Policy Research Working Paper 10754 News Sentiment in Destination Countries and Migration Choices Evidence from Libya Michele Di Maio Nelly Elmallakh Valerio Leone Sciabolazza Middle East and North Africa Region Office of the Chief Economist April 2024 Policy Research Working Paper 10754 Abstract Changes in the sentiment of migration-related news pub- The results indicate that changes in news sentiment have lished in destination countries affect the timing of migrants’ a significant impact only for some groups of migrants and journeys to these countries. Using geo-localized data on under specific conditions, suggesting a limited effect on migrants in Libya and the complete record of news articles overall migrant movements. Finally, the paper provides in their country of destination, this paper shows that a suggestive evidence that a worsening news sentiment in worsening news sentiment leads to migrants staying longer the preferred destination induces substitution across desti- in Libya, slowing down their journeys to their final destina- nation countries, yet it does not make migrants return to tions. The paper validates these results by showing that the their country of origin. effect is concentrated in locations with internet connections. This paper is a product of the Office of the Chief Economist, Middle East and North Africa Region. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at michele.dimaio@uniroma1.it, nelmallakh@worldbank.org, and valerio.leonesciabolazza@uniroma1.it. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team News Sentiment in Destination Countries and Migration Choices: Evidence from Libya∗ Michele Di Maio† Nelly Elmallakh‡ Valerio Leone Sciabolazza§ Sapienza University and IZA World Bank Sapienza University Keywords: Migration, news sentiment, GDELT, Libya, Europe JEL codes: F22, J61, L82. ∗ We thank the participants to the 16th AFD-World Bank Conference on Migration and Development (Boston University), the Conflicts and the Economy Workshop (Universit` a Cattolica of Milan), and the EBRD Research Seminar (London) for their comments and suggestions. We also thank Margherita Bove, Francesca Calamunci, and Cecilia Nardi for excellent research assistance. The authors gratefully acknowledge the financial support from the Office of the Chief Economist for the Middle East and North Africa (MNACE) under the regional Labor and Gender Research Programs. The views expressed in this paper are those of the authors and do not necessarily reflect the position of the World Bank. All errors are our own. † michele.dimaio@uniroma1.it, Sapienza University of Rome (Italy) and IZA ‡ nelmallakh@worldbank.org, World Bank: Corresponding Author § valerio.leonesciabolazza@uniroma1.it, Sapienza University of Rome (Italy) 1 Introduction Migration is one of the most politically polarizing issues in advanced countries. This is reflected in the way migration and its effects are described in the news. While some studies look at how migration-related news influences public opinion and votes in destination countries,1 little is known about the effect of changes in the sentiment of migration-related news in destination countries on en route migrants and their migration choices. Investigating this matter is impor- tant to better understand the impact of the increasing use of an aggressive tone by populist leaders trying to discourage migrants from seeking entry into Europe or the US.2 This paper studies how changes in the sentiment of migration-related news published in mi- grants’ preferred destination countries affect migrants’ movements within Libya and the timing of their journey to these countries. Libya is the country with the largest number of international migrants in all of North Africa and the major gateway from Africa to Europe.3 By showing to what extent changes in the news sentiment in destination countries impact the choices of mi- grants in Libya, we contribute to a better understanding of the determinants of the movements along the most important irregular migration route to Europe. Our analysis combines two main data sources. First, we use data on migrants located in Libya during the period 2017-2020 collected by the International Organization for Migration (IOM). Even though some migrants plan to remain in the country, for most of them Libya is a transit country towards their final destination. While in Libya, migrants often stay for long periods in the same location, live in rented houses, and have an (informal) job (IOM, 2018a). During their stay in the country, migrants often move across locations several times, using different internal routes according to their nationality (Di Maio et al., 2023). To study migrants’ movements within Libya, we use geo-localized monthly data collected by the IOM in 167 Flow Monitoring Points (FMPs) in Libya. FMPs are located at important points of permanence and transit for migrants. For each FMP, we have information on the country of origin of migrants located in the area, their destination countries, and their length of stay at the FMP. Second, we use data from the Global Database of Events, Language, and Tone project (GDELT)4 to identify all the news articles related to migration published in any of the destination countries of the migrants located in all the FMPs in Libya. Importantly, GDELT provides a variable measuring the sentiment of each news article, allowing us to construct, for each FMP in each month, a measure of the average tone of all migration-related news articles published in the destination countries of migrants located near the FMP. Our main result shows that a reduction in the measure of the sentiment of migration-related news articles published in destination countries increases the share of migrants remaining in their current location in Libya for a long period. The characteristics of migration flows in Libya 1 See for instance, Benesch et al. (2019); Meltzer et al. (2021); Couttenier et al. (2023); Djourelova (2023). 2 This is the case of the message broadcasted by Italian PM Giorgia Meloni on September 16 2023, https: //youtu.be/gFrSJyF7xfU?feature=shared or that by Florida Gov. Ron DeSantis https://www.foxnews.com/ politics/desantis-warns-migrants-bused-texas-dc-florida-do-not-come 3 One important reason why Libya is the preferred transit country to Europe is the lack of border and internal controls, which has characterized the country since the beginning of the Civil War in 2011 (UNHCR, 2019b). 4 GDELT represents a rich source of information to study news coverage of migration events. The database registers events worldwide by collecting all online news released every 15 minutes, and categorizing each news item by location, subject, and sentiment, among other things. 1 allow us to interpret this result as also informative about how a change in the tone of the news influences the timing of migrants’ movements toward their final destinations. Migration in Libya is in fact largely unidirectional: most migrants move from South to North-West, i.e. toward the coastal areas where there are more employment opportunities, and from where - those directed to Europe - then attempt to cross the Mediterranean Sea.5 It follows that our result, i.e. a more negative tone in the migration-related news at destination makes more migrants stay longer in Libya, can be read as indicating that a worsening news sentiment slows down the migrants’ journey by delaying the next step of their travel to destination countries. Our interpretation is corroborated by results showing that the effect of a change in the tone of migration-related news on migrant’s choices is driven by articles related to discrimination. This finding is consistent with the idea that the tone of news articles on discrimination episodes predicts the attitude of the local population in destination countries towards migration and the level of difficulty that migrants may encounter in terms of job opportunities, access to housing, and social integration in these countries (World Bank, 2023; Aksoy et al., 2023). A more negative news sentiment in articles related to these events indicates that conditions at destination are worsening, reducing the incentive for the individual to make the next step of the migration journey towards the final destination. To validate our findings, we provide suggestive evidence that the effect of changes in the tone of the news is significant only for migrants located in areas where it is more likely that they have access to the Internet and that our results are not driven by reverse causality. We also show that our findings are robust to several checks, including: an alternative way of measuring the tone of the news; the use of different samples; an alternative estimator; increasingly demanding model regression specifications; the inclusion of additional controls for the political orientation of the governments in destination countries, and the tightness of the Mediterranean Sea patrolling activity; and the use of spatial regression models to account for spatial autocorrelation in migrants’ movement choices. Next, we characterize the conditions under which changes in the tone of the news are more likely to have an effect. Our results show that a worsening news sentiment at destination slows down migration movements in Libya only for migrants coming from West African countries (who are those usually making a step-by-step journey to Europe) or located in the Western region of Libya (where there are more economic opportunities), and when migration-related news articles are more numerous and mostly positive, or economic conditions in destination countries are not favorable. While these results corroborate our main findings by showing that the effect of changes in the tone of the news is significant when, where, and for whom we expect it to be so, they also indicate an overall limited impact on migrants’ movements within Libya and towards destination countries. Finally, we test for the possibility that changes in the news sentiment generate spillover effects across countries. Results indicate that when the tone of migration-related news becomes more negative in a country, migrants tend to switch to other countries as preferred destina- tions. Yet, we find that this substitution does not occur between European and non-European 5 This characteristic of the migration flows in Libya (i.e., they mainly move in the same direction) is similar to that of migration flows from Central America to the U.S., suggesting that our approach can also be used in other contexts if similar data are available. 2 countries. We interpret these findings as suggesting that - while there is substitution across des- tinations that are similar - a worsening news sentiment in European countries does not induce migrants to return to their origin country. Our findings have important policy implications. Attempts to discourage migrants from reaching their preferred destination countries based on the use of negative, aggressive, and discriminatory tone in the public discourse on migration are unlikely to be successful. While a more negative tone of the news makes some migrants stay longer in Libya, delaying their journey to their final destination, this happens only for migrants located in specific areas of the country or traveling using a specific route, and only when there are specific conditions in destination countries. Moreover, the strategy employed by populist governments to reduce migration to their country by creating an unwelcoming environment for migrants is politically unsustainable due to spillover effect of the tone of the news across destination countries. Given that a more negative tone of the news in one country increases migration flows to similar destinations (but not the willingness of migrants in Libya to return to their country of origin), the use of this populist strategy undermines the basis for the cooperation between European governments in the management of migration flows from Africa. The paper is organized as follows. Section 2 discusses the related literature. Section 3 provides background information on migration in Libya. Section 4 describes the data. Section 5 presents the empirical strategy, and Section 6 discusses the results. Section 7 concludes. 2 Literature Our paper is related to the small - but rapidly growing - literature looking at the effects of media exposure on migration choices. A few studies document a positive association between Internet access and migration intentions and aspirations (Pesando et al., 2021; Grubanov-Boskovic et al., ohme et al. (2020) show how geo-referenced online search data on migration-related 2021). B¨ terms can be used to measure migration intentions in origin countries and to predict bilateral migration flows. Adema et al. (2022) uses individual-level data from 112 countries to provide causal evidence that an increase in mobile internet access increases the desire to migrate and actual migration by lowering the cost of acquiring information on potential destinations. Other e and studies look at the effect of TV and newspaper articles on migration. For instance, Farr´ Fasani (2013) find that exposure to TV reduces internal migration in Indonesia. Wilson (2021) shows that access to newspaper articles and TV news providing information about potential labor market opportunities increases migration to areas mentioned in the news.6 More generally, our paper is related to the literature on the effect of information on migrants’ decisions. Shrestha (2019) shows that information on migrants’ mortality rates decreases mi- gration flows from Nepal to Malaysia and to the Gulf countries. Baseler (2023) documents that providing information on urban earnings in Kenya increases migration to the capital. Some pa- pers use randomized controlled trials to understand the role of risk information and perceptions in the decision-making process of migrants. Bah et al. (2023) show that providing information 6 Another important source of information is migration networks at destination. These provide prospective migrants with information about the migration process and economic opportunities, helping them to shape their migration decision (McKenzie and Rapoport, 2010; Beine et al., 2015; Giulietti et al., 2018). 3 and testimonials about the risks of backway migration from The Gambia to Europe reduces migration intentions in the five years after the information treatment. Similarly, Tjaden and Dunsch (2021) find that peer-to-peer information transmission on the dangers and potential risks associated with irregular migration raised risk awareness and reduced irregular migration intentions in Senegal. Importantly, Bertoli et al. (2020) emphasize that acquiring information about destinations can be costly for migrants. As a consequence, migration flows from countries in which collecting information is more costly or people have stronger priors are less responsive to variations in economic conditions in destination countries.7 Our paper differs from these previous studies in two main aspects. First, this paper is the first to empirically assess the impact of migration-related news articles in destination countries on the movement decision of migrants located in a (transit) developing country.8 Second, in our analysis, we study the effect of changes in the content of migration-related news (i.e., their sentiment) on migrants’ behavior, thus improving on previous studies using metrics for media access and exposure. We are also the first to use the GDELT dataset to study migration issues.9 Finally, our paper contributes to the understanding of the evolution of the Libyan economy (Irhiam et al., 2023; Del Prete et al., 2023; Rahman and Di Maio, 2020; Di Maio et al., 2023) and of the migration phenomenon in the country after the fall of the Gaddafi regime (Di Maio et al., 2023). Moreover, by providing novel evidence on the determinants and the timing of migrants’ movements within Libya, our analysis is also related to studies looking at the dynamics of migration flows to Europe (Battiston, 2022; Deiana et al., 2024) and of international migration flows in Northern and Sub-Saharan African countries (Friebel et al., 2023; Di Maio et al., 2023). 