Private information and the commitment problem are the fundamental causes of war in the rationalist literature (Fearon Reference Fearon1995). However, despite their theoretical prominence, there are few direct empirical tests of rationalist explanations for war. Do private information and the commitment problem cause conflict, and are their effects large? Do dyads blessed with both perfect information and credible commitments enjoy perpetual peace?
I offer evidence on these questions. The rationalist literature on war has made major advances theoretically, but empirical work has been severely limited by methodological challenges. It is hard to rule out the effects of unobserved confounders. It is also unclear how private information or the commitment problem may be accurately measured with observational data. Moreover, private information always exists in international politics, and we can never know every piece of private information held by each actor (as they are private). These issues do not have clear solutions. Yet it is important for scholars and policymakers to assess the empirical significance of rationalist explanations for war.
I use randomized experiments to isolate the causal effects of private information and the commitment problem on conflict. I translate the formal theory to a laboratory setting with real people, with some simplifications to make the theory experimentally testable. Experimentation enables direct manipulation of the information and commitment environments, whereas randomization controls for all observed and unobserved factors that may contaminate the causal inference. I also conduct a case study of the decisionmaking process in a historical crisis to investigate the realism of the positive experimental findings.
I find that the commitment problem sharply increased the incidence of conflict. Despite the costs of war, decisionmakers with a commitment problem gravitated toward war. The causal effect was robust in both sequential and simultaneous-moves interactions. The effect of private information, however, was surprisingly subdued: the overall incidence of conflict in the shadow of shifting power was almost the same with or without private information.
I also design a field verification of the positive experimental results with a case study of Japan’s policy deliberations leading to the Pacific War. The case includes a shock that created an unambiguous perception of an impending power shift. The United States had imposed a total oil embargo on Japan, which sharply reduced Japan’s future relative power. Based on archival records of the secret deliberations before and after the embargo, I show how Japanese leaders quickly shifted to an active pro-war position against the United States when they were confronted by a commitment problem induced by the shadow of shifting power. The case study shows the realism of the commitment-problem mechanism in actual decisionmaking. It also detects important realistic complications unmodeled in the mechanism.
Together, the formal, experimental and historical findings point in the same direction: the commitment problem has a strong positive relationship with the decision for conflict. The mix of methods also shows how experiments can bridge two research traditions—formal models and historical studies—that typically have little contact.
Related Literature
Why do rational states fight each other despite their incentives for peaceful settlements that would avoid the costs of war? Fearon (Reference Fearon1995) highlighted two rationalist explanations: private information with the incentive to misrepresent makes it harder for states to locate a mutually acceptable deal, whereas commitment problems prevent two sides from striking a peaceful bargain in the present when one side has an incentive to renege in the future. These two explanations for war are believed to be “the full set of rationalist explanations that are both theoretically coherent and empirically plausible” (Fearon Reference Fearon1995, 380).Footnote 1 Since Fearon (Reference Fearon1995), various formal models in international relations (IR) have incorporated and extended the implications of private information and commitment problems on war (e.g., Powell Reference Powell1999; Wagner Reference Wagner2000; Filson and Werner Reference Filson and Werner2002; Kydd Reference Kydd2003; Slantchev Reference Slantchev2003; Smith and Stam Reference Smith and Stam2004; Powell Reference Powell2006; Leventoglu and Slantchev Reference Leventoglu and Slantchev2007). Both explanations are also prominent in the wider literature on conflict and cooperation in political science and economics.Footnote 2 In particular, the commitment problem appears to be pivotal in a vast variety of political phenomena, ranging from regime change and civil wars to international economic relations and the protection of human rights (e.g., Greif, Milgrom and Weingast Reference Greif, Milgrom and Weingast1994; Broz Reference Broz2002; Walter Reference Walter2002; Fearon Reference Fearon2004; Acemoglu and Robinson Reference Acemoglu and Robinson2006; Jeanne Reference Jeanne2009; Simmons Reference Simmons2009).
This study is the first to jointly evaluate Fearon’s (Reference Fearon1995) twin rationalist explanations in the laboratory. I focus on these two mechanisms as they are the most prominent in the rationalist literature on war, as well as the most well understood. The existing literature, however, offers no empirical guidance on what might happen when the two mechanisms interact: do these mechanisms mutually reinforce each other in reality? Is the effect of the commitment problem sharpened in the existence of private information—or does it not matter empirically? Using a 2×2 experimental design, I explore whether the commitment problem influences the risk of conflict in different information environments. It is useful to test the commitment-problem treatment in an idealized world of perfect information, but it is also practically relevant to understand whether and how the treatment operates in a setting with private information, which better approximates the world we live in. The 2×2 factorial design balances both considerations by testing the treatment separately under both information conditions. With the same design, we can also test whether private information has a significant effect on conflict with and without a commitment problem.
