Trade Attitudes in Latin America: Evidence from a

Multi-Country Survey Experiment∗

Marisol Rodríguez Chatruc Ernesto Stein Razvan Vlaicu

September 2019

Abstract

This paper examines individual-level support for trade, relates it to beliefs about trade, and measures its sensitivity to positive and negative framing. The data comes from the 2018 Latinobarometro survey of eighteen countries, in which we embed a survey experiment to study framing effects. We find that respondents are generally favorable to increased trade with other countries, based on perceived positive impacts on employment and consumption. Support for trade is unaffected by consumption benefits framing, but is highly sensitive downward to employment loss framing. Positive framing does shift upward respondent beliefs that trade reduces consumption prices, but also raises concerns about low wages. Negative framing substantially weakens the prevailing beliefs that trade is associated with higher employment, and there is no offsetting effect on the consumption side. Trade attitudes reflect behavioral patterns but also display variation across education levels consistent with traditional skill-based theories of trade.

JEL Codes: F13, D72. Keywords: trade preferences, trade beliefs, skills, framing, survey experiment.

∗Contact information: Research Department, Inter-American Development Bank, 1300 New York Ave NW, Washington, DC 20577, United States. Rodríguez Chatruc ([email protected]), Stein ([email protected]), Vlaicu ([email protected]). We benefitted from conversations with Ana Ines Basco, Gustavo Beliz, Matias Busso, Jeffry Frieden, Marta Lagos, Fabiana Machado, Anna Maria Mayda, Marian Moszoro, Victoria Murillo, Emanuel Ornelas, Carlos Scartascini, and seminar and conference audiences. We thank Kurt Birson, Sergio Perilla, Camila Valencia, and Victor Zuluaga for excellent research assistance. The findings and interpretations in this paper are those of the authors and do not necessarily reflect the views of the Inter-American Development Bank or the governments it represents.

1 1 Introduction

The notion that countries benefit from trading with each other is a main tenet of trade theory. However, trade policy does not always reflect this view. One reason is that while trade liberalization may increase aggregate welfare, at the individual level there will be winners and losers. The conflicting trade preferences are then resolved through the political process. In normal times, trade policy is not a salient electoral issue, in which case policy is decided between policymakers and private sector firms and other interest groups (Grossman and Helpman 1994). Voters often benefit from trade liberalization through lower prices and access to a wider range of goods and services, while at the same time being affected by employment losses in some industries or occupations (U.S.: Autor, Dorn, and Hanson 2013, Brazil: Dix-Carneiro and Kovak 2017). But they are typically not active participants in policymaking, as they are a diffuse group with lower stakes than other actors. However, when trade policy becomes politically salient, the voting public’spreferences for trade matter (McLaren 2016). Standard economic models of trade have clear implications about an individual’sprefer- ences for trade, starting from the premise that material self-interest is the key determinant of preferences (Rodrik 1995). One of the most discussed models in this literature is the Heckscher-Ohlin model and its key implication, the Stolper-Samuelson Theorem. The the- orem says that under perfect factor mobility, trade makes owners of relatively abundant factors better off, due to increased prices of the goods that use these factors more intensely, and hurts the owners of relatively scarce factors. For example, in a country with a relative abundance of unskilled labor, unskilled workers would benefit and skilled workers would lose from trade. The related Ricardo-Viner model, which drops the assumption of perfect factor mobility, predicts that gains will be concentrated in sectors that benefit from trade, such as export-oriented industries. Thus, workers in these sectors will support trade, whereas work- ers in import-competing sectors would oppose it. More recent models of trade emphasize that trade benefits the most productive companies, those that produce high-quality goods and are export-oriented, resulting in higher demand for skilled workers in all sectors of the economy, and leading to an increase in the skill premium (Melitz 2003, Verhoogen 2008). In this case, skilled workers should be the ones in favor of trade. The distributive tensions resulting from trade provide a platform for interest groups and political actors to exploit existing cleavages in their favor. The recent rise in anti-trade public discourse in developed countries is an example of these tensions becoming salient.

2 The stylized models mentioned above imply that voters form their preferences based on the impact they expect trade will have on their labor market outcomes. But are trade shocks that operate through the labor market the main determinants of individual trade preferences? Are trade preferences stable or do they shift at times when trade becomes more salient in public debates? How does the sensitivity of trade preferences depend on individual determinants of labor market outcomes, such as education?1 This paper studies individual-level trade preferences and beliefs using data collected through a module included in the 2018 Latinobarometro, a nationally-representative survey of eighteen Latin American countries. We asked respondents whether they support increased trade between their country and other countries and added questions eliciting beliefs about the consequences of trade on employment, wages, prices, product choice, and respondent’s personal economic situation. To study the malleability of preferences and beliefs, we designed a survey experiment for the Latinobarometro survey. We randomly assign three common pro- and anti-trade arguments in the preface to the trade support question. One treatment frames the question by mentioning the positive consumption benefits of trade, namely product diversity and lower retail prices. Another treatment frames the question by mentioning the potential negative employment effects of trade in import-competing industries. A third treatment combines these positive and negative frames; we call this mixed framing. The control group receives an unframed question. To our knowledge, this is the first trade-related survey experiment conducted on a multi-country nationally-representative sample. Our research design yields several interesting results. Untreated respondents are gener- ally very favorable to increased trade with other countries, based on perceived benefits to employment and consumption. In this context, positive framing does not significantly raise support for trade liberalization. Positive framing does induce an upward revision in beliefs that trade lowers consumption prices, but it also raises concerns about low wages. Negative framing substantially reduced trade support, driven by a large downward change in the pre- vailing view that trade has positive employment effects, with no offsetting upward change in beliefs about consumption benefits. Mixed framing also reduces support for trade, but to a lesser extent, suggesting that pro-trade arguments can significantly mitigate the impact of anti-trade arguments, even though they may not work by themselves. We argue that framing impacts are consistent with common behavioral patterns, such as loss aversion, negativity bias, and reason-based choice.

1 Empirical evidence that has linked electoral outcomes to trade shocks affecting the labor market includes Autor, Dorn, Hanson, and Majlesi (2016) and Che, Lu, Pierce, Schott, and Tao (2016).

3 A novel contribution to the current literature is to explicitly measure beliefs in addition to preferences. This allows us to unpack the reasons why framing affects preferences as well as to estimate which beliefs have more influence on trade preferences. We find that labor market considerations, in particular beliefs about employment and wages, are the primary determinants of trade preferences in Latin America in 2018.2 Support for trade, as well as sensitivity to framing, vary across countries and across individual characteristics, such as age, gender, and particularly education. Education, a key determinant in tests of standard economic models of trade, is strongly associated with support for trade. At first sight, this seems consistent with the predictions of new trade models based on Melitz (2003) where trade increases the skill premium across the board. However, we find a difference between the control group, where the more educated are more supportive of trade in virtually all countries, and the mixed framing group, where the more educated are less supportive of trade in the relatively skill-scarce countries. Alternatively, the association between education and support for trade could be driven by factors other than material self-interest, such as ideology (O’Rourke and Sinott 2001, Mayda and Rodrik 2005) or culture, knowledge, and values (Mansfield and Mutz 2009). Our research design can distinguish between the skill-based explanation and the culture-based explanation because it randomly assigns an informational treatment that should reduce differences in knowledge between the more educated and the less educated, but leaves intact their differences in skills. Our results show that the more educated are more sensitive to framing, and particularly in the relatively skill-scarce countries. This suggests a mechanism where in relatively skill- scarce countries trade decreases the labor market return to skill, but provides consumption benefits through lower prices. Thus, for the more skilled the employment and consumption effects of trade go in opposite directions, weakening their priors about trade, thus leaving them more susceptible to framing. We provide evidence that this mechanism works through changing beliefs about how trade impacts employment and wages. Understanding the determinants of individual trade preferences, and the impact issue framing can have in shaping these preferences in a region like Latin America, is important for several reasons. First, trade in developing countries is not as salient an issue in public debates as it is in developed ones. Therefore, developing countries provide a more fertile ground to test the impact of framing, since individuals are not as exposed to anti-trade discourse. Second, opinions about trade in developed countries are associated with opinions

2 This contrasts with the findings of Baker (2003) from an earlier period, which suggested the primacy of consumption considerations.

4 about offshoring, since both issues are frequently brought together in public debates as sources of domestic job losses to globalization. An advantage of focusing on Latin America is that offshoring production of domestic firms is marginal so the economic factors driving opinions are less intertwined with opinions on offshoring activities, a related but different phenomenon. Lastly, Latin American countries differ widely in their level of trade openness and economic development but the region is more homogeneous culturally (language, religion, race, etc.) than other regions of the world, providing an ideal setup to run a multi-country survey experiment. The literature on individual trade preferences has initially focused on testing the predic- tions of economic models of trade, either in country data (Canada: Balistreri 1997, Beaulieu 2002, U.S.: Scheve and Slaughter 2001, Blonigen 2011, Di Tella and Rodrik 2019) or in in- ternational surveys (O’Rourke and Sinott 2001, Mayda and Rodrik 2005, Mayda 2008, Jakel and Smolka 2017). This literature has found evidence for labor-market-related determinants of trade preferences. At the same time, it has revealed that economic incentives are only part of the story. Personal characteristics, such as gender and age, or values, such as nationalism or neighborhood attachment, explain a significant part of the variation in trade attitudes (O’Rourke and Sinnott 2001, Mayda and Rodrik 2005). Or, economic considerations may play a limited role because individuals have beliefs about the effects of trade on their welfare that do not correspond to standard models of material interests, e.g., because individuals may care about a broader notion of welfare (Mansfield and Mutz 2009, Rho and Tomz 2017). However, as Hiscox (2006) has noted, the survey data used in this research may be unreliable due to the implied negative framing contained in several standard survey questions, e.g., that restricting imports "protects" the national economy. Individuals appear susceptible to whether the question is framed in terms of the benefits, as opposed to the costs of trade. These framing effects can be considerable. Moreover, as education reduces framing sensitivity and education should affect trade preferences, then inferences from standard analyses of key trade preference determinants like education can be biased. Ardanaz, Murillo, and Pinto (2013), using a survey experiment in Argentina similar to Hiscox’s(2006) done in the U.S., found a similar dampening impact of framing on support for trade but the effect was driven by skill-based imports exposure rather than intellectual sophistication. Our research design provides a different way to test for the mechanisms behind the education effects on trade attitudes, by exploiting cross-country variation in skill abundance in our sample. We can test how framing sensitivity varies with education differentially in skill-abundant vs. skill- scarce countries. Another important departure from the few studies that relied on framing

5 experiments in that we measure the mediating role of beliefs about trade. In addition to testing whether trade preferences match a particular trade theory, we can estimate how trade arguments impact beliefs about, e.g., consumption and employment, that underlie trade preferences.3 The structure of the paper is as follows. The next section describes the data and the empirical models used. The following section presents the main results regarding the effects of framing on preferences for trade, and beliefs about trade, and links preferences to beliefs. Next we study economic and non-economic determinants of preferences and beliefs, and whether framing effects vary with these determinants. The final section draws conclusions from the findings and discusses implications for future research.

