Learning by Ruling: A Dynamic Model of Trade Disputes

Transcription

1 Learning by Ruling: A Dynamic Model of Trade Disputes Giovanni Maggi and Robert W. Staiger Yale and Dartmouth June 2016 Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

2 Introduction There have been over 500 disputes in the WTO since Sometimes govs settle early, sometimes they fight it out to a court (DSB) ruling. In GATT+WTO, about 50% of disputes settle early. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

3 Introduction There have been over 500 disputes in the WTO since Sometimes govs settle early, sometimes they fight it out to a court (DSB) ruling. In GATT+WTO, about 50% of disputes settle early. Stakes of trade disputes can be large, so important to understand what determines dispute initiation and resolution. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

4 Introduction There have been over 500 disputes in the WTO since Sometimes govs settle early, sometimes they fight it out to a court (DSB) ruling. In GATT+WTO, about 50% of disputes settle early. Stakes of trade disputes can be large, so important to understand what determines dispute initiation and resolution. Some interesting dynamic patterns: countries fight less as time goes by. Plots 1+2. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

5 Plot 1 RULING DISPUTE Number (wto disputes) #10-6 Plot 2 5 RULING DISPUTE Age of Dyad in Years Note: The vertical axis records the ratio X/Z, where X is the average number of disputes or rulings involving dyads of a given age, and Z is the average trade volume of dyads in this age group.

6 Introduction There have been over 500 disputes in the WTO since Sometimes govs settle early, sometimes they fight it out to a court (DSB) ruling. In GATT+WTO, about 50% of disputes settle early. Stakes of trade disputes can be large, so important to understand what determines dispute initiation and resolution. Some interesting dynamic patterns: countries fight less as time goes by. Plots 1+2. The judicial system is being used less and less... Is this bad news? Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

7 Introduction We propose a theory that can attribute this trend to the effects of judicial learning (good news, not bad news). A key model prediction: If there is court learning by ruling, the frequency of disputes and rulings should decline with court experience. We attempt to gauge empirically the importance, scope and form of judicial learning in the WTO dispute settlement system. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

8 Introduction In the empirical literature on learning by doing, econometricians (try to) measure productivity directly. We can t measure court accuracy directly, but we can try to infer learning effects indirectly: We check if WTO data exhibits imprint of court learning in accordance with our model predictions. We seek to shed light on the scope and form of court learning as viewed through the lens of our model. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

9 Introduction: the model Key ingredients of our model: Importing gov chooses a policy, then exporting gov chooses whether to dispute; If a dispute is initiated, govs bargain in the shadow of the law, subject to negotiation costs; If invoked, DSB issues a ruling to maximize govs joint payoff based on noisy information; Learning by ruling: court accuracy increases with experience, but at diminishing rate; Govs are large players that repeatedly engage in disputes (so they internalize benefits of court learning). Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

10 Introduction: preview of theoretical results In a static setting, there is never a ruling. But a dispute can arise, and is more likely if DSB less accurate. In a dynamic setting, the presence of court learning can give rise to equilibrium rulings. Likelihood of rulings and disputes decreases w/ cumulative rulings, at least if govs are patient enough. Results above largely extend to a multi-country setting with learning spillovers; Also to setting where disputed policy is discrete and transfers are costly (and baseline rate of rulings exists even in static setting). Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

11 Introduction: preview of empirical results Focus on the prediction that likelihood of rulings/disputes decreases with cumulative rulings. Allow scope of court learning to be general, specific to the disputing countries, or specific to the issue area (e.g. GATT/WTO Article). We find evidence of article-specific and disputant-specific learning, but only weak evidence of general-scope learning. Once learning by ruling taken into account, estimated time trend of disputes and rulings turns positive. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

12 Introduction: related literature Maggi and Staiger (2011): does not allow for bargaining/settlement, does not consider learning effects (and focuses on institutional design). Maggi and Staiger (2015a): govs can settle, but model is static, with no dispute-initiation stage. Focuses only on how contract form ( property vs liability ) impacts prob of settlement. Other models that generate trade disputes in equilibrium: Park (2011), Beshkar (2016), Staiger and Sykes (forthcoming). Main focus not on determinants of dispute outcome. Judicial learning: Baker and Mezzetti (2012), Beim (2014). Empirical work on trade dispute outcomes: Guzman and Simmons (2002), Busch and Reinhardt (2001, 2006), Bown and Reynolds (2014), Maggi and Staiger (2015b). Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

