Arguments for a Positive Rally Effect

The first plausible effect of Trump’s surprise victory is a higher popularity of the EU in Europe. The central mechanism behind such a positive impact could be a rally effect. The term, in its full notation—“rally-’round-the-flag effect”—was originally used to describe rises in the U.S. president’s popularity in the wake of international crises.^{Footnote 27} By now, scholars have already uncoupled the “rally effect” from this initially strict focus and replaced the popularity of the U.S. president with trust in government, the ruling party, other leaders, or general public opinion in the United States and other countries as the dependent variable.^{Footnote 28} It is possible to dissociate this idea further from its narrow original context to make the general argument that a perceived external threat can bring members of a social entity to unite. A similar proposition can be found in Karl Deutsch’s transactionalist theory of integration.^{Footnote 29} Admittedly, Deutsch and colleagues speak of unifying effects of an external military threat, and in present days Europeans certainly do not fear a direct attack by the United States against Europe. However, many Europeans are indeed worried about the unpredictability of Donald Trump as the commander-in-chief of the world’s largest military force.^{Footnote 30} Yet more importantly, no reason is immediately apparent why this mechanism should not work similarly for sociopolitical threats more generally. In the present case, the sudden election of Trump as an American nationalist could be subjectively experienced to pose an external threat to the stability and prosperity of Europe. As mentioned earlier, observers suspected that his election would affect transatlantic relations and prospects of European integration. Such concerns were fueled prior to the election by Trump’s nationalist credo “America first,” condemnatory statements about NATO and TTIP, and consentient comments regarding the looming partial breakup of the EU in the wake of Brexit.^{Footnote 31} One fear was thus the need to reinvent Europe’s role in the world upon realizing that the United States is no longer a reliable partner. As stated by Angela Merkel in May 2017 after meeting with Trump, Europe “really must take our fate into our own hands.”^{Footnote 32} Thus, it is possible that already in immediate response to Trump’s unexpected victory, Europeans rallied around the European “flag”—creating a feeling of unity that could be measurable through an increased popularity of the EU.

Notably, Hannah Arendt made a similar argument in the first half of the twentieth century, stating, “If it is true that each nationalism . . . begins with a real or fabricated common enemy, then the current image of America in Europe may well become the beginning of a new pan-European nationalism.”^{Footnote 33} She evaluated this “anti-American Europeanism”^{Footnote 34} negatively, as nationalistic with ties to fascism and in opposition to a liberal federalism. Hence, a unifying effect through an external threat may work for liberal cosmopolitans just as for conservative nationalists. The former may move toward increased EU support, seeing it as a stronghold of an open, post-nationalist world, while the latter may embrace Europe as a bastion in a world of strong and nationalistic regional powers, such as the United States, China, and Russia.

More recent research has also looked deeper into the question of who is most susceptible to rally effects. Baum argues that while different social groups may have varying propensities to engage in rallying, individuals who are closest to the point of ambivalence between approval and disapproval on the issue in question are most likely to change their opinion.^{Footnote 35} Colaresi suggests that rally effects do not require an emotional or irrational public but can be modeled as a rational response to international crises.^{Footnote 36} Baker and Oneal additionally find that the size of a rally effect is influenced by how the media covers the event in question, potentially making differences in media consumption a significant factor.^{Footnote 37}

Recently, it has been argued that Brexit—a victory for right-wing populism in many ways comparable to Trump’s election—may be responsible for a rise in Eurobarometer respondents who say they “feel like citizens of the EU.”^{Footnote 38} The adequacy of this claim has not yet been corroborated through a rigorous causal analysis. But if the assessment is correct, it could suggest that a similar positive effect, in which Europeans rally around “their flag” in defiance of nationalist sentiment, may be at work in the case of Trump. We can thus formulate as a first hypothesis that

At the same time, however, arguments for the opposite case, a negative domino effect, have been made.

