Sample

We use data from Round 8 of the ESS, which includes eighteen countries. Donald Trump's election fell inside the survey fieldwork period for thirteen of them. The fieldwork periods, detailed in the Appendix, typically lasted three to four months. The survey is constructed using highly rigorous translation protocols and conditional monetary incentives are granted to respondents upon the completion of face-to-face interviews.

Empirical Model

To identify racially biased attitudes, we use questions B38 and B39 of the ESS. Question B38 reads, ‘To what extent do you think the country should allow people of the same race or ethnic group as most people of the country to come and live here?’ Question B39 asks, ‘How about people of a different race or ethnic group from most people?’ Answers to both questions range from (1) ‘Allow many’ to (4) ‘Allow none’.

Because the two questions only differ in the race dimension, the differences in answers can only be driven by differences in the perception of migrants according to their race. By giving different answers to the two questions, respondents therefore knowingly reveal a racial bias. Most of all, because interviews are conducted face to face, respondents are subject to a stronger social pressure than in internet surveys, where racially biased opinions can be revealed anonymously (Seth Reference Seth2013).

Denoting respondent i's opposition to different- and same-race immigration by y _1i and y _2i, respectively, the dependent variable Self-reported Racial Bias in Policy Attitudes, y _i, is defined as a dummy variable taking a value of 1 if y _1i >y _2i and 0 otherwise. In the relevant sample, 32.70 per cent of individuals report stronger opposition to different-race immigration than same-race immigration, while 64.80 per cent report equal opposition to different-race immigrants. While such operationalization of the dependent variable has the advantage of simplicity, we discuss its limits and test the robustness of our results to using two alternatives in the Appendix. Reports of racist attitudes (y _i = 1) account for 31.49 per cent in the control group and 34.06 per cent in the treatment group.Footnote ¹

Defining $Y_{i,c} = \ln [(\Pr (y_{i,c} = 1))/(1-\Pr (y_{i,c} = 1))]$, we use the following specification:

$$Y_{i,c} = \alpha + \beta T_i + {\gamma} ^{\prime}\,{\boldsymbol X}_{i,c} + \mu _c + \epsilon _{i,c}$$

T _i is the treatment variable. It takes a value of 1 if respondent i was interviewed after 8 November 2016 (election day) and 0 otherwise. Even though asymmetric levels of respondents' reachability as well as geographic imbalance may induce a non-random selection of respondents among the control and treatment groups (Legewie Reference Legewie2013; Munoz, Falco-Gimeno and Hernandez Reference Munoz, Falco-Gimeno and Hernandez2018), the timing of each interview is as good as random with respect to the timing of the US election. T _i can therefore be interpreted as an exogenous signal of the decrease in the social desirability of reporting racially neutral attitudes. Accordingly, measures the effect of Donald Trump's win on the propensity to report racist attitudes, α, is a constant.

X_i,c summarizes individual-level characteristics. In a first model, we only control for demographic characteristics including age, age squared, sex, household status (having at least one child living at home) and ethnic minority status (0 if majority, 1 if minority). We then add socioeconomic characteristics: highest educational attainment (1–7), a dummy capturing economic insecurity (0 if the respondent experienced short-run unemployment during the previous year, and 1 otherwise) and household income (1 = living comfortably; 4 = living with strong difficulties). We subsequently add a dummy equal to 1 if the respondent voted in the latest general election to capture interest in politics. Note that we only select proper covariates – those that could not be affected by the treatment. To control for unobserved country heterogeneity, we include country fixed effects μ _c. ε _i,c is an idiosyncratic error term, with:

$$E\left[ {\epsilon \vert T,{\boldsymbol X},\mu } \right] = 0.$$

Finally, we weight observations by the design weights provided by the ESS to control for the relative likelihood of each observation to be sampled.

Respondents in the treatment and control groups may differ in the distribution of key covariates. While the ESS is meant to be representative of each country's population in the overall period, there is no guarantee that representativeness holds within particular sub-periods, for instance because of reachability issues. Following Hainmueller (Reference Hainmueller2012), we therefore weight control units such that the distribution of covariates in the control group matches the moment conditions (until skewness) of the treatment group. After this pre-processing, covariate imbalance between control and treatment groups becomes negligible.

We fit the model with a binary logit estimator and report the average marginal effect. Since both the treatment and the output variables are dummies, the marginal effect is easy to interpret: it provides the difference in percentage points between the treated and control groups in the probability that a respondent will exhibit a racial bias. We base our main analysis on an interval of ±15 days before and after the election. This bandwidth choice reflects a trade-off between statistical power, which is greater the larger the bandwidth, and attribution, which is more accurate the smaller the bandwidth. In the Appendix we discuss this rather arbitrary choice in greater detail and report treatment effects for alternative bandwidths.

The ideal dataset to study a ‘global contagion’ should include the race-related attitudes of each individual in each country both before and after the election. Instead, we had to run the analysis on a sample of thirteen countries. We therefore face sample uncertainty (Abadie et al. Reference Abadie2017). Moreover, race-related reports are observed either before or after the election. Consequently, we also face design uncertainty (Abadie et al. Reference Abadie2017). For these reasons, we cluster errors at the country level.