In IR, three recent game-theoretic experiments preceded this study. On the information side, Butler, Bellman and Kichiyev (Reference Butler, Bellman and Kichiyev2007) examined whether a status-quo-advantaged actor can achieve a better bargain under different information conditions, and found that the advantaged actor gets more when there is private information. Tingley and Wang (Reference Tingley and Wang2010) experimentally manipulated the cost of backing down to study how players with incomplete information update their beliefs in a crisis bargaining game. These studies, while related, did not investigate the private-information mechanism proposed in Fearon (Reference Fearon1995). On the commitment side, Tingley (Reference Tingley2011) compared bargaining behavior at two different probabilities of repeated interaction (δ), and insightfully showed how bargaining might differ when the shadow of the future is high (δ=0.7) or low (δ=0.3). Given its specific research focus, however, the subjects’ perception of a commitment problem is not fully eliminated. A subject can perceive a potential commitment problem in both experimental conditions, albeit at two different probabilities (0.3 or 0.7), as a positive probability of repeated interaction exists in both groups.
I construct the treatment condition under which players confront the commitment problem with certainty. For the control group, the key challenge is to shut down the commitment problem but keep the shadow of shifting power the same as in the treatment group. This is achieved through the removal of the proposer’s ability to change its first-period promise in the second period, which constructs a perfect-enforcement device that eliminates all possibilities of reneging.
In economics, the most relevant studies involve bargaining experiments such as the ultimatum-game experiment (e.g., Roth et al. Reference Roth, Prasnikar, Okuno-Fujiwara and Zamir1991; Cameron Reference Cameron1999; Costa-Gomes and Zauner Reference Costa-Gomes and Zauner2001). There is a sizeable literature on the effects of information asymmetry on bargaining. In general, there are more bargaining breakdowns when information is asymmetric, but the support for particular equilibrium predictions is sometimes weak (see Forsythe, Kennan and Sopher Reference Forsythe, Kennan and Sopher1991). The significance of asymmetric information appears to depend on the particularities of the information environment and the specific form of interaction. The experiments here focus specifically on private information on the cost of war in the shadow of shifting power.Footnote 3
In contrast, there are few experiments that focus on the commitment problem in bargaining games. The closest strands of literature in experimental economics are three. The first consists of experiments that focus on the communication of commitment, particularly the making of a promise and its generally positive effect on cooperation (e.g., Ellingsen and Johannesson Reference Ellingsen and Johannesson2004; Charness and Dufwenberg Reference Charness and Dufwenberg2006; Vanberg Reference Vanberg2008). These experiments are mostly silent on the commitment problem. The second strand focuses on experimental manipulations of the “shadow of the future” (e.g., Dal Bó Reference Dal Bó2005; Duffy and Ochs Reference Duffy and Ochs2009; Tingley Reference Tingley2011). These manipulations can induce a commitment problem, although different in nature from that created by non-enforcement in the shadow of shifting power.Footnote 4 The third strand of literature involves experimental manipulations of bargaining power, which involve variations in the payoffs for conflict between players that translate to variations in their bargaining strength (Binmore et al. Reference Binmore, Morgan, Snaked and Sutton1991; Fischer, Güth and Pull Reference Fischer, Güth and Pull2007; Anbarci and Feltovich Reference Anbarci and Feltovich2011). Although my experiments also involve variations in bargaining power, I focus on commitment elicitation, enforcement conditions and intertemporal power shifts, which are not emphasized in this literature.
I focus on a commitment problem in the shadow of shifting power, which scholars believe to be a potent cause of many wars in history (see Vagts Reference Vagts1956, Ch. 8; Powell Reference Powell1999; Van Evera Reference Van Evera1999, Ch. 4; Powell Reference Powell2006). To my knowledge, there is no preceding work in experimental economics or political science that eliminates the perception of a commitment problem under shifting power to isolate its causal effect, nor is there any existing study that combines and jointly evaluates Fearon’s (Reference Fearon1995) twin explanations experimentally. This study attempts to fill these gaps in the literature.
The strengths and limitations of experiments with regard to internal and external validity are well rehearsed in political science. Appendix 1 summarizes and evaluates the debate in the context of IR experiments.
Model Predictions
I combine the information problem and the commitment problem in a joint model by integrating Fearon’s (Reference Fearon1995) models.Footnote 5 The goal is to facilitate experimental testing based on the simplest possible model that combines both the information problem and the commitment problem under one structure. The results and intuitions of the joint model are similar to those in Fearon’s (Reference Fearon1995) separate models.
