The Effectiveness of International Trade Boycotts

The Effectiveness of International Trade Boycotts Kilian Heilmann First Draft: June 6, 2014 Abstract I estimate the impact of international consumer boycotts by studying two specific boycotts that I claim are exogenous to unobserved shocks: the boycott of Danish goods by Muslim countries following the Muhammad Comic crisis in 2005/2006 and the Chinese boycott of Japanese goods in response to the Senkaku/Diaoyu Island conflict in 2012. Results from the synthetic control method suggest strong heterogeneity in the response to boycott calls between countries with an average trade disruption of about 15% in the case of the Muslim countries and 4% in the case of China. In both cases, this is only a minor share of overall exports over this period with 0.4% and 0.8% respectively. Product-level results show that the boycott is most effective for consumer goods and especially highly-branded signature export goods, and at most temporary for intermediates and capital goods. An event study on Japanese stock market returns suggest that the Chinese boycott did not significantly depress stock values of explicitly boycotted Japanese firms. JEL classification: F14 1 Introduction International trade boycotts as a special form of conflict between countries are not a new phenomenon. They have been used throughout history to punish or coerce a specific behavior. Early examples of international consumer boycotts include the repeated boycotts of Japanese goods by China throughout the 1930s in response to the Japanese invasion (for a very early account see Lauterpacht, 1933) and the worldwide boycott movement in protest of South Africa s apartheid system in the late 1950s. The 21st century has seen a US-American boycott of French products during the Iraq War in 2003, the boycott of Danish goods by Muslim countries in reaction to the Muhammad Comic Crisis of Department of Economics, University of California San Diego. Email: kheilman@ucsd.edu. I would like to thank Gordon Hanson, Marc Muendler, Eli Berman, Lawrence Broz, and participants of the UCSD Third Year Paper sequence for helpful comments. 1

2006, and a Chinese consumer boycott against Japan in 2012 after the conflict about the Senkaku/Diaoyu Islands. These events share the common characteristic that they are not motivated by economic rationale, for example by inferior product quality or containing hazardous parts, but by trade-unrelated political events. In contrast to the more frequent boycotts against specific firms, such as the boycott against Shell in 1995, they are directed against entire countries. These boycotts can therefore be seen as policy means to carry out international conflicts. They seem to become an option when other means of coercion, such as war or the severing of diplomatic relationship appear to be infeasible. They also tend to be exercised by the people of a country without or with only few official government support, often when the people think that their government is unable or unwilling to address the conflict. Yet the usefulness of this means strongly depends on whether these boycotts are effective. Their intent to punish countries can fail in many dimensions. At first, boycotts intent to hurt the conflict partner by not importing goods from it and thus depriving it from its export revenues and thus gains from trade. Yet if the boycotted country s exporters can easily redirect their sales to domestic or other foreign markets, the potential economic loss might be relatively small. Secondly, even if disrupted exports do hurt the exporting country significantly, boycotts are a costly tool, since the boycotting country is also giving up its gains from trade by refusing to import. This is even more true in a world characterized by increasing international integration of production, where trade is not primarily in final goods anymore, but where the share of processing trade is rising. If production of the boycotting country crucially depends on imports from the boycotted country, this will raise the costs of the boycott, it might render it an incredible threat not being carried out at all. Furthermore, trade boycotts rely on collective action that can be difficult to organize and therefore fail to materialize. The aim of this paper is to evaluate the effectiveness of international trade boycotts and to quantify their impact on international trade relationships. To do so, I make use of two instances of consumer boycotts that I claim are exogenous to unobserved traderelated shocks and can be used to identify the effect of these boycotts on exports from the boycotted to the boycotting country. These two events are the boycott of Danish goods after the Muhammad Comic crisis in 2005/2006 and the Chinese boycott of Japanese goods in the aftermath of the Senkaku/Diaoyu Island conflict in 2012. Both these conflicts were sparked by random events carried out by private people, came along rather unexpected, and are thus unrelated to previous trade relationships. To measure the impact of boycotts on the exports of the boycotted country, I use the synthetic control group method to construct counterfactual import values and compare them to the actual imports. The results suggest that there is a strong heterogeneity in the 2

response between the boycotting country, with an average one-year reduction in imports of about 15.1% and 4.3% of total trade in the Muhammad boycott case and the Senkaku case respectively. Product-level analysis shows that the impact is concentrated in consumer goods with only minor effects for intermediates and capital goods, being consistent with the notion that international trade boycotts are mainly carried out by consumers and not by firms. While the estimated disruption in imports from the boycotted country can be large, the fraction as total exports of the boycotted country is very low in both boycott cases (0.4% for Denmark and 0.8% for Japan). This suggests that even though an individual firm of the boycotted country might be hit hard, the overall effect on the export sector is negligible, rendering the punishment effect as mostly ineffective. The paper is organized as follows: Section II reviews the existing literature, section III describes the event studies used and section IV presents the data. Section V then explains the methodology while Section VI presents the empirical results. Section VII concludes. 2 Literature Review Trade boycotts and especially consumer boycotts have received extensive treatment not only in economics, but also in the fields of law and psychology. One important branch of this literature is concerned with the mechanisms of consumer boycotts and the main focus of this research is to explain the individual motivation for participating in a consumer boycott, e.g. Friedman (1999). John and Klein (2003) study consumer boycotts as instances of collective action that are inherently faced with the small-agent problem, i.e. the success of the boycott depends on a mass of participants, but every individual s impact and motivation to join in is low. To explain that consumer boycotts do regularly happen, they propose a variety of other reasons to participate, such as psychological motivations like guilt and self-esteem or simply an exaggerated sense of one own s effectiveness. This literature serves as the theoretical foundation of any quantitative analysis of consumer boycotts. The question whether there is a significant impact of consumer boycotts has received more attention in the last few years. Yet thorough quantitative studies of international consumer boycotts against an entire country have to my best knowledge been restricted to a single boycott, the US-American boycott of French products in the aftermath of France s opposition to the invasion of Iraq in 2003. Michaels and Zhi (2010) estimate that trade deteriorated by about 9% in 2002-03 when France s favorability rating in the US fell sharply. Pandya and Venkatesan (2013), using supermarket scanner data, find that brands that are perceived as being French lose market shares in weeks with high media attention of the boycott. They estimate the implied costs of this boycott to be similar to the costs of an average product recall. For the same period, Clerides et al. (2012) find a significant but 3