Arguments for a Negative Domino Effect

Another plausible outcome is a decrease in the EU’s popularity in the wake of Trump’s surprise victory. The underlying mechanism could be a domino effect in which the United States’ “fall” for right-wing populism constitutes the start of a chain reaction in which other countries—e.g., in Europe—successively “tilt over,” resulting in rising levels of nationalism and anti-supranationalism. Although domino theory originated in the cold war context of countries supposedly acceding consecutively to communism, it has been applied to other contexts, such as democratization,^{Footnote 39} regionalism,^{Footnote 40} and, most recently, populism after Brexit.^{Footnote 41} Thus, it does not appear far-fetched to apply domino theory to the spread of right-wing populism (and an according decline of supranationalism) after Trump’s success. In the present case, the domino effect could mean lower popularity levels for the EU due to a combination of resignation among cosmopolitans and reinforcement of nationalists in their beliefs.

In Europe, fears of such a domino effect were visible in public discourses prior to the national elections in Austria (December 2016), the Netherlands (March 2017), France (April-May 2017), and Germany (September 2017). Many observers anticipated significant political shifts to the right, often naming Trump’s victory as one ground for their expectance. Mudde, for instance, argued, writing immediately after the U.S. election, that “the surprise win of Donald Trump is a gift from heaven for the far right around the globe.” Their victory in the United States, he suspected, “gives them a narrative of hope and success.”^{Footnote 42} New impetus to right-wing populism should go hand in hand with increased Euroscepticism^{Footnote 43} and thus a less popular EU. Research shows that this link has become even stronger during recent years.^{Footnote 44} Thus, a competing hypothesis to H₁ would be that

To some extent, however, converse trends became observable in the months after the U.S. presidential election, making concerns over a domino effect appear somewhat exaggerated in hindsight.^{Footnote 45} Yet it is still unclear which form the Trump effect on the EU’s popularity took immediately after the election. Was it a positive rally effect, a negative domino effect—or rather no effect at all?

Arguments for a Non-Effect

A non-effect could simply be the outcome of no effect, if

A more complex “non-effect” could result from positive effects and negative effects (partially) offsetting each other. To arrive at an understanding of how such a canceling out could come about, it is necessary to compare different social subgroups. We look at two^{Footnote 46} specific potentially relevant societal divides: first, subgroups defined by the (perceived) economic situation of the respondents’ country. Here, the way Europeans rate the economic situation of their country serves as a proxy for the degree to which they feel their country has been affected by the economic crises in the EU. The underlying idea is that in economically struggling countries, the EU will likely have low initial popularity levels but potentially high upward momentum, whereas in economically thriving countries that are hardly affected by the crises, the EU likely has comparatively high initial popularity but low potential for upward momentum (ceiling effect). We thus expect that

Second, we look at Europeans’ political orientation. As discussed in the preceding sections, people with different political orientations could react differently to Trump’s surprise victory. It is plausible that the EU has high initial popularity levels among the political center and lower ones among the left and the right.^{Footnote 47} This could again be connected to little upward momentum among the center (which likely tends to be very much in favor of the EU already prior to the election) whereas the right and the left may be more susceptible to changing their views on the EU in response to the unexpected coming to power of a right-wing nationalist in the United States (cf. the argument by Baum presented earlier). Dissecting the Trump effect on the EU’s popularity by political orientation will shed light on these potentially diverging effects and to test whether they (partially) cancel each other out. We thus assume that

Solving the Fundamental Problem of Causal Inference

In order to estimate the causal effect of a treatment T, a study setup is required that allows the same (or very similar) observations to be exposed to treatment and control simultaneously.^{Footnote 48} However, observing treatment and control outcomes for the same observation at the same time, that is, the counterfactuals of each state, is impossible. To be able to distinguish correlation from causation despite this “fundamental problem of causal inference,”^{Footnote 49} a few conditions must be fulfilled. First, the assumption of independence must be satisfied, ensuring that differences in outcome between treatment and control group are only due to the treatment. One possibility to meet this assumption is randomization. If the treatment is assigned randomly to participants, no observable or unobservable factors can bias potential outcomes. While randomization is easily achieved in laboratory experiments given that assignment to treatment can be manipulated by the researcher, studies based on observational data rarely meet this strict assumption. Sometimes, however, randomization-like conditions occur naturally, allowing researchers to draw causal inferences based on observational data. In such natural experiments, the treatment is randomized not through manipulation by the researcher, but by an event that is exogenous to the outcome in question. Thus, in natural experiments, individuals are exposed to the treatment as-if random.^{Footnote 50}

Our research design meets the as-if random criterion of a natural experiment and thus satisfies the independence assumption: The U.S. presidential election took place between the sixth and the seventh day of the Eurobarometer’s 12-day fieldwork (see figure 2). Being assigned to an interview date before or after the presidential election was random, since it did not depend on the respondents’ political preferences, socio-economic characteristics, or other observable or unobservable confounders. Thereby, the outcome (change in the popularity of the EU among respondents) should be related to nothing but the assignment to treatment, namely, being interviewed before or after the election.

