Tariffs and Politics: Evidence from Trump's Trade Wars

We use the recent trade escalation between the US, China, the European Union (EU), Canada and Mexico to study whether retaliatory tariffs are politically targeted. Using aggregate and individual-level data we find evidence that the retaliatory tariffs disproportionally targeted areas that swung to Trump in 2016, but not to other Republican candidates. We propose a novel simulation approach to construct counterfactual retaliation responses. This allows us to both quantify the extent of political targeting and assess the general feasibility. Further, the counterfactual retaliation responses allow us to shed light on the potential trade-offs between achieving a high degree of political targeting and managing the risks to ones own economy. China, while being constrained in its retaliation design, appears to put large weight on achieving maximal political targeting. The EU seems successful in maximizing the degree of political targeting, while at the same time minimizing the potential damage to its own economy and consumers.


Introduction
Work by political scientists and economists suggests that a common factor linking the election of Donald Trump, the UK's Brexit vote and the wider populist surge in Western Europe may reflect a long-delayed political backlash against globalization (see Autor et al., 2016;Colantone and Stanig, 2018a,b). It is thus not surprising that US trade policy would see marked shifts under a Trump administration. Nevertheless, the announcement on March 1, 2018 that the US would impose a 25% tariff on steel and a 10% tariff on aluminium imports still came as a surprise. Initially exempt, Canada, Mexico, and the EU became subject to the steel and aluminium tariffs from May 31, 2018. Additionally, the Trump administration set a tariff of 25% on 818 categories of goods imported from China worth $50 billion on July 6. Following the announcement tariffs, President Donald Trump asserted that "Trade wars are good, and easy to win." Despite this assertion, the dispute involving China, the European Union (EU), Canada and Mexico escalated with reciprocal tariffs targeting imports from the US.
As only few of the 316 GATT disputes the WTO (2018) lists from 1948 to 1995 reach this stage of escalation in which threatened tariffs are actually imposed and retaliation is triggered, we know little about how trade disputes are actually fought. In this paper, we use the recent trade-escalation to study how trading partners engage the US in this trade dispute. In particular, this paper tackles two interrelated questions using both aggregate and individual-level data.
In the first part, we ask whether the retaliatory tariffs are designed to target Trump's voter base. The large literature on pork barrel spending (e.g. Levitt and Snyder, 1995;Bickers and Stein, 1996;Canes-Wrone and Shotts, 2004;Berry et al., 2010;Berry and Fowler, 2016) suggest that politicians see value in spending money to gain the support of their voter base. By the same logic, negative shocks hitting Donald Trump's voter base may generate political pressure to remove tariffs and deter future protectionism. 1 While the question whether adverse economic shocks can produce political effects is a field of active and extensive research, so far findings suggest that the effects are highly heterogenous and contextdependent (see Margalit, 2019 for an excellent review). Margalit (2011), finds a distinct anti-incumbent effect of trade-related job losses vis-a-vis other types of economic shocks, while Scheve and Slaughter (2004) suggests that trade-integration increases perceived insecurity. Feigenbaum and Hall (2015) shows that politicians from districts most exposed to the "China-shock" became more protectionist. This lends credence to the idea that economic effects of retaliatory tariffs can shift political support. Yet, a likely necessary condition is that retaliation is sufficiently targeted (see Kavaklı et al., 2017;Marinov, 2005).
In the second part, we investigate both the feasibility of political targeting and study the extent to which countries are managing the harm retaliation may afford on their own economy. In the context of trade disputes the structure of trade between countries may afford an important constraint (see Kavaklı et al., 2017).
Given the large literature on the welfare enhancing effects of trade (e.g. Frankel and Romer, 1999;Baldwin, 2004), it is widely accepted that tariffs, while able to help certain individual industries, are not only harmful for trading partners but also constitute an act of self-harm (Bown, 2004). As such, the design of retaliation by countries may not only reflect the desire to induce economic and political pressure, but also may reflect domestic political and economic considerations (Davis, 2004;Barari et al., 2019).
Our findings suggest that political targeting appears to play an important role in the retaliation design. To assess the degree of political targeting, we construct a county-specific exposure measure similar to Autor et al. (2016) and Colantone and Stanig (2018a,b). Based on this exposure measure, we find that retaliatory tariffs target areas that swung to Trump in the 2016 presidential election. In contrast, areas that swung behind other Republican candidates in the House or Senate elections held on the same day where not target of retaliatory tariffs. Using individuallevel opinion polling data, we show that even among self-identified Republicans retaliation appears to be distinctly targeted towards areas in which Republicans favored Donald Trump over other Republican contestants for the 2016 presidential nomination. Further, we document that the degree of political targeting appears to pick up a distinct shift in geographic patterns in Republican party affiliationbut only after Donald Trump entered the 2016 Presidential race in 2015.
To asses both the feasibility and the degree of political targeting that retaliating countries could have achieved we propose a novel simulation approach constructing counterfactual retaliation responses. This simulation approach further allows us to shed light on which other considerations are likely to play a role in the design of retaliation responses. For the EU, for example, it is known that policy preferences and national politics may impact the bloc's stance in international negotiations (Kleine and Minaudier, 2019;Wratil, 2019;Meunier, 2000). In the context of trade disputes the EU is transparently stating its objectives of retaliation: retaliation should induce compliance of the US with trade rules, while mitigating harm to EU consumers and firms. Motivated by this example, we construct reducedform measures proxying the potential for domestic harm of retaliation. In our analysis we compare the actual chosen retaliation response relative to the counterfactual baskets. The results confirm that, in line with the objectives the EU's achieved a high degree of political targeting, while ensuring that the US is not the dominant supplier of targeted targeted products. For the Chinese tariffs we find that, due to feasibility constraints afforded by the structure of China-US trade, any retaliation response that produces some degree of political targeting requires imposition of tariffs on goods for which the US is a dominant supplier. This suggests that Chinese retaliation may be particularly harmful for Chinese consumers.
Our results contribute to the literature on the politics of protectionist trade policies. The imposition of domestic tariffs has been attributed to either the influence of interest groups (Grossman and Helpman, 1994), peoples inequity aversion (Lü et al., 2012), the importance of tariffs as a source of revenue (Hansen, 1990), the structure of consumer tastes (Baker, 2005) along with the relative factor endowments (Scheve and Slaughter, 2001). Existing research further suggests that democracies are more likely to lower tariff barriers, but are more likely to protect their agricultural sectors and make use of non-tariff barriers (NTB) (Kono, 2006;Barari et al., 2019, e.g.). Cameron and Schuyler (2007) investigates the determinant of protectionism in the agricultural sector. In closely related work Gawande and Hansen (1999) investigate the deterrence effect of NTB and how retaliatory NTB can be used to reduce foreign trade barriers. Our findings shed light on how other countries react to protectionism and the US's aggressive trade policy.

