Foreign aid, bilateral asylum immigration and development

This paper measures the links between aid from 14 rich to 113 developing economies and bilateral asylum applications during the years 1993 to 2013. Dynamic panel models and Sys-Generalized Method of Moments are used. The results show that asylum applications are related to aid nonlinearly in a U-shaped fashion with respect to the level of development of origin countries, although only the downward segment proves to be robust to all specifications. Asylum inflows from poor countries are significantly and negatively associated with aid in the short run, with mixed evidence of more lasting effects, while inflows from less poor economies show a positive but non-robust relationship to aid. Moreover, aid leads to negative cross-donor spillovers. Applications linearly decrease with humanitarian aid. Voluntary immigration is not related to aid. Overall, the reduction in asylum inflows is stronger when aid disbursements are conditional on economic, institutional and political improvements in the recipient economy.

source via economic policy measures. One potential policy instrument is foreign aid but views differ widely on its effectiveness and especially on its real effects. Aid may help countries to overcome the political and economic crises that cause the asylum inflows and hence help to deter them, or aid may allow resource-constrained people in poor countries to afford the costs of migration and hence boost applications.
Both views find support in the economic literature, which provides differing answers. A clear example is the collection of studies on the relationship between aid and migration edited by Böhning and Schloeter-Paredes (1994). In it, scholars concertedly try to uncover the basic relationship linking aid and migration but reach contrasting results (a concluding evaluation is in Martin 1994). Views and studies on forced migration also differ. Thielemann (2004) finds that aid has a positive influence on asylum inflows in 20 OECD countries during the 1985-1999period, whereas Neumayer (2005 finds that aid has no effect on applications in Western Europe during the years 1982-1999. Dreher et al. (2019) find the impact of aid on refugee flows to the world and to the aggregate of OECD countries to be initially positive but then to become negative after some years. One reason for these diverse results is the nature of aid itself: its main purpose is that of improving living and economic conditions in the recipient country, and improved living conditions can both prevent and encourage migration and refugee flows.
This paper measures the links between bilateral aid and asylum seeker applications from 113 developing countries in 14 OECD destination economies for each year over the period 1993-2013. Given the a priori uncertain sign of the impact of aid, I hypothesise that it depends on the level of development of the recipient country and use average income as a rough proxy for development. Specifically, I test whether the links between aid and asylum inflows vary with the level of per capita income in applicants' home countries. As aid transferred with the aim of influencing asylum inflows can unintentionally lead to voluntary immigration, I also test this potential secondary relationship. This paper adds to the existent literature in three main ways: it focuses explicitly on foreign aid as a policy tool designed to influence forced migration; it tests whether this influence varies with the level of development of the recipient country; and it measures aid's overall association with inflows, forced and voluntary.
International norms and agreements outline the difference between refugee and voluntary migration. A refugee, or forced migrant, is a person who 'owing to a wellfounded fear of being persecuted for reasons of race, religion, nationality, membership of a particular social group or political opinion, is outside the country of his nationality, and is unable to, or owing to such fear, is unwilling to avail himself of the protection of that country' (Refugee Convention, 1951). A voluntary migrant is a person who leaves the country for any other reason. There is not only a grey area between the two types of migration, but there is also a significant difference: the refugee flees the home country to escape an extremely critical situation but would rather not leave. She neither does choose the destination nor does the destination choose her, as would happen in a totally voluntary setting. On the other hand, a migrant chooses where to move given a clearly defined set of alternatives and opportunities, including the destination country's policies on immigration (Dustmann et al. 2016).
While forced migration is as old as human history, international aid is officially recognised as a transfer of resources from one country to another only since the end of the Second World War. The majority of aid transfers take the form of donations; a minor proportion consists of grants for specific projects. OECD states that 'Official Development Aid (ODA) is administered with the promotion of the economic development and welfare of developing countries as its main objective'. Hence, aid is expected to positively affect development. 1 The empirical literature is inconclusive on the effects of aid disbursements on the economy, institutions and social norms of the recipient country, but directly, and through them, bilateral aid can be expected to be associated with subsequent asylum inflows.
