Comments on "Causal Inference without Counterfactuals" by A.P. Dawid
Glenn R. Shafer · 1999
In recent years, a number of statisticians and computer scientists have suggested that casual reasoning requires that questions with hypotheses counter to fact have well-defined answers. Phil Dawid’s elegant and insightful article is the first critical examination of this suggestion. As such, it is an essential contribution to the philosophy of probability and causality. It moves the discussion of causality in statistics to a new level of sophistication. The article should prove an effective exercise in persuasion, because Dawid meets the proponents of counterfactuals on their own ground. He begins with the counterfactual variables Yt and Yc that appear in the models formulated by Neyman (1923), Rubin (1974, 1978), and Holland (1986), and he makes every effort to understand how much sense and how much use can be made of these variables. Dawid’s central theme is that counterfactuals should be held up to de Finetti’s observability criterion. This criterion says that it is legitimate to assess a probability distribution for a quantity Y only if Y is observable at least in principle. On this criterion, it is legitimate to assess probabilities for Yt(u), because we can apply the treatment t to the unit u and then observe ∗These comments benefited from research supported by NSF Grant SES-98199116. I have also benefited from discussions of causality with Phil Dawid over many years. These discussions were most recently pursued in the context of a workshop generously supported by the Fields Institute for Research in Mathematical Sciences, which brought together a number of students of causality, including Vanessa Didelez, Mervi Eerola, Michael Eichler, Paul Holland, Steffen Lauritzen, Wayne Oldford, James Robins, Don Rubin, Richard Scheines, and Ross Shachter, in addition to Dawid and myself. Dawid discussed his article at this workshop, and I am grateful to all the participants for their discussion of the issues it raises.