3 Migration in Libya Libya is the country with the largest number of migrants in North Africa. In 2020, Libya hosted approximately 571,400 migrants, which corresponds to 12% of its total population (UN-DESA, 2020). Migrants in Libya are very heterogeneous in terms of nationality, background, motivation for migration, and preferred final destination (IOM, 2018b; UNHCR, 2019b). Libya has been an important destination for migrants since the 1970s. For decades, people from Sub-Saharan Africa, the Middle East and North Africa, and Asia have moved to Libya to work (UNHCR, 2019b). During the Gaddafi regime, economic migrants accounted for more than 50% of the total Libyan labor force (World Bank, 2015). Since 2011, the security situation in Libya has been critical. The country has experienced a prolonged period of conflict which started with the fall of the Gaddafi regime and the First Libyan Civil War, and continued with the Second Libyan Civil War in 2014.10 The conflict 7 There is also some evidence that migrants’ knowledge of the political orientation at destination influences their location choices (see e.g., Bove et al., 2023). Bracco et al. (2018) show that the election of an anti-immigration mayor resulted in a decline in the number of immigrants entering the same town. 8 For a discussion of the characteristics of transit migration, see Djaji´ c (2017) and Artuc and Ozden (2018). 9 Until recently, GDELT has been mainly used to study macroeconomic or financial aspects (Bayer et al., 2020; Consoli et al., 2021). An exception is Deiana et al. (2023), who use the Pope’s message about the refugees’ situation to analyze how persuasion modifies beliefs. 10 The Second Libyan Civil War is one of the geopolitically most relevant conflicts of our times (Fitzgerald and Toaldo, 2016). The conflict - which is still ongoing - is characterized by two rival governments claiming authority over Libya, one (based in Tripoli) controlling the Western part of the country and the other (based in Tobruk) 4 transformed the characteristics of migration in Libya, deeply affecting flows to, within, and from the country (Cummings et al., 2015). While Libya is still an important destination country for foreign workers, Libya is now also a transit country for migrants heading to Europe. Due to the dissolution of the Libyan state and the lack of a government able to control the territory, the migration routes through the country remained unguarded and Libya has become the main gateway to Europe for irregular migrants (Friebel et al., 2023; UNHCR, 2019b). Libya is the arrival point of two important irregular migrant routes, the Western and the Eastern route (see Figure A1). The Western route to enter Libya is used by migrants from West and Central Africa traveling from Niger and reaching the southern district of Sebha. The Eastern route is instead used predominantly by East Africans who usually cross the Sudanese border to enter the district of Kufra in the southeast (UNHCR, 2019b). Once migrants arrive in Libya, most head north to the coastline. This is the part of the country where more employment opportunities exist (Wittenberg, 2017) and also from where it is easier to attempt reaching the other side of the Mediterranean Sea (Mixed Migration Hub, 2015; IOM, 2015; UNHCR, 2017a). The trip to Europe is expensive. This is why - once in Libya - many migrants stay for long periods in hubs along their route, working to afford the next leg of their trip (Adesina, 2021). Information on and analysis of the internal movements of migrants within Libya is very scarce. One exception is Di Maio et al. (2023), which documents the large migrants movements within the country (see Figure A2). They identify the different migrants’ routes in Libya and their evolution across time, showing that migrants tend to move north, and that they sort into these routes according to their nationality. Migrants in Libya travel in stages, turning to smugglers for the difficult legs of the journey.11 Ethnographic research documents that the relationship between migrants and smugglers in Libya is more complex than usually thought (Sanchez, 2020). Identifying smugglers as criminals and migrants as victims oversimplify a complex and often symbiotic relationship (World Bank, 2023).12 In the Libyan context, smugglers are often ordinary people, living in border areas along migration pathways, and in coastal towns and cities, who facilitate the movement of migrants through the country (Sanchez, 2020). Migrants survey data indicates that smugglers are in most of the case considered “travel agents” (59% of respondents) and only in few cases “criminals” (11%) (Murphy-Teixidor et al., 2020). Other studies emphasize that smugglers and migrants have a common interest that their interaction is successful. Moreover, the various roles of migrants (who sometimes contribute to the organization of the journey) blurs the boundary between smugglers and their clients (Achilli, 2021). Taken together, the evidence from these studies suggests that migrants do not perform a passive role in their journeys: they have agency and are essential actors of their mobility (Sanchez, 2020). Based on these arguments, migrants in Libya should be thought of as (most often) being able to choose when and where to move in controlling the Eastern part of the country. 11 In a survey of 5,159 migrants conducted in Libya in 2019, 32% reported not using any smuggler, while 37% used one smuggler, and 31% used several smugglers along their journey (Murphy-Teixidor et al., 2020). 12 Smuggling and human trafficking are two very distinct phenomena (Adesina, 2021). Smuggling refers to a voluntary movement of migrants (within or between countries) facilitated by an agent (smuggler) who receives payment to take them to a destination. Instead, human trafficking refers to a movement of migrants into another country that includes an element of extortion, exploitation, or coercion. In the case of Libya, only rarely smuggling overlap with human trafficking (Achilli, 2021). 5 their step-by-step journey through Libya. There are differences between the migration journey through Libya along the Western and the Eastern routes. Migration through the Western route almost always takes place in stages (Wittenberg, 2017). Migrants use smugglers to avoid the large number of checkpoints that have to be crossed, though there are few reports of detention and deportation along the routes heading north. Anecdotal evidence indicates that migrants from West African countries often stop in a locality until they save enough money (often by being employed informally) to be able to continue their journey (UNHCR, 2017b). The journey through the Eastern route is instead non-stop, characterized by strict cooperation between smugglers and human traffickers, and often involves cases of violence. In recent years, this journey has become very short, between ten days and three weeks (Micallef, 2017). 4 Data 4.1 Migration data Data on migrants and their movements are retrieved from the IOM’s Libya Displacement Track- ing Matrix (DTM). The IOM-DTM is the only source providing reliable data on migration flows in Libya. The DTM tracks the movements of migrants in Libya using the data collected at the IOM Flow Monitoring Points (FMPs) in the Libyan territory.13 FMPs are located at important points of transit and permanence of migrants such as border crossings, bus stations, harbors, markets, public squares, areas where migrants look for work, etc. (IOM, 2016). Migrants included in the DTM dataset live close by a FMP and are able to move across the country. These migrants are not victims of human trafficking or imprisoned in detention centers.14 Most of them often remain for long periods in a given location, live in rented houses, and have a job (often informal) (IOM, 2018b, 2019). In our empirical analysis, we use DTM data for the period June 2017 to February 2020.15 The DTM collects information on migrants located in, arriving to, and departing from the area around each FMP. For each FMP, it reports the three most common countries of origin, the three most common countries of destination, and the share of migrants by length-of-stay category (i.e., less than 2 weeks, less than 3 months, between 3 and 6 months, more than 6 months). Data are collected daily and reported weekly. In our analysis, we aggregate them at the monthly level. The primary method of data collection at the Flow Monitoring Point (FMP) is through Key 13 The IOM definition of migrant does not distinguish between legal and illegal migration. Yet, as discussed in Di Maio et al. (2023), migration flows recorded by the DTM are likely to be mostly made of irregular migrants. 14 According to the Libyan immigration law, individuals entering the country without a permit are at risk of detention at any time. The Directorate to Combat Illegal Migration (DCIM) (established in 2012 to tackle irregular migration flows into the country) is in charge of arresting and organizing the deportations of irregular migrants, and managing the detention centers. While immigration law is rarely enforced, migrants arrested and detained often face severe abuses by security forces and militias, including torture, rape, extortion, forced labor, and extra-judicial execution (Mixed Migration Centre, 2019). The living conditions in the detention centers remain an issue of grave concern with serious humanitarian consequences for the migrants detained (UNSMIL, 2018). In 2019, 4,700 migrants (around 1% of the total migrant population) were in detention centers. 15 The period of analysis begins with the first wave of the DTM data for Libya (2017) and ends with the beginning of the COVID-19 pandemic (March 2020). 6 Informants (KIs) interviews (IOM, 2016).16 While information is not directly collected from migrants, DTM data have been collected repeatedly and cross-checked since 2017 and have been used in several official reports. These data are considered the best source of information—in terms of accuracy and coverage—available at the moment on migration flows in Libya. Figure 1 shows the physical map of Libya and the location of the 167 FMPs used in our analysis. The FMPs are located across all of Libya. Importantly, there are FMPs in both the areas controlled by the (Tripoli-based) Government of National Accord and by the (Tobruk- based) House of Representatives (see footnote 10). There are FMPs at all the entry points to Libya, at all the important crossroads, close to the largest cities (e.g., Tripoli, Misurata, Banghazi), and in both the desert areas (South) and the coastal areas (North). Most migrants located in the FMPs in Libya come from sub-Saharan African countries. During the period under analysis, the most common origin countries are: Niger, the Arab Republic of Egypt, Chad, Sudan, and Nigeria. The destination countries for migrants in Libya are mostly European ones. Yet, some migrants want to remain in Libya or to return home. In our data, the list of most common destination countries includes: Italy, France, Germany, Libya, and Niger. 4.2 News articles data Data on news articles are retrieved from the Global Database of Events, Language, and Tone (GDELT) (https://www.gdeltproject.org/). GDELT is an open-source news database built combining advanced natural language and data mining techniques to monitor news at the global level in real time. The project gathers meta-data from online news sources globally (every 15 minutes) and translates it into English from more than 65 languages In our analysis, we use GDELT 2.0 Global Knowledge Graph (GKG), the most compre- hensive version of the GDELT dataset. GDELT identifies for each news article its source, its nationality, and its main topics.17 We define a news article to be migration-related if one of its top 3 topics is one of the following: migration, refugees, borders, displacement. GDELT thus allows us to identify all the news articles related to migration for the period 2017-2020 for any of the destination countries of the migrants located in all the FMPs in Libya. An important feature of GDELT is that it provides a numerical variable measuring the tone of each news article. The tone is calculated as the positive score minus the negative score, where the positive (negative) score is the share of all words in the article found to have a positive (negative) connotation. In practice, a news article has a positive (negative) tone score if the share of words with a positive connotation is larger (smaller) than that with a negative connotation. This tone variable has been shown to perform similarly to human-coded sentiment 16 A Key Informant (KI) is a “person within the community who, due to his or her position, has access to specialized knowledge about the situation and context of the populations residing within. KIs can be local officials, religious, community, or tribal leaders, government administrators, local humanitarian workers, representatives of displaced groups, or others.” As an example, during Round 18 (February-March 2018), 1360 KIs interviews were conducted, in all 100 Baladiyas (i.e., districts) of Libya (IOM, 2018a). 17 The GDELT algorithm extracts from the content of the news the major topics discussed, categorize them into topics using the taxonomy provided by the World Bank Topical Ontology, and then rank topics in order of importance, so to make it possible to identify the primary focus of each news article. For more details, see https://blog.gdeltproject.org/mapping-the-media-a-geographic-lookup-of-gdelts-sources/ 7 scores and other sentiment indicators (Ribeiro et al., 2016). Using GDELT data, we construct the variable T one N ews Destinationsi,m−1,t , which mea- sures the average tone of all migration-related news articles published in month m-1 in year t in the destination countries j of migrants in FMP i. This is done with the formula: Nz,j Nji ixz,ji ,m−1,t z =1 j =1 Nz,ji T one N ews Destinations i,m−1,t = Nji where Nji is the number of top destination countries indicated by migrants in FMP i, xz,ji ,m−1,t is the tone (i.e., the sentiment) of migration-related news article z published in country j in month m-1 in year t, and Nz,ji is the number of migration-related news article published in destination country j in that period of time.18 The variable T one N ews Destinations i,m−1,t has mean -2.39 and standard deviation 0.59. We construct the same index also for different subsets of topics in migration-related news published in destination countries. To this end, we calculate the following index: Nzj i xztopic ,ji ,m−1,t Nji z =1 j =1 Nzji T one N ews Destinations T opic i,m−1,t = Nji where xztopic ,ji ,m−1,t is the sentiment of a migration-related news article discussing a specific topic published in destination country j of migrants in FMP i in month m-1 in year t, while the other terms are the same as in the previous formula. In our analysis, we consider the following topics: 1) discrimination (at destination); 2) sea incidents; 3) economic conditions (at destination); 4) crime (at destination); 5) government announcements (related to migration). Finally, we create a variable measuring the tone of the migration-related news articles referring to any other topic not included in this list.19 Summary statistics for these variables are shown in Table A1. 