This study's main findings are that the relation of bilateral aid with asylum inflows varies with average income in the origin country: asylum inflows from poor countries are negatively and significantly associated with bilateral aid disbursements in previous years, while applications from medium-income developing economies are positively but weakly related to aid transfers. Only the negative relation between bilateral aid and asylum inflows is robust to all specifications and cofactors. In addition, aid generates negative cross-donor spillovers: asylum applications in the OECD destination are negatively associated to aid disbursements from other countries. A further result is that aid has no association with voluntary immigration. Hence, bilateral aid transfers are followed by less asylum applications but have no effects on immigrant inflows. I use a dynamic panel empirical model with a rich array of fixed effects and specifications-including System Generalized Method of Moments (GMM) in levels and differences-to control for potential endogeneity and the robustness and sensitivity of results. The rest of the paper is organised as follows. Section 2 reviews and resumes the related literature, Sect. 3 presents data sources and descriptive statistics, Sect. 4 describes the estimation strategy, Sect. 5 presents and discusses results and Sect. 6 concludes.

Related literature
A general finding of the empirical literature on the determinants of forced and voluntary migration is that economic factors tend to be more important for voluntary migration and political factors for forced migration (Neumayer 2005). According to the neoclassical model, migration should respond to the difference between average income in origin and destination countries. Empirical evidence supports this prediction (e.g., Hatton and Williamson 2005;Mayda 2010; Grogger and Hanson 2011;Ortega and Peri 2013). Migration flows are also positively related to income in the destination economy (Hartog and Vriend 1989;Katseli and Glystos 1989;Lundborg 1991;Bauer and Zimmermann 1998). Given everything else, emigration should decrease with development in the origin economy (Lucas 2006). Ortega and Peri (2013) find that per capita income 1 The real impact of aid on growth and institutions remains elusive, as it depends on diverse factors, among which the incentives of recipients and donors in transferring and receiving aid. Several studies find that a substantial part of the aid provided by rich economies is unrelated to the real needs of recipient countries (Boone 1996;Alesina and Dollar 2000;Collier and Dollar 2002;Lancaster 2007;Fuchs et al. 2014;Jones 2015). However, there is a certain degree of agreement on a change of approach of Western donors after the end of the cold war. While during the Cold War, the political allegiance of recipients was decisive, afterwards, their economic and institutional development became more important. negatively affects emigration, while it only marginally influences emigration to OECD countries in Dao et al. (2018a). However, the neoclassical hypothesis is not supported by findings that migration first rises with income growth in the origin but then falls after a certain level of development is reached. This bellshaped pattern of emigration, called the 'mobility transition', is present in Martin andTaylor (1996), de Haas (2007) and Clemens (2014), among others. Several studies evidencing this inverted 'U' relationship are based on cross-country data. Clemens (2014) claims that the pattern might hold also in the long run, a timespan that goes beyond that of most panel databases. Other potential determinants of migration tested in empirical studies concern economic factors such as unemployment and political and institutional characteristics of countries (among others, Hatton and Williamson 2005;Docquier et al. 2014).
Foreign aid is one potential determinant of migration decisions. The empirical literature on the link between aid and migration reaches diverse results. Focusing on overall migration, Faini and Venturini (1993) hypothesise that emigration is related to aid as it is to income: in a bell-shaped fashion. They assume that aid consists mostly of income transfers that loosen the resource constraints of would-be migrants. Schiff (1994), Vogler et al. (1997) and Vogler and Rotte (2000) also hypothesise a bellshaped relationship between aid, development and migration. Studies on the link aidmigration collected in Aid in Place of Migration? (Böhning and Schloeter-Paredes 1994) reach contrasting findings. Berthélemy et al. (2009), using cross-country data from a wide set of countries, find that bilateral aid encourages migration from the poorest economies and reduces it from less poor ones. Belloc (2015), also using a cross-section of countries, finds a positive relationship, in this case linear, between foreign aid and total emigration from South Saharan countries. In Nyberg-Sørensen et al. (2003), aid to poor countries has no homogenous effects on migration to rich economies. The authors also test the impact of aid transfers to neighbouring economies of countries with a political crisis. Clemens and Sandefur (2015) argue that the aiddevelopment-migration nexus is positive: more aid to poor countries boosts immigrant flows to rich economies. Most studies on aid and migration are based on migrant stocks. Using data on migration flows, Lanati and Thiele (2017) find a negative relationship between aid and migration, which holds even for the poorest countries.
