Causality in Time Series - ClopiNet

Causality in Time SeriesVolume 5:Causality in Time SeriesFlorin Popescu, Isabelle Guyon and Nicola TalbotFlorin Popescu and Isabelle GuyonProduction Editor: Nicola Talbot

ForewordFlorin PopescuFraunhofer Institute FIRSTi

PrefaceThe following is a print version of articles submitted by speakers at the Mini Symposiumon Time Series Causality which was part of the 2009 Neural Information ProcessingSystems (NIPS) 2009 Conference in Vancouver, CA.March 2011The Editorial Team:Florin PopescuFraunhofer Institute FIRST, BerlinFlorin.popescu@first.fraunhofer.deIsabelle GuyonClopinet, San Franciscoguyon@clopinet.comiii

Table of ContentsForewordPrefaceiiiiLinking Granger Causality and the Pearl Causal Model with Settable Systems 1H. White, K. Chalak & X. Lu; JMLR W&CP 12:1–29 .Robust statistics for describing causality in multivariate time-series. 33F. Popescu; JMLR W&CP 12:30–64, 2011.Causal Time Series Analysis of functional Magnetic Resonance Imaging Data 73A. Roebroeck, A.K. Seth & P. Valdes-Sosa; JMLR W&CP 12:65–94, 2011.Causal Search in Structural Vector Autoregressive Models 105A. Moneta, N. Chlaß, D. Entner & P. Hoyer; JMLR W&CP 12:95–114, 2011.Time Series Analysis with the Causality Workbench 131I. Guyon, A. Statnikov & C. Aliferis; JMLR W&CP 12:115–139, 2011.v

JMLR: Workshop and Conference Proceedings 12:1–29Linking Granger Causality and the Pearl Causal Model withSettable SystemsHalbert WhiteDepartment of EconomicsUniversity of California, San DiegoLa Jolla, CA 92093Karim ChalakDepartment of EconomicsBoston College140 Commonwealth AvenueChestnut Hill, MA 02467Xun LuDepartment of EconomicsHong Kong University of Science and TechnologyClear Water Bay, Hong Konghwhite@ucsd.educhalak@bc.eduxunlu@ust.hkEditors: Florin Popescu and Isabelle GuyonAbstractThe causal notions embodied in the concept of Granger causality have been arguedto belong to a different category than those of Judea Pearl’s Causal Model, and sofar their relation has remained obscure. Here, we demonstrate that these conceptsare in fact closely linked by showing how each relates to straightforward notions ofdirect causality embodied in settable systems, an extension and refinement of the PearlCausal Model designed to accommodate optimization, equilibrium, and learning. Wethen provide straightforward practical methods to test for direct causality using testsfor Granger causality.Keywords: Causal Models, Conditional Exogeneity, Conditional Independence, GrangerNon-causality1. IntroductionThe causal notions embodied in the concept of Granger causality (“G−causality") (e.g.,Granger, C.W.J., 1969; Granger, C.W.J. and P. Newbold, 1986) are probabilistic, relatingto the ability of one time series to predict another, conditional on a given informationset. On the other hand, the causal notions of the Pearl Causal Model (“PCM")(e.g., Pearl, J., 2000) involve specific notions of interventions and of functional ratherthan probabilistic dependence. The relation between these causal concepts has so farremained obscure. For his part, Granger, C.W.J. (1969) acknowledged that G−causalitywas not “true" causality, whatever that might be, but that it seemed likely to be an importantpart of the full story. On the other hand, Pearl, J. (2000, p. 39) states thatc○ H. White, K. Chalak & X. Lu.

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 in 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 arising naturally in dynamicstructural systems, defined in the context of settable systems. The settable systemsframework, introduced by White, H. and K. Chalak (2009, “WC”), extends andrefines the PCM to accommodate optimization, equilibrium, and learning. In this paper,we explore the relations between direct structural causality in the settable systemsframework and notions of direct causality in the PCM for both recursive andnon-recursive systems. The close correspondence between these concepts in the recursivesystems relevant to G−causality then enables us to show that there is in fact aclose linkage between G−causality and PCM notions of direct causality. This enablesus to provide straightforward practical methods to test for direct causality using testsfor Granger causality.In a related paper, Eichler, M. and V. Didelez (2009) also study the relation betweenG−causality and interventional notions of causality. They give conditions under whichGranger non-causality implies that an intervention has no effect. In particular, Eichler,M. and V. Didelez (2009) use graphical representations as in Eichler, M. (2007)of given G−causality relations satisfying the “global Granger causal Markov property"to provide graphical conditions for the identification of effects of interventionsin “stable" systems. Here, we pursue a different route for studying the interrelationsbetween G−causality and interventional notions of causality. Specifically, we see thatG−causality and certain 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, in discretetime, to work by other authors in this volume (for example Roebroeck, A., Seth,A.K., and Valdes-Sosa, P., 2011) on combining 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 discussing certain limitations of the PCM usinga series of examples involving optimization, equilibrium, and learning. 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 define direct structuralcausality for settable systems and relate this to corresponding notions in the PCM.The close correspondence between these concepts in recursive systems establishes thefirst step in linking G−causality and the PCM. In Section 5, we discuss how the resultsof WL complete the chain by linking direct structural causality and G−causality. Thisalso involves a conditional form of exogeneity. Section 6 constructs convenient practi-2

Linking Granger Causality and the Pearl Causal Model with Settable Systemscal tests for structural causality based on proposals of WL, using tests for G−causalityand conditional exogeneity. Section 7 contains a summary and concluding remarks.2. Pearl’s Causal ModelPearl’s definition of a causal model (Pearl, J., 2000, def. 7.1.1, p. 203) provides a formalstatement of elements supporting causal reasoning. The PCM is a triple M := (u,v, f ),where u := {u 1 ,...,u m } contains “background" variables determined outside the model,v := {v 1 ,...,v n } contains “endogenous" variables determined within the model, and f :={ f 1 ,..., f n } contains “structural" functions specifying how each endogenous variable isdetermined 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 containing every element of v but v i . The integers m and n are finite.The elements of u and v are system “units."Finally, the PCM requires that for each u, f yields a unique fixed point. 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 point requirement is crucial to the PCM, as this is necessary fordefining 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 in the PCM. A variantof the PCM (Halpern, J., 2000) does not require a fixed point, but if any exist, there maybe multiple collections of functions g yielding a fixed point. We call this a GeneralizedPearl Causal Model (GPCM). As GPCMs also do not possess an analog of the potentialresponse function in the absence of a unique fixed point, causal discourse in the GPCMis similarly restricted.In presenting 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 forsupporting causal discourse. As a simple game-theoretic example, consider a marketin which there are exactly two firms producing similar but not identical products (e.g.,Coke and Pepsi in the cola soft-drink market). Price determination in this market is atwo-player game known as “Bertrand duopoly."In deciding its price, each firm maximizes its profit, taking into account the prevailingcost and demand conditions it faces, as well as the price of its rival. A simplesystem representing price determination in 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 prevailing cost and demand conditions.3

White Chalak LuWe see that this maps directly to the PCM with n = 2, endogenous variables v =(p 1 , p 2 ), background variables u = (a 1 ,b 1 ,a 2 ,b 2 ), and structural functionsf 1 (v 2 ,u) = a 1 + b 1 p 2f 2 (v 1 ,u) = a 2 + b 2 p 1 .These functions are the Bertrand "best response" or "reaction" functions.Further, provided b 1 b 2 1, this system has a unique fixed point,p 1 = g 1 (u) = (a 1 + b 1 a 2 )/(1 − b 1 b 2 )p 2 = g 2 (u) = (a 2 + b 2 a 1 )/(1 − b 1 b 2 ).This fixed point represents the Nash equilibrium for this two-player game.Clearly, the PCM applies perfectly, supporting causal discourse for this Bertrandduopoly game. Specifically, we see that p 1 causes p 2 and vice-versa, and that the effectof p 2 on p 1 is b 1 , whereas that of p 1 on p 2 is b 2 .In fact, the PCM applies directly to a wide variety of games, provided that the gamehas a unique equilibrium. But there are many important cases where there may be noequilibrium or multiple equilibria. This limits the applicability of the PCM. We exploreexamples of this below, as well as other features of the PCM that limit its applicability.3. Settable Systems3.1. Why Settable Systems?WC motivate the development of the settable system (SS) framework as an extension ofthe PCM that accommodates optimization, equilibrium, and learning, which are centralfeatures of the explanatory structures of interest in economics. But these features areof interest more broadly, especially in machine learning, as optimization correspondsto any intelligent or rational behavior, whether artificial or natural; equilibrium (e.g.,Nash equilibrium) or transitions toward equilibrium characterize stable interactions betweenmultiple interacting systems; and learning corresponds to adaptation and evolutionwithin and between interacting systems. Given the prevalence of these features innatural and artificial systems, it is clearly desirable to provide means for explicit andrigorous causal discourse relating to systems with these features.To see why an extension of the PCM is needed to handle optimization, equilibrium,and learning, we consider a series of examples that highlight certain limiting featuresof the PCM: (i) in the absence of a unique fixed point, causal discourse is undefined;(ii) background variables play no causal role; (iii) the role of attributes is restricted;and (iv) only a finite rather than a countable number of units is permitted. WC discussfurther relevant aspects of the PCM, but these suffice for present purposes.Example 3.1 (Equilibria in Game Theory) Our first example concerns generaltwo-player games, extending the discussion that we began above in considering Bertrandduopoly.4

Linking Granger Causality and the Pearl Causal Model with Settable SystemsLet two players i = 1,2 have strategy sets S i and utility functions u i , such that π i =u i (z 1 ,z 2 ) gives player i’s payoff when player 1 plays z 1 ∈ S 1 and player 2 plays z 2 ∈ S 2 .Each player solves the optimization problemmax u i (z 1 ,z 2 ).z i ∈S iThe solution to this problem, when it exists, is player i’s best response, denotedy i = r e i (z (i);a),where ri e is player i’s best response function (the superscript “e" stands for “elementary,"conforming to notation formally introduced below); z (i) denotes the strategy played bythe player other than i; and a := (S 1 ,u 1 ,S 2 ,u 2 ) denotes given attributes defining thegame. For simplicity here, we focus on “pure strategy" games; see Gibbons, R. (1992)for an accessible introduction to game theory.Different configurations for a correspond to different games. For example, one ofthe most widely known games is prisoner’s dilemma, where two suspects in a crime areseparated and offered a deal: if one confesses and the other does not, the confessor isreleased and the other goes to jail. If both confess, both receive a mild punishment. Ifneither confesses, both are released. The strategies are whether to confess or not. Eachplayer’s utility is determined by both players’ strategies and the punishment structure.Another well known game is hide and seek. Here, player 1 wins by matching player2’s strategy and player 2 wins by mismatching player 1’s strategy. A familiar exampleis a penalty kick in soccer: the goalie wins by matching the direction (right or left) ofthe kicker’s kick; the kicker wins by mismatching the direction of the goalie’s lunge.The same structure applies to baseball (hitter vs. pitcher) or troop deployment in battle(aggressor vs. defender).A third famous game is battle of the sexes. In this game, Ralph and Alice are tryingto decide how to spend their weekly night out. Alice prefers the opera, and Ralphprefers boxing; but both would rather be together than apart.Now consider whether the PCM permits causal discourse in these games, e.g., aboutthe effect of one player’s action on that of the other. We begin by mapping the elementsof the game to the elements of the PCM. First, we see that a corresponds to PCM backgroundvariables u, as these are specified outside the system. The variables determinedwithin the system, i.e., the PCM endogenous variables are z := (z 1 ,z 2 ) corresponding tov, provided that (for now) we drop the distinction between y i and z i . Finally, we see thatthe best response functions ri e correspond to the PCM structural functions f i .To determine whether the PCM permits causal discourse in these games, we cancheck whether there is a unique fixed point for the best responses. In prisoner’s dilemma,there is indeed a unique fixed point (both confess), provided the punishments are suitablychosen. The PCM therefore applies to this game to support causal discourse. Butthere is no fixed point for hide and seek, so the PCM cannot support causal discoursethere. On the other hand, there are two fixed points for battle of the sexes: both Ralphand Alice choose opera or both choose boxing. The PCM does not support causal discoursethere either. Nor does the GPCM apply to the latter games, because even though5

White Chalak Luit does not require a unique fixed point, the potential response functions required forcausal discourse are not defined.The importance of game theory generally in describing the outcomes of interactionsof goal-seeking agents and the fact that the unique fixed point requirement prohibits thePCM from supporting causal discourse in important cases strongly motivates formulatinga causal framework that drops this requirement. As we discuss below, the SSframework does not require a unique fixed point, and it applies readily to games generally.Moreover, recognizing and enforcing the distinction between y i (i’s best responsestrategy) and z i (an arbitrary setting of i’s strategy) turns out to be an important componentto eliminating this requirement.Another noteworthy aspect of this example is that a is a fixed list of elements thatdefine the game. Although elements of a may differ across players, they do not vary fora given player. This distinction should be kept in mind when referring to the elementsof a as background “variables."Example 3.2 (Optimization in Consumer Demand) The neoclassical theory ofconsumer demand posits that consumers determine their optimal goods consumption bymaximizing utility subject to a budget constraint (see, e.g., Varian, H., 2009). Supposefor simplicity that there are just two goods, say beer and pizza. Then a typical consumersolves the problemmaxz 1 ,z 2U(z 1 ,z 2 ) s.t. m = z 1 + pz 2 ,where z 1 and z 2 represent quantities consumed of goods 1 (beer) and 2 (pizza) respectivelyand U is the utility function that embodies the consumer’s preferences for thetwo goods. For simplicity, let the price of a beer be $1, and let p represent the price ofpizza; m represents funds available for expenditure, “income" for short 1 . The budgetconstraint m = z 1 + pz 2 ensures that total expenditure on beer and pizza does not exceedincome (no borrowing) and also that total expenditure on beer and pizza is not less thanm. (As long as utility is increasing in consumption of the goods, it is never optimal toexpend less than the funds available.)Solving the consumer’s demand problem leads to the optimal consumer demandsfor beer and pizza, y 1 and y 2 . It is easy to show that these can be represented asy 1 = r a 1 (p,m;a) and y 2 = r a 2 (p,m;a),where r1 a and ra 2are known as the consumer’s market demand functions for beer andpizza. The “a" superscript stands for “agent," corresponding to notation formally introducedbelow. The attributes a include the consumer’s utility function U (preferences)and the admissible values for z 1 , z 2 , p, and m, e.g., R + := [0,∞).Now consider how this problem maps to the PCM. First, we see that a and (p,m)correspond to the background variables u, as these are not determined within the system.Next, we see that y := (y 1 ,y 2 ) corresponds to PCM endogenous variables v. Finally,1. Since a beer costs a dollar, it is the “numeraire," implying that income is measured in units of beer.This is a convenient convention ensuring that we only need to keep track of the price ratio betweenpizza and beer, p, rather than their two separate prices.6

Linking Granger Causality and the Pearl Causal Model with Settable Systemswe see that the consumer demand functions ri a correspond to the PCM structural functionsf i . Also, because the demand for beer, y 1 , does not enter the demand functionfor pizza, r2 a , and vice versa, there is a unique fixed point for this system of equations.Thus, the PCM supports causal discourse in this system.Nevertheless, this system is one where, in the PCM, the causal discourse naturalto economists is unavailable. Specifically, economists find it natural to refer to “priceeffects" and “income effects" on demand, implicitly or explicitly viewing price p andincome m as causal drivers of demand. For example, the pizza demand price effectis (∂/∂p)r2 a (p,m;a). This represents how much optimal pizza consumption (demand)will change as a result of a small (marginal) increase in the price of pizza. Similarly,the pizza demand income effect is (∂/∂m)r2 a (p,m;a), representing how much optimalpizza consumption will change as a result of a small increase in income. But in thePCM, causal discourse is reserved only for endogenous variables y 1 and y 2 . The factthat background variables p and m do not have causal status prohibits speaking abouttheir effects.Observe that the “endogenous" status of y and “exogenous" status of p and m isdetermined in SS by utility maximization, the “governing principle" here. In contrast,there is no formal mechanism in the PCM that permits making these distinctions. Althoughcausal discourse in the PCM can be rescued for such systems by “endogenizing"p and m, that is, by positing additional structure that explains the genesis of p and min terms of further background variables, this is unduly cumbersome. It is much morenatural simply to permit p and m to have causal status from the outset, so that price andincome effects are immediately meaningful, without having to specify their determiningprocesses. The SS framework embodies this direct approach. Those familiar withtheories of price and income determination will appreciate the considerable complicationsavoided in this way. The same simplifications occur with respect to the primitivevariables appearing in any responses determined by optimizing behavior.Also noteworthy here is the important distinction between a, which represents fixedattributes of the system, and p and m, which are true variables that can each take a rangeof different possible values. As WC (p.1774) note, restricting the role of attributesby “lumping together" attributes and structurally exogenous variables as backgroundobjects without causal status creates difficulties for causal discourse in the PCM:[this] misses the opportunity to make an important distinction betweeninvariant aspects of the system units on the one hand and counterfactualvariation admissible for the system unit values on the other. Among otherthings, assigning attributes to u interferes with assigning natural causalroles to structurally exogenous variables.By distinguishing between attributes and structurally exogenous variables, settable systemspermit causal status for variables determined outside a given system, such as whenprice and income drive consumer demand.Example 3.3 (Learning in Structural Vector Autoregressions) Structural vectorautoregressions (VARs) are widely used to analyze time-series data. For example,7

White Chalak Luconsider the structural VARy 1,t = a 11 y 1,t−1 + a 12 y 2,t−1 + u 1,ty 2,t = a 21 y 1,t−1 + a 22 y 2,t−1 + u 2,t , t = 1,2,...,where y 1,0 and y 2,0 are given scalars, a := (a 11 ,a 12 ,a 21 ,a 22 ) ′ is a given real “coefficient"vector, and {u t := (u 1,t ,u 2,t ) : t = 1,2,...} is a given sequence. This system describes theevolution of {y t := (y 1,t ,y 2,t ) : t = 1,2,...} through time.Now consider how this maps to the PCM. We see that y 0 := (y 1,0 ,y 2,0 ), {u t }, and acorrespond to the PCM background variables u, as these are not determined within thesystem. Further, we see that the sequence {y t } corresponds to the endogenous variablesv, and that the PCM structural functions f i correspond tor 1,t (y t−1 ,u t ;a) = a 11 y 1,t−1 + a 12 y 2,t−1 + u 1,tr 2,t (y t−1 ,u t ;a) = a 21 y 1,t−1 + a 22 y 2,t−1 + u 2,t , t = 1,2,...,where y t−1 := (y 0 ,...,y t−1 ) and u t := (u 1 ,...,u t ) represent finite “histories" of the indicatedvariables. We also see that this system is recursive, and therefore has a unique fixedpoint.The challenge to the PCM here is that it permits only a finite rather than a countablenumber of units: both the number of background variables (m) and endogenous variables(n) must be finite in the PCM, whereas the structural VAR requires a countableinfinity of background and endogenous variables. In contrast, settable systems permit(but do not require) a countable infinity of units, readily accommodating structuralVARs.In line with our previous discussion, settable systems distinguish between systemattributes a (a fixed vector) and structurally exogenous causal variables y 0 and {u t }. Thedifference in the roles of y 0 and {u t } on the one hand and a on the other are particularlyclear in this example. In the PCM, these are lumped together as background variablesdevoid of causal status. Since a is fixed, its lack of causal status is appropriate; indeed,a represents effects here 2 , not causes. But the lack of causal status is problematic forthe variables y 0 and {u t }; for example, this prohibits discussing the effects of structural“shocks" u t .Observe that the structural VAR represents u 1,t as a causal driver of y 1,t , as is standard.Nevertheless, settable systems do not admit “instantaneous" causation, so eventhough u 1,t has the same time index as y 1,t , i.e. t, we adopt the convention that u 1,t isrealized prior to y 1,t . That is, there must be some positive time interval δ > 0, no matterhow small, separating these realizations. For example, δ can represent the amount oftime it takes to compute y 1,t once all its determinants are in place. Strictly speaking,then, we could write u 1,t−δ in place of u 1,t , but for notational convenience, we leavethis implicit. We refer to this as “contemporaneous" causation to distinguish it frominstantaneous causation.2. For example, (∂/∂y 1,t−1 )r 1,t (y t−1 ,e t ;a) = a 11 can be interpreted as the marginal effect of y 1,t−1 on y 1,t .8

Linking Granger Causality and the Pearl Causal Model with Settable SystemsA common focus of interest when applying structural VARs is to learn the coefficientvector a. In applications, it is typically assumed that the realizations {y t } areobserved, whereas {u t } is unobserved. The least squares estimator for a sample of sizeT, say â T , is commonly used to learn (estimate) a in such cases. This estimator is astraightforward function of y T , say â T = r a,T (y T ). If {u t } is generated as a realizationof a sequence of mean zero finite variance independent identically distributed (IID)random variables, then â T generally converges to a with probability one as T → ∞,implying that a can be fully learned in the limit. Viewing â T as causally determined byy T , we see that we require a countable number of units to treat this learning problem.As these examples demonstrate, the PCM exhibits a number of features that limitits applicability to systems involving optimization, equilibrium, and learning. Theselimitations motivate a variety of features of settable systems, extending the PCM inways that permit straightforward treatment of such systems. We now turn to a morecomplete description of the SS framework.3.2. Formal Settable SystemsWe now provide a formal description of settable systems that readily accommodatescausal discourse in the foregoing examples and that also suffices to establish the desiredlinkage between Granger causality and causal notions in the PCM. The material thatfollows is adapted from Chalak, K. and H. White (2010). For additional details, seeWC.A stochastic settable system is a mathematical framework in which a countablenumber of units i, i = 1,...,n, interact under uncertainty. Here, n ∈ ¯N + := N + ∪{∞}, whereN + denotes the positive integers. When n = ∞, we interpret i = 1,...,n as i = 1,2,....Units have attributes a i ∈ A; these are fixed for each unit, but may vary across units.Each unit also has associated random variables, defined on a measurable space (Ω,F ).It is convenient to define a principal space Ω 0 and let Ω := × n i=0 Ω i, with each Ω i acopy of Ω 0 . Often, Ω 0 = R is convenient. A probability measure P a on (Ω,F ) assignsprobabilities to events involving random variables. As the notation suggests, P a candepend on the attribute vector a := (a 1 ,...,a n ) ∈ A := × n i=1 A.The random variables associated with unit i define a settable variable X i for thatunit. A settable variable X i has a dual aspect. It can be set to a random variabledenoted by Z i (the setting), where Z i : Ω i → S i . S i denotes the admissible setting valuesfor Z i , a multi-element subset of R. Alternatively, the settable variable can be freeto respond to settings of other settable variables. In the latter case, it is denoted bythe response Y i : Ω → S i . The response Y i of a settable variable X i to the settingsof other settable variables is determined by a response function, r i . For example, r ican be determined by optimization, determining the response for unit i that is best insome sense, given the settings of other settable variables. The dual role of a settablevariable X i : {0,1} × Ω → S i , distinguishing responses X i (0,ω) := Y i (ω) and settingsX i (1,ω) := Z i (ω i ), ω ∈ Ω, permits formalizing the directional nature of causal relations,whereby settings of some variables (causes) determine responses of others.9

White Chalak LuThe principal unit i = 0 also plays a key role. We let the principal setting Z 0 andprincipal response Y 0 of the principal settable variable X 0 be such that Z 0 : Ω 0 → Ω 0is the identity map, Z 0 (ω 0 ) := ω 0 , and we define Y 0 (ω) := Z 0 (ω 0 ). The setting Z 0 ofthe principal settable variable may directly influence all other responses in the system,whereas its response Y 0 is unaffected by other settings. Thus, X 0 supports introducingan aspect of “pure randomness” to responses of settable variables.3.2.1. Elementary Settable SystemsIn elementary settable systems, Y i is determined (actually or potentially) by the settingsof all other system variables, denoted Z (i) . Thus, in elementary settable systems, Y i =r i (Z (i) ;a). The settings Z (i) take values in S (i) ⊆ Ω 0 × ji S j . We have that S (i) is a strictsubset of Ω 0 × ji S j if there are joint restrictions on the admissible settings values, forexample, when certain elements of S (i) represent probabilities that sum to one.We now give a formal definition of elementary settable systems.Definition 3.1 (Elementary Settable System) Let A be a set and let attributes a ∈ Abe given. Let n ∈ ¯N + be given, and let (Ω,F , P a ) be a complete probability space suchthat Ω := × n i=0 Ω i, with each Ω i a copy of the principal space Ω 0 , containing at leasttwo elements.Let the principal setting Z 0 : Ω 0 → Ω 0 be the identity mapping. For i = 1,2,...,n,let S i be a multi-element Borel-measurable subset of R and let settings Z i : Ω i → S i besurjective measurable functions. Let Z (i) be the vector including every setting except Z iand taking values in S (i) ⊆ Ω 0 × ji S j , S (i) ∅. Let response functions r i ( · ;a) : S (i) → S ibe measurable functions and define responses Y i (ω) := r i (Z (i) (ω);a). Define settablevariables X i : {0,1} × Ω → S i asX i (0,ω) := Y i (ω) and X i (1,ω) := Z i (ω i ), ω ∈ Ω.Define Y 0 and X 0 by Y 0 (ω) := X 0 (0,ω) := X 0 (1,ω) := Z 0 (ω 0 ), ω ∈ Ω.Put X := {X 0 ,X 1 ,...}. The triple S := {(A,a),(Ω,F , P a ),X} is an elementary settablesystem.An elementary settable system thus comprises an attribute component, (A,a), astochastic component, (Ω,F , P a ), and a structural or causal component X, consistingof settable variables whose properties are crucially determined by response functionsr := {r i }. It is formally correct to write X a instead of X; we write X for simplicity.Note the absence of any fixed point requirement, the distinct roles played by fixedattributes a and setting variables Z i (including principal settings Z 0 ), and the countablenumber of units allowed.Example 3.1 is covered by this definition. There, n = 2. Attributes a := (S 1 ,u 1 ,S 2 ,u 2 )belong to a suitably chosen set A. Here, S i = S i . We take z i = Z i (ω i ), ω i ∈ Ω i andy i = Y i (ω) = r e i (Z (i)(ω);a) = r e i (z (i);a), i = 1,2. The “e" superscript in r e iemphasizes thatthe response function is for an elementary settable system. In the example games, theresponses y i only depend on settings (z 1 ,z 2 ). In more elaborate games, dependence onz 0 = ω 0 can accommodate random responses.10

Linking Granger Causality and the Pearl Causal Model with Settable Systems3.2.2. Partitioned Settable SystemsIn elementary settable systems, each single response Y i can freely respond to settingsof all other system variables. We now consider systems where several settable variablesjointly respond to settings of the remaining settable variables, as when responses representthe solution to a joint optimization problem. For this, partitioned settable systemsgroup jointly responding variables into blocks. In elementary settable systems, everyunit i forms a block by itself. We now define general partitioned settable systems.Definition 3.2 (Partitioned Settable System) Let (A,a),(Ω,F , P a ), X 0 , n, and S i , i =1,...,n, be as in Definition 3.1. Let Π = {Π b } be a partition of {1,...,n}, with cardinalityB ∈ ¯N + (B := #Π).For i = 1,2,...,n, let ZiΠZi Π ,i Π b , and taking values in S Π (b) ⊆ Ω 0 × iΠb S i , S Π (b)be settings and let Z Π (b) be the vector containing Z 0 and ∅, b = 1,..., B. For b = 1,..., Band i ∈ Π b , suppose there exist measurable functions r Π i ( · ;a) : SΠ (b) → S i, specific to Πsuch that responses Y Π i(ω) are jointly determined asY Π i:= r Π i (ZΠ (b) ;a).Define the settable variables X Π i: {0,1} × Ω → S i asX Π i (0,ω) := YΠ i (ω) and X Π i (1,ω) := ZΠ i (ω i ) ω ∈ Ω.Put X Π := {X 0 ,X Π 1 ,XΠ 2 ...}. The triple S := {(A,a),(Ω,F ),(Π,XΠ )} is a partitionedsettable system.The settings Z(b) Π may be partition-specific; this is especially relevant when the admissibleset S Π (b) imposes restrictions on the admissible values of ZΠ (b). Crucially, responsefunctions and responses are partition-specific. In Definition 3.2, the joint responsefunction r[b] Π := (rΠ i ,i ∈ Π b) specifies how the settings Z(b) Π outside of block Π bdetermine the joint response Y[b] Π := (YΠ i,i ∈ Π b ), i.e., Y[b] Π = rΠ [b] (ZΠ (b);a). For conveniencebelow, we let Π 0 = {0} represent the block corresponding to X 0 .Example 3.2 makes use of partitioning. Here, we have n = 4 settable variables withB = 2 blocks. Let settable variables 1 and 2 correspond to beer and pizza consumption,respectively, and let settable variables 3 and 4 correspond to price and income. Theagent partition groups together all variables under the control of a given agent. Let theconsumer be agent 2, so Π 2 = {1,2}. Let the rest of the economy, determining price andincome, be agent 1, so Π 1 = {3,4}. The agent partition is Π a = {Π 1 ,Π 2 }. Then for block2,y 1 = Y a 1 (ω) = ra 1 (Z 0(ω 0 ),Z a 3 (ω 3),Z a 4 (ω 4);a) = r a 1 (p,m;a)y 2 = Y a 2 (ω) = ra 2 (Z 0(ω 0 ),Z a 3 (ω 3),Z a 4 (ω 4);a) = r a 2 (p,m;a)represents the joint demand for beer and pizza (belonging to block 2) as a function ofsettings of price and income (belonging to block 1). This joint demand is unique under11

White Chalak Lumild conditions. Observe that z 0 = Z 0 (ω 0 ) formally appears as an allowed argumentof ria after the second equality, but when the consumer’s optimization problem has aunique solution, there is no need for a random component to demand. We thus suppressthis argument in writing ri a (p,m;a), i = 1,2. Nevertheless, when the solution to theconsumer’s optimization problem is not unique, a random component can act to ensurea unique consumer demand. We do not pursue this here; WC provide related discussion.We write the block 1 responses for the price and income settable variables asy 3 = Y a 3 (ω) = ra 3 (Z 0(ω 0 ),Z a 1 (ω 1),Z a 2 (ω 2);a) = r a 3 (z 0;a)y 4 = Y a 4 (ω) = ra 4 (Z 0(ω 0 ),Z a 1 (ω 1),Z a 2 (ω 2);a) = r a 4 (z 0;a).In this example, price and income are not determined by the individual consumer’sdemands, so although Z a 1 (ω 1) and Z a 2 (ω 2) formally appear as allowed arguments of r a iafter the second equality, we suppress these in writing r a i (z 0;a), i = 3,4. Here, priceand income responses (belonging to block 1) are determined solely by block 0 settingsz 0 = Z 0 (ω 0 ) = ω 0 . This permits price and income responses to be randomly distributed,under the control of P a .It is especially instructive to consider the elementary partition for this example,Π e = {{1},{2},{3},{4}}, so that Π i = {i}, i = 1,...,4. The elementary partition specifieshow each system variable freely responds to settings of all other system variables. Inparticular, it is easy to verify that when consumption of pizza is set to a given level,the consumer’s optimal response is to spend whatever income is left on beer, and viceversa. Thus,y 1 = r1 e (Z 0(ω 0 ),Z2 e (ω 2),Z3 e (ω 3),Z4 e (ω 4);a) = r1 e (z 2, p,m;a) = m − pz 2y 2 = r2 e (Z 0(ω 0 ),Z1 e (ω 2),Z3 e (ω 3),Z4 e (ω 4);a) = r2 e (z 1, p,m;a) = (m − z 1 )/p.Replacing (y 1 ,y 2 ) with (z 1 ,z 2 ), we see that this system does not have a unique fixedpoint, as any (z 1 ,z 2 ) such that m = z 1 + pz 2 satisfies bothz 1 = m − pz 2 and z 2 = (m − z 1 )/p.Causal discourse in the PCM is ruled out by the lack of a fixed point. Nevertheless,the settable systems framework supports the natural economic causal discourse hereabout effects of prices, income, and, e.g., pizza consumption on beer demand. Further,in settable systems, the governing principle of optimization (embedded in a) ensuresthat the response functions for both the agent partition and the elementary partition aremutually consistent.3.2.3. Recursive and Canonical Settable SystemsThe link between Granger causality and the causal notions of the PCM emerges from aparticular class of recursive partitioned settable systems that we call canonical settablesystems, where the system evolves naturally without intervention. This corresponds towhat are also called “idle regimes" in the literature (Pearl, J., 2000; Eichler, M. and V.Didelez, 2009; Dawid, 2010).12

Linking Granger Causality and the Pearl Causal Model with Settable SystemsTo define recursive settable systems, for b ≥ 0 define Π [0:b] := Π 0 ∪ ... ∪ Π b−1 ∪ Π b .Definition 3.3 (Recursive Partitioned Settable System) Let S be a partitioned settablesystem. For b = 0,1,..., B, let Z[0:b] Π denote the vector containing the settings ZΠ ifor i ∈ Π [0:b] and taking values in S [0:b] ⊆ Ω 0 × i∈Π[1:b] S i , S [0:b] ∅. For b = 1,..., B andi ∈ Π b , suppose that r Π := {ri Π} is such that the responses YΠ i= X Π i(1,·) are determinedas:= ri Π (ZΠ [0:b−1] ;a).Y Π iThen we say that Π is a recursive partition, that r Π is recursive, and that S :={(A,a),(Ω,F ), (Π,X Π )} is a recursive partitioned settable system or simply that S isrecursive.Example 3.2 is a recursive settable system, as the responses of block 1 depend onthe settings of block 0, and the responses of block 2 depend on the settings of block 1.Canonical settable systems are recursive settable systems in which the settings fora given block equal the responses for that block, i.e.,Z[b] Π = YΠ [b] := rΠ [b] (ZΠ [0:b−1];a), b = 1,..., B.Without loss of generality, we can represent canonical responses and settings solely asa function of ω 0 , so thatZ Π [b] (ω 0) = Y Π [b] (ω 0) := r Π [b] (ZΠ [0:b−1] (ω 0);a), b = 1,..., B.The canonical representation drops the distinction between settings and responses; wewriteY[b] Π = rΠ [b] (YΠ [0:b−1];a), b = 1,..., B.It is easy to see that the structural VAR of Example 3.3 corresponds to the canonicalrepresentation of a canonical settable system. The canonical responses y 0 and {u t }belong to the first block, and canonical responses y t = (y 1,t ,y 2,t ) belong to block t + 1,t = 1,2,... Example 3.3 implements the time partition, where joint responses for a giventime period depend on previous settings.4. Causality in Settable Systems and in the PCMIn this section we examine the relations between concepts of direct causality in settablesystems and in the PCM, specifically the PCM notions of direct cause and controlleddirect effect (Pearl, J. (2000, p. 222); Pearl, J. (2001, definition 1)). The close correspondencebetween these notions for the recursive systems relevant to Granger causalityenables us to take the first step in linking Granger causality and causal notions in thePCM. Section 5 completes the chain by linking direct structural causality and Grangercausality.13

White Chalak Lu4.1. Direct Structural Causality in Settable SystemsDirect structural causality is defined for both recursive and non-recursive partitionedsettable systems. For notational simplicity in what follows, we may drop the explicitpartition superscript Π when the specific partition is clearly understood. Thus, we maywrite Y, Z, and X in place of the more explicit Y Π , Z Π , and X Π when there is no possibilityof confusion.Let X j belong to block b ( j ∈ Π b ). Heuristically, we say that a settable variableX i , outside of block b, directly causes X j in S when the response for X j differs fordifferent settings of X i , while holding all other variables outside of block b to the samesetting values. There are two main ingredients to this notion. The first ingredient is anadmissible intervention. To define this, let z * (b);idenote the vector otherwise identicalto z (b) , but replacing elements z i with z * i . An admissible intervention z (b) → z * (b);i :=(z (b) ,z * (b);i ) is a pair of distinct elements of S (b). The second ingredient is the behavior ofthe response under this intervention.We formalize this notion of direct causality as follows.Definition 4.1 (Direct Causality) Let S be a partitioned settable system. For givenpositive integer b, let j ∈ Π b . (i) For given i Π b , X i directly causes X j in S if thereexists an admissible intervention z (b) → z * (b);isuch thatr j (z * (b);i ;a) − r j(z (b) ;a) 0,and we write X iD⇒S X j . Otherwise, we say X i does not directly cause X j in S andwrite X iDS X j . (ii) For i, j ∈ Π b ,X iDS X j .We emphasize that although we follow the literature in referring to “interventions,”with their mechanistic or manipulative connotations, the formal concept only involvesthe properties of a response function on its domain.By definition, variables within the same block do not directly cause each other. Inparticular X iDS X i . Also, Definition 4.1 permits mutual causality, so that X iD⇒S X jand X jD⇒S X i without contradiction for i and j in different blocks. Nevertheless, inrecursive systems, mutual causality is ruled out: if X iD⇒S X j then X jDS X i .We call the response value difference in Definition 4.1 the direct effect of X i on X jin S of the specified intervention. Chalak, K. and H. White (2010) also study variousnotions of indirect and total causality.These notions of direct cause and direct effect are well defined regardless of whetheror not the system possesses a unique fixed point. Further, all settable variables, includingX 0 , can act as causes and have effects. On the other hand, attributes a, being fixed,do not play a causal role. These definitions apply regardless of whether there is a finiteor countable number of units. It is readily verified that this definition rigorouslysupports causal discourse in each of the examples of Section 3.As we discuss next, in the recursive systems relevant for G−causality, these conceptscorrespond closely to notions of direct cause and “controlled” direct effect in14

Linking Granger Causality and the Pearl Causal Model with Settable SystemsPearl, J. (2000, 2001). To distinguish the settable system direct causality conceptfrom Pearl’s notion and later from Granger causality, we follow WL and refer to directcausality in settable systems as direct structural causality.4.2. Direct Causes and Effects in the PCMPearl, J. (2000, p. 222), drawing on Galles, D. and J. Pearl (1997), gives a succinctstatement of the notion of direct cause, coherent with the PCM as specified in Section2:X is a direct cause of Y if there exist two values x and x ′ of X and a valueu of U such that Y xr (u) Y x ′ r(u), where r is some realization of V∖{X,Y}.To make this statement fully meaningful requires applying Pearl’s (2000) definitions7.1.2 (Submodel) and 7.1.4 (Potential Response) to arrive at the potential response,Y xr (u). For brevity, we do not reproduce Pearl’s definitions here. Instead, it sufficesto map Y xr (u) and its elements to their settable system counterparts. Specifically, ucorresponds to (a,z 0 ); x corresponds to z i ; r corresponds to the elements of z (b) otherthan z 0 and z i , say z (b)(i,0) ; and, provided it exists, Y xr (u) corresponds to r j (z (b) ;a).The caveat about the existence of Y xr (u) is significant, as Y xr (u) is not defined inthe absence of a unique fixed point for the system. Further, even with a unique fixedpoint, the potential response Y xr (u) must also uniquely solve a set of equations denotedF x (see Pearl, J., 2000, eq. (7.1)) for a submodel, and there is no general guarantee ofsuch a solution. Fortunately, however, this caveat matters only for non-recursive PCMs.In the recursive case relevant for G−causality, the potential response is generally welldefined.Making a final identification between x ′ and z * i, and given the existence of potentialresponses Y x ′ r(u) and Y xr (u), we see that Y x ′ r(u) Y xr (u) corresponds to the settablesystems requirement r j (z * (b);i ;a) − r j(z (b) ;a) 0.Pearl, J. (2001, definition 1) gives a formal statement of the notion stated above,saying that if for given u and some r, x, and x ′ we have Y xr (u) Y x ′ r(u) then X has acontrolled direct effect on Y in model M and situation U = u. In definition 2, Pearl,J. (2001) labels Y x ′ r(u) − Y xr (u) the controlled direct effect, corresponding to the directstructural effect r j (z * (b);i ;a) − r j(z (b) ;a) defined for settable systems.Thus, although there are important differences, especially in non-recursive systems,the settable systems and PCM notions of direct causality and direct effects closely correspondin recursive systems. These differences are sufficiently modest that the resultsof WL linking direct structural causality to Granger causality, discussed next, also serveto closely link the PCM notion of direct cause to that of Granger causality.5. G−Causality and Direct Structural CausalityIn this section we examine the relation between direct structural causality and Grangercausality, drawing on results of WL. See WL for additional discussion and proofs of allformal results given here and in Section 6.15

White Chalak Lu5.1. Granger CausalityGranger, C.W.J. (1969) defined G−causality in terms of conditional expectations. Granger,C.W.J. and P. Newbold (1986) gave a definition using conditional distributions. Wework with the latter, as this is what relates generally to structural causality. In what follows,we adapt Granger and Newbold’s notation, but otherwise preserve the conceptualcontent.For any sequence of random vectors {Y t , t = 0,1,...}, let Y t := (Y 0 ,...,Y t ) denoteits “t−history," and let σ(Y t ) denote the sigma-field (“information set") generated byY t . Let {Q t ,S t ,Y t } be a sequence of random vectors. Granger, C.W.J. and P. Newbold(1986) say that Q t−1 does not G-cause Y t+k with respect to σ(Q t−1 ,S t−1 ,Y t−1 ) if for allt = 0,1,...,F t+k ( · | Q t−1 ,S t−1 ,Y t−1 ) = F t+k ( · | S t−1 ,Y t−1 ), k = 0,1,..., (2)where F t+k ( · | Q t−1 ,S t−1 ,Y t−1 ) denotes the conditional distribution function of Y t+kgiven Q t−1 ,S t−1 ,Y t−1 , and F t+k ( · | S t−1 ,Y t−1 ) denotes that of Y t+k given S t−1 ,Y t−1 .Here, we focus only on the k = 0 case, as this is what relates generally to structuralcausality.As Florens, J.P. and M. Mouchart (1982) and Florens, J.P. and D. Fougère (1996)note, G non-causality is a form of conditional independence. Following Dawid (1979),we write X ⊥ Y | Z when X and Y are independent given Z. Translating (2) gives thefollowing version of the classical definition of Granger causality:Definition 5.1 (Granger Causality) Let {Q t ,S t ,Y t } be a sequence of random vectors.Suppose thatY t ⊥ Q t−1 | Y t−1 ,S t−1 t = 1,2,... . (3)Then Q does not G−cause Y with respect to S. Otherwise, Q G−causes Y with respectto S.As it stands, this definition has no necessary structural content, as Q t , S t , and Y t can beany random variables whatsoever. This definition relates solely to the ability of Q t−1 tohelp in predicting Y t given Y t−1 and S t−1 .In practice, researchers do not test classical G−causality, as this involves data historiesof arbitrary length. Instead, researchers test a version of G−causality involvingonly a finite number of lags of Y t , Q t , and S t . This does not test classical G−causality,but rather a related property, finite-order G−causality, that is neither necessary nor sufficientfor classical G−causality.Because of its predominant practical relevance, we focus here on finite-order ratherthan classical G−causality. (See WL for discussion of classical G−causality.) To definethe finite-order concept, we define the finite histories Y t−1 := (Y t−l ,...,Y t−1 ) and Q t :=(Q t−k ,..., Q t ).Definition 5.2 (Finite-Order Granger Causality) Let {Q t ,S t ,Y t } be a sequence ofrandom variables, and let k ≥ 0 and l ≥ 1 be given finite integers. Suppose thatY t ⊥ Q t | Y t−1 ,S t , t = 1,2,... .16