short-lived drop in sales of US-American soft drinks in the Middle East, but cannot find a similar effect on other goods. Both these studies are based on local sales and do not investigate the effect on imports. Davis and Meunier (2011) study the quarterly trade relationships between the US and France as well as between China and Japan for the years 1990-2006, thus including the boycott of French goods. They do not find any significant link between negative events involving these countries and the level of goods exchanged, but find that trade as well as foreign direct investment continued to grow sharply in the period studied. Besides the narrow focus on explicitly announced boycotts, there is a new literature studying the relationship between other political conflicts and international economic relations. Fuchs and Klann (2013) for example study countries trade with China if they officially receives the Dalai Lama. China perceives any formal relations with the Tibetan spiritual leader as an interference into internal political affairs and threatens countries that do so with a reduction of trade. The authors test whether China carries out its threat in order to deter those countries from receiving the Dalai Lama. They find a significant negative short-term effect on trade volumes and confirm that, even though the effect dies out quickly after one year, countries are willing to use trade as a tool to enforce their political will. The main shortcoming of Fuchs and Klann s paper, however, is the focus on annual trade levels and the resulting short time span to test their hypothesis. Fisman et al (2014) and Govella and Newland (2010) study the effects of Sino-Japanese conflicts on the stock market value of Japanese firms using an event study approach. They find that stocks of Japanese companies with a high share of sales to China lose value compared to companies with a low exposure to China. This paper tries to fill in a gap in the understanding of consumer boycotts by linking the literature on boycotts with the literature studying the impact of conflicts on trade. The novelty of this paper that distinguishes it from previous studies is the use of monthly data allowing for a more thorough analysis of short and long-term effects of international consumer boycotts on international trade. 3 Background 3.1 Muhammad Cartoon Crisis On September 30, 2005 the Danish newspaper Jyllands-Posten published a series of cartoons depicting Islamic prophet Muhammad in an arguably unfavorable manner, the 4

most striking one showing him with a bomb in his turban. 1 This caused a huge outcry by Muslims in Denmark as not only the depiction of the prophet is forbidden in several branches of Islam but also because they felt that the comics equated Muslims to terrorists. Even though Danish Muslims protested the publication from the very beginning, it was not until early 2006 that the controversy became an international conflict. After the comics had been reprinted in several Arabic newspapers, violent protests sparked in many Middle Eastern countries, leading the ambassadors of several Muslim countries to demand an official apology by the Danish government and prosecution of the cartoon artists. The Danish government however refused any involvement into the crisis and argued that the private newspaper s actions are protected by freedom of speech. The months of January and February 2006 saw further escalation of the conflict with Western embassies being attacked, leaving several dozen people dead. With the Danish government still refusing any official apology, religious leaders in Saudi Arabia called for a boycott of Danish on January 26, 2006 in order to hurt Denmark economically, and published a boycott list of Danish firms. 2 Soon other countries joined the boycott including for example Indonesia and Syria The French supermarket chain Carrefour preemptively removed Danish goods from its shelves and several Danish food producers such as Arla Foods reported huge losses in the Middle East. 3 At the same time, a counter-boycott campaign called Buy Danish was called for, but it remains unclear whether this campaign gained enough media attention to have any large scale effects. 4 The scandal about the Muhammad cartoons eventually lost public attention and the protests calmed down, though several incidents in later years were linked to the cartoons. In 2008 and 2010 attempts to kill the creator of the most controversial of the cartoons were made, but could be prevented by police. It is therefore unclear how long the conflict and the boycott continued to affect trade relationships between the Middle East and Denmark. 3.2 Senkaku/Diaoyu Islands Conflict The Senkaku (in Japanese) or Diaoyu (in Chinese) Islands are a small group of islets set in the East China Sea approximately 170 km North-East of Taiwan. None of the five islands in the archipelago is inhabited and with a maximum size of 4.3 square kilometers the islets are not suited for permanent settlement. In the aftermath of the First Sino-Japanese War (1984-85) and the subsequent invasion of Taiwan, Japan began to survey the islands and claimed it as Her territory. After the Treaty of San Francisco formally established 1 For a detailed time line of the events, see Jensen (2008). 2 Examples of these lists can be found on http://shariahway.com/boycott/index.htm 3 http://www.nytimes.com/2006/01/31/international/middleeast/31danish.html? r=0. 4 http://www.foxnews.com/story/2006/02/16/muslim-boycotts-hurt-danish-firms/. 5

peace after World War II, Japan ceded all Her claims to Taiwan and the nearby inhabited Okinawa islands came under US control. When the Okinawa islands were returned to Japanese control in 1972, Japan tacitly took control of the Senkaku islands as well and remains a military presence on the islands until today. In 1968, possible oil reserves were found in the area surrounding the Diaoyu/Senkaku islands leading to claims of both Mainland China and the Republic of China (Taiwan) to the islets. Japan rejected those claims denying that China had ever exerted control over the islands and the territorial conflict remained unsolved. It was not until the 2000s when several incidents brought the Senkaku/Diaoyu conflict back to public attention. Between 2006 and 2011 several activist groups from Mainland China, Taiwan and Hong Kong arrived at the islands to proclaim Chinese sovereignty, just to be found to be expelled by the Japanese navy immediately. The most severe incident happened in 2010 when a Chinese fishing boat was destroyed by Japanese forces and its captain held captive for three days. While these events definitely worsened Japanese-Chinese relationships, the conflict only escalated after Japan announced to purchase the islands from its private owner in August 2012 and de facto established sovereignty over the archipelago. This led to furious anti- Japanese protests in several Chinese cities that later turned violent. Japanese businesses in China were attacked and protesters called for a boycott of Japanese goods. Japanese- Chinese relations deteriorated drastically when further naval standoffs near the disputed islands occurred, leading to worldwide fears over a military conflict of the two biggest powers in East Asia. While the conflict has calmed down and lost media attention, the major issue is still unresolved and remains a major problem in Japanese-Chinese relations. A major objection against the use of the conflict as a natural experiment is that the conflict could have been intentionally triggered by the Chinese government in order to distort public attention from other political questions. However, this would only be a problem if the intention was in response to economic issues that also affect trade. I argue that the conflict evolved rather accidentally through the actions of several private activists, but acknowledge the possibility of a political intention behind the dispute. Yet if the conflict was planned and sparked by the Chinese government, I believe that it was only in response to the upcoming change in the political leadership of the party and state in late 2012. This change was planned long in advance and is a regular feature of the Chinese political system and therefore exogenous to any economic shocks that also affect the trade relations between Japan and China. 6

4 Data 4.1 Data Sources This paper makes use of two distinct datasets that both provide monthly trade data, but differ significantly in certain dimensions. For the boycott of Danish goods due to the Muhammad Crisis, I use data from the online portal of Statistics Denmark. 5 This dataset covers Danish export values in local Danish krona (DKK) to a multitude of countries and other political entities at monthly frequency for total trade volumes as well as for both the two-digit (66 different industries) and five-digit SITC classification (3316 different products). The time series run from the late 1980s to currently August 2013, yet I only use data from January 1994 onwards to abstract from effects caused by the transition of Eastern Europe after the collapse of the Soviet Union. To avoid inconsistencies in the data I drop observations with trading partners that are extremely small, remote or experienced any changes in territory during the period of study. The data for the Senkaku/Diaoyu conflict comes from the newly created Monthly Comtrade dataset that, in addition to the standard Comtrade data, reports trade flows at monthly frequency for all Harmonized System (HS) product codes (about 5300 different products). The very fine disaggregation of the data is however offset by other problems as the Monthly Comtrade database is still in development and has not passed the beta testing phase yet. This results in very limited availability of trade data depending strongly on the reporting country. For example, the Japanese-reported time series for monthly trade volumes with China runs from January 2010 to January 2014 only, while the mirror data reported by China only covers the period July 2011 to September 2012, that is it stops exactly at the height of the Senkaku/Diaoyu Island crisis. This prohibits verification of the Japanese export data by Chinese import data. 6 Data prior to 2010 is at the moment only available at annual frequency. In addition to export values, both datasets provide information on net weight for a small amount of the product codes at the most disaggregated level (SITC5 and HS6 respectively). I employ a variety of variables to control for confounding effects on trade. These data include measures of economic production and geographic distance which are good predictors of trade levels in gravity regressions. For GDP levels at yearly frequency (in current US dollars), I use data from the World Bank indicators. For data on population-weighted 5 http://www.dst.dk 6 Import values are usually considered to be more accurate than export figures since state organizations are thought to be more interested in accurately accounting tariff income-generating imports. 7