Several studies find that migration by refugees and asylum seekers depends strongly on political and institutional factors, such as protest, oppression, conflict and genocide in the origin country (Marfleet 2006;Schmeidl 1997;Davenport et al. 2003;Moore and Shellman 2007;and Hatton 2009). Among the economic determinants, in Neumayer (2005), Hatton (2009) and Hatton and Moloney (2015), refugee and asylum seeker flows diminish with higher income in the home country; this contrasts the mobility transition hypothesis. Other potential determinants of refugee and asylum flows are migrant networks, which can lower the costs of international movements and facilitate further inflows from the home country. In Neumayer (2004), stocks of asylum seekers exert a pull effect on new asylum applicants. Hatton (2016) finds a positive influence of immigrant stocks on asylum applications. In Davenport et al. (2003), past refugee migration positively influences refugee stocks. The empirical literature on the direct link between aid and asylum seekers or refugees is scarce and results diverge (Thielemann 2004;Neumayer 2005). Dreher et al. (2019) find significant differences between the short and medium run correlations of aid with refugee flows.

Data and descriptive statistics
The United Nations High Commissioner for Refugees (UNHCR) provides standardised cross-country data on refugees and asylum seekers since 1950, and the OECD statistics division is the main source of standardised data on Official Development Aid (ODA) since 1969. I built a panel database by using data from UNHCR extracted from OECD Statistics on asylum applications submitted by people from 113 developing countries in 14 destination economies-Australia, Austria, Belgium, Canada, Denmark, France, Germany, Italy, the Netherlands, Norway, Spain, Sweden, the UK and the USA-each year during 1993-2013. The period starts when data from former ex-communist countries in Eastern Europe became available. Asylum seekers are individuals who have sought international protection and whose claims for refugee status have not yet been determined. Asylum applications from the list of 113 origin countries account for almost 80% of all asylum application in the selected OECD destinations (and 70% of asylum applications in all Western OECD countries) during the period considered. Data on foreign aid, regarding the Official Development Aid (net disbursements) from each donor (destination country for asylum applicants) economy to each recipient (origin country of asylum applicants), are extracted from OECD Statistics. A complete list of variables and sources, and the list of developing countries, is in Table 5. Figure 1 shows a high number of asylum applications at the beginning of the period considered, which is partly due to the fall of the Berlin wall in 1989, and a rapid decrease afterwards with the consolidation of the new world order. Many refugees from the former republics of the Soviet Union returned home, and new applications in the selected OECD destinations decreased. Another important wave of asylum applications-still underway and expected to last for the next few years-started with the terrorist attacks of September 2001 and the subsequent military conflicts in Afghanistan and Iraq. Substantial increases in asylum applications also followed the Balkans' ethnic conflicts, the 'Arab spring' in Middle Eastern and North African countries and political turmoil in sub-Saharan African and central Asian countries. During the period 1993-2013, bilateral aid initially decreased, then grew from 2000 to 2006 and decreased again afterwards. These turns partly coincide with changes of the composition of aid recipients and of countries of origin of applicants: from 2002, there were fewer asylum applications from Eastern Europe and more from the Middle East and sub-Saharan Africa.