Linking Granger Causality and the Pearl Causal Model with Settable SystemsThen we say Q does not finite-order G−cause Y with respect to S . Otherwise, we sayQ finite-order G−causes Y with respect to S.We call max(k,l − 1) the “order" of the finite-order G non-causality.Observe that Q t replaces Q t−1 in the classical definition, that Y t−1 replaces Y t−1 , andthat S t replaces S t−1 . Thus, in addition to dropping all but a finite number of lags in Q t−1and Y t−1 , this version includes Q t . As WL discuss, however, the appearance of Q t neednot involve instantaneous causation. It suffices that realizations of Q t precede thoseof Y t , as in the case of contemporaneous causation discussed above. The replacementof S t−1 with S t entails first viewing S t as representing a finite history, and second therecognition that since S t plays purely a conditioning role, there need be no restrictionwhatever on its timing. We thus call S t “covariates." As WL discuss, the covariatescan even include leads relative to time t. When covariate leads appear, we call this the“retrospective" case.In what follows, when we refer to G−causality, it will be understood that we arereferring to finite-order G−causality, as just defined. We will always refer to the conceptof Definition 5.1 as classical G−causality to avoid confusion.5.2. A Dynamic Structural SystemWe now specify a canonical settable system that will enable us to examine the relationbetween G−causality and direct structural causality. As described above, in suchsystems “predecessors" structurally determine “successors," but not vice versa. In particular,future variables cannot precede present or past variables, enforcing the causaldirection of time. We write Y ⇐ X to denote that Y succeeds X (X precedes Y ). When Yand X have identical time indexes, Y ⇐ X rules out instantaneous causation but allowscontemporaneous causation.We now specify a version of the causal data generating structures analyzed by WLand White, H. and P. Kennedy (2009). We let N denote the integers {0,1,...} and define¯N := N ∪ {∞}. For given l,m,∈ N, l ≥ 1, we let Y t−1 := (Y t−l ,...,Y t−1 ) as above; we alsodefine Z t := (Z t−m ,...,Z t ). For simplicity, we keep attributes implicit in what follows.Assumption A.1 Let {U t ,W t ,Y t ,Z t ; t = 0,1,...} be a stochastic process on (Ω,F , P), acomplete probability space, with U t ,W t ,Y t , and Z t taking values in R k u,R k w,R k y, andR k zrespectively, where k u ∈ ¯N and k w ,k y ,k z ∈ N, with k y > 0. Further, suppose thatY t ⇐ (Y t−1 ,U t ,W t ,Z t ), where, for an unknown measurable k y × 1 function q t , and forgiven l,m,∈ N,l ≥ 1, {Y t } is structurally generated assuch that, with Y t := (Y ′ 1,t ,Y′ 2,t )′ and U t := (U ′ 1,t ,U′ 2,t )′ ,Y t = q t (Y t−1 , Z t ,U t ), t = 1,2,..., (4)Y 1,t = q 1,t (Y t−1 , Z t ,U 1,t ) Y 2,t = q 2,t (Y t−1 , Z t ,U 2,t ).Such structures are well suited to representing the structural evolution of time-seriesdata in economic, biological, or other systems. Because Y t is a vector, this covers the17

White Chalak Lucase of panel data, where one has a cross-section of time-series observations, as in fMRIor EEG data sets. For practical relevance, we explicitly impose the Markov assumptionthat Y t is determined by only a finite number of its own lags and those of Z t and U t . WLdiscuss the general case.Throughout, we suppose that realizations of W t ,Y t , and Z t are observed, whereasrealizations of U t are not. Because U t ,W t , or Z t may have dimension zero, their presenceis optional. Usually, however, some or all will be present. Since there may be acountable infinity of unobservables, there is no loss of generality in specifying that Y tdepends only on U t rather than on a finite history of lags of U t .This structure is general: the structural relations may be nonlinear and non-monotonicin their arguments and non-separable between observables and unobservables. This systemmay generate stationary processes, non-stationary processes, or both. AssumptionA.1 is therefore a general structural VAR; Example 3.3 is a special case.The vector Y t represents responses of interest. Consistent with a main applicationof G−causality, our interest here attaches to the effects on Y 1,t of the lags of Y 2,t . Wethus call Y 2,t−1 and its further lags “causes of interest." Note that A.1 specifies that Y 1,tand Y 2,t each have their own unobserved drivers, U 1,t and U 2,t , as is standard.The vectors U t and Z t contain causal drivers of Y t whose effects are not of primaryinterest; we thus call U t and Z t “ancillary causes." The vector W t may contain responsesto U t . Observe that W t does not appear in the argument list for q t , so it explicitlydoes not directly determine Y t . Note also that Y t ⇐ (Y t−1 ,U t ,W t ,Z t ) ensures that W t isnot determined by Y t or its lags. A useful convention is that W t ⇐ (W t−1 ,U t ,Z t ), sothat W t does not drive unobservables. If a structure does not have this property, thensuitable substitutions can usually yield a derived structure satisfying this convention.Nevertheless, we do not require this, so W t may also contain drivers of unobservablecauses of Y t .For concreteness, we now specialize the settable systems definition of direct structuralcausality (Definition 4.1) to the specific system given in A.1. For this, let y s,t−1 bethe sub-vector of y t−1 with elements indexed by the non-empty set s ⊆ {1,...,k y } × {t −l,...,t − 1}, and let y (s),t−1 be the sub-vector of y t−1 with elements of s excluded.Definition 5.3 (Direct Structural Causality) Given A.1, for given t > 0, j ∈ {1,...,k y },and s, suppose that for all admissible values of y (s),t−1 , z t , and u t , the function y s,t−1 →q j,t (y t−1 , z t ,u t ) is constant in y s,t−1 . Then we say Y s,t−1 does not directly structurallyDcause Y j,t and write Y s,t−1 S Y j,t . Otherwise, we say Y s,t−1 directly structurallycauses Y j,t and write Y s,t−1D⇒S Y j,t .We can similarly define direct causality or non-causality of Z s,t or U s,t for Y j,t , butwe leave this implicit. We write, e.g., Y s,t−1D⇒S Y t when Y s,t−1D⇒S Y j,t for somej ∈ {1,...,k y }.Building on work of White, H. (2006a) and White, H. and K. Chalak (2009), WLdiscuss how certain exogeneity restrictions permit identification of expected causal effectsin dynamic structures. Our next result shows that a specific form of exogeneity18

Linking Granger Causality and the Pearl Causal Model with Settable Systemsenables us to link direct structural causality and finite order G−causality. To state thisexogeneity condition, we write Y 1,t−1 := (Y 1,t−l ,...,Y 1,t−1 ), Y 2,t−1 := (Y 2,t−l ,...,Y 2,t−1 ),and, for given τ 1 ,τ 2 ≥ 0, X t := (X t−τ1 ,..., X t+τ2 ), where X t := (W ′ t ,Z′ t )′ .Assumption A.2 For l and m as in A.1 and for τ 1 ≥ m,τ 2 ≥ 0, suppose that Y 2,t−1 ⊥U 1,t | (Y 1,t−1 , X t ), t = 1,...,T − τ 2 .The classical strict exogeneity condition specifies that (Y t−1 , Z t ) ⊥ U 1,t , which impliesY 2,t−1 ⊥ U 1,t | (Y 1,t−1 , Z t ). (Here, W t can be omitted.) Assumption A.2 is a weakerrequirement, as it may hold when strict exogeneity fails. Because of the conditioninginvolved, we call this conditional exogeneity. Chalak, K. and H. White (2010) discussstructural restrictions for canonical settable systems that deliver conditional exogeneity.Below, we also discuss practical tests for this assumption.Because of the finite numbers of lags involved in A.2, this is a finite-order conditionalexogeneity assumption. For convenience and because no confusion will arisehere, we simply refer to this as “conditional exogeneity."Assumption A.2 ensures that expected direct effects of Y 2,t−1 on Y 1,t are identified.As WL note, it suffices for A.2 that U t−1 ⊥ U 1,t | (Y 0 ,Z t−1 , X t ) and Y 2,t−1 ⊥ (Y 0 ,Z t−τ 1−1 ) |(Y 1,t−1 , X t ). Imposing U t−1 ⊥ U 1,t | (Y 0 ,Z t−1 , X t ) is the analog of requiring that serialcorrelation is absent when lagged dependent variables are present. Imposing Y 2,t−1 ⊥(Y 0 ,Z t−τ 1−1 ) | (Y 1,t−1 , X t ) ensures that ignoring Y 0 and omitting distant lags of Z t fromX t doesn’t matter.Our first result linking direct structural causality and G−causality shows that, givenA.1 and A.2 and with proper choice of Q t and S t , G−causality implies direct structuralcausality.Proposition 5.4 Let A.1 and A.2 hold. If Y 2,t−1finite order G−cause Y 1 with respect to X, i.e.,DS Y 1,t , t = 1,2,..., then Y 2 does notY 1,t ⊥ Y 2,t−1 | Y 1,t−1 , X t , t = 1,...,T − τ 2 .In stating G non-causality, we make the explicit identifications Q t = Y 2,t−1 and S t = X t .This result leads one to ask whether the converse relation also holds: does directstructural causality imply G−causality? Strictly speaking, the answer is no. WL discussseveral examples. The main issue is that with suitably chosen causal and probabilisticrelationships, Y 2,t−1 can cause Y 1,t , but Y 2,t−1 and Y 1,t can be independent, conditionallyor unconditionally, i.e. Granger non-causal.As WL further discuss, however, these examples are exceptional, in the sense thatmild perturbations to their structure destroy the Granger non-causality. WL introduce arefinement of the notion of direct structural causality that accommodates these specialcases and that does yield a converse result, permitting a characterization of structuraland Granger causality. Let supp(Y 1,t ) denote the support of Y 1,t , i.e., the smallest setcontaining Y 1,t with probability 1, and let F 1,t (· | Y 1,t−1 , X t ) denote the conditional distributionfunction of U 1,t given Y 1,t−1 , X t . WL introduce the following definition:19

White Chalak LuDefinition 5.5 Suppose A.1 holds and that for given τ 1 ≥ m,τ 2 ≥ 0 and for each y ∈supp(Y 1,t ) there exists a σ(Y 1,t−1 , X t )−measurable version of the random variable∫︁1{q 1,t (Y t−1 , Z t ,u 1,t ) < y} dF 1,t (u 1,t | Y 1,t−1 , X t ).DThen Y 2,t−1 S(Y1,t−1 ,X t ) Y 1,t (direct non-causality−σ(Y 1,t−1 , X t ) a.s.). If not, Y 2,t−1D⇒ S(Y1,t−1 ,X t ) Y 1,t .For simplicity, we refer to this as direct non-causality a.s. The requirement that theintegral in this definition is σ(Y 1,t−1 , X t )−measurable means that the integral does notdepend on Y 2,t−1 , despite its appearance inside the integral as an argument of q 1,t . Forthis, it suffices that Y 2,t−1 does not directly cause Y 1,t ; but it is also possible that q 1,tand the conditional distribution of U 1,t given Y 1,t−1 , X t are in just the right relation tohide the structural causality. Without the ability to manipulate this distribution, thestructural causality will not be detectable. One possible avenue to manipulating thisdistribution is to modify the choice of X t , as there are often multiple choices for X tthat can satisfy A.2 (see White, H. and X. Lu, 2010b). For brevity and because hiddenstructural causality is an exceptional circumstance, we leave aside further discussion ofthis possibility here. The key fact to bear in mind is that the causal concept of Definition5.5 distinguishes between those direct causal relations that are empirically detectableand those that are not, for a given set of covariates X t .We now give a structural characterization of G−causality for structural VARs:Theorem 5.6 Let A.1 and A.2 hold. Then Y 2,t−1DS(Y1,t−1 ,X t ) Y 1,t , t = 1,...,T− τ 2 , if andonly ifY 1,t ⊥ Y 2,t−1 | Y 1,t−1 , X t , t = 1,...,T − τ 2 ,i.e., Y 2 does not finite-order G−cause Y 1 with respect to X.Thus, given conditional exogeneity of Y 2,t−1 , G non-causality implies direct non-causalitya.s. and vice-versa, justifying tests of direct non-causality a.s. in structural VARs usingtests for G−causality.This result completes the desired linkage between G−causality and direct causalityin the PCM. Because direct causality in the recursive PCM corresponds essentiallyto direct structural causality in canonical settable systems, and because the latter isessentially equivalent to G−causality, as just shown, direct causality in the PCM isessentially equivalent to G−causality, provided A.1 and A.2 hold.5.3. The Central Role of Conditional ExogeneityTo relate direct structural causality to G−causality, we maintain A.2, a specific conditionalexogeneity assumption. Can this assumption be eliminated or weakened? Weshow that the answer is no: A.2 is in a precise sense a necessary condition. We alsogive a result supporting tests for conditional exogeneity.20

Linking Granger Causality and the Pearl Causal Model with Settable SystemsFirst, we specify the sense in which conditional exogeneity is necessary for theequivalence of G−causality and direct structural causality.DProposition 5.7 Given A.1, suppose that Y 2,t−1 S Y 1,t , t = 1,2,... . If A.2 does nothold, then for each t there exists q 1,t such that Y 1,t ⊥ Y 2,t−1 | Y 1,t−1 , X t does not hold.That is, if conditional exogeneity does not hold, then there are always structures thatgenerate data exhibiting G−causality, despite the absence of direct structural causality.Because q 1,t is unknown, this worst case scenario can never be discounted. Further, asWL show, the class of worst case structures includes precisely those usually assumedin applications, namely separable structures (e.g., Y 1,t = q 1,t (Y 1,t−1 , Z t ) + U 1,t ), as wellas the more general class of invertible structures. Thus, in the cases typically assumedin the literature, the failure of conditional exogeneity guarantees G−causality in theabsence of structural causality. We state this formally as a corollary.Corollary 5.8 Given A.1 with Y 2,t−1DS Y 1,t , t = 1,2,..., suppose that q 1,t is invertiblein the sense that Y 1,t = q 1,t (Y 1,t−1 , Z t ,U 1,t ) implies the existence of ξ 1,t such that U 1,t =ξ 1,t (Y 1,t−1 , Z t ,Y 1,t ), t = 1,2,... . If A.2 fails, then Y 1,t ⊥ Y 2,t−1 | Y 1,t−1 , X t fails, t = 1,2,....Together with Theorem 5.6, this establishes that in the absence of direct causality andfor the class of invertible structures predominant in applications, conditional exogeneityis necessary and sufficient for G non-causality.Tests of conditional exogeneity for the general separable case follow from:Proposition 5.9 Given A.1, suppose that E(Y 1,t ) < ∞ and thatq 1,t (Y t−1 , Z t ,U 1,t ) = ζ t (Y t−1 , Z t ) + υ t (Y 1,t−1 , Z t ,U 1,t ),where ζ t and υ t are unknown measurable functions. Let ε t := Y 1,t − E(Y 1,t |Y t−1 , X t ). IfA.2 holds, thenε t = υ t (Y 1,t−1 , Z t ,U 1,t ) − E(υ t (Y 1,t−1 , Z t ,U 1,t ) | Y 1,t−1 , X t )E(ε t |Y t−1 , X t ) = E(ε t |Y 1,t−1 , X t ) = 0andY 2,t−1 ⊥ ε t | Y 1,t−1 , X t . (5)Tests based on this result detect the failure of A.2, given separability. Such tests arefeasible because even though the regression error ε t is unobserved, it can be consistentlyestimated, say as ˆε t := Y 1,t − Ê(Y 1,t |Y t−1 , X t ), where Ê(Y 1,t |Y t−1 , X t ) is a parametric ornonparametric estimator of E(Y 1,t |Y t−1 , X t ). These estimated errors can then be used totest (5). If we reject (5), then we must reject A.2. We discuss a practical procedure inthe next section. WL provide additional discussion.WL also discuss dropping the separability assumption. For brevity, we maintainseparability here. Observe that under the null of direct non-causality, q 1,t is necessarilyseparable, as then ζ t is the zero function.21

White Chalak Lu6. Testing Direct Structural CausalityHere, we discuss methods for testing direct structural causality. First, we discuss a generalapproach that combines tests of G non-causality (GN) and conditional exogeneity(CE). Then we describe straightforward practical methods for implementing the generalapproach.6.1. Combining Tests for GN and CETheorem 5.6 implies that if we test and reject GN, then we must reject either directstructural non-causality (SN) or CE, or both. If CE is maintained, then we can directlytest SN by testing GN; otherwise, a direct test is not available.Similarly, under the traditional separability assumption, Corollary 5.8 implies thatif we test and reject CE, then we must reject either SN or GN (or both). If GN ismaintained, then we can directly test SN by testing CE; otherwise, a direct test is notavailable.When neither CE nor GN is maintained, no direct test of SN is possible. Nevertheless,we can test structural causality indirectly by combining the results of the GN andCE tests to isolate the source of any rejections. WL propose the following indirect test:(1) Reject SN if either:(i) the CE test fails to reject and the GN test rejects; or(ii) the CE test rejects and the GN test fails to reject.If these rejection conditions do not hold, however, we cannot just decide to “accept"(i.e., fail to reject) SN. As WL explain in detail, difficulties arise when CE and GN bothfail, as failing to reject SN here runs the risk of Type II error, whereas rejecting SN runsthe risk of Type I error. We resolve this dilemma by specifying the further rules:(2) Fail to reject SN if the CE and GN tests both fail to reject;(3) Make no decision as to SN if the CE and GN tests both reject.In the latter case, we conclude only that CE and GN both fail, thereby obstructing structuralinference. This sends a clear signal that the researcher needs to revisit the modelspecification, with particular attention to specifying covariates sufficient to ensure conditionalexogeneity.Because of the structure of this indirect test, it is not enough simply to consider itslevel and power. We must also account for the possibility of making no decision. Forthis, definep : = P[ wrongly make a decision ]= P[ fail to reject CE or GN | CE is false and GN is false ]q : = P[ wrongly make no decision ]= P[ reject CE and GN | CE is true or GN is true ].22

Linking Granger Causality and the Pearl Causal Model with Settable SystemsThese are the analogs of the probabilities of Type I and Type II errors for the “nodecision" action. We would like these probabilities to be small. Next, we considerα * : = P[ reject SN or make no decision | CE is true and GN is true ]π * : = P[ reject SN | exactly one of CE and GN is true ].These quantities correspond to notions of level and power, but with the sample spacerestricted to the subset on which CE is true or GN is true, that is, the space where adecision can be made. Thus, α * differs from the standard notion of level, but it doescapture the probability of taking an incorrect action when SN (the null) holds in therestricted sample space, i.e., when CE and GN are both true. Similarly, π * captures theprobability of taking the correct action when SN does not hold in the restricted samplespace. We would like the “restricted level" α * to be small and the “restricted power" π *to be close to one.WL provide useful bounds on the asymptotic properties (T → ∞) of the sample-sizeT values of the probabilities defined above, p T , q T , α * T , and π* T :Proposition 6.1 Suppose that for T = 1,2,... the significance levels of the CE and GNtests are α 1T and α 2T , respectively, and that α 1T → α 1 < .5 and α 2T → α 2 < .5. Supposethe powers of the CE and GN tests are π 1T and π 2T , respectively, and that π 1T → 1 andπ 2T → 1. Thenp T → 0, limsupq T ≤ max{α 1 ,α 2 },|α 1 − α 2 | ≤ liminf α * T ≤ limsupα* T ≤ α 1 + α 2 + min{α 1 ,α 2 }, andmin{1 − α 1 ,1 − α 2 } ≤ liminf π * T ≤ limsupπ* T ≤ max{1 − α 1,1 − α 2 }.When π 1T → 1 and π 2T → 1, one can also typically ensure α 1 = 0 and α 2 = 0 by suitablechoice of an increasing sequence of critical values. In this case, q T → 0, α * T→ 0, andπ * T→ 1. Because GN and CE tests will not be consistent against every possible alternative,weaker asymptotic bounds on the level and power of the indirect test hold for thesecases by Proposition 8.1 of WL. Thus, whenever possible, one should carefully designGN and CE tests to have power against particularly important or plausible alternatives.See WL for further discussion.6.2. Practical Tests for GN and CETo test GN and CE, we require tests for conditional independence. Nonparametrictests for conditional independence consistent against arbitrary alternatives are readilyavailable (e.g. Linton, O. and P. Gozalo, 1997; Fernandes, M. and R. G. Flores, 2001;Delgado, M. A. and W. Gonzalez-Manteiga, 2001; Su, L. and H. White, 2007a,b, 2008;Song, K., 2009; Huang, M. and H. White, 2009). In principle, one can apply any ofthese to consistently test GN and CE.But nonparametric tests are often not practical, due to the typically modest numberof time-series observations available relative to the number of relevant observable23

White Chalak Luvariables. In practice, researchers typically use parametric methods. These are convenient,but they may lack power against important alternatives. To provide convenientprocedures for testing GN and CE with power against a wider range of alternatives,WL propose augmenting standard tests with neural network terms, motivated by the“QuickNet" procedures introduced by White, H (2006b) or the extreme learning machine(ELM) methods of Huang, G.B., Q.Y. Zhu, and C.K. Siew (2006). We now provideexplicit practical methods for testing GN and CE for a leading class of structuresobeying A.1.6.2.1. Testing Granger Non-CausalityStandard tests for finite-order G−causality (e.g., Stock, J. and M. Watson, 2007, p. 547)typically assume a linear regression, such as 3E(Y 1,t |Y t−1 , X t ) = α 0 + Y ′ 1,t−1 ρ 0 + Y ′ 2,t−1 β 0 + X ′ tβ 1 .For simplicity, we let Y 1,t be a scalar here. The extension to the case of vector Y 1,t iscompletely straightforward. Under the null of GN, i.e., Y 1,t ⊥ Y 2,t−1 | Y 1,t−1 , X t , we haveβ 0 = 0. The standard procedure therefore tests β 0 = 0 in the regression equationY 1,t = α 0 + Y ′ 1,t−1 ρ 0 + Y ′ 2,t−1 β 0 + X ′ tβ 1 + ε t . (GN Test Regression 1)If we reject β 0 = 0, then we also reject GN. But if we don’t reject β 0 = 0, care is needed,as not all failures of GN will be indicated by β 0 0.Observe that when CE holds and if GN Test Regression 1 is correctly specified, i.e.,the conditional expectation E(Y 1,t |Y t−1 , X t ) is indeed linear in the conditioning variables,then β 0 represents precisely the direct structural effect of Y 2,t−1 on Y 1,t . Thus,GN Test Regression 1 may not only permit a test of GN, but it may also provide aconsistent estimate of the direct structural effect of interest.To mitigate specification error and gain power against a wider range of alternatives,WL propose augmenting GN Test Regression 1 with neural network terms, as inWhite’s (2006b, p. 476) QuickNet procedure. This involves testing β 0 = 0 inY 1,t = α 0 + Y ′ 1,t−1 ρ 0 + Y ′ 2,t−1 β 0 + X ′ tβ 1 +r∑︁ψ(Y ′ 1,t−1 γ 1, j + X ′ tγ j )β j+1 + ε t .j=1(GN Test Regression 2)Here, the activation function ψ is a generically comprehensively revealing (GCR) function(see Stinchcombe, M. and H. White, 1998). For example, ψ can be the logistic cdfψ(z) = 1/(1 + exp(−z)) or a ridgelet function, e.g., ψ(z) = (−z 5 + 10z 3 − 15z)exp(−.5z 2 )(see, for example, Candès, E. (1999)). The integer r lies between 1 and ¯r, the maximumnumber of hidden units. We randomly choose (γ 0 j ,γ j ) as in White, H (2006b, p. 477).3. For notational convenience, we understand that all regressors have been recast as vectors containingthe referenced elements.24

Linking Granger Causality and the Pearl Causal Model with Settable SystemsParallel to our comment above about estimating direct structural effects of interest,we note that given A.1, A.2, and some further mild regularity conditions, such effectscan be identified and estimated from a neural network regression of the formY 1,t = α 0 + Y ′ 1,t−1 ρ 0 + Y ′ 2,t−1 β 0 + X ′ tβ 1r∑︁+ ψ(Y ′ 1,t−1 γ 1, j + Y ′ 2,t−1 γ 2, j + X ′ tγ 3, j )β j+1 + ε t .j=1Observe that this regression includes Y 2,t−1 inside the hidden units. With r chosen sufficientlylarge, this permits the regression to achieve a sufficiently close approximationto E(Y 1,t |Y t−1 , X t ) and its derivatives (see Hornik, K. M. Stinchcombe, and H. White,1990; Gallant, A.R. and H. White, 1992) that regression misspecification is not such anissue. In this case, the derivative of the estimated regression with respect to Y 2,t−1 wellapproximates(∂/∂y 2 )E(Y 1,t | Y 1,t−1 ,Y 2,t−1 = y 2 , X t )= E[ (∂/∂y 2 )q 1,t (Y 1,t−1 , y 2 , Z t ,U 1,t ) | Y 1,t−1 , X t ].This quantity is the covariate conditioned expected marginal direct effect of Y 2,t−1 onY 1,t .Although it is possible to base a test for GN on these estimated effects, we do notpropose this here, as the required analysis is much more involved than that associatedwith GN Test Regression 2.Finally, to gain additional power WL propose tests using transformations of Y 1,t ,Y 1,t−1 , and Y 2,t−1 , as Y 1,t ⊥ Y 2,t−1 | Y 1,t−1 , X t implies f (Y 1,t ) ⊥ g(Y 2,t−1 ) | Y 1,t−1 , X t forall measurable f and g. One then tests β 1,0 = 0 inψ 1 (Y 1,t ) = α 1,0 + ψ 2 (Y 1,t−1 ) ′ ρ 1,0 + ψ 3 (Y 2,t−1 ) ′ β 1,0 + X ′ tβ 1,1r∑︁+ ψ(Y ′ 1,t−1 γ 1,1, j + X ′ tγ 1, j )β 1, j+1 + η t . (GN Test Regression 3)j=1We take ψ 1 and the elements of the vector ψ 3 to be GCR, e.g., ridgelets or the logisticcdf. The choices of γ,r, and ψ are as described above. Here, ψ 2 can be the identity(ψ 2 (Y 1,t−1 ) = Y 1,t−1 ), its elements can coincide with ψ 1 , or it can be a different GCRfunction.6.2.2. Testing Conditional ExogeneityTesting conditional exogeneity requires testing A.2, i.e., Y 2,t−1 ⊥ U 1,t | Y 1,t−1 , X t . SinceU 1,t is unobservable, we cannot test this directly. But with separability (which holdsunder the null of direct structural non-causality), Proposition 5.9 shows that Y 2,t−1 ⊥U 1,t | Y 1,t−1 , X t implies Y 2,t−1 ⊥ ε t | Y 1,t−1 , X t , where ε t := Y 1,t − E(Y 1,t |Y t−1 , X t ). Withcorrect specification of E(Y 1,t |Y t−1 , X t ) in either GN Test Regression 1 or 2 (or someother appropriate regression), we can estimate ε t and use these estimates to test Y 2,t−1 ⊥25

White Chalak Luε t | Y 1,t−1 , X t . If we reject this, then we also must reject CE. We describe the procedurein detail below.As WL discuss, such a procedure is not “watertight," as this method may miss certainalternatives to CE. But, as it turns out, there is no completely infallible method. Byoffering the opportunity of falsification, this method provides crucial insurance againstbeing naively misled into inappropriate causal inferences. See WL for further discussion.The first step in constructing a practical test for CE is to compute estimates of ε t ,say ˆε t . This can be done in the obvious way by taking ˆε t to be the estimated residualsfrom a suitable regression. For concreteness, suppose this is either GN Test Regression1 or 2.The next step is to use ˆε t to test Y 2,t−1 ⊥ ε t | Y 1,t−1 , X t . WL recommend doing thisby estimating the following analog of GN Test Regression 3:ψ 1 (ˆε t ) = α 2,0 + ψ 2 (Y 1,t−1 ) ′ ρ 2,0 + ψ 3 (Y 2,t−1 ) ′ β 2,0 + X ′ tβ 2,1r∑︁+ ψ(Y ′ 1,t−1 γ 2,1, j + X ′ tγ 2, j )β 2, j+1 + η t .j=1(CE Test Regression)Note that the right-hand-side regressors are identical to those of GN Test Regression 3;we just replace the dependent variable ψ 1 (Y 1,t ) for GN with ψ 1 (ˆε t ) for CE. Nevertheless,the transformations ψ 1 , ψ 2 , and ψ 3 here may differ from those of GN Test Regression3. To keep the notation simple, we leave these possible differences implicit. To test CEusing this regression, we test the null hypothesis β 2,0 = 0 : if we reject β 2,0 = 0, then wereject CE.As WL explain, the fact that ˆε t is obtained from a “first-stage" estimation (GN) involvingpotentially the same regressors as those appearing in the CE regression meansthat choosing ψ 1 (ˆε t ) = ˆε t can easily lead to a test with no power. For CE, WL thus recommendchoosing ψ 1 to be GCR. Alternatively, non-GCR choices may be informative,such asψ 1 (ˆε t ) = |ˆε t |, ψ 1 (ˆε t ) = ˆε t (λ − 1{ˆε t < 0}), λ ∈ (0,1), or ψ 1 (ˆε t ) = ˆε 2 t .Significantly, the asymptotic sampling distributions needed to test β 2,0 = 0 will generallybe impacted by the first-stage estimation. Handling this properly is straightforward,but somewhat involved. To describe a practical method, we denote the first-stage(GN) estimator as ˆθ 1,T := ( ˆα 1,T , ˆρ 1,T , ˆβ ′ 1,0,T , ˆβ ′ 1,1,T ,..., ˆβ 1,r+1,T ) ′ , computed from GN TestRegression 1 (r = 0) or 2 (r > 0). Let the second stage (CE) regression estimator be ˆθ 2,T ;this contains the estimated coefficients for Y 2,t−1 , say ˆβ 2,0,T , which carry the informationabout CE. Under mild conditions, a central limit theorem ensures that√T(ˆθ T − θ 0 ) d → N(0,C 0 ),where ˆθ T := (ˆθ ′ 1,T , ˆθ ′ 2,T )′ , θ 0 := plim(ˆθ T ), convergence in distribution as T → ∞ is denotedd →, and N(0,C 0 ) denotes the multivariate normal distribution with mean zero and26

Linking Granger Causality and the Pearl Causal Model with Settable Systemscovariance matrix C 0 := A −10 B 0A −1′0, where[︃ ]︃A011 0A 0 :=A 022A 021is a two-stage analog of the log-likelihood Hessian and B 0 is an analog of the informationmatrix. See White, H. (1994, pp. 103–108) for specifics. 4 This fact can then be useto construct a well behaved test for β 2,0 = 0.Constructing this test is especially straightforward when the regression errors of theGN and CE regressions, ε t and η t , are suitable martingale differences. Then B 0 has theform[︃ E[ Zt εB 0 :=t ε ′ t Z′ t ] E[ Z t ε t η ′ t Z′ t ] ]︃E[ Z t η t ε ′ t Z′ t ] E[ Z t η t η ′ t Z′ t ] ,where the CE regressors Z t are measurable-σ(X t ), X t := (vec[Y t−1 ] ′ ,vec[X t ] ′ ) ′ , ε t :=Y 1,t − E(Y 1,t | X t ), and η t := ψ 1 (ε t ) − E[ψ 1 (ε t ) | X t ]. For this, it suffices that U 1,t ⊥(Y t−l−1 , X t−τ1−1 ) | X t , as WL show. This memory condition is often plausible, as itsays that the more distant history (Y t−l−1 , X t−τ1−1 ) is not predictive for U 1,t , given themore recent history X t of (Y t−1 , X t+τ 2). Note that separability is not needed here.The details of C 0 can be involved, especially with choices like ψ 1 (ˆε t ) = |ˆε t |. But thisis a standard m−estimation setting, so we can avoid explicit estimation of C 0 : suitablebootstrap methods deliver valid critical values, even without the martingale differenceproperty (see, e.g., Gonçalves, S. and H. White, 2004; Kiefer, N. and T. Vogelsang,2002, 2005; Politis, D. N., 2009).An especially appealing method is the weighted bootstrap (Ma, S. and M. Kosorok,2005), which works under general conditions, given the martingale difference property.To implement this, for i = 1,...,n generate sequences {W t,i ,t = 1,...,T} of IID positivescalar weights with E(W t,i ) = 1 and σ 2 W := var(W t,i) = 1. For example, take W t,i ∼χ 2 1 /√ 2 + (1 − 1/ √ 2), where χ 2 1is chi-squared with one degree of freedom. The weightsshould be independent of the sample data and of each other. Then compute estimatorsˆθ T,i by weighted least squares applied to the GN and CE regressions using (the same)weights {W t,i ,t = 1,...,T}. By Ma, S. and M. Kosorok (2005, theorem 2), the randomvariables √T(ˆθ T,i − ˆθ T ), i = 1,...,ncan then be used to form valid asymptotic critical values for testing hypotheses aboutθ 0 .To test CE, we test β 2,0 = 0. This is a restriction of the form S 2 θ 0 = 0, where S 2 is theselection matrix that selects the elements β 2,0 from θ 0 . Thus, to conduct an asymptoticlevel α test, we can first compute the test statistic, sayT T := T ˆθ ′ T S′ 2 S 2 ˆθ T = T ˆβ ′ 2,0,T ˆβ 2,0,T .4. The regularity conditions include plausible memory and moment requirements, together with certainsmoothness and other technical conditions.27

White Chalak LuWe then reject CE if T T > ĉ T,n,1−α , where, with n chosen sufficiently large, ĉ T,n,1−α isthe 1 − α percentile of the weighted bootstrap statisticsT T,i := T (ˆθ T,i − ˆθ T ) ′ S ′ 2 S 2 (ˆθ T,i − ˆθ T ) = T (ˆβ 2,0,T,i − ˆβ 2,0,T ) ′ (ˆβ 2,0,T,i − ˆβ 2,0,T ),i = 1,...,n.This procedure is asymptotically valid, even though T T is based on the “unstudentized"statistic S 2 ˆθ T = ˆβ 2,0,T . Alternatively, one can construct a studentized statisticT * T := T ˆθ ′ T S′ 2 [S 2 ĈT,n S ′ 2 ]−1 S 2 ˆθ T ,where Ĉ T,n is an asymptotic covariance estimator constructed from √ T(ˆθ T,i − ˆθ T ), i =1,...,n. The test rejects CE if T * T > c 1−α, where c 1−α is the 1 − α percentile of the chisquareddistribution with dim(β 0,2 ) degrees of freedom. This method is more involvedbut may have better control over the level of the test. WL provide further discussionand methods.Because the given asymptotic distribution is joint for ˆθ 1,T and ˆθ 2,T , the same methodsconveniently apply to testing GN, i.e., β 1,0 = S 1 θ 0 = 0, where S 1 selects β 1,0 fromθ 0 . In this way, GN and CE test statistics can be constructed at the same time.WL discuss three examples, illustrating tests for direct structural non-causalitybased on tests of Granger non-causality and conditional exogeneity. A matlab module,testsn, implementing the methods described here is available at http://ihome.ust.hk/~xunlu/code.htm.7. Summary and Concluding RemarksIn this paper, we explore the relations between direct structural causality in the settablesystems framework and direct causality in the PCM for both recursive and nonrecursivesystems. The close correspondence between these concepts in recursive systemsand the equivalence between direct structural causality and G−causality establishedby WL enable us to show the close linkage between G−causality and PCM notionsof direct causality. We apply WL’s results to provide straightforward practicalmethods for testing direct causality using tests for Granger causality and conditionalexogeneity.The methods and results described here draw largely from work of WC and WL.These papers contain much additional relevant discussion and detail. WC provide furtherexamples contrasting settable systems and the PCM. Chalak, K. and H. White(2010) build on WC, examining not only direct causality in settable systems, but alsonotions of indirect causality, which in turn yield implications for conditional independencerelations, such as those embodied in conditional exogeneity, which plays a keyrole here. WL treat not only the structural VAR case analyzed here, but also the “timeseriesnatural experiment" case, where causal effects of variables D t , absorbed here intoZ t , are explicitly analyzed. The sequence {D t } represents external stimuli, not driven by{Y t }, whose effects on {Y t } are of interest. For example, {D t } could represent passivelyobserved visual or auditory stimuli, and {Y t } could represent measured neural activity.Interest may attach to which stimuli directly or indirectly affect which neurons or28

Linking Granger Causality and the Pearl Causal Model with Settable Systemsgroups of neurons. WL also examine the structural content of classical Granger causalityand a variety of related alternative versions that emerge naturally from differentversions of Assumption A.1.AcknowledgmentsWe express our deep appreciation to Sir Clive W.J. Granger for his encouragement ofthe research underlying the work presented here.ReferencesE. Candès. Ridgelets: Estimating with Ridge Functions. Annals of Statistics, 31:1561–1599, 1999.K. Chalak and H. White. Causality, Conditional Independence, and Graphical Separationin Settable Systems. Technical report, Department of Economics, BostonCollege, 2010.A. P. Dawid. Conditional Independence in Statistical Theory. Journal of the RoyalStatistical Society, Series B, 41:1–31, 1979.A. P. Dawid. Beware of the DAG! Proceedings of the NIPS 2008 Workshop on Causality,Journal of Machine Learning Research Workshop and Conference Proceedings,6:59–86, 2010.M. A. Delgado and W. Gonzalez-Manteiga. Significance Testing in NonparametricRegression Based on the Bootstrap. Annals of Statistics, 29:1469–1507, 2001.M. Eichler. Granger Causality and Path Diagrams for Multivariate Time Series. Journalof Econometrics, 137:334-353, 2007.M. Eichler and V. Didelez. Granger-causality and the Effect of Interventions in TimeSeries. Lifetime Data Analysis, forthcoming.R. Engle, D. Hendry, and J.F. Richard. Exogeneity. Econometrica 51:277–304, 1983.M. Fernandes and R. G. Flores. Tests for Conditional Independence, Markovian Dynamics,and Noncausality. Technical report, European University Institute, 2001.J.P. Florens and D. Fougère. Non-causality in Continuous Time. Econometrica,64:1195–1212, 1996.J.P. Florens and M. Mouchart. A Note on Non-causality. Econometrica, 50:583–591,1982.A.R. Gallant and H. White. On Learning the Derivatives of an Unknown Mapping withMultilayer Feedforward Networks. Neural Networks, 5:129–138, 1992.29

White Chalak LuD. Galles and J. Pearl. Axioms of Causal Relevance. Artificial Intelligence, 97:9–43,1997.R. Gibbons. Game Theory for Applied Economists. Princeton University Press, Princeton,1992.S. Gonçalves and H. White. Maximum Likelihood and the Bootstrap for NonlinearDynamic Models. Journal of Econometrics, 119:199–219, 2004.C. W. J. Granger. Investigating Causal Relations by Econometric Models and CrossspectralMethods. Econometrica, 37:424–438, 1969.C. W. J. Granger and P. Newbold. Forecasting Economic Time Series (2nd edition).Academic Press, New York, 1986.J. Halpern. Axiomatizing Causal Reasoning. Journal of Artificial Intelligence Research,12:317-337, 2000.K. Hornik, M. Stinchcombe, and H. White. Universal Approximation of an UnknownMapping and its Derivatives Using Multilayer Feedforward Networks. Neural Networks,3:551–560, 1990.G. B. Huang, Q.Y. Zhu, and C.K. Siew. Extreme Learning Machines: Theory andApplications. Neurocomputing, 70:489–501, 2006.M. Huang and H. White. A Flexible Test for Conditional Independence. Technicalreport, Department of Economics, University of California, San Diego.N. Kiefer and T. Vogelsang. Heteroskedasticity-autocorrelation Robust Testing UsingBandwidth Equal to Sample Size. Econometric Theory, 18:1350–1366, 2002.N. Kiefer and T. Vogelsang. A New Asymptotic Theory for HeteroskedasticityautocorrelationRobust Tests. Econometric Theory, 21:1130–1164, 2005.O. Linton and P. Gozalo. Conditional Independence Restrictions: Testing and Estimation.Technical report, Cowles Foundation for Research, Yale University, 1997.S. Ma and M. Kosorok. Robust Semiparametric M-estimation and the Weighted Bootstrap.Journal of Multivariate Analysis, 96:190-217, 2005.J. Pearl. Causality. Cambridge University Press, New York, 2000.J. Pearl. Direct and Indirect Effects. In Proceedings of the Seventeenth Conference onUncertainty in Artificial Intelligence, pages 411-420, 2001.D.N. Politis. Higher-Order Accurate, Positive Semi-definite Estimation of Large-Sample Covariance and Spectral Density Matrices. Technical report, Departmentof Economics, University of California, San Diego, 2009.30

Linking Granger Causality and the Pearl Causal Model with Settable SystemsA. Roebroeck, A.K. Seth, and P. Valdes-Sosa. Causality Analysis of Functional MagneticResonance Imaging Data. Journal of Machine Learning Research, (this issue),2011.K. Song. Testing Conditional Independence via Rosenblatt Transforms. Annals ofStatistics, 37:4011-4015, 2009.M. Stinchcombe and H. White. Consistent Specification Testing with Nuisance ParametersPresent Only Under the Alternative. Econometric Theory, 14:295-324, 1998.J. Stock and M. Watson. Introduction to Econometrics. Addison-Wesley, Boston, 2007.L. Su and H. White. A Consistent Characteristic Function-Based Test for ConditionalIndependence. Journal of Econometrics, 141:807-834, 2007a.L. Su and H. White. Testing Conditional Independence via Empirical Likelihood. Technicalreport, Department of Economics, University of California, San Diego, 2007b.L. Su and H. White. A Nonparametric Hellinger Metric Test for Conditional Independence.Econometric Theory, 24:829–864, 2008.H. Varian. Intermediate Microeconomics (8th edition). Norton, New York, 2009.H. White. Estimation, Inference, and Specification Analysis. Cambridge UniversityPress, New York, 1994.H. White. Time-series Estimation of the Effects of Natural Experiments. Journal ofEconometrics, 135:527-566, 2006a.H. White. Approximate Nonlinear Forecasting Methods. In G. Elliott, C.W.J. Granger,and A. Timmermann, editors, Handbook of Economic Forecasting, pages 460–512,Elsevier, New York, 2006b.H. White and K. Chalak. Settable Systems: An Extension of Pearl’s Causal Model withOptimization, Equilibrium, and Learning. Journal of Machine Learning Research,10:1759-1799, 2009.H. White and P. Kennedy. Retrospective Estimation of Causal Effects Through Time. InJ. Castle and N. Shephard editors, The Methodology and Practice of Econometrics:A Festschrift in Honour of David F. Hendry, pages 59–87, Oxford University Press,Oxford, 2009.H. White and X. Lu. Granger Causality and Dynamic Structural Systems. Journal ofFinancial Econometrics, 8:193-243, 2010a.H. White and X. Lu. Causal Diagrams for Treatment Effect Estimation with Applicationto Selection of Efficient Covariates. Technical report, Department of Economics,University of California, San Diego, 2010b.31

JMLR: Workshop and Conference Proceedings 12:30–64, 2011Causality in Time SeriesRobust statistics for describing causality in multivariatetime-series.Florin PopescuFraunhofer Institute FIRSTKekulestr. 7, Berlin 12489 Germanyflorin.popescu@first.fraunhofer.deEditors: Florin Popescu and Isabelle GuyonAbstractA widely agreed upon definition of time series causality inference, established in theseminal 1969 article of Clive Granger (1969), is based on the relative ability of the historyof one time series to predict the current state of another, conditional on all otherpast information. While the Granger Causality (GC) principle remains uncontested,its literal application is challenged by practical and physical limitations of the processof discretely sampling continuous dynamic systems. Advances in methodology fortime-series causality subsequently evolved mainly in econometrics and brain imaging:while each domain has specific data and noise characteristics the basic aims andchallenges are similar. Dynamic interactions may occur at higher temporal or spatialresolution than our ability to measure them, which leads to the potentially false inferenceof causation where only correlation is present. Causality assignment can be seenas the principled partition of spectral coherence among interacting signals using bothauto-regressive (AR) modelling and spectral decomposition. While both approachesare theoretically equivalent, interchangeably describing linear dynamic processes, thepurely spectral approach currently differs in its somewhat higher ability to accuratelydeal with mixed additive noise.Two new methods are introduced 1) a purely auto-regressive method named CausalStructural Information is introduced which unlike current AR-based methods is robustto mixed additive noise and 2) a novel means of calculating multivariate spectrafor unevenly sampled data based on cardinal trigonometric functions is incorporatedinto the recently introduced phase slope index (PSI) spectral causal inference method(Nolte et al., 2008). In addition to these, PSI, partial coherence-based PSI and existingAR-based causality measures were tested on a specially constructed data-set simulatingpossible confounding effects of mixed noise and another additionally testing theinfluence of common, background driving signals. Tabulated statistics are provided,in which true causality influence is subjected to an acceptable level of false inferenceprobability. The new methods as well as PSI are shown to allow reliable inferencefor signals as short as 100 points and to be robust to additive colored mixed noise andto the influence commonly coupled driving signals, as well as provide for a usefulmeasure of strength of causal influence.Keywords: Causality, spectral decomposition, cross-correlation, auto regressive models.c○ 2011 F. Popescu.