The joint model has three key features. First, it combines the information and commitment problems in a two-period model, instead of a single-period model for the information problem and an infinite-period model for the commitment problem, as in Fearon’s work. A two-period model offers the simplest possible structure that generates the commitment problem, but avoids the practical difficulties of implementing infinite-period bargaining with human subjects. The second feature is the use of “perfect enforcement.” The enforcement condition is defined such that Player A cannot make any change to its Stage 1 demand in Stage 2. Hence, the commitment problem is completely shut down in the enforcement condition. The use of a perfect-enforcement device eliminates any noise that might arise from the expectation of even a minuscule possibility of reneging. Third, I black box the war technology. Players end up with an absolute payoff if war is chosen, without having to deal with expected value calculations involving the probabilities of winning wars. With these changes, the experiment becomes cleaner in design as well as simpler for the players to play.
Consider a two-stage game in which two players bargain over a prize with the value v. In Stage 1, Player A makes a demand x 1∈[0, v]. Player B observes x 1 and decides to agree or to fight. If B fights, the game ends with payoff (w 1A−c A, w 1B−c B), with w 1A=w 1B, where w 1i and c i are Player i’s war payoff in Stage 1 and cost of war, respectively. If B agrees, the game enters Stage 2 in which A confirms its demand x 2∈[0, v]. B observes x 2 and decides to agree or to fight. If B agrees to x 2, the game ends with payoff (x 2, v−x 2). If B fights, the game ends with payoff (w 2A−c A, w 2B−c B), with v=w 1A+w 1B=w 2A+w 2B and w 2A>w 1A>w 2B>c. Thus, A and B have equal bargaining power in Stage 1, but the bargaining power shifts in A’s favor in Stage 2.
In the experiment, v is fixed at 10 and c i is fixed at 2. For the war payoffs, w 1A=w 1B=5, w 2A=7 and w 2B=3. These parameter values are chosen to minimize noise in the experiment. Payoffs are set within the integer range of [0, 10] to simplify the mathematics for subjects and to reduce the likelihood of calculation error. Parameter c i is set at 2 as it is the highest integer value by which all payoffs in the model remain in positive integer domain, given w 2A=7 and w 2B=3. The values of w 2A and w 2B are chosen to achieve equidistance from the midpoint. Note that other combinations of integer values will complicate the experiment by either breaking the equidistance or generating potential payoffs that are 0 or negative given the requirements of the model.Footnote 6
I analyze the incentive structures implemented in the experiment condition by condition. The formal analysis is described in Appendix 2. It yields four predictions for experimental testing:
Prediction 1: War will be avoided if there is enforcement in the public-information condition.
Prediction 2: War will occur with certainty if there is no enforcement in the public-information condition.
Prediction 3 There is always a positive risk of war in the private-information condition with enforcement.
Prediction 4: War will occur with certainty in the private-information condition without enforcement.
Experimental Design
I apply a 2×2 factorial design with controlled variation in information and enforcement. The first treatment variable captures the difference in information conditions; the second captures the difference in enforcement conditions.
Random Assignment
Subjects are randomly divided into public-information and private-information groups. Subjects in each group are then randomly allocated to either the enforcement condition or the no-enforcement condition. In each round, subjects are randomly assigned as decisionmakers for either Country A or B and they play with a randomly assigned opponent within the same information/enforcement conditions. To control for reciprocity effects, the matching of anonymous players is subject to a “stranger-matching” rule that no two players be matched with each other twice in the same scenario.
Experimental Conditions
-
∙ Enforcement: Player A cannot change its Stage 1 demand (x 1) in Stage 2. The computer will directly equate x 1=x 2. This is perfect enforcement and it eliminates the commitment problem.
-
∙ No enforcement: Player A can change its Stage 1 demand in Stage 2. Hence, the commitment problem remains.
-
∙ Public information: The cost of war c A=c B=2 is known to all players. Each player is told: “You know your Opponent’s cost of war. Your Opponent knows your cost of war.”
-
∙ Private information: Each player knows its own cost of war (c i =2), but does not know its opponent’s cost of war.Footnote 7 Each player is told: “You do NOT know your Opponent’s cost of war. Your Opponent does NOT know your cost of war.”Footnote 8 Players are told that the cost will be assigned once in Round 1 and thereafter fixed for all subsequent rounds in the game.
There are two important details to highlight for the private-information condition. First, players are explicitly told that the computer will assign the cost once and only once in Round 1, and thereafter the cost will be fixed for all rounds in the game.Footnote 9 This is important to support the belief that costs are drawn from a discrete uniform distribution. With the present design, we can hold constant the effect of the cost of war, while making it natural for subjects to play with the same cost in different rounds. Second, in the private-information condition, there is an incentive for misrepresentation but no channels for communication. This is to ensure a clean experiment consistent with the model in the earlier section and with Fearon’s (Reference Fearon1995) original model, in which there is an incentive to misrepresent but no opportunity to communicate. As Fearon (Reference Fearon1995) emphasized, the private-information argument is not driven by the lack of opportunities to communicate.Footnote 10 Future research may explore the effectiveness of different types of communication channels in dealing with private information under shifting power.
