Networks and migrants’ intended destination∗

Simone Bertolia and Ilse Ruyssenb aUniversity of Auvergne† and CNRS bSHERPPA, Ghent University and IRES, Universit´eCatholique de Louvain‡

March 7, 2016

Abstract

Social networks are known to influence migration decisions, but it is hard to observe connections across individuals in the data. We rely on individual-level surveys con- ducted by Gallup in 148 countries that provide information on migration aspirations and on the existence of distance-one connections for all respondents in each of the poten- tial countries of (intended) destination. The origin-specific distribution of distance-one connections across destinations from Gallup closely mirrors the actual distribution of migrant stocks. This unique data source allows estimating origin-specific RUM-based location decision models that shed light on the value of having a friend in a given coun- try on the attractiveness of (intended) migration to that destination. We also test the validity of the distributional assumptions that underpin the estimation. The results of the econometric analysis also provide suggestive evidence that a direct connection to social networks at destination is less valuable for origin countries with a diaspora that is concentrated in few destinations.

∗The authors are grateful to Fr´ed´eric Docquier, Jes´us Fern´andez-Huertas Moraga and Jo¨el Machado and to participants at seminar presentations at CERDI and at the University of Ghent for their comments, to Robert Manchin and the Gallup Institute for Advanced Behavioural Studies for providing access to the data for the purpose of this project, to Olivier Santoni for providing excellent research assistance. Simone Bertoli acknowledges the support received from the Agence Nationale de la Recherche of the French government through the program “Investissements d’avenir” (ANR-10-LABX-14-01) the usual disclaimers apply. †CERDI, Bd. F. Mitterrand, 65, F-63000, Clermont-Ferrand - ; email: [email protected]. ‡Department of Economics, Tweekerkenstraat 2, 9030 Ghent - ; email: [email protected].

1 1 Introduction

Social networks are expected to exert an important influence on migration decisions, as connections with individuals that have already moved contribute to improve job prospects at destination of a potential migrant (Munshi, 2003), and they can reduce the multifaceted costs of crossing a border. The existing empirical evidence is based on rather coarse measures of networks, such as the share of households with a migrant at the village (McKenzie and Rapoport, 2010) or at the county level (Bertoli, 2010), or the size of the diaspora in each des- tination country (see, for instance, Pedersen, Pytlikova, and Smith, 2008; Beine, Docquier, and Ozden,¨ 2011, 2015; Beine and Salomone, 2013; Bertoli and Fern´andez-Huertas Moraga, 2015). The implicit assumption behind this approach, which reflects binding data con- straints, is that all potential migrants have an equal access to the networks at destination.1 This assumption is at odds with theoretical representations of social networks (see Jackson, 2010) and with how members of a migrant network interact with each other (Comola and Mendola, 2015). Our objective is to contribute to gaining a deeper understanding of how social networks influence international migration by using a dataset that provides unique information on the individual-level connections to networks in each potential destination. Specifically, we draw on the data from the Gallup World Polls (see Gallup, 2013) conducted in 148 countries of the world between 2007 and 2013. For each respondent, we have information on whether she has household members, relatives or friends who reside abroad, with information on the (up to four) countries in which they reside.2 Reassuringly, the geographical distribution of distance-one connections for each country matches closely the actual bilateral distribution of migrant across destinations for 2010 from the World Bank (2015a). We combine the information on the countries in which a respondent has a distance-one

1The estimation of gravity equations derived from underlying random utility maximization models on aggregate data has to rest on this assumption, as the equivalence of the estimates obtained on aggregate or on individual-level data depends on the absence of individual-specific regressors (Guimaraes, Figueiredo, and Woodward, 2003). Munshi (2015) reviews additional concerns related to the identification of network effects from gravity equation on aggregate data on bilateral migration flows. 2This destination-specific dimension of the information is what distinguishes the data that we use from the dataset on internal Chinese migration used by Giulietti, Wahba, and Zenou (2014), who have information about whether each individual has a friend residing in an (unspecified) Chinese urban area.

2 connection with information on whether she aspires to emigrate to another country,3 and if so to which destination. The Gallup World Polls do not provide information about actual moves, but we provide econometric evidence that the bilateral number of aspiring migrants by year is significantly associated with the yearly scale of actual bilateral migration flows to OECD destinations.4 There is a large literature in sociology and demography focusing on the determinants of intentions to emigrate (as in Becerra, 2012; Carling, 2002; Creighton, 2013; Van Dalen, Groenewold, and Fokkema, 2005; van Dalen, Groenewold, and Schoorl, 2005; Drinkwater and Ingram, 2009; J´onsson, 2008; Papapanagos and Sanfey, 2001; Wood, C.L., Ribeiro, and Hamsho-Diaz, 2010; Yang, 2000). Most of these papers use country-specific surveys, making it difficult to dig out comparative conclusions. On the contrary, the Gallup database covers a large number of countries of all sizes, regions and income levels, which makes it exceptional. Only a few studies have relied on Gallup World Polls to investigate the patterns and de- terminants of migration aspirations, disregarding the bilateral dimension. Esipova, Ray, and Pugliese (2011) present detailed descriptive statistics on the willingness to migrate across countries; Manchin, Manchin, and Orazbayev (2014) analyzes the effect of individual sat- isfaction on the desire to emigrate; Dustmann and Okatenko (2014) identify the effects of wealth constraints and local amenities. Other studies explore the bilateral dimension of the database after aggregating individual data, i.e. after computing the aggregate proportion of individuals who express a desire to leave their country and their bilateral shares. Docquier, Machado, and Sekkat (2015) and Delogu, Docquier, and Machado (2015) use the Gallup World Polls to quantify the short-run and long-run effects of removing migration barriers internationally, assuming that all aspiring migrants would be able to move after a complete liberalization. More related to our work, Docquier, Peri, and Ruyssen (2014) empirically analyze the country-specific and bilateral factors governing the size and skill composition of the pool of aspiring migrants, as well as the probability to realize aspirations. They find that the average income at destination and the presence of networks of previous migrants are ro-

3The Gallup World Polls include the following question: “Ideally, if you had the opportunity, would you like to move to another country, or would you prefer to continue living in this country?”; the answers to this question have been interpreted by Docquier, Machado, and Sekkat (2015) as reflecting whether an individual would move in a world with no policy-induced restrictions on human mobility. 4See also Chort (2014) for empirical evidence on the relationship between stated intentions to move and actual migration.

3 bust and quantitatively significant determinants of potential migration rates. However, they rely on macrodata to construct their network variable and ignore individual heterogeneity in network connections. To the best of our knowledge, our paper is the first to examine the effect of individual network on intended destination choices. We estimate - separately for each country - a conditional logit model that describes the choice of aspiring migrants among the alternative (intended) destinations.5 The estimation reveals that having a distance-one connection in a country is associated with a substantially higher conditional probability of choosing that country. Specifically, having a friend or a relative in a potential destination increases by a factor of eight the log odds of opting for that destination. The estimated coefficients of the variable of interest are compared with those that we obtain when re-estimating the model over restricted choice sets, as the distributional assumptions that underpin the conditional logit model imply that the coefficients should remain stable when the choice set is modified. The estimated coefficients for distance-one connections, which prove to be stable to vari- ations in the size of the choice set, are significantly negatively correlated with measures of concentration of the diaspora of a country. Specifically, the estimated coefficient is substan- tially lower for countries such as , whose diasporas are highly concentrated in one or in a few destinations compared to countries such as France, that have a more even distribution of their migrants across destinations. This indirectly suggests that the influence exerted by a distance-one connection on the attractiveness of a potential destination is a hump-shaped function of the size of the diaspora in that destination.