Popescu1. IntroductionCausality is the sine qua non of scientific inference methodology, allowing us, amongother things to advocate effective policy, diagnose and cure disease and explain brainfunction. While it has recently attracted much interest within Machine Learning, itbears reminding that a lot of this recent effort has been directed toward static data ratherthan time series. The ‘classical’ statisticians of the early 20 th century, such as Fisher,Gosset and Karl Pearson, aimed at a rational and general recipe for causal inference anddiscovery (Gigerenzer et al., 1990) but the tools they developed applied to simple typesof inference which required the pres-selection, through consensus or by design, of ahandful of candidate causes (or ‘treatments’) and a handful of subsequently occurringcandidate effects. Numerical experiments yielded tables which were intended to serveas a technician’s almanac (Pearson, 1930; Fisher, 1925), and are today an essential partof the vocabulary of scientific discourse, although tables have been replaced by preciseformulae and specialized software. These methods rely on removing possible causallinks at a certain ‘significance level’, on the basic premise that a twin experiment ondata of similar size generated by a hypothetical non-causal mechanism would yield aresult of similar strength only with a known (small) probability. While it may havebeen hoped that a generalization of the statistical test of difference among populationmeans (e.g. the t-test) to the case of time series causal structure may be possible usinga similar almanac or recipe book approach, in reality causality has proven to be a muchmore contentious - and difficult - issue.Time series theory and analysis immediately followed the development of classicalstatistics (Yule, 1926; Wold, 1938) and was spurred thereafter by exigence (a severeeconomic boom/bust cycle, an intense high-tech global conflict) as well as opportunity(the post-war advent of a machine able to perform large linear algebra calculations).From a wide historical perspective, Fisher’s ‘almanac’ has rendered the industrialage more orderly and understandable. It can be argued, however, that the ‘scientificmethod’, at least in its accounting/statistical aspects, has not kept up with the explosivegrowth of data tabulated in history, geology, neuroscience, medicine, population dynamics,economics, finance and other fields in which causal structure is at best partiallyknown and understood, but is needed in order to cure or to advocate policy. While itmay have been hoped that the advent of the computer might give rise to an automaticinference machine able to ‘sort out’ the ever-expanding data sphere, the potential of acomputer of any conceivable power to condense the world to predictable patterns haslong been proven to be shockingly limited by mathematicians such as Turing (Turing,1936) and Kolmogorov (Kolmogorov and Shiryayev, 1992) - even before the ENIACwas built. The basic problem reduces itself to the curse of dimensionality: being forcedto choose among combinations of members of a large set of hypotheses (Lanterman,2001). Scientists as a whole took a more positive outlook, in line with post-war boomoptimism, and focused on accessible automatic inference problems. One of these wasscientists was Norbert Wiener, who, besides founding the field of cybernetics (the precursorof ML), introduced some of the basic tools of modern time-series analysis, a lineof research he began during wartime and focused on feedback control in ballistics. The34

Robust Statistics for Causalitytime-series causality definition of Granger (1969) owes inspiration to earlier discussionof causality by Wiener (1956). Granger’s approach blended spectral analysis withvector auto-regression, which had long been basic tools of economics (Wold, 1938;Koopmans, 1950), and appeared nearly at the same time as similar work by Akaike(1968) and Gersch and Goddard (1970).It is useful to highlight the differences in methodological principle and in motivationfor static vs. time series data causality inference, starting with the former asit comprises a large part of the pertinent corpus in Machine Learning and in data mining.Static causal inference is important in the sense that any classification or regressionpresumes some kind of causality, for the resulting relation to be useful in identifying elementsor features of the data which ‘cause’ or predict target labels or variables and areto be selected at the exclusion of other confounding ‘features’. In learning and generalizationof static data, sample ordering is either uninformative or unknown. Yet order isimplicitly relevant to learning both in the sense that some calculation occurs in the physicalworld in some finite number of steps which transform independent inputs (stimuli)to dependent output (responses), and in the sense that generalization should occur onexpected future stimuli. To ably generalize from a limited set of samples implies makingaccurate causal inference. With this priority in mind prior NIPS workshops haveconcentrated on feature selection and on graphical model-type causal inference (Guyonand Elisseeff, 2003; Guyon et al., 2008, 2010) inspired by the work of Pearl (2000) andSpirtes et al. (2000). The basic technique or underlying principle of this type of inferenceis vanishing partial correlation or the inference of static conditional independenceamong 3 or more random variables. While it may seem limiting that no unambiguous,generally applicable causality assignment procedure exists among single pairs of randomvariables, for large ensembles the ambiguity may be partially resolved. Statisticaltests exist which assign, with a controlled probability of false inference, random variableX 1 as dependent on X 2 given no other information, but as independent on X 2 givenX 3 , a conceptual framework proposed for time-series causality soon after Granger’s1969 paper using partial coherence rather than static correlation (Gersch and Goddard,1970). Applied to an ensemble of observations X 1 ..X N , efficient polynomial time algorithmshave been devised which combine information about pairs, triples and othersub-ensembles of random variables into a complete dependency graph including, butnot limited to, a directed acyclical graph (DAG). Such inference algorithms operatein a nearly deductive manner but are not guaranteed to have unique, optimal solution.Underlying predictive models upon which this type of inference can operate includeslinear regression (or structural equation modeling) (Richardson and Spirtes, 1999; Lacerdaet al., 2008; Pearl, 2000) and Markov chain probabilistic models (Scheines et al.,1998; Spirtes et al., 2000). Importantly, a previously unclear conceptual link betweenthe notions of time series causality and static causal inference has been formally described:see White and Lu (2010) in this volume.Likewise, algorithmic and functional relation constraints, or at least likelihoodsthereof, have been proposed as to assign causality for co-observed random variablepairs (i.e. simply by analyzing the scatter plot of X 1 vs. X 2 ) (Hoyer et al., 2009). In35

Popescugeneral terms, if we are presented a scatter plot X 1 vs. X 2 which looks like a noisysine wave, we may reasonably infer that X 2 causes X 1 , since a given value of X 2 ‘determines’X 1 and not vice versa. We may even make some mild assumptions about thenoise process which superimposes on a functional relation ( X 2 = X 1 + additive noisewhich is independent of X 1 ) and by this means turn our intuition into a proper asymmetricstatistic, i.e. a controlled probability that X 1 does not determine X 2 , an approachthat has proven remarkably successful in some cases where the presence of a causalrelation was known but the direction was not (Hoyer et al., 2009). The challenge hereis that, unlike in traditional statistics, there is not simply the case of the null hypothesisand its converse, but one of 4 mutually exclusive cases. A) X 1 is independent ofX 2 B) X 1 causes X 2 C) X 2 causes X 1 and D) X 1 and X 2 are observations of dependentand non-causally related random variables (bidirectional information flow or feedback).The appearance of a symmetric bijection (with additive noise) between X 1 and X 2 doesnot mean absence of causal relation, as asymmetry in the apparent relations is merely aclue and not a determinant of causality. Inference over static data is not without ambiguitieswithout additional assumptions and requires observations of interacting triples(or more) of variables as to allow somewhat reliable descriptions of causal relations orlack thereof (see Guyon et al. (2010) for a more comprehensive overview). Statisticalevaluation requires estimation of relative likelihood of various candidate models orcausal structures, including a null hypothesis of non-causality. In the case of complexmultidimensional data theoretical derivation of such probabilities is quite difficult, sinceit is hard to analytically describe the class of dynamic systems we may be expected toencounter. Instead, common ML practice consists in running toy experiments in whichthe ‘ground truth’ (in our case, causal structure) is only known to those who run theexperiment, while other scientists aim to test their discovery algorithms on such data,and methodological validity (including error rate) of any candidate method rests on itsability to predict responses to a set of ‘stimuli’ (test data samples) available only tothe scientists organizing the challenge. This is the underlying paradigm of the CausalityWorkbench (Guyon, 2011). In time series causality, we fortunately have far moreinformation at our disposal relevant to causality than in the static case. Any type ofreasonable interpretation of causality implies a physical mechanism which accepts amodifiable input and performs some operations in some finite time which then producean output and includes a source of randomness which gives it a stochastic nature, be itinherent to the mechanism itself or in the observation process. Intuitively, the structureor connectivity among input-output blocks that govern a data generating process are relatedto causality no matter (within limits) what the exact input-output relationships are:this is what we mean by structural causality. However, not all structures of data generatingprocesses are obviously causal, nor is it self evident how structure correspondsto Granger (non) causality (GC), as shown in further detail by White and Lu (2010).Granger causality is a measure of relative predictive information among variables andnot evidence of a direct physical mechanism linking the two processes: no amount ofanalysis can exclude a latent unobserved cause. Strictly speaking the GC statistic is36

Robust Statistics for Causalitynot a measure of causal relation: it is the possible non-rejection of a null hypothesis oftime-ordered independence.Although time information helps solve many of the ambiguities of static data severalproblems, and despite the large body of literature on time-series modeling, severalproblems in time-series causality remain vexing. Knowledge of the structure of theoverall multivariate data generating process is an indispensable aid to inferring causalrelationships: but how to infer the structure using weak a priori assumptions is an openresearch question. Sections 3, 4 and 5 will address this issue. Even in the simplest case(the bivariate case) the observation process can introduce errors in time-series causalinference by means of co-variate observation noise (Nolte et al., 2010). The bivariatedataset NOISE in the Causality Workbench addresses this case, and is extended inthis study to the evaluation datasets PAIRS and TRIPLES. Two new methods are introduced:an autoregressive method named Causal Structural Information (Section 7) anda method for estimating spectral coherence in the case of unevenly sampled data (Section8.1). A principled comparison of different methods as well as their performance interms of type I, II and III errors is necessary, which addresses both the presence/absenceof causal interaction and directionality. In discussing causal influence in real-world processes,we may reasonably expect that not inferring a potentially weak causal link maybe acceptable but positing one where none is missing may be problematic. Sections 2,6, 7 and 8 address robustness of bivariate causal inference, introducing a pair of novelmethods and evaluating them along with existing ones. Another common source ofargument in discussions of causal structure is the case of false inference by neglectingto condition the proposed causal information on other background variables which mayexplain the proposed effect equally well. While the description of a general deductivemethod of causal connectivity in multivariate time series is beyond the scope of thisarticle, Section 9 evaluates numerical and statistical performance in the tri-variate case,using methods such as CSI and partial coherence based PSI which can apply to bivariateinteractions conditioned by an arbitrary number of background variables.2. Causality statisticCausality inference is subject to a wider class of errors than classical statistics, whichtests independence among variables. A general hypothesis evaluation framework canbe:Null Hypothesis = No causal interaction H 0 = A ⊥ C B |CHypothesis 1a = A drives B H a = A → B |CHypothesis 1b = B drives A H b = B → A |CType I error prob. α = P (︁ Ĥ a or Ĥ b | H 0)︁(1)Type II error prob. β = P (︁ Ĥ 0 | H a or H b)︁37

PopescuType III error prob. γ = P (︁ Ĥ a |H b or Ĥ b |H a)︁The notation Ĥ means that our statistical estimate of the estimated likelihood of Hexceeds the threshold needed for our decision to confirm it. This formulation carriessome caveats the justification for which is pragmatic and will be expounded upon inlater sections. The main one is the use of the term ‘drives’ in place of ‘causes’. Thenull hypothesis can be viewed as equivalent to strong Granger non-causality (as it willbe argued is necessary), but it does not mean that the signals A and B are independent:they may well be correlated to one another. Furthermore, we cannot realisticallyaim at statistically supporting strict Granger causality, i.e. strictly one-sided causalinteraction, since asymmetry in bidirectional interaction may be more likely in realworldobservations and is equally meaningful. By ‘driving’ we mean instead that thehistory of one time series element A is more useful to predicting the current state ofB than vice-versa, and not that the history of B is irrelevant to predicting A. In thelatter case we would specify ‘G-causes’ instead of ‘drives’ and for H 0 we would employnon-parametric independence tests of Granger non causality (GNC) which havealready been developed as in Su and White (2008) and Moneta et al. (2010). Note thatthe definition in (1) is different from that recently proposed in White and Lu (2010),which goes further than GNC testing to make the point that structural causality inferencemust also involve a further conditional independence test: Conditional Exogeneity(CE). In simple terms, CE tests whether the innovations process of the potential effectis conditionally independent of the cause (or, by practical consequence, whether theinnovations processes are uncorrelated). White and Lu argue that if both GNC and CEfail we ought not make any decision regarding causality, and combine the power ofboth tests in a principled manner such that the probability of false causal inference, ornon-decision, is controlled. The difference in this study is that the concurrent failureof GNC and CE is precisely the difficult situation requiring additional focus and it willbe argued that methods that can cope with this situation can also perform well for thecase of CE, although they require stronger assumptions. In effect, it is assumed thatreal-world signals feature a high degree of non-causal correlation, due to aliasing effectsas described in the following section, and that strong evidence to the contrary isrequired, i.e. that non-decision is equivalent to inference of non-causality. The precisemeaning of ’driving’ will also be made explicit in the description of Causal StructuralInformation, which is implicitly a proposed definition of H 0 . Also different in Definition(1) than in White and Lu is the accounting of potential error in causal directionassignment under a framework which forces the practitioner to make such a choice ifGNC is rejected.One of the difficulties of causality inference methodology is that it is difficult to ascertainwhat true causality in the real world (‘ground truth’) is for a sufficiently comprehensiveclass of problems (such that we can reliably gage error probabilities): hence theneed for extensive simulation. A clear means of validating a causal hypothesis wouldbe intervention Pearl (2000), i.e. modification of the presumed cause, but in instancessuch as historic and geological data this is not feasible. The basic approach will be toassume a non-informative probability distribution of the degree degree of mixing, or38

Robust Statistics for Causalitynon-causal dynamic interactions, as well as over individual spectra and compile inferenceerror probabilities over a wide class of coupled dynamic systems. In constructing a‘robust causality’ statistic there is more than simply null-hypothesis rejection and accuratedirectionality to consider, however. In scientific practice we are not only interestedto know that A and B are causally related or not, but which is the main driver in case ofbidirectional coupling, and among a time series vector A, B, C, D... it is important todetermine which of these factors are the main causes of the target variable, say A. Therelative effect size and relative causal influence strength, lest the analysis be misused(Ziliak and McCloskey, 2008). The rhetorical and scientific value of effect size in noway devalues the underlying principle of robust statistics and controlled inference errorprobabilities used to quantify it.3. Auto-regression and aliasingA simple multivariate time series model is the multivariate auto-regressive model (abbreviatedas MVAR or VAR). It assumes that the data generating process (DGP) thatcreated the observations is a linear dynamic model and, as such, it contains poles onlyi.e. the numerator of the transfer function between innovations process and observationis a scalar. The more complex auto-regressive moving average model (ARMA)includes zeros as well. Despite the rather stringent assumptions of VAR, a time-seriesextension of ordinary least squares linear regression, it has been hugely successful inapplications from neuroscience to engineering to sociology and economics. Its familiarVAR (or VARX) formulation is:y i =K∑︁A k y i−k + Bu + w i (2)k=1Where {y i,d=1..D } is a real valued vector of dimension D. Notice the absence of asubscript in the exogenous input term u. This is because a general treatment of exogenousinputs requires a lagged sum, i.e. Kk=1 ∑︀B k u i−k . Since exogenous inputs are notexplicitly addressed in the following derivations the general linear operator placeholderBu is used instead and can be re-substituted for subsequent use.Granger non-causality for this system, expressed in terms of conditional independence,would place a relation among elements of y subject to knowledge of u. If D = 2, for all iy 1,i ⊥ y 2,i−1..i−K | y 1,i−1..i−K (3)If the above is true, we would say that y 2 does not finite-order G cause y 1 . If theworld was made exclusively of linear VARs, it would not be terribly difficult to devise areliable statistic for G causality. We would, given a sequence of N data points, identifythe maximum-likelihood parameters A and B via ordinary least squares (OLS) linearregression after having, via some model selection criterion, determined the order K.Furthermore we would choose another criterion (e.g. test and p-value) which tells uswhether any particular coefficient is likely to be statistically indistinguishable from 0,39

Popescuwhich would correspond to a vanishing partial correlation. If all A’s are lower triangularG non-causality is satisfied (in one direction but not the converse). It is however veryrare that the physical mechanism we are observing is indeed the embodiment of a VAR,and therefore even in the case in which G non-causality can be safely rejected, it isnot likely that the best VAR approximation of the data observed is strictly lower/uppertriangular. The necessity of a distinction between strict causality, which has a structuralinterpretation, and a causality statistic, which does not measure independence inthe sense of Granger-non causality, but rather relative degree of dependence in bothdirections among two signals (driving) is most evident in this case. If the VAR in questionhad very small (and statistically observable) upper triangular elements would adiscussion of causality of the observed time series be rendered moot?One of the most common physical mechanisms which is incompatible with VARis aliasing, i.e. dynamics which are faster than the (shortest) sampling interval. Thestandard interpretation of aliasing is the false representation of frequency componentsof a signal due to sub-Nyquist frequency sampling: in the multivariate time-series casethis can also lead to spurious correlations in the observed innovations process (Phillips,1973). Consider a continuous bivariate VAR of order 1 with Gaussian innovations inwhich the sampling frequency is several orders of magnitude smaller than the Nyquistfrequency. In this case we would observe a covariate time independent Gaussian processsince for all practical purposes the information travels ‘instantaneously’. In economics,this effect could be due to social interactions or market reactions to news whichhappen faster than the sampling interval (be it daily, hourly or monthly). In fMRI analysissub- sampling interval brain dynamics are observed over a relatively slow timeconvolution process of hemodynamic response of neural activity (for a detailed expositionof causality inference in fMRI see Roebroeck et al. (2011) in this volume).Although ‘aliasing’ normally refers to temporal aliasing, the same process can occurspatially. In neuroscience and in economics the observed variables are summations(dimensionality reductions) of a far larger set of interacting agents, be they individualsor neurons. In electroencephalography (EEG) the propagation of electrical potentialfrom cortical axons arrives via multiple pathways to the same recording location onthe scalp: the summation of micrometer scale electric potentials on the scalp at centimeterscale. Once again there are spurious observable correlations: this is known asthe mixing problem. Such effects can be modeled, albeit with significant informationloss, by the same DGP class which is a superset of VAR and known in econometrics asSVAR (structural vector auto-regression, the time series equivalent of structural equationmodeling (SEM), often used in static causality inference (Pearl, 2000)). Anotherbasic problem in dynamic system identification is that we not only discard much informationfrom the world in sampling it, but that our observations are susceptible toadditive noise, and that the randomness we see in the data is not entirely the randomnessof the mechanism we intend to study. One of the most problematic of additivenoise models is mixed colored noise, in which there are structured correlations both intime and across elements of the time-series, but not in any causal way: there is only alinear transformation of colored noise, sometimes called mixing, due to spatial aliasing.40

Robust Statistics for CausalityMixing may occur due to temporal aliasing in sampling a coupled continuous-variableVAR system. In EEG analysis mixed colored noise models the background electricalactivity of the brain. In other domains such as economics, one can imagine the influenceof unpredictable events such as natural cataclysms or macroenomic cycles whichare not white noise and which are reflect nearly ‘instantaneously’ but to varying degreein all our measurements. In this case, since each additive noise component is colored (ithas temporal autocorrelation), its past helps predict its current value. Since the observationis a linear mixture of noise components, all current observations are correlated,and the past of any component can help predict the current state of any other. In thiscase, the strict definition of Granger causality would not make practical sense, sincethis cross-predictability is not meaningful.It should be noted on this point that the literature contains (sometimes inconsistent)sub-classifications of Granger Causality, such as weak and strong Granger causality.One definition which is particularly pertinent to this work is that given in Caines (1976)and Solo (2006) and is that strong Granger causality allows instantaneous dependenceand that weak Granger causality does not (i.e. it is strictly time ordered). We are aimingin this work at strong Granger causality inference, i.e. one which is robust to aliasingeffects such as colored noise. While we should account for instantaneous interactions,we do not have to assign causal interpretations to them, since they are symmetric (thecross-correlation of independent mixed signals is symmetric).4. Auto-regression, learning and Granger CausalityLearning is the process of discovering predictable patterns in the real world, where a‘pattern’ is described by an algorithm or an automaton. Besides the object of learning,i.e. the algorithm which we infer and which maps stimuli to responses, we need to considerthe algorithm which performs the learning process and outputs the former. Thethird algorithm we should consider is the algorithm embodied in the real world, whichwe do not know, which generates the data we observe, and which we hope to be able torecover, or at least approximate. How can we formally describe it? A Data GeneratingProcess (DGP) can be a machine or automaton: an algorithm that performs every operationdeterministically in a finite number of steps, but which contains an oracle thatgenerates perfectly random numbers. It is sufficient that this oracle generate 1’s and0’s only: all other computable probability distributions can be calculated from it. ADGP contains rational valued parameters (rational as to comply with finite computability),in this case the integer K and all elements of the matrices A. Last but not leasta DGP specification may limit the set of admissible parameter values and probabilitydistributions of the oracle-generated values. The set of all possible outputs of a DGPcorresponds to the set of all probability distributions generated by it over all admissibleparameter values, which we shall call the DGP class.Definition 1 Let i ∈ N and let s a , s w , p w be finite length prefix-free binary strings.Furthermore let y and u be rational valued matrices of size N × i and M × i, and t berational valued vector with distinct elements, of length i. Let a also be a finite rational41

Popescuvalued vector. A Data Generating Process is a quintuple {s a ,p w ,T a ,T w } where T a , T ware finite time Turing machines which perform the following operations: Given an inputof the incompressible string p w the machine T w calculates a rational valued matrix w.The machine T a when given matrices y, a, u, t, w and a positive rational ∆t outputs avector y i+1 which is assigned for future operations to the time t i+1 = max(t) + ∆tThe definition is somewhat unusual in terms of the definition of stochastic systemsas embodiments of Turing machines, but it is quite standard in terms of defining aninnovations term w, a probability distribution thereof p w , a state y, a generating functionp a with parameters a and an exogenous input u. The motivation for using theterminology of algorithmic information theory is to analyse causality assignment as acomputational problem. For reasons of finite description and computability our variablesare rational, rather than real valued. Notice that there is no real restriction onhow the time series is to be generated, recursively or otherwise. The initial conditionin case of recursion is implicit, and time is specified as distinct and increasing but otherwisearbitrarily distributed - it does not necessarily grow in constant increments (itis asynchronous). The slight paradox about describing stochastic dynamical systemsin algorithmic terms is the necessity of postulating a random number generator (an oracle)which in some ways is our main tool for abstracting the complexity of the realworld, but yet is a physical impossibility (since such an oracle would require infinitecomputational time see Li and Vitanyi (1997) for overview). Also, the Turing machineswe consider have finite memory and are time restricted (they implement a predefinedmaximum number of operations before yielding a default output). Otherwise the rulesof algebra (since they perform algebraic operations) apply normally. The cover of aDGP can be defined as:Definition 2 The cover of a Data Generating Process (DGP) class is the cover of theset of all outputs y that a DGP calculates for each member of the set of admissibleparameters a,u,t,w and for each initial condition y 1 . Two DGPs are stochasticallyequivalent if the cover of the set of their possible outputs (for fixed parameters) is thesame.Let us now attempt to define a Granger Causality statistic in algorithmic terms.Allowing for the notation j..k = { j − 1, j − 2..,k + 1,k} if j > k and in reverse order ifj < k1i∑︁K(y 1, j | y 1, j−1..1 ,u j−1..1 ) − K(y 1, j | y 2, j−1..1 ,y 1, j−1..1 ,u j−1..1 ) (4)ij=1This differs from Equation (3) in two elemental ways: it is not a statement of independencebut a number (statistic), namely the average difference (rate) of conditional(or prefix) Kolmogorov complexity of each point in the presumed effect vector whengiven both vector histories or just one, and given the exogenous input history. It is ageneralized conditional entropy rate, and may be reasonably be normalized as such:42

Robust Statistics for CausalityF K2→1|u = 1 ii∑︁j=1(︃1 − K(y )︃1, j | y 2, j−1..1 ,y 1, j−1..1 ,u j−1..1 )K(y 1, j | y 1, j−1..1 ,u j−1..1 )which is a fraction ranging from 0 - meaning no influence of y 1 by y 2 - to 1, correspondingto complete determination of y 1 by y 2 and can be transformed into a statisticcomparing different data sets and processes, and which gives probabilities of spuriousresults. Another difference with Equation (3) is that we do not refer to finite-order Gcausality but simply G causality (in the general case we do not know the maximum lagorder but must infer it). For a more in depth look at DGPs, structure and G-causality,see White and Lu (2010). The larger the value F2→1|u K , the more likely that y 2 G-causesy 1 . The definition is one of conditional information and it is one of an averaged processrather than a single instance (time point). However, Kolmogorov complexity is incomputable,and as such Granger (non) causality must also be, in general, incomputable.A detailed look at this issue is beyond the scope of this article, but in essence, we cannever test all possible models that could tell us wether the history of a time series helpsor does not help predict (compress) another, and the set of finite running time Turingmachines is not enumerable. We’ve partially circumvented the halting problem sincewe’ve specified finite-state, finite-operation machines as the basis of DGPs but havenot specified a search procedure over all DGPs that enumerates them. Even if we limitourselves to DGPs which are MVAR, the necessary computational time to calculate thedescription length (instead of K(.)) is NP-complete, i.e. it requires an enumeration ofall possible parameters of a DGP class, barring any special properties thereof: findingthe optimal model order requires such a search (keep in mind VAR estimation is convexonly once we know the model order and AR structure).In practice, we should limit the class of DGPs we consider within out statistic to onewhich allows the possibility of polynomial time computation. Let us take Equation (2),and further make the common assumption that the input vector w is an i.i.d. normallydistributed sequence independent along dimension d, we’ve specified the linear VARGaussian DGP class (which we shall shorten as VAR class). This DGP class, again, hasproven remarkably useful in cases where nothing else except the time series vector y isknown. Re-writing (2):y i =K∑︁A k y i−k + D w i−1 ,D ii > 0,D i j = 0 (6)k=1The matrix D is a positive diagonal matrix containing the scaling, or effective standarddeviations of the innovation terms. The standard deviation of each element of theinnovations term w is assumed hereafter to be equal to 1.5. Equivalence of auto-regressive data generation processes.In econometrics the following formulation is familiar (SVAR):(5)43

Popescuy i =K∑︁A k y i−k + Bu + Dw i (7)k=0The difference between this and Equation (6) is the presence of a 0-lag matrix A 0which, for easy tractability has zero diagonal entries and is sometimes present on theLHS. This 0-lag matrix is meant to model the sub-sampling interval dynamic interactionsamong observations, which appear instantaneous, see Moneta et al. (2011) in thisvolume. Let us call this form zero lag SVAR. In electric- and magnetoencephalography(EEG/MEG) we often encounter the following form:x i =K∑︁µA k x i−k + µ Bu + Dw i ,k=1y i = Cx i (8)Where C represents the observation matrix, or mixing matrix and is determinedby the conductivity/permeability of tissue, and accounts for the superposition of theelectromagnetic fields created by neural activity, which happens at nearly the speed oflight and therefore appears instantaneous. Let us call this mixed output SVAR. Finally,in certain engineering applications we may see structured disturbances:y i =K∑︁θA k y i−k + θ Bu + D w w i (9)k=1Which we shall call covariate innovations SVAR (D w is a general nonsingular matrixunlike D which is diagonal). Another final SVAR form to consider would be onein which the 0-lag matrix ⊳A 0 is strictly upper triangular (upper triangular zero lagSVAR):y i = ⊳A 0 y i +K∑︁A k y i−k + ⊳Bu + Dw i (10)k=1Finally, we may consider a upper or lower triangular co-variate innovations SVAR:y i =K∑︁A k y i−k + Bu + ⊳Dw i (11)k=0Where ⊳D is upper/lower triangular. The SVAR forms (6)-(10) may look different,and in fact each of them may uniquely represent physical processes and allow for directinterpretation of parameters. From a statistical point of view, however, all four SVARDGPs introduced above are equivalent since they have identical cover.Lemma 3 The Gaussian covariate innovations SVAR DGP has the same cover as theGaussian mixed output SVAR DGP. Each of these sets has a redundancy of 2 N N! for44

Robust Statistics for Causalityinstances in which the matrices D w is the product of and unitary and diagonal matrices,the matrix C is a unitary matrix and the matrix A 0 is a permutation of an uppertriangular matrix.Proof Staring with the definition of covariate innovations SVAR in Equation (9) weuse the variable transformation y = D w x and obtain the mixed-output form (trivial).The set of Guassian random variables is closed under scalar multiplication (and hencesign change) and addition. This means that the variance if the innovations term inEquation (9) can be written as:Σ w = D T wD w = D T wU T UD wWhere U is a unitary (orthogonal, unit 2-norm) matrix. Since all innovations termelements are zero mean, the covariance matrix is the sole descriptor of the Gaussianinnovations term. This in turn means that any other matrix D ′ w = D T wU T substituted intothe DGP described in Equation (9) amounts to a stochastically equivalent DGP. Thematrix D ′ w can belong to a number of general sets of matrices, one of which is the set ofnonsingular upper triangular matrices (the transformation is achievable through the QRdecomposition of Σ w ). Another such set is lower triangular matrix set. Both are subsetsof the set of matrices sometimes named ‘psychologically upper triangular’, meaning apermutation of an upper triangular matrix.If we constrain D w to be of the form D w = UD, i.e. such that (by polar decomposition)it is the product of a unitary and a diagonal positive definite matrix, the onlystochastically equivalent transformations of D w are a symmetry preserving permutationof its rows/columns and a sign change in one of the columns (this is a property oforthogonal matrices such as U). There are N! such permutations and 2 N possible signchanges. For the general case, in which the input u has no special properties, there areno other redundancies in the SVAR model (since changing any parameter in A and Bwill otherwise change the output). Without loss of generality then, we can write thetransformation from covariate innovations to mixed output SVAR form as:x i =y i =K∑︁θA k y i−k + θ Bu + UD w w ik=1K∑︁U T ( θ A k )Ux i−k + U T ( θ B)u + D w w ik=1y i = U T x iSince the transformation U is one to one and invertible, and since this transformationis what allows a (restricted) a covariate noise SVAR to map, one to one, onto amixed output SVAR, the cardinality of both covers is the same.Now consider the zero-lag SVAR form:45

Popescuy i =K∑︁A k y i−k + Bu + Dw ik=0D −1 (1 − A 0 )y i =K∑︁D −1 A k y i−k + D −1 Bu + wk=1Taking the singular value decomposition of the (nonsingular) matrix coefficient on theLHS:V T 0 y i = S −1 U T 0U 0 S V T 0 y i =K∑︁D −1 A k y i−k + D −1 Bu + w ik=1K∑︁D −1 A k y i−k + S −1 U0 T D−1 Bu + S −1 U0 T w ik=1Using the coordinate transformation z = V0 T y. The unitary transformation UT 0can beignored due closure properties of the Gaussian. This leaves us with the mixed-outputform:z i =K∑︁S −1 U0 T D−1 A k V 0z i−k + S −1 U0 T D−1 Bu + S −1 w ′ ik=1y = V 0 zSo far we’ve shown that for every zero-lag SVAR there is at least one mixed-outputVAR. Let us for a moment consider the covariate noise SVAR (after pre-multiplication)D −1w y i =K∑︁k=1D −1w θA k y i−k + D −1w θBu + w iWe can easily then write it in terms of zero lag:y i = (︁ )︁I − D −1w yi +However, the entries of I − D −1wdone by scaling by the diagonal:K∑︁k=1diag(D −1w )y i = (diag(D −1w ) − D −1w )y i +D −1w θA k y i−k + D −1w θBu + w iare not zero (as required by definition). This can beK∑︁k=1D 0 diag(D −1w )D −1w θA k y i−k + D −1w θBu + w i46

Robust Statistics for Causalityy i = (I − D −10 D−1 w )y i +K∑︁k=1D −10 D−1 w θA k y i−k + D −10 D−1 w θBu + D −10 w iA 0 = (I − D −10 D−1 w )D −1w= diag(D −1w )(I − A 0 )While the following constant relation preserves DGP equivalence:(D T wD w ) −1 = Σ −1w= D −1w D −Tw = D 0 (I − A 0 )(I − A 0 ) T D 0A 0 = (I − D −10 D−1 w ) T (I − D −10 D−1 w )The zero lag matrix is a function of D −1w , the inverse of which is an eigenvalueproblem. However, as long as the covariance matrix or its inverse is constant, the DGPis unchanged and this allows N(N − 1)/2 degrees of freedom. Let us consider onlymixed input systems for which the innovations terms are of unit variance. There is noreal loss of generality since a simple row-division by each element of D 0 normalized thecovariate noise form (to be regained by scaling the output). In this case the equivalenceconstraint on is one of in which:(I − ⊳ A 0 ) T (I − ⊳ A 0 ) = (I − A 0 ) T (I − A 0 )If (I − A 0 ) is full rank, a strictly upper triangular matrix ⊳ A 0 may be found that isequivalent (this would be the Cholesky decomposition of the inverse covariance matrixin reverse order). As D Wis equivalent to a unitary transformation UD Wthis willinclude permutations and orthogonal rotations. Any permutation of D Wwill imply acorresponding permutation of A 0 , which (along with rotations) has 2 N N! solutions.The non-uniqueness of SVAR and the problematic interpretation of AR coefficientswith respect to variable permutation is a known problem Sims (1981), as is the factthat modeling zero-lag matrices is equivalent to covariance estimation for the Gaussiancase in the other lag coefficients are zero. In fact, statistically vanishing elements ofthe covariance matrix are used in Structural Equation Modeling and are given causalityinterpretations Pearl (2000). It is not clear how robust such inferences are with respectto equivalent permutations. The point of the lemma above is to illustrate the ambiguityof interpretation if the structure of (sparse or full) AR systems in the case of covariateinnovations, zero-lag, or mixed output, which are equivalent to each other. In the case ofSVAR, one approach is to perform standard AR followed by a Cholesky decompositionof the covariance of the residuals and then pre-multiplying. In Popescu (2008), theupper triangular SVAR estimation is done directly by singular value decompositionafter regression and the innovations covariance estimated from the zero-lag matrix.47

PopescuFigure 1: SVAR causality and equivalence. Structural VAR equivalence and causality.A) direct structural Granger causality (both directions shown). z standsfor the delay operator. B) equivalent covariate innovations (left) and mixedoutput systems. Neither representation shows dynamic interaction C) sparse,one sided covariate innovations DAG is non sparse in the mixed output case(and vice-versa). D) upper triangular structure of the zero-lag matrix is notinformative in the 2 variable Gaussian case, and is equivalent to a full mixedoutput system.48

Robust Statistics for CausalityGranger, in his 1969 paper, suggests that ‘instantaneous’ (i.e. covariate) effects beignored and only the temporal structure be used. Whether or not we accept instantaneouscausality depends on prior knowledge: in the case of EEG, the mixing matrixcannot have any physical ‘causal’ explanation even if it is sparse. Without additionala priori assumptions, either we infer causality on unseen and presumably interactinghidden variables (mixed output form, the case of EEG/MEG) or we assume a a noncausalmixed innovations input. Note also that the zero-lag system appears to be causalbut can be written in a form which suggest the opposite difference causal influence(hence it is sometimes termed ‘spurious causality’). In short, since instantaneous interactionin the Gaussian case cannot be resolved causally purely in terms of predictionand conditional information (as intended by Wiener and Granger), it is proposed thatsuch interactions be accounted for but not given causal interpretation (as in ‘strong’Granger non-causality) .There are at least four distinct overall approaches to dealing with aliasing effectsin time series causality. 1) is to make prior assumptions about covariance matrices andlimit inference to domain relevant and interpretable posteriors, as in Bernanke et al.(2005) in economics and Valdes-Sosa et al. (2005) in neuroscience. 2) to allow forunconstrained graphical causal model type inference among covariate innovations, byeither assuming Gaussianity or non-Gaussianity, the latter allowing for stronger causalinferences (see Moneta et al. (2011) in this volume). One possible drawback of thisapproach is that DAG-type inference, at least in the Gaussian case in which there isso-called ’Markov equivalence’ among candidate graphs, is non-unique. 3) a physicallyinterpretable mixed output or co-variate innovations is assumed and the inferredsparsity structure (or the intersection thereof over the nonzero lag coefficient matrices)as the connection graph. Popescu (2008) implemented such an approach by using theminimum description length principle to provide a universal prior over rational-valuedcoefficients, and was able to recover structure in the majority of simulated co-variateinnovations processes of arbitrary sparsity. This approach is computationally laborious,as it is NP and non-convex, and moreover a system that is sparse in one form (covariateinnovations or mixed-ouput) is not necessarily sparse in another equivalent SVARform. Moreover completely dense SVAR systems may be non-causal (in the strongGC sense). 4) Causality is not interpreted as a binary value, but rather direction ofinteraction is determined as a continuous valued statistic, and one which is theoreticallyrobust to covariate innovations or mixtures. This is the principle of the recentlyintroduced phase slope index (PSI), which belongs to a class of methods based on spectraldecomposition and partition of coherency. Although auto-regressive, spectral andimpulse response convolution are theoretically equivalent representation of linear dynamics,they do differ numerically and spectral representations afford direct access tophase estimates which are crucial to the interpretation of lead and lag as it relates tocausal influence. These methods are reviewed in the next section.49