bilateral distance between trading partners, I use the gravity dataset from CEPII. 7 Data on the Muslim population for each country is provided by the Pew Research Center. 8 4.2 Descriptive Statistics 4.2.1 Denmark At first it is to notice that exports to the Muslim world as a share of total Danish exports is relatively small. The 34 countries with a Muslim population of 75% or more for which trade data was available account for not more than 2.66% of Danish exports to all trading partners in 2004. For example, Denmark exported more to Finland (with a share of 2.95%) than to all these countries together. Even the biggest Muslim trading partner, Saudi Arabia, accounted for less than half a percent of Danish exports in 2004. Examining export values from Denmark to the boycotting countries shows that monthly trade data is characterized by high volatility, seasonal patterns, and possibly changing time trends. It is not uncommon that Danish exports to these countries increase by a multitude over one month or that trade completely collapses even in the pre-boycott period. In Uzbekistan for example the total trade value from March to April 2005 surged by a massive 7000% from 20,000 to 1,423,000 DKK, only to drop by 95% to 59,000 DKK the following month. The strong month-to-month swings are more prominent for the smaller export partners, but are still significant for the three major Danish importers. Table 1 summarizes the descriptive statistics of the time series for the three treatment countries with the largest imports from Denmark: Saudi Arabia, Turkey, and the United Arab Emirates. Some of the smallest Muslim countries trade with Denmark is characterized by long spells of no imports with an occasional spike. Since estimating an impact of the boycott for these countries is both difficult to implement and economically uninteresting (for example, even a total collapse of trade in the case of Kyrgyztan would reduce Danish exports by only 0.004%), I omit countries from the further analysis that show zero trade flows in the data and only include them when I use aggregates of the whole Muslim worlds. The strong volatility might hint towards important seasonality patterns that complicate the quantitative analysis. I test for monthly seasonality patterns by a simple F-test in a linear regression with indicator variables for each month and find evidence for seasonal patterns only for a handful smaller countries (Algeria, Bangladesh, Djibouti). 7 www.cepii.fr/anglaisgraph/bdd/gravity.htm. 8 http://www.pewforum.org/2009/10/07/mapping-the-global-muslim-population/. 8

Table 1: Descriptive Statistics Country Saudi Arabia Turkey UAE Aggregate Mean 176,710 205,965 129,765 1,036,268 Standard Deviation 31,738 87,048 37,468 188,399 Std Dev as Mean 18.0% 42.3% 28.9% 18.2% Minimum 100,143 77,810 82,677 710,530 Maximum 272,301 462,941 422,072 1,596,562 Min % Change -33.0% -60.2% -54.4% -29.1% Max % Change 67.8% 92.6% 194.7% 46.9% F-test p-value 0.37 0.32 0.68 N/A Statistics over the pre-boycott period October 2000 to September 2005. Seasonality p-value is the p-value of a F-test testing for joint significance of monthly indicator variables in a linear time series regression. 4.2.2 China-Japan Unlike the Danish-Muslim boycott where all the boycotting countries take up only a small share of total exports, for Japan the People s Republic of China (from now on referred simply as China ) is in fact the largest export partner when considering the preboycott period from January 2000 to August 2012. China alone accounts for 19.23% of all Japanese exports. The Special Administrative Regions of Hong Kong and Macao, even though technically part of the PRC, but with a very different political and economic system, and Taiwan 9 report separate trade statistics. Including the trade with those entities, the total percentage of exports to the Chinese-speaking world amounts to 30.8% For the Japanese-Chinese trade data, the month-to-month fluctuations are lower but can still reach percentage changes of more than 30% in both directions. In general, Taiwan s imports from Japan are the most stable while Macao s is the most volatile, most likely due to its small share of Japanese exports. The time series is marked by a stark drop in March 2011, the effect of the devastating Japanese earthquake and tsunami that resulted in more than 50,000 deaths. Seasonality might be an issue especially in the winter months in which trade appears to slow down and the F-test testing for the joint significance of the monthly indicator variables suggest seasonal patterns for all entities except Macao. 9 For political reasons, monthly trade data for Taiwan is not officially available in the Comtrade Monthly dataset, but can be inferred from the country code 490, Other Asia, nes. 9

Table 2: Descriptive Statistics Country China Taiwan Hong Kong Macao Aggregate Mean 12,800,000 4,175,000 3,502,000 19,964 20,496,964 Standard Deviation 1,378,000 363,400 362,700 3,568 1,949,993 Std Dev as Mean 10.8% 8.7% 10.4% 17.9% 9.5% Minimum 9,626,000 3,090,000 2,674,000 13,871 15,570,000 Maximum 15,420,000 4,770,000 4,149,000 29,157 24,350,000 Min % Change -30.4% -25.2% -29.4% -32.9% -28.2% Max % Change 32.6% 22.9% 39.2% 56.8% 29.0% Seasonality (p-value) 0.01 0.00 0.00 0.28 N/A Statistics over the pre-boycott period January 2010 to August 2012. Seasonality p-value is the p-value of a F-test testing for joint significance of monthly indicator variables in a linear time series regression. 5 Methodology 5.1 Synthetic Control Group To identify the effect of a trade boycott it is necessary to construct counterfactual export levels that one can compare the actual export figures to. Simply comparing the trade levels of the boycott period to the pre-boycott months will not account for any idiosyncratic shocks to exports that would have been present even without the boycott. Then we might attribute shocks to the boycott which have nothing to do with it. Using a difference-indifference approach to compare boycotting countries to non-boycotting countries requires choosing an appropriate control group. In my dataset with few boycotting countries and a large pool of non-boycotting countries, choosing the correct one is not a trivial task. The results may strongly depend on the choice of the control group and may not be robust to alternative choices. Trade theory gives us little guidance in determining the control group. Even though the gravity equation shows that bilateral distance and partner GDP are strong predictors for overall trade volumes, this relationship describes long-time averages and is less applicable to monthly trade data. In the short-term, there may be many more confounding factors that are unobserved. To construct counterfactual trade values for the treated countries, I therefore follow the synthetic control group method first used in Abadie and Gardeazabal (2003) and later further developed in Abadie et al. (2010, 2014). The synthetic control group method follows a pragmatic data-driven approach to choose the right control group by creating an artificial control unit using a weighted average of all the available control units. The weights are chosen such that the synthetic control 10