Context and Data
The international trading system after the second World War was first institutionalized through the General Agreement on Tariffs and Trade (GATT) in 1948.
It was a direct result of the failings of the international trade system during the Great Depression. In 1930, the Smoot-Hawley Act increased tariffs on more than 20,000 products imported by the US. This set off a tit-for-tat retaliation. Irwin (1998) estimates that nearly a quarter of the observed 40% decline in imports can be attributed to the rise in the US tariff and thereby contributed to the lengthening of the Great Depression.
Through multiple GATT rounds from 1948 onwards, average tariff rates were reduced significantly. One of the most important features of the international trading system which is now regulated by the WTO -the successor organisation to the GATT established in 1995 -is a formal Dispute Resolution System. In principle, governments are still able to restrict trade to foster non-economic social policy objectives, to ensure "fair competition", or to support preferential treatment of developing countries, regional free trade areas and customs unions. But measures of this kind are subject to scrutiny, should adhere to the broad principles of the WTO and can be contested by WTO member countries by invoking the WTO's Dispute Resolution mechanism. Rosendorff (2005) and Sattler et al. (2014) provide evidence that the WTO's Dispute Resolution mechanism helps to enforce stable trade relationships. The Dispute Resolution mechanism also regulates the imposition of retaliation measures.

Retaliatory Tariffs as a Political Tool
The most recent precedent in which the international trading system came close to a similar escalation were steel tariffs imposed by President George W. While this does not proof that the threat of retaliation was the reason why tariffs were abandoned, it does suggest that it may have played a role. The European Commission stands out in terms of transparently the objectives it aims to achieve in the context of trade disputes (see Baccini, 2010;Stasavage, 2004  (c) availability of alternative sources of supply for the goods or services concerned, in order to avoid or minimise any negative impact on downstream industries, contracting authorities or entities, or final consumers within the Union; In other words, trade policy should aim to change the trade policy of the opposing country, while minimizing harm to the own economy. To design the retaliation response, the European Commission is known to use an algorithm to select products against which retaliatory tariffs are targeted. This algorithm is naturally a safely guarded secret. 2 The Chinese government does not publish their policy objectives in the trade dispute, but there is evidence that they also try to target their tariffs against the electoral base of Donald Trump and the Republican party. For example, the Chinese as well as the EU's retaliation targeted bourbon whiskey produced in Ken- These examples suggest that the design of retaliatory tariffs shares some similarities with political sanctions. The growing literature on sanctions (see for ex-2 One of the authors of this paper had a conversation with an anonymous senior EU commission source, who referred to the algorithm as the EU's "weapon of war" in the context of the trade dispute, indicating why it is a closely guarded secret. ample (e.g. Elliott and Hufbauer, 1999;Eaton and Engers, 1992;Ahn and Ludema, 2017)) understands sanctions as a political tool to induce compliance. In a closely related paper, Kavaklı et al. (2017) find that comparative advantage in exports and domestic production capabilities determine a countries' ability to maximize the economic impact while minimizing the domestic costs of sanctions. In this literature, Dashti-Gibson et al. (1997) studies the success factors of economic sanctions, while Marinov (2005) and Allen (2008) provide evidence that sanctions increase the probability of leadership change. In other related work, Draca et al. (2018) show that US sanctions against Iran are indeed targeting politically connected firms and actors. 3 In contrast, the political dimensions of tariffs so far has been widely ignored. In our analysis, we investigate to what degree the retaliating countries indeed systematically politically targeted their retaliation. For our analysis, we construct a measure of exposure to retaliatory tariffs for each US county, which we discuss next.

Descriptives of the retaliation measures
The retaliation measures against the US tariffs take the form of a list of products with descriptions and (typically) the Harmonized System (HS) code along with an (additional) tariff rate to be imposed on imports of these goods stemming from the US. These lists are prepared through a consultative process in the case of the EU and Canada. They are lodged and registered with the WTO and, there is typically a delay prior to the tariffs being implemented. For our analysis, we have obtained 3 Whether sanctions are effective in inducing compliance is a different question: Grossman et al. (2018) find that the EU's labelling of products from the West Bank -in the relative short-termproduced a backlash in Israel and increased support for hardline policies. Similarly, Peeva (2019) suggest that sanctions against Putin following the Crimea annexation actually backfired and helped Putin's approval ratings. retaliatory tariff lists from the EU, China, Mexico and Canada. We do not analyze the retaliation of other countries such as India and Turkey, as the overall trade volume and therefore the retaliation is far smaller. 4 Appendix Figure A2 visualizes the distribution of the retaliation measures across coarse economic sectors. The figure suggests similarly, that manufacturing sector outputs, as well as agricultural commodities were significant features in the retaliation lists. We next describe how we use the retaliation list to construct a county's exposure to tariffs.