13 The basic model A single industry; importing gov (Home) chooses tariff T ; exporting gov (Foreign) passive in this industry. Home s payoff: ω(t, θ) = CS(T ) + R(T ) + θ PS(T ). Foreign s payoff: ω (T ) = CS (T ) + PS (T ). Joint-payoff-maximizing policy: T fb = arg max T [ω(t, θ) + ω (T )]. Veil of ignorance: ex-ante each gov equally likely to be importer or exporter (i.e. claimant or defendant). Absent international transfers, Pareto frontier is concave. With transfers, Pareto frontier is linear with slope -1. Figure 1a. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

14 Figure 1: Continuous policy (static setting) ω N FB slope=-1 ω*

15 The basic model Political shock θ ( state of the world ) not verifiable. Incomplete contract: does not specify policy T, but the court (DSB) can fill gaps ex-post. If invoked, DSB observes a noisy signal of T fb, denoted T dsb = T fb + ε, with Var(ε) = σ 2, and issues a ruling T dsb to maximize govs joint payoff. Timing: (0) θ realized and observed by govs; (1) Importer chooses T ; (2) Exporter acquiesces or initiates dispute; (3) If dispute initiated, govs negotiate over policy T and a transfer; (4) If govs disagree, DSB ruling is triggered. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

16 The dispute subgame Bargaining protocol: each gov makes take-or-leave offer with prob 1/2. Govs automatically incur symmetric litigation cost C L if court is triggered. Gov i s net disagreement payoff: ω ND i = ω D i C L, where = E [ω i (T dsb, θ) θ]. ω D i Negotiation costs: part of bargaining surplus melts away. If ω B i is player i s bargaining ( payoff absent ) negotiation costs, her payoff gain from bargaining is κ ω B i ω ND i with κ (0, 1). Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

17 Static benchmark Backward induction: start with the dispute subgame, where govs bargain in the shadow of the law. Given concave frontier, disagreement point D is below frontier, and Southeast of FB point (b/c uncertainty about court ruling hurts importer, benefits exporter). Litigation costs further worsen threat point for both govs (ND). Figure 1b. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

18 Figure 1: Continuous policy (static setting) ω N FB slope=-1 ND D ω*

19 Static benchmark If govs negotiate, given the negotiation cost κ they move part-way toward negotiation frontier. Net bargaining payoff point Bnet. Moving backwards (Figure 1c): (i) if Bnet point below no-transfer frontier, Home chooses T s.t. Foreign is indifferent b/w complaining and not, hence no dispute (B 0 ). (ii) If Bnet above no-transfer frontier, Home triggers a dispute (by choosing a bad T ). Clearly there is never a ruling in equilibrium. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

20 Figure 1: Continuous policy (static setting) ω N FB B 0 slope=-1 ND D B net ω*

21 Static benchmark Proposition 1: In the static setting: (i) there is never a DSB ruling; (ii) the likelihood of a dispute is increasing in σ. Graphical intuition for part (ii): as σ, point Bnet moves Southeast starting below no-transfer frontier, and crosses this frontier from the left. Figure 2. Remark 1: In the static setting, the equilibrium joint payoff is decreasing and piecewise concave in the DSB noise σ. Note, decreasing σ increases joint surplus through off-equilibrium effects, because there is no ruling in equilibrium: Lower σ improves disagreement point in case of dispute. If there is no dispute, lower σ improves the would-be negotiation outcome, thus inducing Home to choose a better T (off-off-equilibrium effect). Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

22 Figure 2: Continuous policy (static setting) ω N FB ND(σ) B net (σ) ω*

23 Learning by ruling Basic idea: DSB becomes more accurate with experience. Two periods, t = 1, 2. Discount factor δ (0, ). The same game takes place in each period, and θ is iid, so learning-by-ruling is the only source of dynamics. We model learning-by-ruling similarly as in standard models of learning-by-doing for firms: ability of DSB increases with cumulative past rulings. If there has been a ruling at t = 1, DSB is more accurate at t = 2, so σ is lower. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

24 Learning by ruling At t = 2 of course there is no ruling. But learning can give rise to equilibrium rulings at t = 1. Recall that at t = 2, joint payoff (Ω) is decreasing in σ. Given veil of ignorance, going to court at t = 1 implies common future payoff gain > 0. Learning is beneficial ( > 0) even though there is no ruling in equilibrium at t = 2, because it improves disagreement point for tomorrow s negotiation. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