Estimation strategy
In order to study the correlations between foreign aid and asylum applications, I use dynamic panel regressions. The dependent variable is the number of asylum applications made during each year in the destination country. The base regression is where lnY odt is the (log of the) number of asylum applications made by individuals from country o in country d during year t; lnY odt−1 is its value lagged 1 year; it should capture the influence of former asylum seekers on new inflows. T odt includes other dyadic variables. Among them is the variable of interest, aid provided by OECD country d to developing country o. Aid should have a positive or negative coefficient depending on whether it provides incentives to remain or to leave the country and, in the second case, on how it influences individuals' preferences across possible destinations. 2 Another group of dyadic cofactors is migrant networks. Studies generally test the impact of only one network, asylum applicants, refugees or voluntary migrants (Davenport et al. 2003;Neumayer 2004;Hatton 2016). The implicit assumption is that the three have similar characteristics, defined only by the country of origin. However, in the real world, these networks and their effects may remain distinct. For example, voluntary migrants and refugees or asylum seekers may reach the host country in different periods, belong with different social classes or ethnic groups in the home country, or hold different political, religious or cultural orientations. This can make their allegiance to the origin country weaker than their fidelity to their own specific group. Hence, immigrant and refugee networks can exert different pull effects on asylum applicants. 3 Based on this paper's data, the correlation coefficient between Asylum seekers, Bilateral refugees and Bilateral immigrant inflows is below 0.3, too low to hypothesise that they are reunited in a unique network. Therefore, they will be included as distinct cofactors. Distance between origin and destination, a time-invariant dyadic variable, should capture the effective cost of international migration and of cultural dissimilarities between countries. Bilateral trade agreements between origin and destination (a dummy taking value 1 in the presence of agreements and 0 otherwise) can increase reciprocal knowledge between partners and are expected to lower the costs of bilateral movements of people. O ot concerns factors regarding origin countries. Per capita GDP, the main economic proxy for the country's level of development can facilitate or deter asylum seeker inflows: not only higher income implies more resources to escape, but also weaker incentives to leave. Population accounts for the size of the country. Higher levels of political terror and lack of civil liberties should both be strong push factors (Hatton 2004;Neumayer 2004). Natural disasters (proxied by the number of deaths) can boost outflows (Naudé 2010;Neumayer 2005). The number of refugees from the origin country to all destinations except d should be positively correlated to applications in d; Moore and Shellman (2007) and Hatton and Williamson (2005) find that some countries are more prone to 'produce' refugees than others. Similar to bilateral transfers, aid from all other countries (all countries except d) can provide incentives to remain or opportunities to leave. It, however, includes an 'attraction for the donor' component that can divert asylum seekers from d.
D dt concerns characteristics of the destination country. Per capita GDP at destination is a proxy for expected earnings and potentially a pull factor (Neumayer 2004). Population approximates the extension of the labour market. The unemployment rate signals the difficulty of finding a job and has an expected negative coefficient. Policies at destination concerning asylum seekers should also significantly influence the number of applications, but no standardised indicators on these policies are available. Hence, I use two proxies. The first is the rate of rejection of asylum demands from origin o in country d at year t. This is a weak proxy since it includes both pull and push elements. The second, which is more reliable, is an index built by Hatton and Moloney (2015) based on yearly changes in the tightness of refugee policies in the selected countries.
The impact of aid on asylum applications can depend on how individuals react to the improved environment at home and the extra opportunities to leave. To test whether the level of development influences this response, subsequent specifications include the interaction between bilateral aid and per capita income in the origin country: where (ln Bilateral aid odt-1 ) × (ln pc GDP origin ot ) is the interacted term, expected to be significantly associated with applications; as for the main term, Bilateral aid odt−1 , no hypothesis is formulated on the sign of this coefficient. Wider sets of fixed effects will be used to control for multilateral resistance to migration (MRM): people's bilateral flows do not depend solely on the relative attractiveness of origin and destination countries, but also on alternative destinations (Bertoli and Fernández-Huertas Moraga 2013). Endogeneity and reverse causation can be an issue if aid and asylum seekers influence each other. Using a panel dataset on 18 donor and 148 recipient countries during the period 1992-2003, Czaika and Mayer (2011) find that asylum seekers and refugees in the destination economy positively influence bilateral aid. To account for potential endogeneity, I use System GMM specifications in levels and differences based on Blundell and Bond (1998) and Roodman (2009a, b). I also check for the suitability of instrumental variables used in the empirical literature on aid and refugee migration (Dreher et al. 2019).