Measuring exposure to retaliation
We use two data sources to construct a county level measure of exposure to retaliation measures. First, we use data from the Brookings (2017) Export Monitor.
This data contain a measure of county-level exports across a set of 131 NAICS industries. 5 We denote X c,i the export of industry i for each county c. The data also provides an estimate of the total exports at the county level and the number of export dependent jobs. The latter will be used to weight the regressions.
Secondly, we use the individual retaliation lists L r for r ∈ {EU, MX, CA, CN}.
These are matched at the 8-digit HS level to the US trade data using export volume. 6 -cmpcaa-eng.asp, accessed 18.08.2018. 5 The data incorporates a host of data, including US goods trade data, service-sector export data from the Bureau of Economic Analysis (BEA), Internal Revenue Service (IRS) data for royalties, Moody's Analytics production estimates at the county level, and foreign students' expenditures from NAFSA. More details on Brookings (2017) can be found https://www.brookings.edu/ research/export-nation-2017/.
6 While technically the codes of products are provided at the 10 digit level, the matching results are best at the 8-digit HS level due to slight discrepancies in the coding standard across countries. affected by tariffs with the official WTO submissions. For this exercise, we make use of HS-level U.S. import and export data from the U.S. Census Bureau. 7 In the case of the EU, the retaliation measures officially target trade worth USD 7.2 billion. Matching the EU list to the US trade data for 2017, we find that US exports worth USD 7.6 billion are affected by retaliation, suggesting that the overall magnitude is similar.
To link the targeted exports to the different six digit NAICS sectors that produce the goods (HS10 codes), we use the concordances between HS codes and NAICS/SIC codes from Schott (2008). These concordances provide up to 10 digit commodity codes, which map into the Harmonized System codes used, together with SIC and NAICS codes. This allows us to merge the tariffs lists with the employment data. In case multiple sectors are linked to an HS10 code, we retain the NAICS sector listed first in the concordance. As an illustration, consider the example of the EU's rebalancing measures, which includes the item "10059000 Maize (excluding seeds for growing)". This HS code is mapped to the NAICS industry 111150, which stands for "Corn Farming". This procedure results in a list of tariff exposed industries.
Next, we collapse the total volume of affected trade to the four digit industry level. This gives us a measure of export T i,r affected by retaliatory tariffs of country r for each four digit industry i. We break this measure down to the county level using X c,i , the amount of production of industry i in county c as measured by the Export Monitor data. In other words, the total export volume affected by tariffs is broken down to the county-level using the share of a county in overall exports This introduces only a small amount of inconsequential noise. 7 These data can be found here https://usatrade.census.gov/.
from industry i. The final exposure measures τ c,r for county c and list of retaliatory tariffs r is given as: This measures approximates the exposure of counties to retaliation measures of each retaliating country r. The measure is bounded between 0 and 1. If all industries in a county are unaffected by tariffs the measure is 0. If the entire production of a good subject to retaliation were to take place in a single county and if that county were to export only this good, the exposure measure would be 1.
Our approach is similar to the Autor et al. (2013)-type labor market shocks.
The main difference is that rather than constructing this measure based on sector level employment figures, our measure is based on sector level output figures.
This should come closer to capturing the economic impact more broadly. As a robustness check we consider an alternative exposure measure based following Autor et al. (2013) and Kovak (2013), which uses the County Business Patterns (CBP) employment data to construct a county-level retaliation exposure based on sector-level employment shares. In Appendix Table A3 we show that results are similar when using this alternative measure. 8 Since the added tariff rate was set at 25% for 85% of the products, our retaliation exposure measure ignores the actual added tariff rate. This also implies that the variation in our county-level exposure measure τ c,r is driven by the choice of products and not the choice of tariff rates. While this is only a small deviation from the actual data, it greatly simplifies the simulation of counterfactual retaliation baskets in Section 4. In Appendix Figure A1 we compare our exposure measure with the exposure measure that would result if we incorporate the actual added tariff rate. The two measures are statistically virtually identical.

Main political outcome measures
In the following, we describe the aggregate and individual-level data sources used to measure the extent of political targeting. we provide some auxiliary evidence that complements their work suggesting that retaliation was indeed effective in reducing US exports and lead to drop in export prices, suggesting that exposure to retaliation produced indeed an economic shock.

Aggregate election results
3 Was the retaliation politically targeted?

Descriptive evidence
We first provide descriptive evidence that counties with a stronger support for the Republican party were more heavily targeted by tariffs. Figure 1 In this specification, y c,s measures the vote share of the Republican party in county c in state s in 2016. As an alternative outcome we use ∆y c,s , the change in the Republication party vote share between the 2012 Presidential election and the 2016 Presidential election at the county level. τ c,r is the county level exposure measure for retaliatory tariffs list r (for more details see Section 2.3). All regressions includes state fixed effects, hence we exploit within-state variation in retaliation exposure. Standard errors are clustered at the county level.

Results
The results from the estimation of model 1 are presented in Table 1.
The results suggest that counties which are more exposed to retaliatory tariff had higher levels of support for Trump in the 2016 presidential election. Further, as indicated in Panel B, counties exposed to retaliation also saw larger swings in sup-port from the 2012 Presidential election to the 2016 Presidential election. The point estimate in column (2) suggests that the counties most exposed to EU retaliation saw an average swing in the Republican candidates' vote share of 22 percentage points vis-a-vis counties not exposed to EU retaliation.
As the retaliation exposure measures τ c,r are bounded between zero and one the coefficients are directly comparable. We find, that the degree of political targeting is strongest for the EU and Mexico's retaliation. We will revisit this result in our simulation study in Section 4. Before turning to the individual-level data, we next conduct further robustness checks for our baseline findings.