Popescu6. Spectral methods and phase estimationCross- and auto spectral densities of a time series, assuming zero-mean or de-trendedvalues, are defined as:ρ Li j (τ) = E (︁ y i (t)y j (t − τ) )︁S i j (ω) = F (ρ Li j (τ)) (12)Note that continuous, linear, raw correlation values are used in the above definitionas well as the continuous Fourier transform. Bivariate coherency is defined as:C i j (ω) =S i j (ω)√︀ S ii(ω)S j j (ω)(13)Which consists of a complex numerator and a real-valued denominator. The coherenceis the squared magnitude of the coherency:c i j (ω) = C i j (ω) * C i j (ω) (14)Besides various histogram and discrete (fast) Fourier transform methods availablefor the computation of coherence, AR methods may be also used, since they are alsolinear transforms, the Fourier transform of the delay operator being simply z k = e − j2πωτ Swhere τ S is the sampling time and k = ωτ S . Plugging this into Equation (9) we obtain:⎛⎞K∑︁X( jω) = ⎜⎝ A k e − j2πωτ S kk=1⎛⎞K∑︁Y( jω) = C ⎜⎝ I − A k e − j2πωτ S k⎟⎠k=1⎟⎠ X( jω) + BU( jω) + DY( jω) = CX( jω) (15)−1(BU( jω) + DW( jω)) (16)In terms of a SVAR therefore (as opposed to VAR) the mixing matrix C does notaffect stability, nor the dynamic response (i.e. the poles). The transfer functions fromith innovations to jth output are entries of the following matrix of functions:⎛⎞K∑︁H( jω) = C ⎜⎝ I − A k e − j2πωτ S k⎟⎠k=1−1D (17)The spectral matrix is simply (having already assumed independent unit Gaussiannoise):S ( jω) = H( jω) * H( jω) (18)The coherency as the coherence following definitions above. The partial coherenceconsiders the pair (i, j) of signals conditioned on all other signals, the (ordered) set ofwhich we denote (i, j):50

Robust Statistics for CausalityS i, j|(i, j)( jω) = S (i, j),(i, j) + S (i, j),(i, j)S −1(i, j),(i, j) S (i, j),(i, j)(19)Where the subscripts refer to row/column subsets of the matrix S ( jω). The partialspectrum, substituted into Equation (13) gives us partial coherency C i, j|(i, j)( jω) andcorrespondingly, partial coherence c i, j|(i, j)( jω) . These functions are symmetric andtherefore cannot indicate direction of interaction in the pair (i, j). Several alternativeshave been proposed to account for this limitation. Kaminski and Blinowska (1991);Blinowska et al. (2004) proposed the following normalization of H( jω) which attemptsto measure the relative magnitude of the transfer function from any innovations processto any output (which is equivalent to measuring the normalized strength of Grangercausality) and is called the directed transfer function (DTF):γ i j ( jω) =H i j ( jω)√︁ ∑︀k |H ik ( jω)| 2γ 2 i j ( jω) = ⃒ ⃒⃒Hij ( jω) ⃒ ⃒ ⃒2∑︀k |H ik ( jω)| 2 (20)A similar measure is called directed coherence Baccalá et al. (Feb 1991), later elaboratedinto a method complimentary to DTF, called partial directed coherence (PDC)Baccalá and Sameshima (2001); Sameshima and Baccalá (1999), based on the inverseof H:π i j ( jω) =H −1i j√︁ ( jω)∑︀ ⃒k ⃒Hik −1(jω)⃒ ⃒2The objective of these coherency-like measures is to place a measure of directionalityon the otherwise information-symmetric coherency. While SVAR is not generallyused as a basis of the autoregressive means of spectral and coherence estimation, or ofDTF/PDC is is done so in this paper for completeness (otherwise it is assumed C = I).Granger’s 1969 paper did consider a mixing matrix (indirectly, by adding non-diagonalterms to the zero-lag matrix), and suggested ignoring the role of that part of coherencywhich depends on mixing terms as non-informative ‘instantaneous causality’. Notethat the ambiguity of the role and identifiability of the full zero lag matrix, as describedherein, was fully known at the time and was one of the justifications given for separatingsub-sampling time dynamics. Another measure of directionality, proposed bySchreiber (2000) is a Shannon-entropy interpretation of Granger Causality, and thereforewill be referred to as GC herein. The Shannon entropy, and conditional Shannonentropy of a random process is related to its spectrum. The conditional entropy formulationof Granger Causality for AR models in the multivariate case is (where (i) denotes,as above, all other elements of the vector except i ):H GCj→i|u = H(y i,t+1|y :,t:t−K ,u :,t:t−K ) − H(y i,t+1 |y ( j),t:t−K,u :,t:t−K )51

PopescuH GCj→i|u = logD i − logD ( j)i(21)The Shannon entropy of a Gaussian random variable is the logarithm of its standarddeviation plus a constant. Notice than in this paper the definition of Granger Causalityis slightly different than the literature in that it relates to the innovations process of amixed output SVAR system of closest rotation and not a regular MVAR. The secondterm D ( j)iis formed by computing a reduced SVAR system which omits the jth variable.Recently Barrett et al. have proposed an extension of GC, based on prior work byGeweke (1982) from interaction among pairs of variables to groups of variables, termedmultivariate Granger Causality (MVGC) Barrett et al. (2010). The above definition isstraightforwardly extensible to the group case, where I ad J are subsets of 1..D, sincetotal entropy of independent variables is the sum of individual entropies.H GCJ→I|u = ∑︁i∈I(︂logD i − logD (J)iThe Granger entropy can be calculated directly from the transfer function, using theShannon-Hartley theorem:H GCHj→i)︂(22)⎛ ⃒∑︁ ⃒⃒Hi= − ∆ω ln⎜⎝ 1 − j (ω) ⃒ 2 ⎞⃒S ii (ω)⎟⎠ (23)ωFinally Nolte (Nolte et al., 2008) introduced a method called Phase Slope Indexwhich evaluates bilateral causal interaction and is robust to mixing effects (i.e. zerolag, observation or innovations covariance matrices that depart from MVAR):⎛⎞∑︁PS I i j→i = Im⎜⎝C i * j (ω) C i j(ω + dω) ⎟⎠ (24)ωPSI, as a method is based on the observation that pure mixing (that is to say, alleffects stochastically equivalent to output mixing as outlined above) does not affectthe imaginary part of the coherency C i j just as (equivalently) it does not affect theantisymmetric part of the auto-correlation of a signal. It does not place a measure thephase relationship per se, but rather the slope of the coherency phase weighted by themagnitude of the coherency.7. Causal Structural InformationCurrently, Granger causality estimation based on linear VAR modeling has been shownto be susceptible to mixed noise, in the presence of which it may produce false causalityassignment Nolte et al. (2010). In order to allow for accurate causality assignment inthe presence of instantaneous interaction and aliasing the Causal Structural Information(CSI) method and statistic for causality assignment is introduced below.Consider the SVAR lower triangular form in (10) for a set of observations y. Theinformation transfer from i to j may be obtained by first defining the index re-orderings:52

Robust Statistics for Causalityi j* {i, j,i j}i* {i,i j}This means that we reorder the (identified) mixed-innovations system by placingthe target time series first and the driver series second, followed by all the rest. Thesame ordering, minus the driver is also useful. We define CSI asCS I( j → i|i j) log( ⊳ i j*D 11 ) − log( ⊳i* D 11 ) (25)CS I(i, j|i j) CS I( j → i|i j) −CS I(i → j|i j) (26)Where the D is upper-triangular form in each instance. This Granger Causality formulationrequires the identification of 3 different SVAR models, one for the entire timeseries vector, and one each for all elements except i and all elements except j. ViaCholesky decomposition, the logarithm of the top vertex of the triangle is proportionalto the entropy rate (conditional information) of the innovations process for the targetseries given all other (past and present) information including the innovations process.While this definition is clearly an interpretation of the core idea of Granger causality,it is, like DTF and PDC, not an independence statistic but a measure of (causal) informationflow among elements of a time-series vector. Note the anti-symmetry (bydefinition) of this information measure CS I(i, j|i j) = −CS I( j,i|i j) . Note also thatCS I( j → i|i j) and CS I(i → j|i j) may very conceivably have the same sign: the varioustriangular forms used to derive this measure are purely for calculation purposes,and do not carry intrinsic meaning. As a matter of fact other re-orderings and SVARforms may be employed for convenient calculation as well. In order to improve the explanatorypower of the CSI statistic the following normalization is proposed, mirroringthat defined in Equation (5) :F CS Ij→i|i j CS I(i, j|i j)log( ⊳i* D 11 ) + log( ⊳ j* D 11 ) + ζThis normalization effectively measures the ratio of causal to non-causal information,where ζ is a constant which depends on the number of dimensions and quantizationwidth and is necessary to transform continuous entropy to discrete entropy.8. Estimation of multivariate spectra and causality assignmentIn Section 6 and a series of causality assignment methods based on spectral decompositionof a multivariate signal were described. In this section spectral decompositionitself will be discussed, and a novel means of doing so for unevenly sampled data willbe introduced and evaluated along with the other methods for a bivariate benchmarkdata set.(27)53

Popescu8.1. The cardinal transform of the autocorrelationCurrently there are few commonly used methods for cross- power spectrum estimation(i.e. multivariate spectral power estimation) as opposed to univariate power spectrumestimation, and these methods average over repeated, or shifting, time windows andtherefore require a lot of data points. Furthermore all commonly used multivariate spectralpower estimation methods rely on synchronous, evenly spaced sampling, despitethe fact that much of available data is unevenly sampled, has missing values, and canbe composed of relatively short sequences. Therefore a novel method is presented belowfor multivariate spectral power estimation which can be estimated on asynchronousdata.Returning to the definition of coherency as the Fourier transform of the auto-correlation,which are both continuous transforms, we may extend the conceptual means of its estimationin the discrete sense as a regression problem (as a discrete Fourier transform,DFT) in the evenly sampled case as:Ω n n, n = −⌊N/2⌋...⌊N/2⌋ (28)2τ 0 (N − 1)Ĉ i j (ω)| ω=Ω = a i j,n + jb i j,n (29)ρ ji (−kτ) = ρ i j (kτ) = E(x i (t)x j (t + kτ)) (30)ρ i j (kτ 0 ) 1N − k∑︁q=1:N−kx i (q)x j (q + k) (31){︁ }︁ ∑︁N/2ai j ,b i j = argmink=−N/2(︁ρi j (kτ 0 ) − a i j,n cos(2πΩ n τ 0 k) − b i j,n sin(2πΩ n τ 0 k) )︁ 2(32)where τ 0 is the sampling interval. Note that for an odd number of points the regressionabove is actually a well determined set of equations, corresponding to the 2-sidedDFT. Note also that by replacing the expectation with the geometric mean, the aboveequation can also be written (with a slight change in weighting at individual lags) as:{︁ai j ,b i j}︁= argmin∑︁(︁xi,p x j,q − a i j,n cos(2πΩ k (t i,p − t j,q )) − b i j,n sin(2πΩ n (t i,p − t j,q )) )︁ 2p,q∈1..N(33)The above equation holds even for time series sampled at unequal (but overlapping)times (x i ,t i ) and (x j ,t j ) as long as the frequency basis definition is adjusted (for exampleτ 0 = 1). It represents a discrete, finite approximation of the continuous, infinite autoregressionfunction of an infinitely long random process. It is a regression on the outerproduct of the vectors x i and x j . Since autocorrelations for finite memory systems tend54

Robust Statistics for Causalityto fall off to zero with increasing lag magnitude, a novel coherency estimate is proposedbased on the cardinal sine and cosine functions, which also decay, as a compact basis:Ĉ i j (ω) = a i j,n∑︁C(Ω n ) + jb i j,n S(Ω n ) (34)∑︁p,q∈1..N{︁ai j ,b i j}︁= argmin(︁xi,p x j,q − a i j,n cosc(2πΩ k (t i,p − t j,q )) − b i j,n sinc(2πΩ n (t i,p − t j,q )) )︁ 2Where the sine cardinal is defined as sinc(x) = sin(πx)/x, and its Fourier transformis S( jω) = 1,| jω| < 1 and S( jω) = 0 otherwise. Also the Fourier transform of the cosinecardinal can be written as C( jω) = jω S ( jω). Although in principle we could chooseany complete basis as a means of Fourier transform estimation, the cardinal transformpreserves the odd-even function structure of the standard trigonometric pair. Computationallythis means that for autocorrelations, which are real valued and even, only sincneeds to be calculated and used, while for cross-correlation both functions are needed.As linear mixtures of independent signals only have symmetric cross-correlations, anynonzero values of the cosc coefficients would indicate the presence of dynamic interaction.Note that the Fast Fourier Transform earns its moniker thanks to the orthogonalityof sin and cos which allows us to avoid a matrix inversion. However their orthogonalityholds true only for infinite support, and slight correlations are found for finite windows -in practice this effect requires further computation (windowing) to counteract. The cardinalbasis is not orthogonal, requires full regression and may have demanding memoryrequirements. For moderate size data this not problematic and implementation detailswill be discussed elsewhere.8.2. Robustness evaluation based on the NOISE datasetA dataset named NOISE, intended as a benchmark for the bivariate case, has beenintroduced in the preceding NIPS workshop on causality Nolte et al. (2010) and canbe found online at www.causality.inf.ethz.ch/repository.php, alongwith the code that generated the data. It was awarded best dataset prize in the previousNIPS causality workshop and challenge Guyon et al. (2010). For further discussion ofCausality Workbench and current dataset usage see Guyon (2011). NOISE is createdby the summation of the output of a strictly causal VAR DGP and a non-causal SVARDGP which consists of mixed colored noise:y C,i =x N,i =K∑︁k=1K∑︁k=1[︃ ]︃a11 a 120 a 22[︃ ]︃a11 00 a 22C,kN,k(35)y C,i−k + w C,i (36)x N,i−k + w N,i55

Popescuy N,i = Bx N,i (37)y = (1 − |β|)y N + |β|y C‖y N ‖ F‖y C ‖ F(38)The two sub-systems are pictured graphically as systems A and B in Figure 1.If β < 0 the AR matrices that create y C are transposed (meaning that y 1C causes y 2Cinstead of the opposite). The coefficient β is represented in Nolte et al. (2010) by ’γ’ where β = sgn(γ)(1 − |γ |). All coefficients are generated as independent Gaussianrandom variables of unit variance, and unstable systems are discarded. While boththe causal and noise generating systems have the same order, note that the system thatwould generate the sum thereof requires an infinite order SVAR DGP to generate (itis not stochastically equivalent to any SVAR DGP but instead is a SVARMA DGP,having both poles and zeros). Nevertheless it is an interesting benchmark since the exactparameters are not fully recoverable via the commonly used VAR modeling procedureand because the causality interpretation is fairly clear: the sum of a strictly causal DGPand a stochastic noncausal DGP should retain the causality of the former.In this study, the same DGPs were used as in NOISE but as one of the currentaims is to study the influence of sample size on the reliability of causality assignment,signals of 100, 500, 1000 and 5000 points were generated (as opposed to the original6000). This is the dataset referred to as PAIRS below, which only differs in numbersof samples per time series. For each evaluation 500 time series were simulated, withthe order for each system of being uniformly distributed from 1 to 10. The followingmethods were evaluated:∙ PSI (Ψ) using Welch’s method, and segment and epoch lengths being equal andset to ⌈︁ √N⌉︁and otherwise is the same as Nolte et al. (2010).∙ Directed transfer function DTF. estimation using an automatic model order selectioncriterion (BIC, Bayesian Information Criterion) using a maximum modelorder of 20. DTF has been shown to be equivalent to GC for linear AR models(Kaminski et al., 2001) and therefore GC itself is not shown . The covariancematrix of the residuals is also included in the estimate of the transfer function.The same holds for all methods described below.∙ Partial directed coherence PDC. As described in the previous section, it is similarto DTF except it operates on the signal-to-innovations (i.e. inverse) transferfunction.∙ Causal Structural Information. As a described above this is based on the triangularinnovations equivalent to the estimated SVAR (of which there are 2 possibleforms in the bivariate case) and which takes into account instantaneous interaction/ innovations process covariance.56

Robust Statistics for CausalityAll methods were statistically evaluated for robustness and generality by performinga 5-fold jackknife, which gave both a mean and standard deviation estimate for eachmethod and each simulation. All statistics reported below are mean normalized by standarddeviation (from jackknife). For all methods the final output could be -1, 0, or 1,corresponding to causality assignment 1→2, no assignment, and causality 2 → 1. A truepositive (TP) was the rate of correct causality assignment, while a false positive (FP)was the rate of incorrect causality assignment (type III error), such that TP+FP+NA=1,where NA stands for rate of no assignment (or neutral assignment). The TP and FPrates are be co-modulated by increasing/decreasing the threshold of the absolute valueof the mean/std statistic, under which no causality assignment is made:S T AT = rawS T AT/std(rawS T AT), rawSTAT=PSI, DTF, PDC ..c = sign(S T AT) if S T AT > TRES H, 0 otherwiseIn Table 1 we see results of overall accuracy and controlled True Positive rate forthe non-mixed colored noise case (meaning the matrix B above is diagonal). In Table 1and Table 2 methods are ordered according to the mean TP rate over time series length(highest on top).Table 1: Unmixed colored noise PAIRSMax. Accuracy TP , FP < 0.10100 500 1000 5000 100 500 1000 5000Ψ 0.62 0.73 0.83 0.88 0.25 0.56 0.75 0.85DTF 0.58 0.79 0.82 0.88 0.18 0.58 0.72 0.86CSI 0.62 0.72 0.79 0.89 0.23 0.53 0.66 0.88Ψ C 0.57 0.68 0.81 0.88 0.19 0.29 0.70 0.87PDC 0.64 0.67 0.75 0.78 0.23 0.33 0.48 0.57In Table 2 we can see results for a PAIRS, in which the noise mixing matrix B isnot strictly diagonal.Table 2: Mixed colored noise PAIRSMax. Accuracy TP , FP< 0.10N=100 500 1000 5000 N=100 500 1000 5000Ψ C 0.64 0.74 0.81 0.83 0.31 0.49 0.64 0.73Ψ 0.66 0.76 0.78 0.81 0.25 0.59 0.61 0.71CSI 0.63 0.77 0.79 0.80 0.27 0.62 0.59 0.66PDC 0.64 0.71 0.69 0.66 0.24 0.30 0.29 0.24DTF 0.55 0.61 0.66 0.66 0.11 0.10 0.09 0.1257

Popescu(a) Unmixed colored noise(b) Mixed colored noiseFigure 2: PSI vs. DTF Scatter plots of β vs. STAT (to the left of each panel), TP vs.FP curves for different time series lengths (100, 500,1000 and 500) (right).a) colored unmixed additive noise. b) colored mixed additive noise. DTFis equivalent to Granger Causality for linear systems. All STAT values arejackknife mean normalized by standard deviation.58

Robust Statistics for Causality(a) Unmixed colored noise(b) Mixed colored noiseFigure 3: PSI vs. CSI Scatter plots of β vs. STAT (to the left of each panel), TP vs. FPcurves for different time series lengths (right). a) unmixed additive noise. b)mixed additive noise59

PopescuAs we can see in both Figure 2 and Table 1, all methods are almost equally robustto unmixed colored additive noise (except PDC). However, while addition of mixedcolored noise induces a mild gap in maximum accuracy, it creates a large gap in termsof TP/FP rates. Note the dramatic drop-off of the TP rate of VAR/SVAR based methodsPDC and DTF. Figure 3 shows this most clearly, by a wide scatter of STAT outputsfor DTF around β = 0 that is to say with no actual causality in the time series and acorresponding fall-off of TP vs. FP rates. Note also that PSI methods still allow a fairlyreasonable TP rate determination at low FP rates of 10% even at 100 points per timeseries,while the CSI method was also robust to the addition of colored mixed noise,not showing any significant difference with respect to PSI except a higher FP rate forlonger time series (N=5000). The advantage of PSIcardinal was near PSI in overallaccuracy. In conclusion, DTF (or weak Granger causality) and PDC are not robust withrespect to additive mixed colored noise, although they perform similarly to PSI and CSIfor independent colored noise. 19. Conditional causality assignmentIn multivariate time series analysis we are often concerned with inference of causalrelationship among more than 2 variables, in which the role of a potential commoncause must be accounted for, analogously to vanishing partial correlation in the staticdata case. For this reason the PAIRS data set was extended into a set called TRIPLESin which the degree of common driver influence versus direct coupling was controlled.In effect, the TRIPLES DGP is similar to PAIRS, in that additive noise is mixedcolored noise (in 3 dimensions) but in this case another variable x 3 may drive the pairx 1 , x 2 independently of each other, also with random coefficients (but either one set to1/10 of the other randomly). That is to say, the signal is itself a mixture of one wherethere is a direct one sided causal link among x 1 , x 2 as in PAIRS and one where they areactually independent but commonly driven, according to a parameter χ which at 0 iscommonly driven and at 1 is causal.β < 0 y C,i =⎡K∑︁k=1⎢⎣a 11 a 12 00 a 22 0 y⎤⎥⎦ C,i−k + w C,i0 0 a 33 C,k1. Correlation and rank correlation analysis was performed (for N=5000) to shed light on the reasonfor the discrepancy between PSI and CSI. The linear correlation between rawSTAT and STATwas .87 and .89 for PSI and CSI. No influence of model order K of the simulated system wasseen in the error of either PSI or CSI, where error is estimated as the difference in rank ofrankerr(S T AT) = |rank(β) − rank(S T AT)|. There were however significant correlations betweenrank(|β|) and rankerr(S T AT), -.13 for PSI and -.27 for CSI. Note that as expected, standard Grangercausality (GC) performed the same as DTF (TP=0.116 for FP

Robust Statistics for CausalityFigure 4: Diagram of TRIPLES dataset with common driver61

Popescuβ > 0 y C,i =x N,i =⎡K∑︁k=1⎢⎣⎡K∑︁k=1⎢⎣a 11 0 0a 12 a 22 0 y⎤⎥⎦ C,i−k + w C,i (39)0 0 a 33 C,ka 11 0 00 a 22 0 x⎤⎥⎦ N,i−k + w N,i0 0 a 22 N,ky N,i = Bx N,i (40)x D,i =⎡K∑︁k=1⎢⎣a 11 0 a 130 a 22 a 23 x⎤⎥⎦ D,i−k + w D,i0 0 a 22 D,ky MC = (1 − |β|)y N + |β|y C‖y N ‖ F‖y C ‖ F(41)y DC = (1 − χ )y MC + χ y D‖y MC ‖ F‖y D ‖ F(42)The Table 3, similar to the tables in the preceding section, shows results for allusual methods, except for PSIpartial which is PSI calculated on the partial coherenceas defined above and calculated from Welch (cross-spectral) estimators in the case ofmixed noise and a common driver.Table 3: TRIPLES: Commonly driven, additive mixed colored noiseMax. Accuracy TP , FP< 0.10100 500 1000 5000 100 500 1000 5000Ψ p 0.53 0.61 0.71 0.75 0.12 0.31 0.49 0.56Ψ 0.54 0.60 0.70 0.72 0.10 0.25 0.40 0.52CSI 0.51 0.60 0.69 0.76 0.09 0.27 0.38 0.45PDC 0.55 0.54 0.60 0.58 0.13 0.12 0.16 0.13DTF 0.51 0.56 0.59 0.61 0.12 0.09 0.09 0.11Notice that the TP rates are lower for all methods with respect to Table 2 whichrepresents the mixed noise situation without any common driver.10. DiscussionIn a recent talk, Emanuel Parzen (Parzen, 2004) proposed, both in hindsight and forfuture consideration, that aim of statistics consist in an ‘answer machine’, i.e. a moreintelligent, automatic and comprehensive version of Fisher’s almanac, which currentlyconsists in a plenitude of chapters and sections related to different types of hypotheses62

Robust Statistics for Causalityand assumption sets meant to model, insofar as possible, the ever expanding variety ofdata available. These categories and sub-categories are not always distinct, and furthermorethere are competing general approaches to the same problems (e.g. Bayesian vs.frequentist). Is an ‘answer machine’ realistic in terms of time-series causality, prerequisitesfor which are found throughout this almanac, and which has developed in parallelin different disciplines?This work began by discussing Granger causality in abstract terms, pointing out theimplausibility of finding a general method of causal discovery, since that depends onthe general learning and time-series prediction problem, which are incomputable. However,if any consistent patterns that can be found mapping the history of one time seriesvariable to the current state of another (using non-parametric tests), there is sufficientevidence of causal interaction and the null hypothesis is rejected. Such a determinationstill does not address direction of interaction and relative strength of causal influence,which may require a complete model of the DGP. This study - like many others - reliedon the rather strong assumption of stationary linear Gaussian DGPs but otherwise madeweak assumptions on model order, sampling and observation noise. Are there, instead,more general assumptions we can use? The following is a list of competing approachesin increasing order of (subjectively judged) strength of underlying assumption(s):∙ Non-parametric tests of conditional probability for Granger non-causality rejection.These directly compare the probability distributions P(y 1, j | y 1, j−1..1 ,u j−1..1 )P(y 1, j | y 1, j−1..1 ,u j−1..1 ) to detect a possible statistically significant difference. Proposedapproaches (see chapter in this volume by (Moneta et al., 2011) for adetailed overview and tabulated robustness comparison) include product kerneldensity with kernel smoothing (Chlaß and Moneta, 2010), made robust by bootstrappingand with density distances such as the Hellinger (Su and White, 2008),Euclidean (Szekely and Rizzo, 2004), or completely nonparametric differencetests such Cramer-Von Mises or Kolmogorov-Smirnov. A potential pitfall ofnonparametric approaches is their loss of power for higher dimensionality of thespace over which the probabilities are estimated - aka the curse of dimensionality(Yatchew, 1998). This can occur if the lag order K needed to be considered ishigh, if the system memory is long, or the number of other variables over whichGC must be conditioned (u j−1..1 ) is high. In the case of mixed noise, strongGC estimation would require accounting for all observed variables (which inneuroscience can number in the hundreds). While non-parametric non-causalityrejection is a very useful tool (and could be valid even if the lag considered inanalysis is much smaller than the true lag K), in practice we would require robustestimated of causal direction and relative strength of different factors, which impliesa complete accounting of all relevant factors. As was already discussed, inmany cases Granger non-causality is likely to be rejected in both directions: it isuseful to find the dominant one.∙ General parametric or semi-parametric (black-box) predictive modeling subjectto GC interpretation which can provide directionality, factor analysis and inter-63

Popescupretation of information flow. A large body of literature exists on neural networktime series modeling (in this context see White (2006) ), complemented in recentyears by support vector regression and Bayesian processes. The major concernwith black-box predictive approaches is model validation: does the fact that agiven model features a high cross-validation score automatically imply the implausibilityof another predictive model with equal CV-score that would lead todifferent conclusions about causal structure? A reasonable compromise betweennonlinearity and DGP class restriction can be seen in (Chu and Glymour, 2008)and Ryali et al. (2010), in which the VAR model is augmented by additive nonlinearfunctions of the regressed variable and exogenous input. Robustness tonoise, sample size influence and accuracy of effect strength and direction determinationare open questions.∙ Linear dynamic models which incorporate (and often require) non-Gaussianityin the innovations process such as ICA and TDSEP (Ziehe and Mueller, 1998).See Moneta et al. (2011) in this volume for a description of causality inferenceusing ICA and causal modeling of innovation processes (i.e. independent components).Robustness under permutation is necessary for a principled accounting ofdynamic interaction and partition of innovations process entropy. Note that manyICA variants assume that at most one of the innovations processes is Gaussian,a strong assumption which requires a posteriori checks. To be elucidated is therobustness to filtering and additive noise.∙ Non-stationary Gaussian linear models. In neuroscience non-stationarity is important(the brain may change state abruptly, respond to stimuli, have transientpathological episodes etc). Furthermore accounting for non-stochastic exogenousinputs needs further consideration. Encouragingly, the current study showsthat even in the presence of complex confounders such as common driving signalsand co-variate noise, segments as small as 100 points can yield accurate causalityestimation, such that changes in longer time series can be adaptively tracked.Note that in establishing statistical significance we must take into account signalbandwidth: up-sampling the same process would arbitrarily increase the numberof samples but not the information contained in the signal. See Appendix A for aproposal on non-parametric bandwidth estimation.∙ Linear Gaussian dynamic models: in this work we have considered SVAR butnot wider classes of linear DGPs such as VARMA and heteroskedastic (GARCH)models. In comparing PSI and CSI note that overall accuracy of directionalityassignment was virtually identical, but PSI correlated slightly better with effectsize. While CSI made slightly more errors at low strengths of ‘causality’, PSImade slightly more errors at high strengths. Nevertheless, PSI was most robustto (colored, mixed) noise and hidden driving/conditioning signal (tabulated significanceresults are provided in Appendix A). Jackknifed, normalized estimatescan help establish causality at low strength levels, although a large raw PSI statisticvalue may also suffice. A potential problem with the jackknife (or bootstrap)64

Robust Statistics for Causalityprocedure is the strong stationarity assumption which allows segmentation andrearrangement of the data.Although AR modeling was commonly used to model interaction in time seriesand served as a basis for (linear) Granger causality modeling (Blinowska et al., 2004;Baccalá and Sameshima, 2001), robustness to mixed noise remained a problem, whichthe spectral method PSI was meant to resolve (Nolte et al., 2008). While ‘phase’, ifstructured, already implies prediction, precedence and mutual information among timeseries elements, it was not clear how SVAR methods would reconcile with PSI performance,until now. This prompted the introduction in this article of the causal ARmethod (CSI) which takes into account ‘instantaneous’ causality . A prior study hadshown that strong Granger causality is preserved under addition of colored noise, asopposed to weak (i.e. strictly time ordered) causality Solo (2006). This is consistentwith the results obtained herein. The CSI method, measuring strong Granger Causality,was in fact robust with respect to a type of noise not studied in (Solo, 2006), which ismixed colored noise; other VAR based methods and (weak) Granger causality measureswere not. While and SVAR DGP observed under additive colored noise is a VARMAprocess (the case of the PAIRS and TRIPLES datasets), SVAR modeling did not resultin a severe loss of accuracy. AR processes of longer lags can approximate VARMAprocesses by using higher orders and more parameters, even if doing so increases exposureto over-fit and may have resulted in a small number of outliers. Future workmust be undertaken to ascertain what robustness and specificity advantages result fromVARMA modeling, and if it is worth doing so considering the increased computationaloverload. One of the common ‘defects’ of real-life data are missing/outlier samples,or uneven sampling in time, or that the time stamps of two time series to be comparedare unrelated though overlapping: it is for these case that the method PSIcardinal wasdeveloped and shown to be practically equal in numerical performance to the Welchestimate-based PSI method (though it is slower computationally). Both PSI estimateswere robust to common driver influence even when not based on partial but direct coherencybecause it is the asymmetry in influence of the driver on phase that is measuredrather than its overall strength. While 2-way interaction with conditioning was considered,future work must organize multivariate signals using directed graphs, as in DAGtypestatic causal inference. Although only 1 conditioning signal was analysed in thispaper, the methods apply to higher numbers of background variables. Directed TransferFunction and Partial Directed Coherence did not perform as well under additive colorednoise, but their formulation does address a theoretically important question, namely thepartition of strength of influence among various candidate causes of an observation;CSI also proposes an index for this important purpose. Whether the assumptions aboutstationarity or any other data properties discussed are warranted may be checked byperforming appropriate a posteriori tests. If these tests justify prior assumptions anda correspondingly significant causal effect is observed, we can assign statistical confidenceintervals to the causality structure of the system under study. The ‘almanac’chapter on time series causality is rich and new alternatives are emerging. For the entirecorpus of time-series causality statistics to become an ‘answer machine’, however,65

Popescuit is suggested that a principled bottom-up investigation be undertaken, beginning withthe simple SVAR form studied in this paper and all proposed criteria be quantified:type I, II and III errors, accurate determination of causality strength and direction androbustness in the presence of conditioning variables and colored mixed noise.AcknowledgmentsI would like to thank Guido Nolte for his insightful feedback and informative discussions.ReferencesH. Akaike. On the use of a linear model for the identification of feedback systems.Annals of the Institute of statistical mathematics, 20(1):425–439, 1968.L. A Baccalá and K. Sameshima. Partial directed coherence: a new concept in neuralstructure determination. Biological cybernetics, 84(6):463–474, 2001.L. A. Baccalá, M. A. Nicolelis, C. H. Yu, and M. Oshiro. Structural analysis of neuralcircuits using the theory of directed graphs. Computers and Biomedical Research,an International Journal, 24:7–28, Feb 1991. URL http://www.ncbi.nlm.nih.gov/pubmed/2004525.A. B. Barrett, L. Barnett, and A. K. Seth. Multivariate granger causality and generalizedvariance. Physical Review E, 81(4):041907, April 2010. doi: 10.1103/PhysRevE.81.041907. URL http://link.aps.org/doi/10.1103/PhysRevE.81.041907.B. S. Bernanke, J. Boivin, and P. Eliasz. Measuring the Effects of Monetary Policy: AFactor-Augmented Vector Autoregressive (FAVAR) Approach. Quarterly Journal ofEconomics, 120(1):387–422, 2005.K. J. Blinowska, R. Kuś, and M. Kamiński. Granger causality and informationflow in multivariate processes. Physical Review E, 70(5):050902, November 2004.doi: 10.1103/PhysRevE.70.050902. URL http://link.aps.org/doi/10.1103/PhysRevE.70.050902.P. E. Caines. Weak and strong feedback free processes. IEEE. Trans. Autom . Contr,21:737–739, 1976.N. Chlaß and A. Moneta. Can Graphical Causal Inference Be Extended to NonlinearSettings? EPSA Epistemology and Methodology of Science, pages 63–72, 2010.T. Chu and C. Glymour. Search for additive nonlinear time series causal models. TheJournal of Machine Learning Research, 9:967–991, 2008.66

Robust Statistics for CausalityR. A. Fisher. Statistical Methods for Research Workers. Macmillan Pub Co, 1925.ISBN 0028447301.W. Gersch and G. V. Goddard. Epileptic focus location: spectral analysis method.Science (New York, N.Y.), 169(946):701–702, August 1970. ISSN 0036-8075. URLhttp://www.ncbi.nlm.nih.gov/pubmed/5429908. PMID: 5429908.J. Geweke. Measurement of linear dependence and feedback between multiple timeseries. Journal of the American Statistical Association, 77:304–313, 1982.G. Gigerenzer, Z. Swijtink, T. Porter, L. Daston, J. Beatty, and L. Kruger. The Empireof Chance: How Probability Changed Science and Everyday Life. CambridgeUniversity Press, October 1990. ISBN 052139838X.C. W. J. Granger. Investigating causal relations by econometric models and crossspectralmethods. Econometrica, 37(3):424–438, August 1969. ISSN 00129682.URL http://www.jstor.org/stable/1912791.I. Guyon. Time series analysis with the causality workbench. Journal of MachineLearning Research, Workshop and Conference Proceedings, XX. Time SeriesCausality:XX–XX, 2011.I. Guyon and A. Elisseeff. An introduction to variable and feature selection. JMLR, 3:1157–1182, March 2003.I. Guyon, C. Aliferis, G. Cooper, A. Elisseeff, J.-P. Pellet, P. Spirtes, and A. Statnikov.Design and analysis of the causation and prediction challenge. wcci2008 workshopon causality, hong kong, june 3-4 2008. Journal of Machine Learning ResearchWorkshop and Conference Proceedings, 3:1–33, 2008.I. Guyon, D. Janzing, and B. Schölkopf. Causality: Objectives and assessment. JMLRW&CP, 6:1–38, 2010.P. O. Hoyer, D. Janzing, J. M. Mooij, J. Peters, and B. Schölkopf. Nonlinear causaldiscovery with additive noise models. In D. Koller, D. Schuurmans, Y. Bengio, andL. Bottou, editors, Advances in Neural Information Processing Systems 21, pages689–696, 2009.M Kaminski, M Ding, W A Truccolo, and S L Bressler. Evaluating causal relationsin neural systems: granger causality, directed transfer function and statistical assessmentof significance. Biological Cybernetics, 85(2):145–157, August 2001. ISSN0340-1200. URL http://www.ncbi.nlm.nih.gov/pubmed/11508777.PMID: 11508777.M. J. Kaminski and K. J. Blinowska. A new method of the description of the informationflow in the brain structures. Biological Cybernetics, 65(3):203–210, 1991. ISSN0340-1200. doi: 10.1007/BF00198091. URL http://dblp.uni-trier.de/rec/bibtex/journals/bc/KaminskiB91.67

PopescuA. N. Kolmogorov and A. N. Shiryayev. Selected Works of A.N. Kolmogorov: Probabilitytheory and mathematical statistics. Springer, 1992. ISBN 9789027727978.T. C. Koopmans. Statistical Inference in Dynamic Economic Models, Cowles CommissionMonograph, No. 10. New York: John Wiley & Sons, 1950.G. Lacerda, P. Spirtes, J. Ramsey, and P. O. Hoyer. Discovering Cyclic Causal Modelsby Independent Component Analysis. In Proceedings of the 24th Conference onUncertainty in Artificial Intelligence (UAI-2008), Helsinki, Finland, 2008.A. D. Lanterman. Schwarz, wallace and rissanen: Intertwining themes in theories ofmodel selection. International Statistical Review, 69(2):185–212, 2001.M. Li and P. M. B. Vitanyi. An introduction to Kolmogorov complexity and its applications,2nd edition. Springer-Verlag, 1997.A. Moneta, D. Entner, P.O. Hoyer, and A. Coad. Causal inference by independent componentanalysis with applications to micro-and macroeconomic data. Jena EconomicResearch Papers, 2010:031, 2010.A. Moneta, N. Chlaß, D. Entner, and P.Hoyer. Causal search in structural vector autoregressionmodels. Journal of Machine Learning Research, Workshop and ConferenceProceedings, XX. Time Series Causality:XX–XX, 2011.G. Nolte, A. Ziehe, V.V. Nikulin, A. Schlögl, N. Krämer, T. Brismar, and K.-R. Müller.Robustly estimating the flow direction of information in complex physical systems.Physical Review Letters, 00(23):234101, 2008.G. Nolte, A. Ziehe, N. Kraemer, F. Popescu, and K.-R. Müller. Comparison of grangercausality and phase slope index. Journal of Machine Learning Research Workshop& Conference Proceedings., Causality: Objectives and Assessment:267:276, 2010.E. Parzen. Long memory of statistical time series modeling. presented at the 2004nber/nsf time series conference at smu, dallas, usa. Technical report, Texas A&MUniversity, http://www.stat.tamu.edu/~eparzen/Long%20Memory%20of%20Statistical%20Time%20Series%20Modeling.pdf, 2004.J. Pearl. Causality: models, reasoning and inference. Cambridge University Press,Cambridge, 2000.K. Pearson. Tables for statisticians and biometricians,. University Press, UniversityCollege, London, [Cambridge Eng., 1930.Peter C.B. Phillips. The problem of identification in finite parameter continous timemodels. Journal of Econometrics, 1:351–362, 1973.F. Popescu. Identification of sparse multivariate autoregressive models. Proceedings ofthe European Signal Processing Conference (EUSIPCO 2008), Lausanne, Switzerland,2008.68

Robust Statistics for CausalityT. Richardson and P. Spirtes. Automated discovery of linear feedback models. InComputation, causation and discovery. AAAI Press and MIT Press, Menlo Park,1999.A. Roebroeck, A. K. Seth, and P. Valdes-Sosa. Causal time series analysis of functionalmagnetic resonance imaging data. Journal of Machine Learning Research, Workshopand Conference Proceedings, XX. Time Series Causality:XX–XX, 2011.S. Ryali, K. Supekar, T. Chen, and V. Menon. Multivariate dynamical systems modelsfor estimating causal interactions in fmri. Neuroimage, 2010.K. Sameshima and L. A Baccalá. Using partial directed coherence to describe neuronalensemble interactions. Journal of Neuroscience Methods, 94(1):93:103, 1999.R. Scheines, P. Spirtes, C. Glymour, C. Meek, and T. Richardson. The TETRADproject: Constraint based aids to causal model specification. Multivariate BehavioralResearch, 33(1):65–117, 1998.T. Schreiber. Measuring information transfer. Physical Review Letters, 85(2):461, July2000. doi: 10.1103/PhysRevLett.85.461. URL http://link.aps.org/doi/10.1103/PhysRevLett.85.461.C. A. Sims. An autoregressive index model for the u.s. 1948-1975. In J. Kmenta andJ.B. Ramsey, editors, Large-scale macro-econometric models: theory and practice,pages 283–327. North-Holland, 1981.V. Solo. On causality i: Sampling and noise. Proceedings of the 46th IEEE Conferenceon Decision and Control, pages 3634–3639, 2006.P. Spirtes, C. Glymour, and R. Scheines. Causation, prediction, and search. MIT Press,Cambridge MA, 2nd edition, 2000.L. Su and H. White. A nonparametric Hellinger metric test for conditional independence.Econometric Theory, 24(04):829–864, 2008.G. J. Szekely and M. L. Rizzo. Testing for equal distributions in high dimension.InterStat, 5, 2004.A. M. Turing. On computable numbers, with an application to the entscheidungsproblem.Proceedings of the London Mathematical Society, 42:230–65, 1936.J. M Valdes-Sosa, P. A aand Sanchez-Bornot, A. Lage-Castellanos, M. Vega-Hernandez, J. Bosch-Bayard, L. Melie-Garca, and E. Canales-Rodriguez. Estimatingbrain functional connectivity with sparse multivariate autoregression. Neuroinformatics,360(1457):969, 2005.H White. Approximate nonlinear forecasting methods. In G. Elliott, C. W. J. Granger,and A. Timmermann, editors, Handbook of Economic Forecasting, chapter 9, pages460–512. Elsevier, New York, 2006.69