Experimental Sequence
Participants begin the experiment by reading the instructions on their computer screens. They are placed in the role of a decisionmaker bargaining with another country for a valuable prize worth 10 points. I frame the situation as one involving crisis bargaining and war to enhance the relevance of the experiment, as my research question deals directly with rationalist explanations for war. The war framing also helps to strengthen the inducement of utilities by reducing the potential other-regarding behavior widely documented in many bargaining experiments.Footnote 11 This is important as rationalist explanations for war do not assume other-regarding behavior.
Participants are informed that they will be randomly divided into dyads. They will be randomly assigned as a decisionmaker for either Country A or B in each round, and hence their paired opponent will also change at random after every round. The two-stage bargaining game (Figure 1) is carefully explained to the subjects, with the payoff difference between Stages 1 and 2 highlighted. In Stage 1, both countries are equally powerful: if they fight a war, each country will seize 50 percent of the prize (5 points) for itself. In Stage 2, Country A becomes more powerful, and if they fight a war, A will seize 70 percent of the prize (7 points). As war is costly, each country will lose points given the cost of war. The cost of war is known to all subjects in the public-information condition, whereas subjects in the private-information condition know their own cost but not their opponent’s cost. If a peaceful bargain is reached, both sides avoid the costs of war.
Crisis bargaining game
Subjects stay in the same information condition for Rounds 1–10. For the enforcement manipulation, I use a crossover design with a sudden switch in the enforcement condition in Round 6. Hence, half the subject pool plays five rounds with enforcement before switching to five rounds with no enforcement; the other half plays in the reverse sequence. The crossover design is used often in clinical trials but rarely in political science. It offers two benefits in this study: the collection of both within-subjects and between-subjects data on the enforcement variable; and a replication test of the results with a sudden exchange of treatments. Findings that survive the crossover test are more robust than those that do not.
To summarize the experimental sequence:
-
∙ Main experiment: Participants play ten rounds. They play five rounds with enforcement and five without. They stay in the same information treatment as randomly assigned in Round 1.
-
∙ Extensions: After Round 10, participants play five more rounds under the public-information condition with an option for simultaneous moves.Footnote 12 This is an extension to test whether the effect of enforcement is generalizable and robust beyond pure sequential bargaining. After Round 15, participants play one round of the same game in the no-enforcement condition with timer treatments that induce differential time pressures (see Part 4 of Appendix 5 for details).
-
∙ End: The session ends with a risk-aversion game in the final round that measures individual risk preferences.
The full experimental instructions are reproduced in Appendix 3.
Implementation
70 students from Harvard University were recruited through the Harvard Decision Science Laboratory. The experiment was programmed on z-Tree, which allowed subjects to interact anonymously through computers (Fischbacher Reference Fischbacher2007). The experiment was conducted at Harvard University in February 2011. There were three sessions in total. Each session had 22–24 subjects. Each subject participated in only one session.
Special attention was paid to ensure that subjects fully understood the game. The instructions were written in neutral language, with questions at the end to check subject understanding. No practice rounds were implemented before the experiment. Participants were paid solely based on their performance in the game. The computer paid US $0.50 for each point earned in nine randomly chosen rounds. On average, each participant earned US $19.87 in the 1-hour session.
Results
Result 1: The existence of a commitment problem causes a large increase in the incidence of war.
The incidence of war (war outcomes as a percentage of all outcomes) is 59 percent in the no-enforcement condition (with commitment problem) compared with 20 percent in the enforcement condition (without commitment problem) (two-tailed test of proportion, p<0.0001, n=347). Table 1 shows the incidence of war across the four experimental conditions in the first ten rounds. Table 2 shows the average number of wars per player in each condition. In the public-information condition, the average number of wars is more than thrice as high in the no-enforcement group than in the enforcement group. In the private-information condition, the average in the no-enforcement group is more than twice as high as the average in the enforcement group. Both differences are statistically significant at p<0.001 (two-tailed t-test: n=72 for public-information condition; n=62 for private-information condition).Footnote 13 Player average is used as the unit of observation as round outcomes may not be independent. Round outcomes will be analyzed with different model specifications later in this section.
Incidence of War Across Conditions
Note: maximum possible value for the incidence of war is 1.00 (100 percent).
Average Number of Wars Per Player
Note: maximum possible value is 5.00 for each player.
Together, the data show strong tendencies toward war when there is a commitment problem in both public and private-information conditions (Predictions 2 and 4), and toward a peaceful bargain when there is enforcement in the public-information condition (Prediction 1).