Robustness
We first explore whether the targeting was stronger for the presidential election than for the House and Senate election held on the same day (Tuesday, November 8, 2016). The results of this exercise can be found in Appendix Table   A1. Panel A explores Republican party vote shares. Throughout, there is a strong positive correlation -yet, we find no evidence for differences in targeting across election types. In Panel B we compare the changes in Republican candidate vote share vis-a-vis the elections held in 2012 (for Presidential and House elections).
For the Senate election, we compare the change with the most recent prior Senate election for Senators (as only 1/3 or Senators are up for election each time). In this specification it appears that the regression coefficient for retaliation exposure is markedly larger for the Presidential election but not for Republican candidates across other election types. This hold true despite the fact that voters could vote on the same day in 2016. This provides some additional evidence that retaliation may have been targeted to hit areas that swung behind Trump in 2016. A potential rational behind such a strategy could be that these voters may conceivably swing back (see Alesina and Rosenthal, 1995, 2006or Scheve and Tomz, 1999 for work studying the dynamics of US presidential and midterm elections).
In Appendix Table A2 we highlight that the correlation between retaliation exposure and (shifts in) support of Republican presidential candidates is distinctly stronger for the 2016 election. We investigate this observation further in the individual level analysis. Our finding are similar when we use an alternative exposure measure based on the sector-level employment shares inspired by Autor et al.
Lastly, in Appendix Table A4 we show that our results are robust to the inclusion of additional control variables. 9 First, we control for a county-level measure of the China shock used in Autor et al. (2013). This control is motivated by the work of Autor et al. (2016) who find that Trump performed better in counties that were more exposed to Chinese import competition. 10 In line with this result, we find that the estimated coefficient on the China shock is positive and significant.
Yet, our retaliation exposure coefficient hardly changes.This is not surprising for two reasons. Naturally, a county's exposure to retaliation depends on the structure of trade between the US and the trading partner. Retaliation exposure is driven by US exports, while the China shock is based on US imports. In addition, tradedispute induced retaliation can only produce economic shocks in regions and parts of the US in which the tradable-goods producing sectors have survived the "China shock". We also control for the level (and changes) in turnout in the 2016 presidential election. Guiso et al. (2018) suggest that the ability for populist candidates 9 Note that we focus on the combined retaliation exposure measure. The patterns are very similar when analyzing country-by-country. 10 Similar effects have been documented in the context of the UK and Western Europe more broadly (Colantone and Stanig, 2018a,b); Feigenbaum and Hall (2015) shows that politicians from districts most exposed to the "China-shock" became more protectionist direction. to affect turnout may be a key feature to understanding their success. Indeed, in Appendix Table A5 we document that places more exposed to retaliation saw, on average, lower levels of turnout. Yet, our observation suggesting that retaliation was politically targeted remains intact.

Cross-sectional individual-level data
We use repeated individual-level cross-sectional data from the Gallup Daily Tracker. This allows us to study the extent of support for Donald Trump using individual-level micro data allowing us to control for a set of potential confounders. Further, we can exploit variation over time and draw comparison to other Republican candidates.

Empirical specification
To leverage individual level data we modify our above regression specification in the following way: In this regression y i,c,t measures whether an individual i in county c in year t has a favorable view of Donald Trump as candidate. In our analysis, we focus on the period from June 2015 to March 2016 prior to the election and prior to Donald Trump becoming the presumptive nominee. This allows us to compare the degree of targeting for other Republican candidates who were (still) in the race at the same time. The specification controls for state fixed effects α s as well as a set of individual controls X i . In particular, we control for the respondents race across five categories, income across 12 categories, gender and the year of the survey. In specifications where the dependent variable is not party-affiliation, we also control for an indicator whether a respondent describes themselves as Republican or leaning Republican.
Since the Republican party affiliation is observed consistently from 2012 onwards, we can further estimate a flexible difference-in-difference specification: Since the regression contains county fixed effects α c and time fixed effects γ t , the coefficient β r,t captures the differential changes in individuals' leaning towards the Republican party and our county-level measure of retaliation exposure. In other words, β r,t picks up whether areas more exposed to retaliatory tariffs exhibited shifts in the support for the Republican party relative to previous years. If retaliation was indeed targeted to counties with a Republican voter base that identifies with Donald Trump, we would expect the correlation between individual respondents self-reported affinity towards the Republican party and the county-level retaliation exposure measure to increase with Trump's presidential run. Further, this analysis will also show whether there were changes in Republican support before Trump's campaign started. In this way, we can disentangle general shifts in political preferences or party affiliation from support for Trump as a candidate.

Results
In Table 2 A potential concern with these findings could be that the retaliation patterns simply capture geographic differences of republican versus democratic support. To highlight that retaliatory tariffs indeed appear to target areas with strong support for Donald Trump, we analyze the period in which the Republican nomination was still open and included Donald Trump as a candidate (July 2015 onwards until March 2016). 11 We further focus on the subset of respondents who self-identify as Republican (≈ 23.4% of the sample). With this analysis we aim to capture whether retaliation exposure was distinctively targeted against areas who supported Trump instead of another Republican presidential candidate.
The results are presented in Table 3 for any of the other presidential candidates. For the Mexican and the Canadian retaliation the correlation is also positive, but statistical insignificance. This finding suggests that retaliation was carefully chosen to target areas with Republican supporters with an affinity for Donald Trump. The specific targeting of Trump's voter base, exhibits parallels to the targeting of politically connected firms by economic sanctions in Iran (Draca et al., 2018), both of which are likely to increase the pressure on the respective political leader.
Lastly, in Figure 3 we present the estimated difference-in-difference coefficients from specification 3. The figure suggests that the correlation between a county's exposure to retaliation and indviduals' leaning towards the Republican party becomes distinctly stronger from 2016 onwards. This suggest that retaliation was targeted against areas which increased their support for the Republican party relative towards the 2015 baseline level. In other words, areas that during Trump's presidential run were swayed to support Republicans were more strongly targeted than areas that always exhibited a strong support of the Republican party. It is also worth noting that trends prior to 2015 are flat. This suggests that our retaliation measure is not confounding other latent trends in the geography of Republican party affiliation that pre-date Donald Trump's candidacy. 12 If we were simply picking up the trade-induced manufacturing sector decline (Autor et al., 2016), for example, this trends should be visible before the 2016 presidential election We next explore a short individual-level panel highlighting that retaliation, especially from the EU and China, was targeted to the US that saw sizable swings from Obama in 2012 to supporting Trump in 2016.