PopescuH. White and X. Lu. Granger Causality and Dynamic Structural Systems. Journal ofFinancial Econometrics, 8(2):193, 2010.N. Wiener. The theory of prediction. Modern mathematics for engineers, Series, 1:125–139, 1956.H. O. Wold. A Study in the Analysis of Stationary Time Series. Stockholm: Almqvistand Wiksell., 1938.A. Yatchew. Nonparametric regression techniques in economics. Journal of EconomicLiterature, 36(2):669–721, 1998.G. U. Yule. Why do we sometimes get nonsense correlations between time series?a study in sampling and the nature of time series. Journal of the Royal StatisticalSociety, 89:1–64, 1926.A. Ziehe and K.-R. Mueller. Tdsep- an efficient algorithm for blind separation usingtime structure. ICANN Proceedings, pages 675–680, 1998.S. T. Ziliak and D. N. McCloskey. The Cult of Statistical Significance: How the StandardError Costs Us Jobs, Justice, and Lives. University of Michigan Press, February2008. ISBN 0472050079.Appendix A. Statistical significance tables for Type I and Type III errorsIn order to assist practitioners in evaluating the statistical significance of bivariatecausality testing, tables were prepared for type I and type III error probabilities as definedin (1) for different values of the base statistic. Below tables are provided for boththe jacknifed statistic Ψ/std(Ψ) and for the raw statistic Ψ, which is needed in case thenumber of points is too low to allow a jackknife/cross-validation/bootstrap or computationalspeed is at a premium. The spectral evaluation method is Welch’s method asdescribed in Section 6. There were 2000 simulations for each condition. The tables inthis Appendix differ in one important aspect with respect to those in the main text. Inorder to avoid non-informative comparison of datasets which are, for example, analysesof the same physical process sampled at different sampling rates, the number of pointsis scaled by the ‘effective’ number of points which is essentially the number of samplesrelative to a simple estimate of the observed signal bandwidth:N * = Nτ S /̂︂BŴ︂BW =‖X‖ F‖∆X/∆T‖ FThe values marked with an asterisk have values of both α and γ which are less than5%. Note also that Ψ is non-dimensional index.70

Robust Statistics for CausalityTable 4: α vs. Ψ/std(Ψ)N * → 50 100 200 500 750 1000 1500 2000 50000.125 0.82 0.83 0.86 0.87 0.88 0.90 0.89 0.89 0.890.25 0.67 0.69 0.73 0.77 0.76 0.78 0.78 0.78 0.770.5 0.41 0.44 0.50 0.54 0.55 0.56 0.56 0.59 0.570.75 0.26 0.26 0.32 0.36 0.36 0.37 0.38 0.40 0.391 0.15 0.15 0.20 0.23 0.23 0.25 0.25 0.26 0.261.25 0.09 0.09 0.11 0.13 0.14 0.16 0.14 0.15 0.161.5 0.06 0.05 0.06 0.07 0.08 0.09 0.08 0.10 0.101.75 0.04 0.03 0.04 0.05 0.05 * 0.05 * 0.04 * 0.06 0.062 0.03 0.02 0.03 0.02 * 0.02 * 0.03 * 0.02 * 0.03 * 0.03 *2.5 0.01 0.01 0.01 0.01 * 0.01 * 0.01 * 0.00 * 0.01 * 0.01 *3 0.01 0.00 0.00 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *4 0.00 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *8 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *16 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *Table 5: γ vs. Ψ/std(Ψ)N * → 50 100 200 500 750 1000 1500 2000 50000.125 0.46 0.41 0.35 0.26 0.23 0.21 0.19 0.18 0.160.25 0.46 0.39 0.33 0.24 0.21 0.19 0.18 0.16 0.140.5 0.44 0.36 0.29 0.21 0.18 0.15 0.14 0.13 0.120.75 0.43 0.35 0.28 0.17 0.14 0.12 0.11 0.10 0.091 0.43 0.31 0.23 0.13 0.12 0.09 0.08 0.07 0.061.25 0.42 0.31 0.20 0.09 0.08 0.06 0.05 0.05 0.041.5 0.40 0.26 0.20 0.06 0.05 0.04 0.04 0.04 0.021.75 0.42 0.26 0.16 0.05 0.04 * 0.02 * 0.02 * 0.03 0.012 0.41 0.23 0.11 0.03 * 0.03 * 0.02 * 0.02 * 0.02 * 0.01 *2.5 0.41 0.20 0.06 0.02 * 0.02 * 0.01 * 0.01 * 0.01 * 0.01 *3 0.33 0.09 0.06 0.03 * 0.01 * 0.00 * 0.00 * 0.00 * 0.00 *4 0.33 0.00 * 0.00 * 0.02 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *8 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *16 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *71

PopescuTable 6: α vs. ΨN * → 50 100 200 500 750 1000 1500 2000 50000.125 0.25 0.22 0.25 0.24 0.24 0.24 0.21 0.21 0.150.25 0.09 0.09 0.09 0.09 0.10 0.09 0.09 0.08 0.05 *0.5 0.02 * 0.01 * 0.01 * 0.02 * 0.02 * 0.02 * 0.01 * 0.01 * 0.01 *0.75 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *1 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *Table 7: γ vs. ΨN * → 50 100 200 500 750 1000 1500 2000 50000.125 0.21 0.17 0.14 0.10 0.08 0.07 0.06 0.05 0.030.25 0.11 0.08 0.06 0.04 0.03 0.03 0.02 0.02 0.01 *0.5 0.03 * 0.01 * 0.01 * 0.01 * 0.01 * 0.00 * 0.00 * 0.01 * 0.00 *0.75 0.02 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *1 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 * 0.00 *72

JMLR: Workshop and Conference Proceedings 12:65–94, 2011Causality in Time SeriesCausal Time Series Analysis of functional MagneticResonance Imaging DataAlard RoebroeckFaculty of Psychology & NeuroscienceMaastricht University, the NetherlandsAnil K. SethSackler Centre for Consciousness ScienceUniversity of Sussex, UKPedro Valdes-SosaCuban Neuroscience Centre, Playa, Cubaa.roebroeck@maastrichtuniversity.nla.k.seth@sussex.ac.ukpeter@cneuro.edu.cuEditors: Florin Popescu and Isabelle GuyonAbstractThis review focuses on dynamic causal analysis of functional magnetic resonance(fMRI) data to infer brain connectivity from a time series analysis and dynamicalsystems perspective. Causal influence is expressed in the Wiener-Akaike-Granger-Schweder (WAGS) tradition and dynamical systems are treated in a state space modelingframework. The nature of the fMRI signal is reviewed with emphasis on theinvolved neuronal, physiological and physical processes and their modeling as dynamicalsystems. In this context, two streams of development in modeling causalbrain connectivity using fMRI are discussed: time series approaches to causality in adiscrete time tradition and dynamic systems and control theory approaches in a continuoustime tradition. This review closes with discussion of ongoing work and futureperspectives on the integration of the two approaches.Keywords: fMRI, hemodynamics, state space model, Granger causality, WAGS influence1. IntroductionUnderstanding how interactions between brain structures support the performance ofspecific cognitive tasks or perceptual and motor processes is a prominent goal in cognitiveneuroscience. Neuroimaging methods, such as Electroencephalography (EEG),Magnetoencephalography (MEG) and functional Magnetic Resonance Imaging (fMRI)are employed more and more to address questions of functional connectivity, interregioncoupling and networked computation that go beyond the ‘where’ and ‘when’ oftask-related activity (Friston, 2002; Horwitz et al., 2000; McIntosh, 2004; Salmelin andKujala, 2006; Valdes-Sosa et al., 2005a). A network perspective onto the parallel anddistributed processing in the brain - even on the large scale accessible by neuroimagingmethods - is a promising approach to enlarge our understanding of perceptual, cognitiveand motor functions. Functional Magnetic Resonance Imaging (fMRI) in particular isc○ 2011 A. Roebroeck, A.K. Seth & P. Valdes-Sosa.

Roebroeck Seth Valdes-Sosaincreasingly used not only to localize structures involved in cognitive and perceptualprocesses but also to study the connectivity in large-scale brain networks that supportthese functions.Generally a distinction is made between three types of brain connectivity. Anatomicalconnectivity refers to the physical presence of an axonal projection from one brainarea to another. Identification of large axon bundles connecting remote regions in thebrain has recently become possible non-invasively in vivo by diffusion weighted Magneticresonance imaging (DWMRI) and fiber tractography analysis (Johansen-Berg andBehrens, 2009; Jones, 2010). Functional connectivity refers to the correlation structure(or more generally: any order of statistical dependency) in the data such that brain areascan be grouped into interacting networks. Finally, effective connectivity modelingmoves beyond statistical dependency to measures of directed influence and causalitywithin the networks constrained by further assumptions (Friston, 1994).Recently, effective connectivity techniques that make use of the temporal dynamicsin the fMRI signal and employ time series analysis and systems identification theoryhave become popular. Within this class of techniques two separate developments havebeen most used: Granger causality analysis (GCA; Goebel et al., 2003; Roebroecket al., 2005; Valdes-Sosa, 2004) and Dynamic Causal Modeling (DCM; Friston et al.,2003). Despite the common goal, there seem to be differences between the two methods.Whereas GCA explicitly models temporal precedence and uses the concept ofGranger causality (or G-causality) mostly formulated in a discrete time-series analysisframework, DCM employs a biophysically motivated generative model formulatedin a continuous time dynamic system framework. In this chapter we will give a generalcausal time-series analysis perspective onto both developments from what we havecalled the Wiener-Akaike-Granger-Schweder (WAGS) influence formalism (Valdes-Sosa et al., in press).Effective connectivity modeling of neuroimaging data entails the estimation of multivariatemathematical models that benefits from a state space formulation, as we willdiscuss below. Statistical inference on estimated parameters that quantify the directedinfluence between brain structures, either individually or in groups (model comparison)then provides information on directed connectivity. In such models, brain structures aredefined from at least two viewpoints. From a structural viewpoint they correspond to aset of “nodes" that comprise a graph, the purpose of causal discovery being the identificationof active links in the graph. The structural model contains i) a selection of thestructures in the brain that are assumed to be of importance in the cognitive process ortask under investigation, ii) the possible interactions between those structures and iii)the possible effects of exogenous inputs onto the network. The exogenous inputs maybe under control of the experimenter and often have the form of a simple indicator functionthat can represent, for instance, the presence or absence of a visual stimulus in thesubject’s view. From a dynamical viewpoint brain structures are represented by statesor variables that describe time varying neural activity within a time-series model of themeasured fMRI time-series data. The functional form of the model equations can em-74

Causal analysis of fMRIbed assumptions on signal dynamics, temporal precedence or physiological processesfrom which signals originate.We start this review by focusing on the nature of the fMRI signal in some detailin section 2, separating the treatment into neuronal, physiological and physical processes.In section 3 we review two important formal concepts: causal influence in theWiener-Akaike-Granger-Schweder tradition and the state space modeling framework,with some emphasis on the relations between discrete and continuous time series models.Building on this discussion, section 4 reviews time series modeling of causality infMRI data. The review proceeds somewhat chronologically, discussing and comparingthe two separate streams of development (GCA and DCM) that have recently begun tobe integrated. Finally, section 5 summarizes and discusses the main topics in generaldynamic state space models of brain connectivity and provides an outlook on futuredevelopments.2. The fMRI SignalThe fMRI signal reflects the activity within neuronal populations non-invasively withexcellent spatial resolution (millimeters down to hundreds of micrometers at high fieldstrength), good temporal resolution (seconds down to hundreds of milliseconds) andwhole-brain coverage of the human or animal brain (Logothetis, 2008). Although fMRIis possible with a few different techniques, the Blood Oxygenation Level Dependent(BOLD) contrast mechanism is employed in the great majority of cases. In short, theBOLD fMRI signal is sensitive to changes in blood oxygenation, blood flow and bloodvolume that result from oxidative glucose metabolism which, in turn, is needed to fuellocal neuronal activity (Buxton et al., 2004). This is why fMRI is usually classified as a‘metabolic’ or ‘hemodynamic’ neuroimaging modality. Its superior spatial resolution,in particular, distinguishes it from other functional brain imaging modalities used inhumans, such as EEG, MEG and Positron Emission Tomography (PET). Although itstemporal resolution is far superior to PET (another ‘metabolic’ neuroimaging modality)it is still an order of magnitude below that of EEG and MEG, resulting in a relativelysparse sampling of fast neuronal processes, as we will discuss below. The final fMRIsignal arises from a complex chain of processes that we can classify into neuronal,physiological and physical processes (Uludag et al., 2005), each of which contain somecrucial parameters and variables and have been modeled in various ways as illustratedin Figure 1. We will discuss each of the three classes of processes to explain the intricaciesinvolved in trying to model this causal chain of events with the ultimate goal ofestimating neuronal activity and interactions from the measured fMRI signal.On the neuronal level, it is important to realize that fMRI reflects certain aspectsof neuronal functioning more than others. A wealth of processes are continuously inoperation at the microscopic level (i.e. in any single neuron), including maintaininga resting potential, post-synaptic conduction and integration (spatial and temporal) ofgraded excitatory and inhibitory post synaptic potentials (EPSPs and IPSPs) arrivingat the dendrites, subthreshold dynamic (possibly oscillatory) potential changes, spikegeneration at the axon hillock, propagation of spikes by continuous regeneration of75

Roebroeck Seth Valdes-SosaFigure 1: The neuronal, physiological and physical processes (top row) and variablesand parameters involved (middle row) in the complex causal chain of eventsthat leads to the formation of the fMRI signal. The bottom row lists somemathematical models of the sub-processes that play a role in the analysis andmodeling of fMRI signals. See main text for further explanation.76

Causal analysis of fMRIthe action potential along the axon, and release of neurotransmitter substances into thesynaptic cleft at arrival of an action potential at the synaptic terminal. There are manydifferent types of neurons in the mammalian brain that express these processes in differentdegrees and ways. In addition, there are other cells, such as glia cells, that performimportant processes, some of them possibly directly relevant to computation or signaling.As explained below, the fMRI signal is sensitive to the local oxidative metabolismin the brain. This means that, indirectly, it mainly reflects the most energy consumingof the neuronal processes. In primates, post-synaptic processes account for the greatmajority (about 75%) of the metabolic costs of neuronal signaling events (Attwell andIadecola, 2002). Indeed, the greater sensitivity of fMRI to post-synaptic activity, ratherthan axon generation and propagation (‘spiking’), has been experimentally verified. Forinstance, in a simultaneous invasive electrophysiology and fMRI measurement in theprimate, Logothetis and colleagues (Logothetis et al., 2001) found the fMRI signal tobe more correlated to the mean Local Field Potential (LFP) of the electrophysiologicalsignal, known to reflect post-synaptic graded potentials, than to high-frequency andmulti-unit activity, known to reflect spiking. In another study it was shown that, bysuppressing action potentials while keeping LFP responses intact by injecting a serotoninagonist, the fMRI response remained intact, again suggesting that LFP is a betterpredictor for fMRI activity (Rauch et al., 2008). These results confirmed earlier resultsobtained on the cerebellum of rats (Thomsen et al., 2004).Neuronal activity, dynamics and computation can be modeled at a different levelsof abstraction, including the macroscopic (whole brain areas), mesoscopic (sub-areasto cortical columns) and microscopic level (individual neurons or groups of these).The levels most relevant to modeling fMRI signals are at the macro- and mesoscopiclevels. Macroscopic models used to represent considerable expanses of gray matter tissueor sub-cortical structures as Regions Of Interest (ROIs) prominently include singlevariable deterministic (Friston et al., 2003) or stochastic (autoregressive; Penny et al.,2005; Roebroeck et al., 2005; Valdes-Sosa et al., 2005b) exponential activity decaymodels. Although the simplicity of such models entail a large degree of abstractionin representing neuronal activity dynamics, their modest complexity is generally wellmatched to the limited temporal resolution available in fMRI. Nonetheless, more complexmulti-state neuronal dynamics models have been investigated in the context offMRI signal generation. These include the 2 state variable Wilson-Cowan model (Marreiroset al., 2008), with one excitatory and one inhibitory sub-population per ROI andthe 3 state variable Jansen-Rit model with a pyramidal excitatory output populationand an inhibitory and excitatory interneuron population, particularly in the modeling ofsimultaneously acquired fMRI and EEG (Valdes-Sosa et al., 2009).The physiology and physics of the fMRI signal is most easily explained by startingwith the physics. We will give a brief overview here and refer to more dedicatedoverviews (Haacke et al., 1999; Uludag et al., 2005) for extended treatment. The hallmarkof Magnetic Resonance (MR) spectroscopy and imaging is the use of the resonancefrequency of magnetized nuclei possessing a magnetic moment, mostly protons(hydrogen nuclei, 1H), called ‘spins’. Radiofrequency antennas (RF coils) can measure77

Roebroeck Seth Valdes-Sosasignal from ensembles of spins that resonate in phase at the moment of measurement.The first important physical factor in MR is the main magnetic field strength (B 0 ),which determines both the resonance frequency (directly proportional to field-strength)and the baseline signal-to-noise ratio of the signal, since higher fields make a larger proportionof spins in the tissue available for measurement. The most used field strengthsfor fMRI research in humans range from 1,5T (Tesla) to 7T. The second importantphysical factor – containing several crucial parameters – is the MR pulse-sequence thatdetermines the magnetization preparation of the sample and the way the signal is subsequentlyacquired. The pulse sequence is essentially a series of radiofrequency pulses,linear magnetic gradient pulses and signal acquisition (readout) events (Bernstein et al.,2004; Haacke et al., 1999). An important variable in a BOLD fMRI pulse sequence iswhether it is a gradient-echo (GRE) sequence or a spin-echo (SE) sequence, whichdetermines the granularity of the vascular processes that are reflected in the signal, asexplained later this section. These effects are further modulated by the echo-time (timeto echo; TE) and repetition time (time to repeat; TR) that are usually set by the end-userof the pulse sequence. Finally, an important variable within the pulse sequence is thetype of spatial encoding that is employed. Spatial encoding can primarily be achievedwith gradient pulses and it embodies the essence of ‘Imaging’ in MRI. It is only withspatial encoding that signal can be localized to certain ‘voxels’ (volume elements) in thetissue. A strength of fMRI as a neuroimaging technique is that an adjustable trade-offis available to the user between spatial resolution, spatial coverage, temporal resolutionand signal-to-noise ratio (SNR) of the acquired data. For instance, although fMRI canachieve excellent spatial resolution at good SNR and reasonable temporal resolution,one can choose to sacrifice some spatial resolution to gain a better temporal resolutionfor any given study. Note, however, that this concerns the resolution and SNR of thedata acquisition. As explained below, the physiology of fMRI can put fundamental limitationson the nominal resolution and SNR that is achieved in relation to the neuronalprocesses of interest.On the physiological level, the main variables that mediate the BOLD contrast infMRI are cerebral blood flow (CBF), cerebral blood volume (CBV) and the cerebralmetabolic rate of oxygen (CMRO2) which all change the oxygen saturation of the blood(as usefully quantified by the concentration of deoxygenated hemoglobin). The BOLDcontrast is made possible by the fact that oxygenation of the blood changes its magneticsusceptibility, which has an effect on the MR signal as measured in GRE and SEsequences. More precisely, oxygenated and deoxygenated hemoglobin (oxy-Hb anddeoxy-Hb) have different magnetic properties, the former being diamagnetic and thelatter paramagnetic. As a consequence, deoxygenated blood creates local microscopicmagnetic field gradients, such that local spin ensembles dephase, which is reflected in alower MR signal. Conversely oxygenation of blood above baseline lowers the concentrationof deoxy-Hb, which decreases local spin dephasing and results in a higher MRsignal. This means that fMRI is directly sensitive to the relative amount of oxy- and deoxyHb and to the fraction of cerebral tissue that is occupied by blood (the CBV), whichare controlled by local neurovascular coupling processes. Neurovascular processes, in78

Causal analysis of fMRIturn, are tightly coupled to neurometabolic processes controlling the rate of oxidativeglucose metabolism (the CMRO2) that is needed to fuel neural activity.Naively one might expect local neuronal activity to quickly increase CMRO2 andincrease the local concentration of deoxy-Hb, leading to a lowering of the MR signal.However, this transient increase in deoxy-Hb or the initial dip in the fMRI signal is notconsistently observed and, thus, there is a debate whether this signal is robust, elusive orsimply not existent (Buxton, 2001; Ugurbil et al., 2003; Uludag, 2010). Instead, earlyexperiments showed that the dynamics of blood flow and blood volume, the hemodynamics,lead to a robust BOLD signal increase. Neuronal activity is quickly followedby a large CBF increase that serves the continued functioning of neurons by clearingmetabolic by-products (such as CO2) and supplying glucose and oxy-Hb. This CBFresponse is an overcompensating response, supplying much more oxy-Hb to the localblood system than has been metabolized. As a consequence, within 1-2 seconds, theoxygenation of the blood increases and the MR signal increases. The increased flowalso induces a ‘ballooning’ of the blood vessels, increasing CBV, the proportion ofvolume taken up by blood, further increasing the signal.Figure 2: A: Simplified causal chain of hemodynamic events as modeled by the balloonmodel. Grey arrows show how variable increases (decreases) tend to relate toeach other. The dynamic changes after a brief pulse of neuronal activity areplotted for CBF (in red), CBV (in purple), deoxyHb (in green) and BOLDsignal (in blue). B: A more abstract representation of the hemodynamic responsefunction as a set of linear basis functions acting as convolution kernels(arbitrary amplitude scaling). Solid line: canonical two-gamma HRF; Dottedline: time derivative; Dashed line: dispersion derivative.A mathematical characterization of the hemodynamic processes in BOLD fMRI at1.5-3T has been given in the biophysical balloon model (Buxton et al., 2004, 1998),schematized in Figure 2A. A simplification of the full balloon model has become importantin causal models of brain connectivity (Friston et al., 2000). In this simplifiedmodel, the dependence of fractional fMRI signal change ∆SS, on normalized cerebral79

Roebroeck Seth Valdes-Sosablood flow f , normalized cerebral blood volume v and normalized deoxyhemoglobincontent q is modeled as:∆S[︂S = V 0 · k 1 · (1 − q) + k 2 · (1 − q ]︂v ) + k 3 · (1 − v)(1)˙v t = 1 (︁ft − v 1/α )︁t(2)τ 0(︁ )︁˙q t = 1 f t 1 − (1 − E0 )τ 0⎛⎜⎝1/ f ⎞t− q tE 0⎟⎠ (3)v 1−1/αtThe term E 0 is the resting oxygen extraction fraction, V 0 is the resting blood volumefraction, τ 0 is the mean transit time of the venous compartment, α is the stiffness componentof the model balloon and {k 1 ,k 2 ,k 3 } are calibration parameters. The main simplificationsof this model with respect to a more complete balloon model (Buxton et al.,2004) are a one-to-one coupling of flow and volume in (2), thus neglecting the actualballoon effect, and a perfect coupling between flow and metabolism in (3). Fristonet al. (2000) augment this model with a putative relation between the a neuronal activityvariable z, a flow-inducing signal s, and the normalized cerebral blood flow f .They propose the following relations in which neuronal activity z causes an increase ina vasodilatory signal that is subject to autoregulatory feedback:ṡ t = z t − 1 τ ss t − 1τ f2 ( f t − 1) (4)˙ f t = s t (5)Here τ s is the signal decay time constant, τ f is the time-constant of the feedback autoregulatorymechanism 1 , and f is the flow normalized to baseline flow. The physiologicalinterpretation of the autoregulatory mechanism is unspecified, leaving us witha neuronal activity variable z that is measured in units of s −2 . The physiology of thehemodynamics contained in differential equations (2) to (5), on the other hand, is morereadily interpretable, and when integrated for a brief neuronal input pulse shows thebehavior as described above (Figure 2A, upper panel). This simulation highlights afew crucial features. First, the hemodynamic response to a brief neural activity eventis sluggish and delayed, entailing that the fMRI BOLD signal is a delayed and lowpassfiltered version of underlying neuronal activity. More than the distorting effects ofhemodynamic processes on the temporal structure of fMRI signals per se, it is the differencein hemodynamics in different parts of the brain that forms a severe confound fordynamic brain connectivity models. Particularly, the delay imposed upon fMRI signalswith respect to the underlying neural activity is known to vary between subjects andbetween different brain regions of the same subject (Aguirre et al., 1998; Saad et al.,2001). Second, although CBF, CBV and deoxyHb changes range in the tens of percents,the BOLD signal change at 1.5T or 3T is in the range of 0.5-2%. Nevertheless,1. Note that we have reparametrized the equation here in terms of τ f 2 to make τ f a proper time constantin units of seconds80

Causal analysis of fMRIthe SNR of BOLD fMRI in general is very good in comparison to electrophysiologicaltechniques like EEG and MEG.Although the balloon model and its variations have played an important role indescribing the transient features of the fMRI response and inferring neuronal activity,simplified ways of representing the BOLD signal responses are very often used. Mostprominent among these is a linear finite impulse response (FIR) convolution with asuitable kernel. The most used single convolution kernel characterizing the ‘canonical’hemodynamic reponse is formed by a superposition of two gamma functions (Glover,1999), the first characterizing the initial signal increase, the second the later negativeundershoot (Figure 2B, solid line):h(t) = m 1 t τ 1e (−l 1t) − cm 2 t τ 2e (−l 2t)m i = max (︁ t τ ie (−l it) )︁ (6)With times-to-peak in seconds τ 1 = 6, τ 2 = 16, scale parameters l i (typically equal to 1)and a relative amplitude of undershoot to peak of c = 1/6.Often, the canonical two-gamma HRF kernel is augmented with one or two additionalorthogonalized convolution kernels: a temporal derivative and a dispersionderivative. Together the convolution kernels form a flexible basis function expansionof possible HRF shapes, with the temporal derivative of the canonical accounting forvariation in the response delay and the dispersion derivative accounting for variations intemporal response width (Henson et al., 2002; Liao et al., 2002). Thus, the linear basisfunction representation is a more abstract characterization of the HRF (i.e. further awayfrom the physiology) that still captures the possible variations in responses.It is an interesting property of hemodynamic processes that, although they are characterizedby a large overcompensating reaction to neuronal activity, their effects arehighly local. The locality of the hemodynamic reponse to neuronal activity limits theactual spatial resolution of fMRI. The path blood inflow in the brain is from large arteriesthrough arterioles into capillaries where exchange with neuronal tissue takes placeat a microscopic level. Blood outflow takes place via venules into the larger veins. Themain regulators of blood flow are the arterioles that are surrounded by smooth muscle,although arteries and capillaries are also thought to be involved in blood flow regulation(Attwell et al., 2010). Different hemodynamic parameters have different spatialresolutions. While CBV and CBF changes in all compartments but mostly venules,oxygenation changes mostly in the venules and veins. Thus, the achievable spatial resolutionwith fMRI is limited by its specificity to the smaller arterioles and venules andmicroscopic capillaries supplying the tissue, rather than the larger supplying arteriesdraining veins. The larger vessels have a larger domain of supply or extraction and, as aconsequence, their signal is blurred and mislocalized with respect to active tissue. Here,physiology and physics interact in an important way. It can be shown theoretically – bythe effects of thermal motion of spin diffusion over time and the distance of the spinsto deoxy-Hb – that the origin of the BOLD signal in SE sequences at high main fieldstrengths (larger than 3T) is much more specific to the microscopic vasculature than tothe larger arteries and veins. This does not hold for GRE sequences or SE sequences81

Roebroeck Seth Valdes-Sosaat lower field strengths. The cost of this greater specificity and higher effective spatialresolution is that SE-BOLD has a lower intrinsic SNR than GRE-BOLD. The balloonmodel equations above are specific to GRE-BOLD at 1.5T and 3T and have been extendedto reflect diffusion effects for higher field strengths (Uludag et al., 2009).In summary, fMRI is an indirect measure of neuronal and synaptic activity. Thephysiological quantities directly determining signal contrast in BOLD fMRI are hemodynamicquantities such as cerebral blood flow and volume and oxygen metabolism.fMRI can achieve a excellent spatial resolution (millimeters down to hundreds of micrometersat high field strength) with good temporal resolution (seconds down to hundredsof milliseconds). The potential to resolve neuronal population interactions at ahigh spatial resolution is what drives attempts at causal time series modeling of fMRIdata. However, the significant aspects of fMRI that pose challenges for such attemptsare i) the enormous dimensionality of the data that contains hundreds of thousands ofchannels (voxels) ii) the temporal convolution of neuronal events by sluggish hemodynamicsthat can differ between remote parts of the brain and iii) the relatively sparsetemporal sampling of the signal.3. Causality and state-space modelsThe inference of causal influence relations from statistical analysis of observed data hastwo dominant approaches. The first approach is in the tradition of Granger causality orG-causality, which has its signature in improved predictability of one time series by another.The second approach is based on graphical models and the notion of intervention(Glymour, 2003), which has been formalized using a Bayesian probabilistic frameworktermed causal calculus or do-calculus (Pearl, 2009). Interestingly, recent work hascombined of the two approaches in a third line of work, termed Dynamic StructuralSystems (White and Lu, 2010). The focus here will be on the first approach, initiallyfirmly rooted in econometrics and time-series analysis. We will discuss this tradition ina very general form, acknowledging early contributions from Wiener, Akaike, Grangerand Schweder and will follow (Valdes-Sosa et al., in press) in refering to the crucialconcept as WAGS influence.3.1. Wiener-Akaike-Granger-Schweder (WAGS) influenceThe crucial premise of the WAGS statistical causal modeling tradition is that a causemust precede and increase the predictability of its effect. In other words: a variableX 2 influences another variable X 1 if the prediction of X 1 improves when we use pastvalues of X 2 , given that all other relevant information (importantly: the past of X1itself) is taken into account. This type of reasoning can be traced back at least toHume and is particularly popular in analyzing dynamical data measured as time series.In a formal framework it was originally proposed (in an abstract form) by Wiener(Wiener, 1956), and then introduced into practical data analysis and popularized byGranger (Granger, 1969). A point stressed by Granger is that increased predictabilityis a necessary but not sufficient condition for a causal relation between time series.82

Causal analysis of fMRIIn fact, Granger distinguished true causal relations – only to be inferred in the presenceof knowledge of the state of the whole universe – from “prima facie" causal relationsthat we refer to as “influence" in agreement with other authors (Commenges andGegout-Petit, 2009). Almost simultaneous with Grangers work, Akaike (Akaike, 1968),and Schweder (Schweder, 1970) introduced similar concepts of influence, prompting(Valdes-Sosa et al., in press) to coin the term WAGS influence (for Wiener-Akaike-Granger-Schweder). This is a generalization of a proposal by placeAalen (Aalen, 1987;Aalen and Frigessi, 2007) who was among the first to point out the connections betweenGranger’s and Schweder’s influence concepts. Within this framework we candefine several general types of WAGS influence, which are applicable to both Markovianand non-Markovian processes, in discrete or continuous time.For three vector time series X 1 (t), X 2 (t), X 3 (t) we wish to know if time series X 1 (t)is influenced by time series X 2 (t) conditional on X 3 (t). Here X 3 (t) can be consideredany set of relevant time series to be controlled for. Let X [a,b] = {X (t),t ∈ [a,b]} denotethe history of a time series in the discrete or continuous time interval [a,b] The firstcategorical distinction is based on what part of the present or future of X 1 (t) can be predictedby the past or present of X 2 (τ 2 ) τ 2 ≤ t. This leads to the following classification(Florens, 2003; Florens and Fougere, 1996):1. If X 2 (τ 2 ) : τ 2 < t , can influence any future value of X 1 (t) it is a global influence.2. If X 2 (τ 2 ) : τ 2 < t , can influence X 1 (t) at time t it is a local influence.3. If X 2 (τ 2 ) : τ 2 = t , can influence X 1 (t) it is a contemporaneous influence.A second distinction is based on predicting the whole probability distribution (stronginfluence) or only given moments (weak influence). Since the most natural formal definitionis one of independence, every influence type amounts to the negation of anindependence statement. The two classifications give rise to six types of independenceand corresponding influence as set out in Table 1.To illustrate, X 1 (t) is strongly, conditionally, and globally independent of X 2 (t)given X 3 (t), ifP( X 1 (∞,t]| X 1 (t,−∞], X 2 (t,−∞], X 3 (t,−∞]) = P( X 1 (∞,t]| X 1 (t,−∞], X 3 (t,−∞])That is: the probability distribution of the future values of X 1 does not depend on thepast of X 2 , given that the influence of the past of both X 1 and X 3 has been taken intoaccount. When this condition does not hold we say X 2 (t) strongly, conditionally, andglobally influences (SCGi) X 1 (t) given X 3 (t). Here we use a convention for intervals[a,b) which indicates that the left endpoint is included but not the right and thatb precedes a. Note that the whole future of X t is included (hence the term “global").And the whole past of all time series is considered. This means these definitions accommodatenon-Markovian processes (for Markovian processes, we only consider theprevious time point). Furthermore, these definitions do not depend on an assumptionof linearity or any given functional relationship between time series. Note also that83

Roebroeck Seth Valdes-SosaTable 1: Types of Influence defined by absence of the corresponding independence relations.See text for acronym definitions.Global( All horizons)Local(Immediate future)ContemporaneousStrong(Probability Distribution)By absence ofstrong, conditional, global independence:X 2 (t)SCGi X 1 (t)||X 3 (t)By absence ofstrong, conditional, local independence:X 2 (t)SCLi X 1 (t)||X 3 (t)By absence ofstrong, conditional, contemporaneousindependence:X 2 (t)SCCi X 1 (t)||X 3 (t)Weak(Expectation)By absence ofweak, conditional, global independence:X 2 (t)WCGi X 1 (t)||X 3 (t)By absence ofweak, conditional, local independence:X 2 (t)WCLi X 1 (t)||X 3 (t)By absence ofweak, conditional, contemporaneousindependence:X 2 (t)WCCi X 1 (t)||X 3 (t)this definition is appropriate for point processes, discrete and continuous time series,even for categorical (qualitative valued) time series. The only problem with this formulationis that it calls on the whole probability distribution and therefore its practicalassessment requires the use of measures such as mutual information that estimate theprobability densities nonparametrically.As an alternative, weak concepts of influence can be defined based on expectations.Consider weak conditional local independence in discrete time, which is defined:E [ X 1 [t + ∆t]| X 1 [t,−∞], X 2 [t,−∞], X 3 [t,−∞]] = E [ X 1 [t + ∆t]| X 1 [t,−∞], X 3 [t,−∞]](7)When this condition does not hold we say X 2 weakly, conditionally and locally influences(WCLi) X 1 given X 3 . To make the implementation this definition insightful,consider a discrete first-order vector auto-regressive (VAR) model for X = [X 1 X 2 X 3 ]:X [t + ∆t] = AX [t] + e[t + ∆t] (8)For this case E [ X[t + ∆t]| X[t,−∞]] = AX [t], and analyzing influence reduces to findingwhich of the autoregressive coefficients are zero. Thus, many proposed operationaltests of WAGS influence, particularly in fMRI analysis, have been formulated as testsof discrete autoregressive coefficients, although not always of order 1. Within the samemodel one can operationalize weak conditional instantaneous independence in dis-84

Causal analysis of fMRIcrete time as zero off-diagonal entries in the co-variance matrix of the innovations e[t]:Σ e = cov[X [t + ∆t]|X [t,−∞]] = E [︀ X [t + ∆t] X ′ [t + ∆t]|X [t,−∞] ]︀In comparison weak conditional local independence in continuous time is defined:E [Y 1 [t]|Y 1 (t,−∞],Y 2 (t,−∞],Y 3 (t,−∞]] = E [Y 1 [t]|Y 1 (t,−∞],Y 3 (t,−∞]] (9)Now consider a first-order stochastic differential equation (SDE) model for Y = [Y 1 Y 2 Y 3 ]:dY = BYdt + dω (10)Then, since ω is a Wiener process with zero-mean white Gaussian noise as a derivative,E [Y[t]|Y(t,−∞]] = BY (t)and analysing influence amounts to estimating the parametersB of the SDE. However, if one were to observe a discretely sampled versionX[k] =Y (k∆t) at sampling interval ∆tand model this with the discrete autoregressive modelabove, this would be inadequate to estimate the SDE parameters for large ∆t, since theexact relations between continuous and discrete system matrices are known to be:A = e B∆t = I + ∞ ∑︀Σ e = ∫︀ t+∆tti=1∆t ii! Bie Bs∑︀ ω e Bs dsThe power series expansion of the matrix exponential in the first line shows A to bea weighted sum of successive matrix powers B i of the continuous time system matrix.Thus, the Awill contain contributions from direct (in B) and indirect (in i steps inB i )causal links between the modeled areas. The contribution of the more indirect links isprogressively down-weighted with the number of causal steps from one area to anotherand is smaller when the sampling interval ∆t is smaller. This makes clear that multivariatediscrete signal models have some undesirable properties for coarsely sampled signals(i.e. a large ∆t with respect to the system dynamics), such as fMRI data. Critically,entirely ruling out indirect influences is not actually achieved merely by employing amultivariate discrete model. Furthermore, estimated WAGS influence (particularly therelative contribution of indirect links) is dependent on the employed sampling interval.However, the discrete system matrix still represents the presence and direction ofinfluence, possibly mediated through other regions.When the goal is to estimate WAGS influence for discrete data starting from a continuoustime model, one has to model explicitly the mapping to discrete time. Mappingcontinuous time predictions to discrete samples is a well known topic in engineeringand can be solved by explicit integration over discrete time steps as performed in (11)above. Although this defines the mapping from continuous to discrete parameters, itdoes not solve the reverse assignment of estimating continuous model parameters fromdiscrete data. Doing so requires a solution to the aliasing problem (Mccrorie, 2003) incontinuous stochastic system identification by setting sufficient conditions on the matrixlogarithm function to make Babove identifiable (uniquely defined) in terms of A.Interesting in this regard is a line of work initiated by Bergstrom (Bergstrom, 1966,(11)85

Roebroeck Seth Valdes-Sosa1984) and Phillips (Phillips, 1973, 1974) studying the estimation of continuous timeAutoregressive models (McCrorie, 2002), and continuous time Autoregressive MovingAverage Models (Chambers and Thornton, 2009) from discrete data. This work restson the observation that the lag zero covariance matrix Σ e will show contemporaneouscovariance even if the continuous covariance matrix Σ ω is diagonal. In other words,the discrete noise becomes correlated over the discrete time-series because the randomfluctuations are aggregated over time. Rather than considering this a disadvantage, thisapproach tries to use both lag information (the AR part) and zero-lag covariance informationto identify the underlying continuous linear model.Notwithstanding the desirability of a continuous time model for consistent inferenceon WAGS influence, there are a few invariances of discrete VAR models, or moregenerally discrete Vector Autoregressive Moving Average (VARMA) models that allowtheir carefully qualified usage in estimating causal influence. The VAR formulation ofWAGS influence has the property of invariance under invertible linear filtering. Moreprecisely, a general measure of influence remains unchanged if channels are each premultipliedwith different invertible lag operators (Geweke, 1982). However, in practicethe order of the estimated VAR model would need to be sufficient to accommodatethese operators. Beyond invertible linear filtering, a VARMA formulation has furtherinvariances. Solo (2006) showed that causality in a VARMA model is preserved undersampling and additive noise. More precisely, if both local and contemporaneous influenceis considered (as defined above) the VARMA measure is preserved under samplingand under the addition of independent but colored noise to the different channels. Finally,Amendola et al. (2010) shows the class of VARMA models to be closed underaggregation operations, which include both sampling and time-window averaging.3.2. State-space modelsA general state-space model for a continuous vector time-series y(t)can be formulatedwith the set of equations:ẋ(t) = f (x(t),v(t),Θ) + ω(t)y(t) = g(x(t),v(t),Θ) + ε(t)This expresses the observed time-series y(t)as a function of the state variables x(t),which are possibly hidden (i.e. unobserved) and observed exogenous inputs v(t), whichare possibly under control. All parameters in the model are grouped into Θ. Notethat some generality is sacrificed from the start since f and gdo not depend on t (Themodel is autonomous and generates stationary processes) or on ω(t) or ε(t), that is:noise enters only additively. The first set of equations, the transition equations or stateequations, describe the evolution of the dynamic system over time in terms of stochasticdifferential equations (SDEs, though technically only when ω(t) = Σẇ(t) with w(t) aWiener process), capturing relations among the hidden state variables x(t) themselvesand the influence of exogenous inputs v(t). The second set of equations, the observationequations or measurement equations, describe how the measurement variables y(t) areobtained from the instantaneous values of the hidden state variables x(t) and the inputs(12)86

Causal analysis of fMRIv(t). In fMRI experiments the exogenous inputs v(t) mostly reflect experimental controland often have the form of a simple indicator function that can represent, for instance,the presence or absence of a visual stimulus. The vector-functions f and g can generallybe non-linear.The state-space formalism allows representation of a very large class of stochasticprocesses. Specifically, it allows representation of both so-called ‘black-box’ models,in which parameters are treated as means to adjust the fit to the data without reflectingphysically meaningful quantities, and ‘grey-box’ models, in which the adjustable parametersdo have a physical or physiological (in the case of the brain) interpretation. Aprominent example of a black-box model in econometric time-series analysis and systemsidentification is the (discrete) Vector Autoregressive Moving Average model withexogenous inputs (VARMAX model) defined as (Ljung, 1999; Reinsel, 1997):F (B)y t = G (B)v t + L(B)e t ⇔∑︀ pi=0 F iy t−i = ∑︀ sj=0G j v t− j + ∑︀ qk=0 L ke t−k(13)Here, the backshift operator B is defined, for any η t as B i η t = η t−i and F, G and Lare polynomials in the backshift operator, such that e.g. F (B) = ∑︀ pi=0 F iB i and p, sand q are the dynamic orders of the VARMAX(p,s,q) model. The minimal constraintson (13) to make it identifiable are F 0 = L 0 = I, which yields the standard VARMAXrepresentation. The VARMAX model and its various reductions (by use of only oneor two of the polynomials, e.g. VAR, VARX or VARMA models) have played a largerole in time-series prediction and WAGS influence modeling. Thus, in the context ofstate space models it is important to consider that the VARMAX model form can beequivalently formulated in a discrete linear state space form:x k+1 = Ax k + Bv k + Ke ky k = Cx k + Dv k + e k(14)In turn the discrete linear state space form can be explicitly accommodated by the continuousgeneral state-space framework in (12) when we define:f (x(t),v(t),Θ) ≃ Fx(t) +Gv(t)g(x(t),v(t),Θ) ≃ Hx(t) + Dv(t)ω(t) = ˜Kε(t)Θ = {︁ F,G, H, D, ˜K,Σ e}︁ (15)Again, the exact relations between the discrete and continuous state space parametermatrices can be derived analytically by explicit integration over time (Ljung, 1999).And, as discussed above, wherever discrete data is used to model continuous influencerelations the problems of temporal aggregation and aliasing have to be taken into account.Although analytic solutions for the discretely sampled continuous linear systemsexist, the discretization of the nonlinear stochastic model (12) does not have a uniqueglobal solution. However, physiological models of neuronal population dynamics andhemodynamics are formulated in continuous time and are mostly nonlinear while fMRIdata is inherently discrete with low sampling frequencies. Therefore, it is the discretizationof the nonlinear dynamical stochastic models that is especially relevant to causal87