Base specifications
The base results of the estimation of Eqs.
(1) and (2) are in Table 1. The variable of interest, Bilateral aid, is expected to affect asylum movements from origin (recipient), o, to destination (donor), d. Its coefficient is not signed a priori, but I hypothesise that it is correlated with the average income of the origin country. The dependent variable is the log of the annual applications for asylum-plus one-for each country-pair. Adding one allows me to keep zero-flow observations. 4 The variables Bilateral aid, Bilateral refugees and Bilateral immigrants are lagged one period to allow their influence to affect asylum seekers. All regressions include a time trend and year fixed effects. In column 3, country effects are controlled for. In column 4, the 2SLS specification is used. Tests in columns 5-7 are based on the Sys-GMM specification in levels and differences. In column 6, the interacted term is not included, but Bilateral aid is split in five parts, corresponding to the quintiles of the income of the origin country; therefore, its effects are linear. The pooled ordinary least squares (OLS) estimates provide a first idea of how the data are correlated without controlling for country fixed effects and therefore overestimate the coefficient on the lagged dependent variable. This coefficient ranges from 0.82 in the OLS specification of column 1 to 0.56 in the OLS-FE specification of column 3. It is always significant at the 1 % level, indicating that past asylum seekers have a robust influence on new applications. As expected, the value of the coefficient in the Sys-GMM specification lies between the values of the OLS and OLS-FE coefficients (Roodman 2009a).
The variable of interest, Bilateral aid, is negatively and significantly related to asylum applications in column 1. 5 Column 2 suggests that this aggregate coefficient may vary with income levels in the origin country: the estimated coefficients on Bilateral aid and on the interacted variable (Bilateral aid) × (pc GDP origin) are not significant, but they suggest that bilateral aid has a negative association with asylum applications from low-income countries and a positive one for mediumincome countries. Except for column 4, where the relation is linear, subsequent 4 Part of foreign aid is concessional in character and conveys a grant element (OECD). As an effect of interest repayment, some figures are negative. However, they are a very small proportion of total observations and have been substituted by zeros. 5 Following Alesina et al. (2013), the share of variation in asylum application explained by foreign aid can be calculated by excluding Bilateral aid from the regression of column1. This makes the R 2 to shrink from 0.871 to 0.868. Hence, Bilateral aid accounts for 0.3% of the total variation in asylum applications and 2.3% of the residual variation left unexplained by the control variables. The latter is calculated as (0.871-0.868)/(1-0.868). The same procedure shows that the exclusion of Bilateral aid and (Bilateral aid) × (pc GDP origin) produces almost identical figures.

Table 1
Baseline regression results for bilateral asylum applications Foreign aid, bilateral asylum immigration and development Table 1 (continued) (1)  Robust standard errors clustered by country-pairs in parentheses. Corrected standard covariance matrix, robust to panel-specific autocorrelation and heteroskedasticity. Constant not reported. The sample comprises data from 1993 to 2013. Variables are in logs. Columns 5-7: Sys-GMM, in levels and differences: LDV, bilateral aid, refugees and immigrants are treated as potentially endogenous; all other control variables are treated as predetermined. AR (2) and AR (3) are second-and third-step Arellano-Bond tests for serial correlation. AR (1), not reported, = 0. Column 6: bilateral aid: split in quintiles of pc GDP origin. Column 7: aid is a percentage of pc GDP origin. Total effect of pc GDP origin is calculated by summing the coefficients on pc GDP origin and on (bilateral aid t−1 ) × (pc GDP origin), evaluated at the average level of bilateral aid regressions confirm this U-shaped relationship: coefficients on Bilateral aid are always negative and significant, and are positive and smaller on the interacted variable in columns 3, 5 and 7. The 2SLS specification of column 4 is a first attempt to deal with potential endogeneity. Dreher et al. (2019) measure the impact of aid on asylum flows to the world and to OECD countries by using an instrumental variable (IV) based on an index of government fractionalization in the destination country and the probability of the developing economy of receiving aid (defined as the average of past aid transfers). I built the IV following their procedure. 6 Results in column 4, based on the Dreher et al.