group resembles the actual treatment unit in both the outcome variable as well as well in as any known explanatory characteristics in the pre-treatment period. An estimate for the treatment effect can then be calculated by the difference in the actual outcome of the treatment unit and the synthetic control unit in the post-treatment period. The following exposition follows closely the one in Abadie et al. (2014) Suppose that there are J + 1 units in a balanced dataset with T observations which consists of one treatment unit and J potential control units. Denote the number of pretreatment periods as T 0 and the first period with the treatment in place as T 0 +1. Without loss of generality, we can define the first unit to be the treated unit and specify the units 2... J + 1 to be the control units. If there are more treated units one can simply remove them from the data and repeat the same procedure for these units. Assume that log export values Y j,t of the boycotted country to its trading partners j are given by the following factor model log Y j,t = δ t + θ t X j + λ t µ j + β t boycott j,t + ɛ j,t (1) where δ t is a time trend common to all export partners, θ t and λ t are (1 r) and (1 F ) vectors of common factors, X j and µ j are and (r 1) and (F 1) vectors of factor loadings, and ɛ j,t is an iid error term with mean zero that captures idiosyncratic shocks. The parameters of interest are {β t } T t=t 1 that measure the dynamic impact of a boycott that is captured by the dummy variable boycott j,t. The difference between X j and µ j is that the former one is known to the econometrician and includes observable trade determinants like GDP and bilateral distance while the latter factors are unobserved and might include variables like industry composition and consumer preferences. The ideal experiment to estimate the impact of the boycott would be to compare the outcome of the boycotting country to the outcome of a non-boycotting country that has the same factor loadings X j and µ j. However, this is infeasible for two reasons: Firstly, most likely no such unit exists and secondly, µ j is unobserved. While the former problem could be resolved by a regression-based approach, the unknown factor loadings cause the regression to be necessarily misspecified. Instead, the synthetic control group method constructs the counter-factual as a weighted average of the availabl control countries. Denote as Yj,t I the counterfactual value of the exports in case the boycott had not happened. Let T 0 the period of the intervention. For all the countries in the control group, I assume that Y I j,t = Y j,t for j = 2... J and all periods t = 1... T. This implicitly assumes that there are no substitutions of exports from boycott to non-boycott countries and is not realistic. If these substitutions happen, the control group will be affected positively and the treatment effect will be upward biased. However, the estimated treatment effect can 11

still be interpreted as an upper bound of the true causal effect of the boycott on exports. The goal is to construct the counterfactual export levels for the boycott country so that we can obtain the estimator β t = Y 1,t Y1,t I for the treatment effect at time t. A synthetic control group is defined by a set of J weights w j for j = 2... J + 1 that determine a weighted average of the control units. The ideal synthetic control group would match both the factors X and µ of the treatment unit, yet this is impossible since µ is unobserved. However, Abadie et al. (2010) show that under mild regularity conditions, the synthetic control group can only match µ if it also matches a long period of pre-treatment outcome variables Y j,t. This motivates to choose the weights to minimize both the deviation in the known characteristics X as well as the pre-treatment oucomes Y j,t. Assume that there exist weights w j with J+1 j=2 w j = 1 and 0 < w j < 1 j = 2... J + 1 such that J+1 wj X j = X 1 (2) j=2 J+1 wj Y j,t = Y 1,t t = 1... T 0, (3) j=2 that is the synthetic control group resembles both the pre-treatment outcomes as well as the known explanatory variables of the treatment unit perfectly. The restrictions on w j ensure that no extrapolation outside the support of the data takes place. 10 The model in (1) implies that for the synthetic control group it holds that J+1 J+1 J+1 J+1 wj Y j,t = δ t + θ t wj X j + λ t wj µ j + wj ɛ j,t j=2 j=2 j=2 j=2 and that the difference between the actual treatment unit and the synthetic control group in the pre-treatment period t = 1... T 0 is J+1 Y 1,t wj Y j,t j=2 } {{ } = 0 J+1 J+1 J+1 = θ t X 1 wj X j +λ t µ 1 wj µ j + wj (ɛ 1,t ɛ j,t ) j=2 } {{ } = 0 j=2 j=2 10 This is the crucial difference to a regression-based construction of the counterfactual. Abadie et al. (2014) show that a regression-based counterfactual can be interpreted as a weighted average of the controls with weights that also sum up to one, but allow for negative values or values larger than one. 12

Rearranging, summing over all time periods, and dividing by T 0 yields J+1 µ 1 wj µ j 1 T 0 J+1 λ t = T 0 j=2 t=1 w 1 T 0 j T j=2 0 t=1 (ɛ 1,t ɛ j,t ) Note that as the number of pre-treatment periods T 0 becomes large, the right hand side of the equation goes to zero. As long as 1 T T t=1 λ t 0 (that is the average effect of the unobserved factors is not equal to zero over time), this implies that the difference in the unobserved characteristics goes to zero as well. This suggests to use J+1 j=2 w j Y j,t as the counter-factual and subsequently calculate the treatment effect as J+1 β t = Y 1,t wj Y j,t t > T 0 j=2 In practice however, one will not be able to find weights such that equations (2) and (3) hold exactly. This is the case if the characteristics of the treatment country are not in the convex hull of the characteristics of the control countries and thus cannot be replicated with the restrictions on w j. In this case the weights are chosen such that the equations hold approximately. Formally, define Z j = (Y j,1, Y j,2... Y j,t0, X j ) as the column vector that stacks all export values for the pre-treatment period 1... T 0 and the known characteristics of country j. Similarly, define the matrix Z C = [Z 2 Z 3... Z J+1 ] that collects these column vectors for all the potential control countries. The (J 1) vector W is then the solution to the following minimization problem W = arg min Z 1 Z C W = arg min (Z1 Z C W ) V (Z 1 Z C W ) (4) W W for a given weighting matrix V. The choice of V allows to assign different importance to the explanatory variables or specific pre-treatment outcomes. In order to reduce the deviation between treatment and synthetic control group the factors with the largest predictive power should be given the highest relative weights. If the number of pre-treatment periods T 0 is small and the number of potential controls J is large, the characteristics of the treatment unit will be mechanically replicated by a combination of the control units. To circumvent this problem of overfitting, Abadie et al. (2014) suggest restricting the pool of potential controls to countries that have similar characteristics as the treatment country to break the spurious fit. Restricting the controls 13

to units that are close in terms of the characteristics also helps to reduce potential interpolation bias. 5.2 An Idea for Inference In practice, the fit between the treatment group and the synthetic control group in the pre-boycott period will not be perfect, but subject to the idiosyncratic shocks captured in the error term ɛ j,t. This will bring randomness into the estimate of the treatment effect β t. The exact distribution of the estimate depends on the unobserved parameter vector λ t and can therefore not be computed. A pragmatic approach to evaluate the significance in the parameter estimates is to compare them to the prediction error in the pre-boycott period. The intuition behind this is the notion that if the synthetic control group fits the actual treatment unit poorly before the boycott happened, this would undermine the confidence in the estimate of the treatment effect. If the fit between the actual treatment country and its synthetic control however is close in the pre-boycott period, we can be more confident in assuming that any divergence after the treatment is actually caused by the boycott and not due to unrelated shocks. Assume that the divergence ν t between the treatment and the control group follows an iid distribution with mean zero and constant variance σ 2 ν J+1 ν t = Y 1,t wj Y j,t i.i.d. (0, σν) 2 j=2 The variance σν 2 can be estimated by the sample variance on the pre-treatment differences between treatment and synthetic control group. Taking the square root, this is the root mean squared prediction error for the pre-treatment period and can be interpreted as the expected deviation between the actual treatment group and the synthetic control group due to country-specific shocks. I assume that this error is not affected by the boycott, so that ν t is also present in the estimate of the treatment effect as well: J+1 β t = Y 1,t wj Y j,t + ν t = β t + ν t t > T 0 j=2 The estimate then follows the distribution β t (β t, σ 2 ) which allows me to compute asymptotic standard errors for the estimate. These standard errors should be interpreted 14