Individual-level panel data
As an additional piece of evidence, we leverage the 2016 CCES study which asked the respondents if and for whom individuals voted in the 2012 and 2016 Presidential elections. The advantage of the CCES in comparison to the Gallup data is that it directly measures voting behavior instead of approval or party affiliation. In this way, the CCES data allow us to study whether individuals switched their party support vis-a-vis the 2012 election. We estimate regression specification 2. The set of individual-level controls X i includes race, gender, age, income and political party affiliation. As we estimate the regression in first-differences, we implicitly accounting for time-invariant individual-level characteristics (similar to individual fixed effects). In particular, we study the direction of the switchi.e. whether retaliation was concentrated in counties with voters that swung from supporting Barack Obama to supporting Donald Trump. We present the results from this analysis in Table 4. The estimates are statistically significant for the EU and the Chinese retaliation exposure measures. The point estimates suggest that in counties most targeted by EU retaliation, the likelihood of an individual voter to be a swing voters that switched from supporting Obama to Trump is 7.6 percentage points higher. For counties exposed to Chinese retaliation at the same level, the likelihood is 3.8 percentage points higher.
In Appendix Table A6 we confirm the results for the subset of voters for which their voting status has been validated based on official voter lists. The patterns remain broadly the same, even though we do lose some statistical precision.
Taken together, the results suggests that retaliation appears to have been chosen to target counties in which Trump had a particular appeal and voters increased their support for the Republican party. The patterns documented across three different data sources are remarkably consistent. Additionally, the fact that the Trump administration provided billions of dollars in farm aid packages (see for example NYT, 2018), suggest that the effect of retaliatory tariffs was felt in the targeted areas. In Appendix B we provide auxiliary evidence for the economic consequences of the tariffs. In line with the findings of Levitt and Snyder (1995); A remaining concern is that the underlying patterns could be spurious in a fashion that can not be accounted for with individual level or other county-level control variables. Specifically, one might worry that the specific mix of products that countries purchase from the US may mechanically constrain the structure of any retaliation response. To address this concern we exploit the fact that for the initial wave of tariffs -which we study in this paper -the constraints on the retaliation response are quite well defined. This allows us to construct counterfactual baskets countries could have chosen and evaluate the degree of political targeting against these counterfactuals. These counterfactual baskets additionally allow us to investigate other constraints on the retaliation response.

Counterfactual retaliation baskets
Is the observed targeting of Republican counties a mere artefact of the US export mix with specific trading partners and do trading-partners face trade-offs due to domestic constraints? In this section we attempt to answer these questions by proposing a simulation approach exploiting retaliation design constraints to construct feasible counterfactual retaliation baskets.

Retaliation design constraints
In our simulations, we leverage the fact that trading rules impose constraints on the design of retaliation (or more formally, rebalancing measures). The key constraint is that that the applied retaliatory tariffs should be commensurate with the US tariffs. For example, the tariffs imposed by the US on steel and aluminium affected around USD 7.2 billion of EU exports to the US with an expected added overall tariff revenue volume of USD 1.6 billion. 13 To comply with WTO rules, the EU's expected tariff revenues from the retaliation should not exceed this amount.
Our aim therefor is to identify a vector of products i among all traded HS goods categories for which there is non-zero imports, M i,r > 0, into retaliating country r along with a vector non-zero tariff rates t i,r > 0 to be applied such that the combined expected tariff revenues ∑ i∈S r t i,r M i,r is less than the expected tariff revenues that the US levies on imports from country r, T r . As previously discussed, the choice of the tariff rates is secondary for the retaliation wave we study: for 85% of product classes included in the actual retaliation the added tariff rate was fixed at 25%. For the counterfactual construction we therefore ignore the choice of the added tariff rate t i,r implicitly assuming a fixed rate t. 14 With a fixed tariff rate the above problem becomes a subset sum problem. (see Figure A3). This potentially leaves an uncountably large set of combinations of products for which the combined affected imports from the US is approximately equal to the US tariffs. To overcome this challenge, we use a probabilistic simulation approach to identify a set of alternative baskets

Simulation approach
In particular we use the following sampling procedure for each country's retaliation list L: while less than 1000 alternative retaliation baskets L r,i have been found: do 1. Randomly draw an integer N i indicating length of retaliation list in terms of HS10 codes -allow for a 20% deviation around list length of actual retaliation N r * 2. Draw a sample list L i,r of HS10 codes of length N i on which there is some exports from the US in 2017 3. Compute the volume of exports from US to country r that would be affected by retaliation if the sample list L i,r were chosen ∑ i∈L i,r E i,US,r 4. if 0.9 < ∑ i∈L i,r E i,US,r ∑ i∈L * r E i,US,r < 1.1 then Accept the candidate list L i,r ; end end As indicated in the pseudo-code we construct counterfactual retaliation list by first choosing a similar number of products to target (allowing for 20% deviation).
We then sample a set of products to target and calculate the effected export volume.
Lastly, we accept any list which effects a similar amount of exports as the actulal list (allowing for a 10% deviation). The result of our sampling procedure is a set of retaliation lists that are similar to the original list in many dimensions, but target a different set of US exports. While the simulation approach trace out some aspects of the "retaliation possibilities frontier", it ignores two potential strategic elements. First, retaliation lists may be designed in a way to preserve an option value to hit back in case of a further escalation. Second, retaliating countries may coordinate their retaliation responses to maximize the effectiveness. It is also import to note, that the counterfactual retaliation bundles are not orthogonal to the actual retaliation basket (see Appendix Figure A4). The observed positive correlation mechanically results because the simulated baskets overlap with the actual retaliation basket.