4.3 Other data Economic activity in Libya To compute a proxy for economic activity around the FMP, we use the intensity of night lights in the 5 km radius from the FMP’s location in the corresponding month. Night lights data are often used to measure economic activity when data to construct economic indicators are missing or badly measured, as in our case.20 In our analysis, we use 18 The tone of a given news article reflects the attitude towards migration of the media outlet on which it is published, e.g., its political orientation. As long as the number of outlets does not vary systematically across periods, our measure captures - given the average political orientation of all the outlets in a given country - the variation in the tone of migration-related news across time. 19 For more details on the construction of these variables, see Appendix C. The correlation be- tween the variable T one N ews Destinations and a variable given by the sum of the tone of news articles by topic (i.e., T one N ews Destinations Discrimination + T one N ews Destinations Crime + T one N ews Destinations Economic conditions + T one N ews Destinations Sea incidents + T one N ews Destinations Government announcements + T one N ews Destinations Other) is 64%. This depends on two factors: i) some news could be double counted (e.g., they refer to both discrimination and crime) and used in the construction of more than one variable; ii) the average tone of the news in a destination country is averaged with the average tone of the news in other destination countries indicated in that FMP. 20 Data on the Libyan economy are extremely limited. Official data on economic activities were collected only until 2011. After that, statistics on the Libyan economy have been largely unreliable due to the limited capacity 8 night light data distributed by the Visible Infrared Imaging Radiometer Suite (VIIRS). Conflict intensity exposure For each FMP, we construct a time-varying measure of conflict intensity exposure as the number of conflict events that occurred in the 5 km radius of the FMP’s location during the corresponding month. Data on conflict events are from the PRIO/Uppsala Armed Conflict and Location Event (ACLED) dataset which covers conflict events worldwide, providing the geo-localization, the date, and the type of event (Raleigh et al., 2010). Internet coverage Data on the geo-localization and the speed connection of cell towers in Libya are collected by OpenCellid (https://opencellid.org), the world’s largest open database of cell towers, and distributed by GSMA (https://www.gsma.com), a non-profit indus- try organization which provides source of data and analysis for the mobile industry worldwide. Data from Libya cover all the network operators present in the country: Almadar Ajadeed, Libyana, LibyaPhone Mobile. Data is available only for a single year: 2021 for the operators Almadar Ajadeed and LibyaPhone Mobile, and 2019 for Libyana. Sea patrolling activities and Mediterranean Sea crossings Data on monthly patrolling activities in the Mediterranean Sea are not available. Yet, data exist on the number of arrivals from Libya to Italy. We retrieve these data from the Operational Data Portal of the United Nations High Commissioner for Refugees (UNHCR).21 We use these data to construct a dummy variable which takes value one when patrolling activities are intense, i.e., when arrivals during a month are below the median in the period 2018-2020 (1150 individuals), and zero otherwise.22 Economic data for destination countries Data on country-level employment rate and per- capita income are sourced from the International Labour Organization (ILO) and the World Bank (WB), respectively. Values for these variables are provided at the annual level. From them, we impute the variables Employment Rate (monthly) and GDP per capita (monthly). For each FMP i, we use this data to calculate the average employment rate and GDP per-capita for the top destination countries of migrants located at FMP i during month m in year t. Political data for destination countries We measure the political orientation of the gov- ernment coalitions in the main European destination countries (Italy, France, Germany, Spain, and United Kingdom) using the Rile index. The Rile index locates a party on a right-left scale by measuring how much the party mentions left or right issues in its electoral program.23 Using this measure, we construct the variable Rile Index Coalitioni,m,t for the government coalition of government services (Rahman and Di Maio, 2020). 21 Data are available at: https://data2.unhcr.org/en/situations/mediterranean/location/5205. 22 To build this variable, we consider data after 2017. The reason is that the introduction of the Minniti Compact in 2017 caused the crossing to decrease by more than 80% the next year, from 117,153 to 23,037. Instead, in the following years, the number of crossings remained fairly similar, 11,4171 in 2019 and 34,154 in 2020. For more details on the effects of the Minniti Compact see Marchesi (2021) and Deiana et al. (2024). 23 Data are from the Manifesto project: https://manifesto-project.wzb.eu/. The Rile index is a continuous variable obtained with the formula R − L, where R and L are the shares of issues in an electoral program commonly associated to a right-wing and a left-wing political platform, respectively. 9 in country i during month m of year t as follows: Rilep,gi ,m,t ∗ seatp,gi ,m,t Rile Index Coalitioni,m,t = seatp,gi ,m,t where Rilep,gi is the Rile index associated to party p in government gi of country i during month m of year t, and seatp,gi is the number of seats that party p obtained in the Parliament of country i in the last elections.24 For the same set of countries, we also register changes of cabinet. Specifically, we construct the dummy variable Change of cabineti,m,t which takes value one if the government of country i is replaced in month m of year t, and zero otherwise. 5 Empirical strategy Our main regression model is: Yi,m,t = β0 + β1 T one N ews Destinationsi,m−1,t + β2 N ightlightsi,m−1,t + β3 Conf licti,m−1,t + + β4 Remaini,m,t + β5 Returni,m,t + θi + φm,t + γo + µd + i,m,t (1) where Yi,m,t is the share of migrants in the FMP i in month m in year t who are long stayers, i.e., report being in that location for more than 6 months. Long stayers are the large majority of migrants in Libya (IOM, 2019) and in all FMPs in our sample (see Figure A3). We use this variable to capture the change in the pace of the movements of migrants across Libya and in their migration journey.25 T one N ews Destinationsi,m−1,t is the average tone of all migration-related article news published in month m − 1 in year t in the top destination countries for migrants in FMP i in month m. N ightlightsi,m−1,t is the intensity of night lights within a 5 km radius of FMP i in month m − 1 and year t and Conf licti,m−1,t is the number of conflict events within a 5 km radius of FMP i in month m − 1 and year t. We control both for the possibility that migrants are not planning to leave Libya using Remaini,m−1,t , a dummy taking value one if Libya is one of the top destination countries for migrants in FMP i in month m, and for the possibility that migrants are planning to return home using Returni,m−1,t , a dummy taking value one if one of the countries of origin is also one of the top destination countries for migrants in FMP i in month m. θi is the FMP fixed effects and φm,t is the month-year fixed effects. γo is the countries of origin fixed effect and µd is the destination countries fixed effect. By including country-of-origin fixed effects, we compare - for a given FMP, the effect of the tone of the news on the migrants’ decision to move between two months of the same year when the FMP’s composition of migrants is the same. By including country-of-destination fixed effects instead, we compare - for a given FMP, the effect of the tone of the news on the migrants’ decision to move during two months of the same year when the FMP’s preferred country-of-destination is the same. As a result, by adding both sets of fixed effects, we sort out from our estimates the 24 The list of parties part of each government coalition and the number of seats obtained by them in each election is retrieved from the ParlGov project (www.parlgov.org). Data from the Manifesto and the ParlGov project are matched using information from PartyFacts (https://partyfacts.herokuapp.com/). 25 Note that the IOM-DTM data does not distinguish departing migrants from a FMP by their length of stay at the FMP. It follows that using the number of departing migrants from the FMP as an outcome would not be informative of the effect of changes in the tone of the news on the pace of the migrants’ journey. 10 possible confounding effects due to the time-invariant characteristics of the FMP, the common characteristics of all FMPs in Libya in a given month, and the time-invariant characteristics of both migrants’ nationality and their destination country. As robustness check, we will also include country of destination-specific time trends and country of origin-specific time trends. Finally, i,m,t is the error term. Standard errors are clustered at the FMP level. Descriptive statistics for all variables used in the analysis are reported in Tables A1 and A2. 6 Results 6.1 Main results: Tone of the news and migrants’ movements Table 1 reports the estimation results from our regression model 1. The outcome variable is the share of migrants in the FMP who are long-term stayers, i.e., reporting that they are in that location for more than 6 months. Column 1 shows the results of the baseline specification in which we control only for the full set of fixed effects. Results indicate that an increase (decrease) in the variable measuring the tone of migration-related news articles in the destination countries decreases (increases) the share of migrants who stay in a FMP in Libya for a long period.26 This result is also confirmed when we control for the local-level of economic activity and of conflict intensity in the neighborhood of the FMP (column 2),27 and for the presence of migrants who plan to remain in Libya or to return to their origin country (column 3).28 We interpret our results on the effect of the tone of the news on the timing of migrants’ movements across FMPs within Libya as also informative of its impact on the timing of their mi- gration journey. Migration in Libya is largely unidirectional: most migrants tend to move from South to North-West, i.e. toward the coastal areas where there are more employment opportu- nities and from where - those directed to Europe - then attempt to cross the Mediterranean Sea (see Section 3). It follows that our result, i.e., a worsening tone of the news at destination makes more migrants stay longer in their FMP, can be interpreted as indicating that a more negative sentiment in the migration-related news delays the migrants’ journey toward their destination countries.29 Finally, in Table 1 column 4, we explore the possible mechanisms explaining the effect 26 Table A3 shows that the tone of the news has no effect on the movements of short-stayer migrants at the FMP, i.e., migrants who have been at the FMPs for less than 2 weeks, between 2 weeks and 3 months, and between 3 and 6 months. 27 While all these controls are not significant at conventional levels, the signs are in line with our expectations: a higher local economic activity and lower conflict intensity induce migrants to stay longer in the FMP. 28 A more negative tone of the news may also have an impact on the choice of migrants to remain in Libya as the final destination, i.e., the variable “FMP with migrants planning to remain in Libya” may be endogenous with respect to the main explanatory variable, T one N ews Destinationsi,m−1,t . Reassuringly, when the latter variable is regressed against the former, controlling for both FMP and month-year fixed effects, its estimated effect is small (0.010) and not statistically significant (p-value = 0.622), suggesting that the choice to remain in Libya is not determined by a change in the tone of the news in destination countries. 29 A possible concern with our interpretation of these findings is related to the role of smugglers in determining migrants’ movement decisions in Libya. Our data in fact does not allow us to distinguish between voluntary or forced choices in the timing (and direction) of the migration journey. However, as discussed in Section 3, two elements support our interpretation of these findings as being the result of migrants choosing when and where to move within Libya. First, our sample of migrants does not include migrants in detention centers or victims of human trafficking. Second, the ethnographic literature on smuggling documents that migrants in Libya (most often) choose when and where to do the next step of their migration journey toward these countries. 11 of changes in the sentiment news in destination countries on migrants’ movement choices in Libya. To this end, we modify our baseline specification by replacing our main explanatory variable with a set of five variables, each measuring the tone of (migration-related) news articles referring to one specific topic, and one variable measuring the tone of the news of all the other (migration-related) news articles not covering these five topics. The five specific topics are: 1) discrimination; 2) sea incidents; 3) economic conditions; 4) crime; 5) (migration-related) government announcements. Results indicate that among all these topics, the only one with a statistically significant effect is discrimination : a more negative tone in migration-related news articles on discrimination episodes induces migrants to postpone the next step in their migration journey and stay longer in their current location. Changes in the tone of news articles focusing on all the other topics, including sea incidents or (migration-related) government announcements, do not seem to have an effect. One possible interpretation of this finding is that the sentiment of news articles on discrimination episodes predicts the attitude of the local population in the destination country towards migration and the level of difficulty that migrants may encounter in terms of job opportunities, access to housing, and social integration (World Bank, 2023; Aksoy et al., 2023).30 A more negative tone of the news indicates that these conditions at destination may be getting worse and that integration is becoming more difficult, reducing the incentive for the individual to make the next step in the migration journey.31 Validation To corroborate these results, we conduct three validation exercises. First, we show that the effect of tone of the news is significant only in FMPs located in areas in which it is more likely that migrants have access to the Internet. Second, we conduct two different placebo tests showing that our results are not spurious. Finally, we provide suggestive evidence that our results are unlikely to be driven by reverse causality. Internet access Our finding that migrants’ movement choices in Libya are influenced by the tone of news published in their preferred destination countries rests on the idea that migrants are able to access and read the news. While we do not have information on mobile ownership nor on individual-level Internet usage, Table 2 provides suggestive evidence that the tone of the news affects only migrants who are more likely to be able to access them. Column 1 shows that the effect of the tone of the news in the destination countries increases with the level of economic activity around the FMP (as proxied by night lights). As long as economic activity positively correlates with Internet access, these results indicate that the effect is stronger in areas where migrants are more likely to read the news and to use it to take their decisions. Columns 2 and 3 offer additional support to this interpretation by using information on internet coverage in Libya. Results show that the effect of the tone of the news is significant only where there is Internet coverage and there is a high-speed connection. The fact that the changes in the tone of the news have an effect only in the locations where migrants are more likely to actually access 30 Aksoy et al. (2023) show that attitudes towards migrants (proxied by a “sentiment index” built using Twitter data) are as important as local unemployment rates in shaping refugees’ integration outcomes in Germany. 