2 Data and Empirical Strategy

The data was collected as part of the Latinobarometro survey conducted between June- August 2018. Latinobarometro is a nationally-representative survey conducted since 1995 in the same eighteen countries in Latin America. Table 1 shows the country coverage of the data. The sample is designed to be representative of each country using a multistage probability sampling design, stratifying within each country by region, sub-stratifying each region by municipality size and urban/rural areas within the municipality, and then selecting households within blocks. The survey design then uses frequency matching to obtain a sample with similar age and gender distributions to each country’snational census. With a few exceptions, respondents were interviewed in their homes and data were collected either with paper surveys or with handheld electronic devices. In 2018 the total sample size was 20,204 respondents from 2,679 blocks, 1,064 municipalities, 268 regions, 18 countries. The survey elicits individual-level preferences and beliefs on diverse social, economic, political, environmental, and other topics, and collects a battery of socio-economic charac- teristics of the respondents, such as age, gender, marital status, race, employment status. We embedded a survey experiment in the 2018 Latinobarometro to randomly vary exposure to common arguments regarding trade liberalization.4

3 Our paper is thus related to an important literature in economics that uses survey data to understand the formation of individual preferences, e.g., Luttmer (2001) and Alesina and La Ferrara (2005) study preferences for redistribution. 4 This was the first time Latinobarometro included a survey experiment in its annual questionnare. We had the experiment piloted in the Spring of 2018 to help develop wording for the informational frames that would be easily understandable and to improve response rates.

6 Within each block, Latinobarometro respondents were randomly assigned to one of four groups. Each group was asked a different version of a trade support question. Table 1 presents the sampling frequencies of the four randomized groups. In every country each randomized group contains a fourth of the sample. A group’s question version differed in how it was framed with additional information. Specifically, the four experimental conditions were:5

Control Group (C): Are you in favor or against (your country) increasing trade with other countries?

Positive Framing (T1): Are you in favor or against (your country) increasing trade with other countries so that prices fall and the variety of products you may buy increases?

Negative Framing (T2): Are you in favor or against (your country) increasing trade with other countries even if increased trade causes employment losses in import competing sectors?

Mixed Framing (T3): Are you in favor or against (your country) increasing trade with other countries so that prices fall and the variety of products you may buy increases, even if increased trade causes employment losses in import competing sectors?

The answer options were: "In favor" and "Opposed," with the possibility of responding "I don’tknow." In the few cases in which no answer is recorded, that appears in the data with a "No response" code. We converted the answers into an indicator variable Trade equal to one if the response was in favor, and zero otherwise.6 Prior large-scale surveys that elicited attitudes toward trade, such as the International Social Survey Programme (ISSP) and the World Values Survey (WVS), phrase the question in terms of the respondent’sagreement with placing "limits on imports." We choose to refer to "increasing trade" which covers both imports and exports, and avoided references to

5 See section A3 in the Online Appendix for original question wording in Spanish. 6 The question design is related to the Hiscox (2006) survey experiment in the United States and the Ardanaz, Murillo, and Pinto (2013) survey experiment in Argentina. Our question design differs from these two studies in that the frame does not precede the question in a prompt, but rather is integrated in the text of the question. Also, we separate the consumption (T1) and the employment (T2) effects of trade, whereas the two papers mentioned include positive employment effects in the positive framing together with positive consumption effects, and include negative employment effects in the negative framing together with negative effects on businesses.

7 specific policy instruments, such as tariffs, duties, quotas, or free trade agreements, to make the question accessible to a wider range of respondent backgrounds.7,8 Immediately after the trade support question that features one of the four experimental conditions, all respondents were asked about their beliefs regarding the consequences of increased trade with other countries. Specifically the question for all respondents was the following.9

Which of the following do you think are consequences of increased trade with other countries? (Mention all the consequences you agree with).

The answer options were: "Higher employment," "Higher wages," "Better product va- riety," "Lower prices," "More and better access to technology," "Better personal economic situation," "Lower wages," "Lower employment," "Has no consequences." We convert the answers to this question into eight indicator variables that take the value one when a conse- quence of increased trade is selected by a respondent, and zero otherwise. Each consequence is indicated in a separate indicator variable, labeled as follows Higher Employment, Higher Wages, Product Variety, Lower Prices, Access Technology, Personal Situation, Lower Wages, and Lower Employment. Note that the responses "Better product variety" and "Lower prices" are related to the arguments included in the positive framing treatment. Table 2 presents summary statistics for these key variables measuring preferences and beliefs about trade liberalization. It also includes, in the second half, various individual characteristics of the respondents. We make several observations. A majority of respondents, 61.8%, were in favor of increased trade with other countries. Variation in trade preferences is about five times larger within countries than across countries. Among the beliefs about trade consequences, only Higher Employment garnered a majority, 53.5%. The other trade consequences that came close were Product Variety, Lower Prices, and Higher Wages. While a majority of respondents were in favor of increased trade, only 24.2% believed trade improves their personal economic situation; this may suggest that reported trade support reflects both individual and sociotropic motivations.

7 The AmericasBarometer surveys conducted by the Latin American Public Opinion Project (LAPOP) asked the following question in the 2004—2012 waves: "To what extent do you believe that free trade agree- ments help to improve the economy?" 8 Latinobarometro has traditionally contained a question about "support for the economic integration of countries in Latin America." While economic integration is related to trade, we preferred to design a new question about "trade with other countries" (stated above in experimental conditions C,T1,T2,T3) that is more squarely focused on trade and that covers exchanges with countries outside of Latin America. 9 See section A3 in the Online Appendix for original question wording in Spanish.

8 The individual characteristics include personal, social, and economic variables and display significant variation both across and particularly within countries. We selected those that have been found in the previous literature to be determinants of preferences for trade. Among non-binary covariates, Education is measured on a 1-4 scale, with 1 meaning primary school or less and 4 meaning completed college/technical school. Media measures on a 0-4 scale exposure to mass media by counting the number of media channels, e.g., radio, tv, used to access political news. Ideology is measured on a 0 (left) to 10 (right) scale. Income is measured on a 1-4 scale, where 1 means insuffi cient income with serious financial problems, and 4 means suffi cient income that allows saving. Economy measures the respondent’s perception of the current economic situation of the country, from very bad (1) to very good (5). Market records the respondent’slevel of agreement with the claim that a free market is the only system that can promote development in the respondent’scountry.10 Note that average trade support in Table 2 may be affected by framing which was applied to three-fourths of the sample. In Table 3 we present a breakdown by treatment status. The first column shows the means of trade preferences and beliefs for the control group, i.e., the group in the no framing condition. A sizable majority of respondents, around 72%, are in favor of trade. Framing can change this fraction considerably. Figure 1 plots the raw means overall and also by country, to be discussed below. The control group means of the eight beliefs variables are more similar to their unconditional means; differences do not exceed five percentage points. If we define Price/Variety as one if either Price or Variety are one, zero otherwise, we point out that this together with Higher Employment were the only variables with agreement exceeding half the sample in the control group. Recall that the issues of prices and product variety were included in the positive framing; employment was mentioned in the negative framing.11 Given that the four trade question versions were randomly assigned, we can identify the causal effects of framing on trade preferences and beliefs with the following regression:

yij = β0 + β1T1ij + β2T2ij + β3T3ij + γ0Xij + δkj + εij (1)

where yij is an indicator variable for being in favor of trade, or agreeing with a given con- sequence of increased trade, by individual i from country j; Tnij is an indicator variable for receiving treatment Tn, where n = 1, 2, 3; Xij is a vector of individual-level covariates, such

10 Full variable definitions are in section A4 of the Online Appendix. 11 The questions about beliefs were asked immediately after the survey experiment question. Respondents take the entire survey in one sitting, in the same fixed order of questions.

9 as age, gender, education; δkj is a fixed effect for country j or a subdivision k of country

j; and εij is the error term. No framing, i.e., being in the control group, is the excluded

category. Each coeffi cient βn measures the average difference in the response yij between the treatment group n and the control group; in other words, it measures the average treatment effect of exposing a respondent to that particular type of framing.

For the fixed effects δkj we use either country, region, city, or block, this last subdivision being the unit of randomization. Within each country the survey was administered to re- spondents through several interviewers. Each interviewer is identified by a unique code in the dataset and is always nested within the same county. Including interviewer fixed effects controls for potential interviewer-specific influence on responses. We present OLS or FE estimates of equation (1), assuming a linear probability model. Probit estimates, not re- ported here, are similar. We report standard errors robust to heteroskedasticity for country fixed effects, or clustered at the level of the sub-country fixed effects, whether region, city, interviewer, or block. Given that the Latinobarometro sample is representative at the country level, and ran- domization of treatment assignment was done within a country at the block level, we can study the effects of framing separately for each country. The empirical specification, for each country j, is:

yij = β0j + β1jT1ij + β2jT2ij + β3jT3ij + γj0 Xij + δkj + uij (2)

where in this case δkj are block fixed effects and uij is the error term. The coeffi cients

β1j, β2j, β3j are the country-specific average treatment effects. Table 4 presents evidence of covariate balance across the four experimental conditions for the full sample. It reports the coeffi cients from a regression of each individual-level covariate xij on the three treatment indicators T1ij,T2ij,T3ij with no additional controls and robust standard errors. These covariates are either factual, e.g., age, education, or measure opin- ions, e.g., ideology, trust, that were elicited before asking the survey experiment questions. Thus, none should be affected by the randomized treatments. The estimates show that the ˆ ˆ ˆ treatment coeffi cients β1, β2, β3 are small in magnitude relative to the dependent variable average and none are statistically different from zero at the conventional 5% significance level; only the coeffi cient on employment status is different at the 10% significance level. We highlight that media exposure is balanced across treatment groups, i.e., the treated are not more exposed to media than the controls. Lack of significant differences in observed factors

10 between each treatment group and the control group suggests that unobserved factors are also unlikely to bias coeffi cient estimates.12 The overall nonresponse rate on the trade support question was 4.94%. We studied whether treatment status affects nonresponse by regressing a non-response indicator on the

treatment indicators T1ij,T2ij,T3ij and country fixed effects, results not reported here. For Trade, the positive framing condition decreases nonresponse by one percentage point relative to the control group; the negative and mixed framing do not have different nonresponse rates relative to the control group. Based on this, imputing zero support for nonresponders is unlikely to materially affect the estimated treatment effects. In other words, this imputation underestimates the control-group average trade support from about 76% if we dropped them from the data, to about 72% if we imputed zeros, but should not affect the differences in trade support between framing conditions. Treatment effect estimates based on dropping the nonresponders, not reported here, are very similar.13

3 Framing Effects

This section presents average treatment effects from the framing experiment we embedded in the 2018 Latinobarometro survey. The next section discusses heterogenous effects. We begin with average effects on preferences and then move on to average effects on beliefs. After that we estimate correlations between preferences and beliefs to learn which consequences of trade are the most salient for individual support for trade liberalization.