The remainder of the paper is structured as follows. Section 2 introduces the data from the Gallup World Polls. Section 3 briefly describes the random utility model that describes the (potential) location-decision problem that individuals face. Section 4 presents some basic descriptive statistics, and Section 5 contains the benchmark estimates, the tests that we conduct on them and our interpretation on what they actually suggest. Finally, Section 6 draws the main conclusions.

5The specification that we bring to the data is consistent with the idea that a larger diaspora can increase the attractiveness of a destination, as suggested by the literature (Pedersen, Pytlikova, and Smith, 2008; Beine, Docquier, and Ozden,¨ 2011; Beine and Salomone, 2013; Bertoli and Fern´andez-Huertas Moraga, 2015).

4 2 The Gallup World Polls

We have individual-level data for 148 countries where Gallup has conducted at least one Gallup World Poll between 2007 and 2013.6 The surveys conducted by Gallup typically have a sample of around 1,000 randomly selected respondents per country,7 and the data are collected either through face-to-face interviews or through phone calls in countries where at least 80 percent of the population has a telephone land-line.

2.1 Aspiring migrants

The Gallup World Polls include two related questions on intentions to migrate, asked in all countries between 2007 and 2013: (i) “Ideally, if you had the opportunity, would you like to move to another country, or would you prefer to continue living in this country?”, and (ii) “To which country would you like to move?” for the individuals who provide a positive answer to question (i). We refer to the individuals who express their intention to leave their country of residence as aspiring migrants (see also Van Dalen, Groenewold, and Fokkema, 2005; van Dalen, Groenewold, and Schoorl, 2005; J´onsson, 2008; Drinkwater and Ingram, 2009; Becerra, 2012; Creighton, 2013; Dustmann and Okatenko, 2014).

6For a description of the methodology and codebook, see Gallup (2013); further details on the data source can be found in Section 4.1 below. 7The number of respondents is actually larger in countries with a large population, such as China, India or Russia.

5 Figure 1: Share of aspiring migrants among natives (percent)

Notes: the figure reports the percentage of natives aged 15 to 49 aspiring to migrate in each country; data are pooled across different waves of the survey, and sampling weights are used. Source: Authors’ elaboration on Gallup World Polls.

Figure 1 reports the percentage of aspiring migrants among the native-born respondents aged between 15 and 49 years in each of the 148 countries covered by the Gallup World Polls.8 The (weighted) average of the share of aspiring migrants stands at 19.52 percent.9 The ten countries with the highest shares of intended migrants among natives are either Sub-Saharan African and Latin American and Caribbean countries, with the Dominican Republic (68.0 percent) having the largest share of aspiring migrants, followed by Liberia (66.4 percent), Sierra Leone (65.1 percent) and Guyana (61.7 percent). Five out of the ten countries with the lowest shares of intended migrants are Gulf countries, namely United Arab Emirates (3.5 percent), Bahrain (3.9 percent), Kuwait (4.5 percent), Saudi Arabia (5.9 percent) and Qatar (10.6 percent).10 The share of natives that intend to migrate declines with income per capita, as the bivariate correlation between the share of aspiring migrants and the logarithm

8Figure A.1 in the Appendix reports the corresponding Kernel density plot. 9The weights given by the native population in each country are computed based on World Bank (2015a,b). 10India (5.9 percent), Indonesia (7.7 percent), Thailand (8.7 percent), Rwanda (11.3 percent) and Malaysia (11.6 percent) are the other countries with the lowest shares of intended migrants.

6 of real GDP per capita in 2010 from World Bank (2015b) stands at -0.35.11

Figure 2: Distribution of intended migrants by destination (percent)

Notes: the figure reports the distribution across destinations of intended migrants aged 15 to 49; data are pooled across different countries and waves of the survey, and sampling weights are used to compute the distribution. Source: Authors’ elaboration on Gallup World Polls and World Bank (2015a,b).

The respondents in each of the 148 countries in our sample differ with respect to the number of countries they aspire to move to; on average, respondents in each country report 42.97 intended destinations, ranging from 14 for Trinidad and Tobago to 81 for Chad. Figure 2 reports the distribution of aspiring migrants across countries of (intended) destination.12 The natives aged 15 to 49 in our sample aspire to migrate towards 187 different countries in the world, with a (highly) uneven distribution of aspiring migrants across (intended) des- tinations. Specifically, 29.1 percent of the individuals in our sample aspire to migrate to

11Figure A.2 in the Appendix presents the scatter plot of the two variables. 12Country-specific figures are aggregated using weights given by the native population in each country in 2010 are computed from World Bank (2015a,b), i.e., the size of the resident population minus the total number of foreign-born residents. Ideally, we would have used figures for the population aged 15 to 49, but these are not available neither for the resident population nor for the immigrant stocks. World Bank (2015a) does not provide an estimate of the total foreign-born population in Taiwan and in the Occupied Palestinian Territories, which we thus set to zero.

7 the , followed by the (8.2 percent), (6.0 percent), France (5.4 percent) and Saudi Arabia (4.8 percent), with the first five (intended) destina- tions totaling 53.6 percent of the preferences of aspiring migrants. The first 25 (intended) destinations are chosen by 89.7 percent of all aspiring migrants, while the total share of the 92 countries at the bottom of the list stands at just 1.0 percent. The (pooled) distribution of aspiring migrants across countries is closely positively correlated with the distribution of actual migrant stocks, but it is more concentrated than the latter. The first five intended destinations, which account for 53.6 percent of all aspiring migrants, hosted 35.9 percent of the actual migrants from the origin countries in our sample according to World Bank (2015a). The data from the Gallup World Polls can be aggregated to obtain the number of natives of country j intending to move to country k in each year in which the survey is conducted,

which we denote as qjkt. The OECD International Migration Database provides us with

information about the size of the actual gross bilateral migration flow mjkt from j to k by year for 32 of the 187 destination countries mentioned as preferred destinations by the 13 respondents to the Gallup World Polls. We can thus test whether mjkt correlates with

the (lagged) value of qjkt, i.e., whether past number of bilateral aspiring migrants contain information about the current the size of actual gross migration flows. We thus rely on a Poisson Pseudo Maximum Likelihood estimation of the following regression:

′ mjkt = exp[α ln(qjkt−1)+ β xjk + djt + dkt + ǫjkt] (1)

where xjk is a vector of dyadic controls including, as in Beine, Docquier, and Ozden¨ (2011) or Bertoli and Fern´andez-Huertas Moraga (2015), the logarithm of distance, and dummies

for contiguity, common colonial history and a common language, and djt and dkt represent origin-year and destination-year dummies respectively. The estimation of (1) on a sample of 8,258 observations gives an estimate of α = 0.550, with an associated z-statistic of 27.28.14 b This strongly suggests that migration intentions are significantly associated with actual migration flows: an 10 percent increase in the number of individuals aspiring to move from j to k in year t − 1 is associated with a 5.5 percent increase in the gross bilateral migration flows from j to k in year t.15 13These 32 destination countries represent the preferred destination for 72.8 percent of the our sample of natives aged 15 to 49 who intend to migrate. 14The full set of estimation results is available from the Authors upon request. 15This result is robust to the inclusion of the logarithm of (one plus) the size of the bilateral migrant

8 2.2 Distance-one connections in the intended destinations

The questionnaire of the Gallup World Polls includes the following question: (iii) “Have any members of your household gone to live in a foreign country permanently or temporarily in the past five years?”, with (iv) information on the country of residence for those who provide an affirmative answer to question (iii). Furthermore, the questionnaire also asks: (v) “Do you have relatives or friends who are living in another country whom you can count on to help you when you need them, or not?”. For the individuals who answer affirmatively to this question, the data provide (vi) information on the (up to three) countries of residence of these relatives or friends. Thus, questions (iii) to (vi) gives us information about up to four countries in which each individual is connected to someone who could provide help to him or her. If we consider the individuals who provide an affirmative answer to question (v), 63 percent of them report a distance-one connection in just one country, and 22 percent of them in two countries. This implies that for 85 percent of the respondents the limit of three countries in question (vi) is certainly not binding, so that we observe in the data all the countries in which they have a distance-one connection with relatives or friends, while the limit might be binding for (a part of) the 15 percent the respondents that report three countries. Thus, the Gallup World Polls give us information about the foreign countries in which each individual has at least one distance-one connection. This variable is not defined for the country of birth itself, but it is reasonable to assume that all respondents have at least a distance-one connection in their country of birth (and residence).16 Notice that a respondent might have more than one distance-one connection in each of the countries that he or she reports, and that the distance-one connections might refer to individuals who are not born in the same country as the respondent.17 Keeping these two caveats in mind, it is interesting to compare for each country the distribution the distance- one connections from the Gallup World Polls with the actual distribution of its migrants across destinations, and we can rely on World Bank (2015a) for data on bilateral migrant stocks (defined on the basis of country of birth) for 2010.18

stock in 2000, as in Beine, Docquier, and Ozden¨ (2011), as an additional control, we obtain an estimate of α = 0.441, with an associated z-statistic of 22.45. b 16 This plausible assumption implies that we have no variability across individuals in this variable for the country of birth, which poses a problem in the estimation of the conditional logit (see below). 17For instance, a Belgian respondent might have a French and an Italian friend residing in France. 18Recall that the Gallup World Polls have been conducted around 2010.

9 Figure 3: Spearman’s rank correlation coefficient, distance-one connections and migrants

Notes: the figure reports country-specific Spearman’s rank correlation coefficient between the distribution of the bilateral migrant stocks in 2010 and the distribution of the distance-one connections from our individual- level data; sampling weights are used; the Occupied Palestinian Territories, Serbia and Taiwan are not included; the weighted average and standard deviations, with weights given by the native population in each country computed from World Bank (2015a,b), for Spearman’s rank correlation coefficients stand at 0.429 and 0.073 percent respectively. Source: Authors’ elaboration on Gallup World Polls and World Bank (2015a).

For each country j, we compute the Spearman’s rank correlation coefficient ρj between the distributions of distance-one connections and actual migrants, with ρj ∈ [−1, 1]. The

Spearman’s rank correlation coefficient ρj is always positive, and significantly so for 142 out of 145 countries,19 and the (weighted) average of ρ stands at 0.429.20 Figure 3 plots the Spearman’s rank correlation coefficient for the 145 countries in our sample.21 The high value of the Spearman’s rank correlation coefficient is reassuring with respect to the fact that the data coming out of the surveys from Gallup match well with the distribution of actual

19We do not have data on bilateral migrant stocks for the Occupied Palestinian Territories, Serbia and Taiwan from World Bank (2015a); the countries for which the Spearman’s rank correlation coefficient is not significantly different from zero are Bahrain (p-value 0.109), Kuwait (0.045) and Namibia (0.168). 20Similar evidence is obtained when relying on the Pearson’s correlation coefficient. 21Figure A.3 in the Appendix reports the corresponding Kernel density plot.

10 migrants across destinations.

3 RUM model and estimation

Consider an individual i residing in country j. The utility that this individual would obtain from locating in country k belonging to the choice set D is given by:

Uijk = Vijk + ǫijk, (2)

′ where Vijk = xijk βjk represents the deterministic component of utility, net of moving costs, and ǫijk is the stochastic component of utility. If ǫijk follows an independently and identically − distributed Extreme Value Type-1 distribution, with F (x)= e−e x , then the probability that country k represents the utility-maximizing alternative is given by (McFadden, 1974):

x ′ e ijk βjk pijk ≡ Prob(Uijk > Uijl , ∀l ∈ D/{k})= ′ (3) xijl βjl Pl∈D e The separate estimation of a conditional logit model for each origin j allows us to recover 22 the vectors of parameters βjk. The location-specific deterministic component of utility de-

pends on: (i) a dummy variable dijk that signals whether the individual i has a distance-one connection to destination k, i.e., she has a relative or friend living in k; the coefficient of this individual alternative-specific variable dijk varies across origins but not across destina- tions; (ii) a dyadic dummy djk that absorbs the effect of all time-invariant dyadic, origin or destination-specific variables that exert an identical influence on the attractiveness of

the (intended) destination k; (iii) a vector zij of individual characteristics (e.g. age, gender, ′ ′ level of education), whose coefficients vary across destinations. Letting xijk =(dijk,djk, zij ) ,

we denote by β1jk = β1j, ∀k ∈ D, the parameter of the individual and alternative-specific β network variable dijk. Standard errors for bjk can be obtained through bootstrapping. The choice set over which we estimated (3) does not include the origin j itself, because

the variable dijk is not defined when k = j, as discussed in Section 2.3 above, so that our estimation is restricted to the sub-sample of individuals that state the intention to migrate.

Notice that the estimation of (3) on the choice set Dj ≡ D/{j} entails that our estimation is actually consistent with the distributional assumptions introduced by Bertoli, Fern´andez- Huertas Moraga, and Ortega (2013) and Ortega and Peri (2013), who allow for a common

22 The vector of coefficients that refers to the origin country, i.e., βjj, is normalized to zero.

11 variance component across all countries but the origin, as this component does not influence the choice across destinations.23

3.1 The IIA property

The estimation of (3) rests on the independence of irrelevant alternatives property across destinations, which implies that the relative probability of choosing between two alternative options depends exclusively on the attractiveness of these two options, i.e. ln(pijk)−ln(pijh)=

Vijk − Vijh, and it is independent from the presence of other alternatives in the choice set 24 Dj. An implication of this property is that the estimated coefficients should be stable

when the choice set Dj is modified; we thus follow Head, Ries, and Swenson (1995) and Grogger and Hanson (2011) by re-estimating (3) on a series of restricted samples that are

obtained by progressively dropping larger sets of destinations from Dj, and comparing the Rj estimated coefficient βb1j obtained on the subsample Rj ⊂ Dj with the one obtained from the estimation on the entire choice set Dj, i.e., βb1j. More specifically, for each country j we compute the share of the estimations conducted on the various restricted samples Rj for Rj 25 which we do not reject the null hypothesis that βb1j = βb1j.