as a lower bound on the true standard errors as they ignore potential serial correlation in the true deviations, but they can still provide an idea of the reliability of the estimates. To complement the analysis, I also perform placebo tests that traditionally have been used in the context of synthetic control groups. There are two dimensions where a placebo test can detect wrongful inference: Within a single time series, a random assignment of a treatment time should not break the close fit between actual and synthetic control group and should not produce large estimates of the treatment effect. If both series deviate even though there is no boycott, then this should warn us that the synthetic control group is merely picking up unrelated idiosyncratic effects. Furthermore, we can estimate the same treatment effect for the control countries. If these countries are indeed unaffected by the boycott, the synthetic control group method should not find large treatment effects. If however the control countries seem to be negatively affected by the boycott, this would hint to mis-specification in the model and would greatly undermine our confidence in the method. 6 Empirical Implementation 6.1 Muslim Boycott of Danish Goods 6.1.1 Country-level Results Since the publication of the Muhammad Comics was offensive to all Muslims, there are many potential countries that could be induced to boycott Danish goods. I could find evidence of a boycott call in many countries, but even if a country did not announce a boycott, it could still have participated. I therefore allow for silent boycotts and use the share of Muslim population in a country as a guidance to choose potential treatment countries. I assign all countries that have a share of Muslim population of the total population of more than 75% into the treatment group. Conversely, I assign countries for which this share is less than 10% into the control group. I drop all other countries to avoid a contamination of the control group. This leaves me with 34 countries in the treatment group and 100 countries in the control group. (See Table 15 and Figure 10 in the Appendix.) Since the comics were published on the last day of September 2005 and a same-day effect is unlikely, I define October 2005 to be the first treatment period in the sample. This is considerably earlier than the official announcement of the consumer boycott in January 2006, but allows for undeclared boycotts as an immediate reaction to the insult. I separately fit the pre-treatment trend for each potential boycott country by minimizing the 15

prediction error for all monthly export levels five years prior to October 2005 as well as for the averages of the GDP level for this time period and the population weighted distance. Since the number of potential control units is large, I restrict the potential control countries to countries that are close in both distance and GDP in the month prior to the boycott to avoid overfitting. In specific, I allow the GDP to differ by 75% in both directions and distance to deviate by 5,000 km. This avoids that the relatively small economies of the Middle East are replicated by large and distant countries like Japan and the US. I experiment with the time dimension and change the pre-treatment period to three years as well as seven years respectively and the results are generally robust. The goodness of fit between the different boycott countries and their synthetic control in general varies between 0.6 and 0.7. Table 3 summarizes the pre-treatment congruence by employing two different measures of fit, the simple correlation coefficient and the root mean squared prediction error. Table 3: Pre-Treatment Fit Three Years Five Years Seven Years Country Corr. RMSPE Corr. RMSPE Corr. RMSPE Saudi Arabia 0.872 14322 0.793 17005 0.701 21677 Turkey 0.651 25010 0.674 29909 0.674 30385 United Arab Emirates 0.416 15290 0.473 16495 0.684 14975 Average 0.726 6793 0.654 7707 0.619 7846 Corr: Correlation Coefficient RMSPE: Root Mean Squared Prediction Error (in thousands of DKK) To calculate the value of the foregone trade, I simply add up the treatment effects of all treatment countries for each month and calculate the percentage loss as a share of total trade levels. Table 4 shows the estimated percentage reduction for all countries for a period a three, twelve, and 24 months and the associated prediction errors. The results suggest a strong heterogeneity between the different Muslim countries. Some larger export partners like Egypt, Libya, and Saudi Arabia see a strong and persistent negative effect, while some even show a positive reaction, which is likely due to the very small and irregular trade flows. Most notably, the second and third largest Danish trading partners Turkey and UAE show no reaction to the boycott. Figure 1 shows the geographical distribution of the estimated treatment effect. Adding up the estimated treatment effects for all countries, I can calculate the total disruption of trade due to the boycott. I estimate the total costs to be about 0.51 billion DKK after three months, 2.86 billion DKK after twelve months, and 4.28 billion DKK after two years. This corresponds to a 14.8%, 20.9%, and 15.1% reduction in overall trade to the Muslim countries. The US-Dollar equivalents after taking into account fluctuations of the 16

Table 4: Estimated Percentage Reduction in Trade by Country Country 3m 12m 24m Country 3m 12m 24m Albania -32.7% -29.3% -18.6% Maldives 325.1% 158.3% 144.8% (39.4%) (21.0%) (15.6%) (177.6%) (105.4%) (71.0%) Algeria -44.2% -17.7% -6.8% Mali -52.9% -70.1% -50.0% (23.7%) (12.4%) (9.3%) (33.8%) (14.9%) (13.4%) Azerbaijan -49.8% -54.4% -39.1% Morocco -49.1% -19.8% -12.3% (25.5%) (15.2%) (11.7%) (42.8%) (22.2%) (14.7%) Bahrain -16.9% -0.2% 10.6% Mauritania -12.4% 274.0% 63.2% (13.0%) (6.6%) (4.9%) (145.1%) (72.6%) (43.0%) Bangladesh 25.5% 34.4% 32.2% Oman 3.5% -29.7% -9.3% (33.5%) (16.4%) (11.1%) (17.8%) (8.5%) (5.9%) Djibouti -42.0% -34.2% -46.8% Pakistan 131.2% 48.2% 40.5% (97.5%) (55.3%) (32.4%) (22.8%) (12.4%) (9.3%) Egypt -43.6% -29.3% -20.5% Qatar -2.2% 2.1% 49.0% (17.4%) (10.2%) (6.3%) (37.4%) (19.8%) (14.2%) UAE -1.3% -7.3% -1.4% Saudi Arabia -31.6% -41.9% -34.5% (7.7%) (3.7%) (2.6%) (5.4%) (2.7%) (1.8%) Gambia 131.3% 85.2% 114.5% Senegal -61.5% 83.2% 45.3% (76.3%) (46.1%) (33.1%) (70.8%) (42.7%) (27.1%) Guinea -7.9% 34.1% 67.4% Syria -14.8% -29.0% -14.5% (54.0%) (33.7%) (27.6%) (16.5%) (8.4%) (6.0%) Indonesia 31.6% 5.1% -1.2% Tunisia -20.2% -11.4% 24.3% (11.5%) (5.3%) (3.6%) (30.3%) (15.0%) (10.4%) Iran -6.6% -34.5% -31.2% Turkey -10.4% -3.6% 0.0% (20.2%) (10.0%) (7.2%) (8.8%) (4.4%) (2.9%) Jordan -32.1% -38.4% -37.4% Uzbekistan 2.8% 15.8% 1.8% (16.9%) (8.5%) (5.5%) (151.8%) (87.7%) (49.8%) Kuwait -20.2% -46.4% -48.9% Yemen -27.4% -40.7% -36.4% (14.4%) (6.9%) (4.6%) (16.4%) (8.5%) (5.3%) Libya -61.7% -80.3% -56.1% Total -14.8% -20.9% -15.1% (25.6%) (12.8%) (9.3%) (3.6%) (1.8%) (1.3%) Prediction errors in parentheses. exchange rate are, 198 million USD after three months, 444 million USD after one year, and 758 million USD after two years. Figure 2 plots the estimated cumulative trade loss for all treatment countries combined. The graph reveals that the boycott constantly impacts Danish imports at about the same rate and dies out after about 16 months when the curve becomes flatter. After that time, the realized imports increase at the same rate as the implied counterfactuals, so that we can conclude that the boycott is a one time reduction in trade that does not change the long-term growth rates of trade. It is however permanent in the sense that it appears that there is no catch-up effect. The dynamic response is robust to changes in the pre-treatment period. While the percentage loss for all the Muslim countries combined is sizable, this loss is marginal when compared to the total exports of Denmark. Over the period from October 2005 to September 2007, Danish exports with all its trading partners summed to 1.08 17

