Theoretical framework

Migration information campaigns aim to influence migration by providing issue-related information. This does not imply that all migration decisions lead to actual migration and will be observable as migration movement accumulated at the macro level (Docquier et al. Reference Docquier, Peri and Ruyssen2014). The nonpossession or possession of required resources may prevent this. The measurement of migration intentions, as in the present study, can be understood as an upper bound of potential migration. However, in academia, migration intentions are themselves becoming of interest. Recent literature studies the impact of the immobile aspiring migrant on society, reducing for example their common-good provision (Carling & Schewel Reference Carling and Schewel2018).

In my perspective, when investigating how an information campaign on migration shapes decision-making, two aspects must be considered: (1) how information shapes the migration decision-making process and (2) chronologically preceding, how that information is conveyed and processed. In the following, I first focus on the characteristics of irregular migration decision-making (1), as this forms the basis for making statements about the outcome of interest. Second, I discuss the information processing model (2), where I account for the unique situation in information campaigns. In information campaigns an external source provides information, meaning that individuals do not perform any practices that enhance processing motivation, like the search for information or intrinsic interest.

Following RC theory, the dominant approach used to explain migration decision-making (Castles et al. Reference Castles, De Haas and Miller2014), individuals make their decision based on a cost–benefit calculation of relevant factors, also known as push and pull factors. Hence, information about the challenges of irregular migration, which is the focus of most migration information campaigns (Tjaden et al. Reference Tjaden, Morgenstern and Laczko2018), is expected to be included as a cost, that is, a negative estimate in the RC calculation. Given the few opportunities for the average citizen in developing countries to migrate to Europe legally, the decision is most often a binary, one-stage decision between nonmigration and irregular migration (Koser Reference Koser2010). In contrast to regular migration, the decision to migrate irregularly is an individual one. Hence, the decision-making model suggests that the information provided by the migration information campaign will negatively affect the intentions to migrate.

Since the degree of new information available for the RC migration decision-making calculus depends on the processing of the information, I integrate the previously introduced RC decision-making model with a preceding information processing model. One speciality of information campaigns is that the audience did not seek this information themselves; hence, the general processing motivation is low. However, following AIT, processing motivation is elevated via emotions. Triggering anxiety in the target audience is a widely practiced yet highly controversial issue in migration information campaigns. They do so by using messages about highly uncertain situations with only limited human control, aspects that are also known from appraisal theories to foster anxiety (Lazarus Reference Lazarus2001; Lerner & Keltner Reference Lerner and Keltner2000). Following AIT, anxiety triggers the surveillance system at the neurological level, which is shown to enhance attention, increase learning, reconsider new information, and encourage novel behavioral practices (Brader & Marcus Reference Brader, Marcus, Levy, Huddy and Sears2013). In other words, “Anxious feelings awaken the attentive and open-minded citizen” (Brader Reference Brader2005). Thus, triggering anxiety in the target audience is expected to enhance the processing motivation and lead to increased absorption of information. The state of anxiety opposes the state of calmness, as these are defined as two poles on a single axis (Brader & Marcus Reference Brader, Marcus, Levy, Huddy and Sears2013). Hence, I expect an anxiety-triggering message to have a more negative effect on individuals’ intentions to migrate irregularly and, respectively, on any hypothetical or actual migration behavior, than a message triggering calmness.

Research design

This theoretical frame translates into a mediation model. I expect the campaign message to trigger either anxiety, calmness, or no particular emotion, which then leads to different levels of impact of the campaign based on the information processing. For rigorous causal mediation testing, I would need to manipulate the independent variable and the mediator. In practical terms, the campaign message is an inherent part of the information campaign, and the emotions cannot be manipulated independently without a much more artificial field study design. However, closeness to reality is crucial in generating the first rigorous evidence about this understudied policy. Therefore, I follow the model of a design-based mediation approach as described in Gerber and Green (Reference Gerber and Green2012), meaning that mediation is studied via design. In this case, for reasons of limited feasibility, no independent mediator manipulation is undertaken. To refer to the example of design-based mediation analysis from the text, the campaign reflects the lemon and the campaign message reflects the variation in vitamin C.

