General methodological considerations

I wish to thank Alok Bhargava for organizing a discussion of this interesting paper by Adams et al. My discussion will focus on quite generic philosophical, statistical, and modelling issues that arise when inferring causal e0ects from complex panel and time series data. The analysis of Adams et al. appears to be driven by the following philosophical proposition: when the sample size is large and a test of the statistical hypothesis of Granger non-causality accepts, one can validly infer lack of causation based on observational panel and/or time series data, even in the absence of subject-matter-speci5c background knowledge; however, because of the possibility of hidden (unmeasured) common causes, it is never possible to infer causality in the absence of background knowledge. In this discussion, I will argue that the authors’ apparent philosophical belief that it is possible to infer lack of causation from lack of association in the absence of substantive subject-matter knowledge implicitly uses the “faithfulness analysis” of causation made explicit by Spirtes et al. (1993) in their book Causation, Prediction and Search and implemented in their computer program Tetrad. In Section 2, I review the assumptions underlying a “faithfulness analysis”. In Section 3, I will describe some of the surprising implications of these analyses due to Spirtes et al. (1993) and Pearl and Verma (1991). Speci5cally, I will show that in certain speci5c settings and contrary to the intuitions of the authors’, it is also possible to infer causality using a faithfulness analysis. Sections 2 and 3 are devoted to describing how a faithfulness analysis can be used to infer causation and non-causation based on dependencies and independencies found in observational data in the absence of subject-matter speci5c background knowledge. In Section 4, I argue, following Robins (1997) and Robins and Wasserman (1999), that a faithfulness analysis is not appropriate when, as in most econometric and epidemiologic studies, the probability that any two given study variables have no common cause is close to zero. As a consequence, in most econometric and epidemiologic studies,