Roebroeck Seth Valdes-Sosaanalysis of fMRI data. A local linearization approach was proposed by (Ozaki, 1992)as bridge between discrete time series models and nonlinear continuous dynamical systemsmodel. Considering the nonlinear state equation without exogenous input:ẋ(t) = f (x(t)) + ω(t). (16)The essential assumption in local linearization (LL) of this nonlinear system is to considerthe Jacobian matrix J (l,m) = ∂ f l(X)∂X mas constant over the time period [t + ∆t,t]. ThisJacobian plays the same role as the autoregressive matrix in the linear systems above.Integration over this interval gives the solution:x k+1 = x k + J −1 (e J∆t − I) f (x k ) + e k+1 (17)where I is the identity matrix. Note integration should not be computed this way sinceit is numerically unstable, especially when the Jacobian is poorly conditioned. A listof robust and fast procedures is reviewed in (Valdes-Sosa et al., 2009). This solution islocally linear but crucially it changes with the state at the beginning of each integrationinterval; this is how it accommodates nonlinearity (i.e., a state-dependent autoregressionmatrix). As above, the discretized noise shows instantaneous correlations due tothe aggregation of ongoing dynamics within the span of a sampling period. Once again,this highlights the underlying mechanism for problems with temporal sub-sampling andaggregation for some discrete time models of WAGS influence.4. Dynamic causality in fMRI connectivity analysisTwo streams of developments have recently emerged that make use of the temporaldynamics in the fMRI signal to analyse directed influence (effective connectivity):Granger causality analysis (GCA; Goebel et al., 2003; Roebroeck et al., 2005; Valdes-Sosa, 2004) in the tradition of time series analysis and WAGS influence on the one hand,and Dynamic Causal Modeling (DCM; Friston et al., 2003) in the tradition of systemcontrol on the other hand. As we will discuss in the final section, these approacheshave recently started developing into an integrated single direction. However, initiallyeach was focused on separate issues that pose challenges for the estimation of causalinfluence from fMRI data. Whereas DCM is formulated as an explicit grey box statespace model to account for the temporal convolution of neuronal events by sluggishhemodynamics, GCA analysis has been mostly aimed at solving the problem of regionselection in the enormous dimensionality of fMRI data.4.1. Hemodynamic deconvolution in a state space approachWhile having a long history in engineering, state space modeling was only introducedrecently for the inference of neural states from neuroimaging signals. The earliest attemptstargeted estimating hidden neuronal population dynamics from scalp-level EEGdata (Hernandez et al., 1996; Valdes-Sosa et al., 1999). This work first advanced theidea that state space models and appropriate filtering algorithms are an important tool to88

Causal analysis of fMRIestimate the trajectories of hidden neuronal processes from observed neuroimaging dataif one can formulate an accurate model of the processes leading from neuronal activityto data records. A few years later, this idea was robustly transferred to fMRI data in theform of DCM (Friston et al., 2003). DCM combines three ideas about causal influenceanalysis in fMRI data (or neuroimaging data in general), which can be understood interms of the discussion of the fMRI signal and state space models above (Daunizeauet al., 2009a).First, neuronal interactions are best modeled at the level of unobserved (latent)signals, instead of at the level of observed BOLD signals. This requires a state spacemodel with a dynamic model of neuronal population dynamics and interactions. Theoriginal model that was formulated for the dynamics of neuronal states x = {x 1 ,..., x N }is a bilinear ODE model:∑︁ẋ = Ax + v j B j x + Cv (18)That is, the noiseless neuronal dynamics are characterized by a linear term (with entriesin A representing intrinsic coupling between populations), an exogenous term (with Crepresenting driving influence of experimental variables) and a bilinear term (with B jrepresenting the modulatory influence of experimental variables on coupling betweenpopulations). More recent work has extended this model, e.g. by adding a quadraticterm (Stephan et al., 2008), stochastic dynamics (Daunizeau et al., 2009b) or multiplestate variables per region (Marreiros et al., 2008).Second, the latent neuronal dynamics are related to observed data by a generative(forward) model that accounts for the temporal convolution of neuronal events by slowand variably delayed hemodynamics. This generative forward model in DCM for fMRIis exactly the (simplified) balloon model set out in section 2. Thus, for every selectedregion a single state variable represents the neuronal or synaptic activity of a local populationof neurons and (in DCM for BOLD fMRI) four or five more represent hemodynamicquantities such as capillary blood volume, blood flow and deoxy-hemoglobincontent. All state variables (and the equations governing their dynamics) that servethe mapping of neuronal activity to the fMRI measurements (including the observationequation) can be called the observation model. Most of the physiologically motivatedgenerative model in DCM for fMRI is therefore concerned with an observation modelencapsulating hemodynamics. The parameters in this model are estimated conjointlywith the parameters quantifying neuronal connectivity. Thus, the forward biophysicalmodel of hemodynamics is ‘inverted’ in the estimation procedure to achieve a deconvolutionof fMRI time series and obtain estimates of the underlying neuronal states. DCMhas also been applied to EEG/MEG, in which case the observation model encapsulatesthe lead-field matrix from neuronal sources to EEG electrodes or MEG sensors (Kiebelet al., 2009).Third, the approach to estimating the hidden state trajectories (i.e. filtering andsmoothing) and parameter values in DCM is cast in a Bayesian framework. In short,Bayes’ theorem is used to combine priors p(Θ|M)and likelihood p(y|Θ, M)into the89

Roebroeck Seth Valdes-Sosamarginal likelihood or evidence:∫︁p(y|M) =p(y|Θ, M) p(Θ|M)dΘ (19)Here, the model M is understood to define the priors on all parameters and the likelihoodthrough the generative models for neuronal dynamics and hemodynamics. Aposterior for the parameters p(Θ|y, M) can be obtained as the distribution over parameterswhich maximizes the evidence (19). Since this optimization problem has no analyticsolution and is intractable with numerical sampling schemes for complex models,such as DCM, approximations must be used. The inference approach for DCM relieson variational Bayes methods (Beal, 2003) that optimize an approximation density q(Θ)to the posterior. The approximation density is taken to have a Gaussian form, which isoften referred to as the “Laplace approximation” (Friston et al., 2007). In addition tothe approximate posterior on the parameters, the variational inference will also resultinto a lower bound on the evidence, sometimes referred to as the “free energy”. Thislower bound (or other approximations to the evidence, such as the Akaike InformationCriterion or the Bayesian Information Criterion) are used for model comparison(Penny et al., 2004). Importantly, these quantities explicitly balance goodness-of-itagainst model complexity as a means of avoiding overfitting.An important limiting aspect of DCM for fMRI is that the models M that are comparedalso (implicitly) contain an anatomical model or structural model that contains i)a selection of the ROIs in the brain that are assumed to be of importance in the cognitiveprocess or task under investigation, ii) the possible interactions between those structuresand iii) the possible effects of exogenous inputs onto the network. In other words, eachmodel M specifies the nodes and edges in a directed (possibly cyclic) structural graphmodel. Since the anatomical model also determines the selected part y of the totaldataset (all voxels) one cannot use the evidence to compare different anatomical models.This is because the evidence of different anatomical models is defined over differentdata. Applications of DCM to date invariably use very simple anatomical models (typicallyemploying 3-6 ROIs) in combination with its complex parameter-rich dynamicalmodel discussed above. The clear danger with overly simple anatomical models is thatof spurious influence: an erroneous influence found between two selected regions thatin reality is due to interactions with additional regions which have been ignored. Prototypicalexamples of spurious influence, of relevance in brain connectivity, are thosebetween unconnected structures A and B that receive common input from, or are intervenedby, an unmodeled region C.4.2. Exploratory approaches for model selectionEarly applications of WAGS influence to fMRI data were aimed at counteracting theproblems with overly restrictive anatomical models by employing more permissiveanatomical models in combination with a simple dynamical model (Goebel et al., 2003;Roebroeck et al., 2005; Valdes-Sosa, 2004). These applications reflect the observation90

Causal analysis of fMRIthat estimation of mathematical models from time-series data generally has two importantaspects: model selection and model identification (Ljung, 1999). In the modelselection stage a class of models is chosen by the researcher that is deemed suitablefor the problem at hand. In the model identification stage the parameters in the chosenmodel class are estimated from the observed data record. In practice, model selectionand identification often occur in a somewhat interactive fashion where, for instance,model selection can be informed by the fit of different models to the data achieved inan identification step. The important point is that model selection involves a mixtureof choices and assumptions on the part of the researcher and the information gainedfrom the data-record itself. These considerations indicate that an important distinctionmust be made between exploratory and confirmatory approaches, especially in structuralmodel selection procedures for brain connectivity. Exploratory techniques useinformation in the data to investigate the relative applicability of many models. Assuch, they have the potential to detect ‘missing’ regions in structural models. Confirmatoryapproaches, such as DCM, test hypotheses about connectivity within a set ofmodels assumed to be applicable. Sources of common input or intervening causes aretaken into account in a multivariate confirmatory model, but only if the employed structuralmodel allows it (i.e. if the common input or intervening node is incorporated inthe model).The technique of Granger Causality Mapping (GCM) was developed to exploreall regions in the brain that interact with a single selected reference region using autoregressivemodeling of fMRI time-series (Roebroeck et al., 2005). By employing asimple bivariate model containing the reference region and, in turn, every other voxel inthe brain, the sources and targets of influence for the reference region can be mapped.It was shown that such an ‘exploratory’ mapping approach can form an important toolin structural model selection. Although a bivariate model does not discern direct fromindirect influences, the mapping approach locates potential sources of common inputand areas that could act as intervening network nodes. In addition, by settling fora bivariate model one trivially avoids the conflation of direct and indirect influencesthat can arise in discrete AR model due to temporal aggregation, as discussed above.Other applications of autoregressive modeling to fMRI data have considered full multivariatemodels on large sets of selected brain regions, illustrating the possibility toestimate high-dimensional dynamical models. For instance, Valdes-Sosa (2004) andValdes-Sosa et al. (2005b) applied these models to parcellations of the entire cortex inconjunction with sparse regression approaches that enforce an implicit structural modelselection within the set of parcels. In another more recent example (Deshpande et al.,2008) a full multivariate model was estimated over 25 ROIs (that were found to be activatedin the investigated task) together with an explicit reduction procedure to pruneregions from the full model as a structural model selection procedure. Additional variantsof VAR model based causal inference that has been applied to fMRI include timevarying influence (Havlicek et al., 2010), blockwise (or ‘cluster-wise’) influence fromone group of variables to another (Barrett et al., 2010; Sato et al., 2010) and frequencydecomposedinfluence (Sato et al., 2009).91

Roebroeck Seth Valdes-SosaThe initial developments in autoregressive modeling of fMRI data led to a numberof interesting applications studying human mental states and cognitive processes, suchas gestural communication (Schippers et al., 2010), top-down control of visual spatialattention (Bressler et al., 2008), switching between executive control and default-modenetworks (Sridharan et al., 2008), fatigue (Deshpande et al., 2009) and the resting state(Uddin et al., 2009). Nonetheless, the lack of AR models to account for the varyinghemodynamics convolving the signals of interest and aggregation of dynamics betweentime samples has prompted a set of validation studies evaluating the conditions underwhich discrete AR models can provide reliable connectivity estimates. In (Roebroecket al., 2005) simulations were performed to validate the use of bivariate AR models inthe face of hemodynamic convolution and sampling. They showed that under these conditions(even without variability in hemodynamics) AR estimates for a unidirectionalinfluence are biased towards inferring bidirectional causality, a well known problemwhen dealing with aggregated time series (Wei, 1990). They then went on to show thatinstead unbiased non-parametric inference for bivariate AR models can be based on adifference of influence terms (X → Y − Y → X). In addition, they posited that inferenceon such influence estimates should always include experimental modulation of influence,in order to rule out hemodynamic variation as an underlying reason for spuriouscausality. In Deshpande et al. (2010) the authors simulated fMRI data by manipulatingthe causal influence and neuronal delays between local field potentials (LFPs) acquiredfrom the macaque cortex and varying the hemodynamic delays of a convolving hemodynamicresponse function and the signal-to-noise ratio (SNR) and the sampling periodof the final simulated fMRI data. They found that in multivariate (4 dimensional) simulationswith hemodynamic and neuronal delays drawn from a uniform random distributioncorrect network detection from fMRI was well above chance and was up to 90%under conditions of fast sampling and low measurement noise. Other studies confirmedthe observation that techniques with intermediate temporal resolution, such as fMRI,can yield good estimates of the causal connections based on AR models (Stevensonand Kording, 2010), even in the face of variable hemodynamics (Ryali et al., 2010).However, another recent simulation study, investigating a host of connectivity methodsconcluded low detection performance of directed influence by AR models undergeneral conditions (Smith et al., 2010).4.3. Toward integrated modelsDavid et al. (2008) aimed at direct comparison of autoregressive modeling and DCMfor fMRI time series and explicitly pointed at deconvolution of variable hemodynamicsfor causality inferences. The authors created a controlled animal experiment wheregold standard validation of neuronal connectivity estimation was provided by intracranialEEG (iEEG) measurements. As discussed extensively in Friston (2009b) and Roebroecket al. (2009a) such a validation experiment can provide important informationon best practices in fMRI based brain connectivity modeling that, however, need to becarefully discussed and weighed. In David et al.’s study, simultaneous fMRI, EEG, andiEEG were measured in 6 rats during epileptic episodes in which spike-and-wave dis-92

Causal analysis of fMRIcharges (SWDs) spread through the brain. fMRI was used to map the hemodynamicresponse throughout the brain to seizure activity, where ictal and interictal states werequantified by the simultaneously recorded EEG. Three structures were selected by theauthors as the crucial nodes in the network that generates and sustains seizure activityand further analysed with i) DCM, ii) simple AR modeling of the fMRI signal andiii) AR modeling applied to neuronal state-variable estimates obtained with a hemodynamicdeconvolution step. By applying G-causality analysis to deconvolved fMRItime-series, the stochastic dynamics of the linear state-space model are augmented withthe complex biophysically motivated observation model in DCM. This step is crucialif the goal is to compare the dynamic connectivity models and draw conclusions onthe relative merits of linear stochastic models (explicitly estimating WAGS influence)and bilinear deterministic models. The results showed both AR analysis after deconvolutionand DCM analysis to be in accordance with the gold-standard iEEG analyses,identifying the most pertinent influence relations undisturbed by variations in HRF latencies.In contrast, the final result of simple AR modeling of the fMRI signal showedless correspondence with the gold standard, due to the confounding effects of differenthemodynamic latencies which are not accounted for in the model.Two important lessons can be drawn from David et al.’s study and the ensuingdiscussions (Bressler and Seth, 2010; Daunizeau et al., 2009a; David, 2009; Friston,2009b,a; Roebroeck et al., 2009a,b). First, it confirms again the distorting effects ofhemodynamic processes on the temporal structure of fMRI signals and, more importantly,that the difference in hemodynamics in different parts of the brain can form aconfound for dynamic brain connectivity models (Roebroeck et al., 2005). Second,state-space models that embody observation models that connect latent neuronal dynamicsto observed fMRI signals have a potential to identify causal influence unbiasedby this confound. As a consequence, substantial recent methodological work has aimedat combining different models of latent neuronal dynamics with a form of a hemodynamicobservation model in order to provide an inversion or filtering algorithm for estimationof parameters and hidden state trajectories. Following the original formulationof DCM that provides a bilinear ODE form for the hidden neuronal dynamics, attemptshave been made at explicit integration of hemodynamics convolution with stochasticdynamic models that are interpretable in the framework of WAGS influence.For instance in (Ryali et al., 2010), following earlier work (Penny et al., 2005;Smith et al., 2009), a discrete state space model with a bi-linear vector autoregressivemodel to quantify dynamic neuronal state evolution and both intrinsic and modulatoryinteractions is proposed:x k = Ax k−1 + ∑︀ j=1 v j k B j x k−1 + Cv j k + ε kx m k = [︁ xk m, xm k−1 ,··· , ]︁xm k−L+1(20)y m k = βm Φx m k + em kHere, we index exogenous inputs with j and ROIs with min superscripts. The entriesin the autoregressive matrix A, exogenous influence matrix C and bi-linear matricesB j have the same interpretation as in deterministic DCM. The relation between93

Roebroeck Seth Valdes-Sosaobserved BOLD-fMRI data yand latent neuronal sources x is modeled by a temporalembedding of into x m for each region or ROI m. This allows convolution with a flexiblebasis function expansion of possible HRF shapes to be represented by a simple matrixmultiplication β m Φxk m in the observation equation. Here Φ contains the temporal basisfunctions in Figure 2B and β m the basis function parameters to be estimated. By estimatingbasis function parameters individually per region, variations in the HRF shapebetween region can be accounted for and the confounding effects of these on WAGSinfluence estimate can be avoided. Ryali et al. found that robust estimates of parametersΘ = {︁ }︁A,B j ,C,β m ,Σ ε ,Σ e and states xk can be obtained from a variational Bayesianapproach. In their simulations, they show that a state-space model with interactionsmodeled at the latent level can compensate well for the effects of HRF variability, evenwhen relative HRF delays are opposed to delayed interactions. Note, however, that subsamplingof the BOLD signal is not explicitly characterized in their state-space model.A few interesting variations on this discrete state-space modeling have recentlybeen proposed. For instance in (Smith et al., 2009) a switching linear systems modelfor latent neuronal state evolution, rather than a bi-linear model was used. This modelrepresents experimental modulation of connections as a random variable, to be learnedfrom the data. This variable switches between different linear system instantiationsthat each characterize connectivity in a single experimental condition. Such a schemehas the important advantage that an n-fold cross validation approach can be used toobtain a measure of absolute model-evidence (rather than relative between a selectedset of models). Specifically, one could learn parameters for each context-specific linearsystem with knowledge of the timing of changing experimental conditions in a trainingdata set. Then the classification accuracy of experimental condition periods in a testdata set based on connectivity will provide a absolute model-fit measure, controlled formodel complexity, which can be used to validate overall usefulness of the fitted model.In particular, this can point to important brain regions missing from the model incaseof poor classification accuracy.Another related line of developments instead has involved generalizing the ODEmodels in DCM for fMRI to stochastic dynamic models formulated in continuous time(Daunizeau et al., 2009b; Friston et al., 2008). An early exponent of this approach usedlocal linearization in a (generalized) Kalman filter to estimate states and parameters ina non-linear SDE models of hemodynamics (Riera et al., 2004). Interestingly, the inclusionof stochastics in the state equations makes inference on coupling parameters ofsuch models usefully interpretable in the framework of WAGS influence. This hints atthe ongoing convergence, in modeling of brain connectivity, of time series approachesto causality in a discrete time tradition and dynamic systems and control theory approachesin a continuous time tradition.5. Discussion and OutlookThe modeling of an enormously complex biological system such as the brain has manychallenges. The abstractions and choices to be made in useful models of brain connectivityare therefore unlikely to be accommodated by one single ‘master’ model that94

Causal analysis of fMRIdoes better than all other models on all counts. Nonetheless, the ongoing developmentefforts towards improved approaches are continually extending and generalizing thecontexts in which dynamic time series models can be applied. It is clear that state spacemodeling and inference on WAGS influence are fundamental concepts within this endeavor.We end here with some considerations of dynamic brain connectivity modelsthat summarize some important points and anticipate future developments.We have emphasized that WAGS influence models of brain connectivity have largelybeen aimed at data driven exploratory analysis, whereas biophysically motivated statespace models are mostly used for hypothesis-led confirmatory analysis. This is especiallyrelevant in the interaction between model selection and model identification.Exploratory techniques use information in the data to investigate the relative applicabilityof many models. As such, they have the potential to detect ‘missing’ regions inanatomical models. Confirmatory approaches test hypotheses about connectivity withina set of models assumed to be applicable.As mentioned above, the WAGS influence approach to statistical analysis of causalinfluence that we focused on here is complemented by the interventional approachrooted in the theory of graphical models and causal calculus. Graphical causal modelshave been recently applied to brain connectivity analysis of fMRI data (Ramseyet al., 2009). Recent work combining the two approaches (White and Lu, 2010) possiblyleads the way to a combined causal treatment of brain imaging data incorporatingdynamic models and interventions. Such a combination could enable incorporation ofdirect manipulation of brain activity by (for example) transcranial magnetic stimulation(Pascual-Leone et al., 2000; Paus, 1999; Walsh and Cowey, 2000) into the current statespace modeling framework.Causal models of brain connectivity are increasingly inspired by biophysical theories.For fMRI this is primarily applicable in modeling the complex chain of eventsseparating neuronal population activity from the BOLD signal. Inversion of such amodel (in state space form) by a suitable filtering algorithm amounts to a model-baseddeconvolution of the fMRI signal resulting in an estimate of latent neuronal populationactivity. If the biophysical model is appropriately formulated to be identifiable (possiblyincluding priors on relevant parameters), it can take variation in the hemodynamicsbetween brain regions into account that can otherwise confound time series causalityanalyses of fMRI signals. Although models of hemodynamics for causal fMRI analysishave reached a reasonable level of complexity, the models of neuronal dynamicsused to date have remained simple, comprising one or two state variables for an entirecortical region or subcortical structure. Realistic dynamic models of neuronal activityhave a long history and have reached a high level of sophistication (Deco et al., 2008;Markram, 2006). It remains an open issue to what degree complex realistic equationsystems can be embedded in analysis of fMRI – or in fact: any brain imaging modality– and result in identifiable models of neuronal connectivity and computation.Two recent developments create opportunities to increase complexity and realismof neuronal dynamics models and move the level of modeling from the macroscopic(whole brain areas) towards the mesoscopic level comprising sub-populations of areas95

Roebroeck Seth Valdes-Sosaand cortical columns. First, the fusion of multiple imaging modalities, possibly simultaneouslyrecorded, has received a great deal of attention. Particularly, several attemptsto model-driven fusion of simultaneousy recorded fMRI and EEG data, by inverting aseparate observation model for each modality while using the same underlying neuronalmodel, have been reported (Deneux and Faugeras, 2010; Riera et al., 2007; Valdes-Sosaet al., 2009). This approach holds great potential to fruitfully combine the superior spatialresolution of fMRI with the superior temporal resolution of EEG. In (Valdes-Sosaet al., 2009) anatomical connectivity information obtained from diffusion tensor imagingand fiber tractography is also incorporated. Second, advances in MRI technology,particularly increases of main field strength to 7T (and beyond) and advances in parallelimaging (de Zwart et al., 2006; Heidemann et al., 2006; Pruessmann, 2004; Wiesingeret al., 2006), greatly increase the level spatial detail that are accessible with fMRI. Forinstance, fMRI at 7T with sufficient spatial resolution to resolve orientation columns inhuman visual cortex has been reported (Yacoub et al., 2008).The development of state space models for causal analysis of fMRI data has movedfrom discrete to continuous and from deterministic to stochastic models. Continuousmodels with stochastic dynamics have desirable properties, chief among them a robustinference on causal influence interpretable in the WAGS framework, as discussedabove. However, dealing with continuous stochastic models leads to technical issuessuch as the properties and interpretation of Wiener processes and Ito calculus (Friston,2008). A number of inversion or filtering methods for continuous stochastic modelshave been recently proposed, particularly for the goal of causal analysis of brain imagingdata, including the local linearization and innovations approach (Hernandez et al.,1996; Riera et al., 2004), dynamic expectation maximization (Friston et al., 2008) andgeneralized filtering (Friston et al., 2010). The ongoing development of these filteringmethods, their validation and their scalability towards large numbers of state variableswill be a topic of continuing research.AcknowledgmentsThe authors thank Kamil Uludag for comments and discussion.ReferencesOdd O. Aalen. Dynamic modeling and causality. Scandinavian Actuarial journal,pages 177–190, 1987.O.O. Aalen and A. Frigessi. What can statistics contribute to a causal understanding?Board of the Foundation of the Scandinavian journal of Statistics, 34:155–168, 2007.G. K. Aguirre, E. Zarahn, and M. D’Esposito. The variability of human, bold hemodynamicresponses. Neuroimage, 8(4):360–9, 1998.H Akaike. On the use of a linear model for the identification of feedback systems.Annals of the Institute of statistical mathematics, 20(1):425–439, 1968.96

Causal analysis of fMRIA Amendola, M Niglio, and C Vitale. Temporal aggregation and closure of VARMAmodels: Some new results. In F. Palumbo et al., editors, Data Analysis and Classification:Studies in Classification, Data Analysis, and Knowledge Organization,pages 435–443. Springer Berlin/Heidelberg, 2010.D. Attwell and C. Iadecola. The neural basis of functional brain imaging signals. TrendsNeurosci, 25(12):621–5, 2002.D. Attwell, A. M. Buchan, S. Charpak, M. Lauritzen, B. A. Macvicar, and E. A. Newman.Glial and neuronal control of brain blood flow. Nature, 468(7321):232–43,2010.A. B. Barrett, L. Barnett, and A. K. Seth. Multivariate granger causality and generalizedvariance. Phys Rev E Stat Nonlin Soft Matter Phys, 81(4 Pt 1):041907, 2010.M.J. Beal. Variational Algorithms for Approximate Bayesian Inference. PhD thesis,University College London, 2003.A R Bergstrom. Nonrecursive models as discrete approximations to systems of stochasticdifferential equations. Econometrica, 34:173–182, 1966.A R Bergstrom. Continuous time stochastic models and issues of aggregation. InZ. Griliches and M.D. Intriligator, editors, Handbook of econometrics, volume II.Elsevier, 1984.M.A. Bernstein, K.F. King, and X.J. Zhou. Handbook of MRI Pulse Sequences. ElsevierAcademic Press, urlington, 2004.S. L. Bressler and A. K. Seth. Wiener-granger causality: A well established methodology.Neuroimage, 2010.S. L. Bressler, W. Tang, C. M. Sylvester, G. L. Shulman, and M. Corbetta. Top-downcontrol of human visual cortex by frontal and parietal cortex in anticipatory visualspatial attention. J Neurosci, 28(40):10056–61, 2008.R. B. Buxton. The elusive initial dip. Neuroimage, 13(6 Pt 1):953–8, 2001.R. B. Buxton, E. C. Wong, and L. R. Frank. Dynamics of blood flow and oxygenationchanges during brain activation: the balloon model. Magn Reson Med, 39(6):855–64,1998.R. B. Buxton, K. Uludag, D. J. Dubowitz, and T. T. Liu. Modeling the hemodynamicresponse to brain activation. Neuroimage, 23 Suppl 1:S220–33, 2004.Marcus J Chambers and Michael A Thornton. Discrete time representation of continuoustime arma processes, 2009.97

Roebroeck Seth Valdes-SosaDaniel Commenges and Anne Gegout-Petit. A general dynamical statistical modelwith possible causal interpretation. journal of the Royal Statistical Society: Series B(Statistical Methodology), 71(3):1–43, 2009.J. Daunizeau, O. David, and K. E. Stephan. Dynamic causal modelling: A criticalreview of the biophysical and statistical foundations. Neuroimage, 2009a.J. Daunizeau, K. J. Friston, and S. J. Kiebel. Variational bayesian identification andprediction of stochastic nonlinear dynamic causal models. Physica D, 238(21):2089–2118, 2009b.O. David. fmri connectivity, meaning and empiricism comments on: Roebroeck et al.the identification of interacting networks in the brain using fmri: Model selection,causality and deconvolution. Neuroimage, 2009.O. David, I. Guillemain, S. Saillet, S. Reyt, C. Deransart, C. Segebarth, and A. Depaulis.Identifying neural drivers with functional mri: an electrophysiological validation.PLoS Biol, 6(12):2683–97, 2008.J. A. de Zwart, P. van Gelderen, X. Golay, V. N. Ikonomidou, and J. H. Duyn. Acceleratedparallel imaging for functional imaging of the human brain. NMR Biomed, 19(3):342–51, 2006.G. Deco, V. K. Jirsa, P. A. Robinson, M. Breakspear, and K. Friston. The dynamicbrain: from spiking neurons to neural masses and cortical fields. PLoS Comput Biol,4(8):e1000092, 2008.T. Deneux and O. Faugeras. Eeg-fmri fusion of paradigm-free activity using kalmanfiltering. Neural Comput, 22(4):906–48, 2010.G. Deshpande, X. Hu, R. Stilla, and K. Sathian. Effective connectivity during hapticperception: a study using granger causality analysis of functional magnetic resonanceimaging data. Neuroimage, 40(4):1807–14, 2008.G. Deshpande, S. LaConte, G. A. James, S. Peltier, and X. Hu. Multivariate grangercausality analysis of fmri data. Hum Brain Mapp, 30(4):1361–73, 2009.G. Deshpande, K. Sathian, and X. Hu. Effect of hemodynamic variability on grangercausality analysis of fmri. Neuroimage, 52(3):884–96, 2010.J Florens. Some technical issues in defining causality. journal of Econometrics, 112:127–128, 2003.J.P. Florens and D. Fougere. Noncausality in continuous time. Econometrica, 64(5):1195–1212, 1996.K. Friston. Functional and effective connectivity in neuroimaging: A synthesis. HumBrain Mapp, 2:56–78, 1994.98

Causal analysis of fMRIK. Friston. Beyond phrenology: what can neuroimaging tell us about distributed circuitry?Annu Rev Neurosci, 25:221–50, 2002.K. Friston. Dynamic causal modeling and granger causality comments on: The identificationof interacting networks in the brain using fmri: Model selection, causalityand deconvolution. Neuroimage, 2009a.K. Friston, J. Mattout, N. Trujillo-Barreto, J. Ashburner, and W. Penny. Variational freeenergy and the laplace approximation. Neuroimage, 34(1):220–34, 2007.K. J. Friston, A. Mechelli, R. Turner, and C. J. Price. Nonlinear responses in fmri:the balloon model, volterra kernels, and other hemodynamics. Neuroimage, 12(4):466–77, 2000.K. J. Friston, L. Harrison, and W. Penny. Dynamic causal modelling. Neuroimage, 19(4):1273–302, 2003.K. J. Friston, N. Trujillo-Barreto, and J. Daunizeau. Dem: a variational treatment ofdynamic systems. Neuroimage, 41(3):849–85, 2008.Karl Friston. Hierarchical models in the brain. PLoS Computational Biology, 4, 2008.Karl Friston. Causal modelling and brain connectivity in functional magnetic resonanceimaging. PLoS biology, 7:e33, 2009b.Karl Friston, Klaas Stephan, Baojuan Li, and Jean Daunizeau. Generalised filtering.Mathematical Problems in Engineering, 2010:1–35, 2010.J. F. Geweke. Measurement of linear dependence and feedback between multiple timeseries. journal of the American Statistical Association, 77(378):304–324, 1982.G. H. Glover. Deconvolution of impulse response in event-related bold fmri. Neuroimage,9(4):416–29, 1999.C. Glymour. Learning, prediction and causal bayes nets. Trends Cogn Sci, 7(1):43–48,2003.R. Goebel, A. Roebroeck, D. S. Kim, and E. Formisano. Investigating directed corticalinteractions in time-resolved fmri data using vector autoregressive modeling andgranger causality mapping. Magn Reson Imaging, 21(10):1251–61, 2003.C. W. J. Granger. Investigating causal relations by econometric models and crossspectralmethods. Econometrica, 37(3):424–438, 1969.E.M. Haacke, R.W. Brown, M.R. Thompson, and R. Venkatesan. Magnetic ResonanceImaging: Physical Principles and Sequence Design. John Wiley and Sons, Inc, NewYork, 1999.99

Roebroeck Seth Valdes-SosaM. Havlicek, J. Jan, M. Brazdil, and V. D. Calhoun. Dynamic granger causality basedon kalman filter for evaluation of functional network connectivity in fmri data. Neuroimage,53(1):65–77, 2010.R. M. Heidemann, N. Seiberlich, M. A. Griswold, K. Wohlfarth, G. Krueger, and P. M.Jakob. Perspectives and limitations of parallel mr imaging at high field strengths.Neuroimaging Clin N Am, 16(2):311–20, 2006.R. N. Henson, C. J. Price, M. D. Rugg, R. Turner, and K. J. Friston. Detecting latencydifferences in event-related bold responses: application to words versus nonwordsand initial versus repeated face presentations. Neuroimage, 15(1):83–97, 2002.J. L. Hernandez, P. A. Valdés, and P. Vila. Eeg spike and wave modelled by a stochasticlimit cycle. NeuroReport, 1996.B. Horwitz, K. J. Friston, and J. G. Taylor. Neural modeling and functional brainimaging: an overview. Neural Netw, 13(8-9):829–46, 2000.H. Johansen-Berg and T.E.J Behrens, editors. Diffusion MRI: From quantitative measurementto in-vivo neuroanatomy. Academic Press, London, 2009.D.K. Jones, editor. Diffusion MRI: Theory, Methods, and Applications. Oxford UniversityPress, Oxford, 2010.S. J. Kiebel, M. I. Garrido, R. Moran, C. C. Chen, and K. J. Friston. Dynamic causalmodeling for eeg and meg. Hum Brain Mapp, 30(6):1866–76, 2009.C. H. Liao, K. J. Worsley, J. B. Poline, J. A. Aston, G. H. Duncan, and A. C. Evans.Estimating the delay of the fmri response. Neuroimage, 16(3 Pt 1):593–606, 2002.L. Ljung. System Identification: Theory for the User. Prentice-Hall, New Jersey, 2ndedition, 1999.N. K. Logothetis. What we can do and what we cannot do with fmri. Nature, 453(7197):869–78, 2008.N. K. Logothetis, J. Pauls, M. Augath, T. Trinath, and A. Oeltermann. Neurophysiologicalinvestigation of the basis of the fmri signal. Nature, 412(6843):150–7, 2001.H. Markram. The blue brain project. Nat Rev Neurosci, 7(2):153–60, 2006.A. C. Marreiros, S. J. Kiebel, and K. J. Friston. Dynamic causal modelling for fmri: atwo-state model. Neuroimage, 39(1):269–78, 2008.J. Roderick McCrorie. The likelihood of the parameters of a continuous time vector autoregressivemodel. Statistical Inference for Stochastic Processes, 5:273–286, 2002.J. Roderick Mccrorie. The problem of aliasing in identifying finite parameter continuoustime stochastic models. Acta Applicandae Mathematicae, 79:9–16, 2003.100

Causal analysis of fMRIA. R. McIntosh. Contexts and catalysts: a resolution of the localization and integrationof function in the brain. Neuroinformatics, 2(2):175–82, 2004.T Ozaki. A bridge between nonlinear time series models and nonlinear stochastic dynamicalsystems: A local linearization approach. Statistica Sinica, 2:113–135, 1992.A. Pascual-Leone, V. Walsh, and J. Rothwell. Transcranial magnetic stimulation in cognitiveneuroscience–virtual lesion, chronometry, and functional connectivity. CurrOpin Neurobiol, 10(2):232–7, 2000.T. Paus. Imaging the brain before, during, and after transcranial magnetic stimulation.Neuropsychologia, 37(2):219–24, 1999.J. Pearl. Causality: Models, Reasoning and Inference. Cambridge University Press,New York, 2nd edition, 2009.W. Penny, Z. Ghahramani, and K. Friston. Bilinear dynamical systems. Philos Trans RSoc Lond B Biol Sci, 360(1457):983–93, 2005.W. D. Penny, K. E. Stephan, A. Mechelli, and K. J. Friston. Comparing dynamic causalmodels. Neuroimage, 22(3):1157–72, 2004.Peter C.B. Phillips. The problem of identification in finite parameter continuous timemodels. journal of Econometrics, 1:351–362, 1973.Peter C.B. Phillips. The estimation of some continuous time models. Econometrica,42:803–823, 1974.K. P. Pruessmann. Parallel imaging at high field strength: synergies and joint potential.Top Magn Reson Imaging, 15(4):237–44, 2004.J. D. Ramsey, S. J. Hanson, C. Hanson, Y. O. Halchenko, R. A. Poldrack, and C. Glymour.Six problems for causal inference from fmri. Neuroimage, 49(2):1545–58,2009.A. Rauch, G. Rainer, and N. K. Logothetis. The effect of a serotonin-induced dissociationbetween spiking and perisynaptic activity on bold functional mri. Proc NatlAcad Sci U S A, 105(18):6759–64, 2008.G.C. Reinsel. Elements of Multivariate Time Series Analysis. Springer-Verlag, NewYork, 2nd edition, 1997.J. J. Riera, J. Watanabe, I. Kazuki, M. Naoki, E. Aubert, T. Ozaki, and R. Kawashima. Astate-space model of the hemodynamic approach: nonlinear filtering of bold signals.Neuroimage, 21(2):547–67, 2004.J. J. Riera, J. C. Jimenez, X. Wan, R. Kawashima, and T. Ozaki. Nonlinear localelectrovascular coupling. ii: From data to neuronal masses. Hum Brain Mapp, 28(4):335–54, 2007.101

Roebroeck Seth Valdes-SosaA. Roebroeck, E. Formisano, and R. Goebel. Mapping directed influence over the brainusing granger causality and fmri. Neuroimage, 25(1):230–42, 2005.A. Roebroeck, E. Formisano, and R. Goebel. The identification of interacting networksin the brain using fmri: Model selection, causality and deconvolution. Neuroimage,2009a.A. Roebroeck, E. Formisano, and R. Goebel. Reply to friston and david after commentson: The identification of interacting networks in the brain using fmri: Modelselection, causality and deconvolution. Neuroimage, 2009b.S. Ryali, K. Supekar, T. Chen, and V. Menon. Multivariate dynamical systems modelsfor estimating causal interactions in fmri. Neuroimage, 2010.Z. S. Saad, K. M. Ropella, R. W. Cox, and E. A. DeYoe. Analysis and use of fmriresponse delays. Hum Brain Mapp, 13(2):74–93, 2001.R. Salmelin and J. Kujala. Neural representation of language: activation versus longrangeconnectivity. Trends Cogn Sci, 10(11):519–25, 2006.J. R. Sato, D. Y. Takahashi, S. M. Arcuri, K. Sameshima, P. A. Morettin, and L. A.Baccala. Frequency domain connectivity identification: an application of partialdirected coherence in fmri. Hum Brain Mapp, 30(2):452–61, 2009.J. R. Sato, A. Fujita, E. F. Cardoso, C. E. Thomaz, M. J. Brammer, and Jr. E. Amaro.Analyzing the connectivity between regions of interest: an approach based on clustergranger causality for fmri data analysis. Neuroimage, 52(4):1444–55, 2010.M. B. Schippers, A. Roebroeck, R. Renken, L. Nanetti, and C. Keysers. Mapping theinformation flow from one brain to another during gestural communication. ProcNatl Acad Sci U S A, 107(20):9388–93, 2010.T Schweder. Composable markov processes. journal of Applied Probability, 7(2):400–410, 1970.J. F. Smith, A. Pillai, K. Chen, and B. Horwitz. Identification and validation of effectiveconnectivity networks in functional magnetic resonance imaging using switchinglinear dynamic systems. Neuroimage, 52(3):1027–40, 2009.S. M. Smith, K. L. Miller, G. Salimi-Khorshidi, M. Webster, C. F. Beckmann, T. E.Nichols, J. D. Ramsey, and M. W. Woolrich. Network modelling methods for fmri.Neuroimage, 2010.V Solo. On causality i: Sampling and noise. Proceedings of the 46th IEEE Conferenceon Decision and Control, pages 3634–3639, 2006.D. Sridharan, D. J. Levitin, and V. Menon. A critical role for the right fronto-insularcortex in switching between central-executive and default-mode networks. Proc NatlAcad Sci U S A, 105(34):12569–74, 2008.102

Causal analysis of fMRIK. E. Stephan, L. Kasper, L. M. Harrison, J. Daunizeau, H. E. den Ouden, M. Breakspear,and K. J. Friston. Nonlinear dynamic causal models for fmri. Neuroimage, 42(2):649–62, 2008.I. H. Stevenson and K. P. Kording. On the similarity of functional connectivity betweenneurons estimated across timescales. PLoS One, 5(2):e9206, 2010.K. Thomsen, N. Offenhauser, and M. Lauritzen. Principal neuron spiking: neithernecessary nor sufficient for cerebral blood flow in rat cerebellum. J Physiol, 560(Pt1):181–9, 2004.L. Q. Uddin, A. M. Kelly, B. B. Biswal, F. Xavier Castellanos, and M. P. Milham.Functional connectivity of default mode network components: correlation, anticorrelation,and causality. Hum Brain Mapp, 30(2):625–37, 2009.K. Ugurbil, L. Toth, and D. S. Kim. How accurate is magnetic resonance imaging ofbrain function? Trends Neurosci, 26(2):108–14, 2003.K. Uludag. To dip or not to dip: reconciling optical imaging and fmri data. Proc NatlAcad Sci U S A, 107(6):E23; author reply E24, 2010.K. Uludag, D. J. Dubowitz, and R. B. Buxton. Basic principles of functional mri. InR. Edelman, J. Hesselink, and M. Zlatkin, editors, Clinical MRI. Elsevier, San Diego,2005.K. Uludag, B. Muller-Bierl, and K. Ugurbil. An integrative model for neuronal activityinducedsignal changes for gradient and spin echo functional imaging. Neuroimage,48(1):150–65, 2009.P Valdes-Sosa, J C Jimenez, J Riera, R Biscay, and T Ozaki. Nonlinear eeg analysisbased on a neural mass model. Biological cybernetics, 81:415–24, 1999.P. Valdes-Sosa, A. Roebroeck, J. Daunizeau, and K. Friston. Effective connectivity:Influence, causality and biophysical modeling. Neuroimage, in press.P. A. Valdes-Sosa. Spatio-temporal autoregressive models defined over brain manifolds.Neuroinformatics, 2(2):239–50, 2004.P. A. Valdes-Sosa, R. Kotter, and K. J. Friston. Introduction: multimodal neuroimagingof brain connectivity. Philos Trans R Soc Lond B Biol Sci, 360(1457):865–7, 2005a.P. A. Valdes-Sosa, J. M. Sanchez-Bornot, A. Lage-Castellanos, M. Vega-Hernandez,J. Bosch-Bayard, L. Melie-Garcia, and E. Canales-Rodriguez. Estimating brain functionalconnectivity with sparse multivariate autoregression. Philos Trans R Soc LondB Biol Sci, 360(1457):969–81, 2005b.P. A. Valdes-Sosa, J. M. Sanchez-Bornot, R. C. Sotero, Y. Iturria-Medina, Y. Aleman-Gomez, J. Bosch-Bayard, F. Carbonell, and T. Ozaki. Model driven eeg/fmri fusionof brain oscillations. Hum Brain Mapp, 30(9):2701–21, 2009.103

Roebroeck Seth Valdes-SosaV. Walsh and A. Cowey. Transcranial magnetic stimulation and cognitive neuroscience.Nat Rev Neurosci, 1(1):73–9, 2000.W. W. S. Wei. Time Series Analysis: Univariate and Multivariate Methods. Addison-Wesley, Redwood City, 1990.Halbert White and Xun Lu. Granger causality and dynamic structural systems. Journalof Financial Econometrics, 8(2):193–243, 2010.N. Wiener. The theory of prediction. In E.F. Berkenbach, editor, Modern Mathematicsfor Engineers. McGraw-Hill, New York, 1956.F. Wiesinger, P. F. Van de Moortele, G. Adriany, N. De Zanche, K. Ugurbil, and K. P.Pruessmann. Potential and feasibility of parallel mri at high field. NMR Biomed, 19(3):368–78, 2006.E. Yacoub, N. Harel, and K. Ugurbil. High-field fmri unveils orientation columns inhumans. Proc Natl Acad Sci U S A, 105(30):10607–12, 2008.104

JMLR: Workshop and Conference Proceedings 12:95–114, 2011Causality in Time SeriesCausal Search in Structural Vector Autoregressive ModelsAlessio MonetaMax Planck Institute of EconomicsJena, GermanyNadine ChlaßFriedrich Schiller University of Jena, GermanyDoris EntnerHelsinki Institute for Information Technology, FinlandPatrik HoyerHelsinki Institute for Information Technology, Finlandmoneta@econ.mpg.denadine.chlass@uni-jena.dedoris.entner@cs.helsinki.fipatrk.hoyer@helsinki.fiEditors: Florin Popescu and Isabelle GuyonAbstractThis paper reviews a class of methods to perform causal inference in the framework ofa structural vector autoregressive model. We consider three different settings. In thefirst setting the underlying system is linear with normal disturbances and the structuralmodel is identified by exploiting the information incorporated in the partial correlationsof the estimated residuals. Zero partial correlations are used as input of a searchalgorithm formalized via graphical causal models. In the second, semi-parametric,setting the underlying system is linear with non-Gaussian disturbances. In this casethe structural vector autoregressive model is identified through a search procedurebased on independent component analysis. Finally, we explore the possibility ofcausal search in a nonparametric setting by studying the performance of conditionalindependence tests based on kernel density estimations.Keywords: Causal inference, econometric time series, SVAR, graphical causal models,independent component analysis, conditional independence tests1. Introduction1.1. Causal inference in econometricsApplied economic research is pervaded by questions about causes and effects. Forexample, what is the effect of a monetary policy intervention? Is energy consumptioncausing growth or the other way around? Or does causality run in both directions? Areeconomic fluctuations mainly caused by monetary, productivity, or demand shocks?Does foreign aid improve living standards in poor countries? Does firms’ expenditurein R&D causally influence their profits? Are recent rises in oil prices in part caused byspeculation? These are seemingly heterogeneous questions, but they all require someknowledge of the causal process by which variables came to take the values we observe.c○ 2011 A. Moneta, N. Chlaß, D. Entner & P. Hoyer.