(2019) IV, show that Bilateral aid has a negative and linear relation with asylum inflows, stronger than in other specifications in Table 1, and in further tests in this study, to be seen below. With the IV specification, a 1% increase in aid transfers decreases applications by 0.4%, independently of the level of development of the home country. Adding the interacted variable makes the instruments to fail the first-stage tests. More generally, this approach shares the general weakness of fixed effects dynamic models: coefficients are biased, even controlling for endogeneity with 2SLS (Nickell 1981).
Hence, in what follows I use the Sys-GMM specification in levels and differences, which leads to consistent and unbiased results. In columns 5-7, Bilateral aid, Bilateral refugees, Bilateral immigrants and the lagged dependent variable are specified as potentially endogenous regressors; all other variables are treated as predetermined and instrumented with their own lags and differences. Sys-GMM results confirm the U-shaped relationship of previous tests: coefficients on aid and on the interacted variable in columns 5 and 7, respectively, are negative and positive; significance is at the 5% level. More specifically, in column 5, the turning point takes place at an average income of about 3000$ at constant 2011 US$. The reported values of tests for serial correlation and over-identification restrictions confirm the validity of results. 7 In column 6, the linear effect of aid on applications is tested in relation to different income levels of developing countries. 8 Specifically, the per capita income is split in quintiles, and aid is multiplied by a dummy taking value 1 for each quintile and 0 otherwise. Coefficients on Bilateral aid are negative for poor countries in the first three quintiles and positive for higher income levels. This confirms previous results, with a difference: in column 6 only negative coefficients of the lower two quintiles are significant (at the 5% level), while in columns 5 and 7, coefficients (with opposite signs) and significance are similar in the two segments of the U-shaped relationship. In column 7, the variable of interest, Bilateral aid, is a share of the pc GDP origin. 9 Several empirical studies use this measure, as well as that of aid/GDP. Results are as above: a negative coefficient on Bilateral aid/pc GDP and a positive and smaller one on the interacted term, both 6 In regressions, not shown to save space, I also added the 'affinity' IV proposed by Alesina and Dollar (2000), based on dyads' coincidence of votes at the United Nations, but it failed first-stage tests. 7 The size of the panel is N = 1582 (country-pairs), T = 21 (years). Although there might seem to be a high number of number of instruments, it is always lower than N (Roodman 2009b). 8 I thank an anonymous reviewer for suggesting this test. 9 Bilateral aid as a share of the per-capita income of the recipient country decreases with the country's level of development. It varies from 0.13% in the lowest income quintile to less than 0.003% in the highest quintile. Detailed figures are available from the author upon request. with significance at the 5% level. To test for the robustness of these results, in column 1 of Table 7, logs on Bilateral aid were taken without adding one. The presence of zero asylum flows between countries in certain years makes observations to drop by about 20%, but previous results are confirmed.