25 Learning by ruling At t = 1, disagreement payoffs are (ω ND + δ, ω ND + δ ). Graphically, label this point ND + δ. If ND + δ is above negotiation frontier, then a dispute ends in ruling. And going backwards, Home chooses a T that triggers a dispute. Figure 3. Thus the presence of court learning (together with large players and negotiation costs) can explain equilibrium rulings. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

26 Figure 3: Continuous policy (two period setting) ω N FB ND+ δ Δ D ND ω*

27 Impact of court experience on current outcomes How does court experience affect the likelihood of rulings and disputes? Keep two periods, t = 1, 2, but suppose there is an initial stock of rulings x, inherited from a past period t = 0. Will examine how likelihood of rulings and disputes at t = 1 depends on x. Learning-by-ruling curve σ(x) decreasing and convex, with lim x σ(x) > 0. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

28 Impact of court experience on current rulings Static effect: x reduces today s ineffi ciency from going to court, because as σ the disagreement point ND gets closer to negotiation frontier. Dynamic effect: x reduces future gain from going to court ( ). This follows because E Ω t=2 is concave in σ (Remark 1) and σ > 0. If δ suffi ciently large, dynamic effect dominates static effect and Pr(Ruling) decreases with x. Even if δ is small, Pr(Ruling) decreasing in x for x suffi ciently large, because as learning gets exhausted 0, so Pr(Ruling) must hit zero at some point. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

29 Figure 3: Continuous policy (two period setting) ω N FB ND+ δ Δ D ND ω*

30 Impact of court experience on current rulings Proposition 2a: At t = 1, Pr(ruling) is decreasing in x for x suffi ciently large, and is globally decreasing in x if δ is high enough. Note, frequency of DSB use not a good measure of the institution s effectiveness. A declining ruling frequency is not bad news, in fact it s a symptom of beneficial learning. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

31 Impact of court experience on current disputes Next consider the impact of x on the probability of a dispute at t = 1. As in static setting, there is dispute iff net bargaining payoff point (B net ) above no-transfer frontier, except that now disagreement point is ND + δ. Proposition 2b: At t = 1, Pr(dispute) is decreasing in x for x suffi ciently large, and is globally decreasing in x if δ is high enough. Intuition for high δ case: The dynamic effect (increasing x reduces ) dominates. This worsens disagreement point, and hence (due to negotiation costs) worsens net bargaining payoffs, making dispute less appealing. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

32 Impact of court experience on current settlement rate Note, the model does not yield sharp predictions on the conditional likelihood of settlement: Remark 2: At t = 1 the likelihood of settlement conditional on a dispute may go up or down with x, even if δ is high. Intuition: Effect of x on ruling margin (i.e. when ND + δ is on negotiation frontier) may be stronger or weaker than effect of x on dispute margin (i.e. when B net is on no-transfer frontier), depending on prob distribution of θ. So Pr(ruling)/Pr(dispute) can go either way. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

33 Empirical evidence We focus on a key implication of the model: the likelihood of current rulings and disputes should tend to decrease with cumulative rulings. We show this implication generalizes to a many-country extension of our model, where learning could be specific to an issue area, or to the disputant countries, or it could be general. So we explore the effect of cumulative rulings at various levels (country dyad, article, general). Dual objective: (1) Check if key model prediction is consistent with data; (2) If so, gauge the strength and scope of learning by ruling. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

34 Regressions Focus first on the impact of court experience on the likelihood of rulings. Start with undirected-dyads, then use directed dyads. Table 2. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

35 Regressions Focus first on the impact of court experience on the likelihood of rulings. Start with undirected-dyads, then use directed dyads. Table 2. Results are consistent with: Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

36 Regressions Focus first on the impact of court experience on the likelihood of rulings. Start with undirected-dyads, then use directed dyads. Table 2. Results are consistent with: article-specific learning; Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