Figure 1: Spatial Distribution of Treatment Effect Legend Estimated Treatment Effect less than -55% -40 to -55% -20 to -40% -15 to -20% -5 to -15% -5 to 5% 5 to 10% 10 to 30% 30-50% 50-99% Controls Excluded No Data >100% trillion DKK (185 billion USD). The implied overall disruption of trade caused by the boycott is then only 0.4% of all Danish exports during this period. While the boycott might have hit individual Danish companies hard, the effect on the total Danish export sector is negligible. 6.1.2 Robustness Checks and Placebo Tests To check whether the results depend strongly on the parametrization of the synthetic control group, I deploy a variety of robustness checks. Since most countries in the analysis have such a small trade volume with Denmark that even a total boycott would barely have an effect, I restrict the robustness checks to the three largest export partners Saudi Arabia, Turkey, and the United Arab Emirates. I assign placebo treatment times and estimate the trade disruption for these false boycott instances. I estimate the cumulative treatment effect over six months for the 30 months 18

Figure 2: Cumulative Treatment Effect over Time to Muslim Countries (Aggregate) 0.5 0 x 106 Estimated Cumulative Treatment 95% Confidence Interval 1 1.5 in billion DKK 2 2.5 3 3.5 4 4.5 Boycott officially announced 5 0 3 6 9 12 15 18 21 24 Months since Boycott preceding the publication of the comics. 25 of these placebo treatment times are not related to the boycott, but the five 6-month estimates prior to the actual treatment month will contain at least one of the actual treatment months respectively. Figure 11 shows the distribution of estimated treatment effects. Some of the placebo treatments do create negative treatment effects, but in general are of smaller magnitude and not as persistent as the estimated trade disruption of the actual treatment. For Saudi Arabia and UAE, the countries that showed significant trade disruption, all six-month estimates including the actual treatment month are negative and large. For Turkey, the estimate of the actual treatment is still negative, but at a much smaller scale especially compared to previous large negative and positive effects. These random fluctuations are in line with Turkey s estimated, non-significant effect of almost 0%. In addition to the synthetic control group methodology, I also estimate a differencein-difference model of imports from Denmark Y j,t by country j at time t. I use the share of Muslim population as a continuous treatment and control for GDP as well as bilateral distance, the usual gravity model explanatory variables, and also include a time trend The regression equation is given by log Y j,t = α+β 1 muslim j +β 2 post t +β 3 muslim j post t +β 4 GDP t +β 5 dist j +β 6 t+ɛ j,t. (5) 19

In addition, I include country-fixed effects to control for heterogeneity in the industry composition. This specification does not allow to identify the effect of distance or the share of muslim population separately, but the two time-invariant coefficients will be merged into the fixed effect. The results in Table 5 indicate that the treatment effect is negative, but it only becomes statistically significant if country-fixed effects are included. GDP and distance have the expected positive and negative sign respectively. The large negative coefficient on the share of Muslim population indicates that the treatment countries in general import less from Denmark. My preferred fixed effects estimate suggest that the average decline in exports due to the boycott is 12.2%, which is in line with the estimates of the synthetic control group. Table 5: Fixed Effects Results Log Imports from Denmark (1) (2) GDP 0.000 0.000 (0.000) (0.000) Distance Muslim -0.000310 (0.00000364) -1.268 (0.0412) Post 0.115 0.279 (0.0357) (0.0138) Post Muslim -0.00365-0.122 (0.0774) (0.0291) Constant 11.04 8.145 (0.0521) (0.0839) Fixed Effects no yes Time Trend yes yes N 30500 30500 adj. R 2 0.319 0.908 Standard errors in parentheses. p < 0.05, p < 0.01, p < 0.001 6.1.3 Industry-level Results To analyze the potentially heterogeneous effect on different product groups, I break up the analysis into three main product types: Consumer goods, intermediate goods, capital 20

goods. 11 This allows us to gain some insight about who is the main driver behind the boycott. If the boycott is mainly consumer-driven, we should expect a higher trade disruption in consumer goods as compared to non-consumer goods. If the boycott however is state-organized, we would expect a similar disruption in all goods. I repeat the methodology for each product type by classifying all SITC-5 product code into either consumer, intermediate, capital goods and others. Where available, I use the Broad Economic Categories (BEC) classification developed by the UN Statistics Department. 12 SITC-5 codes that are not available in the BEC were coded by my own judgment in close concordance with the logic of the BEC classification. The conversion table can be found in the online appendix. Figure 3: Realized and Counterfactuals by Class Log Consumer Exports 20 19.5 19 Realized Consumer Synthetic Consumer Boycott Comics 18.5 60 40 20 0 20 40 Log Intermediate Exports 20.5 20 19.5 Realized Intermediate Synthetic Intermediate Boycott Comics 19 60 40 20 0 20 40 Log Capital Exports 20.5 20 19.5 19 18.5 Realized Capital Synthetic Capital Boycott Comics 18 60 40 20 0 20 40 Months since Boycott Log Other Exports 18 17.5 17 16.5 16 15.5 15 Realized Others Synthetic Others Boycott 14.5 Comics 14 60 40 20 0 20 40 Months since Boycott 11 The remaining goods are classified in a category called Others which only makes up a tiny share of overall trade and contains unclassified goods such as coin and charitable donations. 12 http://unstats.un.org/unsd/trade/bec%20classification.htm. 21