31 Note that for our argument to be valid, migrants do not need to be able to read the news articles (e.g., to know the language of the destination country) nor to actually read them. It is enough to assume that someone in contact with the migrants can read the news articles published in the destination country and then communicate to them the sentiment of the news articles. 12 and read the news articles corroborates our interpretation of the main findings. Placebo tests Figure A4 shows the results of two placebo tests. In the first one (displayed in panel a), we randomly assign to migrants in a given FMP the tone of the news published in the preferred destination countries of the migrants in another FMP. If migrants’ movements are influenced by the tone of the news in their destination countries, we should expect a null effect of the tone of the news in other countries. In the second test (displayed in panel b), we assign to migrants in a given FMP the tone of the news in their destination countries but published in a random month. If migrants’ movements are influenced by the tone of the news in destination countries in the month before, we should expect a null effect for the news articles published in any other period. Figure A4 shows that, in both cases, the average placebo effect is almost zero and is considerably smaller with respect to the observed effect of the tone of the news.32 Reverse causality One possible concern with our results is that of reverse causality, namely that migration movements within Libya influence the tone of the news in destination countries. Two possible mechanisms may lead to this. The first is that the tone of the news in destination countries is determined by the number of migrants potentially moving to Europe. A reduction in the share of long-stayers - which we interpret as an increase in the number of migrants moving North, possibly leading to more attempts to reach Europe - is thus expected to worsen the tone of the news in destination countries. In this case, the two variables would be positively correlated. In fact, our analysis shows the opposite result, thus excluding the existence of this potential source of reverse causality. Another possibility is that the tone of the news is influenced by the size of migration in Libya. The more migrants are in Libya, the larger the number of those that – at some point – will attempt to reach Europe. In this case, the tone of the news would be influenced by the overall number of migrants in the country, in particular, the larger the number of migrants the more negative the tone of the news. To test for this, we look at the correlation between the number of migrants in all FMPs in a given month and the tone of the news in the following month in the most important destination countries (Italy, France, and Germany). We also do the same with only the FMPs located close to the northeastern coastal border, i.e., the area from where migrants most often attempt to reach Europe. As shown in Figures A5 and A6, the correlation between these variables in both cases is practically zero. 6.2 Robustness Our results are robust to a number of checks. We show them in Table 3. Alternative measure of the tone of the news Column 1 shows that the magnitude of the effect is unchanged if we use as explanatory variable the median value of the tone (i.e., the most common tone) rather than the average tone of migration-related news articles. 32 For Panel (a), 5 observations out of 10k are lower than the observed effect: i.e., probability of error is 0.0005%. For Panel (b), 2 observations are lower than the observed effect: i.e., probability of error is 0.0002%. 13 Measurement error in the tone of the news The variable T one N ews Destinationsi,m−1,t is constructed considering the news article published in the destination countries of the migrants located at the FMP. Because the majority of migrants are long stayers (see Fig. A3), our analysis makes the plausible assumption that these destinations reflect the preferences of this group of migrants. However, this may not always be the case. In order to deal with this issue, we restrict the analysis to those cases in which the top preferred destinations in the FMP are most likely to represent those of long stayers: i.e., when the percentage of long stayers in the FMP is very high at month m (equal to or more than 75%), and the number of arrivals at month m+1 is equal to zero. Column 2 shows that our results are qualitatively unchanged. Change in the number of migrants in the FMP One possible concern with our results is that the reduction in the share of long-stayer migrants in the FMP may be mechanically driven by arrivals. Column 3 shows that this is not the case: our results still hold when we restrict the analysis only to FMPs in which there are no arrivals at month m. Alternative estimator When the dependent variable is a share, OLS estimates can be po- tentially biased. In these situations, one alternative is to use a fractional probit estimator. Results in column 4 show that our findings are confirmed also when using this estimator. Time trends in origin and destination countries Our results are robust to augmenting our model specification with country-of-origin specific time-trends (see column 5). These fixed effects control for all the (linearly) time-varying characteristics in the origin country that may influence the movement decision of any national group of migrants. Our results also hold when we add to our baseline specification country-of-destination specific time-trends accounting for all the characteristics which (linearly) change across time in the destination country that may influence the movement decision of any national group of migrants (see column 6). While Libya is still an important destination country for foreign workers, for a large number of migrants it has become a transit country toward Europe. Most of these migrants reach the coastal Northern border of the country to then attempt the crossing of the Mediterranean Sea (see Section 3). One possible concern with our results is that changes in the tone of the news in fact capture changes in migration policies or in the intensity of patrolling of the Mediterranean Sea rather than in the news sentiment in destination countries. Results in Table 1 column 4 already show that news related to Sea incidents and Government announcements is not what is driving our results. In the following, we provide additional evidence that this is not the case. Political orientation of the government in destination countries One possible con- founder of the effect of tone of the news is a change in the political orientation towards migra- tion of the government of the destination country. To test for this, we include, as additional controls in our baseline specification, a variable measuring the Rile Index of the parties in the government coalition and a dummy to capture changes in the cabinet compositions for each of the five most important destination countries. Table 3, column 7 shows that the effect of 14 changes in the tone of the news is significant above and beyond the changes in the political orientation towards migration of the governments in destination countries. Sea patrolling intensity and Mediterranean Sea crossing A possible concern with our results is that changes in the tone of the news could be mirroring the tightness of the patrolling activities in the Mediterranean Sea having the objective of stopping migrants’ attempts to reach Europe. In other words, the slowing down of migrants’ movements in Libya could be due to the increase in the difficulty and riskiness in crossing, reflected in a more negative tone of the news articles on migration. During our period of analysis, there have been several changes in the extent and intensity of both sea patrolling activities and Search and Rescue Operations (SARs) in the Mediterranean Sea conducted by both the Libyan and the European authorities, and the NGOs. The (net) effects of these different operations are unclear, being highly dependent on the data used and the period considered. Among these various phases, the 2017 Minniti Compact represents a turning point in the Italian government’s attempt to reduce the arrivals on its coasts (Marchesi, 2021; Deiana et al., 2023). Given the large discontinuity in the modality of the patrolling activity due to this policy change, it is difficult to compare the dynamics of the crossing of the Mediterranean Sea pre and post the Minniti Compact. For this reason, in our analysis of the patrolling activity as a mediator of the effect of tone of the news on migrants’ choices, we exclude the implementation period of the Minniti Compact and we focus on the period 2018-2020. For our test, we define months with high or low-intensity of patrolling using the actual flows of migrants arriving in Italy in this period (see Section 4.3). Column 8 shows that the effect of the tone of the news on movements in Libya is significant in both periods and the magnitude of the effect is not stronger in months with high intensity patrolling. Spatial analysis Our results are robust to the use of alternative regression models to account for the possible presence of spatial autocorrelation in the migrants’ movement choices across FMPs. To this end, we re-estimate equation 1 using a Spatial Error Model (SEM) and a Spatial Lag Mode (SLM). Results reported in Table A5 show that for both models the effect is statistically significant (and higher than in our baseline model), confirming that changes in the tone of the news affect the movement decisions of migrants (for more details, see Section D). 6.3 Heterogeneity In the following, we discuss a set of heterogeneity results that help us characterize the conditions under which changes in the tone of the news affect migrants’ movement choices. 6.3.1 Country of origin and FMP location Table 4 looks at the differential effect of the tone of the news depending on the migrants’ country of origin and their current location in Libya. As discussed in Section 3, migrants from West African countries travel through Libya differently with respect to migrants from East African countries: the former group most often follows a multi-step journey towards the North, while the latter usually makes only a few stops before reaching the coastal areas (see UNHCR, 2017b; Di Maio et al., 2023). Results reported in Table 4, column 1, are in line with this anecdotal 15 evidence: the effect of the tone of the news is significant only for migrants coming from West African countries. For the others, the sentiment toward migrants in destination countries, no matter how negative it is, does not influence their journey through Libya. This is also consistent with the fact that migrants coming from East African countries are often escaping violence and conflict, heading directly to Europe to apply for a refugee status (UNHCR, 2019a). Next, we consider the differential effects of the tone of the news on the movements of migrants located in different areas of Libya. Results in Table 4 column 2 show that a more negative tone of the news slows down the movements of migrants only in FMPs located in the Western part of the country. This differential effect is explained by two elements. First, the main migration route to Europe passes through the Western region of Libya (while migration is low from the Eastern one, see Section 3). Second, this is the part of the country where economic activity is higher (Del Prete et al., 2023). The availability of job opportunities allows migrants to (temporarily) interrupt their journey by choosing to work and live there for some time. It follows that these locations are where we expect changes in news sentiment to influence migrants’ movement. 6.3.2 Prevailing sentiment and numerosity of migration-related news Table 5 column 1 looks at the differential effect of changes in the tone of the news depending on the prevailing sentiment (positive vs negative) of migration-related news in that period. The results indicate that migration choices are only influenced by changes in the tone of the news in months during which migration-related news are mostly positive. Instead, in months in which news are mostly negative or there is an equal number of positive and negative news, the change in the tone of the news has no effect on migrants’ decisions. Column 2 further explores this heterogeneity by checking if the effect of the tone of the news varies depending on how relevant migration is in the public discourse in that period. Results show that the effect is significant only in months in which there are many news items on migration. Finally, column 3, which combines these two aspects, shows that migrants’ journey plans are influenced by changes in the tone of the news only in periods during which there are many news items on migration and they are mostly positive. Taken together, these results speak to the effectiveness of the strategy - pursued by several right-wing political parties in Europe - to incite a negative sentiment toward migrants in destination countries as a mean of discouraging them from heading to Europe. Our results indicate that this strategy is unlikely to have a lasting effect: the exacerbation of an already negative sentiment toward migration does not affect migrants’ plans, possibly because they may have already discounted the existence of a negative sentiment against them when making their initial migration decision. 6.3.3 Economic conditions in destination countries Economic conditions in destination countries are an important driver of migration flows and thus may influence how the tone of the news impacts migrants’ choices. To account for this, in Table 6 column 1 we include GDP per-capita and employment rate in the destination countries as additional controls. As expected, better economic conditions in destination countries tend to reduce the duration of migrants’ stay in their current location in Libya, i.e., suggesting that these factors tend to make migrants move faster toward their final destination. Importantly, 16 the effects of the tone of the news is still significant (at 10%) and its magnitude is only slightly reduced with respect to the baseline specification. Next, columns 2 and 3 show that a worsening tone of the news induces migrants to delay their next migration movement only when the economic conditions in the destination country are unfavorable (low employment rate or low GDP per capita). Conversely, during periods of favorable economic conditions, a worsening tone of the news has no effect. We interpret these results as indicating that economic conditions in destination countries are a strong pull factor: good economic opportunities can offset the (expected) high discrimination associated with a negative sentiment towards migrants, the latter becoming more costly and binding when the economic conditions are instead unfavorable. Summary of the results Taken together, these heterogeneity results characterize the con- ditions under which changes in news sentiment affect migrants’ choices. A worsening tone of the news slows down the migrants’ movements in Libya - delaying their journey towards the preferred destination countries - only for migrants coming from West African countries (who are those following a step-by-step journey) or located in the Western part of Libya (where there are more economic opportunities), and when the migration-related news articles are numerous and mostly with a positive tone, or economic conditions in destination countries are not favorable. 