3.1 Trade Preferences

Returning to Table 3, we can see how different types of framing affect the fraction of respon- dents favorable to trade. From the first row, we see that trade support in the positive framing group is half of a percentage point higher than in the control group, 73.11% compared to 72.64%. Trade support in the negative framing group is sharply lower than in the control group, down more than 25 percentage points; the drop brings support down below the fifty

12 Covariate balance tests country by country are presented graphically in Figure A2 of the Online Ap- pendix. Out of 648 coeffi cients (12 covariates 3 treatments 18 countries), on average 32.4 coeffi cients or less should have p-values below 0.05, or on average× 1.8 per country.× 13 By keeping the nonresponders we also preserve the sample representativeness, and thus the external validity of the results, for a small risk of nonresponse bias. This choice seems justified given the magnitudes of the estimates below.

11 percent mark. In the mixed framing group trade support is lower by about 17 percentage points compared to the control group. See also Figure 1. We estimate framing effects using the regression model in equation (1) and report them in Table 5. We begin with a simple specification that includes only the treatment indica- tors. The next columns augment this model gradually with country fixed effects, followed by region, city, interviewer, and block fixed effects. Below the estimates, in parantheses, are robust and/or clustered standard errors. The results are remarkably stable across spec- ifications, and confirm the simple mean comparisons in the previous paragraph. Positive framing has a positive but small and statistically indistinguishable from zero effect on trade support. Negative framing reduces average support for trade by about 26 percentage points. Mixed framing reduces average support for trade by about 17 percentage points; the 8.8 (= 26.3 17.5) percentage point difference between the negative framing effect and the − mixed framing effect is statistically and practically significant.14 Overall, the experimental results so far indicate that across the 18 countries in the sample a sizable majority favors trade liberalization, nevertheless stated views shift easily toward protectionism when common anti-trade arguments frame the question. It is perhaps sur- prising that positive framing by itself seems to be ineffective in shifting preferences in a more pro-trade direction. However, this finding resonates with the related studies by Hiscox (2006) in the United States, and Ardanaz, Murillo, and Pinto (2013) in Argentina, where the effect of positive framing was actually reversed. In the U.S. study, positive framing reduced average support for trade by almost 5 percentage points, although not statistically different from zero; in the Argentina study, positive framing reduced average support for trade by 9 percentage points, a statistically significant difference. Several potential explanations could account for a weak effect of positive framing in our sample in conjunction with a strong effect of negative framing. First, individuals may only respond to framing when the information provided challenges prior beliefs. Negative framing goes against most respondents’prior beliefs that trade is associated with high employment from Table 3. Second, individuals may react less strongly to a gain (of consumption benefits) than to a loss (of job security), in other words, they may display "loss aversion" as predicted by prospect theory (Kahneman and Tversky 1979). A related psychological asymmetry potentially relevant here is "negativity bias." Suppose we can interpret the average trade

14 Given the asymmetry in the estimated impacts of the positive and negative framing, it seems unlikely that acquiescence bias affected these responses, i.e., the respondents’tendency to agree rather than disagree with the position presented by an interviewer.

12 support level in the control group of 72.64% from Table 3 as the prior belief of the average survey respondent that trade is good. Then negativity bias says that, when individuals who expect good things to happen are presented with credible evidence that challenge their prior, they over-weight the bad news when updating their prior beliefs than when receiving good news that confirm their prior beliefs (Skowronski and Carlston 1989). Third, mixed framing that combines good and bad news significantly mitigates the negative effect of only bad news, i.e., anti-trade arguments, on preferences. This is consistent with a reason-based model of choice where salience of a given reason declines with the number of opposing reasons (Shafir, Simonson, and Tversky 1993). Adding the pro-trade arguments appears to reduce the salience of the anti-trade arguments that go against the average respondent’sprior.15

3.2 Trade Beliefs

Different respondents may have different understandings of how trade affects their personal situation or the overall economy. Studying framing effects on various beliefs about trade can shed additional light on the findings about trade preferences reported above. Prior work on framing has largely ignored this mediating role of beliefs. Without understanding the underlying assumptions behind preferences for trade it is diffi cult to interpret the direction and size of framing effects.16 Going back to the mean comparisons of Table 3, we find that nearly 58 percent of control group respondents associate trade with high employment.17 Around a third of control group respondents associate trade with high wages, more product variety, and low prices. Around a fourth associate trade with improved personal situation and access to technology. Only about 10 percent of untreated respondents believe trade results in low employment and low wages. Table 6 reports estimates of framing effects on all eight beliefs variables. Introducing framing generally comes with a decline in beliefs that trade has desirable employment and consumption consequences. The exceptions are in columns (3) and (4) where positive framing

15 Neither Hiscox (2006) nor Ardanaz, Murillo, and Pinto (2013) find this mitigating effect of their mixed framing treatments. One possible reason is that their good and bad news may be interpreted as mutually exclusive in the employment dimension, i.e., trade creates jobs vs. trade destroys jobs, which may lead some respondents to ignore the positive component of the prompt. 16 The 2014 wave of the Pew Global Attitudes Survey elicits beliefs about trade impacts on jobs, wages and prices for 44 developed and emerging countries, but to our knowledge that data has not been linked to support for trade. 17 This is consistent with the 56 percent share reported in the 2014 Pew Global Attitudes Survey for the case of emerging economies. The share is smaller for developed economies, at 46 percent. See Pavcnik (2017).

13 shifts upward the average respondent’sbeliefs about product variety and lower prices. This should not be too surprising, given that the positive framing focused precisely on those issues. What is surprising is that positive framing also makes respondents more likely to believe increased trade is associated with lower wages, in columns (2) and (7). It is unclear what in the positive framing question leads to this shift. Perhaps some respondents believe that trade leads to a reallocation of factors toward low-skill jobs which pay less. Alternatively, respondents may be (correctly) linking lower prices and greater variety with increased market competition, which means less revenue for local firms and thus lower wages. Whatever the reason, the increased beliefs in low wages could explain why the positive framing has a subdued impact on support for trade. As we will see below, belief in low wages has a strong negative correlation with support for trade on par with the correlations that beliefs about product variety and low prices have with support for trade. Negative framing has large downward impacts on beliefs that trade has positive employ- ment as well as consumption benefits. The impacts are particularly large for the employment- related variables, jobs and wages, as shown in columns (1),(8) and (2),(7). Being exposed to this type of framing reduces the beliefs that trade leads to higher employment by over 8 percentage points, and reduces the beliefs that trade leads to higher wages by more than 4.5 percentage points. The impact of mixed framing on beliefs is mostly negative, with the exception of an upward shift in beliefs that trade leads to lower prices in column (4). This may be driving the attenuated impact of mixed framing on support for trade. The magni- tudes of the effects are somewhat lower relative to negative framing, except for the impact on beliefs about employment and wages, which have comparable magnitudes. Overall, these results reveal high sensitivity of respondents to the job loss framing contained in both the negative and mixed framing.18

3.3 Preferences and Beliefs

Trade preferences should be grounded in beliefs about the effects of trade. If a respondent believes trade has on balance positive effects, then they should be more likely to state they are in favor of increasing trade. In this section we explore the link between trade beliefs and trade preferences. Linking beliefs and preferences may offer important insights regarding what respondents care about when forming trade preferences. Moreover, in combination

18 Interestingly, the mixed framing does not have any discernible impact on the belief that trade increases product variety, even when the framing explicitly refers to this aspect.

14 with the results on the impact of framing on beliefs from the previous section, the link between beliefs and preferences may help illuminate the mechanisms through which framing affects support for trade. Table 7 reports regression estimates for a model where support for trade is a function of the beliefs variables discussed above. Odd-numbered columns control for country fixed effects. Even-numbered columns control for region fixed effects. We present results for the full sample as well as the control group only. The pattern of coeffi cients is similar between the two samples and supportive of the notion that preferences are grounded in beliefs. Positive beliefs are always associated with higher support for trade. Negative beliefs about lower wages and lower employment are always strongly associated with lower trade support. The magnitudes of the coeffi cients in Table 7 are also informative since they provide insights regarding the extent to which individuals care about employment vs. consumption outcomes. Respondents seem to care more about employment than they do about any other outcome, including wages. Within the consumption-related outcomes, respondents seem to attach more importance to expanded product variety than they do about lower prices. Table 7 includes both positive (columns 1-4) and negative beliefs (columns 5-8) regarding two of the outcome variables, employment and wages. In addition to serving as a consistency check, the comparison across positive and negative labor market implications appears con- sistent with loss aversion. Employment losses associate more strongly with trade preferences than employment gains.19 The coeffi cient for lower employment ( .281) is about 30 percent − larger in magnitude vis a vis the coeffi cient for higher employment (.219). The asymmetry is even greater when it comes to wages. The association between support for trade and negative beliefs on wages is about four times larger in magnitude than that between trade support and positive beliefs on wages. These large correlations, together with the large effect of negative framing on beliefs about wages and in particular employment from Table 6, suggest that the large impact of negative framing on trade preferences is driven by high respondent sensitivity to the wage and employment consequences of trade. In particular, the concern about employment seems to be the most important mechanism at play. It is perhaps surprising that the personal economic situation does not correlate more strongly with preferences for trade. This may suggest that both self-interest and sociotropic motivations play a role in the formation of

19 In cognitive psychology and decision theory, loss aversion refers to people’stendency to prefer avoiding losses rather than to acquiring equivalent gains. This asymmetric response to gains and losses was first identified in Kahneman and Tversky (1979).

15 trade preferences.20

4 Heterogeneity

In this section we investigate whether the average effects reported above display interesting heterogeneity. First we report effects estimated separately by country. Then we study eco- nomic and non-economic determinants of trade preferences as well as sensitivity to different framing about trade impacts. As before, we also examine the mediating effect of beliefs about trade.

4.1 Country Variation

To what extent do framing effects vary by country? Figure 1 displays the within-country sample means of our main dependent variable, by treatment status. Countries are arranged in increasing order of mean support in the control group; for easier visualization, the control group means are connected through a dotted line. We see significant variation across coun- tries in support for trade among control group respondents, ranging from just under 60% to over 85%, as well as significant variation in framing impacts, in particular for negative and mixed framing. Framing effects are low in Costa Rica (CRI) and the Dominican Republic (DOM), and large in Brazil (BRA) and Mexico (MEX). In 2007 Costa Rica had a referendum on CAFTA that dominated political life. The previous presidential election, which led to the Arias government, as well as the ones that followed, hinged to a large extent on candidates’ positions on trade. Our interpretation is that, because of the centrality of this issue in Costa Rica’spolitical landscape, respondents from this country have firmer preferences and beliefs, and thus are swayed less by framing.21 Table 8 reports the experimental framing effects by country, using the most demanding model specification from the average effects analysis in the previous section, namely the one

20 This is a interesting direction for future research, particularly linking trade support with perceptions of trade-driven income inequality (Goldberg and Pavcnik 2007). Using a survey experiment that varies infor- mation about the identity of winners and losers from trade intended to alter beliefs about the consequences of trade, Rho and Tomz (2017) find that both egoistic and altruistic motivations are behind preferences for trade. Theoretical treatments on self-interest vs. other factors in the formation of trade preferences include Grossman and Helpman (2018) and Mukand and Rodrik (2018). 21 The Dominican Republic is in fact the only country in which framing does not seem to have any discernible impact, with the exception of a very small positive impact of the mixed framing. It is diffi cult to rationalize this result. Our conjecture is that there may have been implementation problems with the randomized experiment in this country, but we have not been able to identify what these were.