Evaluating the degree of political targeting
The simulation approach is particularly useful as it allows us to quantify the degree of political targeting relative to the counterfactual baskets. More specifically, we evaluate whether retaliation appears at the upper-or lower end of the potential retaliation distribution. We also investigate the underling trade-offs that countries face in their retaliation design. For this analysis, we estimate the regression models studied in the previous Tables 1, 2 and Table 4 Table 5 presents the share of the counterfactual estimatesβ r that would imply 15 Note that there is a non-negligible cross-correlation across retaliation bundles. Appendix Figure A4 highlights that the implied measures and the actually chosen retaliation response have a positive correlation almost across each of the 1000 counterfactual bundles. This is a mere direct result of the retaliation response that meet the criteria to be quite similar will produce some overlap, implying a mechanical cross correlation. a higher level of political targeting. In column (1) and (2) we focus on the the outcomes studied in Table 1. Column (1) suggests that for China, there exist hardly any feasible and comparable retaliation response that would produce a stronger  Tables 2 and Table 4. They observed patterns are broadly similar.
The finding that Canadian and Mexican retaliation, while being quite robustly associated with support for Donald Trump, does not appear to be at the upper end of the achievable targeting distribution, suggests that other considerations may have played just as important a role. We next aim to investigate the which other objectives countries might include in their considerations.

Retaliation trade-offs
The previous section suggests that retaliation appears to specifically target parts of the US that swung to support Donald Trump. Yet, relative to a set of counterfactual retaliation responses, especially for Canada and Mexico, we observed that the implemented choice seems suboptimal. What may explain this observation?
As our discussion of the EU's retaliation design objectives suggested, countries designing the retaliation have multiple objectives. In the EU regulation constraining retaliation design the mitigation of harm to consumers and firms features prominently along with the political effectiveness. In this section we construct a set of relevant measures that might constrain the retaliation choice. In particular, we investigate the role of the revealed comparative advantage, the import demand elasticities and the dominance of US exports.

Revealed comparative advantage
The first measure is an index of the revealed comparative advantage (henceforth, RCA) as introduced by Balassa (1965). The intuition for this index, which is constructed based on export data, is that countries appear to have a revealed comparative advantage for a good h if a higher share of the countries export is accounted for by this good relative to the export share of this good across all trading countries. Formally, an RCA value above 1 for a specific good h indicates that a country has a revealed comparative advantage (see Kavaklı et al., 2017 for a recent example using RCA measures in the context of economic sanctions). When designing their retaliation response countries reasonably might want to avoid goods for which the US has a Revealed Comparative Advantage.
We denote the implied average RCA for each retaliation list as RCA i,r , which we weight by the implied volume of trade. 16 As the construction of the RCA indices requires trade data between all countries, we can only construct the RCA at the HS6 digit level, based on data from UN Comtrade. Import demand elasticities Whether a specific good is chosen for retaliation may also depend on the associated (import) demand elasticities. Presumably, in order 16 While the sum of the weights across baskets will be the same as our counterfactual baskets target a similar volume of trade, the distribution of weights differ. for retaliation to be effective, goods for which import-demand is found to be particularly price elastic would proof to be more effective. Further, tariffs on goods with a high import demand elasticity are less likely to affect domestic consumers.
We therefore use the import-demand elasticity estimates constructed by Soderbery (2018) at the HS4 level for each of the retaliating countries. As before, we compute the retaliation-specific weighted average import-demand elasticity specific to a counterfactual retaliation list i for country r, σ i,r and evaluate this against the elasticities associated with the actually chosen retaliation response, σ r * . 17

Dominance of US exports
Countries may also want to avoid retaliating and raising the cost of a specific imports for which the US is the predominant source. To measure this, we construct the share of imports I i,h,r of a good h on a retaliation list i of country r that stems from the US relative to the rest of the world, s h,i,r = We compute the trade-volume weighted average implied share of US imports, s i,r for each good in the retaliation lists, across each of the counterfactual retaliation lists i for country r. We then again evaluate the corresponding shares associated with the actual retaliation list s r * compared to the counterfactual lists. This analysis is conducted at the HS6 level (based on UN Comtrade data).
In Table 6, we report summary statistics for the three measures and how they compare across retaliation baskets. Ideally, countries in order to minimize harm to their own economy would favor retaliating against goods with a low RCA, a large import-demand elasticity and a low US import market share. In Table 7 we contrast how the distribution of counterfactual baskets compares with the actual retaliation response. The EU's retaliation appears to be targeting goods for which 17 For the EU, we use Germany as the US's biggest trading-partner's estimated elasticities from Soderbery (2018). The results qualitatively similar if we use other EU countries as reference. the US has a weaker RCA and goods for which US is less dominant supplier.
The Mexican response, on the other hand, appears to be targeted goods basket with a relatively high import demand elasticity and a lower revealed comparative advantage. We next shed light on the underlying trade-offs visually.

