International Journal of Epidemiology 2002;31:1030–1037

• Other Authors: The Pennsylvania State University CiteSeerX Archives
• Format: Text
• Language: English
• Subjects: THEORY AND METHODS; Greenland
• Online Access: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.574.5670; http://ije.oxfordjournals.org/content/31/5/1030.full.pdf

Following a long history of informal use in path analysis, causal diagrams (graphical causal models) saw an explosion of theor-etical development during the 1990s,1–3 including elaboration of connections to other methods for causal modelling. The latter connections are especially valuable for those familiar with some but not all methods, as certain background assumptions and sources of bias are more easily seen with certain models, whereas practical statistical procedures may be more easily derived under other models. We provide here a brief overview of graphical causal models,1–6 the sufficient-component cause (SCC) models of Rothman,7,8 Ch. 2 the potential-outcome (counterfactual) models now popular in statistics, health, and social sciences,9–15 and the structural-equations models long established in social sciences.11–14 We focus on special insights facilitated by each approach, translations among the approaches, and the level of detail specified by each approach. Graphical models The following is a brief summary of terms and concepts of causal graph theory; see Greenland et al.4 and Robins5 for more detailed explanations. Figure 1 provides the graphs used for illustration below. An arc or edge is any line segment (with or without arrowheads) connecting two variables. If there is an arrow from a variable X to another variable Y in a graph, X is called a parent of Y and Y is called a child of X. If a variable has an arrow into it (i.e. it has a parent in the graph) it is called endogenous; otherwise it is exogenous. A path between two variables X and Y is a sequence of arcs connecting X and Y. A back-door path from X to Y is a path whose