Next we turn to the crossover test. If the causal effects are robust, we should expect an increase in the incidence of war when players switch from the enforcement condition to no-enforcement, as well as a decrease when they switch from no-enforcement to enforcement.
This is indeed what we see: the sudden introduction of a commitment problem caused a jump in the incidence of war, whereas the war incidence slumped when the commitment problem was removed. Figure 2 shows a sharp difference across Rounds 5 and 6 when a sudden exchange of treatments was introduced. The crossover test reversed the direction of the treatment effects and revealed a striking symmetry in the effects measured: the incidence of war increased from 22 percent (Round 5) to 67 percent (Round 6) when a commitment problem was introduced, and fell from 63 to 25 percent when the commitment problem was removed (two-tailed tests of proportion: p=0.0073, n=36 dyadic outcomes; p=0.0325, n=32, respectively). Splitting the sample by information conditions, introducing a commitment problem in the public-information condition increased the incidence of war from 33 percent (Round 5) to 67 percent (Round 6), whereas the removal of the commitment problem caused it to fall from 56 to 11 percent.Footnote 14 In the private-information condition, war incidence increased from 11 to 67 percent when the commitment problem was introduced, and fell from 71 to 43 percent when it was removed.Footnote 15 Figure 2 also shows that the average incidence of war in the final rounds (Rounds 8–10) converged to 6 percent with an extremely low range of 0–12 percent when there is enforcement. By contrast, without enforcement, the incidence of war in the final rounds is 65 percent with a high range of 50–78 percent. Again the same story emerges: with enforcement, the dominant behavior revolved around the peaceful bargain; without enforcement, the dominant behavior was war.
Incidence of war in Rounds 1–10
Although the observations agree with the model predictions, Result 1 does not tell us whether the predictions were observationally correct because of the equilibria-generating calculations underlying the model, or because of other systematic factors that were not captured by the model. Hence, I test two observable implications derived from the strategic calculations behind Predictions 1, 2 and 4. If the hypothesized mechanisms are correct, we should expect that in the no-enforcement condition (1) players with a relative power shift against their favor will choose to fight in Stage 1 and (2) players with a relative power shift in their favor will renege on their Stage 1 agreement in Stage 2. These are indeed what happened.
Result 2: In the no-enforcement condition, wars in Stage 1 comprised 89 percent of all wars fought, whereas players with a power shift in their favor reneged on their promises 93 percent of the time.
Figure 3 shows the percentage of players who reneged on their promises across ten rounds in the no-enforcement condition.Footnote 16 A player is coded as a “reneger” if he or she made a Stage 2 offer lower than his or her Stage 1 offer. Figure 4 shows that wars in Stage 1 comprised 89 percent of all wars fought in the no-enforcement condition.Footnote 17 Taken together, the data suggest that the observed behaviors supporting Predictions 1, 2 and 4 were very likely motivated by the strategic calculations underlying the model.Footnote 18
Result 3: The existence of private information has no significant effect on the incidence of war in the shadow of shifting power.
Percentage of reneged offers (no enforcement)
Percentage of wars in Stage 1 (no enforcement)
Pooling the data across enforcement conditions, the incidence of war in the private-information condition (39 percent) is almost identical to that in the public-information condition (40 percent) (two-tailed test of proportion, p=0.8374, n=347). From Table 1 earlier, war incidence is slightly higher under private information than under public information in the enforcement condition, and lower in the no-enforcement condition. Both differences are insignificant (two-tailed test of proportion: p=0.3033, n=173; p=0.2503, n=174, respectively).Footnote 19 Figure 5 also indicates that differences in information conditions have no systematic association with the incidence of war across ten rounds.
Incidence of war (by enforcement conditions)
For robustness check, I fit logit models to analyze the round outcomes. Each model uses one dyadic observation per round and session, with robust standard errors corrected for clustering at the subject level. The dependent variable is the decision for war in each round. The treatment variables are binary variables with the value 1 if there is enforcement (for the enforcement treatment) or public information (for the information treatment), and 0 if otherwise. I use different model specifications with and without round and session fixed effects, and with and without an interaction term for the treatment variables, a control for the size of Player A’s initial offer and a control for individual risk preference. Risk preference is measured on a summed score based on the decisions made in the risk-aversion game (see Appendix 3) at the end of the experiment: the higher the score, the greater the individual willingness to take risk.
The information treatment has no significant relationship with the decision for war across all model specifications (p>0.30). In contrast, the enforcement coefficient is negative and highly significant in all specifications (p<0.001). Meanwhile, the negative and significant coefficient (p<0.01) for the initial-offer variable suggests that Player B is less likely to decide for war when Player A makes a higher initial offer. The interaction term for the treatment variables is insignificant (p≥0.20) across all models. Table 3 shows the logit estimates with different combinations of control variables.