Moneta Chlass Entner HoyerA traditional approach to address such questions hinges on the explicit use of apriori economic theory. The gist of this approach is to partition a causal process ina deterministic, and a random part and to articulate the deterministic part such as toreflect the causal dependencies dictated by economic theory. If the formulation of thedeterministic part is accurate and reliable enough, the random part is expected to displayproperties that can easily be analyzed by standard statistical tools. The touchstoneof this approach is represented by the work of Haavelmo (1944), which inspired theresearch program subsequently pursued by the Cowles Commission (Koopmans, 1950;Hood and Koopmans, 1953). Therein, the causal process is formalized by means ofa structural equation model, that is, a system of equations with endogenous variables,exogenous variables, and error terms, first developed by Wright (1921). Its coefficientswere given a causal interpretation (Pearl, 2000).This approach has been strongly criticized in the 1970s for being ineffective inboth policy evaluation and forecasting. Lucas (1976) pointed out that the economictheory included in the SEM fails to take economic agents’ (rational) motivations andexpectations into consideration. Agents, according to Lucas, are able to anticipate policyintervention and act contrary to the prediction derived from the structural equationmodel, since the model usually ignores such anticipations. Sims (1980) puts forth anothercritique which runs parallel to Lucas’ one. It explicitly addresses the status ofexogeneity which the Cowles Commission approach attributes (arbitrarily, accordingto Sims) to some variables such that the structural model can be identified. Sims arguesthat theory is not a reliable source for deeming a variable as exogenous. More generally,the Cowles Commission approach with its strong a priori commitment to theory,risks falling into a vicious circle: if causal information (even if only about direction)can exclusively be derived from background theory, how do we obtain an empiricallyjustified theory? (Cfr. Hoover, 2006, p.75).An alternative approach has been pursued since Wiener (1956) and Granger’s (1969)work. It aims at inferring causal relations directly from the statistical properties of thedata relying only to a minimal extent on background knowledge. Granger (1980) proposesa probabilistic concept of causality, similar to Suppes (1970). Granger definescausality in terms of the incremental predictability (at horizon one) of a time seriesvariable {Y t } (given the present and past values of {Y t } and of a set {Z t } of possible relevantvariables) when another time series variable {X t } (in its present and past values) isnot omitted. More formally:{X t } Granger-causes {Y t } if P(Y t+1 |X t , X t−1 ,...,Y t ,Y t−1 ,...,Z t ,Z t−1 ,...) P(Y t+1 |Y t ,Y t−1 ,...,Z t ,Z t−1 ,...)(1)As pointed out by Florens and Mouchart (1982), testing the hypothesis of Grangernoncausality corresponds to testing conditional independence. Given lags p, {X t } doesnot Granger cause {Y t }, ifY t+1 ⊥ (X t , X t−1 ,..., X t−p ) | (Y t ,Y t−1 ,...,Y t−p ,Z t ,Z t−1 ,...,Z t−p ) (2)106

Causal Search in SVARTo test Granger noncausality, researchers often specify linear vector autoregressive(VAR) models:Y t = A 1 Y t−1 + ... + A p Y t−p + u t , (3)in which Y t is a k × 1 vector of time series variables (Y 1,t ,...,Y k,t ) ′ , where () ′ is thetranspose, the A j ( j = 1,..., p) are k × k coefficient matrices, and u t is the k × 1 vectorof random disturbances. In this framework, testing the hypothesis that {Y i,t } does notGranger-cause {Y j,t }, reduces to test whether the ( j,i) entries of the matrices A 1 ,...,A pare vanishing simultaneously. Granger noncausality tests have been extended to nonlinearsettings by Baek and Brock (1992), Hiemstra and Jones (1994), and Su and White(2008), using nonparametric tests of conditional independence (more on this topic insection 4).The concept of Granger causality has been criticized for failing to capture ‘structuralcausality’ (Hoover, 2008). Suppose one finds that a variable A Granger-causesanother variable B. This does not necessarily imply that an economic mechanism existsby which A can be manipulated to affect B. The existence of such a mechanism inturn does not necessarily imply Granger causality either (for a discussion see Hoover2001, pp. 150-155). Indeed, the analysis of Granger causality is based on coefficientsof reduced-form models, like those incorporated in equation (3), which are unlikely toreliably represent actual economic mechanisms. For instance, in equation (3) the simultaneouscausal structure is not modeled in order to facilitate estimation. (However, notethat Eichler (2007) and White and Lu (2010) have recently developed and formalizedricher structural frameworks in which Granger causality can be fruitfully analyzed.)1.2. The SVAR frameworkStructural vector autoregressive (SVAR) models constitute a middle way between theCowles Commission approach and the Granger-causality approach. SVAR models aimat recovering the concept of structural causality, but eschew at the same time the strong‘apriorism’ of the Cowles Commission approach. The idea is, like in the Cowles Commissionapproach, to articulate an unobserved structural model, formalized as a dynamicgenerative model: at each time unit the system is affected by unobserved innovationterms, by which, once filtered by the model, the variables come to take the valueswe observe. But, differently from the Cowles Commission approach, and similarly tothe Granger-VAR model, the data generating process is generally enough articulatedso that time series variables are not distinguished a priori between exogenous and endogenous.A linear SVAR model is in principle a VAR model ‘augmented’ by thecontemporaneous structure:Γ 0 Y t = Γ 1 Y t−1 + ... + Γ p Y t−p + ε t . (4)This is easily obtained by pre-multiplying each side of the VAR modelY t = A 1 Y t−1 + ... + A p Y t−p + u t , (5)by a matrix Γ 0 so that Γ i = Γ 0 A i , for i = 1,...,k and ε t = Γ 0 u t . Note, however, that notany matrix Γ 0 will be suitable. The appropriate Γ 0 will be that matrix corresponding107

Moneta Chlass Entner Hoyerto the ‘right’ rotation of the VAR model, that is the rotation compatible both with thecontemporaneous causal structure of the variable and the structure of the innovationterm. Let us consider a matrix B 0 = I − Γ 0 . If the system is normalized such that thematrix Γ 0 has all the elements of the principal diagonal equal to one (which can be donestraightforwardly), the diagonal elements of B 0 will be equal to zero. We can write:Y t = B 0 Y t + Γ 1 Y t−1 + ... + Γ p Y t−p + ε t (6)from which we see that B 0 (and thus Γ 0 ) determines in which form the values of a variableY i,t will be dependent on the contemporaneous value of another variable Y j,t . The‘right’ rotation will also be the one which makes ε t a vector of authentic innovationterms, which are expected to be independent (not only over time, but also contemporaneously)sources or shocks.In the literature, different methods have been proposed to identify the SVAR model(4) on the basis of the estimation of the VAR model (5). Notice that there are moreunobserved parameters in (4), whose number amounts to k 2 (p + 1), than parametersthat can be estimated from (5), which are k 2 p + k(k + 1)/2, so one has to impose atleast k(k − 1)/2 restrictions on the system. One solution to this problem is to get arotation of (5) such that the covariance matrix of the SVAR residuals Σ ε is diagonal,using the Cholesky factorization of the estimated residuals Σ u . That is, let P be thelower-triangular Cholesky factorization of Σ u (i.e. Σ u = PP ′ ), let D be a k × k diagonalmatrix with the same diagonal as P, and let Γ 0 = DP −1 . By pre-multiplying (5) byΓ 0 , it turns out that Σ ε = E[Γ 0 u t u ′ t Γ′ 0 ] = DD′ , which is diagonal. A problem withthis method is that P changes if the ordering of the variables (Y 1t ,...,Y kt ) ′ in Y t and,consequently, the order of residuals in Σ u , changes. Since researchers who estimate aSVAR are often exclusively interested on tracking down the effect of a structural shockε it on the variables Y 1,t ,...,Y k,t over time (impulse response functions), Sims (1981)suggested investigating to what extent the impulse response functions remain robustunder changes of the order of variables.Popular alternatives to the Cholesky identification scheme are based either on theuse of a priori, theory-based, restrictions or on the use of long-run restrictions. Theformer solution consists in imposing economically plausible constraints on the contemporaneousinteractions among variables (Blanchard and Watson, 1986; Bernanke,1986) and has the drawback of ultimately depending on the a priori reliability of economictheory, similarly to the Cowles Commission approach. The second solution isbased on the assumptions that certain economic shocks have long-run effect to othervariables, but do not influence in the long-run the level of other variables (see Shapiroand Watson, 1988; Blanchard and Quah, 1989; King et al., 1991). This approach hasbeen criticized as not being very reliable unless strong a priori restrictions are imposed(see Faust and Leeper, 1997).In the rest of the paper, we first present a method, based on the graphical causalmodel framework, to identify the SVAR (section 2). This method is based on conditionalindependence tests among the estimated residuals of the VAR estimated model.Such tests rely on the assumption that the shocks affecting the model are Gaussian.108

Causal Search in SVARWe then relax the Gaussianity assumption and present a method to identify the SVARmodel based on independent component analysis (section 3). Here the main assumptionis that shocks are non-Gaussian and independent. Finally (section 4), we explorethe possibility of extending the framework for causal inference to a nonparametric setting.In section 5 we wrap up the discussion and conclude by formulating some openquestions.2. SVAR identification via graphical causal models2.1. BackgroundA data-driven approach to identify the structural VAR is based on the analysis of theestimated residuals û t . Notice that when a basic VAR model is estimated (equation 3),the information about contemporaneous causal dependence is incorporated exclusivelyin the residuals (being not modeled among the variables). Graphical causal models,as originally developed by Pearl (2000) and Spirtes et al. (2000), represent an efficientmethod to recover, at least in part, the contemporaneous causal structure moving fromthe analysis of the conditional independencies among the estimated residuals. Once thecontemporaneous causal structure is recovered, the estimation of the lagged autoregressivecoefficients permits us to identify the complete SVAR model.This approach was initiated by Swanson and Granger (1997), who proposed to testwhether a particular causal order of the VAR is in accord with the data by testing all thepartial correlations of order one among error terms and checking whether some partialcorrelations are vanishing. Reale and Wilson (2001), Bessler and Lee (2002), Demiralpand Hoover (2003), and Moneta (2008) extended the approach by using the partialcorrelations of the VAR residuals as input to graphical causal model search algorithms.In graphical causal models, the structural model is represented as a causal graph (aDirected Acyclic Graph if the presence of causal loops is excluded), in which each noderepresents a random variable and each edge a causal dependence. Furthermore, a setof assumptions or ‘rules of inference’ are formulated, which regulate the relationshipbetween causal and probabilistic dependencies: the causal Markov and the faithfulnessconditions (Spirtes et al., 2000). The former restricts the joint probability distributionof modeled variables: each variable is independent of its graphical non-descendantsconditional on its graphical parents. The latter makes causal discovery possible: all ofthe conditional independence relations among the modeled variables follow from thecausal Markov condition. Thus, for example, if the causal structure is represented asY 1t → Y 2t → Y t,3 , it follows from the Markov condition that Y 1,t ⊥ Y 3,t |Y 2,t . If, on theother hand, the only (conditional) independence relation among Y 1,t ,Y 2,t ,Y 3,t is Y 1,t ⊥Y 3,t , it follows from the faithfulness condition that Y 1,t → Y 3,t

Moneta Chlass Entner Hoyersecond step, conditional independence relations (or d-separations, which are the graphicalcharacterization of conditional independence) are merely used to erase edges and,in further steps, to direct edges. The output of such algorithms are not necessarily onesingle graph, but a class of Markov equivalent graphs.There is nothing neither in the Markov or faithfulness condition, nor in the constraintbasedalgorithms that limits them to linear and Gaussian settings. Graphical causalmodels do not require per se any a priori specification of the functional dependence betweenvariables. However, in applications of graphical models to SVAR, conditionalindependence is ascertained by testing vanishing partial correlations (Swanson andGranger, 1997; Bessler and Lee, 2002; Demiralp and Hoover, 2003; Moneta, 2008).Since normal distribution guarantees the equivalence between zero partial correlationand conditional independence, these applications deal de facto with linear and Gaussianprocesses.2.2. Testing residuals zero partial correlationsThere are alternative methods to test zero partial correlations among the error termsû t = (u 1t ,...,u kt ) ′ . Swanson and Granger (1997) use the partial correlation coefficient.That is, in order to test, for instance, ρ(u it ,u kt |u jt ) = 0, they use the standard t statisticsfrom a least square regression of the model:u it = α j u jt + α k u kt + ε it , (7)on the basis that α k = 0 ⇔ ρ(u it ,u kt |u jt ) = 0. Since Swanson and Granger (1997) imposethe partial correlation constraints looking only at the set of partial correlations of orderone (that is conditioned on only one variable), in order to run their tests they considerregression equations with only two regressors, as in equation (7).Bessler and Lee (2002) and Demiralp and Hoover (2003) use Fisher’s z that isincorporated in the software TETRAD (Scheines et al., 1998):z(ρ XY.K ,T) = 1 (︃ )︃√︀ |1 + ρXY.K |T − |K| − 3 log2|1 − ρ XY.K | , (8)where |K| equals the number of variables in K and T the sample size. If the variables(for instance X = u it , Y = u kt , K = (u jt ,u ht )) are normally distributed, we have thatz(ρ XY.K ,T) − z(ˆρ XY.K ,T) ∼ N(0,1) (9)(see Spirtes et al., 2000, p.94).A different approach, which takes into account the fact that correlations are obtainedfrom residuals of a regression, is proposed by Moneta (2008). In this case it is usefulto write the VAR model of equation (3) in a more compact form:Y t = Π ′ X t + u t , (10)110

Causal Search in SVARwhere X ′ t = [Y′ t−1 , ...,Y′ t−p ], which has dimension (1 × kp) and Π′ = [A 1 ,...,A p ], whichhas dimension (k × kp). In case of stable VAR process (see next subsection), the conditionalmaximum likelihood estimate of Π for a sample of size T is given byMoreover, the ith row of ˆΠ ′ is⎡ ⎤⎡⎤T∑︁ T∑︁ˆΠ ′ = ⎢⎣ Y t X ′ t⎥⎦⎢⎣ X t X ′ t⎥⎦t=1t=1t=1⎤⎡⎤T∑︁ T∑︁ˆπ ′ i⎡⎢⎣ = Y it X ′ t⎥⎦⎢⎣ X t X ′ t⎥⎦which coincides with the estimated coefficient vector from an OLS regression of Y it onX t (Hamilton 1994: 293). The maximum likelihood estimate of the matrix of varianceand covariance among the error terms Σ u turns out to be ˆΣ u = (1/T) ∑︀ Tt=1 û t û ′ t , whereû t = Y t − ˆΠ ′ X t . Therefore, the maximum likelihood estimate of the covariance betweenu it and u jt is given by the (i, j) element of ˆΣ u : ˆσ i j = (1/T) ∑︀ Tt=1 û it û jt . Denoting by σ i jthe (i, j) element of Σ u , let us first define the following matrix transform operators: vec,which stacks the columns of a k × k matrix into a vector of length k 2 and vech, whichvertically stacks the elements of a k × k matrix on or below the principal diagonal intoa vector of length k(k + 1)/2. For example:[︃ ]︃σ11 σvec 12σ 21 σ 22⎡=⎢⎣t=1−1−1σ 11[︃ ]︃σ 21 σ11 σ, vech 12σ 12 σ⎤⎥⎦21 σ 22σ 22.,⎡= ⎢⎣σ 11σ 21σ 22⎤⎥⎦ .The process being stationary and the error terms Gaussian, it turns out that:√ dT [vech( ˆΣ u ) − vech(Σ u )] −→ N(0, Ω), (11)where Ω = 2D + k (Σ u ⊗ Σ u )(D + k )′ , D + k ≡ (D′ k D k) −1 D ′ k , D k is the unique (k 2 × k(k +1)/2) matrix satisfying D k vech(Ω) = vec(Ω), and ⊗ denotes the Kronecker product (seeHamilton 1994: 301). For example, for k = 2, we have,√T⎡⎢⎣ˆσ 11 − σ 11ˆσ 12 − σ 12⎤⎥⎦ˆσ 22 − σ 22d⎛⎡−→ N ⎜⎝ ⎢⎣000⎤⎥⎦ , ⎡⎢⎣2σ 2 112σ 11 σ 12 2σ 2 122σ 11 σ 12 σ 11 σ 22 + σ 2 122σ 12 σ 222σ 2 122σ 12 σ 22 2σ 2 22Therefore, to test the null hypothesis that ρ(u it ,u jt ) = 0 from the VAR estimated residuals,it is possible to use the Wald statistic:⎤⎞⎥⎦ ⎟⎠T ( ˆσ i j ) 2ˆσ ii ˆσ j j + ˆσ 2 i j≈ χ 2 (1).111

Moneta Chlass Entner HoyerThe Wald statistic for testing vanishing partial correlations of any order is obtainedby applying the delta method, which suggests that if X T is a (r × 1) sequenceof vector-valued random-variables and if [ √ T(X 1T − θ 1 ),..., √ T(X rT − θ r )] −→ N(0,Σ)and h 1 ,...,h r are r real-valued functions of θ = (θ 1 ,...,θ r ), h i : R r → R, defined andcontinuously differentiable in a neighborhood ω of the parameter point θ and such thatthe matrix B = ||∂h i /∂θ j || of partial derivatives is nonsingular in ω, then:[ √ T[h 1 (X T ) − h 1 (θ)],..., √ T[h r (X T ) − h r (θ)]] −→ N(0, BΣB ′ )(see Lehmann and Casella 1998: 61).Thus, for k = 4, suppose one wants to test corr(u 1t ,u 3t |u 2t ) = 0. First, notice thatρ(u 1 ,u 3 |u 2 ) = 0 if and only if σ 22 σ 13 − σ 12 σ 23 = 0 (by definition of partial correlation).One can define a function g : R k(k+1)/2 → R, such that g(vech(Σ u )) = σ 22 σ 13 − σ 12 σ 23 .Thus,∇g ′ = (0, −σ 23 , σ 22 , 0, σ 13 , −σ 12 , 0, 0, 0, 0).Applying the delta method:√ dT[( ˆσ22 ˆσ 13 − ˆσ 12 ˆσ 23 ) − (σ 22 σ 13 − σ 12 σ 23 )] −→ N(0,∇g ′ Ω∇g).The Wald test of the null hypothesis corr(u 1t ,u 3t |u 2t ) = 0 is given by:ddT( ˆσ 22 ˆσ 13 − ˆσ 12 ˆσ 23 ) 2∇g ′ Ω∇g≈ χ 2 (1).Tests for higher order correlations and for k > 4 follow analogously (see also Moneta,2003). This testing procedure has the advantage, with respect to the alternative methods,to be straightforwardly applied to the case of cointegrated data, as will be explainedin the next subsection.2.3. Cointegration caseA typical feature of economic time series data in which there is some form of causaldependence is cointegration. This term denotes the phenomenon that nonstationaryprocesses can have linear combinations that are stationary. That is, suppose that eachcomponent Y it of Y t = (Y 1t ,...,Y kt ) ′ , which follows the VAR processY t = A 1 Y t−1 + ... + A p Y t−p + u t ,is nonstationary and integrated of order one (∼ I(1)). This means that the VAR processY t is not stable, i.e. det(I k − A 1 z − A p z p ) is equal to zero for some |z| ≤ 1 (Lütkepohl,2006), and that each component ∆Y it of ∆Y t = (Y t − Y t−1 ) is stationary (I(0)), that isit has time-invariant means, variances and covariance structure. A linear combinationbetween between the elements of Y t is called a cointegrating relationship if there is alinear combination c 1 Y 1t + ... + c k Y kt which is stationary (I(0)).112

Causal Search in SVARIf it is the case that the VAR process is unstable with the presence of cointegratingrelationships, it is more appropriate (Lütkepohl, 2006; Johansen, 2006) to estimate thefollowing re-parametrization of the VAR model, called Vector Error Correction Model(VECM):∆Y t = F 1 ∆Y t−1 + ... + F p−1 ∆Y t−p+1 − GY t−p + u t , (12)where F i = −(I k −A 1 −...−A i ), for i = 1,..., p−1 and G = I k −A 1 −...−A p . The (k×k)matrix G has rank r and thus G can be written as HC with H and C ′ of dimension (k×r)and of rank r. C ≡ [c 1 ,...,c r ] ′ is called the cointegrating matrix.Let ˜C, ˜H, and ˜F i be the maximum likelihood estimator of C, H, F according toJohansen’s (1988, 1991) approach. Then the asymptotic distribution of ˜Σ u , that is themaximum likelihood estimator of the covariance matrix of u t , is:√ dT [vech( ˜Σ u ) − vech(Σ u )] −→ N(0, 2D + k (Σ u ⊗ Σ u )D +′ ), (13)which is equivalent to equation (11) (see it again for the definition of the various operators).Thus, it turns out that the asymptotic distribution of the maximum likelihoodestimator ˜Σ u is the same as the OLS estimation ˆΣ u for the case of stable VAR.Thus, the application of the method described for testing residuals zero partial correlationscan be applied straightforwardly to cointegrated data. The model is estimatedas a VECM error correction model using Johansen’s (1988, 1991) approach, correlationsare tested exploiting the asymptotic distribution of ˜Σ u and finally can be parameterizedback in its VAR form of equation (3).2.4. Summary of the search procedureThe graphical causal models approach to SVAR identification, which we suggest incase of Gaussian and linear processes, can be summarized as follows.Step 1 Estimate the VAR model Y t = A 1 Y t−1 +...+A p Y t−p +u t with the usual specificationtests about normality, zero autocorrelation of residuals, lags, and unit roots (seeLütkepohl, 2006). If the hypothesis of nonstationarity is rejected, estimate the VARmodel via OLS (equivalent to MLE under the assumption of normality of the errors).If unit root tests do not reject I(1) nonstationarity in the data, specify the model asVECM testing the presence of cointegrating relationships. If tests suggest the presenceof cointegrating relationships, estimate the model as VECM. If cointegration is rejectedestimate the VAR models taking first difference.Step 2 Run tests for zero partial correlations between the elements u 1t ,...,u kt of u tusing the Wald statistics on the basis of the asymptotic distribution of the covariancematrix of u t . Note that not all possible partial correlations ρ(u it ,u jt |u ht ,...) need to betested, but only those necessary for step 3.k113

Moneta Chlass Entner HoyerStep 3 Apply a causal search algorithm to recover the causal structure among u 1t ,...,u kt ,which is equivalent to the causal structure among Y 1t ,...,Y kt (cfr. section 1.2 and seeMoneta 2003). In case of acyclic (no feedback loops) and causally sufficient (no latentvariables) structure, the suggested algorithm is the PC algorithm of Spirtes et al. (2000,pp. 84-85). Moneta (2008) suggested few modifications to the PC algorithm in orderto make the orientation of edges compatible with as many conditional independencetests as possible. This increases the computational time of the search algorithm, butconsidering the fact that VAR models deal with a few number of time series variables(rarely more than six to eight; see Bernanke et al. 2005), this slowing down does notcreate a serious concern in this context. Table 1 reports the modified PC algorithm. Incase of acyclic structure without causal sufficiency (i.e. possibly including latent variables),the suggested algorithm is FCI (Spirtes et al. 2000, pp. 144-145). In the caseof no latent variables and in the presence of feedback loops, the suggested algorithmis CCD (Richardson and Spirtes, 1999). There is no algorithm in the literature whichis consistent for search when both latent variables and feedback loops may be present.If the goal of the study is only impulse response analysis (i.e. tracing out the effectsof structural shocks ε 1t ,...,ε kt on Y t ,Y t−1 ,...) and neither contemporaneous feedbacksnor latent variables can be excluded a priori, a possible solution is to apply only steps(A) and (B) of the PC algorithm. If the resulting set of possible causal structures (representedby an undirected graph) contains a manageable number of elements, one canstudy the characteristics of the impulse response functions which are robust across allthe possible causal structures, where the presence of both feedbacks and latent variablesis allowed (Moneta, 2004).Step 4 Calculate structural coefficients and impulse response functions. If the outputof Step 3 is a set of causal structures, run sensitivity analysis to investigate therobustness of the conclusions under the different possible causal structures. Bootstrapprocedures may also be applied to determine which is the most reliable causal order(see simulations and applications in Demiralp et al., 2008).3. Identification via independent component analysisThe methods considered in the previous section use tests for zero partial correlation onthe VAR-residuals to obtain (partial) information about the contemporaneous structurein an SVAR model with Gaussian shocks. In this section we show how non-Gaussianand independent shocks can be exploited for model identification by using the statisticalmethod of ‘Independent Component Analysis’ (ICA, see Comon (1994); Hyvärinenet al. (2001)). The method is again based on the VAR-residuals u t which can be obtainedas in the Gaussian case by estimating the VAR model using for example ordinaryleast squares or least absolute deviations, and can be tested for non-Gaussianity usingany normality test (such as the Shapiro-Wilk or Jarque-Bera test).To motivate, we note that, from equations (3) and (4) (with matrix Γ 0 ) or theCholesky factorization in section 1.2 (with matrix PD −1 ), the VAR-disturbances u t and114

Causal Search in SVARTable 1: Search algorithm (adapted from the PC Algorithm of Spirtes et al. (2000:84-85); in bold character the modifications).Under the assumption of Gaussianity conditional independence is tested by zero partialcorrelation tests.(A): (connect everything):Form the complete undirected graph G on the vertex set u 1t ,...,u kt so that each vertex isconnected to any other vertex by an undirected edge.(B)(cut some edges):n = 0repeat :repeat :select an ordered pair of variables u ht and u it that are adjacent in G such thatthe number of variables adjacent to u ht is equal or greater than n + 1. Select aset S of n variables adjacent to u ht such that u ti S . If u ht ⊥ u it |S delete edgeu ht — u it from G;until all ordered pairs of adjacent variables u ht and u it such that the number ofvariables adjacent to u ht is equal or greater than n + 1 and all sets S of n variablesadjacent to u ht such that u it S have been checked to see if u ht ⊥ u it |S ;n = n + 1;until for each ordered pair of adjacent variables u ht , u it , the number of adjacent variablesto u ht is less than n + 1;(C)(build colliders):for each triple of vertices u ht ,u it ,u jt such that the pair u ht ,u it and the pair u it ,u jt are eachadjacent in G but the pair u ht ,u jt is not adjacent in G, orient u ht — u it — u jt as u ht −→ u it

Moneta Chlass Entner Hoyerthe structural shocks ε t are connected byu t = Γ −10 ε t = PD −1 ε t (14)with square matrices Γ 0 and PD −1 , respectively. Equation (14) has two important properties:First, the vectors u t and ε t are of the same length, meaning that there are asmany residuals as structural shocks. Second, the residuals u t are linear mixtures of theshocks ε t , connected by the ‘mixing matrix’ Γ −10. This resembles the ICA model, whenplacing certain assumptions on the shocks ε t .In short, the ICA model is given by x = As, where x are the mixed components, s theindependent, non-Gaussian sources, and A a square invertible mixing matrix (meaningthat there are as many mixtures as independent components). Given samples from themixtures x, ICA estimates the mixing matrix A and the independent components s, bylinearly transforming x in such a way that the dependencies among the independentcomponents s are minimized. The solution is unique up to ordering, sign and scaling(Comon, 1994; Hyvärinen et al., 2001).By comparing the ICA model x = As and equation (14), we see a one-to-one correspondenceof the mixtures x to the residuals u t and the independent components s tothe shocks ε t . Thus, to be able to apply ICA, we need to assume that the shocks arenon-Gaussian and mutually independent. We want to emphasize that no specific non-Gaussian distribution is assumed for the shocks, but only that they cannot be Gaussian. 1For the shocks to be mutually independent their joint distribution has to factorize intothe product of the marginal distributions. In the non-Gaussian setting, this implies zeropartial correlation, but the converse is not true (as opposed to the Gaussian case wherethe two statements are equivalent). Thus, for non-Gaussian distributions conditionalindependence is a much stronger requirement than uncorrelatedness.Under the assumption that the shocks ε t are non-Gaussian and independent, equation(14) follows exactly the ICA-model and applying ICA to the VAR residuals u tyields a unique solution (up to ordering, sign, and scaling) for the mixing matrix Γ −10and the independent components ε t (i.e. the structural shocks in our case). However,the ambiguities of ICA make it hard to directly interpret the shocks found by ICA sincewithout further analysis we cannot relate the shocks directly to the measured variables.Hence, we assume that the residuals u t follow a linear non-Gaussian acyclic model(Shimizu et al., 2006), which means that the contemporaneous structure is representedby a DAG (directed acyclic graph). In particular, the model is given byu t = B 0 u t + ε t (15)with a matrix B 0 , whose diagonal elements are all zero and, if permuted according tothe causal order, is strictly lower triangular. By rewriting equation (15) we see thatΓ 0 = I − B 0 . (16)From this equation it follows that the matrix B 0 describes the contemporaneous structureof the variables Y t in the SVAR model as shown in equation (6). Thus, if we can1. Actually, the requirement is that at most one of the residuals can be Gaussian.116

Causal Search in SVARidentify the matrix Γ 0 , we also obtain the matrix B 0 for the contemporaneous effects.As pointed out above, the matrix Γ −10(and hence Γ 0 ) can be estimated using ICA up toordering, scaling, and sign. With the restriction of B 0 representing an acyclic system,we can resolve these ambiguities and are able to fully identify the model. For simplicity,let us assume that the variables are arranged according to a causal ordering, sothat the matrix B 0 is strictly lower triangular. From equation (16) then follows that thematrix Γ 0 is lower triangular with all ones on the diagonal. Using this information, theambiguities of ICA can be resolved in the following way.The lower triangularity of B 0 allows us to find the unique permutation of the rows ofΓ 0 , which yields all non-zero elements on the diagonal of Γ 0 , meaning that we replacethe matrix Γ 0 with Q 1 Γ 0 where Q 1 is the uniquely determined permutation matrix.Finding this permutation resolves the ordering-ambiguity of ICA and links the shocksε t to the components of the residuals u t in a one-to-one manner. The sign- and scalingambiguityis now easy to fix by simply dividing each row of Γ 0 (the row-permutedversion from above) by the corresponding diagonal element yielding all ones on thediagonal, as implied by Equation (16). This ensures that the connection strength of theshock ε t on the residual u t is fixed to one in our model (Equation (15)).For the general case where B 0 is not arranged in the causal order, the above argumentsfor solving the ambiguities still apply. Furthermore, we can find the causalorder of the contemporaneous variables by performing simultaneous row- and columnpermutationson Γ 0 yielding the matrix closest to lower triangular, in particular ˜Γ 0 =Q 2 Γ 0 Q ′ 2 with an appropriate permutation matrix Q 2. In case non of these permutationsleads to a close to lower triangular matrix a warning is issued.Essentially, the assumption of acyclicity allows us to uniquely connect the structuralshocks ε t to the components of u t and fully identify the contemporaneous structure.Details of the procedure can be found in (Shimizu et al., 2006; Hyvärinen et al.,2010). In the sense of the Cholesky factorization of the covariance matrix explained inSection 1 (with PD −1 = Γ −10), full identifiability means that a causal order among thecontemporaneous variables can be determined.In addition to yielding full identification, an additional benefit of using the ICAbasedprocedure when shocks are non-Gaussian is that it does not rely on the faithfulnessassumption, which was necessary in the Gaussian case.We note that there are many ways of exploiting non-Gaussian shocks for modelidentification as alternatives to directly using ICA. One such approach was introducedby Shimizu et al. (2009). Their method relies on iteratively finding an exogenous variableand regressing out their influence on the remaining variables. An exogenous variableis characterized by being independent of the residuals when regressing any othervariable in the model on it. Starting from the model in equation (15), this procedurereturns a causal ordering of the variables u t and then the matrix B 0 can be estimatedusing the Cholesky approach.One relatively strong assumption of the above methods is the acyclicity of the contemporaneousstructure. In (Lacerda et al., 2008) an extension was proposed wherefeedback loops were allowed. In terms of the matrix B 0 this means that it is not re-117

Moneta Chlass Entner Hoyerstricted to being lower triangular (in an appropriate ordering of the variables). While ingeneral this model is not identifiable because we cannot uniquely match the shocks tothe residuals, Lacerda et al. (2008) showed that the model is identifiable when assumingstability of the generating model in (15) (the absolute value of the biggest eigenvalue inB 0 is smaller than one) and disjoint cycles.Another restriction of the above model is that all relevant variables must be includedin the model (causal sufficiency). Hoyer et al. (2008b) extended the abovemodel by allowing for hidden variables. This leads to an overcomplete basis ICAmodel, meaning that there are more independent non-Gaussian sources than observedmixtures. While there exist methods for estimating overcomplete basis ICA models,those methods which achieve the required accuracy do not scale well. Additionally,the solution is again only unique up to ordering, scaling, and sign, and when includinghidden variables the ordering-ambiguity cannot be resolved and in some cases leads toseveral observationally equivalent models, just as in the cyclic case above.We note that it is also possible to combine the approach of section 2 with that describedhere. That is, if some of the shocks are Gaussian or close to Gaussian, it maybe advantageous to use a combination of constraint-based search and non-Gaussianitybasedsearch. Such an approach was proposed in Hoyer et al. (2008a). In particular,the proposed method does not make any assumptions on the distributions of the VARresidualsu t . Basically, the PC algorithm (see Section 2) is run first, followed by utilizationof whatever non-Gaussianity there is to further direct edges. Note that there isno need to know in advance which shocks are non-Gaussian since finding such shocksis part of the algorithm.Finally, we need to point out that while the basic ICA-based approach does notrequire the faithfulness assumption, the extensions discussed at the end of this sectiondo.4. Nonparametric setting4.1. TheoryLinear systems dominate VAR, SVAR, and more generally, multivariate time seriesmodels in econometrics. However, it is not always the case that we know how a variableX may cause another variable Y. It may be the case that we have little or no a prioriknowledge about the way how Y depends on X. In its most general form we wantto know whether X is independent of Y conditional on the set of potential graphicalparents Z, i.e.H 0 : Y ⊥ X | Z, (17)where Y, X,Z is a set of time series variables. Thereby, we do not per se require ana priori specification of how Y possibly depends on X. However, constraint basedalgorithms typically specify conditional independence in a very restrictive way. In continuoussettings, they simply test for nonzero partial correlations, or in other words, forlinear (in)dependencies. Hence, these algorithms will fail whenever the data generationprocess (DGP) includes nonlinear causal relations.118

Causal Search in SVARIn search for a more general specification of conditional independency, Chlaß andMoneta (2010) suggest a procedure based on nonparametric density estimation. Therein,neither the type of dependency between Y and X, nor the probability distributions ofthe variables need to be specified. The procedure exploits the fact that if two randomvariables are independent of a third, one obtains their joint density by the product of thejoint density of the first two, and the marginal density of the third. Hence, hypothesistest (17) translates into:f (Y, X,Z)H 0 : = f (YZ)f (XZ) f (Z) . (18)If we define h 1 (·) := f (Y, X,Z) f (Z), and h 2 (·) := f (YZ) f (XZ), we have:H 0 : h 1 (·) = h 2 (·). (19)We estimate h 1 and h 2 using a kernel smoothing approach (see Wand and Jones, 1995,ch.4). Kernel smoothing has the outstanding property that it is insensitive to autocorrelationphenomena and, therefore, immediately applicable to longitudinal or time seriessettings (Welsh et al., 2002).In particular, we use a so-called product kernel estimator:)︁ (︁1ĥ 1 (x,y,z;b) = KYi)︁ (︁−y KZi −z)︁}︁{︁∑︀ (︁ ni=1KZi)︁}︁−zp{︁∑︀ ni=1K (︁ X i −xN 2 b m+d b )︁ (︁ b bbZi −z)︁}︁{︁∑︀ ni=1 KZ K (︁ )︁ (︁Y i −y Zi)︁}︁−z (20)Kp ,1{︁∑︀ĥ 2 (x,y,z;b) = ni=1K (︁ X i −xN 2 b m+d bwhere X i , Y i , and Z i are the i th realization of the respective time series, K denotes thekernel function, b indicates a scalar bandwidth parameter, and K p represents a productkernel 2 .So far, we have shown how we can estimate h 1 and h 2 . To see whether these aredifferent, we require some similarity measure between both conditional densities. Thereare different ways to measure the distance between a product of densities:(i) The weighted Hellinger distance proposed by Su and White (2008):⎧ √︃⎫d H = 1 2n∑︁ ⎪⎨n ⎪⎩ 1 − h 2 (X i ,Y i ,Z i ) ⎪⎬a(X i ,Y i ,Z i ), (21)h 1 (X i ,Y i ,Z i ) ⎪⎭i=1where a(·) is a nonnegative weighting function. Both the weighting function a(·),and the resulting test statistic are specified in Su and White (2008).(ii) The Euclidean distance proposed by Szekely and Rizzo (2004) in their ‘energytest’:d E = 1 n∑︁ n∑︁||h 1i − h 2nj|| − 1 n∑︁ n∑︁||h 1i − h 12nj|| − 1 n∑︁ n∑︁||h 2i − h 22nj||, (22)i=1j=1i=1j=12. I.e. K p ((Z i − z)/b) = ∏︀ dj=1 K((Z ji − z j )/b). For our simulations (see next section) we choose thekernel: K(u) = (3 − u 2 )φ(u)/2, with φ(u) the standard normal probability density function. We use a“rule-of-thumb” bandwidth: b = n −1/8.5 .bbi=1j=1b119

Moneta Chlass Entner Hoyerwhere h 1i = h 1 (X i ,Y i ,Z i ), h 2i = h 2 (X i ,Y i ,Z i ), and || · || is the Euclidean norm. 3Given these test statistics and their distributions, we compute the type-I error, or p-valueof our test problem (19). If Z = ∅, the tests are available in R-packages energy andcramer. The Hellinger distance is not suitable here, since one can only test for Z ∅.For Z ∅, our test problem (19) requires higher dimensional kernel density estimation.The more dimensions, i.e. the more elements in Z, the scarcer the data, andthe greater the distance between two subsequent data points. This so-called Curse ofdimensionality strongly reduces the accuracy of a nonparametric estimation (Yatchew,1998). To circumvent this problem, we calculate the type-I errors for Z ∅ by a localbootstrap procedure, as described in Su and White (2008, pp. 840-841) and Paparoditisand Politis (2000, pp. 144-145). Local bootstrap draws repeatedly with replacementfrom the sample and counts how many times the bootstrap statistic is larger than thetest statistic of the entire sample. Details on the local bootstrap procedure ca be foundin appendix A.Now, let us see how this procedure fares in those time series settings, where othertesting procedures failed - the case of nonlinear time series.4.2. Simulation DesignOur simulation design should allow us to see how the search procedures of 4.1 performin terms of size and power. To identify size properties (type-I error), H 0 (19) must holdeverywhere. We call data generation processes for which H 0 holds everywhere, size-DGPs. We induce a system of time series {V 1,t ,V 2,t ,V 3,t } n t=1whereby each time seriesfollows an autoregressive process AR(1) with a 1 = 0.5 and error term e t ∼ N(0,1), forinstance, V 1,t = a 1 V 1,t−1 + e V1 ,t. These time series may cause each other as illustrated inFig. 1.V 1,t−1 V 2,t−1 V 3,t−1❅❅❅❄ ❅ ❅❘ ❄ ❄V 1,t V 2,t V 3,tFigure 1: Time series DAG.Therein, V 1,t ⊥ V 2,t |V 1,t−1 , since V 1,t−1 d-separates V 1,t from V 2,t , while V 2,t ⊥ V 3,s ,for any t and s. Hence, the set of variables Z, conditional on which two sets of variablesX and Y are independent of each other, contains zero elements, i.e. V 2,t ⊥ V 3,t−1 ,contains one element, i.e. V 1,t ⊥ V 2,t |V 1,t−1 , or contains two elements, i.e. V 1,t ⊥3. An alternative Euclidean distance is proposed by Baringhaus and Franz (2004) in their Cramer test.This distance turns out to be d E /2. The only substantial difference from the distance proposed in (ii)lies in the method to obtain the critical values (see Baringhaus and Franz 2004).120