Results
For every (potential) retaliation list i of retaliating country r, we have now constructed a vector of attributes (β i,r , RCA i,r , s i,r , σ i,r ) . To illustrate the trade-offs and constraints imposed on retaliation design we visualize the joint distribution of the pair (β i,r , RCA i,r ) in ships with the US, there are much fewer constraints on retaliation design. Relative to the counterfactual baskets, we observe that in particular for the EU and China, retaliation appears to have been chosen at the upper end. To the best of our knowledge, we are not aware of another paper that has explored retaliation in this way.
For the EU, there exist very few alternative retaliation bundles that would produce a higher degree of political targeting and a lower RCA value. The same is true for Canada, and, to a lesser extent for Mexico.
In Appendix Figure A5, we study the implied import-demand elasticity. The figure highlights that, for both Canada and Mexico, retaliation appears to targeted towards goods with a high import demand elasticity and a higher degree of political targeting. Appendix Figure A6 studies the implied US market power for specific retaliation baskets. Based on this measure, the EU's retaliation response stands out in achieving a fair degree of political targeting, while avoiding goods for which the US is a dominant supplier.
In Table 8, we computes the shares of retaliation baskets that would imply a higher degree of political targeting while considering our other proxies capturing retaliation effectiveness and domestic economic harm. Throughout, the chosen retaliation appears at the upper end in terms of producing high political targeting but a lower RCA. For the EU only around 1% of the counterfactual retaliation responses would produce a higher degree of political targeting and a lower RCA.
The Chinese retaliation response clearly stands out as it appears to target goods with a high RCA. Much of this is afforded by the specific constraints that Chinese retaliation design faces as the vast majority of other feasible retaliation baskets would produce no political targeting whatsoever.

Conclusion
Based on the recent trade escalation provoked by the administration of Donald Trump, this paper provides empirical evidence for the political targeting of retaliatory tariffs. Using a novel simulation approach, we show that retaliatory tariffs indeed disproportionately targeted more Republican areas. This suggest that retaliatory tariffs appear to have a clear political dimension. We further illustrate that countries face a trade-off between the degree of political targeting and potential harm done their own economy. Our findings suggest that countries appear to put different weights on these two policy objectives. To the best of our knowledge, this paper is the first to empirically document this trade-off.

Future work should hence incorporate whether retaliation is effective in shap-
ing the underlying trade-policy preferences of politicians and the electorate more broadly. This paper suggests that such an empirical study, for example, using difference-in-difference designs will have to find a way to navigate the endogeneity of retaliation exposure that this paper highlights.

43
Electronic copy available at: https://ssrn.com/abstract=3349000   Notes: The dependent variable is an indicator stating whether a respondent holds a favorable view of the candidate indicated. The responses includes don't know, refused and those that hold no view. The patterns are robust to dropping these observations. Regressions include individual level controls: respondents racial identity, income, republican party affiliation, gender and the year of the survey. Regressions are weighted using survey weights provided by Gallup. Standard errors are clustered at the county level and are presented in parentheses, stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1. Notes: The dependent variable is a dummy indicated in the panel label. All regressions control for state FE and are weighted with the provided survey weights. Regressions include individual level controls: respondents racial identity, income, republican party affiliation and gender. Standard errors are clustered at the county level and are presented in parentheses, stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1. The table reports analysis of the implied measures of the extent of political targeting implied by the set of simulated counterfactual retaliation baskets vis-a-vis the actually chosen retaliation response. The figures represent the share of retaliation baskets that imply a retaliation exposure measure above what is implied in the actually chosen retaliation response. Columns (1) -(2) study the county level data explored in Table 1, columns (3) -(5) use the measures leveraged in Table 2, while columns (6)-(8) explore the measures studied in Table 4.     Figure A1: Ignoring tariff-rate does not skew retaliation exposure measure: comparing retaliation exposure measure including or ignoring the applied tariff rate Figure A2: Which sectors were targeted by retaliation measures? Combining the EU, Canada, Turkey, India, and Chinese retaliation lists.

A Additional Figures and Tables
Notes: Pie chart plots the trade-volume weighted distribution of countermeasures across sectors using the 2017 export volume.

B What are the economic effects of retaliation?
As a first measure of economic impact, we study the effects of retaliation on trade flows and export price indices. While reduced trade-flows could purely capture both trade-disruptions as well as trade-diversion, any impact of retaliatory tariffs on export price indices is likely to indicate tangible economic shocks. This data is available for around 90 different four digit NAICS sectors and will help complement the analysis on trade flows. Specifically, since trade flows may simply be re-routed, it could be that the income implications of the tariff may be limited. Hence, studying export price indices may help shed light on whether tariffs actually did produce a negative income shock. Figure ?? presents the year on

B.1 Data on Economic Impact Measures
year changes in the export price indices of the US for agriculture and manufacturing sector outputs. Note that, these figures do not account for the different size of the relevant sectors, but the observed deterioration in export prices following July 2018 is evident, indicating that export prices did indeed collapse. We also test this in a more robust econometric framework later on.

