International Signaling and Economic Sanctions

Abstract

Do economic sanctions serve international signaling purposes? A fully structural statistical model that employs a signaling game as a statistical model is used to investigate the existence of signaling effects of sanctions. Estimation results suggest that sanctions fail to work as a costly signal. The cheapness of sanctions prevents a target state from being able to distinguish a resolute sender state from a sender who is bluffing. When sanctions are imposed, a target rarely updates its initial evaluation of the sender state’s resolve, much less than when a military challenge is observed.

KEYWORDS:

• formal modeling
• sanctions
• statistical methodology

ACKNOWLEDGMENTS

For comments and suggestions, we thank the editor, anonymous reviewers, Kimberly Elliott, Mark Fey, Nehemia Geva, Hein Goemans, Mike Koch, Quan Li, Elena McLean, Kyoungwon Seo, Curt Signorino, Randy Stone, Ahmer Tarar, and participants in the Watson seminar at the University of Rochester and the Program on International Conflict and Cooperation at Texas A&M University. Appendices containing data and replication files are available upon request directly from the authors and are also available on the International Interactions Dataverse page, http://dvn.iq.harvard.edu/dvn/dv/internationalinteractions.

Notes

¹ The cost of sanctions to the sender can be conceptualized not only from an economic perspective but also from domestic political and dyadic perspectives. We interpret the costliness of sanctions for the sender more generally and also specifically in the context of a signaling game by the disutility that the sender receives in relevant outcomes in which sanctions are used.

² For the military coercion model, we rely on the data, variables, and specification used in the work of Lewis and Schultz (2005).

³ At each final node, we write the outcomes and the corresponding payoffs of the sender and target in parentheses (sender’s payoff, target’s payoff).

⁴ We assume that CP occurs only when the target complies after sanctions. Compliance with military force is coded as DL.

⁵ Our model does not capture all the strategic interactions and interesting choices throughout sanctions episodes—for example, the threat of sanctions. Our choice of sanctions-response-escalation interactions is primarily driven by our research focus, which is the signaling effects of sanctions. Therefore, our model is set up to best represent these effects as a version of a widely used signaling game. To this end, our extensive-form game structure takes the simplest form that allows for signaling.

⁶ Because the Status Quo payoff of the target () does not enter into to the equilibrium calculation, we only consider the sender’s Status Quo payoff.

⁷ Our choice of the uncertainty structure pertains to our research focus to study signaling. The QRE-based models (Signorino 2003) are not appropriate for our research purpose since they assume symmetric uncertainty. The early PBE models (Lewis and Schultz 2003; Wand 2006) are also irrelevant due to their restrictions on the stochastic components of the payoffs, which lead to downward bias in estimating the amount of belief updating.

⁸ While there are more than 183 cases, in order to control the characteristics of the sender, we select the cases where the sender is a member of the UN Security Council or the G8. Moreover, sanctions imposed during WWI (1914–1918) and WWII (1938–1945) are not included in the data.

⁹ The military coercion data used by Lewis and Schultz (2005) is ideal for the purpose of comparing the amount of signaling between economic and military coercion. Their data have the same sequential structure as the sanctions data (State A challenges or not, State B resists or not, and State A fights or not) and therefore have the same four final outcomes (SQ, CP, BD, and DL).

¹⁰ Our results remain robust when we use policy result scores instead of success scores to determine the outcome variables (Hufbauer et al. 2008).

¹¹ Our findings do not change when we exclude cases from the deadlock outcome in which the sender opts for covert actions, mounted by intelligence forces, when sanctions fail to work. The excluded cases are not dropped but coded as Back Down because their success scores are still less than 9.

¹² We rely on two sources of SQ observations, the Hufbauer et al. (2008) and EUGene data. Hufbauer et al. (2008) lists cases in which the calculated economic impacts of sanctions are zero because sanctions are not imposed but stopped at the threat stage. We code these cases as SQ (for example US sanctions against East Germany in 1960). For the remaining SQ observations, we use EUGene. The final number of Status Quo observations in the analysis is reduced to 26 due to missing values for the target states.

¹³ As in Lewis and Schultz (2005), we assume variance and covariance of the target’s payoffs to be 1 and 0 respectively. Whang (2010) shows that variance-covariance of the target’s payoffs does not have significant effects on updating.

¹⁴ As shown in the equilibrium probabilities of four outcomes, Capability Ratio and Sender Trade Dependence exert different effects on the outcome probabilities such as and .

¹⁵ The five categories of Hufbauer et al. (2008) are modest policy change, destabilization, major policy change, impairment of military potential, and disruption of military adventure. Because the inclusion of all these categories would raise the number of parameters too much, we collapse them to make a single binary variable such that the modest policy change takes the value of 0 and all other categories take the value of 1. For a detailed analysis of issue salience and success of sanctions, see Ang and Peksen (2007).

¹⁶ Issue can measure an important aspect of the costliness different types of sanction represent. In other words, it seems that there is a correlation between Issue and the type of sanctions that the sender chooses. It is reasonable to expect that the sender opts for the right type or intensity of sanctions that corresponds to the salience of the issue at stake.

¹⁷ Our findings remain robust when we use Sender Democracy in the Compliance payoffs of the sender and target. See the online appendix for the results.

¹⁸ These trade data are drawn from Oneal, Russett, and Berbaum (2003) and IMF data archives.

¹⁹ We include three different models for robustness checks in the online appendix. Our main findings remain robust to the addition and deletion of the regressors; the mean and standard deviation values of belief updating for all observations are close to 0.

²⁰ In order to explain full payoffs, it is necessary to consider their stochastic terms, s, besides their mean components. For example, is the average value of Compliance payoff for the sender, which can be observed by the opponent. When we interpret the results, we say “on average” to indicate the fact that the estimated payoffs do not contain stochastic components.

²¹ We report the 95% confidence interval for these probabilities in the online appendix.

²² While a line-by-line comparison between the two results is not possible, we use the histogram for military coercion to simply show the difference in updating between the two coercive measures.

²³ The prior beliefs, , vary from 0.205 to 0.999, that is, the prior values are not concentrated on a certain range.

²⁴ One may argue that the near-zero updating results from the fact that strong countries are more likely to initiate sanctions. Since the prior probability of escalation is already high for strong countries, there is no room for updating based on the imposition of sanctions. However, this claim is not supported by the results. When we calculate the average amount of prior probability that the sender escalates, it is not high enough to explain negligible updating: The average prior is only 0.416 for all observations and 0.410 for the cases in which sanctions are imposed.

²⁵ Bootstrapping allows us to estimate the sampling distribution of an estimator by resampling with replacement from the original sample. First, we resample with replacement from the sanctions data and make 10,031 data sets. Second, for each data set, we run the same analysis in Table 1, estimate parameters, and calculate the test statistic as specified in the appendix. Finally, from the distribution of test statistics, we construct the 95% confidence interval to determine whether it includes 0.