16 that includes block fixed effects. The framing effects for Trade in Table 8 show substantial variation across countries. Positive framing has effects ranging from over ten percentage point increase, in Brazil (BRA) and Mexico (MEX), to over nine percentage point decrease, in Ecuador (ECU) and Honduras (HND). In the rest of the countries, the impact of the positive framing is not statistically significant. Negative framing has negative effects on trade preferences across the board, with the only exception of the Dominican Republic (DOM); the magnitude of the negative framing effect can be as high as the negative 46 percentage points estimated from Brazil (BRA).22 Mixed framing also has negative effects that vary substantially across countries, but the magnitudes are lower, not exceeding 30 percentage points. Overall, we find that, qualitatively, the impact of framing tends to be similar across countries. At the same time, we do find significant heterogeneity in the magnitudes of framing effects across countries, that raises a question regarding its possible sources.23

4.2 Individual Characteristics

We now turn to heterogeneity with respect to individual characteristics. Table 9 presents estimates from regressions of trade support on the twelve covariates included in the covariate balance analysis presented earlier, together with interactions with the treatment indicators, and region fixed effects; the regressions also include the treatment indicators separately, not reported. The coeffi cient on the covariate measures how it associates with trade support in the control group. The coeffi cients on the three interaction variables measure how sensitivity to each type of framing varies with the covariate. The covariates include personal charac- teristics (age, gender, marital status, religion), social characteristics (interpersonal trust, education, media exposure, political ideology) and economic characteristics (employment status, income, perceptions about the economy, support for free markets). Education could be both a social variable and an economic variable as it reflects both social views and skill set. We discuss this in more detail below. As mentioned above, these variables are either pre-determined, e.g., age, gender, or were obtained through questions asked pre-treatment.24 Overall, the covariates relate more to the level of support and less to framing sensitivity.

22 On average over the eighteen countries, the coeffi cient on negative framing is 26.1, which is consistent with the average treatment effect of 26.3 percent for the sample as a whole, as− shown in Table 5 column (6). − 23 This country heterogeneity in framing effects is reflected in beliefs about the employment consequences of trade. See Figure A1 in the Online Appendix. 24 Region fixed effects afford more within variation in these characteristics, since on average there are 75.38 respondents per region. By comparison, we have 18.98 respondents per city, and 7.54 respondents per block on average.

17 The clear exception is Education, which is strongly related to both. More educated individu- als are more supportive of trade, but are also more sensitive to all types of framing. Income displays a similar pattern, although only responses to mixed framing vary with income. We also find that trade support is higher among those with higher interpersonal trust, more media exposure, more right-wing ideology, and those more supportive of markets. Except for media exposure, these associations are to be expected and are qualitatively similar to results from other data, such as those in Mayda and Rodrik (2005) and Mansfield and Mutz (2009), as discussed in the Introduction. While social and economic factors have dominated the empirical trade literature, individ- ual trade preferences may also be linked to personal factors. Here the theoretical underpin- nings are less solid. However, we explore this dimension of the data. In Table 9 the coeffi cient estimates on the covariates show that men are more supportive of trade than women. The gender factor in trade preferences has been a feature of prior work on trade attitudes and our data confirms its presence in the Latin American sample as well. The estimates also reveal that older and married respondents are less sensitive to negative or mixed framing. Perhaps this reflects the fact that more experienced individuals hold more firm beliefs or opinions.25

5 Mechanisms

The heterogeneity analysis above has revealed that education is a key correlate of both sup- port for trade and also sensitivity to framing. Interestingly, the primacy of education as a determinant of trade preferences is also a noticeable feature of the U.S. data, see Bloni- gen (2011). Here we explore these relationships further to shed light on the underlying mechanisms through which education affects trade preferences. Education proxies for two key characteristics that may affect an individual’strade prefer- ences: skill and culture. Skill is a material and personal characteristic that determines one’s position in the labor market. Culture (knowledge, values, cognition) is a non-material and social characteristic that determines one’sviews about the world. An individual’sskill should make them more supportive of expanding trade if trade increases the labor market return to their skill. An individual’sculture should make them more supportive of expanding trade if their understanding of the world suggests that trade improves social welfare. It is generally

25 In Tables A1a and A1b of the Online Appendix we explore heterogeneity in trade beliefs with respect to the same covariates. There also education stands out as the factor most correlated (in the control group) with holding positive beliefs about trade, particularly about the consumption benefits of trade.

18 assumed that more educated individuals have been more exposed to the ideas of comparative advantage during their formal education years, and thus should be more inclined to agree with the economics profession’snear consensus that trade raises social welfare. Whether we think of education as skill, or education as culture, thus has different im- plications for individual trade preferences. It also has different implications for individual sensitivity to trade framing. The hypotheses can be stated as follows. According to the skill-based explanation: (i) the more educated should be more supportive of trade if the economy is relatively skill-abundant, but less supportive of trade if the economy is relatively skill-scarce. (ii) the more educated should be more sensitive to framing if individual employment and consumption effects go in opposite directions, as their priors are weaker, but less sensitive to framing if individual employment and consumption effects go in the same direction, as their priors are firmer. According to the culture-based explanation: (i) the more educated should always be more supportive of trade because of their exposure to the intellectual consensus on trade. (ii) the more educated should always be less sensitive to framing, due to their higher cognition capacity to resist being swayed by messaging. Standard economic models of trade emphasize the role of human capital (skills) in the formation of individual trade preferences. The Heckscher-Ohlin model assumes that factors of production are perfectly mobile across sectors and their returns depend on their relative abundance within a country: factors in which the country has a comparative advantage expe- rience higher returns from trade than relatively less abundant factors (Stolper and Samuelson 1941).26 Empirical work has found data patterns consistent with the Stolper-Samuelson Theorem. However, Hiscox (2006) challenged the observational analyses of Scheve and Slaughter (2001) and Mayda and Rodrik (2005) where more educated individuals are more favorable to trade - particularly in more developed countries - arguing that more educated individuals are less susceptible to the implied negative framing contained in the American National Election

26 Other models of trade emphasize the point that trade benefits the most productive companies, those that produce high-quality goods and are export-oriented, resulting in higher demand for skilled workers, and leading to an increase in the skill premium (Melitz 2003). The Ricardo-Viner model relaxes the perfect factor mobility assumption of the Heckscher-Ohlin model and predicts that gains will be concentrated in sectors that benefit from trade, such as export-oriented industries. The Latinobarometro dataset has limited information about respondents’sector of employment.

19 Studies (ANES) or International Social Survey Programme (ISSP) survey questions, e.g., the suggestion that expanding trade may "hurt" domestic firms. Indeed, using a survey ex- periment, Hiscox (2006) finds evidence in U.S. data that framing effects decline in education. The interpretation he offers is based on cognition differences, namely that more educated individuals have thought more carefully about the effects of trade and thus have firmer prior beliefs about trade. However, Ardanaz, Murillo, and Pinto (2013), using a survey experi- ment in Argentina, show that while education does reduce susceptibility to framing, once economic interests are accounted for, framing impacts are less related to education. They conclude that economic incentives dominate the potential cognitive effects of education in the response to framing: individuals that lose from trade both in the consumption and em- ployment domains are less susceptible to framing than individuals who gain in consumption but lose in employment - and this pattern applies to all education levels. When the loss is both in consumption and employment, an individual has a firmer (negative) prior about trade, and is thus less susceptible to framing. When the loss is in only one domain, say employment, then the individual has weaker priors and is easier to sway with framing. Our research design can help discriminate between the skills-based and culture-based explanations because it randomly assigns an informational treatment that should reduce differences in knowledge between the more educated and the less educated, but does not affect differences in skills. Specifically, our mixed framing treatment provides the two key economic reasons for supporting or opposing trade, namely consumption and employment. We can then test for differences in trade preferences and framing sensitivity between individuals in the mixed framing group and those in the control group.27 In Figure 2 we plot country-level trade-education correlations against countries’ 2017 GDP per capita, separately for the control group and the mixed framing group. GDP per capita has been used in the seminal multi-country studies by Sinnot and O’Rourke (2001) and Mayda and Rodrik (2005) as a proxy for relative skill abundance. The results we present are similar when using UNDP’s Human Development Index (HDI), a more comprehensive measure of country-level human capital that includes social and educational variables. In the figure, the trade-education correlations in the control group are positive for all countries but one, suggesting that in almost all countries the more educated are more supportive of trade. In the mixed framing group, however, the pattern changes: the more educated are

27 By interpreting mixed framing as providing information we do not rule out the more behavioral inter- pretation that framing makes the individual’salready existing knowledge more salient for the choice at hand. See Nelson, Oxley, and Clawson (1997).

20 more supportive of trade in high-GDP countries, but less supportive of trade in the low- GDP countries. This is consistent with the skill-based hypothesis stated above. Tables 10a and 10b confirm this differential patterns in a regression context. In the control group the Education coeffi cient is positive and the Education GDP interaction is statistically zero. In × the mixed framing group the Education coeffi cient is negative, and the Education GDP × interaction is positive. The GDP per capita level where the education gradient changes sign is about 16 thousand PPP current dollars (.054/.034) which is closest to the GDP per capita of Brazil. If the trade-relevant difference between the more educated and the less educated had been knowledge differences about trade, then providing information through the mixed framing treatment should have driven the trade-education correlations to zero in all countries. Instead we observe a response consistent with skill differences between the more educated and the less educated.28 We now turn to differences in susceptibility to framing between the more educated and the less educated. Figure 3 displays the gradient in trade support with respect to education comparatively for the control group, the dashed line, and the mixed framing group, the solid line. It is evident that the drop in trade support due to mixed framing is monotoni- cally increasing in education. Figure 4 plots this difference in education gradients computed country by country against the countries’GDP per capita. The scatterplot suggests a ten- dency for the education gradient differential to be more negative in lower-GDP countries. In other words, the more educated are more sensitive to framing particularly in the skill-scarce countries. This is consistent with the skill-based explanation stated above: in relatively skill- scarce countries trade decreases the labor market return to skill, but provides consumption benefits through lower prices. Thus, for the more skilled the net effects of trade are am- biguous, weakening their priors about trade, thus leaving them more susceptible to framing. Conversely, in the relatively skill-abundant countries, the more skilled expect positive effects from trade both in employment and consumption, thus their priors about trade are firmer, and so they should be less sensitive to framing. The pattern in the graph, while suggestive, may reflect the influence of confounders. We verify the pattern is robust to controlling for country or region fixed effects, and individual- level characteristics correlated with both education and trade preferences. The results are presented in Table 11. In column (1) we replicate the earlier result that the more educated

28 For completeness, we report the same analysis for the positive and negative framing groups in the Online Appendix, see Figure A3 and Tables A2a and A2b. We note that a one-sided treatment (positive framing or negative framing) does not generate the same pattern as the mixed framing treatment.