More specifically, what is the impact of aid on asylum at different levels of development of home countries? The Total effect of Bilateral aid is the result of the sum of the coefficient on Bilateral aid plus the coefficient on the interacted variable, at each level of pc GDP origin. Considering column 5, the coefficient on Total effect of Bilateral aid for the poorer countries, such as Afghanistan, Burundi, Eritrea, Ethiopia, Liberia, Malawi, the Democratic Republic of Congo and Central African Republic, is − 0.051, with significance level at 5% (standard error 0.023). For less poor economies, the magnitude of the coefficient is − 0.036, with significance at 10% (s.e. 0.019). Aid transfers to countries at intermediate levels of development have no effect on asylum applications. They are again significant, this time with a positive coefficient of 0.048, and significance at the 5% level (s.e. 0.022), for richer developing economies, such as Argentina, Brazil, Chile, Saudi Arabia, Kuwait or United Arab Emirates. Results are quite similar in column 1 of Table 7, where logs on the dependent variable were taken without adding one. Results are also similar in column 6 of Table 1, where aid links with aid are split in income quintiles. Therefore, aid coefficients are related to the country's level of development in all specifications, except in the 2SLS model of column 4. Overall, only the negative relation of bilateral aid with asylum is always significant: a 1% increase in bilateral aid to the recipient country is associated with less asylum applications from that country, in a measure between 0.05 (column 5) and 0.4% (column 4).
Coefficients on bilateral aid are small, but they report short run effects. Their longrun elasticities are defined by (coefficient on Bilateral aid t−1 ) / (1 − coefficient on Y t−1 ). For example, in column 6, where the coefficient on Y t−1 is 0.82, the long run effect of bilateral aid on applications from countries of the first income quintile is − 0.28 (s.e. 0.13); on applications from countries of the second quintile is − 0.21 (s.e. 0.9); and on applications from countries of higher quintiles, significance is below the 10% level. Hence, a 10% increase in bilateral aid to a poor country will permanently decrease applications by almost 3%. Medium-run effects will be considered below.
These findings show that aid to poor countries provides individuals with incentives and resources to stay, while transfers to medium-income economies can be positively related to applications. More generally, this suggests that aid to poor countries has a stronger impact in improving living conditions than in loosening the resource constraints of potential asylum applicants. If the dominant force at work were the resourceconstraint, then the response of inflows from poor countries would be positive and stronger than that from medium-income economies. Rather than the U-shaped pattern evidenced by this paper's results, a bell-shaped relationship in income between aid and asylum applications would emerge (as hypothesised, among others, by Clemens and Sandefur 2015). More generally, this study's results are consistent with the definition of asylum seekers and refugees as individuals who are forced to leave their country but would rather not move (Dustmann et al. 2016). Especially in poor countries, foreign aid can represent the critical support that allows people to remain or to move temporarily to a nearby country. At the other extreme, aid to medium-income developing countries has a less crucial impact on living conditions but, by increasing the knowledge on the donor economy, can attract asylum seekers. On average, they are endowed with higher levels of human capital and would face lower costs of integration at destination than applicants from the poorer countries (Dao et al. 2018a).
Further findings are that both refugees and immigrant networks exert a pull effect on asylum seeker inflows. On average, there is a positive interaction between immigrant, refugee and asylum networks from the same origin country. Moreover, once endogeneity is controlled for, immigrants exert a stronger effect on asylum applicants than refugees (except for column 6). This can be due to immigrants being more settled in the host country and hence being more able to provide effective support. Distance, as expected, has a negative and significant impact: a 1% increase in distance leads to a decrease in applications that ranges from 0.16 to 0.2%. This is consistent with empirical evidence showing that the great majority of world refugees move to nearby countries, with only a minor proportion of them migrating to the more distant OECD destination. 10 None of the origin countries in the sample is in the same region of, or shares a border with, the selected Western economies.