6.4 Additional results: The tone of the news and spillover effects across destination countries Our main analysis looks at the effect of changes in tone of the news in the destination country on migrants’ movements within Libya. At the same time, one may expect that changes in the tone of the news in the destination country could also have spillover effects on other countries, e.g., potentially making them more attractive for migrants. For instance, migrants may substitute across destination countries by excluding those in which the news sentiment toward them is more negative. To test for this possibility, we estimate the following regression model: P ref erredd,m,t = β0 + β1 T one N ews Destination Highd,m−1,t + φm,t + µd + d,m,t (2) where P ref erredd,m,t is the share of FMPs for which country d in month m in year t is a preferred destination. T one N ews Destinations Highd,m−1,t is a dummy variable that takes value one if the tone of the news of country d in month m − 1 is higher than the median value of the distribution of the tone of the news in close-by destination countries, and 0 otherwise.33 φm,t is the month-year fixed effects; µd is the destination country fixed effects, and d,m,t is the error term. To account for the possibility that the set of the preferred destination countries varies only because of the change in the composition of the migrants present in the FMP, we restrict the analysis only to FMPs for which the number of migrants does not change across two periods. Results in Table 7 suggest that changes in the tone of the news in a destination country have spillover effects on other countries. Column 1 shows the OLS results: a more negative tone of the news in a country (with respect to other countries) reduces the share of FMPs for 33 We define close-by countries as those in the same continent, with the idea that a meaningful comparison group is that of countries which are similar and not too much geographically distant from each other. 17 which that country is a preferred destination among migrants in Libya. Column 2 shows that this result is confirmed also when we use a Spatial Lag Model (SLM) to account for the fact that a lower preference for country d may be mechanically correlated with a higher preference for d ’s close-by countries.34 Finally, column 3 shows that the results still hold when controlling for economic conditions in the destination country. A more negative tone of the news in the preferred destination induces migrants to switch to another country. One important related question is whether this substitution occurs between European and non-European destination countries. This seems not to be the case. To begin with, Figure A7 shows no correlation between the tone of the news in European destination countries and the number of FMPs registering an increase in non-European countries as pre- ferred destinations in the following period. This is also confirmed by results in Table A4 showing that there is no correlation between the change (or the difference) between month m and m + 1 in the number of returnee countries which are the preferred destinations of migrants in FMP i and our main explanatory variable, i.e., the tone of the news in the destination countries of migrants located in FMP i in the previous period. Taken together, these results show that a more negative tone of the news in European destination countries is not associated with an increase in returnee countries among the most preferred destinations for migrants in the next period. We interpret these findings as suggesting that a worsening tone of the news - while it induces a substitution across European destinations - does not make migrants return to their country of origin. 7 Concluding remarks This paper provides evidence that migrants’ movements in Libya and the timing of their jour- ney towards their destination countries are influenced by changes in the news sentiment of migration-related articles in these same countries. Importantly, we document that these effects are significant only for some groups of migrants and under specific conditions, indicating an overall limited effect of a more negative tone of the news on migrants’ movements towards destination countries. Relying on IOM data providing information on migrants located in 167 FMP in Libya for the period 2017-2020, we show that a more negative tone of migration-related news articles published in migrants’ destination countries increases their length of stay in Libya, slowing down their migration journey. To validate our findings, we provide evidence that the effect is significant only for migrants located in areas where it is more likely that they have internet access and that our results are not due to reverse causality. Our heterogeneity analysis shows that the effect is significant only for migrants coming from West African countries and located in the Western part of the country. Moreover, a worsening tone of the news has no effect in periods during which most of migration-related news articles have a negative tone or the economic conditions in destination countries are very good. Taken together, these results indicate that the effect of changes in news sentiment on migrants’ decisions is significant when, where, and for whom we expect it to be so. Yet, they also suggest a limited impact on overall migration movements. 34 The negative and significant ρ indicates that this is the case: an increase in the likelihood that d is a preferred country corresponds to a reduction in the preference for another country (for details, see Appendix D). 18 Given that most migrants in Libya have reached the country after a risky and difficult journey, it is not surprising that a more negative news sentiment makes only a few of them change their initial plan and decide not to reach their final destination anymore. Finally, we document that a worsening news sentiment, while it induces a substitution across destinations, does not lead migrants to decide to return to their country of origin. This paper provides the first assessment of how changes in the sentiment of migration-related news in destination countries may affect migrants’ choices in developing countries. More specif- ically, our findings contribute to the understanding of the drivers of migration in Libya and to the policy debate on how to control migration flows to Europe. Our results also provide a contribution to the public debate on migration in European countries. In particular, our results suggest that a strategy to reduce migration based on the deterrence effects of adopt- ing a very negative attitude towards migrants is - at best - short-sighted and likely to have a marginal effect. To begin with, the overall effect of worsening the tone of the news on mi- grants’ choices is limited, being significant only for some groups of migrants and under some specific conditions. For instance, a worsening tone of the news has no effect when most of the migration-related news articles already have a negative tone, suggesting that the deterrence effect of using an increasingly negative tone is small and its marginal effect tends rapidly to zero. Thus a political campaign based on an aggressive anti-migration discourse is unlikely to have a lasting effect in terms of reducing migration flows: the exacerbation of an already negative sentiment toward migration does not affect migrants’ plans, possibly because they may have already discounted the existence of a negative sentiment against them when making their choice of the destination country. Similarly, a more negative news sentiment has no effect when the economic conditions at destination are favorable: good economic opportunities offset the (expected) high discrimination associated with a negative sentiment towards migrants. Again, under these conditions, an aggressive anti-migration discourse is unlikely to have any effect. Finally, the fact that changes in the news sentiment generate spillover effects across countries indicates that such anti-migration strategy is also not politically sustainable: a worsening tone of the news in one destination increases migration flows to other European countries while it does not induce migrants in Libya to return to their country of origin. This implies that an aggressive and negative attitude towards migrants will not reduce the overall migration pres- sure on advanced countries. It may however undermine the possibility of cooperation between destination countries to regulate migration flows from developing countries. References Achilli, L. (2021). Human smuggling in the mediterranean: A comparative analysis of the central and eastern mediterranean smuggling routes. In E. Savona, R. Guerette, and A. Aziani (Eds.), The Evolution of Illicit Flows: Displacement and Convergence among Transnational Crime. Springer. Adema, J. A. H., C. G. Aksoy, and P. Poutvaara (2022). Mobile internet access and the desire to emigrate. CESifo Working Paper No. 9758. 19 Adesina, O. (2021). Libya and african migration to europe. In O. Abegunrin and O. Abidde (Eds.), African Migrants and the Refugee Crisis. Springer. Aksoy, C. G., P. Poutvaara, and F. Schikora (2023). First time around: Local conditions and multi-dimensional integration of refugees. Journal of Urban Economics 137, 103588. Artuc, E. and C. Ozden (2018). Transit migration: All roads lead to America. Economic Journal 128 (612), F306–F334. Bah, T. L., C. Batista, F. Gubert, and D. McKenzie (2023). Can information and alternatives to irregular migration reduce “backway” migration from the gambia? Journal of Development Economics 165, 103153. Baseler, T. (2023). Hidden income and the perceived returns to migration. American Economic Journal: Applied Economics 15, 321–352. Battiston, G. (2022). Rescue on stage: Border enforcement and public attention in the mediter- ranean sea. Mimeo . Bayer, R., R. Hodler, P. A. Raschky, and A. Strittmatter (2020). Expropriations, property con- fiscations and new offshore entities: Evidence from the Panama Papers. Journal of Economic Behavior & Organization 171, 132–152. Beine, M., F. Docquier, and C ¨ ¸ . Ozden (2015). Dissecting network externalities in international migration. Journal of Demographic Economics 81 (4), 379–408. Benesch, C., S. Loretz, D. Stadelmann, and T. Thomas (2019). Media coverage and immigration worries: Econometric evidence. Journal of Economic Behavior and Organization 160, 52–67. Bertoli, S., J. F. H. Moraga, and L. Guichard (2020). Rational inattention and migration decisions. Journal of International Economics 126, 103364. ohme, M. H., A. Gr¨ B¨ ohr (2020). Searching for a better life: Predicting inter- oger, and T. St¨ national migration with online search keywords. Journal of Development Economics 142, 102347. Bove, V., G. Efthyvoulou, and H. Pickard (2023). Government ideology and international migration. The Scandinavian Journal of Economics 125.1. Bracco, E., M. De Paola, C. P. Green, and V. Scoppa (2018). The effect of far right parties on the location choice of immigrants: Evidence from Lega Nord Mayors. Journal of Public Economics 166, 12–26. Consoli, S., L. T. Pezzoli, and E. Tosetti (2021). Emotions in macroeconomic news and their impact on the European bond market. Journal of International Money and Finance 118, 102472. Couttenier, M., S. Hatte, M. Thoenig, and S. Vlachos (2023). Anti-muslim voting and media coverage of immigrant crimes. Review of Economics and Statistics forthcoming. 20 Cummings, C., J. Pacitto, D. Lauro, and M. Foresti (2015). Why people move: Understanding the drivers and trends of migration to Europe. Overseas Development Institute, London. Deiana, C., V. Maheshri, and G. Mastrobuoni (2024). Unintended consequences of search and rescue operations. American Economic Journal: Economic Policy, forthcoming. Deiana, C., G. Mazzarella, E. C. Meroni, and L. Tiozzo Pezzoli (2023). The unexpected influ- encer: Pope Francis and European perceptions of the recent refugee crisis. Oxford Economic Papers 75 (1), 75–95. Del Prete, D., M. Di Maio, and A. Rahman (2023). Firms amid conflict: Performance, produc- tion inputs, and market competition. Journal of Development Economics 154 (103143). Di Maio, M., F. Fasani, V. Leone Sciabolazza, and V. Molini (2023). Facing displacement and a global pandemic: Evidence from a fragile state. Journal of Demographic Economics , 1–26. Di Maio, M., V. Leone Sciabolazza, and V. Molini (2023). Migration in Libya: A spatial network analysis. World Development 163, 106139. c, S. (2017). Transit migration. Review of International Economics 25 (5), 1017–1045. Djaji´ Djourelova, M. (2023). Persuasion through slanted language: Evidence from the media coverage of immigration. American Economic Review 113, 800–835. e, L. and F. Fasani (2013). Media exposure and internal migration—Evidence from Indone- Farr´ sia. Journal of Development Economics 102, 48–61. Fitzgerald, M. and M. Toaldo (2016). A quick guide to libya’s main players. European Council on Foreign Relations . Friebel, G.and Manchin, M., M. Mendola, and G. Prarolo (2023). International migration intentions and illegal costs: Evidence using africa-to-europe smuggling routes. Journal of International Economica forthcoming. Giulietti, C., J. Wahba, and Y. Zenou (2018). Strong versus weak ties in migration. European Economic Review 104, 111–137. Grubanov-Boskovic, S., S. Kalantaryan, S. Migali, and M. Scipioni (2021). The impact of the internet on migration aspirations and intentions. Migration Studies 9 (4), 1807–1822. IOM (2015). Migration Trends Across the Mediterranean: Connecting the Dots. International Organization for Migration, Geneva. IOM (2016). Displacement Tracking Matrix, DTM Round 1. International Organization for Migration, Geneva. IOM (2018a). Displacement Tracking Matrix (DTM). Libya’s Migrant Report (Round 18). March, 2018. International Organization for Migration, Geneva. 