Table 6: Treatment Effect by Product Type Period Consumer Intermediate Capital Others 3 Month 1.5% -9.0% -13.4% -36.1% (8.7%) (4.5%) (11.5%) (11.7%) 12 Month -27.5% -10.9% -12.0% 52.8% (4.3%) (2.3%) (6.3%) (4.4%) 24 Month -24.8% -1.7% -1.7% 43.0% (3.0%) (1.6%) (4.4%) (3.1%) Prediction Error in parentheses. I first combine the Danish exports to the treatment countries and then separate them by product type. Table 3 shows the realized and counterfactual log Danish exports to of each classification. Consistent with the boycott being consumer-driven, I see the largest relative decline in consumer goods imports with long-term reductions in this category of 27.5% and 24.8% after one and two years respectively. This suggests that the publication of the comics itself did not cause a major consumer reaction, but only after the official boycott announcement did imports from Denmark decline. For non-consumer goods, the reaction is less strong and in many cases not statistically significant. Danish capital goods exports to the Muslim world seems to decline marginally in the short and medium run, but the large prediction errors render this result statistically insignificant. Over two years, this decline is reduced to less than 2%, so that any eventual boycott effect in the short run is leveled off by a catch-up effect. For intermediate goods, we do see a significant reduction in imports from Denmark of about 9.0% and 10.9% after 3 and 12 months respectively that is,similar to the capital goods case, reduced to 1.7% after two years. This is inconsistent with the idea of a pure consumer boycott and could be explained by nationalistic sentiment of business owners or official trade restrictions such as complicating the processing of imports at custom offices. 6.2 Chinese Boycott of Japanese Goods 6.2.1 Country-Level Results For the Senkaku Island Crisis case, I identify four political entities that are potentially affected by the boycott announcement: The People s Republic of China, its Special Administrative Regions (SAR) Hong Kong and Macao, as well as the Republic of China (Taiwan). All these entities are quintessentially Chinese and I found evidence for sovereignty claims to the Diaoyu Islands for all of them except Macao. 22

Table 7: Top Japanese Export Destinations Rank Trade Partner Value Share 1 China 409,581 19.23% 2 United States of America 342,378 16.08% 3 Republic of Korea 169,204 7.95% 4 Taiwan 133,596 6.27% 5 Hong Kong SAR 112,078 5.26% 6 Thailand 100,137 4.70% 7 Singapore 68,629 3.22% 8 Germany 57,839 2.72% 80 Macao SAR 638 0.03% Sum of China 655,895 30.80% Sum of Exports from Japan to its trade partners for the period 1/2010 to 8/2012 in million USD. The nature of Japanese trade with Mainland China creates problems with the synthetic control group method. As discussed above, Mainland China is not only Japan s largest export partner over the pre-treatment period but it is also geographically close. It is thus at the end of the distribution of both outcome as well as explaining factors and it will be impossible to replicate its imports from Japan with a weighted average given the strong restrictions on the weights given in equations (2) and (3). The other treatment units, Taiwan and Hong Kong, have a similar level of imports from Japan with shares of 6.2% and 5.2% respectively. Although these shares are much smaller than the share of Mainland China, there are still only two control countries that have more imports (USA and Korea). I therefore relax the conditions of the weights to be in the unit interval and instead allow for arbitrary weights. To avoid overfitting, I restrict the number of control units to countries that have a similar GDP. 13 In general, a small number of countries is able to replicate the Chinese trade patterns rather well. Table 8 shows the composition of the control group and reports statistics of the closeness of fit. Figure 4 shows the realized and counterfactual exports from Japan to China on a log scale. The strong decline in realized exports after the boycott is announced is easily visible and trade levels even fell below those after the devastating earthquake in 2011. Yet one can also see that Chinese imports from Japan were on a downward trend and that only a portion of the decline can be attributed to the boycott, as the counterfactual trade figures implied by the synthetic control group decline as well. The negative effect of the boycott compared 13 In specific, the replicating country s GDP should have at least 20% of GDP of the treatment country and it should not exceed it by the factor 1.8. While arguably arbitrary, this creates control pools of around 10 control countries. 23

Figure 4: Realized and Counterfactual Trade Levels (Mainland China) 23.5 Realized China Synthetic China 23.4 23.3 Log Japanese Exports 23.2 23.1 23 22.9 Boycott 30 25 20 15 10 5 0 3 6 9 12 Months since Boycott to the prediction error is low in the first two months and only starts to be realized in the trade data in the third boycott month, resulting in a 5% reduction of Japanese imports. The boycott effect then accelerates and reaches its maximum after five months so that trade is reduced by 9.5% after half a year. After that the boycott effect seems to fade out and imports from Japan might even be catching up as the realized imports are above the counterfactual values. Given the high prediction errors however, this catch-up effect is highly speculative. The total reduction in Japanese exports within one year of the boycott amounts to 4.3% and is equivalent to 5.82 billion USD. This estimated trade disruption amounts to a share of 0.8% of total Japanese exports over the same time period. As in the case of the Muhammad Comic boycott, this is a rather small percentage of the total Japanese export economy. For the other Chinese entities, there is no negative effect. While trade does decline over the months following the boycott call, this is a general trend affecting the control countries as well. In the case of Hong Kong and Taiwan, one can see positive effects at longer horizons. This might be a substitution of exports from Mainland China towards the other Chinese entities, significantly reducing the overall negative impact of the Mainland boycott. One can conclude that the boycott was effective only in Mainland China and that 24

Figure 5: Realized and Counterfactual Trade Levels (Taiwan, Hong Kong, Macao) 22.5 Log Exports 22 Realized Taiwan Synthetic Taiwan Boycott 21.5 30 25 20 15 10 5 0 Months since Boycott 5 10 15 20 Log Exports 22.5 22 Realized Hong Kong Synthetic Hong Kong Boycott 21.5 30 25 20 15 10 5 0 5 10 15 20 Months since Boycott Log Exports 17.5 17 16.5 Realized Macao Synthetic Macao Boycott 16 30 25 20 15 10 5 0 5 10 15 20 Months since Boycott 25

Figure 6: Estimated Cumulative Loss in Trade over Time (Mainland China) 2 x 109 0 Estimated Cumulative Treatment 95% Confidence Interval Loss in Trade (in bn USD) 2 4 6 8 10 Boycott announced 12 0 2 4 6 8 10 12 Months after the Boycott the movement was unable to encourage Chinese people in Taiwan, Hong Kong, and Macao to participate in the boycott. 6.2.2 Robustness Checks and Placebo Tests The short pre-boycott period does not allow for a sensible placebo assignment of the treatment time. I instead estimate the treatment effect for the control countries that should not be affected by the boycott. I calculate the percentage losses of Japanese imports for the countries of France, Germany, Russia, India, Thailand, the UK, and the US which are all major trading partners of Japan. Table 10 summarized the estimates for these countries. The results show that for the majority of the controls, the boycott did not have a significant effect on imports from Japan. France and Russia are exceptions as they show negative impact over a 6-month period, this effect however disappears at the one-year window. The US shows a positive reaction to the Chinese boycott which is statistically significant at both the 6 and 12 month window, suggesting that the Japanese exporters substituted their goods towards the US. 26