3.2 Interpretation of the estimated coefficients

The estimation on the reduced choice set Dj implies that our estimation allows us to under-

stand how each variable in the vector xijk influences the probability of opting for destination k conditionally upon expressing an intention to migrate, but not how the various elements x of ijk influence the probability of expressing such an intention, i.e., 1 − pijj = Pl∈Dj pijl, in β the first place. For a given estimate of bjk, the responsiveness of 1 − pijj with respect to a variation in any of the elements of xijk depends on the (unknown) extent to which individu-

23See Section 3.2 below for a discussion about the ensuing implications for the interpretation of the estimated coefficients. 24We should recall here that the independence of irrelevant alternatives is a property of the specification of the model that is estimated, rather than an inherent feature of the choice situation, and it depends on the extent to which observables allow capturing heterogeneity across individuals; Bertoli and Fern´andez- Huertas Moraga (2013, 2015) provide evidence that this property is violated in specifications estimated on aggregate data that assume that the deterministic component of utility is not individual-specific, while we relax this assumption in (3). 25See Section 5.2 for more details.

12 als in our sample might have heterogeneous preferences for leaving the origin country j. We could follow Schmidheiny and Br¨ulhart (2011) and Bertoli and Fern´andez-Huertas Moraga

(2015) to derive bounds around the unconditional responsiveness of pijk with respect to a 26 variation in one of the elements of xijk.

4 Descriptive statistics

The Gallup World Polls cover the entire civilian, non-institutionalized population aged 15 years and above, with a sample of around 1,000 individuals in each wave of the survey. As discussed in Section 2 above, we restrict our sample to natives aged 15 to 49 who aspire to migrate abroad. Foreign-born individuals are (very) likely to report a distance-one connec- tion in their country of birth, and they are also likely to report the intention to return to their home country,27 and we do not want to blur the intention to migrate with the intention to return. The number of individuals included in the sample for each of the 148 countries depends on (i) the number of waves of the Gallup World Polls conducted between 2007 and 2014, (ii) the share of foreign-born individuals residing in each country, and (ii) the share of aspiring migrants in each country. Table 1 reports the number of waves of the Gallup World Polls for each country, together with the number of aspiring migrants among the natives aged 15 to 49,28 and the number of intended destinations. The total sample size is 153,538 aspiring migrants, which corresponds to an average of 1,037 per country, with the sample size varying between 85 () and 3,037 (Nigeria).

26This uncertainty would disappear if the model in (3) could be estimated on the entire choice set D for each origin, as in Bertoli, Fern´andez-Huertas Moraga, and Ortega (2013). 27The Gallup World Survey reports whether an individual is a native or a foreign-born, but the information about the country of birth is unfortunately missing for 23 percent of the foreign-born that aspire to migrate. 28We also drop individuals with missing data for any of the individual-level characteristics included as controls in the estimation, or that report to aspire to migrate but then do not disclose their preferred destination.

13 Table 1: Sample size and number of intended destinations

Country Waves Obs. Dest. Country Waves Obs. Dest. Algeria 4 1332 48 Tunisia 6 1976 54 Angola 3 658 52 Uganda 6 2254 59 Benin 4 722 53 Zambia 6 1416 58 Botswana 4 880 44 Zimbabwe 5 1894 57 Burkina Faso 5 1373 55 5 587 37 Burundi 3 364 35 Belize 1 112 20 Cameroon 7 2693 70 Bolivia 6 1106 31 Central African Republic 2 629 39 4 560 39 Chad 6 1396 78 Canada 6 279 46 Comoros 4 1711 40 5 1014 41 Congo (Kinshasa) 3 1046 52 Colombia 6 1535 39 Congo 4 1292 58 5 836 41 Djibouti 4 760 43 Dominican Republic 6 2447 35 Egypt 5 1407 44 Ecuador 5 727 35 Ghana 5 2181 50 El Salvador 6 2005 31 Guinea 4 1268 43 6 1451 32 Ivory Coast 2 454 36 Guyana 1 212 17 Kenya 6 2110 71 Haiti 4 833 39 Liberia 3 1579 46 Honduras 6 2170 29 Libya 2 600 28 Mexico 5 843 44 Madagascar 4 508 28 6 1949 27 Malawi 3 1050 47 Panama 5 728 30 Mali 5 1200 50 Paraguay 4 228 27 Mauritania 7 1955 56 Peru 6 1753 42 Morocco 4 1685 40 Trinidad and Tobago 2 108 14 Mozambique 2 406 27 United States 5 291 45 Namibia 1 157 26 6 450 29 Niger 6 1106 45 Venezuela 6 518 32 Nigeria 6 3037 72 Afghanistan 6 1893 39 Rwanda 3 305 34 Armenia 7 1777 38 Senegal 7 2822 53 Azerbaijan 7 1292 39 Sierra Leone 3 1553 41 Bahrain 4 141 26 Somalia 4 1564 43 Bangladesh 7 1988 49 6 1062 56 Cambodia 7 2052 29 Sudan 4 1623 62 China 4 1674 39 Tanzania 5 1321 62 Georgia 7 1224 37 Togo 2 571 34 Hong Kong 4 539 31 (continued)

14 Table 1: Sample size and number of intended destinations (continued)

Country Waves Obs. Dest. Country Waves Obs. Dest. India 6 1695 47 Bosnia and Herzegovina 6 1083 38 Indonesia 7 451 26 Bulgaria 6 733 38 Iran 5 951 41 Croatia 6 530 29 Iraq 4 598 39 Cyprus 5 543 36 Israel 6 584 38 Czech Republic 6 478 38 Japan 6 801 45 Denmark 7 560 50 Jordan 6 1571 48 Estonia 6 750 40 Kazakhstan 7 1016 45 Finland 5 356 47 Kuwait 4 140 21 France 6 707 59 Kyrgyzstan 7 1615 40 7 1022 69 Laos 3 438 22 Greece 6 658 41 Lebanon 6 1793 62 Hungary 6 825 36 Malaysia 6 485 32 Iceland 3 239 24 Mongolia 5 1353 35 Ireland 6 600 37 Nepal 7 1283 44 Italy 6 964 53 Occupied Palestinian Territory 6 1375 51 Latvia 6 761 41 Pakistan 6 1000 45 Lithuania 7 1108 38 Philippines 7 1728 40 Luxembourg 5 303 36 Qatar 3 138 26 Macedonia 5 930 46 Russia 7 2459 66 Malta 5 556 34 Saudi Arabia 6 414 43 Moldova 7 2055 44 Singapore 6 796 39 Netherlands 5 413 50 South Korea 6 1289 39 Norway 2 176 37 Sri Lanka 7 1189 46 Poland 7 809 43 Syria 6 2062 51 Portugal 6 694 38 Taiwan 4 782 38 Romania 6 910 36 Tajikistan 7 1203 30 Serbia and Montenegro 6 2428 53 Thailand 6 379 36 Slovakia 4 481 26 Turkmenistan 4 683 30 Slovenia 5 589 46 United Arab Emirates 4 99 25 6 578 53 Uzbekistan 6 871 28 Sweden 6 588 50 Vietnam 5 479 31 Switzerland 2 85 32 Yemen 5 1911 37 Turkey 6 706 55 Albania 5 1203 26 Ukraine 7 1277 46 Austria 6 295 41 United Kingdom 7 1119 74 Belarus 7 1266 47 5 271 35 Belgium 6 619 56 New Zealand 5 397 35 Notes: we report the number of waves of Gallup World Polls conducted in each country between 2007 and 2013, the number of natives aged 15 to 49 who aspire to migrate and the number of intended destinations. Source: Authors’ elaboration on Gallup World Polls.