21 are more supportive of trade in the absence of framing, but also more likely to reduce their support for trade due to mixed framing. In column (2) we see that this sensitivity to framing is less pronounced in higher-GDP countries: the coeffi cient on the triple interaction variable is positive and statistically significant. Columns (3)-(4) introduce region fixed effects instead of country fixed effects, as well as additional covariates. We can also verify if the mechanisms behind trade support sensitivity are working through the sensitivity of trade beliefs. Figure 5 displays education gradients for the beliefs that refer to employment (left side) and consumption (right side) impacts of trade. Interestingly, we observe the same pattern of change in the education gradient for employment, but not for consumption. This reinforces our previous finding that beliefs about the trade-employment link are the main driver of support for trade. The figure shows that the more educated respond more negatively to framing that mentions negative employment effects of trade by revising their beliefs by more than the less educated. This pattern differs from the Hiscox (2006) finding that the more educated are less susceptible to framing, and challenges the culture-based explanation that the more educated are more sophisticated and thus less likely to be swayed by framing. Instead, if we look at how this education-related sensitivity varies with country characteristics, we find evidence more consistent with a skill-based explanation. Figure 6 plots the difference in education gradients of trade beliefs by country. For Higher Employment beliefs, the top graph, the scatterplot suggests a positive relationship. This is reversed in the bottom graph for Lower Employment beliefs. In other words, the more educated are more likely to respond to framing by weakening their beliefs that trade increases employment, and strengthening their beliefs that trade lowers employment, and particularly so in the relatively skill-scarce countries. Table 12 verifies the robustness of this pattern in a regression context. The estimates confirm the message from the plots: in both cases, the triple interaction coeffi cient for Mix Fr Educ GDP has the opposite sign from the interaction coeffi cient for Mix Fr Educ, × × × which implies that the more educated are more sensitive to framing particularly in lower- GDP countries, consistent with the skills-based explanation outlined above. If the relevant difference between the more educated and the less educated had been in cognition capacity, then the more educated would have been less sensitive to mixed framing, and that would have been the case regardless of country characteristics like GDP per capita.29

29 Beliefs about wages show a similar pattern as beliefs about employment, although less pronounced, supporting the point that labor market considerations are paramount in the formation of trade attitudes in our sample. See Figure A4 in the Online Appendix.

22 6 Conclusion

International trade should have positive effects on aggregate welfare in the participating countries, through increased productivity, employment in export sectors, and consumer pur- chasing power. However, in the short run, it may also have negative distributional impacts on segments of the population due to changes in the market equilibrium, particularly in the labor market. Thus, public support for trade is often not a generalized phenomenon. In this paper we explored the "demand side" of trade policy by estimating the level of support for trade liberalization and its sensitivity to common pro- and anti-trade arguments. We collected data through a large multi-country survey experiment conducted as part of the nationally-representative 2018 Latinobarometro survey. We studied the interplay between information, beliefs, and preferences and identified key economic and non-economic factors driving trade preferences. We found that public support for trade in Latin America is con- siderable, based on a prevailing belief that it raises employment, but is also sensitive to informational cues mentioning reduced employment in import-competing sectors. Incorpo- rating the mediating role of beliefs in the analysis of preference formation was valuable in tracing out the reasons for reported preferences, and should be a fruitful approach in future research. We find education to be a key driver of support for trade, as well as beliefs about trade, in our data. The significant sensitivity to framing suggests that behavioral effects such as risk aver- sion and reason-based choice can be important. However, our findings also highlight the importance of labor market impacts on individual trade preferences. Differences in edu- cation are therefore key to understanding the mechanisms behind the formation of trade preferences. We proposed a novel empirical approach to distinguish between skills-based and culture-based explanations of the link between education and trade support. The pat- terns of variation in support levels and framing sensitivity with respect to education appear consistent with traditional theories of trade that emphasize differential trade impacts based on skill differences among individuals and relative skill abundance among countries.

23 Figures and Tables

1. Figures

Figure 1: Trade Support, Raw Means

Cntrl Pos Fr Neg Fr Mix Fr 8 . 7 . e d a r T 6 . n a e M 5 . 4 .

Cntrl Pos Fr Tr eat m ent Neg Fr Mix Fr 1 9 . 8 . 7 . e 6 . d a r T 5 . n a e 4 . M 3 . 2 . 1 . 0 I L L L Y Y V A X D U R N C N G M M R I L H O O E R R E R A E N C T R O C N S P B C P P C B V U E H A M G D

a x is

Note: Sample size is 20,204 respondents from eighteen countries. For country sample sizes see Table 1. Ranges around conditional means are 95 percent confidence intervals.

24 Figure 2: Trade-Education Correlations

Cntrl Group (N = 5,051) PER BRA 3 .

ARG MEX 2 . BOL PRY COL ) c

1 ECU URY u . CRI PAN d NIC

E HND CHL , SLVGTM e DOM d 0 a r VEN T ( r r o 1 . C • 2 . • 3 . • .5 1 1.5 2 2.5 Per Capita GDP, PPP, $10K

Mix Fr Group (N = 5,051) 3 . 2 .

MEX ) c 1

. PER URY u

d CRI

E CHL ,

e NIC

d COL

0 ARG

a VEN BRA r GTM DOM PAN T

( ECU

r BOL

r HND o 1 . • C SLV PRY 2 . • 3 . • .5 1 1.5 2 2.5 Per Capita GDP, PPP, $10K

Note: Figure plots country-level correlations between a respondent’s support for trade and their education, against country GDP per capita, separately for the control group (top) and the mixed framing group (bottom).

25 Figure 3: Framing Sensitivity

Trade Support 9 . Cntrl

8 Mix Fr . 7 . 6 . e d 5 a . r T n a 4 . e M 3 . 2 . 1 . 0

1 2 3 4 Education

Note: Figure plots the mean of the trade support indicator, by level of edu- cation, separately for the control and mixed framing groups. Ranges around conditional means are 95 percent confidence intervals.

Figure 4: Framing Sensitivity Determinants

Trade Support 1 . 5 n 0 o . i t a c

u VEN URY d CHL 0 E x

r NIC CRI F DOM MEX 5

x GTM i 0 . • M HND n COL PAN o

. ECU f 1

f SLV . • e

o PER

C BOL 5 1 .

• BRA PRY ARG

.5 1 1.5 2 2.5 Per Capita GDP, PPP, $10K

Note: Sample consists of the control and mixed framing groups, N=10,102. Figure plots interaction coeffi cients, estimated separately for each country, from a regression of the trade support indicator on Education and Mix Fr x Education, controlling for region fixed effects.

26 Figure 5: Trade Beliefs and Framing Sensitivity

Higher Employment Product Variety 9 9 . . Cntrl Cntrl 8 8 Mix Fr Mix Fr . . 7 7 . . l 6 6 . . p y t m e E i r 5 5 . . h a g V i H n 4 4 a . . n e a e M 3 3 M . . 2 2 . . 1 1 . . 0 0

1 2 3 4 1 2 3 4 Education Education

Lower Employment Lower Prices 9 9 . . Cntrl Cntrl 8 8 Mix Fr Mix Fr . . 7 7 . . 6 l 6 . . p s m e E c 5 i 5 . . r w P o n L 4 4 a . . n e a M e 3 M 3 . . 2 2 . . 1 1 . . 0 0

1 2 3 4 1 2 3 4 Education Education

Note: Figures plot the means of each of four variables measuring beliefs about trade, by level of education, separately for the control and mixed framing groups. Ranges around conditional means are 95 percent confidence intervals.

27 Figure 6: Trade Beliefs and Framing Sensitivity Determinants

Higher Employment 1 . PAN 5 n 0 o . i t VEN a

c NIC CHL u

d CRI URY 0 E

x DOM r

F GTM 5 COL x i 0 BRA . MEX • M BOL PER

n HND o ECU PRY . f 1 f . • e

o SLV ARG C 5 1 . •

.5 1 1.5 2 2.5 Per Capita GDP, PPP, $10K

Lower Employment

ARG 1 . HND BOLSLV PRY 5 n 0 o . i

t GTM ECU PER a

c URY VEN u BRA MEX d 0

E COL

x PAN r DOM CHL F 5

x CRI i 0

. NIC • M n o . f 1 f . • e o C 5 1 . •

.5 1 1.5 2 2.5 Per Capita GDP, PPP, $10K

Note: Sample consists of the control and mixed framing groups, N=10,102. Figures plot interaction coeffi cients, estimated separately for each country, from a regression of Higher Employment (top) and Lower Employment (bot- tom) on Mix Fr x Education, Mix Fr, and Education, controlling for region fixed effects.

28 2. Tables Table 1: Sample Coverage Code Cntrl Pos Fr Neg Fr Mix Fr All Argentina ARG 300 300 300 300 1,200 Bolivia BOL 300 300 300 300 1,200 Brazil BRA 301 301 301 301 1,204 Chile CHL 300 300 300 300 1,200 Colombia COL 300 300 300 300 1,200 Costa Rica CRI 250 250 250 250 1,000 Dominican Rep DOM 250 250 250 250 1,000 Ecuador ECU 300 300 300 300 1,200 El Salvador SLV 250 250 250 250 1,000 Guatemala GTM 250 250 250 250 1,000 Honduras HND 250 250 250 250 1,000 Mexico MEX 300 300 300 300 1,200 Nicaragua NIC 250 250 250 250 1,000 Panama PAN 250 250 250 250 1,000 Paraguay PRY 300 300 300 300 1,200 Peru PER 300 300 300 300 1,200 Uruguay URY 300 300 300 300 1,200 Venezuela VEN 300 300 300 300 1,200 Total 18 5,051 5,051 5,051 5,051 20,204 Note: The table reports sample size by country and treatment status: control, positive framing, negative framing, or mixed framing. Tabulation based on the full sample collected for the 2018 Latinobarometro survey.

29 Table 2: Summary Statistics Obs Mean SD Between SD Within Min Max Trade 20,204 0.618 0.101 0.476 0 1 Higher Employment 20,204 0.535 0.114 0.487 0 1 Higher Wages 20,204 0.341 0.106 0.463 0 1 Product Variety 20,204 0.382 0.115 0.473 0 1 Lower Prices 20,204 0.341 0.087 0.466 0 1 Access Technology 20,204 0.241 0.081 0.421 0 1 Personal Situation 20,204 0.242 0.082 0.421 0 1 Lower Wages 20,204 0.116 0.050 0.317 0 1 Lower Employment 20,204 0.138 0.070 0.338 0 1 Control 20,204 0.250 0 0.433 0 1 Positive Framing 20,204 0.250 0 0.433 0 1 Negative Framing 20,204 0.250 0 0.433 0 1 Mixed Framing 20,204 0.250 0 0.433 0 1 Age 20,204 40.56 2.640 16.32 16 100 Male 20,204 0.480 0.017 0.499 0 1 Married 20,142 0.526 0.054 0.497 0 1 Catholic 19,984 0.593 0.164 0.465 0 1 Trust 19,628 0.147 0.047 0.351 0 1 Education 20,204 2.480 0.313 0.842 1 4 Media 20,204 1.907 0.463 1.114 0 4 Ideology 16,713 5.045 0.429 2.953 0 10 Employed 20,204 0.596 0.093 0.482 0 1 Income 19,704 2.507 0.262 0.848 1 4 Economy 20,035 2.552 0.356 0.859 1 5 Market 18,600 2.752 0.095 0.743 1 4 GDP per Cap 20,204 1.462 0.629 0 0.50 2.46 Note: See Section A4 in the Online Appendix for full variable definitions and measurement. Statistics computed for the full sample of eighteen countries included in the 2018 Latino- barometro survey. Sample size differs across variables due to incomplete or invalid responses to survey questions.