Logit Estimates of Determinants for the Decision for War (Rounds 1–10)
Note: In parentheses are heteroskedasticity-robust standard errors corrected for clustering at the subject level. Round and session dummies are used to control for round and session fixed effects.
***p≤0.001, **p≤0.01, *p≤0.05.
Result 3 suggests the relative non-importance of private information in the shadow of shifting power. The pertinent question is why. One possibility is that when players were placed in the two-stage game with shifting power, their attention was captivated by the power shift at the expense of the information asymmetry. As the impending change in power gripped the minds of the players, their mental resources were diverted away from thinking about what the opponent’s cost of war might be, given the limited working memory capacity of the human mind and the ease of attentional capture by a more dynamic and immediate stimulus (Yantis and Jonides Reference Yantis and Jonides1984; Fukuda and Vogel Reference Fukuda and Vogel2009). This neural tendency might have thus weakened the effect of private information in the players’ calculations. Of course, one can also interpret the result without appealing to human neuroscience. One may observe that in the game-theoretic literature the commitment problem often produces a sharp deterministic prediction of conflict, whereas private information does not. Indeed, Prediction 3 in my model is probabilistic, whereas Predictions 1, 2 and 4 are deterministic. Although the model clearly predicts war with certainty in the case of the commitment problem, it predicts only a positive risk of war in the private-information condition with enforcement. This is not unique to my model. Comparable private-information models of war, including that in Fearon (Reference Fearon1995), generate similar probabilistic predictions. Based on differences in the nature of their model predictions, one might expect the effect of private information to be less decisive than the effect of the commitment problem. Hence, all we conclude is that the behavioral data provide no clear support for the causal impact of private information, based on private information on non-catastrophic costs of war in the shadow of shifting power.Footnote 20 This result applies only to this particular form of private information. Future work may investigate if a similar result holds for other types of private information.
Result 4: The treatment effect of the commitment problem is robust with interactions that allow for simultaneous moves.
After Round 10, I modified the game to allow for simultaneous moves. This extension study serves two purposes. Methodologically, it provides a stress test by moving beyond pure sequential bargaining. Substantively, it is motivated by the fact that, in reality, states can always choose to forego the bargaining process if they believe that diplomacy is futile. The modified game is more complex but provides a closer approximation to reality. The game is similar to the public-information variant of the original game, but with the option for simultaneous moves: Player A decides its demand, whereas B decides at the same time whether to wait for the demand (and subsequently accept or reject it) or to wage a war. The model predictions remain unchanged.Footnote 21 Subjects were randomly assigned to either the enforcement or the no-enforcement condition. All subjects played five rounds under public information (i.e., everyone knows that c i =2).Footnote 22
The incidence of war in the no-enforcement condition is 73 percent, which far exceeds the incidence of 8 percent in the enforcement condition. The difference in the average incidence of war across players is statistically significant at p<0.001 (two-tailed t-test, n=70). Regression models analyzing the round outcomes confirm the same result. I fit logit models with the war decision as the dependent variable and the enforcement variable as the regressor. Different model specifications are used, with and without round and session fixed effects, and with and without a control for the size of the initial offer and a control for risk preference. Table A1 in Appendix 4 displays the logit estimates conditioned on different sets of controls, showing that enforcement has a strong negative relationship with war outcomes across all specifications (p<0.001).
Table 4 summarizes the incidence of war in each session. Figure 6 visualizes the large and persistent difference in the incidence of war across the two conditions.
Incidence of war (Rounds 11–15)
Incidence of War (Rounds 11–15)
Note: the maximum value for the incidence of war is 1.00 (100 percent). In Rounds 11–15, subjects played the crisis bargaining game with an option for simultaneous moves.
Taking this result with the earlier observations, we can conclude with confidence that the commitment problem has a strong positive effect on the decision for conflict. The treatment effect is large even in an experiment without large stakes. In real-world crises, the stakes are dramatically increased. All else being equal, it is plausible that decisionmakers will be more payoff sensitive and will try to behave more (rather than less) rationally in a high-stake environment (see Appendix 1). If this is the case, the gravitation to the rational equilibrium—and war—may be even more pronounced.
Case Study Design
A laboratory experiment allows the researcher to construct the ceteris paribus condition often assumed in a scientific theory but rarely achievable in the real world, and is thus a powerful tool for testing mechanisms. Yet, the scientific virtue of the lab also raises a reasonable concern on whether the identified effect—and the mechanism that predicts it—actually matters in the real world when a ceteris paribus setting does not exist. For instance, a mechanism may be valid but its causal effect sensitive or weak. The effect may be significant in a tightly controlled environment, but canceled out, reconditioned or interacted away by the circumstances found in a naturalistic setting. In this case, the experimental result may be theoretically validating but also practically misleading.