Figure 2 presents an overview of the respective design and sample sizes per school and treatment group. The students were allocated randomly to the workshops by distributing colored bracelets in the school court. Although randomly allocated, the treatment and control groups vary in sizes. This is especially the case for school 1, where we observe a variation of about ten percent. This variation, however, is by change and cannot be traced back to any other issue. The meaning of the bracelet colors from the student’s perspective, that is, time slots, was only communicated after distribution to prevent self-selection. Before and after the workshop, the students obtained tasks from the teachers in separate school buildings to prevent noncompliance or spillover mechanisms. Within each workshop, the students first received the information campaign treatment by the NGO, then the small paper booklet of survey questions, that is, the endline survey, and finally, a debriefing. Ethical considerations concerning the research design and the general topic are described in Appendix E.

Figure 2.

Overview of the study design with the number of participants (N) per survey; N = total number of participants, ${N_1}$ and ${N_2}$  = number of participants per school.

I implemented the study design in two secondary schools in Benin City, Nigeria. The school setting and the city–country choice resemble a case that is typical both for implementing these information campaigns and concerning the expected propensity for irregular migration. Nigeria is one of the most frequently named countries of origin among irregular migrants in Europe, most of whom come from Benin City, due to its exceptional migration history (TheWorld Bank Group 2018). Most irregular migrants from Nigeria are adolescents or young adults taking the moment of opportunity in life transitions (see detailed Appendix A). Most migrants are part of the socioeconomic and educational middle class, a broad concept in this setting that puts together the upper-lower class and the lower-upper class of Nigerians. Given that irregular migration decisions of young Nigerians are often made independent of the household and sometimes in disguise, school students are a relevant study group. Sample descriptives and balance tests per treatment group are reported in Appendix G1. The balance tests reveal no notable variations based on the covariates and hence indicate a successfully random group allocation.

In terms of case selection, the NGO selects the locations through direct inquiries and according to the institutions’ feasibility, cooperation, and need, in consultation with their funding organization. On the one hand, this nonrandom case selection limits the generalizability of the study; on the other hand, the NGO selection mirrors real-life circumstances, which is of particular relevance in a study examining a policy implementation.

Operationalizations

Migration decision-making is operationalized via two measurements of irregular migration intentions. Additionally, and this is a deviation from the pre-analysis plan (PAP) (see Table 1), I aimed to capture actual migration behavior. Due to a poor success rate in contacting the students in a first, randomly selected, pilot, this part of the study was not followed through.

Firstly, irregular migration intentions are operationalized through a direct self-evaluation question. In line with the survey question in Afrobarometer Round 7, the following wording is used for participants to rate the likelihood of migrating: How much, if at all, do you consider moving through the backdoor (“illegally”) to another country to live. However, while the Afrobarometer-question asks about migration intentions in general, the question here has the additional component of “through the backdoor,” which is a cultural and age-appropriate way to ask about irregular migration intentions. In contrast to the four-level answer option in the Afrobarometer-question, the answer rating scales on a ten-point range, from 1 “never considered” to 10 “it is constantly on my mind.”

Secondly, irregular migration intentions are measured with a more indirect method, eventually addressing respondents’ strategic answering behavior in self-reports. The students rate their probability of irregular migration in six different scenarios that are described, and the ratings per student make up the indirect measurement of irregular migration intentions. Concentrating on the different scenarios is assumed to reduce participants’ self-control and hence any strategic answering. The scenarios are randomly assigned. This random allocation is key because the dependent variable measurement technically becomes conditional on the scenarios the participant receives. However, like in single-item-effect analyses in conjoint experiments, I assume that the random scenario allocation balances the variation of scenarios. To get a better impression of how the scenarios of the hypothetical migration intention measurement look like, it shall be described in detail in the following.

