variables requires: (i) an understanding of the known functional relationships between each of the variables concerned (informed by a firm grasp of the biological, social and/or clinical mechanisms involved); and (ii) careful consideration of any likely causal relationships based on statistical associations reported by previous empirical studies. These diagrams can also include: (iii) speculative (or hypothesised) relationships between variables (so that these can be specifically tested). Indeed, a key benefit of DAGs is their capacity to provide simple and explicit descriptions of the known, likely and speculative causal relationships under consideration – descriptions which other analysts can use to assess whether the relationships are plausible or correct, and to assess whether the statistical models used to test any speculative relationships adjust for variables that have been correctly identified as confounders or competing exposures.