30 Table 3: Means, by Treatment Status Control Positive Negative Mixed Framing Framing Framing Trade 72.64 73.11 46.33 55.10 Higher Employment 57.85 56.64 49.81 49.65 Higher Wages 36.69 34.73 32.13 33.00 Product Variety 38.53 40.31 35.91 38.15 Lower Prices 32.94 36.90 31.56 34.96 Access Technology 25.66 24.57 22.35 23.98 Personal Situation 25.80 25.12 22.43 23.60 Lower Wages 9.44 10.87 13.05 13.11 Lower Employment 10.67 11.62 16.61 16.37 Obs 5,051 5,051 5,051 5,051 Note: Sample consists of individual-level observations from all eighteen countries in the survey. Table displays percentage of respondents in each treatment status who respond favorably to each question, coding a nonre- sponse as not in favor.

31 Table 4: Covariate Balance Dep Var: Age Male Married Catholic Trust Education Positive Framing .118 -.006 .002 -.001 .005 .009 (.329) (.010) (.010) (.010) (.007) (.018) Negative Framing -.145 -.002 -.004 .001 -.005 .007 (.327) (.010) (.010) (.010) (.007) (.018) Mixed Framing -.073 -.004 -.005 .002 .004 -.014 (.329) (.010) (.010) (.010) (.007) (.018) Countries 18 18 18 18 18 18 Obs 20,204 20,204 20,142 19,984 19,628 20,204 Dep Var: Media Ideology Employed Income Economy Market Positive Framing -.004 .071 .000 .004 -.000 .003 (.024) (.065) (.010) (.018) (.018) (.016) Negative Framing -.015 .054 .016* .019 -.000 -.002 (.024) (.065) (.010) (.018) (.019) (.016) Mixed Framing -.010 .055 -.001 .006 -.003 -.001 (.024) (.066) (.010) (.018) (.019) (.016) Countries 18 18 18 18 18 18 Obs 20,204 16,713 20,204 19,704 20,035 18,600 Note: Each column shows the coeffi cients from a regression of each covariate listed at the top of the column on three treatment indicators and a constant, not reported, without additonal variables or fixed effects. Robust standard errors in parantheses. The sample consists of individual-level observations from the 2018 Latinobarometro survey. * p <0.10.

32 Table 5: Framing Effects, Support for Trade Dep Var: Trade Trade Trade Trade Trade Trade (1) (2) (3) (4) (5) (6) Pos Fr .005 .005 .005 .005 .008 .004 (.009) (.009) (.010) (.010) (.009) (.009) Neg Fr .263*** .263*** .263*** .263*** .260*** .263*** − − − − − − (.009) (.009) (.014) (.011) (.011) (.010) Mix Fr .175*** .175*** .175*** .175*** .174*** .175*** − − − − − − (.009) (.009) (.012) (.011) (.010) (.010) Fixed Effects country region city interviewer block − Countries 18 18 18 18 18 18 Clusters 268 1,064 1,017 2,678 − − Obs 20,204 20,204 20,204 20,204 20,199 20,203 Note: Table reports coeffi cients from regressions of the support for trade indicator on three treatment status indicators and a constant, not reported, or fixed effects as indicated. The regressions do not include additional covariates. Robust standard errors in parentheses, clustered at the level indicated at the bottom of the table. *** p <0.01, ** p <0.05, * p <0.10.

33 Table 6: Framing Effects, Beliefs about Trade Dep Var: High Empl High Wage Variety Prices Tech Pers Sit Low Wage Low Empl (1) (2) (3) (4) (5) (6) (7) (8) Pos Fr .014 .021** .015 .035*** .011 .009 .014** .010* − − − − (.009) (.009) (.009) (.009) (.008) (.008) (.006) (.006) Neg Fr .081*** .046*** .028*** .016* .033*** .034*** .036*** .059*** − − − − − − (.009) (.009) (.009) (.009) (.008) (.008) (.006) (.006) Mix Fr .082*** .037*** .008 .016* .018** .024*** .037*** .057*** − − − − − (.010) (.009) (.009) (.009) (.008) (.008) (.006) (.007) Fixed Effects block block block block block block block block Countries 18 18 18 18 18 18 18 18

34 Clusters 2,678 2,678 2,678 2,678 2,678 2,678 2,678 2,678 Obs 20,203 20,203 20,203 20,203 20,203 20,203 20,203 20,203 Note: Each column reports coeffi cients from a fixed effects regression of the beliefs indicator listed at the top of the column on three treatment status indicators. The regressions do not include additional covariates. Standard errors clustered at the block level in parentheses. *** p <0.01, ** p <0.05, * p <0.10. Table 7: Trade Support and Trade Beliefs Dep Var: Trade (1) (2) (3) (4) (5) (6) (7) (8) High Empl .219*** .211*** .224*** .216*** − − − − (.007) (.010) (.013) (.018) High Wage .031*** .034*** .012 .020 − − − − (.007) (.008) (.013) (.013) Variety .123*** .120*** .133*** .130*** .124*** .120*** .135*** .127*** (.007) (.009) (.013) (.014) (.007) (.009) (.013) (.014) Prices .038*** .042*** .010 .011 .050*** .053*** .030** .030** (.007) (.009) (.013) (.013) (.007) (.008) (.013) (.012)

35 Tech .039*** .047*** .055*** .059*** .052*** .058*** .063*** .063*** (.008) (.011) (.014) (.017) (.008) (.010) (.014) (.015) Pers Sit .023*** .024** .039*** .038*** .043*** .041*** .054*** .049*** (.008) (.010) (.013) (.013) (.008) (.010) (.013) (.012) Low Wage .123*** .121*** .133*** .139*** − − − − − − − − (.011) (.010) (.025) (.024) Low Empl .281*** .274*** .284*** .280*** − − − − − − − − (.011) (.014) (.024) (.026) Fixed Effects country region country region country region country region Sample full full ctrl grp ctrl grp full full ctrl grp ctrl grp Clusters 268 268 268 268 − − − − Obs 20,204 20,204 5,051 5,051 20,204 20,204 5,051 5,051 Note: Each column reports coeffi cients from a fixed effects regression of the support for trade indicator on trade beliefs indicators. The regressions do not include additional covariates. Standard errors in parentheses are robust with country fixed effects, and clustered at the region level for region fixed effects. Sample is full or control group only, as indicated at the bottom of the table. *** p <0.01, ** p <0.05, * p <0.10. Table 8: Framing Effects, by Country Dep Var: Trade ARG BOL BRA CHL COL CRI DOM ECU SLV Pos Fr .061 .039 .103*** .028 .044 .006 .050 .093*** .008 − − − − − (.041) (.036) (.035) (.028) (.053) (.040) (.043) (.035) (.038) Neg Fr .351*** .302*** .460*** .251*** .331*** .100** .024 .153*** .316*** − − − − − − − − (.038) (.040) (.039) (.040) (.043) (.048) (.045) (.039) (.035) Mix Fr .278*** .203*** .231*** .105*** .284*** .042 .070 .130*** .200*** − − − − − − − − (.041) (.038) (.042) (.036) (.044) (.047) (.043) (.036) (.041) Avg Cntrl .587 .747 .674 .760 .670 .660 .672 .797 .760 Clusters 153 240 103 121 33 100 100 150 100

36 Obs 1,200 1,200 1,204 1,200 1,200 1,000 1,000 1,200 1,000 GTM HND MEX NIC PAN PRY PER URY VEN Pos Fr .060 .091*** .100*** .030 .012 .011 .044 .004 .016 − − − (.039) (.031) (.034) (.030) (.040) (.036) (.039) (.027) (.023) Neg Fr .240*** .352*** .277*** .268*** .296*** .197*** .239*** .380*** .205*** − − − − − − − − − (.048) (.041) (.037) (.042) (.045) (.037) (.039) (.033) (.030) Mix Fr .172*** .275*** .257*** .162*** .243*** .184*** .156*** .194*** .076*** − − − − − − − − − (.048) (.038) (.042) (.040) (.046) (.037) (.043) (.033) (.028) Avg Cntrl .668 .864 .643 .856 .680 .733 .587 .860 .870 Clusters 100 100 120 100 100 293 240 306 219 Obs 1,000 1,000 1,200 1,000 1,000 1,200 1,200 1,199 1,200 Note: Table reports for each country coeffi cients from a regression of the support for trade indicator on three treatment status indicators and block fixed effects. The regressions do not include additional covariates. Standard errors in parentheses, clustered at the block level. *** p <0.01, ** p <0.05, * p <0.10. Table 9: Heterogenous Effects Dep Var: Trade Interact Var (Covar): Age Male Married Catholic Trust Education Covar .000 .087*** .014 .005 .043*** .068*** − − − (.000) (.012) (.013) (.013) (.016) (.008) Pos Fr Covar .001 .039** .036* .011 .019 .022** × − − − (.001) (.017) (.019) (.017) (.022) (.010) Neg Fr Covar .001** .015 .053*** .017 .003 .033*** × − (.001) (.018) (.020) (.019) (.028) (.011) Mix Fr Covar .002*** .008 .026 .009 .033 .065*** × − − (.001) (.018) (.019) (.018) (.027) (.011) Clusters 268 268 268 268 268 268 Obs 20,204 20,204 20,142 19,984 19,628 20,204 Dep Var: Trade Interact Var (Covar): Media Ideology Employed Income Economy Market Covar .014** .008*** .012 .047*** .014* .058*** (.006) (.003) (.012) (.008) (.008) (.008) Pos Fr Covar .006 .003 .024 .002 .007 .007 × − − − (.007) (.003) (.018) (.010) (.010) (.011) Neg Fr Covar .013 .004 .039** .011 .015 .005 × − (.009) (.004) (.018) (.011) (.013) (.014) Mix Fr Covar .007 .001 .005 .029*** .007 .019* × − − − (.010) (.003) (.018) (.011) (.011) (.011) Clusters 268 268 268 268 268 268 Obs 20,204 16,713 20,204 19,704 20,035 18,600 Note: Table reports for each of twelve covariates coeffi cients from a regression of the support for trade indicator on the covariate, three interaction variables between the treatment indicators and the covariate, reported, three treatment status indicators, not reported, and region fixed effects. The regressions do not include additional covariates. Standard errors in parentheses, clustered at the region level. *** p <0.01, ** p <0.05, * p <0.10.

37 Table 10a: Control Group: Education and Trade Preferences Dep Var: Trade Trade Trade Trade Trade (1) (2) (3) (4) (5) Education .075*** .041** .042* .044** .042* (.007) (.018) (.023) (.022) (.022) Education GDP .023* .020 .019 .018 × − (.012) (.015) (.014) (.014) Fixed Effects country country region region region Controls no no no yes yes Clusters 268 268 268 − − Obs 5,051 5,051 5,051 5,038 4,867 Note: Sample consists of control group only. Table reports coeffi cients from a regression of the support for trade indicator on education, education interacted with GDP per capita, fixed effects, and controls where indicated. Controls are Age, Male, Married in column (4), and Age, Male, Married, Employed, Income, Economy in column (5). Standard errors in parentheses, robust in columns (1)-(2), clustered at the region level in columns (3)-(5). *** p <0.01, ** p <0.05, * p <0.10.