This concern may be addressed with field experiments (Fisher Reference Fisher1926; Campbell Reference Campbell1998; Levitt and List Reference Levitt and List2009; Gerber and Green Reference Gerber and Green2012). In international politics—where field experiments are practically and ethically difficult—a feasible alternative is to seek field verifications of the experimental results. Specifically, in identifying causes of war—a rare event that can often only be retrospectively investigated—the feasible alternative may be to study the relevant historical processes with the most reliable and high-resolution evidence that we can find. If a decision-based mechanism is being investigated, evidence on the historical decisionmaking process should be gathered, and consistency between the predicted and revealed behaviors carefully evaluated. This is the logic behind the field verification design described in this section.
In the last section, I have shown that the commitment problem increases the risks of conflict in the laboratory. To what extent does the commitment problem shape decisionmaking in a real-world crisis?Footnote 23 I focus on three questions: (1) Does the commitment-problem logic operate in a real-world crisis triggered by an impending power shift? (2) How does the logic affect the decision process leading to war? (3) Does the case reveal important omissions in our theoretical model? Questions (1) and (2) help to direct the field verification exercise, whereas (3) extracts new information from the field verification.
I address the three questions with a case study of Japan’s deliberations on the eve of the Pacific War, which involves a sharp power shift. I analyze the meeting records from the four imperial conferences preceding the Pacific War and the 38th–74th liaison conferences between the imperial conferences (translated in Ike Reference Ike1967). Owing to space constraints, methodological considerations and case analysis are contained in Appendix 5.
The case analysis can be encapsulated in three key points.Footnote 24 First, Japanese leaders did not focus on war with the United States before the latter’s imposition of a total oil embargo starting in August 1941. That focus quickly changed after the US embargo. The evidence shows a clear difference in the strategic calculations articulated at the imperial conferences before and after the embargo. In the July conference, the focus was on plans to establish the Greater East Asia Co-Prosperity Sphere and settle the war in China expeditiously, and the question of war with the Soviet Union. Japan’s belligerence did not center on the United States; the conference heard no specific plan to attack the United States; and there was no clear statement of the commitment-problem logic that would be repeatedly applied against the United States in later conferences.
Second, the oil embargo triggered Japanese perceptions that its relative power would deteriorate, and that the United States cannot credibly commit not to exploit Japan in the future. The embargo left Japan with no doubt about its impending decline as Japan depended on foreign oil imports to sustain its war machine. The reference materials for the September imperial conference emphasized: “We need not repeat that at present oil is the weak point of our Empire’s national strength and fighting power…. As time passes, our capacity to carry on war will decline, and our Empire will become powerless militarily” (Ike Reference Ike1967, 155). Stemming the relative decline was crucial, because “[e]ven if we should make concessions to the United States by giving up part of our national policy for the sake of a temporary peace, the United States, its military position strengthened, is sure to demand more and more concessions on our part; and ultimately our Empire will have to lie prostrate at the feet of the United States” (Ike Reference Ike1967, 152). Hence, the September conference was warned that “military force should be used promptly if there is no prospect of diplomatic success. It is expected that the United States and Great Britain will try to delay us with diplomatic negotiations. We must be careful not to be inveigled into this trap” (Ike Reference Ike1967, 155).
Third, the oil embargo created a strong time pressure that made Japanese leaders push for a quick decision for war. This led to a tight deadline for war and a truncated timeframe for negotiations. As Foreign Minister Togo concluded at the November imperial conference, “negotiations with the United States are very much restricted by the time element; consequently, to our regret, there is little room left for diplomatic maneuvering …. The prospects of achieving an amicable settlement in the negotiations are, to our deepest regret, dim” (Ike Reference Ike1967, 214).
Taken together, three key findings emerge from the archival evidence: (1) Japan’s decision for war involved strategic calculations that repeatedly applied the commitment-problem logic. (2) The perceived speed of shift in power influenced Japan’s calculations under the commitment problem. (3) The speed of power shift made Japan truncate the bargaining timeframe, making it more difficult to achieve a diplomatic settlement. Finding 1 demonstrates the realism of the commitment-problem explanation. Findings 2 and 3 highlight a historically important factor—the “truncation effect” arising from a rapid shift in power—that was omitted in my model and experiment.