Table 1 shows that this as-if random criterion actually holds empirically. It provides a summary of descriptive statistics on a range of widely used socio-economic and demographic variables in our sample, including age, gender, education in years (top-coded at 26), occupation, and a dichotomous variable indicating whether the individual lives in the countryside or in a town. As can be seen from the Δ means column, the sample is mostly well balanced, with no significant differences in education, gender, or place of living between the treatment and the control group. For age and some occupational classes, minor differences can be found. For instance, the mean age in the treatment group is 49.2 years, compared to 50.8 in the control group. To correct for these small divergences, one of our models controls for the variables depicted in table 1 (see equation 2). Furthermore, robustness checks were run in which older respondents were excluded (leading to an entirely balanced sample), and in which the potential influence of between-country differences in the distribution of fieldwork across time (see figure 2) were tested. All robustness checks confirm the validity of our conclusions.^{Footnote 51}

Table 1 Descriptive statistics

Note: Based on Eurobarometer 86.2, own calculations, not weighted;

SD=Standard deviation, education was top-coded at 26 years to reduce the influence of outliers.

* p < 0.05, ** p < 0.01, *** p < 0.001.

Additionally, it is necessary to have plausible evidence that people “complied” with the treatment, that is, that they were aware that Donald Trump was elected president. The Google Trends data depicted in figure 1B reveal that in each and every European country contained in the Eurobarometer, the relative amount of online searches targeting “Donald Trump” spiked to its maximum on the day following the election (dashed lines). The universality and extreme nature of this pattern strongly suggest that most people became aware of Trump’s election as president very quickly. While, of course, not everybody uses the Internet, there are good reasons to assume that the general public soon took notice of the surprise victory of the Republican candidate, making it a legitimate treatment in a natural experiment.

Furthermore, the exclusion restriction needs to be fulfilled.^{Footnote 52} That is, it is necessary to be certain that no other events over and above the election influenced the outcome. For this, Google Trends data can again help. Exploring trending topics on Google during the examined time frame reveals that “election” (relative rank in Google searches: 4), “Trump” (6), “election results” (16), “Donald Trump” (17), “election 2016” (18), “Clinton” (23), and “polls” (24) were the dominating date-specific keywords. All other high-ranking keywords relate to unspecific everyday interests, including “Facebook” (1), “YouTube” (2), “Google” (3), “you” (5), “news” (7), “Gmail” (8), “fb” (9), and “Hotmail” (10). This suggests that during the time-span under study, there were no other relevant events that could distort the results. While not salient in the Google Trends data, it should be noted that several European countries commemorate the end of World War I, the Holocaust, and the end of communism in Eastern Europe during the time frame under study. To exclude the possibility that these holidays are responsible for a potential change in the EU’s popularity (due to the EU becoming salient as a peace project in people’s minds during these days), we ran a robustness check in which we excluded countries that have such public commemorations from the analysis. Results reconfirm the main findings. Furthermore, we replicated the analysis with an earlier Eurobarometer from 2013 and did not find a similar effect, confirming that the observed effect is due to the singular event of Trump’s election.^{Footnote 53}

Finally, it is necessary to show that the premise of considering the presidential election as an exogenous event is actually valid. The election would not have been exogenous if Europeans could have anticipated Trump’s victory and started adapting attitudes on the EU already before the election took place. However, figure 1A illustrates that basically no one—not even political insiders—expected Trump to win, and polls remained in Clinton’s favor until the election night. European media outlets also speak of “one of the most improbable political victories in modern U.S. history”^{Footnote 54} and argue that no one saw the victory of Donald Trump coming.^{Footnote 55} Thus, it is very unlikely that ordinary citizens had adapted their attitudes before the election took place.^{Footnote 56} Another argument in favor of exogeneity is that Europeans were, of course, not the ones who went to the ballot boxes in the U.S. election. Thus, arguing that they could anticipate Trump’s win because they were planning on voting Trump into office is implausible. In the wake of the U.S. election, Europeans merely knew that Trump’s becoming the next president was—according to polls and media reports—highly unlikely. All this provides a high degree of certainty that Trump’s election took Europeans by surprise and thus acted as an exogenous shock.