Table 10b: Mixed Framing Group: Education and Trade Preferences Dep Var: Trade Trade Trade Trade Trade (1) (2) (3) (4) (5) Education .000 .057*** .059*** .054*** .054** − − − − − (.008) (.022) (.020) (.021) (.021) Education GDP .039*** .036*** .036*** .034** × − (.014) (.013) (.013) (.014) Fixed Effects country country region region region Controls no no no yes yes Clusters 267 267 267 − − Obs 5,051 5,051 5,050 5,030 4,883 Note: Sample consists of mixed framing group only. Table reports coeffi cients from a regression of the support for trade indicator on education, education interacted with GDP per capita, fixed effects, and controls where indicated. Controls are Age, Male, Married in column (4), and Age, Male, Married, Employed, Income, Economy in column (5). Standard errors in parentheses, robust in columns (1)-(2), clustered at the region level in columns (3)-(5). *** p <0.01, ** p <0.05, * p <0.10.

38 Table 11: Framing Sensitivity of Trade Support Dep Var: Trade Trade Trade Trade Trade (1) (2) (3) (4) (5) Educ .069*** .070*** .065*** .066*** .064*** (.007) (.007) (.009) (.009) (.008) Mix Fr .019 .063 .016 .054 .040 − − (.027) (.055) (.029) (.054) (.057) Mix Fr Educ .063*** .111*** .064*** .106*** .102*** × − − − − − (.010) (.022) (.011) (.021) (.022) Mix Fr GDP .056 .048 .042 × − − − − − (.036) (.034) (.035) Mix Fr Educ GDP .032** .028** .027** × × − − (.013) (.013) (.013) Age .001*** − − − − (.000) Male .091*** − − − − (.009) Married .002 − − − − (.011) Employed .027*** − − − − − (.009) Income .022*** − − − − (.006) Economy .014** − − − − (.007) Fixed Effects country country region region region Clusters 268 268 268 − − Obs 10,102 10,102 10,102 10,102 9,751 Note: Sample consists of control and mixed framing groups only. Standard errors in parentheses, robust in columns (1)-(2), clustered at the region level in columns (3)-(5). *** p <0.01, ** p <0.05, * p <0.10.

39 Table 12: Framing Sensitivity of Trade Beliefs Dep Var: High Empl High Empl High Empl High Empl High Empl (1) (2) (3) (4) (5) Educ .036*** .036*** .032*** .033*** .022** (.008) (.008) (.008) (.008) (.009) Mix Fr .011 .055 .009 .056 .071 − − (.028) (.057) (.031) (.064) (.065) Mix Fr Educ .029*** .070*** .029** .071*** .076*** × − − − − − (.011) (.023) (.011) (.022) (.022) Mix Fr GDP .044 .044 .052 × − − − − − (.037) (.044) (.044) Mix Fr Educ GDP .027** .028** .031** × × − − (.014) (.014) (.014) Dep Var: Low Empl Low Empl Low Empl Low Empl Low Empl (6) (7) (8) (9) (10) Educ .004 .004 .006 .005 .005 (.005) (.005) (.005) (.005) (.006) Mix Fr .001 .081** .003 .086** .093** − − − (.019) (.037) (.019) (.035) (.036) Mix Fr Educ .023*** .068*** .022** .070*** .072*** × (.007) (.015) (.009) (.018) (.018) Mix Fr GDP .057** .062** .069*** × − − (.025) (.025) (.026) Mix Fr Educ GDP .031*** .032*** .035*** × × − − − − − (.010) (.012) (.012) Fixed Effects country country region region region Controls no no no no yes Clusters 268 268 268 − − Obs 10,102 10,102 10,102 10,102 9,751 Note: Sample consists of control and mixed framing groups only. Standard errors in parentheses, robust in columns (1)-(2),(6)-(7) clustered at the region level in columns (3)-(5),(8)-(10). Controls are Age, Male, Married, Employed, Income, Economy in columns (5) and (10). *** p <0.01, ** p <0.05, * p <0.10.

40 References

Ardanaz, M.M., V. Murillo, and P.M. Pinto (2013) "Sensitivity to Issue Framing on Trade Policy Preferences: Evidence from a Survey Experiment," International Organization 67(2): 411—37. Alesina, A. and E. La Ferrara (2005) "Preferences For Redistribution In The Land Of Opportunities," Journal of Public Economics 89(5-6): 897-931. Autor, D., D. Dorn, G. Hanson (2013) "The China Syndrome: Local Labor Market Effects of Import Competition in the United States," American Economic Review 103(6): 2121-68. Autor, D., D. Dorn, G. Hanson, and K. Majlesi (2016) "Importing Political Polarization? The Electoral Consequences of Rising Trade Exposure," NBER wp 22637. Baker, A. (2003) "Why is Trade Reform So Popular in Latin America? A Consumption- Based Theory of Trade Policy Preferences," World Politics 55(3): 423-455. Balistreri, E.J. (1997) "The Performance of the Heckscher—Ohlin—Vanek Model in Pre- dicting Endogenous Policy Forces at the Individual Level," Canadian Journal of Economics 30(1): 1—17. Beaulieu, E. (2002) "Factor or Industry Cleavages in Trade Policy? An Empirical Analy- sis of the Stolper—Samuelson Theorem" Economics and Politics 14(2): 99-131. Blonigen, B. (2011) "Revisiting the Evidence on Trade Policy Preferences," Journal of International Economics 85: 129-135. Che, Y., Y. Lu, J.R. Pierce, P.K. Schott, and Z. Tao (2016) "Does Trade Liberalization with China influence US Elections?" NBER wp 22178. Di Tella, R. and D. Rodrik (2019) "Labor Market Shocks and the Demand for Trade Protection: Evidence from Online Surveys," working paper, Harvard University. Dix-Carneiro, R. and B.K. Kovak (2017) "Trade Liberalization and Regional Dynamics," American Economic Review 107(10): 2908-46. Goldberg, P.K. and N. Pavcnik (2007) "Distributional Effects of Globalization in Devel- oping Countries," Journal of Economic Literature 45(1): 39-82. Grossman, G.M. and E. Helpman (1994) "Protection for Sale," American Economic Re- view 84(4): 833-850. Grossman, G.M. and E. Helpman (2018) "Identity Politics and Trade Policy," NBER wp 25348. Hiscox, M. (2006) "Through a Glass and Darkly: Attitudes Toward International Trade

41 and the Curious Effects of Issue Framing," International Organization 60(3): 755—80. Jakel, I.C. and M. Smolka (2017) "Trade Policy Preferences and Factor Abundance," Journal of International Economics 106: 1-19. Kahneman, D. and A. Tversky (1979) " Prospect Theory: An Analysis of Decision under Risk," Econometrica 47(2): 263—291. Luttmer, E.F.P. (2001) "Group Loyalty and the Taste for Redistribution," Journal of Political Economy 109 (3): 500—528. Mansfield, E.D. and D.C. Mutz (2009) "Support for Free Trade: Self-Interest, Sociotropic Politics, and Out-Group Anxiety," International Organization 63(3): 425—57. Mayda, A.M. (2008) "Why Are People More Pro-Trade Than Pro-Migration?" Economics Letters 101(3): 160-63. Mayda, A.M. and D. Rodrik (2005) "Why Are Some People (And Countries) More Pro- tectionist Than Others?" European Economic Review 49(6): 1393-1430. McLaren, J. (2016) "The Political Economy of Commercial Policy," in K. Bagwell, R.W. Staiger (eds.) Handbook of Commercial Policy, Vol 1(A): 109-159. Melitz, M.J. (2003) "The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry Productivity," Econometrica 71: 1695-1725. Mukand, S. and D. Rodrik (2018) "The Political Economy of Ideas: On Ideas Versus Interests in Policymaking," NBER wp 24467. Nelson, T.E., Z.M. Oxley, and R.A. Clawson (1997) "Toward a Psychology of Framing Effects," Political Behavior 19(3): 221-246. O’Rourke, K.H. and R. Sinnott (2001) "What Determines Attitudes Towards Protection? Some Cross-Country Evidence," in: Collins, S.M. and D. Rodrik (Eds.) Brookings Trade Forum, Brookings Institute Press, Washington DC: 157—206. Pavcnik, N. (2017) "The Impact of Trade on Inequality in Developing Countries," NBER wp 23878. Rho, S. and M. Tomz (2017) "Why Don’t Trade Preferences Reflect Economic Self- Interest?" International Organization 71(Suppl): S85-S108. Rodrik, D. (1995) "Political Economy of Trade Policy," in Grossman, G.M. and K. Rogoff (Eds.) Handbook of International Economics, Vol. 3, Elsevier Science Publishers, Amster- dam: 1457—1494. Scheve, K. and M. Slaughter (2001) "What Determines Individual Trade Policy Prefer- ences?" Journal of International Economics 54(2): 267-292.

42 Shafir, E., I. Simonson, and A. Tversky (1993) "Reason-Based Choice," Cognition 49: 11-36. Skowronski, J.J. and D.E. Carlston (1989) "Negativity and Extremity Biases in Impres- sion Formation: A Review of Explanations," Psychological Bulletin 105(1): 131-142. Stolper, W.F. and P.A. Samuelson (1941) "Protection and Real Wages," Review of Eco- nomic Studies 9(1): 58-73. Verhoogen, E.A. (2008) "Trade, Quality Upgrading, and Wage Inequality in the Mexican Manufacturing Sector," Quarterly Journal of Economics 123(2): 489-530.

43 Online Appendix (Not for Publication)

A1. Figures

Figure A1: Higher Employment, Raw Means

Cntrl Pos Fr Neg Fr Mix Fr 8 . 7 . l p m E h g 6 i . H n a e M 5 . 4 .

Cntrl Pos Fr Tr eat m ent Neg Fr Mix Fr 1 9 . 8 . 7 . l p m 6 . E h g 5 i . H n a 4 . e M 3 . 2 . 1 . 0 I L L L Y V Y A X D U R N N C G M M R I L H O O E R R E R A E N C T R O C N S P C B P P B C V U E H A M G D

a x is

Note: Sample size is 20,204 respondents from eighteen countries. For country sample sizes see Table 1. Ranges around conditional means are 95 percent confidence intervals.

i Figure A2: Covariate Balance by Country

Argentina Bolivia Brazil Chile Colombia Costa Rica

Age Male Married Catholic Trust Education Media Ideology Employed Income Economy Market

0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 Dominican Rep Ecuador El Salvador Guatemala Honduras Mexico

Age Male Married Catholic Trust Education Media ii Ideology Employed Income Economy Market

0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 Nicaragua Panama Paraguay Peru Uruguay Venezuela

Age Male Married Catholic Trust Education Media Ideology Employed Income Economy Market

0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1

Positive Framing Negative Framing Mixed Framing

Note: Each country plot presents three p-values per covariate corresponding to significance tests of the three treatment indicators. Figure A3: Trade-Education Correlations

Pos Fr Group (N = 5,051) 3 .