21 IOM (2018b). Displacement Tracking Matrix (DTM). Libya’s Migrant Report (Round 23). November - December, 2018. International Organization for Migration, Geneva. IOM (2019). Living and Working in the Midst of Conflict: The Status of Long-term Migrants in Libya. International Organization for Migration, Geneva. Irhiam, H. R., M. G. Schaeffer, and K. E. Watanabe (2023). The Long Road to Inclusive Institutions in Libya: A Sourcebook of Challenges and Needs. International Development in Focus series. Washington, DC: World Bank. Marchesi, M. (2021). Can migration flows be controlled? the effects of the italian 2017–2018 policy against illegal immigration. In E. Savona, R. Guerette, and A. Aziani (Eds.), The Evo- lution of Illicit Flows: Displacement and Convergence among Transnational Crime. Springer. McKenzie, D. and H. Rapoport (2010). Self-selection patterns in Mexico-US migration: The role of migration networks. Review of Economics and Statistics 92 (4), 811–821. omb¨ Meltzer, C., J. Eberl, N. Theorin, T. Heidenreich, J. Str¨ ack, H. G. Boomgaarden, and C. Schemer (2021). Media effects on policy preferences toward free movement: evidence from five eu member states. Journal of Ethnic and Migration Studies 47:15, 3390–3408. Micallef, M. (2017). The Human Conveyor Belt: Trends in human trafficking and smuggling in post-revolution Libya. Global Initiative against Transnational Organized Crime. Mixed Migration Centre (2019). What makes refugees and migrants vulnerable to detention in Libya? A microlevel study of the determinants of detention. MMC Research Report, December. Mixed Migration Hub (2015). Detained Youth: The Fate of Young Migrants, Asylum-Seekers, and Refugees in Libya Today. Murphy-Teixidor, A., A. A. Bonfiglio, and V. Leigh (2020). Migration and risks: smuggling networks and dynamics on the central mediterranean route. In F. Fargues and M. Rango (Eds.), Migration in West and North Africa and across the Mediterranean: Trends, risks, development and governance. Berlin: IOM GMDAC. Pesando, L. M., V. Rotondi, M. Stranges, R. Kashyap, and F. C. Billari (2021). The interneti- zation of international migration. Population and Development Review 47 (1), 79–111. Rahman, A. and M. Di Maio (2020). The private sector amid conflict: The case of Libya. The World Bank. Washington D.C. (USA). Raleigh, C., A. Linke, H. Hegre, and J. Karlsen (2010). Introducing acled: an armed conflict location and event dataset: special data feature. Journal of peace research 47 (5), 651–660. ujo, P. Gon¸ Ribeiro, F. N., M. Ara´ e Gon¸ calves, M. Andr´ calves, and F. Benevenuto (2016). Sentibench-a benchmark comparison of state-of-the-practice sentiment analysis methods. EPJ Data Science 5, 1–29. 22 Sanchez, G. (2020). Migrant smuggling in the libyan context: re-examining the evidence. In P. Fargues and M. Rango (Eds.), Migration in west and north Africa and across the Mediter- ranean: Trends, risks, development and governance, pp. 230–238. Berlin: IOM GMDAC. Shrestha, M. (2019). Death scares: How potential work-migrants infer mortality rates from migrant deaths. Journal of Development Economics 141, 102368. Tjaden, J. and F. A. Dunsch (2021). The effect of peer-to-peer risk information on potential migrants–evidence from a randomized controlled trial in senegal. World Development 145, 105488. UN-DESA (2020). International Migrant Stock Database. Department of Economic and Social Affairs of the United Nations Secretariat, New York (USA). UNHCR (2017a). Global Trends: Forced Displacement in 2016. United Nations Human Rights Council. Geneva (CH). UNHCR (2017b). Mixed Migration Trends in Libya: Changing Dynamics and Protection Chal- lenges. United Nations Human Rights Council. Geneva (CH). UNHCR (2019a). From hand to hand: The migratory experience of refugees and migrants from East Africa across Libya. United Nations Human Rights Council, Geneva, CH. UNHCR (2019b). Mixed Migration Routes and Dynamics in Libya in 2018. United Nations Human Rights Council, Geneva, CH. UNSMIL (2018). Desperate and Dangerous: Report on the human rights situation of migrants and refugees in Libya. Office of the United Nations High Commissioner for Human Rights (OHCHR) and United Nations Support Mission in Libya (UNSMIL). Wilson, R. (2021). Moving to jobs: The role of information in migration decisions. Journal of Labor Economics 39 (4), 1083–1128. Wittenberg, L. (2017). Managing Mixed Migration: The Central Mediterranean Route to Eu- rope. New York: International Peace Institute. World Bank (2015). Labour Markets Dynamics in Libya. The World Bank. Washington D.C. (USA). World Bank (2018). Asylum seekers in the European Union: Building evidence to inform policy making. Washington, D.C.: World Bank. World Bank (2023). World Development Report 2023: Migrants, Refugees, and Societies. Wash- ington, DC: World Bank. 23 Tables Table 1: Tone of migration-related news at destination and migrants’ movements Share of migrants in the FMP who are long-stayers (1) (2) (3) (4) Tone of the news at destination -0.0314** -0.0312** -0.0312** (0.0172) (0.0177) (0.0177) Tone of the news at destination (related to) - Discrimination -0.0290* (0.0165) - Sea incidents -0.0180 (0.0200) - Economic conditions -0.0109 (0.0260) - Crime 0.0089 (0.0227) - Government announcements 0.0128 (0.0203) - Other -0.0073 (0.0222) Intensity of night lights within 5 km 0.0097 0.0099 0.0030 (0.0593) (0.0592) (0.0643) Number of conflicts within 5 km -0.0101 -0.0102 -0.0166 (0.0179) (0.0179) (0.0182) FMP with migrants planning to remain in Libya (1 = Yes) 0.0019 0.0129 (0.0270) (0.0280) return home (1 = Yes) 0.0045 0.0132 (0.0369) (0.0411) FMP FEs Yes Yes Yes Yes Year-month FEs Yes Yes Yes Yes Country of destination FEs Yes Yes Yes Yes Country of origin FEs Yes Yes Yes Yes Number of Observations 1,568 1,568 1,568 1,568 Notes: OLS regression results. Standardized OLS estimated coefficients are reported. Standardization of coeffi- cients is obtained using the formula sd (x) β , where βx is the point estimate associated to control variable x, while sd(y ) x sd(x) and sd(y ) indicate the standard deviation or respectively control variable x and dependent variable y . Stan- dard errors clustered at the FMP level in brackets. The dependent variable Share of migrants in the FMP who are long-stayers is the share of migrants in the FMP i in month m in year t who report being in that location for more than 6 months. The variable Tone of the news at destination (T one N ews Destinationsi,m−1,t ) is the average tone of migration-related news in the top destination countries for migrants in FMP i in month m − 1 in year t. For the definitions of the variables: T one N ews Destinations Discrimination; T one N ews Destinations Crime; T one N ews Destinations Economic conditions; T one N ews Destinations Sea incidents; T one N ews Destinations Government announcements, and T one N ews Destinations Other see Sec- tion C. All other variables are defined in Tables A1 and A2 and additional details are provided in Section 4. *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. 24 Table 2: Validation: Internet access Share of migrants in the FMP who are long-stayers (1) (2) (3) Tone of the news at destination in FMPs where night lights have Low intensity 0.0342 (0.0599) Medium intensity -0.0464 (0.0456) High intensity -0.1234*** (0.0496) FMPs with network coverage Not present -0.0787 -0.0708 (0.1077) (0.1061) Present (with) -0.0309* (0.0172) High-speed connection -0.0437* (0.0228) Low-speed connection 0.0223 (0.0185) Other controls Yes Yes Yes FMP FEs Yes Yes Yes Year-month FEs Yes Yes Yes Country of destination FEs Yes Yes Yes Country of origin FEs Yes Yes Yes Number of Observations 1,568 1,568 1,568 Notes: OLS regression results. Standardized OLS estimated coefficients are reported. Standardization of coefficients is sd(x) obtained using the formula sd(y) βx , where βx is the point estimate associated to control variable x, while sd(x) and sd(y ) indicate the standard deviation or respectively control variable x and dependent variable y . Standard errors clustered at the FMP level in brackets. The dependent variable Share of migrants in the FMP who are long-stayers is the share of migrants in the FMP i in month m in year t who report being in that location for more than 6 months. The variable Tone of the news at destination (T one N ews Destinationsi,m−1,t ) is the average tone of migration-related news articles published in the top destination countries for migrants in FMP i in month m − 1 in year t.“Other controls” include: Intensity of night lights within 5 km ; Number of conflicts within 5 km ; FMP with migrants planning to remain in Libya (1 = Yes) ; FMP with migrants planning to return home (1 = Yes). All other variables are defined in Tables A1 and A2 and additional details are provided in Section 4. *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. 25 Table 3: Tone of migration-related news at destination and migrants’ movement: Robustness checks Share of migrants in FMP who are long-stayers (1) (2) (3) (4) (5) (6) (7) (8) Tone of the news at destination (median) -0.0312* (0.0177) Tone of the news at destination -0.4025*** -0.1052* -0.1582*** -0.0331* -0.0255* -0.0300* (0.1302) (0.0564) (0.0503) (0.0170) (0.0163) (0.0172) Tone of the news at destination when High intensity patrolling -0.13921* (0.07281) Low intensity patrolling -0.1020* (0.0550) Other controls Yes Yes Yes Yes Yes Yes Yes Yes FMP FEs Yes Yes Yes Yes Yes Yes Yes Yes Year-month fixed FEs Yes Yes Yes Yes Yes Yes Yes Yes Country of destination FEs Yes Yes Yes Yes Yes Yes Yes Yes 26 Country of origin FEs Yes Yes Yes Yes Yes Yes Yes Yes Country of destination * time trend No No No No Yes No No No Country of origin * time trend No No No No No Yes No No Government political orientation controls No No No No No No Yes No Wald Test [p.value] 0.24 [0.62] Estimator OLS OLS OLS Fractional OLS OLS OLS OLS Probit Sample Full Resticted No Arrivals Full Full Full Full After 2017 Number of Observations 1,568 201 239 1,568 1,568 1,568 1,568 1,024 Notes: Results from OLS in columns 1 to 3, and 5 to 8. Results from Fractional Probit in column 4. For columns 1-3, and 5-8, standardized estimated coefficients are reported. Standardization of coefficients is obtained using the formula sd (x) β , where βx is the point estimate associated to control variable x, while sd(x) and sd(y ) indicate sd(y ) x the standard deviation or respectively control variable x and dependent variable y . Standard errors clustered at the FMP level in brackets. The dependent variable Share of migrants in the FMP who are long-stayers is the share of migrants in the FMP i in month m in year t who report being in that location for more than 6 months. The variable Tone of the news at destination (T one N ews Destinationsi,m−1,t ) is the average tone of migration-related news articles published in the top destination countries for migrants in FMP i in month m − 1 in year t. Restricted sample includes only FMPs for which the percentage of long stayers in the FMP at month m is equal or more than 75%, and the number of arrivals at month m+1 is equal to zero. No Arrivals sample includes only FMPs in which no migrants arrived at month m. The Wald χ2 test in column (7) evaluates the statistical difference of point estimates between variables Tone of the news at destination when High Intensity Patrolling and Tone of the news at destination when Low Intensity Patrolling. The variable “Tone of the news at destination (median)” is equal to the average value of the median tone of migration-related news published in the destination countries for migrants in FMP i in month m-1. “Other controls” include: Intensity of night lights within 5 km ; Number of conflicts within 5 km ; FMP with migrants planning to remain in Libya (1 = Yes) ; FMP with migrants planning to return home (1 = Yes). “Government orientation controls“ include: the Rile Index Coalitioni,m,t for each of the main destination countries (i.e., Italy, France, Germany, Spain, and UK) and the dummy variable Change of cabineti,m,t . All other variables are defined in Tables A1 and A2 and additional details are provided in Section 4. *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. Table 4: Heterogeneity: Migrants’ country of origin and FMP location Share of migrants in FMP who are long-stayers (1) (2) Tone of the news at destination in FMPs with a major group of migrants from West Africa -0.0710*** (0.0273) East Africa 0.0664 (0.0455) FMPs located in West Libya -0.1570** (0.0698) East Libya -0.0348 (0.0486) Other controls Yes Yes FMP FEs Yes Yes Year-month FEs Yes Yes Country of destination FEs Yes Yes Country of origin FEs Yes Yes Number of Observations 1,568 1,568 Notes: OLS regression results. Standardized OLS estimated coefficients are reported. Standardization of coefficients is sd(x) obtained using the formula sd(y) βx , where βx is the point estimate associated to control variable x, while sd(x) and sd(y ) indicate the standard deviation or respectively control variable x and dependent variable y . Standard errors clustered at the FMP level in brackets. “Other controls” include: Intensity of night lights within 5 km ; Number of conflicts within 5 km ; FMP with migrants planning to remain in Libya (1 = Yes) ; FMP with migrants planning to return home (1 = Yes). All variables are defined in Tables A1 and A2 and additional details are provided in Section 4. *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. 27 Table 5: Heterogeneity: Prevailing news sentiment and numerosity of migration- related news Share of migrants in FMP who are long-stayers (1) (2) (3) Tone of the news at destination (during months when) Prevailing news sentiment is Mostly negative 0.0070 0.0072 (0.0242) (0.0243) Same number of negative and positive news -0.0254 -0.0240 (0.0233) (0.0235) Mostly positive (and) -0.0573* (0.0324) Many news about migration -0.0777** (0.0383) Few news about migration -0.0070 (0.0493) Among published news articles Many news about migration -0.0607** (0.0246) Few news about migration -0.0048 (0.0198) Other controls Yes Yes Yes FMP FEs Yes Yes Yes Year-month FEs Yes Yes Yes Country of destination FEs Yes Yes Yes Country of origin FEs Yes Yes Yes Number of Observations 1,568 1,568 1,568 Notes: OLS regression results. Standardized OLS estimated coefficients are reported. Standardization of coefficients is sd(x) obtained using the formula sd(y) βx , where βx is the point estimate associated to control variable x, while sd(x) and sd(y ) indicate the standard deviation or respectively control variable x and dependent variable y . Standard errors clustered at the FMP level in brackets. The dependent variable Share of migrants in the FMP who are long-stayers is the share of migrants in the FMP i in month m in year t who report being in that location for more than 6 months. The variable Tone of the news at destination (T one N ews Destinationsi,m−1,t ) is the average tone of the migration-related news articles published in the top destination countries for migrants in FMP i in month m − 1 in year t.“Other controls” include: Intensity of night lights within 5 km ; Number of conflicts within 5 km ; FMP with migrants planning to remain in Libya (1 = Yes) ; FMP with migrants planning to return home (1 = Yes). All variables are defined in Tables A1 and A2 and additional details are provided in Section 4. *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. 28 Table 6: Heterogeneity: Economic conditions in destination countries Share of migrants in FMP who are long-stayers (1) (2) (3) Tone of the news at destination -0.0300* (0.0185) Tone of the news at destination (during months when) Employment rate at destination -0.0179 High (0.0176) -0.0292* Low (0.0175) GDP per capita at destination -0.0146 High (0.0170) -0.0268* Low (0.0164) Employment rate in destination countries -0.0310 -0.0296 -0.0282 (0.0663) (0.0667) (0.0668) GDP per-capita in destination countries -0.2813*** -0.2823*** -0.2825*** (0.0981) (0.0976) (0.0979) Other controls Yes Yes Yes FMP FEs Yes Yes Yes Year-month FEs Yes Yes Yes Country of destination FEs Yes Yes Yes Country of origin FEs Yes Yes Yes Number of Observations 1,568 1,568 1,568 Notes: OLS regression results. Standardized OLS estimated coefficients are reported. Standardization of coef- ficients is obtained using the formula sd (x) β , where βx is the point estimate associated to control variable x, sd(y ) x while sd(x) and sd(y ) indicate the standard deviation or respectively control variable x and dependent variable y . Standard errors are clustered at the FMP level in brackets. The dependent variable Share of migrants in the FMP who are long-stayers is the share of migrants in the FMP i in month m in year t who report being in that location for more than 6 months. The variable Tone of the news at destination (T one N ews Destinationsi,m−1,t ) is the average tone of migration-related news articles published in the top destination countries for migrants in FMP i in month m − 1 in year t.