Data and Sample

The Eurobarometer is a large-scale cross-sectional public opinion survey, conducted on behalf of the European Commission. The survey program started in 1974 and is released biannually. Here, we draw on the Eurobarometer 86.2, which contains data from 35 European countries, collected between November 3 and 14, 2016. In the sample for this study, all non-EU member states were excluded, since several central items were not included in these countries’ questionnaires. To increase consistency, Bulgaria was also excluded, because the data collection started and ended two days earlier there than everywhere else (cf. figure 2). All other 27 EU member states are contained in the analysis. Following best practice from a range of previous research,^{Footnote 57} respondents who were interviewed immediately after the external shock (on November 9) were excluded to ensure that respondents in the treatment group actually had time to become aware of the fact that Donald Trump was elected president. The number of observations in the sample was additionally reduced by listwise deletion of missing cases in creating the dependent variable (refer to the next section). The final sample contains 16,285 observations.

Dependent Variable (DV)

The study used 12 items from the Eurobarometer to construct an aggregate measure of EU popularity (see table 2).^{Footnote 58} All 12 items are based on a 4-option Likert scale and measure certain positive attitudes toward the EU (e.g., agreement with statements such as “the EU is modern,” “the EU is efficient,” or “the EU creates jobs”). Employing exploratory factor analysis, only one factor with an Eigenvalue >1 (5.17) emerges, on which all items load with at least .55. Thus, following commonly accepted standards,^{Footnote 59} all 12 items have good to excellent factor loadings. The one-factor solution was confirmed by various adjustments and robustness measures based on different sets of variables, as well as various rotation techniques. Furthermore, given that the variables in question were not discrete but categorical, the results were reconfirmed by using polychoric factor analysis, a procedure that makes it possible to perform factor analysis with categorical variables. Since the results resembled those from the conventional factor analysis, it was decided to adhere to the standard procedure. Employing the respective factor, the DV was extracted by using the Bartlett method.

Table 2 Elements of the EU popularity index

Note: Based on Eurobarometer 86.2

Treatment Variable and Subgroup Splitting

To measure the Trump effect, a treatment variable was created that captures whether the respondent was interviewed prior (T = 0, control group, November 5–8, N = 6,395) or after (T = 1, treatment group, November 10–14, N = 9,890) the U.S. presidential election. For the subgroup analyses, the sample was first divided into four subgroups based on how respondents perceived the economic situation of their country (“very good,” “rather good,” “rather bad,” or “very bad”). Individual-level political orientation, the second variable used to split the sample into subgroups, was inquired in Eurobarometer 86.2 through the question “In political matters people talk of ‘the left’ and ‘the right.’ How would you place your views on this scale?” The scale, which was shown to respondents, had ten categories ranging from 1 (“left”) to 10 (“right”), with the middle categories remaining unspecified. In line with common practice,^{Footnote 60} respondents were grouped into the three meta-categories “left” (1–4), “center” (5–6), and “right” (7–10) for the subgroup analyses. For several additional subgroup analyses and corresponding variable descriptions, see the online appendix, items 5, 7, and 8.

Control Variables

Operating within the framework of a natural experiment, it is not necessary to adjust for covariates, as the treatment is randomly assigned and individual socio-economic and unobserved characteristics should thus not vary before and after the treatment. However, covariates can be included to estimate the treatment effect more precisely and to control for smaller imbalances between treatment and control group.^{Footnote 61} Those variables, however, should be strictly exogenous. As discussed earlier, the sample is mostly well balanced, but some minor differences between treatment and control group do exist (cf. table 1). To correct for these small imbalances and to increase precision, the variables listed in Table 1 were included as controls in one of the models (see equation 2).