37

38 Table 2: Logit VARIABLES Undirected Dyad Directed Dyad (1) (2) VARIABLES (3) (4) DLogit_iiiikt RLogit_iiiikt DLogit_iiiikt RLogit_iiiikt CR_iiiikt 0.186*** CR_iiiikt 0.209*** 0.235** (0.0646) (0.0913) (0.0786) (0.0999) CR_n(iiii)kt * *** CR_ nnnn jjkt 0.117*** ( ) (0.0168) (0.0186) (0.0299) CR_iiiinkt *** *** CR_ii(nnnn)kt ( ) ( ) (0.0277) (0.0502) CR_n(iiii)nkt ** *** CR_jjjjkt ** ( ) ( ) (0.0941) (0.184) CR_oooooooookt *** *** ( ) (0.0193) CR_iiiinnkt *** *** ( ) ( ) CR_ nnnn jjnkt ( ) ( ) CR_ii(nnnn)nkt *** *** ( ) ( ) CR_jjjjnkt *** ( ) ( ) CR_ooooooooonkt *** ( ) ( ) t 0.198*** 1.277*** t 0.235*** 1.409*** (0.0260) (0.111) (0.0258) (0.117) t *** *** t *** *** ( ) ( ) ( ) ( ) Constant *** *** Constant *** *** (1.076) (0.723) (1.077) (0.792) Observations 439, ,560 Observations 545, ,520 Pseudo R Pseudo R iiiife Y Y iiiife Y Y kfe Y Y kfe Y Y Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 iiiife (iiiife) are undirected (directed) dyad fixed effects. kfe are article fixed effects.

39 Regressions Focus first on the impact of court experience on the likelihood of rulings. Start with undirected-dyads, then use directed dyads. Table 2. Results are consistent with: article-specific learning; disputant-specific (directed dyad- and claimaint-specific) learning; Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

40

41 Table 2: Logit VARIABLES Undirected Dyad Directed Dyad (1) (2) VARIABLES (3) (4) DLogit_iiiikt RLogit_iiiikt DLogit_iiiikt RLogit_iiiikt CR_iiiikt 0.186*** CR_iiiikt 0.209*** 0.235** (0.0646) (0.0913) (0.0786) (0.0999) CR_n(iiii)kt * *** CR_ nnnn jjkt 0.117*** ( ) (0.0168) (0.0186) (0.0299) CR_iiiinkt *** *** CR_ii(nnnn)kt ( ) ( ) (0.0277) (0.0502) CR_n(iiii)nkt ** *** CR_jjjjkt ** ( ) ( ) (0.0941) (0.184) CR_oooooooookt *** *** ( ) (0.0193) CR_iiiinnkt *** *** ( ) ( ) CR_ nnnn jjnkt ( ) ( ) CR_ii(nnnn)nkt *** *** ( ) ( ) CR_jjjjnkt *** ( ) ( ) CR_ooooooooonkt *** ( ) ( ) t 0.198*** 1.277*** t 0.235*** 1.409*** (0.0260) (0.111) (0.0258) (0.117) t *** *** t *** *** ( ) ( ) ( ) ( ) Constant *** *** Constant *** *** (1.076) (0.723) (1.077) (0.792) Observations 439, ,560 Observations 545, ,520 Pseudo R Pseudo R iiiife Y Y iiiife Y Y kfe Y Y kfe Y Y Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 iiiife (iiiife) are undirected (directed) dyad fixed effects. kfe are article fixed effects.

42 Regressions Focus first on the impact of court experience on the likelihood of rulings. Start with undirected-dyads, then use directed dyads. Table 2. Results are consistent with: article-specific learning; disputant-specific (directed dyad- and claimaint-specific) learning; only weak evidence of general-scope learning. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

43

44 Table 2: Logit VARIABLES Undirected Dyad Directed Dyad (1) (2) VARIABLES (3) (4) DLogit_iiiikt RLogit_iiiikt DLogit_iiiikt RLogit_iiiikt CR_iiiikt 0.186*** CR_iiiikt 0.209*** 0.235** (0.0646) (0.0913) (0.0786) (0.0999) CR_n(iiii)kt * *** CR_ nnnn jjkt 0.117*** ( ) (0.0168) (0.0186) (0.0299) CR_iiiinkt *** *** CR_ii(nnnn)kt ( ) ( ) (0.0277) (0.0502) CR_n(iiii)nkt ** *** CR_jjjjkt ** ( ) ( ) (0.0941) (0.184) CR_oooooooookt *** *** ( ) (0.0193) CR_iiiinnkt *** *** ( ) ( ) CR_ nnnn jjnkt ( ) ( ) CR_ii(nnnn)nkt *** *** ( ) ( ) CR_jjjjnkt *** ( ) ( ) CR_ooooooooonkt *** ( ) ( ) t 0.198*** 1.277*** t 0.235*** 1.409*** (0.0260) (0.111) (0.0258) (0.117) t *** *** t *** *** ( ) ( ) ( ) ( ) Constant *** *** Constant *** *** (1.076) (0.723) (1.077) (0.792) Observations 439, ,560 Observations 545, ,520 Pseudo R Pseudo R iiiife Y Y iiiife Y Y kfe Y Y kfe Y Y Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 iiiife (iiiife) are undirected (directed) dyad fixed effects. kfe are article fixed effects.