“Other controls” include: Intensity of night lights within 5 km ; Number of conflicts within 5 km ; FMP with migrants planning to remain in Libya (1 = Yes) ; FMP with migrants planning to return home (1 = Yes). All other variables are defined in Tables A1 and A2 and additional details are provided in Section 4. *, **, *** indicate statistical significance at the 10%, 5%, and 1%. 29 Table 7: Tone of the news and changes in preferred destination countries Share of FMPs for which country d is a preferred destination (1) (2) (3) Tone of the news at destination d higher than in other countries 0.7368** 0.7531** 0.7668** (0.3524) (0.3322) (0.3447) Economic conditions in country i No No Yes Country of destination fixed effects Yes Yes Yes Year-month fixed effects Yes Yes Yes Estimator OLS SLM SLM ρ [p.value] -0.0373*** [0.0000] -0.2649*** [0.0079] Sample FMPs with constant population Number of Observations 863 863 846 Notes: Results from OLS in columns 1, results from Spatial Lag Model (SLM) in columns 2 to 3. The dependent variable Tone of the news at destination d higher than in other countries is a dummy variable that takes value one if the average tone of the news in d at month m in year t is higher than the median value of the average tone of the news published in all countries located in the same continent of d during the same period and zero otherwise. The variable Tone of the news at destination d higher than in other countries is constructed as follows. For country i in month m, we compute the distribution of the tone of the news at m in all planned destinations (including i ) belonging to the continent where i is situated. We then create a dummy variable which takes value one if the tone of the news of country i is higher than the median value of the distribution, and zero otherwise. In columns 2-3, spatial autocorrelation is modelled using matrix W , where the generic ij t h is a dummy variable which takes value one if countries i and j are located in the same continent, and zero otherwise. The sample includes only FMPs with constant population. The sample is constructed as follows. For each FMP, we only select observations taken in two subsequent months, if there were no migrants arriving at or departing from the FMPs from month m − 1 to month m. The parameter ρ is the degree of spatial autocorrelation between outcomes of countries within the same continent. Economic conditions in country i include Employment Rate and GDP per capita are monthly and computed as described in the Note to Table 6. All other variables are defined in Tables A1 and A2. *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. 30 Figure Figure 1: Location of the IOM Flow Monitoring Points (FMPs) in Libya (2017-2020) Notes: On the left, the physical map of Libya (Source: Worldometer: https://www.worldometers.info/maps/ libya-map/). On the right, the location of the IOM FMPs in Libya (Source: IOM Data). 31 A Appendix: Additional tables Table A1: Data Descriptives Variable Definition Mean St. Dev. Dependent Variables Share of migrants who stay in FMP... more than 6 months Share of migrants who have been recorded to be located for more than 6 months in FMP i during month m. 57.10 29.45 less than two weeks Share of migrants who have been recorded to be located for less than 2 weeks in FMP i during month m. 14.62 10.63 less than 3 months Share of migrants who have been recorded to be located for less than 3 months in FMP i during month m. 13.93 11.86 between 3 and 6 months Share of migrants who have been recorded to be located between 3 and 6 months in FMP i during month m. 18.45 12.42 Share of FMPs for which country d is a Share of FMPs indicating country d as a preferred destination at month m in year t. Countries considered: 2.03 7.76 preferred destination all those nominated at least once among the set of preferred destinations by a FMP. FMPs considered: for each FMP, we only select observations taken in two subsequent months, if there were no migrants arriving at or departing from the FMPs from month m − 1 to month m (i.e., FMPs with constant population ). Migrant-related news variables Tone of the news at destination Average tone of migration-related news published in the destination countries for migrants in FMP i in month -2.39 0.53 m-1. The tone of a news article is obtained from the difference between the positive and the negative score assigned to that article. A higher value indicates a more positive tone. - Discrimination Average tone of migration-related news discussing discrimination events published in the destination countries -0.32 0.03 for migrants in FMP i in month m-1. The tone of a news article is obtained from the difference between the positive and the negative score assigned to that article. - Sea incidents Average tone of migration-related news discussing sea incidents published in the destination countries for -0.01 0.01 migrants in FMP i in month m-1. The tone of a news article is obtained from the difference between the positive and the negative score assigned to that article. - Economic conditions Average tone of migration-related news discussing economics facts published in the destination countries for -0.01 0.01 migrants in FMP i in month m-1. The tone of a news article is obtained from the difference between the positive and the negative score assigned to that article. - Government announcements Average tone of migration-related news referring to government authorities published in the destination coun- -0.02 0.02 tries for migrants in FMP i in month m-1. The tone of a news article is obtained from the difference between the positive and the negative score assigned to that article. - Economic conditions Average tone of migration-related news not referring to any of the previous topics (Discrimination, Sea inci- -0.07 0.05 dents, Economic conditions, Crime, and Government announcements) published in the destination countries for migrants in FMP i in month m-1. The tone of a news article is obtained from the difference between the positive and the negative score assigned to that article. Prevailing news sentiment is... Mostly negative Dummy variable. It takes value one if the positive score of the news in the top destination countries for 0.23 0.42 migrants in FMP is below its annual median, while the negative score is above its annual median. It takes zero otherwise. Same number of negative and positive news Dummy variable. It takes value one if both the positive and the negative scores of the news in the top 0.55 0.50 destination countries for migrants in FMP are above (or below) their annual median. It takes zero otherwise. Mostly positive Dummy variable. It takes value one if the positive score of the news in the top destination countries for 0.22 0.41 migrants in FMP is above its annual median, while the negative score is below its annual median. It takes zero otherwise. Among published news articles... Many news about migration Dummy variable. It takes value one if the monthly number of news in the top destination countries for migrants 0.32 0.47 in FMP is above its annual median. It takes zero otherwise. Few news about migration Dummy variable. It takes value one if the monthly number of news in the top destination countries for migrants 0.67 0.47 in FMP is below its annual median. It takes zero otherwise. Tone of the news at destination when... High Intensity Patrolling Dummy variable. It takes value one if the number of arrivals from Libya to Italy in a given month is below 0.47 0.50 the median value observed from 2018 to 2020 (1150 individuals). It takes zero otherwise. Low Intensity Patrolling Dummy variable. It takes value one if the number of arrivals from Libya to Italy in a given month is above 0.54 0.50 the median value observed from 2018 to 2020 (1150 individuals). It takes zero otherwise. Tone of the news at destination d higher Dummy variable. It takes value one if the average tone of the news in d at month m in year t is higher than 0.50 0.50 than in other countries the median value of the average tone of the news published in all countries located in the same continent of d during the same period. It takes zero otherwise. Note: only countries registered at least once among the preferred destinations indicated by FMPs are considered. Number of Observations 1,568 32 Table A2: Data Descriptives Variable Definition Mean St. Dev. Economic variables Employment rate in destination countries It is calculated in three steps. First, we retrieve the annual employment rate of preferred destination 11.94 4.31 country j indicated in FMP i during month m in year y. Second, we impute the monthly value of the x −xy employment rate of j at month m with the formula xy ∗ (1 + (#m − 1) ∗ y+1 11 ), where xy and xy+1 indicate the value of Employment Rate (annual) in the year when the observation was recorded and the following one, and #m indicates the number of the month in which the observation was recorded. We repeat this operation for all preferred destination countries indicated in FMP i, during month m in year y. Third, we compute the average value of the metrics obtained in step 2, so to obtain the average monthly employment rate of preferred destination countries FMP i during month m in year y. Data Source: International Labour Organization (ILO). Employment rate at destination High For each FMP i, we take the distribution of the variable Employment Rate (monthly) in year y , and register 0.45 0.50 ¯i,y . Then, we create a dummy variable which takes value one if xi,my > x its median value, call it x ¯i,y , where xi,my is the value the variable Employment Rate (monthly) associated to i at month m in year y , and zero otherwise. Employment rate at destination Low For each FMP i, we take the distribution of the variable Employment Rate (monthly) in year y , and register 0.55 0.50 ¯i,y . Then, we create a dummy variable which takes value one if xi,my <= x its median value, call it x ¯i,y , where xi,my is the value the variable Employment Rate (monthly) associated to i at month m in year y , and zero otherwise. GDP per capita in destination countries It is calculated in three steps. First, we retrieve the annual GDP per capita of preferred destination 18934.49 10418.57 country j indicated in FMP i during month m in year y. Second, we impute the monthly value of the x −xy GDP per capita of j at month m with the formula xy ∗ (1 + (#m − 1) ∗ y+1 11 ), where xy and xy+1 indicate the value of GDP per capita (annual) in the year when the observation was recorded and the following one, and #m indicates the number of the month in which the observation was recorded. We repeat this operation for all preferred destination countries indicated in FMP i, during month m in year y. Third, we compute the average value of the metrics obtained in step 2, so to obtain the average monthly GDP per capita of preferred destination countries FMP i during month m in year y. Data Source: The World Bank. GDP per capita at destination High For each FMP i, we take the distribution of the variable GDP per capita (monthly) in year y , and register 0.46 0.50 ¯i,y . Then, we create a dummy variable which takes value one if xi,my > x its median value, call it x ¯i,y , where xi,my is the value the variable GDP per capita (monthly) associated to i at month m in year y , and zero otherwise. GDP per capita at destination Low For each FMP i, we take the distribution of the variable GDP per capita (monthly) in year y , and register 0.54 0.50 ¯i,y . Then, we create a dummy variable which takes value one if xi,my <= x its median value, call it x ¯i,y , where xi,my is the value the variable GDP per capita (monthly) associated to i at month m in year y , and zero otherwise. Control variables Intensity of night lights within 5 km Intensity of night lights in the 5 km radius around the location of FMP i at month m-1. 12.75 15.99 Number of conflicts within 5 km Number of conflicts in the 5 km radius around the location of FMP i at month m-1. 2.46 8.94 Some migrants in FMP plan to... Remain in Libya (1 = Yes) Dummy variable. It takes value one if one of the top destination countries for migrants in FMP i in month 0.45 0.50 m-1 is Libya, and zero otherwise. Return home (1 = Yes) Dummy variable. It takes value one if at least one country is both a top planned destination and a top 0.37 0.48 origin country for migrants in FMP i in month m-1, and zero otherwise. FMPs where night lights have... Low Intensity Dummy variable. In each year, we calculate the distribution of the intensity of night lights in the 5 km 0.33 0.47 radius around FMPs. The variable assigns value one to FMPs falling in the lowest 33rd percentile of the distribution. It takes zero otherwise. Middle Intensity Dummy variable. In each year, we calculate the distribution of the intensity of night lights in the 5 km 0.33 0.47 radius around FMPs. The variable assigns value one to FMPs falling between the 34th and the 66th percentile of the distribution. It takes zero otherwise. High Intensity Dummy variable. In each year, we calculate the distribution of the intensity of night lights in the 5 km 0.33 0.47 radius around FMPs. The variable assigns value one to FMPs falling in the highest 33rd percentile of the distribution. It takes zero otherwise. FMPs with network coverage... Not present Dummy variable. It takes value one if the FMP is not within a 5 km radius from a cell tower, and zero 0.92 0.26 otherwise. Present Dummy variable. It takes value one if the FMP is within a 5 km radius from a cell tower, and zero 0.08 0.26 otherwise. High-speed connection Dummy variable. It takes value one if the FMP is within a 5 km radius from a 4G cell tower, and zero 0.41 0.49 otherwise. Low-speed connection Dummy variable which takes value one if the FMP is within 5 km radius from a 3G or 2G cell tower, and 0.59 0.49 zero otherwise. FMPs located in... West Libya Dummy variable. It takes value one if the FMP is located in the area of the country west to the Libyan 0.44 0.49 centroid. It takes zero otherwise. East Libya Dummy variable. It takes value one if the FMP is located in the area of the country east to the Libyan 0.56 0.49 centroid. It takes zero otherwise. FMPs with a major group of migrants from... West Africa Dummy variable. It takes value one if one of the major group of migrants located in the FMP is from a 0.92 0.27 West African country. It takes zero otherwise. East Africa Dummy variable. It takes value one if one of the major group of migrants located in the FMP is from a 0.04 0.19 East African country. It takes zero otherwise. Number of Observations 1,568 33 Table A3: News sentiment and migrants’ movement: Other groups of migrants Share of migrants who stay in FMP Less than 2 weeks Less than 3 months Between 3 and 6 months Tone of the news at destination 0.0099 0.0262 0.0312 (0.0119) (0.0216) (0.0237) Other controls Yes Yes Yes FMP FEs Yes Yes Yes Year-month FEs Yes Yes Yes Country of destination FEs Yes Yes Yes Num. Obs. 1,568 1,568 1,568 Notes: OLS regression results. Standard errors clustered at the FMP level in brackets. Variables are defined in Tables A1 and A2. Other Controls includes thew same controls as in Table A11 column 3, namely: Intensity of night lights within 5 km ; Number of conflicts within 5 km ; FMP with migrants planning to remain in Libya (1 = Yes) ; FMP with migrants planning to return home (1 = Yes).. *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. Table A4: News in Europe and preference for returning home Returnee countries among preferred destinations between m and m+1 Change (1 = Yes) Difference Tone of the news at destination 0.0641 0.2285 (0.0639) (0.1485) Country of destination fixed effects Yes Yes Year-month fixed effects Yes Yes Number of Observations 1,165 1,165 Notes: OLS regressions results. Standard errors clustered at the FMP level in brackets. In column 1, the dependent variable is a dummy which takes value one if there is a difference in the number of returnee countries indicated among the preferred destinations in FMP i between month m and m − 1, and zero otherwise. In column 2, the dependent variable is equal to the difference in the number of returnee countries indicated among the preferred destinations in FMP i between month m and m − 1.The variable Tone of the news at destination (T one N ews Destinationsi,m−1,t ) is the average tone of migration-related news articles published in the top destination countries for migrants in FMP i in month m − 1 in year t. *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. 