PER 2 . ECU BRA GTM BOL MEX PAN ) NIC VEN ARG URY CHL c 1 .

u CRI d

E PRY

, HND SLV DOM e d 0 a r

T COL ( r r o 1 . • C 2 . • 3 . • .5 1 1.5 2 2.5 Per Capita GDP, PPP, $10K

Neg Fr Group (N = 5,051) 3 . 2 . PER CHL GTM MEXARG ) c 1

. URY u

d HND E

, BOLSLV VEN COL CRIDOM e ECU PRY

d BRA 0 PAN a r T ( r r o 1 . • C 2

. NIC • 3 . • .5 1 1.5 2 2.5 Per Capita GDP, PPP, $10K

Note: Figure plots country-level correlations between a respondent’ssupport for trade and their education, against country GDP per capita, separately for the positive framing group (top) and the negative framing group (bottom).

iii Figure A4: Trade Beliefs and Framing Sensitivity

Higher Wages Access Technology 9 9 . Cntrl . Cntrl 8 Mix Fr 8 Mix Fr . . 7 7 . . 6 e 6 . . g a h c W 5 5 e . . h T g i n a H 4 4 . . e n a M e 3 3 M . . 2 2 . . 1 1 . . 0 0

1 2 3 4 1 2 3 4 Education Education

Lower Wages Personal Situation 9 9 . Cntrl . Cntrl 8

Mix Fr 8

. Mix Fr . 7 7 . . 6 6 e . . t g i a S s W 5 5 r . . e w o P L n 4 4 . . n a a e e M 3 3 M . . 2 2 . . 1 1 . . 0 0

1 2 3 4 1 2 3 4 Education Education

Note: Figures plot the means of each of four variables measuring beliefs about trade, by level of education, separately for the control and mixed framing groups. Ranges around conditional means are 95 percent confidence intervals.

iv A2. Tables Table A1a: Heterogeneous Effects for Trade Beliefs Dep Var: High Empl Intr Var (Covar): Age Male Married Catholic Trust Educ Covar .002*** .044*** .003 .016 .002 .031*** − − − − (.001) (.014) (.012) (.014) (.023) (.008) Dep Var: High Wage Covar .000 .016 .009 .018 .005 .002 − − − (.000) (.012) (.013) (.013) (.019) (.007) Dep Var: Variety Covar .001*** .033** .006 .003 .018 .072*** − − (.000) (.014) (.012) (.015) (.019) (.008) Dep Var: Prices Covar .001** .000 .023* .009 .003 .028*** − − − (.000) (.013) (.012) (.014) (.020) (.007) Dep Var: Tech Covar .001* .048*** .011 .011 .047*** .068*** − − − (.000) (.013) (.012) (.011) (.017) (.010) Dep Var: Pers Sit Covar .002*** .012 .011 .022* .004 .036*** − − − − (.000) (.013) (.012) (.013) (.019) (.007) Dep Var: Low Wage Covar .000 .008 .012 .009 .004 .001 − − − (.000) (.009) (.008) (.007) (.013) (.005) Dep Var: Low Empl Covar .000 .021** .013 .011 .025** .006 − − − (.000) (.009) (.008) (.008) (.012) (.005) Clusters 268 268 268 268 268 268 Obs 20,204 20,204 20,142 19,984 19,628 20,204 Note: Table reports for each of the six covariates listed at the top of the columns, the covari- ate’s coeffi cient from a regression of the dependent variable on the covariate, three treatment status indicators, not reported, three interaction variables between the treatment indicators and the covariate, not reported, and region fixed effects. The regressions do not include additional covariates. See also Table 9. Standard errors in parentheses, clustered at the region level. *** p <0.01, ** p <0.05, * p <0.10.

v Table A1b: Heterogeneous Effects for Trade Beliefs Dep Var: High Empl Intr Var (Covar): Media Ideology Employed Income Economy Market Covar .024*** .004 .025* .016** .021** .037*** (.006) (.003) (.014) (.007) (.008) (.009) Dep Var: High Wage Covar .025*** .005** .012 .016* .008 .039*** − (.006) (.002) (.016) (.008) (.007) (.010) Dep Var: Variety Covar .058*** .005** .011 .015* .012 .013 − (.007) (.002) (.015) (.009) (.009) (.009) Dep Var: Prices Covar .049*** .005* .000 .010 .013* .015* − − − (.006) (.002) (.012) (.008) (.007) (.009) Dep Var: Tech Covar .056*** .004 .019 .025*** .002 .005 (.005) (.002) (.014) (.008) (.008) (.011) Dep Var: Pers Sit Covar .044*** .006** .006 .000 .001 .018** (.006) (.002) (.013) (.008) (.007) (.009) Dep Var: Low Wage Covar .006* .002 .006 .005 .005 .013** − − − − (.003) (.001) (.008) (.005) (.004) (.006) Dep Var: Low Empl Covar .011*** .003* .014* .012* .007 .035*** − − − − (.004) (.002) (.008) (.007) (.005) (.007) Clusters 268 268 268 268 268 268 Obs 20,204 16,713 20,204 19,704 20,035 18,600 Note: Table reports for each of the six covariates listed at the top of the columns, the covariate’s coeffi cient from a regression of the dependent variable on the covariate, three treatment status indi- cators, not reported, three interaction variables between the treatment indicators and the covariate, not reported, and region fixed effects. The regressions do not include additional covariates. See also Table 9. Standard errors in parentheses, clustered at the region level. *** p <0.01, ** p <0.05, * p <0.10.

vi Table A2a: Positive Framing Group: Education and Trade Preferences Trade Trade Trade Trade Trade (1) (2) (3) (4) (5) Education .057*** .047** .036* .042** .037* (.007) (.019) (.020) (.020) (.021) Education GDP .007 .010 .010 .007 × − (.013) (.013) (.014) (.014) Fixed Effects country country region region region Controls no no no yes yes Clusters 268 268 268 − − Obs 5,051 5,051 5,051 5,040 4,886 Note: Sample consists of positive framing group only. Table reports coeffi cients from a regression of the trade support indicator on education, education interacted with GDP per capita, fixed effects, and controls where indicated. Controls are Age, Male, Married in column (4), and Age, Male, Married, Employed, Income, Economy in column (5). Standard errors in parentheses, robust in columns (1)-(2), clustered at the region level in columns (3)-(5). *** p <0.01, ** p <0.05, * p <0.10.

Table A2b: Negative Framing Group: Education and Trade Preferences Trade Trade Trade Trade Trade (1) (2) (3) (4) (5) Education .032*** .009 .012 .008 .014 − − − − (.008) (.022) (.026) (.026) (.026) Education GDP .029** .029* .032* .029* × − (.014) (.017) (.017) (.017) Fixed Effects country country region region region Controls no no no yes yes Clusters 268 268 267 − − Obs 5,051 5,051 5,051 5,033 4,876 Note: Sample consists of negative framing group only. Table reports coeffi cients from a regression of the trade support indicator on education, education interacted with GDP per capita, fixed effects, and controls where indicated. Controls are Age, Male, Married in column (4), and Age, Male, Married, Employed, Income, Economy in column (5). Standard errors in parentheses, robust in columns (1)-(2), clustered at the region level in columns (3)-(5). *** p <0.01, ** p <0.05, * p <0.10.

vii A3. Survey Questions in Spanish

Below is the original transcript of the trade-related questions used by Latinobarometro survey enumerators. The full survey transcript is available at www.latinobarometro.org.

viii A4. Variables: Definitions and Sources

Below is the complete list of variables used in the paper with details on measurement and sources. The variables appear in the order of Table 2 of summary statistics. Trade Support: Indicator variable that takes the value one if the individual reports being in favor of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question Question P27ND. Higher Employment: Indicator variable that takes the value one if the individual reportss that higher employment is a consequence of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P28ND.A. Higher Wages: Indicator variable that takes the value one if the individual reports that higher wages are a consequence of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P28ND.B. Product Variety: Indicator variable that takes the value one if the individual reports that more product variety is a consequence of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P28ND.C. Lower Prices: Indicator variable that takes the value one if the individual reports that lower prices are a consequence of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P28ND.D. Access Technology: Indicator variable that takes the value one if the individual an- swers that more access to technology is a consequence of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P28ND.E. Personal Situation: Indicator variable that takes the value one if the individual answers that a better personal economic situation is a consequence of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P28ND.F. Lower Wages: Indicator variable that takes the value one if the individual answers that lower wages are a consequence of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P28ND.G. Lower Employment: Indicator variable that takes the value one if the individual an- swers that lower employment is a consequence of trade, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P28ND.H. Control: Indicator variable that takes the value one if the individual was assigned to the control group, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P27ND.C.

ix Positive Framing: Indicator variable that takes the value one if the individual was assigned to the positive framing group, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P27ND.A. Negative Framing: Indicator variable that takes the value one if the individual was assigned to the negative framing group, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P27ND.D. Mixed Framing: Indicator variable that takes the value one if the individual was assigned to the mixed framing group, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P27ND.B. Age: Integer variable recrding the age reported by the individual in the survey. Scale: 16,17,...,100. Source: Latinobarometro (2018) Question EDAD. Male: Indicator variable that takes the value one if the individual reports being a male, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question SEXO. Married: Indicator variable that takes the value one if the individual reports being married, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question S23. Catholic: Indicator variable that takes the value one if the individual reports being catholic, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question S5. Trust: Indicator variable that takes the value one if the individual reports a high level of trust in other people, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question P11STGBS. Education: Categorical variable that groups individuals’ reported years of education into four categories: 0-4 years, 5-10 years, 11-12 years or incomplete college or technical school, and completed college or technical school. Scale: 1,2,...,4. Source: Latinobarometro (2018) Question S10. Media: Categorical variable that counts the total number of channels through which the individual obtains political information. The media channels included are radio, newspapers, social media, and television. Scale: 0,1,...,4. Source: Latinobarometro (2018) Questions P19ST.E, P19ST.F, P19ST.G, P19ST.H. Ideology: Categorical variable showing an individual’s self-placement on the left-right political spectrum where zero (0) is left and ten (10) is right. Scale: 0,1,...,10. Source: Latinobarometro (2018) Question P22ST. Employed: Indicator variable that takes the value one if the individual reports being employed, zero otherwise. Scale: 0,1. Source: Latinobarometro (2018) Question S14A. Income: Categorical variable measuring the level of income of the individual. The lowest

x value represents incomes insuffi cient for basic needs, the highest value represents incomes that allow saving. Scale: 1,2,...,4. Source: Latinobarometro (2018) Question S4. Economy: Categorical variable measuring the respondent’s perception of the current economic situation of the country. Low values represent a very bad perception of economic performance, and high values represent a very good perception. Scale: 1,2,...,5. Source: Latinobarometro (2018) Question P6STGBSC. Market: Categorical variable measuring the respondent’sagreement with the claim that a free market is the only system that can promote development in the respondent’scountry. Low values represent disagreement, and high values represent agreement. Scale: 1,2,...,4. Source: Latinobarometro (2018) Question P26ST. GDP per Cap: Continuous variable measuring Gross Domestic Product per capita in PPP current dollars, ten thousands for the year 2017. Scale: [0,2.5]. Source: IMF World Economic Outlook (2018).

xi