45 Table 3: OLS VARIABLES Undirected Dyad Directed Dyad (1) (2) VARIABLES (3) (4) D_iiiikt R_iiiikt D_iiiikt R_iiiikt CR_iiiikt *** *** CR_iiiikt *** ** ( ) ( ) ( ) ( ) CR_n(iiii)kt *** ** CR_ nnnn jjkt ** ( ) ( ) ( ) ( ) CR_iiiinkt *** *** CR_ii(nnnn)kt ( ) ( ) ( ) ( ) CR_n(iiii)nkt -1.09e e-06 CR_jjjjkt ** (1.32e-05) (1.76e-05) ( ) ( ) CR_oooooooookt *** *** ( ) ( ) CR_iiiinnkt *** *** ( ) ( ) CR_ nnnn jjnkt -5.47e e-06 (1.28e-05) (2.23e-05) CR_ii(nnnn)nkt -5.33e-05** -5.12e-05* (2.13e-05) (2.82e-05) CR_jjjjnkt e-05 ( ) ( ) CR_ooooooooonkt -6.42e e-06 (1.01e-05) (1.55e-05) t ** *** t ** *** ( ) ( ) ( ) ( ) t e t e e-05* (3.17e-05) (6.77e-05) (2.44e-05) (5.00e-05) Constant ** Constant *** -6.12e-06 ( ) ( ) ( ) ( ) Observations 439, ,560 Observations 545, ,520 R R iiiife Y Y iiiife Y Y kfe Y Y kfe Y Y CE iiii iiii CE iiii iiii Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 iiiife (iiiife) are undirected (directed) dyad fixed effects. kfe are article fixed effects. CE are standard errors clustered by undirected ( iiii ) or directed ( iiii ) dyads

46 Regressions Focus first on the impact of court experience on the likelihood of rulings. Start with undirected-dyads, then use directed dyads. Table 2. Results are consistent with: article-specific learning; disputant-specific (directed dyad- and claimaint-specific) learning; only weak evidence of general-scope learning. Note positive coeffi cient of t trend: suggests that court learning can explain declining trend in rulings (Plots 1+2). Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

47 Table 2: Logit VARIABLES Undirected Dyad Directed Dyad (1) (2) VARIABLES (3) (4) DLogit_iiiikt RLogit_iiiikt DLogit_iiiikt RLogit_iiiikt CR_iiiikt 0.186*** CR_iiiikt 0.209*** 0.235** (0.0646) (0.0913) (0.0786) (0.0999) CR_n(iiii)kt * *** CR_ nnnn jjkt 0.117*** ( ) (0.0168) (0.0186) (0.0299) CR_iiiinkt *** *** CR_ii(nnnn)kt ( ) ( ) (0.0277) (0.0502) CR_n(iiii)nkt ** *** CR_jjjjkt ** ( ) ( ) (0.0941) (0.184) CR_oooooooookt *** *** ( ) (0.0193) CR_iiiinnkt *** *** ( ) ( ) CR_ nnnn jjnkt ( ) ( ) CR_ii(nnnn)nkt *** *** ( ) ( ) CR_jjjjnkt *** ( ) ( ) CR_ooooooooonkt *** ( ) ( ) t 0.198*** 1.277*** t 0.235*** 1.409*** (0.0260) (0.111) (0.0258) (0.117) t *** *** t *** *** ( ) ( ) ( ) ( ) Constant *** *** Constant *** *** (1.076) (0.723) (1.077) (0.792) Observations 439, ,560 Observations 545, ,520 Pseudo R Pseudo R iiiife Y Y iiiife Y Y kfe Y Y kfe Y Y Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 iiiife (iiiife) are undirected (directed) dyad fixed effects. kfe are article fixed effects.

48 Plot 1 RULING DISPUTE Number (wto disputes) #10-6 Plot 2 5 RULING DISPUTE Age of Dyad in Years Note: The vertical axis records the ratio X/Z, where X is the average number of disputes or rulings involving dyads of a given age, and Z is the average trade volume of dyads in this age group.