24.1 percent of the 153,538 aspiring migrants in our sample have a distance-one connection

15 in at least one destination,29 and 12.8 percent of the aspiring migrants have a distance-one connection in the destination they intend to move to.

5 Estimation

The specification of the conditional logit model that we bring to the data includes: (i) a dummy variable dijk that signals whether the individual i has a distance-one connection to destination k; (ii) a dyadic dummy djk that absorbs the effect of all time-invariant dyadic, origin or destination-specific variables that exert an identical influence on the attractiveness 30 of the (intended) destination k ;(iii) a vector zij of individual characteristics, including sex, four age cohorts,31 and a dummy that takes the value one for individuals who completed at least four years of post-secondary education. Notice that the inclusion of dyadic dummies

djk also controls for the influence exerted by the size of the diaspora of j-born individuals in destination k on the choice of the (intended) destination, as this variable mostly evolves slowly over time, if this enters additively in the function that describes the deterministic

component of location-specific utility Vijk in (2). The empirical specification is thus consis- tent with the econometric evidence provided with aggregate data by Beine, Docquier, and Ozden¨ (2011) and Bertoli and Fern´andez-Huertas Moraga (2015) on the role of the size of the bilateral diaspora in shaping actual migration flows. The conditional logit model is estimated separately for each of the 148 countries in 32 our sample. If the size of the choice set Dj for country j stands at #Dj = Nj, , then the estimation of the conditional logit model requires estimating one coefficient of the alternative-

specific variable dijk plus six times Nj −1 coefficients for the individual-specific variables and 33 the destination-specific intercepts, i.e., a total of 1 + 6(Nj − 1) coefficients. The standard

29The corresponding figure for all natives aged 15 to 49 stands at 12.0 percent. 30When we have more than one wave of the Gallup World Polls for a country, we also include the (one-year lagged) values of the logarithm of the total population, of real GDP per capita, and of the employment rate of the population aged 15 and above (World Bank, 2015a). Given the short time horizon over which we have

data and the inclusion of dyadic dummies djk, we will not discuss the coefficients of these variables. The results are also robust to the inclusion of dummies for each wave of the Gallup World Polls as additional individual-specific regressors, which control for variations over time in the attractiveness of each destination. 31Specifically, 15 to 19, 20 to 29, 30 to 39 and 40 to 49 years. 32 See the third data column in Table 1 for the size Nj of the choice set for each country. 33 Recall that the vector of coefficients βjj is normalized to zero, and that the dummy for the age cohort

16 errors for the estimated coefficients are obtained through bootstrapping (200 replications with replacement).

5.1 Benchmark specification

The minimal size of the choice set for the countries in our sample is Nj = 14 (for Trinidad and Tobago), and it is thus unfeasible to report the 1 + 6(Nj − 1) ≥ 79 estimated coeffi- cients for each country. We thus focus our attention on the estimated coefficients βb1j, with j = 1,..., 148, for our variable of interest dijk. Figure 4 plots the estimated coefficient for distance-one connections for each country against the corresponding z-score.34 The esti- mated coefficient is always positive, and significantly different from zero for 136 out of 148 countries, and the z-score falls short of the value that allows to reject the null hypothesis at the 1 percent confidence level for countries that (mostly) have a very limited sample size, as Figure 4 reveals.

15 to 49 is omitted; the number of estimated coefficients is 4 + 6(Nj − 1) when we also include the three time-varying destination-specific variables. 34The same information is reported in Table A.1 in the Appendix.

17 Figure 4: Estimated coefficient and z-score for distance-one connections

Notes: the figure reports country-specific point estimate for coefficient of distance-one connections from the conditional logit and the corresponding z-score; the surface of each circle is proportional to the sample size for each country. Source: Authors’ elaboration on Gallup World Polls.

The (weighted) average of the estimated coefficient βb1j stands at 2.099, with a standard deviation of 0.631.35 If we consider a j-born individual, having a distance-one connection in country k increases the relative odds of aspiring to migrate to destination k over destination b 2 099 h by more than eight times, i.e., eβ1j = e . ≈ 8.17. Figure 5 reports the estimated coefficient in a world map, and it reveals that there is no clear geographical pattern in the values of the estimates for the coefficient of distance-one connections. The absence of a clear geographical pattern in the data, however, does not preclude uncovering interesting regularities in the size of the estimated coefficients, as we discuss in Section 5.3 below.

35The corresponding unweighted average and standard deviation stand at 1.845 and 0.649 respectively.

18 Figure 5: Estimated coefficients for distance-one connections

Notes: the figure reports country-specific point estimates for coefficient of distance-one connections from the conditional logit. Source: Authors’ elaboration on Gallup World Polls.

5.2 Testing for the IIA property

The estimation of the conditional logit model rests on the property of the independence of irrelevant alternatives, as discussed in Section 3.1 above. We test whether the estimate of βb1j is stable when we re-estimate the model on a restricted choice set. Specifically, for each replication of the estimation on a restricted sample Rj, we see whether the estimated Rj coefficient βb1j falls within the 95 percent (bootstrapped) confidence interval of βb1j, i.e. Rj βb1j ≈ βb1j; we then compute the share of replications for which this is actually the case. We follow two distinct approaches to define the restricted samples Rj over which the conditional logit is estimated: (i) we draw one (intended) destination at a time, as in Grogger and Hanson (2011), thus re-estimating the model Nj times for each origin, with Nj being the size of the choice set Dj; (ii) we rank the countries in Dj by the ascending number of aspiring migrants; we start by dropping the least important (intended) destination, and then we re-estimate the model dropping one more destination each time. The second approach is

19 36 clearly more demanding, as the size of the restricted sample Rj gets progressively smaller.

Rj Figure 6: Share of replications with βb1j ≈ βb1j (one destination dropped)

N βRj Notes: the figure reports country-specific share of the j replications for which the point estimate for b1j β1 lies in the 95 percent confidence interval for b j. Source: Authors’ elaboration on Gallup World Polls.

Rj 1 i Figure 6 reports the share of replications for which βb1j ≈ βb j based on approach ( ), i.e. dropping one destination at a time. On average, 98.6 percent of the replications produce an estimated coefficient for distance-one connections which is not statistically different from the one obtained in the benchmark specification estimated on the entire choice set.