Table 8: Synthetic Control Groups and Associated Weights China Taiwan Hong Kong Macao Australia 0 Bangladesh 0.08 Azerbaijan 0 Brunei 0.45 Canada 0.03 Indonesia -0.40 Bangladesh 0 Cambodia 0.41 France -0.08 Kazakhstan -0.04 Sri Lanka 0.06 Sri Lanka 0.12 Germany 0.16 Malaysia 0.85 Finland 0.02 Estonia 0.19 Indonesia 0.09 Pakistan 0.05 Kazakhstan -0.12 Georgia -0.06 Italy 0.02 Philippines 0.48 Malaysia 0.64 Laos -0.11 Korea 0.38 Singapore 0.14 Oman 0.12 Mongolia -0.13 Mexico 0.11 Vietnam -0.18 Pakistan 0.02 Nepal 0.08 Netherl. -0.24 Thailand 0.07 Philippines 0.17 New Guinea 0.09 Russia 0.07 Singapore -0.08 Turkmenist. 0.05 India 0.15 Vietnam 0.19 Uzbekistan -0.07 Spain 0.15 Thailand 0.03 Turkey -0.04 UK 0.02 Corr..92 Corr..84 Corr..92 Corr..49 RMSPE.002 RMSPE.002 RMSPE.002 RMSPE.011 Unrestricted country weights for synthetic control group. Corr. is the correlation coefficient between treatment and synthetic control group in the pre-treatment period. RMPSE is the root mean squared prediction error between treatment and synthetic control group in the pre-treatment period standardized by the mean of the time series. In addition, I estimate a difference-in-difference model of Japanese exports given by log Y j,t = α+β 1 Chinese j +β 2 post t +β 3 Chinese j post t +β 4 GDP t +β 5 dist j +β 6 t+ɛ j,t. (6) where Chinese j is an indicator variable that takes the value of one if the country is either China, Taiwan, Hong Kong, or Macao. Since the synthetic control group method suggested that besides Mainland China, the boycott did not have an effect, I experiment with a setup where only China is assigned the treatment. The results in Table 11 show that treatment effect is negative in all specifications, but that the estimates fluctuate widely. As expected, the inclusion of the Taiwan, Hong Kong, Macao render the results insignificant as these entities do not seem to truly belong to the treatment group. My preferred estimate in column (4) indicates a 20.4% reduction in Chinese imports from Japan which is significantly larger than the result obtained by the synthetic control group of 4.3%. This suggests that even though both methods detect a significant reduction in trade due to the boycott announcement, the choice of the control group can influence the result drastically. 27

Table 9: Estimated Trade Disruption Country 3 Months 6 Months 12 Months Mainland China -5.01% -9.50% -4.32% (2.49%) (1.84%) (1.35%) Taiwan 3.30% 4.11% 10.22% (3.16%) (2.30%) (1.79%) Hong Kong SAR 2.71% 4.18% 5.94% (2.33%) (1.70%) (1.31%) Macao SAR 9.31% -2.40% 1.78% (11.09%) (7.44%) (5.29%) Prediction errors in parentheses. Table 10: Placebo Treatment Effect for Control Countries Country Correlation 3 months 6 months 12 months China 0.848-5.01% -9.50% -4.32% (2.49%) (1.84%) (1.35%) France 0.856-7.10% -6.60% -3.58% (4.63%) (3.33%) (2.43%) Germany 0.864 3.47% 3.32% 1.06% (3.81%) (2.74%) (2.02%) Russia 0.852-5.97% -14.03% -3.58% (6.55%) (4.63%) (3.62%) India 0.868 4.57% 7.03% 2.16% (5.39%) (3.93%) (2.82%) Thailand 0.712-4.46% 3.36% 1.92% (3.31%) (2.55%) (1.93%) UK 0.962-6.14% -2.17% -4.05% (9.72%) (6.92%) (5.19%) USA 0.913-0.11% 3.84% 6.60% (1.92%) (1.41%) (1.06%) Correlation refers to the pre-treatment correlation coefficient between treatment and synthetic control unit. Prediction errors in parentheses. 28

Table 11: Diff-in-Diff Japan Log Imports from Japan Full Treatment Mainland China only (1) (2) (3) (4) GDP 8.15e-13 7.70e-14 8.74e-13 8.82e-14 (4.63e-14) (2.80e-14) (5.78e-14) (3.06e-14) Distance -0.000158-0.000158 (0.00000954) (0.00000950) Post -0.208-0.191-0.208-0.191 (0.135) (0.0391) (0.135) (0.0392) Treat 2.312-0.277 (0.330) (0.424) Post Treat -0.517-0.0817-1.172-0.204 (1.026) (0.0541) (0.195) (0.0585) Constant 16.50 12.96 16.56 12.96 (2.282) (0.653) (2.283) (0.658) Fixed Effects no yes no yes Time Trend yes yes yes yes N 6034 6034 5998 5998 adj. R 2 0.245 0.944 0.236 0.943 Robust standard errors in parentheses. p < 0.05, p < 0.01, p < 0.001 6.2.3 Identifying Consumer Industries Instead of dividing trade into consumer, intermediate, and capital goods I look at more disaggregated product-level data. While the Broad Economic Categories classification is available for HS codes, the individual product type series will suffer from the same problem as the total trade values of China, since they will be the largest and cannot be reproduced without non-negative weights. This problem is less likely to appear for product-level HS6 codes, as China might not be the biggest export market for all of them. I make use of publications of the Chinese boycott movement itself to identify consumer goods that are most prone to the boycott. These publications are two flyers that were circulated on the internet at the height of the conflict and contain pictures of Japanese brands that Chinese consumers should avoid (see Figure 7). I collect these brand names and their industry in Table (17). Most of these firms are concentrated in a few industries, namely automotive, consumer electronics, foods, clothing, and cosmetics, while the remaining companies engage in industries as diverse as toys, cigarettes, and airline services. 29

Figure 7: Internet Flyers Calling for Boycott I searched through the companies internet representations and identify the brands major brands and products. I then classify these products into the corresponding HS codes using the official description and the commercial website http://hs.e-to-china.com/ that allows to search for keywords and outputs the relevant HS code. These signature products can be subsumed into seven product codes which show significant amount of trade between Japan and China. Table 12 summarizes the trade codes and their description. These codes contain highly branded goods such as passenger cars, make-up and beauty articles, foods, and a variety of consumer electronics such as cameras and video recording devices. HS Code Table 12: List of Main HS Codes Description 1902 Pasta, prepared or not, couscous, prepared or not 22 Beverages, spirits & vinegar 3304 Beauty, make-up & skin-care prep, manicure etc 8508 Electromechanical tools, working in hand, parts 8521 Video recording or reproducing apparatus 8703 Motor cars & vehicles for transporting persons 9006 Photographic still cameras, flash apparatus etc I estimate the impact of the boycott on these consumer goods and Table 13 summarizes the results. The category that sees the most drastic decline in trade is unsurprisingly 8703 which includes passenger cars. Figures 8 shows the realized and counterfactual trade levels for Mainland China. Clearly visible is the massive drop in car imports and although they catch up to the control group after about nine months, Japanese car exports to China drop by a 32.3% within a single year. While the effect of the boycott is very clear for vehicles, evidence for other product codes is not obvious. The estimated percentage disruption in trade in highly-branded goods like 30