49 Regressions Focus first on the impact of court experience on the likelihood of rulings. Start with undirected-dyads, then use directed dyads. Table 2. Results are consistent with: article-specific learning; disputant-specific (directed dyad- and claimaint-specific) learning; only weak evidence of general-scope learning. Note positive coeffi cient of t trend: suggests that court learning can explain declining trend in rulings (Plots 1+2). Results of dispute regressions are similar, except for the positive defendant-article specific effect: a bandwagon mechanism? Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

50 Table 2: Logit VARIABLES Undirected Dyad Directed Dyad (1) (2) VARIABLES (3) (4) DLogit_iiiikt RLogit_iiiikt DLogit_iiiikt RLogit_iiiikt CR_iiiikt 0.186*** CR_iiiikt 0.209*** 0.235** (0.0646) (0.0913) (0.0786) (0.0999) CR_n(iiii)kt * *** CR_ nnnn jjkt 0.117*** ( ) (0.0168) (0.0186) (0.0299) CR_iiiinkt *** *** CR_ii(nnnn)kt ( ) ( ) (0.0277) (0.0502) CR_n(iiii)nkt ** *** CR_jjjjkt ** ( ) ( ) (0.0941) (0.184) CR_oooooooookt *** *** ( ) (0.0193) CR_iiiinnkt *** *** ( ) ( ) CR_ nnnn jjnkt ( ) ( ) CR_ii(nnnn)nkt *** *** ( ) ( ) CR_jjjjnkt *** ( ) ( ) CR_ooooooooonkt *** ( ) ( ) t 0.198*** 1.277*** t 0.235*** 1.409*** (0.0260) (0.111) (0.0258) (0.117) t *** *** t *** *** ( ) ( ) ( ) ( ) Constant *** *** Constant *** *** (1.076) (0.723) (1.077) (0.792) Observations 439, ,560 Observations 545, ,520 Pseudo R Pseudo R iiiife Y Y iiiife Y Y kfe Y Y kfe Y Y Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 iiiife (iiiife) are undirected (directed) dyad fixed effects. kfe are article fixed effects.

51 Regressions Note lack of evidence for most narrow form of court learning. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

52 Table 2: Logit VARIABLES Undirected Dyad Directed Dyad (1) (2) VARIABLES (3) (4) DLogit_iiiikt RLogit_iiiikt DLogit_iiiikt RLogit_iiiikt CR_iiiikt 0.186*** CR_iiiikt 0.209*** 0.235** (0.0646) (0.0913) (0.0786) (0.0999) CR_n(iiii)kt * *** CR_ nnnn jjkt 0.117*** ( ) (0.0168) (0.0186) (0.0299) CR_iiiinkt *** *** CR_ii(nnnn)kt ( ) ( ) (0.0277) (0.0502) CR_n(iiii)nkt ** *** CR_jjjjkt ** ( ) ( ) (0.0941) (0.184) CR_oooooooookt *** *** ( ) (0.0193) CR_iiiinnkt *** *** ( ) ( ) CR_ nnnn jjnkt ( ) ( ) CR_ii(nnnn)nkt *** *** ( ) ( ) CR_jjjjnkt *** ( ) ( ) CR_ooooooooonkt *** ( ) ( ) t 0.198*** 1.277*** t 0.235*** 1.409*** (0.0260) (0.111) (0.0258) (0.117) t *** *** t *** *** ( ) ( ) ( ) ( ) Constant *** *** Constant *** *** (1.076) (0.723) (1.077) (0.792) Observations 439, ,560 Observations 545, ,520 Pseudo R Pseudo R iiiife Y Y iiiife Y Y kfe Y Y kfe Y Y Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 iiiife (iiiife) are undirected (directed) dyad fixed effects. kfe are article fixed effects.