Each scenario consists of a conjoint table with three attributes describing the situation of an individual. Above the table a brief introduction explains that the situations are hypothetical and describe the situation of a Nigerian in ones age group. An example of how to read the table follows. The three attributes in the table are as follows: chance of dying on the route (high/low), chance of obtaining a residence permit in Europe (high/low), and change of obtaining good employment in Nigeria (high/low). Afterward, the respondents rate what they would do being in the situation of the described person on a scale from 1 definitely not migrate “through the backdoor” to 10 definitely migrate “through the backdoor.” An evaluation of the attribute effects on irregular migration intention is provided as (preregistered) side information in Appendix G.2, revealing that the attribute of a high probability to get a residence permit in the destination has the most relevant (significant and positive) impact, followed by a high probability to get a good job at home (significant negative), while a high probability to die on the journey does not reveal significant (negative) effects on irregular migration intentions.

The concept of knowledge is studied as a registered second-order-dependent variable in a self-report assessment. The preregistered covariate variables are age, gender, ethnic group, religion, education, area of living (rural or urban), socioeconomic characteristics (ownership and closeness of necessities), migration history or experience, migration network, perceived difficulty in crossing borders, potential migration destination, and reason for migration. Details on the operationalization of the used covariates, including question-wording, and variable indices are provided in Appendices C and D. Due to practical limitations, a baseline survey was not possible, and therefore the covariates are measured posttreatment (discussed in Appendix G.3).

Treatment: the migration information campaign

The migration information campaign studied here is one example of migration information campaigns implemented by national governments in Europe and the EU itself. It is funded by the German government and implemented by a German-Nigerian NGO. As stated by the NGO, the campaign objective is to improve and update knowledge among Africans with honest information on the risks and dangers of an irregular journey. The NGO implements the campaign through large workshops in schools and universities. The school embeds the workshop in the student schedule, which excludes self-selection into participation practices. This extensive workshop event type is common in Nigeria for educational settings. The information event takes around three hours and proceeds as follows: first, the head of the NGO introduces himself, the NGO, and facts about migration from Africa to Europe in a 30-minute presentation. In the second core element, the NGO presents self-made videos. The videos are about the challenges of irregular migration but differ in the subtopic and tone of communication.

Treatment assumption and validation

The two campaign messages, the videos, are quite distinct. They differ in frame, design, and, most relevantly, the topic. However, the central assumption of a design-based mediation model is that the mediator varies between treatment groups. Translated to the present case, the variation in the randomly assigned campaign message triggers different emotions. To validate this central assumption, I conducted an additional validation study on MTurk (see detailed Appendix B).

In this validation study, the MTurk participants code the emotions from the facial expressions of the participants in the field studies based on around 200 images that were cut in a 10-second rhythm from the video records made during the video part of the information campaign. Additionally, I added pictures of the same students but clearly smiling or laughing, from another part of the event to check the attention of the Mturk users. The number of participants in the rating study is N = 948, and the ratings per picture vary but on average each picture got rated 163 times. The emotions are categorized into six basic emotions: anger, disgust, anxiety, happiness, sadness, and surprise, plus a neutral condition. The validation relies on two basic assumptions: first, that people can recognize facial expressions across cultures (Elfenbein & Ambady Reference Elfenbein and Ambady2002) and second, that facial expressions of emotion mirror the experienced emotion (Brader & Marcus Reference Brader, Marcus, Levy, Huddy and Sears2013).

Figure 3 shows the results and reveals a statistically significantly higher positive effect of the emotions of fear and disgust expressed in the faces of the anxiety treatment group and a significantly lower expression of the emotion of happiness (negative), compared to the calmness video participant picture ratings. While the statistically significant effect on fear expressions goes along with expectations, the positive, strong, and statistically significant coefficient for disgust is somewhat surprising. The emotion literature regarding disgust is scarce compared to the highly researched emotions of anger, sadness, and happiness. Aarøe et al. (Reference Aarøe, Petersen and Arceneaux2017) suggest a certain relationship between the emotions of anxiety and disgust. According to them, both disgust and fear arise as a reaction to a sense of threat and is thus associated with a certain protective instinct of a person.

Figure 3.

Coefficients of each emotion being in the anxiety video group compared to the calmness video group; results are based on individual logistic regressions identifying the treatment effect on each emotion in comparison with the neutral group; robust standard errors; dependent variable scale from 0 to 1; geometrical forms = point estimates, whisker = 0.95 confidence intervals; detailed description and regression output Appendix B, Table 1.

