The geneticist Sewall Wright, in 1921, was the first to use directed graphs to represent probabilistic cause and effect relationships among a set of variables. He developed path diagrams and path analysis, which later went on to be used in the social sciences in methods such as structural equation modelling in the 1970s.
Path diagrams also led to probabilistic DAGs known as Bayesian networks in the 1980s, with artificial intelligence researcher Judea Pearl one of the leading developers. And soon after, causal path diagrams and probabilistic DAGs were merged by Spirtes, Glymour and Scheines (1993) and Pearl (1995, 2000) into a formal theory of causal diagrams, before its introduction into epidemiology in 1999 by Greenland, Pearl and Robins.
At the same time, a concerted effort by Pearl and others fought against the longstanding prejudice in statistics over causality.
Pearl, especially with his book Causality: Models, Reasoning, and Inference in 2000, developed a detailed structural theory of causation that he claims incorporates and unifies other approaches to causation, namely causal graphs, structural equation modelling, and potential outcomes. It is a mathematical theory and introduces the do(.) operator that represents an intervention in the underlying model. The word ‘structural’ is in reference to the causal structure underlying effects in a research study, as represented in a causal DAG, and Pearl defines a structural causal model as one that represents the causal relationships underlying a dataset.
As such, it represents any assumptions we might make in the analysis of that data.* Each structural causal model is related to a graphical model, usually a DAG, but it is mainly his development of DAGs that have earned widespread application.
Nevertheless, some prominent statisticians still regard causal diagrams as inferior to other options. For example, Donald Rubin states that while these “graphical approaches seem to be a clear advance with respect to causal inference over older, less subtle graphical approaches”, he nevertheless feels that “the framework is inherently less revealing than the potential outcomes framework because it tends to bury essential scientific and design issues”.
Despite such views, however, over the last two decades the use of causal diagrams has grown, and they have even been called the “flagship of the new methods”. Numerous researchers and statisticians are now promoting their use, so continued growth does seem likely.
* ‘data’ is used here in the modern sense as a mass noun rather than the plural of datum