53 Regressions Note lack of evidence for most narrow form of court learning. Could reflect upward bias from unobserved heterogeneity at the (undirected or directed) dyad-and-article level. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

54 Regressions Note lack of evidence for most narrow form of court learning. Could reflect upward bias from unobserved heterogeneity at the (undirected or directed) dyad-and-article level. The countries in dyad ij might have a particularly litigious relationship over article k. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

55 Regressions Note lack of evidence for most narrow form of court learning. Could reflect upward bias from unobserved heterogeneity at the (undirected or directed) dyad-and-article level. The countries in dyad ij might have a particularly litigious relationship over article k. We can check this interpretation by including ij k fixed effect. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

56 Regressions Note lack of evidence for most narrow form of court learning. Could reflect upward bias from unobserved heterogeneity at the (undirected or directed) dyad-and-article level. The countries in dyad ij might have a particularly litigious relationship over article k. We can check this interpretation by including ij k fixed effect. Identification only from within ij k variation over time; diminished ability to assess impact of variables with little within- ij k variation. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

57 Regressions Note lack of evidence for most narrow form of court learning. Could reflect upward bias from unobserved heterogeneity at the (undirected or directed) dyad-and-article level. The countries in dyad ij might have a particularly litigious relationship over article k. We can check this interpretation by including ij k fixed effect. Identification only from within ij k variation over time; diminished ability to assess impact of variables with little within- ij k variation. Introduces incidental parameter problem for rulings regression; focus on dispute regression where this problem does not arise. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

58 Regressions Note lack of evidence for most narrow form of court learning. Could reflect upward bias from unobserved heterogeneity at the (undirected or directed) dyad-and-article level. The countries in dyad ij might have a particularly litigious relationship over article k. We can check this interpretation by including ij k fixed effect. Identification only from within ij k variation over time; diminished ability to assess impact of variables with little within- ij k variation. Introduces incidental parameter problem for rulings regression; focus on dispute regression where this problem does not arise. Table 4 Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

59 Table 4: Dispute Regressions with Dyad-and-Article Fixed Effects VARIABLES Undirected Dyad VARIABLES Directed Dyad (1) (2) (3) (4) LOGIT OLS LOGIT OLS DLogit_iiiikt D_iiiikt DLogit_iiiikt D_iiiikt CR_iiiikt CR_n(iiii)kt CR_iiiinkt CR_n(iiii)nkt *** *** CR_iiiikt *** *** (0.0853) (0.0128) (0.102) (0.0115) *** 3.07e-06 CR_ nnnn jjkt ** * ( ) ( ) (0.0276) ( ) *** * CR_ii(nnnn)kt ** ( ) ( ) (0.0441) ( ) *** -1.68e-05 CR_jjjjkt ( ) (1.38e-05) (0.133) ( ) CR_oooooooookt CR_iiiinnkt CR_ nnnn jjnkt CR_ii(nnnn)nkt CR_jjjjnkt CR_ooooooooonkt e-05 (0.0108) ( ) *** *** ( ) ( ) e-06 ( ) (1.28e-05) *** -3.27e-05* ( ) (1.96e-05) ( ) ( ) ** -1.33e-05 ( ) (1.05e-05) t 0.204*** ** t 0.246*** ** (0.0262) ( ) (0.0262) ( ) t *** -1.84e-05 t *** -1.90e-05 ( ) (3.16e-05) ( ) (2.44e-05) Constant ** Constant ** ( ) ( ) Observations 26, ,584 Observations 29, ,142 (Pseudo) R (Pseudo) R iiiikfe Y Y iiiikfe Y Y CE N iiii CE N iiii Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 iiiikfe (iiiikfe) are undirected (directed) dyad-and-article fixed effects. CE are standard errors clustered by undirected ( iiii ) or directed ( iiii ) dyads

60 Alternative interpretations of empirical evidence Through the lens of our model: court learning of the particular scope and form we emphasize is the only way to account for our empirical findings. Mechanisms outside our model: In principle legal precedent could explain article-specific effects, but it is unlikely to explain disputant-specific effects. Govs learning about each other? Our data does not support this story: likelihood of rulings does not decrease with cumulative settlements. A backlog of cases coming out of GATT, hence the flurry of disputes in early WTO years? If so, we would expect cumulative settlements to have similar impact as cumulative rulings, but this is not the case. The WTO court hit a resource constraint? Hard to square that with the pattern of cumulative ruling effects we find. Bad news story? Not obvious this would predict that both rulings and disputes decline with cumulative rulings as our findings indicate. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27

61 Conclusion We explore the idea that judicial learning may contribute to explain the declining trend in WTO disputes and rulings. At the theoretical level, we study the implications of learning-by-ruling for the dynamics of trade disputes. At the empirical level, we seek to gauge the scope and strength of judicial learning. Road ahead: estimate learning-by-ruling curve and spillovers, and ideally get at welfare implications. Maggi and Staiger (Yale and Dartmouth) Trade disputes June / 27