These findings may provide a useful additional interpretation on why Japanese leaders rushed into the gamble that was Pearl Harbor. Although Findings 1–3 are closely consistent with the archival evidence, traditional historical narratives of the Pacific War often do not develop the theoretical connections explicated in these findings. Meanwhile, historical analyses of the war by IR scholars emphasize the role of shifting power (Sagan Reference Sagan1989; Van Evera Reference Van Evera1999; Trachtenberg Reference Trachtenberg2006), but they do not explicate it within the framework of the commitment problem or highlight how it was the perceived speed of shift in power that made Japan push for a tight deadline, which truncated the bargaining timeframe, which in turn dampened the prospect of a diplomatic settlement (truncation effect).
It is important to be cautious about the causal mechanism suggested by the case findings, as constraints in the historical data do not allow the truncation effect to be cleanly identified. To probe further, I replicated the mechanism in a laboratory setting to test its significance. The experiment offers suggestive evidence that the truncation effect exists. The experimental design and results are summarized under Finding 4 in Appendix 5.
Conclusion
Private information and the commitment problem are central to the rationalist theory of war. The causal evidence, however, is scarce. I use an experiment to show that the commitment problem causes a large increase in conflict incidence in environments both with and without private information. The crossover design embedded in the experiment shows that a sudden introduction of a commitment problem causes a sharp rise in conflict incidence, whereas its sudden removal causes a sharp fall. Furthermore, in the absence of enforcement, players with a power shift in their favor renege on their agreements 90 percent of the time. The results also show that despite the theoretical prominence of the private-information argument, private information on the cost of war has no significant effect on conflict in the shadow of shifting power. This raises the general question of whether and how the private-information mechanism exerts causal impact on war decisions, an important question that remains for further investigation. It would also be useful for future research to explore the efficacy of the private-information mechanism empirically in different forms and across different conditions.
The experiment shows that the commitment problem has a decisive effect on the incidence of conflict. A case study focused on the commitment problem investigates the historical realism of the experimental results. It shows how Japan’s policy deliberations leading to the Pacific War were shaped by calculations based on the commitment-problem logic. The case study also reveals a mechanism for the Pacific War that remains untheorized in the literature. The mechanism highlights how the perceived speed of shift in power made Japanese leaders push for a tight deadline, which truncated the bargaining timeframe and made diplomatic settlement more difficult. A supplementary experimental probe offers suggestive evidence for the mechanism.
Convergent results from the formal, experimental and historical investigations strengthen the causal inference: The commitment problem has a strong and positive relationship with conflict incidence. Despite the costs, decisionmakers with a commitment problem gravitate toward conflict. Thus, impending power shifts can be very dangerous. In general, power shifts are perilous when there is no enforcement structure to suppress the incentive to renege on prior agreements. The existence of this incentive dissolves the willingness to trust and makes costly conflict more likely. I have focused specifically on interstate conflict, but the strategic form of the commitment problem isolated here is extremely general. As a consequence, the findings may have potential implications for other forms of conflict—ranging from civil conflict to contentious politics—that involve the commitment problem in a similar form.
A key policy implication is that enforcement can calm the commitment problem and reduce the risk of costly conflict. In IR—where there is no world government to enforce bargains—the lack of external enforcement is often seen as the root of international anarchic violence. Yet, external enforcement can also occur without a world government, insofar as there is a third-party actor willing and able to play the role. Several cases in history have shown how external intervention can help to calm civil wars and bring about peace in violent situations (e.g., Regan Reference Regan2002; Walter Reference Walter2002). Nevertheless, the nature of international enforcement and its multiple possible forms have yet to be fully theorized and rigorously tested in the existing literature.Footnote 25 It is hoped that future research can bridge the gap and shed light on how commitment problems may be cured in theory and in practice.
A key methodological implication is the importance of crossing methodological lines by designing an integrated hybrid of different methods to investigate the same phenomenon. This study uses the experimental method to bridge two major research traditions—formal models and historical studies—that usually have little real contact. Methods can be combined in different ways, and this study shows the utility of a particular permutation (models–experiments–cases) that has received very little attention in political science. First, formal models clarify the strategic logic and generate equilibrium predictions. Next, experiments isolate the causal effects and detect systematic deviations from the predictions. Finally, case studies verify the substantive significance of the experimental results and highlight important realistic omissions in the modeled mechanism. Theoretical insights are compared and cumulated as we iterate between formal models, randomized experiments and historical cases.
This article demonstrates how this particular permutation may strengthen causal inference in the study of war and peace.Footnote 26 Formal models are often criticized for their deviations from observed behavior, laboratory experiments for their lack of realism and case studies for their causal confounds. A combinatorial hybrid of the three can apply the strength of one to compensate for the weakness of another, and thus make a more robust contribution to our scientific knowledge. Such multi-method integration, however, remains scarce in political science. It is hoped that more would recognize the scientific value of interrogating the same phenomenon with different and independent methods, and offer encouragement to those who cross the methodological lines that divide our discipline.
