Causality in Time Series - ClopiNet
Causality in Time Series - ClopiNet
Causality in Time Series - ClopiNet
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
L<strong>in</strong>k<strong>in</strong>g Granger <strong>Causality</strong> 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,"conform<strong>in</strong>g to notation formally <strong>in</strong>troduced 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 def<strong>in</strong><strong>in</strong>g thegame. For simplicity here, we focus on “pure strategy" games; see Gibbons, R. (1992)for an accessible <strong>in</strong>troduction 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 <strong>in</strong> 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 determ<strong>in</strong>ed by both players’ strategies and the punishment structure.Another well known game is hide and seek. Here, player 1 w<strong>in</strong>s by match<strong>in</strong>g player2’s strategy and player 2 w<strong>in</strong>s by mismatch<strong>in</strong>g player 1’s strategy. A familiar exampleis a penalty kick <strong>in</strong> soccer: the goalie w<strong>in</strong>s by match<strong>in</strong>g the direction (right or left) ofthe kicker’s kick; the kicker w<strong>in</strong>s by mismatch<strong>in</strong>g the direction of the goalie’s lunge.The same structure applies to baseball (hitter vs. pitcher) or troop deployment <strong>in</strong> battle(aggressor vs. defender).A third famous game is battle of the sexes. In this game, Ralph and Alice are try<strong>in</strong>gto decide how to spend their weekly night out. Alice prefers the opera, and Ralphprefers box<strong>in</strong>g; but both would rather be together than apart.Now consider whether the PCM permits causal discourse <strong>in</strong> these games, e.g., aboutthe effect of one player’s action on that of the other. We beg<strong>in</strong> by mapp<strong>in</strong>g 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 determ<strong>in</strong>edwith<strong>in</strong> the system, i.e., the PCM endogenous variables are z := (z 1 ,z 2 ) correspond<strong>in</strong>g tov, provided that (for now) we drop the dist<strong>in</strong>ction between y i and z i . F<strong>in</strong>ally, we see thatthe best response functions ri e correspond to the PCM structural functions f i .To determ<strong>in</strong>e whether the PCM permits causal discourse <strong>in</strong> these games, we cancheck whether there is a unique fixed po<strong>in</strong>t for the best responses. In prisoner’s dilemma,there is <strong>in</strong>deed a unique fixed po<strong>in</strong>t (both confess), provided the punishments are suitablychosen. The PCM therefore applies to this game to support causal discourse. Butthere is no fixed po<strong>in</strong>t for hide and seek, so the PCM cannot support causal discoursethere. On the other hand, there are two fixed po<strong>in</strong>ts for battle of the sexes: both Ralphand Alice choose opera or both choose box<strong>in</strong>g. The PCM does not support causal discoursethere either. Nor does the GPCM apply to the latter games, because even though5