However, while this result accords with the intended emotional manipulation of the treatment, it is still unclear whether the topic or the tone of the message induced the emotion. The crucial part of this study is nevertheless supported: the two campaign messages in the form of videos trigger significantly different emotions in the audience.

Results experimental design emotion mediation

Before engaging with the results, I want to give an overview of deviations from the PAP. The PAP was registered before study implementation. Table 1 structures deviations from the PAP in terms of reporting and implementation deviations. Deviations based on reporting imply that the analysis proposed in the PAP is not in the main paper but the Appendix, while implementation deviations are aspects proposed in the PAP but not carried out. Both types of deviations are relevant as they shift the storyline of the paper. Given that the reason for deviating determines the severity of the implication, I report them in Table 1 showing that reporting deviations here are generally due to duplication of information or lack of additional information and implementation deviations because of practical or methodological feasibility.

Table 2 illustrates the average effect of the anxiety-triggering treatment compared to the calmness-triggering treatment on irregular migration intentions. The results are based on linear ordinary least squares regressions with robust standard errors on a subsample of secondary schools, comparing only the two treatment groups (comparison of each treatment group with the control group in Appendix G.5). Since each respondent assessed several randomly assigned hypothetical scenarios regarding the situational likelihood of irregular migration, the analysis is pooled by individual respondent for the dependent variable “hypothetical irregular migration.” The different scenarios assessed by each respondent are intended to distract from the strategic response behavior and are not analyzed further here. As they are randomly assigned, the variation in scenarios should even out in the aggregate, allowing for an analysis independent of the scenario treatment category (similar to the analysis of individual treatment in conjoint experiments). The variation in the number of observations is based on the multiple scenarios per respondent in the hypothetical measurement and an attrition rate. The test for systematic attrition shows that the rate depends on the age of the respondent, with the response rate increasing with age (Appendix G.7).

Table 2.

Average treatment effect – anxiety versus calmness treatment; calculation based on the two treatment groups, a subset of the sample of secondary schools, full comparisons with control group see Appendix G.5; covariates are not included; dependent variable scale from 1 to 10; robust standard errors

N = number of observations; n = number of observations pooled; pooled ${{\rm{\;}}_i}$  = pooled by the individual respondent with clustered standard errors by respondent. Variation in the number of observations is based on the multiple scenarios per respondent in the hypothetical measurement and an attrition rate based on the age of the respondent. *p < 0.1; **p < 0.05; ***p < 0.01.

The results reveal that the anxiety-triggering message reduces irregular migration intentions compared to the calmness-triggering message for both measurements of migration intentions, yet with varying intensity. Self-indicated irregular migration intentions are reduced by 0.71 points on a scale from 1 “never considered” to 10 “it is constantly on my mind” and hypothetical irregular migration intentions by 0.02 (for standardized coefficient regression see Appendix G.3). The estimates indicate a rather small and only part-wise statistically significant, but consistently negative impact of the anxiety-triggering migration information campaign on irregular migration intentions. While being small in magnitude, the effects must be interpreted in relation to the treatment (De Mesquita & Tyson Reference De Mesquita and Tyson2020), a one-time information provision, and can therefore be seen as noteworthy. Adding the median imputed covariates (Lin et al. Reference Lin, Green and Coppock2016: Green lab) from the PAP does not change the effect coefficient notably (see Appendix G.3), and the comparison to the control group does not change the estimate considerably because the control group (i.e., no video) and calmness-triggering group have similar outcomes on average (see Appendix G.5).

Additional (preregistered) analyses show that the anxiety-triggering message has a positive effect on participants’ self-reported level of knowledge in comparison with the calmness treatment (see Appendix G.6). This result aligns with the theoretical mechanism: the emotion of anxiety is expected to foster primarily rational-analytic information processing, which leads to an increase in knowledge. In deviation from the PAP, no analyses of potential heterogeneous effects are conducted. For a valid analysis, the interaction variable inspected should either be measured before treatment implementation or not be affectable by the treatment; neither was feasible here.