36 The number of replications in this second approach is not higher than Nj − 1, as the conditional logit might fail to converge when just two or three destinations are included in Rj.

20 Rj Figure 7: Share of replications with βb1j ≈ βb1j (more destinations dropped)

βRj Notes: the figure reports country-specific share of the replications for which the point estimate for b1j lies β1 in the 95 percent confidence interval for b j. Source: Authors’ elaboration on Gallup World Polls.

Figure 7 reports the corresponding results when following the more demanding approach

(ii). Notwithstanding the fact that we drop more than half of the choice set Dj, 88.5 percent of the replications produce an estimated coefficient for distance-one connection that lies in the confidence interval of the benchmark specification. Both approaches are thus reassuring about the plausibility of the IIA property that characterize the specification of the location- choice model that we bring to the data.

5.3 Exploring the estimates

Figure 8 plots the estimated coefficients for distance-one connections against the Herfind- ahl Index HIj of the country-specific distribution of migrant stocks across destinations for

2010 from the World Bank (2015a). Letting mjk represent the stock of j-born individ- 2 uals residing in destination k, the Herfindahl Index HIj is defined as HIj = Pk6=j sjk, where sjk = mjk/ Pl6=j mjl. Figure 8 immediately reveals that the size of the estimated coefficient βb1j is negatively related to HIj: countries with a more concentrated diaspora

21 have a lower point estimate of the coefficient of the distance-one connections in the con- ditional logit estimated from the Gallup World Polls data.37 The (weighted) correlation

between βb1j and HIj stands at -0.365. The two countries with the lowest Herfindahl In- dexes of migrant stocks are France and Germany, with HIFRA =0.075 and HIDEU =0.080,

and the estimated coefficients which stand at βb1FRA = 2.477 and βb1DEU = 2.497 respec- tively. The two countries with the highest Herfindahl Indexes are Mexico and Burkina 38 Faso (HIMEX = 0.964 and HIBFA = 0.865), with the estimated coefficients equal to

βb1MEX =0.697 and βb1BFA =1.383.

Figure 8: Estimated coefficients and Herfindahl Index of migrant stocks

Notes: the figure reports country-specific point estimates for coefficient of distance-one connections from the conditional logit and the Herfindahl Index of the country-specific distribution of migrant stocks in 2010; the surface of each circle is proportional to the native population in each country; Occupied Palestinian Territories, Serbia and Taiwan not included. Source: Authors’ elaboration on Gallup World Polls and World Bank (2015a).

A similar strongly negative correlation is obtained if we compute the Herfindahl Index after we normalize the bilateral migrant stocks by the size of the population at destination,

37The correlation between the Herfindahl Index and the ratio between the total stock of emigrants and the size of the native population at origin stands at 0.219, i.e., countries with a higher emigration rate generally have more concentrated diasporas, but the correlation goes to zero once we also control for HIj. 38Mexican and Burkinab´emigrants are concentrated in the United States and Ivory Coast respectively.

22 thus removing the influence that could be exerted by the unequal demographic size of the various destination on the distribution of bilateral migrant stocks. Figure A.4 in the Ap- pendix reports the scatter plot of the two variables, whose weighted correlation -0.474. We also repeat the same exercise using the data on the number of distance-one connections in each country from Gallup World Polls; Figure A.5 in the Appendix reports the scatter plot of the two variables, whose (weighted) correlation stands at -0.575.

What does the evidence of a strongly negative correlation between βb1j and HIj suggest? This evidence indirectly points to the need to consider a more general specification for the deterministic component of location-specific utility Vijk in (2). As discussed above, the specification of the conditional logit model that we bring to the data is consistent with the hypothesis that the individual-specific distance-one connection dijk and (any function of) the size of the diaspora of j-born individuals residing in k enter additively in Vijk, so that the influence of a distance-one connection on Vijk is independent from the size of the diaspora. A more general specification would allow distance-one connections and the size of the diaspora to influence Vijk in a multiplicative way, thus allowing for a relationship of complementarity or substitutability between the two. Figure 8 shows that countries with a diaspora that is more evenly distributed across destinations, i.e., with a lower Herfindahl Index, have a higher estimated coefficient βb1j. A country like France, with HIFRA =0.075, is characterized by a size of its diaspora that does not vary greatly across destinations, while a country like

Mexico, with HIMEX = 0.964, is characterized by greatly varying bilateral diasporas, as (almost) all of its migrants reside in the neighboring United States. The Mexican diaspora attains either very large values (in the United States) or it is very small, while the French diaspora does not take values at either of the two extremes. The fact that βb1MEX < βb1FRA indirectly hints to a complex interaction between distance-one connections and the diaspora: the two variables could be complement when the size of the diaspora is small, and they could be substitutes for larger sizes of the diaspora. How can we rationalize this? Having a personal direct relationship with someone in

country k contributes less to the utility Vijk of an aspiring migrant when the size of the j-born diaspora in k is either very small (in which case the local availability of so-called ethnic goods will be limited making destination k unattractive despite the distance-one connection there) or very large (in which case the individual is likely to have a large number of distance-two connections in k, i.e., having friends of one’s own friends who reside there, so that a distance-

23 one connection matters less). Distance-one connections matter more at intermediate values of the diaspora, when the local availability of ethnic goods makes the (intended) destination more attractive, and an aspiring migrant is less likely to have a distance-two connection there.39 In such a case, we should indeed expect that the destination choice of a Mexican is less influence by distance-one connections than the destination choice of a French aspiring migrant.

6 Concluding remarks

This paper relies on the Gallup World Surveys to provide econometric evidence on the relationship between the connection of an individual to the migrant networks in different countries and her own choice concerning the preferred country of (intended) destination. The data from Gallup give us a much finer measure of migrant networks than those commonly employed in the literature, which allow us to get a deeper understanding of the complex way in which networks influence migration decisions. Exploring further, both from a theoretical and an empirical perspective, the way in which the individual direct connection to the migrant network interacts with the size of the diaspora at destination represents a key direction in which this paper will be extended.

References

Becerra, D. (2012): “The Impact of Anti-Immigration Policies and Perceived Discrimina- tion in the United States on Migration Intentions among Mexican Adolescents,” Interna- tional Migration, 50(4), 20–32.

Beine, M., F. Docquier, and c. Ozden¨ (2011): “Diasporas,” Journal of Development Economics, 95(1), 30–41.

(2015): “Dissecting network externalities in international migration,” Journal of Demographic Economics, 81(4), 379–408.

39 Notice that our arguments do not imply that the effect of distance-one connections on Vijk should be the same at very low or a very high levels of the diaspora at destination, but only that the relationship between this effect and the size of the diaspora in non-monotonic.

24 Beine, M., and S. Salomone (2013): “Network Effects in International Migration: Edu- cation versus Gender,” The Scandinavian Journal of Economics, 115(2), 354–380.

Bertoli, S. (2010): “Networks, sorting and self-selection of Ecuadorian migrants,” Annales d’Economie et de Statistique, 97/98, 261–288.

Bertoli, S., and J. Fernandez-Huertas´ Moraga (2013): “Multilateral Resistance to Migration,” Journal of Development Economics, 102, 79–100.