Causal Search in SVARV 2,t |V 1,t−1 ,V 3,t−1 .In our simulations, we vary two aspects. The first aspect is the functional form ofthe causal dependency. To systematically vary nonlinearity and its impact, we characterizethe causal relation between, say, V 1,t−1 and V 2,t , in a polynomial form, i.e. viaV 2,t = f (V 1,t−1 ) + e, where f = ∑︀ pj=0 b jV j1,t−1. Herein, j reflects the degree of nonlinearity,while b j would capture the impact nonlinearity exerts. For polynomials of anydegree, we set only b p 0. An additive error term e completes the specification.The second aspect is the number of variables in Z conditional on which X and Ycan be independent. Either zero, one, but maximally two variables may form the set Z= {Z 1 ,...,Z d } of conditioned variables; hence Z has cardinality #Z = {0,1,2}.To identify power properties, H 0 must not hold anywhere, i.e. X ⊥/ Y|Z. We calldata generation processes where H 0 does not hold anywhere, power-DGPs. Such DGPscan be induced by (i) a direct path between X and Y which does not include Z, (ii) acommon cause for X and Y which is not an element of Z, or (iii) a “collider” between Xand Y belonging to Z. 4 As before, we vary the functional form f of these causal pathspolynomially where again, only b p 0. Third, we investigate different cardinalities#Z = {0,1,2} of set Z conditional on which X and Y become dependent.4.3. Simulation ResultsLet us start with #Z = 0, that is, H 0 := X ⊥ Y. Table 2 reports our simulation results forboth size and power DGPs. Rejection frequencies are reported for three different tests,for a theoretical level of significance of 0.05, and 0.1.Table 2: Proportion of rejection of H 0 (no conditioned variables)Energy Cramer Fisher Energy Cramer Fisherlevel of significance 5% level of significance 10%Size DGPsS0.1 (ind. time series) 0.065 0.000 0.151 0.122 0.000 0.213Power DGPsP0.1 (time series linear) 0.959 0.308 0.999 0.981 0.462 1P0.2 (time series quadratic) 0.986 0.255 0.432 0.997 0.452 0.521P0.3 (time series cubic) 1 0.905 1 1 0.975 1P0.3 (time series quartic) 1 0.781 0.647 1 0.901 0.709Note: length series (n) = 100; number of iterations = 1000Take the first line of Table 2. For size DGPs, H 0 holds everywhere. A test performsaccurately if it rejects H 0 in accordance with the respective theoretical significancelevel. We see that the energy test rejects H 0 slightly more often than it should (0.065 >0.05;0.122 > 0.1), whereas the Cramer test does not reject H 0 often enough (0.000

Moneta Chlass Entner Hoyerlatter rejects H 0 much more often than it should. The energy test keeps the type-I errormost accurately. Contrary to both nonparametric tests, the parametric procedure leadsus to suspect a lot more causal relationships than there actually are, if #Z = 0.How well do these tests perform if H 0 does not hold anywhere? That is, howaccurately do they reject H 0 if it is false (power-DGPs)? For linear time series, we seethat the nonparametric energy test has nearly as much power as Fisher’s z. For nonlineartime series, the energy test clearly outperforms Fisher’s z 5 . As it did for size, Cramer’stest generally underperforms in terms of power. Interestingly, its power appears to behigher for higher degrees of nonlinearity. In summary, if one wishes to test for marginalindependence without any information on the type of a potential dependence, one wouldopt for the energy test. It has a size close to the theoretical significance level, and haspower similar to a parametric specification.Let us turn to #Z = 1, where H 0 := X ⊥ Y|Z, for which the results are shown inTable 3. Starting with size DGPs, tests based on Hellinger and Euclidian distanceslightly underreject H 0 whereas for the highest polynomial degree, the Hellinger teststrongly overrejects H 0 . The parametric Fisher’s z slightly overrejects H 0 in case oflinearity, and for higher degrees, starts to underreject H 0 .Table 3: Proportion of rejection of H 0 (one conditioned variable)Hellinger Euclid Fisher Hellinger Euclid Fisherlevel of significance 5% level of significance 10%Size DGPsS1.1 (time series linear) 0.035 0.035 0.062 0.090 0.060 0.103S1.2 (time series quadratic) 0.040 0.020 0.048 0.065 0.035 0.104S1.3 (time series cubic) 0.010 0.010 0.050 0.020 0.015 0.093S1.4 (time series quartic) 0.13 0 0.023 0.2 0.1 0.054Power DGPsP1.1 (time series linear) 0.875 0.910 0.999 0.925 0.950 1P1.2 (time series quadratic) 0.905 0.895 0.416 0.940 0.950 0.504P1.3 (time series cubic) 0.990 1 1 1 1 1P1.4 (time series quartic) 0.84 0.995 0.618 0.91 0.995 0.679Note: n = 100; number of iterations = 200; number of bootstrap iterations (I) = 200Turning to power DGPs, Fisher’s z suffers a dramatic loss in power for those polynomialdegrees which depart most from linearity, i.e. quadratic, and quartic relations.Nonparametric tests which do not require linearity have high power in absolute terms,and nearly twice as much as compared to Fisher’s z. The power properties of the nonparametricprocedures indicate that our local bootstrap succeeds in mitigating the Curseof dimensionality. In sum, nonparametric tests exhibit good power properties for #Z = 1whereas Fisher’s z would fail to discover underlying quadratic or quartic relationshipsin some 60%, and 40% of the cases, respectively.5. For cubic time series, Fisher’s z performs as well as the energy test does. This may be due to the factthat a cubic relation resembles more to a line than other polynomial specifications do.122

Causal Search in SVARThe results for #Z = 2 are presented in Table 4. We find that both nonparametrictests have a size which is notably smaller than the theoretical significance level we induce.Hence, both have a strong tendency to underreject H 0 . Turning to power DGPs,we find that the Euclidean test still has over 90% power to correctly reject H 0 . Forthose polynomial degrees which depart most from linearity, i.e. quadratic and quartic,the Euclidean test has three times as much power as Fisher’s z. However, the Hellingertest performs even worse than Fisher’s z. Here, it may be the Curse of dimensionalitywhich starts to show an impact.Table 4: Proportion of rejection of H 0 (two conditioned variables)Hellinger Euclid Fisher Hellinger Euclid Fisherlevel of significance 5% level of significance 10%Size DGPsS2.1 (time series linear) 0.006 0.020 0.050 0.033 0.046 0.102S2.2 (time series quadratic) 0.000 0.010 0.035 0.000 0.040 0.087S2.3 (time series cubic) 0 0.007 0.056 0 0.007 0.109S2.4 (time series quartic) 0.006 0 0.031 0.013 0 0.067Power DGPsP2.1 (time series linear) 0.28 0.92 1 0.4 0.973 1P2.2 (time series quadratic) 0.170 0.960 0.338 0.250 0.980 0.411P2.3 (time series cubic) 0.667 1 1 0.754 1 1P2.4 (time series quartic) 0.086 0.946 0.597 0.133 0.966 0.665Note: n = 100; number of iterations = 150; number of bootstrap iterations (I) = 100To sum up, we can say that both marginal independencies, and higher dimensionalconditional independencies, i.e. (#Z = 1,2) are best tested for using Euclidean tests. TheHellinger test seems to be more affected by the Curse of dimensionality. We see that ourlocal bootstrap procedure mitigates the latter, but we admit that the number of variablesour nonparametric procedure can deal with is very small. Here, it might be promising toopt for semiparametric (Chu and Glymour, 2008), rather than nonparametric procedureswhich combine parametric and nonparametric approaches.5. ConclusionsThe difficulty of learning causal relations from passive, that is non-experimental, observationsis one of the central challenges of econometrics. Traditional solutions involvethe distinction between structural and reduced form model. The former is meant toformalize the unobserved data generating process, whereas the latter aims to describe asimpler transformation of that process. The structural model is articulated hinging ona priori economic theory. The reduced form model is formalized in such a way that itcan be estimated directly from the data. In this paper, we have presented an approach toidentify the structural model which minimizes the role of a priori economic theory andemphasizes the need of an appropriate and rich statistical model of the data. Graphical123

Moneta Chlass Entner Hoyercausal models, independent component analysis, and tests of conditional independenceare the tools we propose for structural identification in vector autoregressive models.We conclude with an overview of some important issues which are left open in thisdomain.1. Specification of the statistical model. Data driven procedures for SVAR identificationdepend upon the specification of the (reduced form) VAR model. Therefore, itis important to make sure that the estimated VAR model is an accurate description ofthe dynamics of the included variables (whereas the contemporaneous structure is intentionallyleft out, as seen in section 1.2). The usual criterion for accuracy is to checkthat the model estimates residuals conform to white noise processes (although serialindependence of residuals is not a sufficient criterion for model validation). This impliesstable dependencies captured by the relationships among the modeled variables,and an unsystematic noise. It may be the case, as in many empirical applications, thatdifferent VAR specifications pass the model checking tests equally well. For example,a VAR with Gaussian errors and p lags may fit the data equally well as a VAR withnon-Gaussian errors and q lags and these two specifications justify two different causalsearch procedures. So far, we do not know how to adjudicate among alternative andseemingly equally accurate specifications.2. Background knowledge and assumptions. Search algorithms are based on differentassumptions, such as, for example, causal sufficiency, acyclicity, the Causal MarkovCondition, Faithfulness, and/or the existence of independent components. Maybe,background knowledge could justify some of these assumptions and reject others. Forexample, institutional or theoretical knowledge about an economic process might informus that Faithfulness is a plausible assumption in some contexts rather than inothers, or instead, that one should expect feedback loops if data are collected at certainlevels of temporal aggregation. Yet, if background information could inform us here,this might again provoke a problem of circularity mentioned at the outset of the paper.3. Search algorithms in nonparametric settings. We have provided some informationon which nonparametric test procedures might be more appropriate in certain circumstances.However, it is not clear which causal search algorithms are most efficientin exploiting the nonparametric conditional independence tests proposed in Section 4.The more variables the search algorithm needs to be informed about at the same pointof the search, the higher the number of conditioned variables, and hence, the slower, orthe more inaccurate, the test.4. Number of shocks and number of variables. To conserve degrees of freedom,SVARs rarely model more than six to eight time series variables (Bernanke et al., 2005,p.388). It is an open question how the procedures for causal inference we reviewed canbe applied to large scale systems such as dynamic factor models. (Forni et al., 2000)5. Simulations and empirical applications. Graphical causal models for identifyingSVARs, equivalent or similar to the search procedures described in section 2, have beenapplied to several sets of macroeconomic data (Swanson and Granger, 1997; Besslerand Lee, 2002; Demiralp and Hoover, 2003; Moneta, 2004; Demiralp et al., 2008;Moneta, 2008; Hoover et al., 2009). Demiralp and Hoover (2003) present Monte Carlo124

Causal Search in SVARsimulations to evaluate the performance of the PC algorithm for such an identification.There are no simulation results so far about the performance of the alternative testson residual partial correlations presented in section 2.2. Moneta et al. (2010) appliedan independent component analysis as described in section 3, to microeconomic USdata about firms’ expenditures on R&D and performance, as well as to macroeconomicUS data about monetary policy and its effects on the aggregate economy. Hyvärinenet al. (2010) assess the performance of independent component analysis for identifyingSVAR models. It is yet to be established how independent component analysis appliedto SVARs fares compared to graphical causal models (based on the appropriate conditionalindependence tests) in non-Gaussian settings. Nonparametric tests of conditionalindependence, as those proposed in section 4, have been applied to test for Grangernon-causality (Su and White, 2008), but there are not yet any applications where thesetest results inform a graphical causal search algorithm. Overall, there is a need for moreempirical applications of the procedures described in this paper. Such applications willbe useful to test, compare, and improve different search procedures, to suggest newproblems, and obtain new causal knowledge.6. Appendix6.1. Appendix 1 - Details of the bootstrap procedure from 4.1.(1) Draw a bootstrap sampling Z * t (for t = 1,...,n) from the estimated kernel densityˆ f (z) = n −1 b −d ∑︀ nt=1K p ((Z t − z)/b).(2) For t = 1,...,n, given Z t *, draw X* t and Y t * independently from the estimated kerneldensity f ˆ(x|Z t *) and f ˆ(y|Z t *) respectively.(3) Using X * t , Y* t , and Z* t , compute the bootstrap statistic S * n using one of the distancesdefined above.(4) Repeat steps (1) and (2) I times to obtain I statistics {S * ni }I i=1 .(5) The p-value is then obtained by:∑︀ Ii=11{Sni * p ≡> S n},Iwhere S n is the statistic obtained from the original data using one of the distancesdefined above, and 1{·} denotes an indicator function taking value one ifthe expression between brackets is true and zero otherwise.ReferencesE. Baek and W. Brock. A general test for nonlinear Granger causality: Bivariate model.Discussin paper, Iowa State University and University of Wisconsin, Madison, 1992.125

Moneta Chlass Entner HoyerL. Baringhaus and C. Franz. On a new multivariate two-sample test. Journal of MultivariateAnalysis, 88(1):190–206, 2004.B. S. Bernanke. Alternative explanations of the money-income correlation. InCarnegie-Rochester Conference Series on Public Policy, volume 25, pages 49–99.Elsevier, 1986.B.S. Bernanke, J. Boivin, and P. Eliasz. Measuring the Effects of Monetary Policy: AFactor-Augmented Vector Autoregressive (FAVAR) Approach. Quarterly Journal ofEconomics, 120(1):387–422, 2005.D. A. Bessler and S. Lee. Money and prices: US data 1869-1914 (a study with directedgraphs). Empirical Economics, 27:427–446, 2002.O. J. Blanchard and D. Quah. The dynamic effects of aggregate demand and supplydisturbances. The American Economic Review, 79(4):655–673, 1989.O. J. Blanchard and M. W. Watson. Are business cycles all alike? The Americanbusiness cycle: Continuity and change, 25:123–182, 1986.N. Chlaß and A. Moneta. Can Graphical Causal Inference Be Extended to NonlinearSettings? EPSA Epistemology and Methodology of Science, pages 63–72, 2010.T. Chu and C. Glymour. Search for additive nonlinear time series causal models. TheJournal of Machine Learning Research, 9:967–991, 2008.P. Comon. Independent component analysis, a new concept? Signal processing, 36(3):287–314, 1994.S. Demiralp and K. D. Hoover. Searching for the causal structure of a vector autoregression.Oxford Bulletin of Economics and Statistics, 65:745–767, 2003.S. Demiralp, K. D. Hoover, and D. J. Perez. A Bootstrap method for identifying andevaluating a structural vector autoregression. Oxford Bulletin of Economics andStatistics, 65, 745-767, 2008.M. Eichler. Granger causality and path diagrams for multivariate time series. Journalof Econometrics, 137(2):334–353, 2007.J. Faust and E. M. Leeper. When do long-run identifying restrictions give reliableresults? Journal of Business & Economic Statistics, 15(3):345–353, 1997.J. P. Florens and M. Mouchart. A note on noncausality. Econometrica, 50(3):583–591,1982.M. Forni, M. Hallin, M. Lippi, and L. Reichlin. The generalized dynamic-factor model:Identification and estimation. Review of Economics and Statistics, 82(4):540–554,2000.126

Causal Search in SVARC. W. J. Granger. Investigating causal relations by econometric models and crossspectralmethods. Econometrica: Journal of the Econometric Society, 37(3):424–438, 1969.C. W. J. Granger. Testing for causality:: A personal viewpoint. Journal of EconomicDynamics and Control, 2:329–352, 1980.T. Haavelmo. The probability approach in econometrics. Econometrica, 12:1–115,1944.C. Hiemstra and J. D. Jones. Testing for linear and nonlinear Granger causality in thestock price-volume relation. Journal of Finance, 49(5):1639–1664, 1994.W. C. Hood and T. C. Koopmans. Studies in econometric method, Cowles CommissionMonograph, No. 14. New York: John Wiley & Sons, 1953.K. D. Hoover. Causality in macroeconomics. Cambridge University Press, 2001.K. D. Hoover. The methodology of econometrics. New Palgrave Handbook of Econometrics,1:61–87, 2006.K. D. Hoover. Causality in economics and econometrics. In The New Palgrave Dictionaryof Economics. London: Palgrave Macmillan, 2008.K.D. Hoover, S. Demiralp, and S.J. Perez. Empirical Identification of the Vector Autoregression:The Causes and Effects of US M2. In The Methodology and Practiceof Econometrics. A Festschrift in Honour of David F. Hendry, pages 37–58. OxfordUniversity Press, 2009.P. O. Hoyer, A. Hyvärinen, R. Scheines, P. Spirtes, J. Ramsey, G. Lacerda, andS. Shimizu. Causal discovery of linear acyclic models with arbitrary distributions. InProceedings of the 24th Conference on Uncertainty in Artificial Intelligence, 2008a.P. O. Hoyer, S. Shimizu, A. J. Kerminen, and M. Palviainen. Estimation of causaleffects using linear non-gaussian causal models with hidden variables. InternationalJournal of Approximate Reasoning, 49:362–378, 2008b.A. Hyvärinen, J. Karhunen, and E. Oja. Independent Component Analysis. Wiley, 2001.A. Hyvärinen, K. Zhang, S. Shimizu, and P. O. Hoyer. Estimation of a StructuralVector Autoregression model using non-Gaussianity. Journal of Machine LearningResearch, 11:1709–1731, 2010.S. Johansen. Statistical analysis of cointegrating vectors. Journal of Economic Dynamicsand Control, 12:231–254, 1988.S. Johansen. Estimation and hypothesis testing of cointegrating vectors in Gaussianvector autoregressive models. Econometrica, 59:1551–1580, 1991.127

Moneta Chlass Entner HoyerS. Johansen. Cointegration: An Overview. In Palgrave Handbook of Econometrics.Volume 1. Econometric Theory, pages 540–577. Palgrave Macmillan, 2006.R. G. King, C. I. Plosser, J. H. Stock, and M. W. Watson. Stochastic trends and economicfluctuations. American Economic Review, 81:819–840, 1991.T. C. Koopmans. Statistical Inference in Dynamic Economic Models, Cowles CommissionMonograph, No. 10. New York: John Wiley & Sons, 1950.G. Lacerda, P. Spirtes, J. Ramsey, and P. O. Hoyer. Discovering cyclic causal modelsby Independent Components Analysis. In Proc. 24th Conference on Uncertainty inArtificial Intelligence (UAI-2008), Helsinki, Finland, 2008.R. E. Lucas. Econometric policy evaluation: A critique. In Carnegie-Rochester ConferenceSeries on Public Policy, volume 1, pages 19–46. Elsevier, 1976.H. Lütkepohl. Vector Autoregressive Models. In Palgrave Handbook of Econometrics.Volume 1. Econometric Theory, pages 477–510. Palgrave Macmillan, 2006.A. Moneta. Graphical Models for Structural Vector Autoregressions. LEM PapersSeries, Sant’Anna School of Advanced Studies, Pisa, 2003.A. Moneta. Identification of monetary policy shocks: a graphical causal approach.Notas Económicas, 20, 39-62, 2004.A. Moneta. Graphical causal models and VARs: an empirical assessment of the realbusiness cycles hypothesis. Empirical Economics, 35(2):275–300, 2008.A. Moneta, D. Entner, P.O. Hoyer, and A. Coad. Causal inference by independent componentanalysis with applications to micro-and macroeconomic data. Jena EconomicResearch Papers, 2010:031, 2010.E. Paparoditis and D. N. Politis. The local bootstrap for kernel estimators under generaldependence conditions. Annals of the Institute of Statistical Mathematics, 52(1):139–159, 2000.J. Pearl. Causality: models, reasoning and inference. Cambridge University Press,Cambridge, 2000.M. Reale and G. T. Wilson. Identification of vector AR models with recursive structuralerrors using conditional independence graphs. Statistical Methods and Applications,10, 49-65, 2001.T. Richardson and P. Spirtes. Automated discovery of linear feedback models. InComputation, causation and discovery. AAAI Press and MIT Press, Menlo Park,1999.128

Causal Search in SVARR. Scheines, P. Spirtes, C. Glymour, C. Meek, and T. Richardson. The TETRADproject: Constraint based aids to causal model specification. Multivariate BehavioralResearch, 33(1):65–117, 1998.M. D. Shapiro and M. W. Watson. Sources of business cycle fluctuations. NBERMacroeconomics annual, 3:111–148, 1988.S. Shimizu, P. O. Hoyer, A. Hyvärinen, and A. Kerminen. A linear non-Gaussianacyclic model for causal discovery. Journal of Machine Learning Research, 7:2003–2030, 2006.S. Shimizu, A. Hyvärinen, Y. Kawahara, and T. Washio. A direct method for estimatinga causal ordering in a linear non-Gaussian acyclic model. In Proceedings of the 25thConference on Uncertainty in Artificial Intelligence, 2009.C. A. Sims. Macroeconomics and Reality. Econometrica, 48, 1-47, 1980.C. A. Sims. An autoregressive index model for the u.s. 1948-1975. In J. Kmenta andJ.B. Ramsey, editors, Large-scale macro-econometric models: theory and practice,pages 283–327. North-Holland, 1981.P. Spirtes, C. Glymour, and R. Scheines. Causation, prediction, and search. MIT Press,Cambridge MA, 2nd edition, 2000.L. Su and H. White. A nonparametric Hellinger metric test for conditional independence.Econometric Theory, 24(04):829–864, 2008.P. Suppes. A probabilistic theory of causation. Acta Philosophica Fennica, XXIV, 1970.N. R. Swanson and C. W. J. Granger. Impulse response function based on a causal approachto residual orthogonalization in vector autoregressions. Journal of the AmericanStatistical Association, 92:357–367, 1997.G. J. Szekely and M. L. Rizzo. Testing for equal distributions in high dimension.InterStat, 5, 2004.M. P. Wand and M. C. Jones. Kernel smoothing. Chapman&Hall Ltd., London, 1995.A. H. Welsh, X. Lin, and R. J. Carroll. Marginal Longitudinal Nonparametric Regression.Journal of the American Statistical Association, 97(458):482–493, 2002.H. White and X. Lu. Granger Causality and Dynamic Structural Systems. Journal ofFinancial Econometrics, 8(2):193, 2010.N. Wiener. The theory of prediction. Modern mathematics for engineers, Series, 1:125–139, 1956.S. Wright. Correlation and causation. Journal of agricultural research, 20(7):557–585,1921.129

Moneta Chlass Entner HoyerA. Yatchew. Nonparametric regression techniques in economics. Journal of EconomicLiterature, 36(2):669–721, 1998.130

JMLR: Workshop and Conference Proceedings 12:115–139, 2011Causality in Time SeriesTime Series Analysis with the Causality WorkbenchIsabelle Guyonisabelle@clopinet.comClopiNet, Berkeley, CaliforniaAlexander StatnikovAlexander.Statnikov@nyumc.orgNYU Langone Medical Center, New York cityConstantin AliferisConstantin.Aliferis@nyumc.orgNYU Center for Health Informatics and Bioinformatics, New York cityEditors: Florin Popescu and Isabelle GuyonAbstractThe Causality Workbench project is an environment to test causal discovery algorithms.Via a web portal (http://clopinet.com/causality), it provides anumber of resources, including a repository of datasets, models, and software packages,and a virtual laboratory allowing users to benchmark causal discovery algorithmsby performing virtual experiments to study artificial causal systems. We regularlyorganize competitions. In this paper, we describe what the platform offers forthe analysis of causality in time series analysis.Keywords: Causality, Benchmark, Challenge, Competition, Time Series Prediction.1. IntroductionUncovering cause-effect relationships is central in many aspects of everyday life in bothhighly industrialized and developing countries: what affects our health, the economy,climate changes, world conflicts, and which actions have beneficial effects? Establishingcausality is critical to guiding policy decisions in areas including medicine andpharmacology, epidemiology, climatology, agriculture, economy, sociology, law enforcement,and manufacturing.One important goal of causal modeling is to predict the consequences of given actions,also called interventions, manipulations or experiments. This is fundamentallydifferent from the classical machine learning, statistics, or data mining setting, whichfocuses on making predictions from observations. Observations imply no manipulationon the system under study whereas actions introduce a disruption in the naturalfunctioning of the system. In the medical domain, this is the distinction made between“diagnosis” and “prognosis” (prediction from observations of diseases or disease evolution)and “treatment” (intervention). For instance, smoking and coughing might be bothpredictive of respiratory disease and helpful for diagnosis purposes. However, if smokingis a cause and coughing a consequence, acting on the cause (smoking) can changeyour health status, but not acting on the symptom or consequence (coughing). Thus itc○ 2011 I. Guyon, A. Statnikov & C. Aliferis.

Guyon Statnikov Aliferisis extremely important to distinguish between causes and consequences to predict theresult of actions like predicting the effect of forbidding smoking in public places.The need for assisting policy making while reducing the cost of experimentationand the availability of massive amounts of “observational” data prompted the proliferationof proposed computational causal discovery techniques (Glymour and Cooper,1999; Pearl, 2000; Spirtes et al., 2000; Neapolitan, 2003; Koller and Friedman, 2009),but it is fair to say that to this day, they have not been widely adopted by scientists andengineers. Part of the problem is the lack of appropriate evaluation and the demonstrationof the usefulness of the methods on a range of pilot applications. To fill this need,we started a project called the ”Causality Workbench”, which offers the possibility ofexposing the research community to challenging causal problems and disseminatingnewly developed causal discovery technology. In this paper, we outline our setup andmethods and the possibilities offered by the Causality Workbench to solve problems ofcausal inference in time series analysis.2. Causality in Time SeriesCausal discovery is a multi-faceted problem. The definition of causality itself haseluded philosophers of science for centuries, even though the notion of causality isat the core of the scientific endeavor and also a universally accepted and intuitive notionof everyday life. But, the lack of broadly acceptable definitions of causality hasnot prevented the development of successful and mature mathematical and algorithmicframeworks for inducing causal relationships.Causal relationships are frequently modeled by causal Bayesian networks or structuralequation models (SEM) (Pearl, 2000; Spirtes et al., 2000; Neapolitan, 2003). Inthe graphical representation of such models, an arrow between two variables A → Bindicates the direction of a causal relationship: A causes B. A node in the graph correspondingto a particular variable X, represents a “mechanism” to evaluate the value ofX, given the “parent” node variable values (immediate antecedents in the graph). ForBayesian networks, such evaluation is carried out by a conditional probability distributionP(X|Parents(X)) while for structural equation models it is carried out by a functionof the parent variables and a noise model.Our everyday-life concept of causality is very much linked to time dependencies(causes precede their effects). Hence an intuitive interpretation of an arrow in a causalnetwork representing A causes B is that A preceded B. 1 But, in reality, Bayesiannetworks are a graphical representation of a factorization of conditional probabilities,hence a pure mathematical construct. The arrows in a “regular” Bayesian network (not a“causal Bayesian network”) do not necessarily represent either causal relationships norprecedence, which often creates some confusion. In particular, many machine learningproblems are concerned with stationary systems or “cross-sectional studies”, which are1. More precise semantics have been developed. Such semantics assume discrete time point or intervaltime models and allow for continuous or episodic “occurrences” of the values of a variable as wellas overlapping or non-overlapping intervals (Aliferis, 1998). Such practical semantics in Bayesiannetworks allow for abstracted and explicit time.132

Causality Workbenchstudies where many samples are drawn at a given point in time. Thus, sometimes thereference to time in Bayesian networks is replaced by the notion of “causal ordering”.Causal ordering can be understood as fixing a particular time scale and considering onlycauses happening at time t and effects happening at time t + δt, where δt can be madeas small as we want. Within this framework, causal relationships may be inferred fromdata including no explicit reference to time. Causal clues in the absence of temporalinformation include conditional independencies between variables and loss of informationdue to irreversible transformations or the corruption of signal by noise (Sun et al.,2006; Zhang and Hyvärinen, 2009).In seems reasonable to think that temporal information should resolve many causalrelationship ambiguities. Yet, the addition of the time dimension simplifies the problemof inferring causal relationships only to a limited extend. For one, it reduces, but doesnot eliminate, the problem of confounding: A correlated event A happening in the pastof event B cannot be a consequence of B; however it is not necessarily a cause becausea previous event C might have been a “common cause” of A and B. Secondly, it opensthe door to many subtle modeling questions, including problems arising with modelingthe dynamic systems, which may or may not be stationary. One of the charters ofour Causality Workbench project is to collect both problems of practical and academicinterest to push the envelope of research in inferring causal relationships from timeseries analysis.3. A Virtual LaboratoryMethods for learning cause-effect relationships without experimentation (learning fromobservational data) are attractive because observational data is often available in abundanceand experimentation may be costly, unethical, impractical, or even plain impossible.Still, many causal relationships cannot be ascertained without the recourse toexperimentation and the use of a mix of observational and experimental data might bemore cost effective. We implemented a Virtual Lab allowing researchers to performexperiments on artificial systems to infer their causal structure. The design of the platformis such that researchers can submit new artificial systems for others to experiment,experimenters can place queries and get answers, the activity is logged, and registeredusers have their own virtual lab space. This environment allows researchers to testcomputational causal discovery algorithms and, in particular, to test whether modelingassumptions made hold in real and simulated data.We have released a first version http://www.causality.inf.ethz.ch/workbench.php. We plan to attach to the virtual lab sizeable realistic simulatorssuch as the Spatiotemporal Epidemiological Modeler (STEM), an epidemiology simulatordeveloped at IBM, now publicly available: http://www.eclipse.org/stem/. The virtual lab was put to work in a recent challenge we organized on theproblem of “Active Learning” (see http://clopinet.com/al). More details onthe virtual lab are given in the appendix.133

Guyon Statnikov Aliferis4. A Data RepositoryPart of our benchmarking effort is dedicated to collecting problems from diverse applicationdomains. Via the organization of competitions, we have successfully channeledthe effort or dozens of researchers to solve new problems of scientific and practicalinterest and identified effective methods. However, competition without collaborationis sterile. Recently, we have started introducing new dimensions to our effort of researchcoordination: stimulating creativity, collaborations, and data exchange. Weare organizing regular teleconference seminars. We have created a data repositoryfor the Causality Workbench already populated by 15 datasets. All the resources,which are the product of our effort, are freely available on the Internet at http://clopinet.com/causality. The repository already includes several time seriesdatasets, illustrating problems of practical and academic interest (see table 1):- Learning the structure of a fairly complex dynamic system that disobeys equilibrationmanipulationcommutability, and predicting the effect of manipulations that donot cause instabilities (the MIDS task) (Voortman et al., 2010);- Learning causal relationships using time series when noise is corrupting data in away that classical “Granger causality” fails (the NOISE task) (Nolte et al., 2010);- Uncovering which promotions affect most sales in a marketing database (thePROMO task) (Pellet, 2010);- Identifying in a manufacturing process (wafer production) faulty steps affectinga performance metric (the SEFTI task) (Tuv, 2010);- Modeling a biological signalling process (the SIGNET task) (Jenkins, 2010).The donor of the dataset NOISE (Guido Nolte) received the best dataset award. Thereviewers appreciated that the task includes both real data from EEG time series andartificial data modeling EEG. We want to encourage future data donors to move in thisdirection.5. Benchmarks and CompetitionsOur effort has been gaining momentum with the organization of two challenges, whicheach attracted over 50 participants. The first causality challenge we have organized(Causation and Prediction challenge, December 15 2007 - April 30 2008) allowed researchersboth from the causal discovery community and the machine learning communityto try their algorithms on sizable tasks of real practical interest in medicine, pharmacology,and sociology (see http://www.causality.inf.ethz.ch/challenge.php). The goal was to train models exclusively on observational data, then make predictionsof a target variable on data collected after intervention on the system understudy were performed. This first challenge reached a number of goals that we had setto ourselves: familiarizing many new researchers and practitioners with causal discoveryproblems and existing tools to address them, pointing out the limitations of current134

Causality WorkbenchTable 1: Time dependent datasets.“TP” is the data type, “NP” the number of participants who returned results and “V” the number ofviews as of January 2011. The semi-artificial datasets are obtained from simulators of real tasks. N is the number of variables,T is the number of time samples (not necessarily evenly spaced) and R the number of simulations with different initial statesor conditions.Name(TP; NP; V)MIDS(Artificial;NA; 794)NOISE(Real +artificial; NA;783)PROMO(Semiartificial;3;1601)SEFTI(Semiartificial;NA;908)SIGNET(Semi-artif.;2; 2663)Size Description ObjectiveT=12 sampled values in time(unevenly spaced); R=10000simulations. N=9 variables.Artificial: T=6000 time points;R=1000 simul.; N=2 var.Real: R=10 subjects.T≃200000 points sampled at256Hz. N=19 channels.T=365*3 days; R=1 simulation;N=1000 promotions + 100products.R=4000 manufacturing lots;T=300 async. operations (pairof values {one of N=25 toolIDs, date of proc.}) + cont.target (circuit perf. for each lot).T=21 asynchronous stateupdates; R=300 pseudodynamicsimulations; N=43 rules.Mixed Dynamic Systems. Simulatedtime-series based on linear Gaussianmodels with no latent common causes,but with multiple dynamic processes.Real and simulated EEG data.Learning causal relationships usingtime series when noise is corruptingdata causing the classical Grangercausality method to fail.Simulated marketing task. Dailyvalues of 1000 promotions and 100product sales for three yearsincorporating seasonal effects.Semiconductor manufacturing.Each wafer undergoes 300 steps eachinvolving one of 25 tools. A regressionproblem for quality control ofend-of-line circuit performance.Abscisic Acid Signaling Network.Model inspired by a true biologicalsignaling network.Use the training data tobuild a model able topredict the effects ofmanipulations on thesystem in test data.Artificial task: find thecausal dir. in pairs of var.Real task: Find whichbrain region influenceeach other.Predict a 1000x100boolean matrix of causalinfluences of promotionson product sales.Find the tools that areguilty of performancedegradation and eventualinteractions and influenceof time.Determine the set of 43boolean rules thatdescribe the network.135

Guyon Statnikov Aliferismethods on some particular difficulties, and fostering the development of new algorithms.The results indicated that causal discovery from observational data is not animpossible task, but a very hard one and pointed to the need for further research andbenchmarks (Guyon et al., 2008). The Causal Explorer package (Aliferis et al., 2003),which we had made available to the participants and is downloadable as shareware,proved to be competitive and is a good starting point for researchers new to the field. Itis a Matlab toolkit supporting “local” causal discovery algorithms, efficient to discoverthe causal structure around a target variable, even for a large number of variables. Thealgorithms are based on structure learning from tests of conditional independence, asall the top ranking methods in this first challenge.The first challenge (Guyon et al., 2008) explored an important problem in causalmodeling, but is only one of many possible problem statements. The second challenge(Guyon et al., 2010) called “competition pot-luck” aimed at enlarging the scope ofcausal discovery algorithm evaluation by inviting members of the community to submittheir own problems and/or solve problems proposed by others. The challenge startedSeptember 15, 2008 and ended November 20, 2008, see http://www.causality.inf.ethz.ch/pot-luck.php. One task proposed by a participant drew a lot ofattention: the cause-effect pair task. The problem was to try to determine in pairs ofvariables (of known causal relationships), which one was the cause of the other. Thisproblem is hard for a lot of algorithms, which rely on the result of conditional independencetests of three or more variables. Yet the winners of the challenge succeeded inunraveling 8/8 correct causal directions (Zhang and Hyvärinen, 2009).Our planned challenge ExpDeCo (Experimental Design in Causal Discovery) willbenchmark methods of experimental design in application to causal modeling. The goalwill be to identify effective methods to unravel causal models, requiring a minimum ofexperimentation, using the Virtual Lab. A budget of virtual cash will be allocated toparticipants to “buy” the right to observe or manipulate certain variables, manipulationsbeing more expensive that observations. The participants will have to spend theirbudget optimally to make the best possible predictions on test data. This setup lendsitself to incorporating problems of relevance to development projects, in particular inmedicine and epidemiology where experimentation is difficult while developing newmethodology.We are planning another challenge called CoMSICo for “Causal Models for SystemIdentification and Control”, which is more ambitious in nature because it will perform acontinuous evaluation of causal models rather than separating training and test phase. Incontrast with ExpDeCo in which the organizers will provide test data with prescribedmanipulations to test the ability of the participants to make predictions of the consequencesof actions, in CoMSICo, the participants will be in charge of making theirown plan of action (policy) to optimize an overall objective (e.g., improve the life expectancyof a population, improve the GNP, etc.) and they will be judged directly withthis objective, on an on-going basis, with no distinction between “training” and “test”data. This challenge will also be via the Virtual Lab. The participants will be givenan initial amount of virtual cash, and, as previously, both actions and observations will136

Causality Workbenchhave a price. New in CoMSICo, virtual cash rewards will be given for achieving goodintermediate performance, which the participants will be allowed to re-invest to conductadditional experiments and improve their plan of action (policy). The winner willbe the participant ending up with the largest amount of virtual cash.6. ConclusionOur program of data exchange and benchmark proposes to challenge the research communitywith a wide variety of problems from many domains and focuses on realisticsettings. Causal discovery is a problem of fundamental and practical interest in manyareas of science and technology and there is a need for assisting policy making in allthese areas while reducing the costs of data collection and experimentation. Hence, theidentification of efficient techniques to solve causal problems will have a widespreadimpact. By choosing applications from a variety of domains and making connectionsbetween disciplines as varied as machine learning, causal discovery, experimental design,decision making, optimization, system identification, and control, we anticipatethat there will be a lot of cross-fertilization between different domains.AcknowledgmentsThis project is an activity of the Causality Workbench supported by the Pascal networkof excellence funded by the European Commission and by the U.S. National ScienceFoundation under Grant N0. ECCS-0725746. Any opinions, findings, and conclusionsor recommendations expressed in this material are those of the authors and do not necessarilyreflect the views of the National Science Foundation. We are very grateful toall the members of the causality workbench team for their contribution and in particularto our co-founders Constantin Aliferis, Greg Cooper, André Elisseeff, Jean-PhilippePellet, Peter Spirtes, and Alexander Statnikov.ReferencesC. F. Aliferis, I. Tsamardinos, A. Statnikov, and L.E. Brown. Causal explorer: A probabilisticnetwork learning toolkit for biomedical discovery. In 2003 InternationalConference on Mathematics and Engineering Techniques in Medicine and BiologicalSciences (METMBS), Las Vegas, Nevada, USA, June 23-26 2003. CSREA Press.Constantin Aliferis. A Temporal Representation and Reasoning Model for MedicalDecision-Support Systems. PhD thesis, University of Pittsburgh, 1998.C. Glymour and G.F. Cooper, editors. Computation, Causation, and Discovery. AAAIPress/The MIT Press, Menlo Park, California, Cambridge, Massachusetts, London,England, 1999.I. Guyon, C. Aliferis, G. Cooper, A. Elisseeff, J.-P. Pellet, P. Spirtes, and A. Statnikov.Design and analysis of the causation and prediction challenge. In JMLR W&CP,137

Guyon Statnikov Aliferisvolume 3, pages 1–33, WCCI2008 workshop on causality, Hong Kong, June 3-42008.I. Guyon, D. Janzing, and B. Schölkopf. Causality: Objectives and assessment. JMLRW&CP, 6:1–38, 2010.Jerry Jenkins. Signet: Boolean rile determination for abscisic acid signaling. InCausality: objectives and assessment (NIPS 2008), volume 6, pages 215–224. JMLRW&CP, 2010.Daphne Koller and Nir Friedman.Techniques. MIT Press, 2009.Probabilistic Graphical Models: Principles andR. E. Neapolitan. Learning Bayesian Networks. Prentice Hall series in Artificial Intelligence.Prentice Hall, 2003.G. Nolte, A. Ziehe, N. Krämer, F. Popescu, and K.-R. Müller. Comparison of grangercausality and phase slope index. In Causality: objectives and assessment (NIPS2008), volume 6, pages 267–276. JMLR W&CP, 2010.J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press,2000.J.-P. Pellet. Detecting simple causal effects in time series. In Causality: objectives andassessment (NIPS 2008). JMLR W&CP volume 6, supplemental material, 2010.P. Spirtes, C. Glymour, and R. Scheines. Causation, Prediction, and Search. The MITPress, Cambridge, Massachusetts, London, England, 2000.X. Sun, D. Janzing, and B. Schölkopf. Causal inference by choosing graphs with mostplausible Markov kernels. In Ninth International Symposium on Artificial Intelligenceand Mathematics, 2006.E. Tuv. Pot-luck challenge: Tied. In Causality: objectives and assessment (NIPS 2008).JMLR W&CP volume 6, supplemental material, 2010.M. Voortman, D. Dash, and M. J. Druzdzel. Learning causal models that make correctmanipulation predictions. In Causality: objectives and assessment (NIPS 2008),volume 6, pages 257–266. JMLR W&CP, 2010.K. Zhang and A. Hyvärinen. Distinguishing causes from effects using nonlinear acycliccausal models. In Causality: objectives and assessment (NIPS 2008), volume 6,pages 157–164. JMLR W&CP, 2009.138

TABLE OF CONTENTS 139

Causality in Time Series - ClopiNet

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?