34 B Appendix: Additional figures Figure A1: Migration routes to Europe Source: World Bank (2018). 35 Figure A2: Network of internal movements of migrants in Libya in 2017 Note: Each node represents a Libyan province. Two provinces (nodes) are connected if a movement of migrants was registered between the provinces in the considered year. The size of flows is proportional to the average number of migrants registered in the province of origin of the flow in the considered year. Elaboration of the authors from DTM data. Source: Di Maio et al. (2023). 36 Figure A3: Share of migrants located at Libyan FMPs by length of stay % Less than 2 weeks % Less than 3 months 100% 75% 50% Percentage of migrants in FMPs 25% 0% % Less than 6 months % More than 6 months 100% 75% 50% 25% 0% 2017 2018 2019 2020 2017 2018 2019 2020 Note: The figure shows the boxplot distribution of the share of migrants located at Libyan FMPs by length of stay during our period of analysis. Figure A4: Spatial and time placebo tests 200 200 150 150 100 100 50 50 0 −0 −0 47 0. 0. 0 00 05 .0 .0 69 15 3 31 7 −0.0432 −0.0317 −0.0045 0.0268 (a) Effect of the news in the top destination countries (b) Effect of the news in the FMP’s top for other FMPs destination countries but in another time Notes: The dotted line indicates the estimated coefficient for the explanatory variable in the main regression specification (see Table 1 column 2). Placebo results obtained with 10,000 replications. 37 Figure A5: Tone of the news in major destination countries vs number of migrants in Libyan FMPs −2.0 Regress. Coefficient: 0.00002 (0.0001). P.value:0.8200 Avg. tone of the news in France, Germany, −2.5 and Italy in month t+1 −3.0 −3.5 0 500 1000 1500 Total number of migrants in Libyan FMPs at month m Note: For each month, we indicate the average tone of the news in France, Germany, and Italy in a given month, recorded on the y-axis, associated to the total number of migrants in Libyan FMPs during the same month, registered in the x-axis. The black line describes a regression line obtained with model ymy = α + βxmy + γy + my , where α is the intercept, ymy indicates the value recorded on the y-axis in a given month, xmy indicates the value recorded on the x-axis in the same month, γy indicates year fixed effects, and my is the error term. Figure A6: Tone of the news in major destination countries vs number of migrants in Libyan FMPs located in the North West −2.0 Regress. Coefficient: 0.000008 (0.00006). P.value:0.9900 Avg. tone of the news in France, Germany, −2.5 and Italy in month t+1 −3.0 −3.5 0 100 200 300 Total number of migrants in Libyan FMPs at month m Note: For each month, we indicate the average tone of the news in France, Germany, and Italy in a given month, recorded on the y-axis, associated to the total number of migrants found in FMPs located in the Nort-West area of Libya during the same month, registered in the x-axis. The black line describes a regression line obtained with model ymy = α + βxmy + γy + my , where α is the intercept, ymy indicates the value recorded on the y-axis in a given month, xmy indicates the value recorded on the x-axis in the same month, γy indicates year fixed effects, and my is the error term. 38 Figure A7: News in Europe and preference for returning home 200 Variation in the number of retunee countries among preferred destinations 100 between m−1 and m 0 −100 −3.2 −3.0 −2.8 −2.6 Avg. Tone of European News at m−1 (lower is worse) Note: The horizontal axis shows the average Tone of the news across all European countries that are a preferred destination for any FMP at month m. The vertical axis reports the variation between month m and m + 1 in the total number (i.e., summed over all FMPs) of preferred destinations that are the country of origin of migrants. In other words, we are plotting the news sentiment in European countries that are preferred destinations for the migrants at a given moment on their preference for returning to the origin country in the following period. The black line describes a regression line obtained with model ymy = α + βxmy + γy + my , where α is the intercept, ymy indicates the value recorded on the y-axis in a given month, xmy indicates the value recorded on the x-axis in the same month, γy indicates year fixed effects, and my is the error term. Returnee countries considered: Algeria, Bangladesh, Benin, Burkina Faso, Cameroon, Chad, Democratic Republic of Congo, Egypt, Eritrea, Ethiopia, Gambia, Ghana, Guinea, Guinea Bissau, Ivory Coast, Jordan, Liberia, Malaysia, Mali, Mauritania, Morocco, Niger, Nigeria, Pakistan, Palestine, Senegal, Sierra Leone, Somalia, Sudan, Syria, Togo, Tunisia. 39 C Appendix: Measuring the tone of migration-related news ar- ticles by specific topics In the following, we describe how we categorize migration-related news articles by topic. First, we identify the top 150 GDELT keywords/topics in migration-related news during our period of analysis and for our sample of countries. Second, among these, we select the keywords that are related to the following topics: 1) discrimination (at destination); 2) sea incidents; 3) economic conditions (at destination); 4) crime (at destination); 5) governments announcements. We define a migration-related news article to be about Discrimination if at least one of its top 3 GDELT keywords is among the following: discrimination, migration f ear, xenophobia, hate, racial, racism, racist, stereotype. On average, 7.95% of all migration-related news articles include at least one of these keywords (min. 7.12, max 35.00). We define a migration article news to about Sea incidents if at least one of its top 3 topics is related to one of the following GDELT keywords: maritime incident, sea, oceans. On average, 3.79% of all migration-related news articles include these keywords (min. 3.38, max 7.29). We define a migration article news to be about Economic conditions if at least one of its top 3 GDELT keywords is among the following: economy , macroeconomic vulnerability and debt, growth, jobs, labor, labor markets, employment, unemployment, jobs strategies, job quality , tax workers. On average, 6.78% of all migration-related news articles include at least one of these keywords (min. 5.91, max 18.12). We define a migration article news to be about Crime if at least one of its top 3 GDELT keywords is among the following: crime, common robbery , organized crime, general crime, crime violence, kill. On average, 8.25% of all migration-related news articles include at least one of these keywords (min. 1.11, max 35.00). We define a migration article news to be about Government announcements if at least one of its top 3 topics is related to one of the following GDELT keywords: gov repatriation, policy ref orm, gov ref orm, general government, leader, tax candidate, tax chancellor, tax judge, tax king , tax leader, tax minister, tax representative, public sector management. On average, 8.32% of all migration-related news articles include at least one of hese keywords (min. 6.55, max 51.61). 40 D Appendix: Robustness - Spatial Analysis The presence of spatial autocorrelation in the migrants’ movement choices across FMPs is a possible threat to our model identification. First, spatial autocorrelation may be present when migrants living in FMPs geographically located in the same area are exposed to common shocks (e.g., an increase in conflict intensity, a drop in economic activity, etc.), and thus they make similar movement choices. In the case of Libya, defining what is to be considered “the same area” is not trivial: for instance, the common choice of using the boundaries of the administrative areas is unlikely to be informative in a context of a conflict-affected country that lacks a ruling government by more than ten years. To circumvent this problem, we assume that exposure to common shocks is a function of the distance of the FMPs from one another. Thus we re-estimate equation 1 using a Spatial Error Model (SEM), with which we test two different hypotheses on the role that common exposure to shocks may play: i.e., 1) all FMPs are exposed to the same shock but its impact depends on FMPs’ distance from it; 2) only FMPs close to one another are exposed to common shocks). To test the first hypothesis, we modify equation 1 so that the error term i,m,t is now equal to λW ui,m,t , where W is an adjacency matrix with the generic ij th entry measuring the distance between FMPs i and j , and λ is an estimated parameter used to endogenously assess the degree of spatial autocorrelation between observations. In this way, SEM allows us to model unobserved common exposure to shocks among FMPs as a function of the distance between them, without making any other assumption on the form of spatial autocorrelation potentially at work. Table A5 column 1 reports the estimates obtained with this model. Results indicate the presence of some form of spatial autocorrelation in the decisions of relocation of migrants residing in different areas. This is indicated by the parameter λ having a negative and statistically significant impact on the presence of migrants in the FMPs. At the same time, our main result is qualitatively unchanged: i.e., the tone of the news affects migrants’ movement decisions. Notably, the statistical significance of the effect is now even stronger than in our baseline model 1. Next, we test the second hypothesis, i.e., only FMPs close to one another are exposed to common shocks. This is done by changing matrix W in the SEM. This time, the generic ij th cell takes value one if FMPs i and j are located close to one another, and zero otherwise. Closeness between two FMPs is defined as a distance of 230 km, that is the minimum distance allowing each FMP to be exposed to a common shock at least with another FMP.35 The estimation of the SEM using the new version of W is reported in Table A5 column 2. Reassuringly, all our results are confirmed. Another situation in which the outcome of one FMP is spatially autocorrelated with the outcome of a different FMP is when the migrants’ movement choices affect the composition of both their FMP and that of other FMPs in the same way.36 We model this form of spatial 35 Choosing a smaller distance to define closeness is not possible, because SEM can only be estimated when W is invertible. A necessary condition for this to hold is that the row sums of W must be higher than zero: i.e., all FMPs must be correlated at least with another FMP. 36 For instance, this is the case when migrants who have been staying for a long time in one FMP decide to leave it (reducing the share of migrants who stay there for more than six months) and move to another FMP (also reducing the share of migrants who stay there for more than six months). Another case is the one in which migrants in different FMPs are in contact with one another, and the choice to postpone the journey to Europe by those in one FMP (thus increasing the share of migrants who stay there for more than six months) influence 41 autocorrelation using the Spatial Lag Model, again testing two different hypotheses: i.e., 1) each FMP influences all other FMPs, but its impact depends on its distance from the other FMPs’; 2) each FMP affects only close-by FMPs). In practice, with this model we augment equation 1 with the term ρW y , where y is our dependent variable (i.e., the share of migrants who stay in one FMP for more than six months), W is the same adjacency matrix used when estimating the SEM, and ρ is the estimated impact of the relocation choices of migrants in one FMP on the relocation choices of migrants in other FMPs. We test this model adopting both versions of W employed in the estimation of the SEM: i.e., that considering the potential influence of one FMP over all others, and that constraining the influence of one FMP only to closer FMPs. The results are reported in Table A5, respectively in columns 3, and 4. Also in this case, we find no evidence that spatial autocorrelation do affect our main finding: also with this model we find that a more negative tone of the news in destination countries makes migrants stay longer in their current location - postponing their next step in their migration journey - and the statistical significance of the effect is now even stronger than in our baseline model. 37 Table A5: News sentiment and migrants’ movement Robustness: Spatial Analysis Share of migrants in FMP who are long-stayers (1) (2) (3) (4) Tone of the news at destination -0.0355*** -0.0293*** -0.0314*** -0.0312*** (0.0127) (0.0127) (0.0125) (0.0125) Estimator SEM SEM SLM SLM Spatial autocorrelation exists between All FMPs Closest FMPs All FMPs Closest FMPs λ[p.value] -0.9425[0.0000] 0.1364[0.0065] - - ρ[p.value] - - 0.3571[0.0004] -0.0012[0.9655] Other controls Yes Yes Yes Yes FMP FEs Yes Yes Yes Yes Year-month FEs Yes Yes Yes Yes Country of destination FEs Yes Yes Yes Yes Country of origin FEs Yes Yes Yes Yes Number of Observations. 1,568 1,568 1,568 1,568 Notes: Results from Spatial Error Model (SEM) in columns 1 and 2, Results from Spatial Lag Model (SLM) in columns 3 and 4. In columns 1 and 3, spatial autocorrelation is modelled as a function of FMPs’ geographic distance between each other. In columns 2 and 4, spatial autocorrelation is assumed to exist only between FMPs closer than 230km: i.e., the minimum distance allowing each FMP to be influenced at least by another FMP. The parameter λ is the degree of spatial autocorrelation between errors estimated by the SEM model. The parameter ρ is the degree of spatial autocorrelation between outcomes estimated by the SLM model. All other variables are defined in Tables A1 and A2 and additional details are provided in Section 4. “Other controls” include: Intensity of night lights within 5 km ; Number of conflicts within 5 km ; Some migrants in FMP plan to remain in Libya (1 = Yes) ; Some migrants in FMP plan to return home (1 = Yes). *, **, *** indicate statistical significance at the 10%, 5%, and 1%, respectively. migrants in other FMPs to do the same (increasing as well the share of migrants who stay there for more than six months). 37 The magnitude of the coefficients obtained using the SLM and the OLS cannot be compared because - as in the case of SEM - the quantification of the coefficients is different with the two methods. 42