B.2 Empirical specification
Impact on exports We first investigate the impact of retaliation on US trade flows.
We use monthly US export data at the HS8 level to measure US exports to China, the EU, Canada and Mexico as well as the rest of the world. We then estimate the following difference-in-difference regression: In this specification y measures US exports and the index r indicates the country which retaliated against the US. T h,r is indicator variable which is 1 if good h was chosen to be included in the retaliation basket of country r. The regressions control for a range of shifters and fixed effects. Most importantly we include HS8 by trading country specific shifters α c,h capturing country r specific tastes for imports from the US of goods h. We also control for destination country r specific time fixed effects as well as additional time fixed effects, indicated here by ν i,t . These additional time effects can be specific to a destination country r or, could account for good-specific seasonality. The latter is particularly relevant as US agricultural exports are highly seasonal.
Impact on export prices Secondly, we estimate the impact of retaliation on export price indices. This analysis is based on export price indices constructed for 46 NAICS4 sectors by the Bureau of Labor Statistics. We study to what extent sectors more exposed to retaliation measures saw a differential change in their export prices. To do so, we construct the exposure of a NAICS4 sector n to retaliation from country r, indicated as E n,r as follows. Having merged the HS8 export data to NAICS codes, we compute the total volume of US exports in 2017 at the 4 digit NAICS level that would become subject to retaliatory tariffs from July 2018 by country r and divide this by the overall export volume. The tariff exposure measure across the 46 four digit industry groups for which it is constructed ranges from 0 to 34.6%, indicating that at the top 34.6% of exports produced by an industry was affected by tariffs. The average exposure measure is 5%. We then estimate the following regression: The dependent variable measures the export price index at the four digit NAICS sector n. The sector fixed effects, α n j , are at the level of the three digit sector or the four digit sector. Hence, we explore both within and between NAICS sector variation. We include time fixed effects throughout. Further, in some more demanding specifications we allow for time by first digit NAICS sector fixed effects. These first digit sectors broadly distinguish agriculture, mining and manufacturing. Standard errors are clustered at the four digit NAICS sector level. The main coefficient of interest is β r . We would expect that this coefficient to be negative, indicating that after retaliatory measures came into effect, export price indices decrease for exports from sectors with a higher retaliation exposure E n,r .

B.3 Results
Impact on exports The regression results are presented in Table B1. The point estimates in panel A suggest that exports that were exposed to retaliation shrank by around 75%. Panel B-E explores to what extent this result is robust to the exclusion of specific trading partner. It becomes obvious that, the Chinese retaliation, accounts for around 50-60% of the estimated contraction of US exports. This is expected since the Chinese retaliation was by far the most extensive given the structure of US trade with China. Nonetheless, also goods targeted by the EU, Canada and Mexico exhibit a significant reduction in exports to these markets.
Overall, the point estimate suggest that each month around USD 2.55 billion worth of exports have either not taken place or were diverted as a result of the tariff measures, amounting to around USD 15.28 billion in aggregate since the retaliation measures became effective in July 2018 until the end of 2018. Panel A in Figure B2 provides an event study version of specification 4, estimating separate coefficient for each pre-and post treatment month. The figure highlights the sharp contraction in export volumes since July 2018, when most retaliation measures became effective. In Appendix Figure B1, we estimate the event studies focusing on pairs of countries, studying the US exports to a specific country that retaliated and to the rest of the world with just these two series. The results highlight a strong degree of seasonality in exports of goods that were subject to retaliation by China, which captures the agricultural crop cycle across the US. Notably, the peak in exports that should occur around the summer failed to materialize as commodity exports fell significantly due to retaliation. The figure suggests significant contractions in bilateral exports relative to trade with the rest of the world across the dyads that were affected by the retaliatory measures.
These results do not preclude the possibility that most of this trade was rerouted and absorbed by other trading partners. Yet using the case of soybeans, a look at aggregate numbers suggests, that there is a net contraction of exports.
In other words, the exports to the rest of the world have not absorbed the tariffinduced reduction in demand.
To show that the retaliatory tariffs likely also had a significant effect on incomes in areas that produce the affected commodities (and not just capture tradererouting), we next provide some evidence suggesting that US export price indices also significantly declined. Table B2 presents the results from this analysis. Since the data are aggregated into far coarser industry sectors, the point estimates are unsurprisingly more noisy. Nevertheless, the results suggest that export prices declined significantly in 4-digit NAICS sectors that were more exposed to retaliatory tariffs. Panel A studies the overall sector level retaliation exposure measure, while in Panels B -E, split the retaliation exposure measure by country. The findings indicate that, at the coarse four digit level, only the retaliation by China, Mexico and Canada had a significant effect on export price indices. To reiterate, this is not surprising given the coarseness of the export price indices data. The results also rely on variation between 4 digit NAICS sectors (in columns 1-3). When wen fully focus on within sector variation over time (columns 4-6), the estimates are even more noisy.

Impact on export prices
To illustrate the timing of the effects, in Panel B of Figure B2 shows that the contraction in export price indices occurs at the time of the introduction of retaliatory measures, with export prices growing strongly in early 2018. This could partly highlight increased demand due to stockpiling.
Taken together, the evidence from exports as well as exports prices, indicates that the retaliatory tariffs did indeed, induce some economic harm on the effected sectors. In that sense the tariffs were effective. As the last piece, of our analysis we now investigate whether tariffs also had a political impact.  Notes: Figure plots estimates from a difference-in-difference regressions. Panel A presents point estimates capturing the evolution of exports from the US to EU, China, Canada, Mexico and the ROW over time on goods targeted by retaliation. The underling regressions control for HS8 code by destination shifters, destination by time fixed effects and targeted sector specific seasonality. Standard errors are clustered at the 4 digit HS code level. The right Panel B presents results from a regression studying 46 export price indices constructed at four digit NAICS level. The plot presents point estimates capturing the evolution of export price indices over time as a function of the 4 digit NAICS sectors exposure to retaliation measures as the share of exports in 2017 at the NAICS4 level that was exposed to retaliation measures. The underlying regressions control for NAICS4 export price index fixed effects and time fixed effects; regressions are weighted by the 2017 overall export volume and standard errors are clustered at the NAICS4 level. 90% confidence bands are indicated. Notes: The dependent variable is the level of US exports at the HS8 level by month. Standard errors are clustered at the 4-digit HS good level with stars indicating *** p < 0.01, ** p < 0.05, * p < 0.1.