White Chalak Lu“econometric concepts such as ‘Granger causality’ (Granger, C.W.J., 1969) and ‘strongexogeneity’ (Engle, R., D. Hendry, and J.-F. Richard, 1983) will be classified as statisticalrather than causal." In practice, especially <strong>in</strong> economics, numerous studies haveused G−causality either explicitly or implicitly to draw structural or policy conclusions,but without any firm foundation.Recently, White, H. and X. Lu (2010a, “WL”) have provided conditions underwhich G−causality is equivalent to a form of direct causality aris<strong>in</strong>g naturally <strong>in</strong> dynamicstructural systems, def<strong>in</strong>ed <strong>in</strong> the context of settable systems. The settable systemsframework, <strong>in</strong>troduced by White, H. and K. Chalak (2009, “WC”), extends andref<strong>in</strong>es the PCM to accommodate optimization, equilibrium, and learn<strong>in</strong>g. In this paper,we explore the relations between direct structural causality <strong>in</strong> the settable systemsframework and notions of direct causality <strong>in</strong> the PCM for both recursive andnon-recursive systems. The close correspondence between these concepts <strong>in</strong> the recursivesystems relevant to G−causality then enables us to show that there is <strong>in</strong> fact aclose l<strong>in</strong>kage between G−causality and PCM notions of direct causality. This enablesus to provide straightforward practical methods to test for direct causality us<strong>in</strong>g testsfor Granger causality.In a related paper, Eichler, M. and V. Didelez (2009) also study the relation betweenG−causality and <strong>in</strong>terventional notions of causality. They give conditions under whichGranger non-causality implies that an <strong>in</strong>tervention has no effect. In particular, Eichler,M. and V. Didelez (2009) use graphical representations as <strong>in</strong> Eichler, M. (2007)of given G−causality relations satisfy<strong>in</strong>g the “global Granger causal Markov property"to provide graphical conditions for the identification of effects of <strong>in</strong>terventions<strong>in</strong> “stable" systems. Here, we pursue a different route for study<strong>in</strong>g the <strong>in</strong>terrelationsbetween G−causality and <strong>in</strong>terventional notions of causality. Specifically, we see thatG−causality and certa<strong>in</strong> settable systems notions of direct causality based on functionaldependence are equivalent under a conditional form of exogeneity. Our conditions arealternative to “stability" and the “global Granger causal Markov property," althoughparticular aspects of our conditions have a similar flavor.As a referee notes, the present work also provides a rigorous complement, <strong>in</strong> discretetime, to work by other authors <strong>in</strong> this volume (for example Roebroeck, A., Seth,A.K., and Valdes-Sosa, P., 2011) on comb<strong>in</strong><strong>in</strong>g structural and dynamic concepts ofcausality.The plan of the paper is as follows. In Section 2, we briefly review the PCM. In Section3, we motivate settable systems by discuss<strong>in</strong>g certa<strong>in</strong> limitations of the PCM us<strong>in</strong>ga series of examples <strong>in</strong>volv<strong>in</strong>g optimization, equilibrium, and learn<strong>in</strong>g. We then specifya formal version of settable systems that readily accommodates the challenges to causaldiscourse presented by the examples of Section 3. In Section 4, we def<strong>in</strong>e direct structuralcausality for settable systems and relate this to correspond<strong>in</strong>g notions <strong>in</strong> the PCM.The close correspondence between these concepts <strong>in</strong> recursive systems establishes thefirst step <strong>in</strong> l<strong>in</strong>k<strong>in</strong>g G−causality and the PCM. In Section 5, we discuss how the resultsof WL complete the cha<strong>in</strong> by l<strong>in</strong>k<strong>in</strong>g direct structural causality and G−causality. Thisalso <strong>in</strong>volves a conditional form of exogeneity. Section 6 constructs convenient practi-2
L<strong>in</strong>k<strong>in</strong>g Granger <strong>Causality</strong> and the Pearl Causal Model with Settable Systemscal tests for structural causality based on proposals of WL, us<strong>in</strong>g tests for G−causalityand conditional exogeneity. Section 7 conta<strong>in</strong>s a summary and conclud<strong>in</strong>g remarks.2. Pearl’s Causal ModelPearl’s def<strong>in</strong>ition of a causal model (Pearl, J., 2000, def. 7.1.1, p. 203) provides a formalstatement of elements support<strong>in</strong>g causal reason<strong>in</strong>g. The PCM is a triple M := (u,v, f ),where u := {u 1 ,...,u m } conta<strong>in</strong>s “background" variables determ<strong>in</strong>ed outside the model,v := {v 1 ,...,v n } conta<strong>in</strong>s “endogenous" variables determ<strong>in</strong>ed with<strong>in</strong> the model, and f :={ f 1 ,..., f n } conta<strong>in</strong>s “structural" functions specify<strong>in</strong>g how each endogenous variable isdeterm<strong>in</strong>ed by the other variables of the model, so that v i = f i (v (i) ,u), i = 1,...,n. Here,v (i) is the vector conta<strong>in</strong><strong>in</strong>g every element of v but v i . The <strong>in</strong>tegers m and n are f<strong>in</strong>ite.The elements of u and v are system “units."F<strong>in</strong>ally, the PCM requires that for each u, f yields a unique fixed po<strong>in</strong>t. Thus, theremust be a unique collection g := {g 1 ,...,g n } such that for each u,v i = g i (u) = f i (g (i) (u),u), i = 1,...,n. (1)The unique fixed po<strong>in</strong>t requirement is crucial to the PCM, as this is necessary fordef<strong>in</strong><strong>in</strong>g the potential response function (Pearl, J., 2000, def. 7.1.4). This provides thefoundation for discourse about causal relations between endogenous variables; withoutthe potential response function, causal discourse is not possible <strong>in</strong> the PCM. A variantof the PCM (Halpern, J., 2000) does not require a fixed po<strong>in</strong>t, but if any exist, there maybe multiple collections of functions g yield<strong>in</strong>g a fixed po<strong>in</strong>t. We call this a GeneralizedPearl Causal Model (GPCM). As GPCMs also do not possess an analog of the potentialresponse function <strong>in</strong> the absence of a unique fixed po<strong>in</strong>t, causal discourse <strong>in</strong> the GPCMis similarly restricted.In present<strong>in</strong>g the PCM, we have adapted Pearl’s notation somewhat to facilitatesubsequent discussion, but all essential elements are present and complete.Pearl, J. (2000) gives numerous examples for which the PCM is ideally suited forsupport<strong>in</strong>g causal discourse. As a simple game-theoretic example, consider a market<strong>in</strong> which there are exactly two firms produc<strong>in</strong>g similar but not identical products (e.g.,Coke and Pepsi <strong>in</strong> the cola soft-dr<strong>in</strong>k market). Price determ<strong>in</strong>ation <strong>in</strong> this market is atwo-player game known as “Bertrand duopoly."In decid<strong>in</strong>g its price, each firm maximizes its profit, tak<strong>in</strong>g <strong>in</strong>to account the prevail<strong>in</strong>gcost and demand conditions it faces, as well as the price of its rival. A simplesystem represent<strong>in</strong>g price determ<strong>in</strong>ation <strong>in</strong> this market isp 1 = a 1 + b 1 p 2p 2 = a 2 + b 2 p 1 .Here, p 1 and p 2 represent the prices chosen by firms 1 and 2 respectively, and a 1 , b 1 ,a 2 , and b 2 embody the prevail<strong>in</strong>g cost and demand conditions.3