(2015): “The size of the cliff at the border,” Regional Science and Urban Economics, 51, 1–6.

Bertoli, S., J. Fernandez-Huertas´ Moraga, and F. Ortega (2013): “Crossing the Border: Self-Selection, Earnings and Individual Migration Decisions,” Journal of Devel- opment Economics, 101(1), 75–91.

Carling, J. (2002): “Migration in the Age of Involuntary Immobility: Theoretical Reflec- tions and Cape Verdean Experiences,” Journal of Ethnic and Migration Studies, 28(1), 5–42.

Chort, I. (2014): “Mexican Migrants to the US: What Do Unrealized Migration Intentions Tell Us About Gender Inequalities?,” World Development, 59, 535–552.

Comola, M., and M. Mendola (2015): “Formation of Migrant Networks,” The Scandi- navian Journal of Economics, 117(2), 592–618.

Creighton, M. (2013): “The Role of Aspirations in Domestic and International Migra- tion,” Social Science Journal, 50(1), 79–88.

Delogu, M., F. Docquier, and J. Machado (2015): “Globalizing Labor and the World Economy: The Role of Human Capital,” Manuscript: UCLouvain.

Docquier, F., J. Machado, and K. Sekkat (2015): “Efficiency Gains from Liberalizing Labor Mobility,” The Scandinavian Journal of Economics, 117(2), 303–346.

Docquier, F., G. Peri, and I. Ruyssen (2014): “The Cross-country Determinants of Potential and Actual Migration,” International Migration Review, 48, S37–S99.

25 Drinkwater, S., and P. Ingram (2009): “How Different are the British in their Willing- ness to Move? Evidence from International Social Survey Data,” Regional Studies, 43(2), 287–303.

Dustmann, C., and A. Okatenko (2014): “Out-Migration, Wealth Constraints, and the Quality of Local Amenities,” Journal of Development Economics, 110, 52–63.

Esipova, N., J. Ray, and A. Pugliese (2011): “Gallup World Poll: The Many Faces of Global Migration,” IOM Migration Research Series 43, International Organization for Migration, Geneva, Switzerland.

Gallup (2013): “Worldwide Research Methodology and Codebook,” Washington, D.C.

Giulietti, C., J. Wahba, and Y. Zenou (2014): “Strong Versus Weak Ties in Migra- tion,” CEPR Discussion Paper No. 10248.

Grogger, J., and G. H. Hanson (2011): “Income maximization and the selection and sorting of international migrants,” Journal of Development Economics, 95(1), 42–57.

Guimaraes, P., O. Figueiredo, and D. Woodward (2003): “A Tractable Approach to the Firm Location Decision Problem,” Review of Economics and Statistics, 85(1), 201–204.

Head, K., J. Ries, and D. Swenson (1995): “Agglomeration benefits and location choice: Evidence from Japanese manufacturing investments in the United States,” Journal of International Economics, 38(3-4), 223–247.

Jackson, M. O. (2010): Social and Economic Networks. Princeton: Princeton University Press.

Jonsson,´ G. (2008): Migration aspirations and immobility in a Malian Soninke village, vol. 10. International Migration Institute, University of Oxford.

Manchin, M., R. Manchin, and S. Orazbayev (2014): “Desire to Migrate Internation- ally and Locally and the Importance of Satisfaction with Amenities,” Paper presented at the FIW-wiiw Seminars in International Economics, 10 April 2014, Austria, Vienna.

McFadden, D. (1974): “Conditional logit analysis of qualitative choice behavior,” in Fron- tiers in Econometrics, ed. by P. Zarembka, pp. 105–142. New York: Academic Press.

26 McKenzie, D. J., and H. Rapoport (2010): “Self-selection patterns in Mexico-U.S. migration: The role of migration networks,” Review of Economics and Statistics, 92(4), 811–821.

Munshi, K. (2003): “Networks in the Modern Economy: Mexican Migrants in the U.S. Labor Market,” Quarterly Journal of Economics, 118(2), 549–599.

(2015): “Community Networks and Migration,” in Oxford Handbook on the Eco- nomics of Networks, ed. by J. Bramoull´e, A. Galeotti, and B. W. Rogers. Oxford: Oxford University Press, forthcoming.

Ortega, F., and G. Peri (2013): “The Role of Income and Immigration Policies in Attracting International Migrants,” Migration Studies, 1(1), 27–46.

Papapanagos, H., and P. Sanfey (2001): “Intention to Emigrate in Transition Countries: The Case of Albania,” Journal of Population Economics, 14(3), 491–504.

Pedersen, P. J., M. Pytlikova, and N. Smith (2008): “Selection and network effects– Migration flows into OECD countries 1990-2000,” European Economic Review, 52(7), 1160–1186.

Schmidheiny, K., and M. Brulhart¨ (2011): “On the equivalence of location choice models: Conditional logit, nested logit and Poisson,” Journal of Urban Economics, 69(2), 214–222.

van Dalen, H., G. Groenewold, and J. Schoorl (2005): “Out of Africa: What Drives the Pressure to Emigrate?,” Journal of Population Economics, 18(4), 741–778.

Van Dalen, H. P., G. Groenewold, and T. Fokkema (2005): “The Effect of Re- mittances on Emigration Intentions in Egypt, Morocco, and Turkey,” Population Studies, 59(3), 375–392.

Wood, C., G. C.L., L. Ribeiro, and P. Hamsho-Diaz (2010): “Crime Victimization in and Intentions to Migrate to the United States,” International Migration Review, 44(1), 3–24.

World Bank (2015a): Bilateral migration matrix 2010. Washington DC: World Bank, available at: http://go.worldbank.org/JITC7NYTT0 (accessed on October 10, 2015).

27 (2015b): World Development Indicators. Washington DC: World Bank.

Yang, X. (2000): “Determinants of Migration Intentions in Hubei Province, China: Indi- vidual versus Family Migration,” Environment and Planning, 32(5), 769–787.

28 A Appendix

Figure A.1: Kernel density plot of the share of aspiring migrants (percent)

Notes: the figure reports the Kernel density plot of the percentage of natives aged 15 to 49 in each country who aspire to migrate (Epanechnikov kernel function, with a bandwidth of 3.47); data are pooled across different waves of the survey, and sampling weights are used. Source: Authors’ elaboration on Gallup World Polls.

29 Figure A.2: Share of aspiring migrants and log real GDP per capita in 2010

Notes: each circle corresponds to a country, and the surface of each circle is proportional to the size of the native population residing in each country. Source: Authors’ elaboration on Gallup World Polls and World Bank (2015a,b).

30 Figure A.3: Kernel density plot of the Spearman’s rank correlation coefficient

Notes: the figure reports the Kernel density plot of the Spearman’s rank correlation coefficient between the distribution of the bilateral migrant stocks in 2010 and the distribution of the distance-one connections from our individual-level data (Epanechnikov kernel function, with a bandwidth of 0.037); sampling weights are used; the Occupied Palestinian Territories, Serbia and Taiwan are not included; data are pooled across different waves of the survey, and sampling weights are used. Source: Authors’ elaboration on Gallup World Polls and World Bank (2015a).

31 Figure A.4: Estimated coefficients and HI of migrant stocks (normalized by population)

Notes: the figure reports country-specific point estimates for the coefficient of distance-one connections from the conditional logit and the Herfindahl Index of the country-specific distribution of migrant stocks in 2010, normalized by population at destination; the surface of each circle is proportional to the native population in each country; Occupied Palestinian Territories, Serbia and Taiwan not included. Source: Authors’ elaboration on Gallup World Polls and World Bank (2015a,b).

32 Figure A.5: Estimated coefficients and Herfindahl Index of distance-one connections

Notes: the figure reports country-specific point estimates for coefficient of distance-one connections from the conditional logit and the Herfindahl Index of the distance-one connections; the surface of each circle is proportional to the native population in each country. Source: Authors’ elaboration on Gallup World Polls and World Bank (2015a,b).

33 Table A.1: Estimated coefficient for distance-one connections

Country Obs. Coeff. s.e. Country Obs. Coeff. s.e. Algeria 1332 1.884 0.433 Tunisia 1976 1.084 0.229 Angola 658 2.091 0.299 Uganda 2254 2.213 0.158 Benin 722 1.809 0.422 Zambia 1416 1.667 0.197 Botswana 880 1.115 0.148 Zimbabwe 1894 1.437 0.093 Burkina Faso 1373 1.383 0.148 Argentina 587 1.731 0.229 Burundi 364 1.112 0.494 Belize 112 1.360 0.465 Cameroon 2693 1.650 0.106 Bolivia 1106 1.547 0.113 Central African Republic 629 1.611 0.241 Brazil 560 1.888 0.324 Chad 1396 1.737 0.143 Canada 279 2.402 0.415 Comoros 1711 0.540 0.160 Chile 1014 1.704 0.158 Congo (Kinshasa) 1046 2.430 0.220 Colombia 1535 1.325 0.119 Congo Brazzaville 1292 0.904 0.228 Costa Rica 836 1.169 0.163 Djibouti 760 1.538 0.165 Dominican Republic 2447 1.005 0.088 Egypt 1407 1.540 0.311 Ecuador 727 1.301 0.155 Ghana 2181 1.835 0.125 El Salvador 2005 0.613 0.102 Guinea 1268 1.538 0.243 Guatemala 1451 0.265 0.135 Ivory Coast 454 1.744 0.271 Guyana 212 1.516 3.683 Kenya 2110 1.972 0.151 Haiti 833 1.215 0.187 Liberia 1579 1.395 0.141 Honduras 2170 0.517 0.094 Libya 600 4.659 0.745 Mexico 843 0.721 0.158 Madagascar 508 0.148 0.543 Nicaragua 1949 1.083 0.092 Malawi 1050 1.048 0.220 Panama 728 0.824 0.190 Mali 1200 1.845 0.170 Paraguay 228 2.113 0.468 Mauritania 1955 2.254 0.182 Peru 1753 1.691 0.112 Morocco 1685 2.111 0.191 Trinidad and Tobago 108 1.773 6.073 Mozambique 406 1.273 0.299 United States 291 2.851 0.461 Namibia 157 2.032 0.498 Uruguay 450 1.581 0.234 Niger 1106 2.216 0.141 Venezuela 518 2.281 0.322 Nigeria 3037 1.515 0.123 Afghanistan 1893 2.466 0.131 Rwanda 305 2.098 0.398 Armenia 1777 1.210 0.111 Senegal 2822 1.584 0.076 Azerbaijan 1292 0.955 0.198 Sierra Leone 1553 1.880 0.196 Bahrain 141 3.454 17.175 Somalia 1564 2.206 0.152 Bangladesh 1988 1.945 0.138 South Africa 1062 2.704 0.374 Cambodia 2052 2.026 0.239 Sudan 1623 1.695 0.144 China 1674 1.571 0.197 Tanzania 1321 2.596 0.216 Georgia 1224 1.920 0.147 Togo 571 1.177 0.286 Hong Kong 539 1.461 0.225 (continued)

34 Table A.1: Estimated coefficient for distance-one connections (continued)

Country Obs. Coeff. s.e. Country Obs. Coeff. s.e. India 1695 2.839 0.278 Bosnia and Herzegovina 1083 2.501 0.123 Indonesia 451 2.811 0.385 Bulgaria 733 3.065 0.306 Iran 951 1.835 0.217 Croatia 530 1.606 0.232 Iraq 598 2.358 0.225 Cyprus 543 1.641 0.191 Israel 584 2.051 0.251 Czech Republic 478 2.448 0.310 Japan 801 1.813 0.241 Denmark 560 2.353 0.211 Jordan 1571 2.518 0.264 Estonia 750 1.892 0.177 Kazakhstan 1016 1.785 0.185 Finland 356 1.947 0.323 Kuwait 140 1.464 6.8×103 France 707 2.477 0.248 Kyrgyzstan 1615 1.575 0.149 Germany 1022 2.497 0.179 Laos 438 1.816 0.434 Greece 658 1.665 0.199 Lebanon 1793 2.111 0.143 Hungary 825 2.192 0.171 Malaysia 485 1.730 0.316 Iceland 239 2.253 1.47×102 Mongolia 1353 1.423 0.191 Ireland 600 1.291 0.188 Nepal 1283 1.804 0.205 Italy 964 1.399 0.249 Occupied Palestinian Territories 1375 2.411 0.220 Latvia 761 1.896 0.162 Pakistan 1000 1.954 0.247 Lithuania 1108 1.889 0.144 Philippines 1728 2.181 0.141 Luxembourg 303 1.962 0.278 Qatar 138 2.004 32.147 Macedonia 930 2.068 0.138 Russia 2459 2.210 0.184 Malta 556 1.448 0.203 Saudi Arabia 414 3.313 0.648 Moldova 2055 1.705 0.105 Singapore 796 1.771 0.228 Netherlands 413 2.254 0.276 South Korea 1289 1.612 0.164 Norway 176 2.287 0.439 Sri Lanka 1189 2.687 0.167 Poland 809 2.345 0.161 Syria 2062 2.339 0.239 Portugal 694 2.094 0.158 Taiwan 782 1.986 0.187 Romania 910 2.394 0.176 Tajikistan 1203 0.644 0.218 Serbia and Montenegro 2428 2.229 0.075 Thailand 379 3.829 0.691 Slovakia 481 2.280 0.282 Turkmenistan 683 0.630 0.374 Slovenia 589 1.293 0.265 United Arab Emirates 99 2.494 4.240 Spain 578 1.473 0.200 Uzbekistan 871 1.612 0.253 Sweden 588 1.909 0.184 Vietnam 479 2.879 0.491 Switzerland 85 2.417 0.901 Yemen 1911 1.571 0.229 Turkey 706 2.057 0.295 Albania 1203 2.087 0.102 Ukraine 1277 2.338 0.195 Austria 295 2.474 0.311 United Kingdom 1119 1.950 0.146 Belarus 1266 1.693 0.134 Australia 271 1.794 0.284 Belgium 619 2.090 0.258 New Zealand 397 0.643 0.254 Notes: standard errors obtained through bootstrapping with replacement, 200 replications. Source: Authors’ elaboration on Gallup World Polls.

35