An important result is that asylum applications in rich countries decrease as origin countries develop; this finding is robust to the different specifications of Table 1, and it will be seen below, through all specifications in this study. The coefficient on pc GDP origin is always negative and significant. The total effect of per capita income is shown in a separate row where the variable is interacted with Bilateral aid. The Total effect of pc GDP origin is the sum of the coefficient on pc GDP origin and on the interacted term (Bilateral aid) × (pc GDP origin) evaluated at the average value of Bilateral aid. It is always negative and significant. Interestingly, the pc GDP of the origin country has a negative, strong and significant effect also in the 2SLS regression of column 6 (oppositely to Dreher et al. 2019). This negative effect of development on asylum supports similar findings in Hatton (2009) andNeumayer (2005). It implies that any positive influence of aid on growth contributes to indirectly deter asylum inflows. More generally, this paper's results contrast the hypothesis that aid transfers to developing countries, directly or indirectly, boost asylum applications (Clemens and Sandefur 2015). 11 The coefficient on Population origin is positive and significant in columns 1-5. Regarding the destination economy, both average income, which can be considered a proxy for the expected wage, and the size of its population have no robust effects on asylum applications. Table 2 shows the results of testing the effects of further cofactors on asylum applications. Column 1 includes variables concerning more characteristics of the destination country and column 2 of the origin economy; column 7 is the most complete specification. As expected, the coefficients on the rate of unemployment in the destination country are negative and significant (columns 1 and 7). If unemployment at destination increases by one percentage point, asylum applications diminish by about 2%, with significance at 1% (column 7). Similar coefficients, not shown to save space, are in Table 3. This result supports previous findings (Thielemann 2004). As in Table 1, the per capita GDP at destination has a non-robust impact on asylum applications. A similar finding is in Hatton (2016) and other studies on refugees and asylum seekers. Combined with the more robust result on unemployment, it suggests that asylum seekers value the prospect of employment above that of wage levels.

Robustness and sensitivity
Also as expected, political conditions in the origin country strongly influence individuals' decisions to move to the OECD destination. An increase in political terror and lower levels of civil liberties substantially affects applications (columns 2 and 7). Hatton (2016) finds political terror to be one of the most important and robust determinants of asylum flows. In Table 2, a one-point increase in the five-point scale of Political terror increases asylum applications by about 12%. In Moore and Shellman (2007), higher levels of dissident violence and government terror increase the number of refugees relative to the number of internally displaced. The tendency of the origin country to be prone to 'produce' refugees, Refugees to other countries, has a small and not significant influence on applications in the OECD destination (columns 2 and 7). Similarly, natural disasters have no influence on the number of asylum seekers. This can suggest that people consider natural disasters as transitory phenomena, which can be overcome without moving to a faraway OECD country. A similar result is in Moore and Shellman (2007) and Clemens (2014). Neumayer (2005) finds that natural disasters and famine generate internal or cross border migration, rather than flight to distant destinations.
What is the influence of aid provided by all other countries on the applications to the Western destination, d? The regressor Aid from all others includes all donors except d. Columns 3 and 7 show that it generates negative and significant cross-donor spillovers: aid transfers from all other countries to o reduce applications from o in d. This can be partly due to aid making living conditions more bearable in the origin country, and partly to its attraction-for-the-donor effect, which, in this case, 'deviates' asylum seekers to non-d donors. 12 A similar question concerns the effect of aggregate aidfrom all donors including d-on the applications from country o to d. Differently from Bilateral aid, the variable Total aid varies only across developing countries. The negative coefficient on Total aid, lagged 1 year, and the positive and smaller one on 12 A world economy where countries minimize the expenditure in aid for given levels of social welfare functions and negative aid spillovers can be characterised by multiple equilibria. Given other countries' transfers to a specific destination, a donor can choose to reduce its own attraction effect by reducing its aid transfers, and benefit from the attraction to the other donors. However, a generalised move of this kind would produce inferior equilibria: by worsening living conditions in poor countries, it would lead to higher aggregate asylum inflows (Table 1). Jones (2015) finds evidence of positive bandwagon effects, especially among larger donors.
(3)   (3) are second-and third-step Arellano-Bond tests for serial correlation. AR (1), not reported, = 0 * p < 0.1; **p < 0.05; ***p < 0.01 Table 3 Robustness of results for bilateral asylum applications. Sys-GMM, PPML (3) Foreign aid, bilateral asylum immigration and development Table 3 (continued) (1) (3) R 2 = 0.759. Covariates not reported *p < 0.1; **p < 0.05; ***p < 0.01 the interacted term (Total aid t−1 ) × (pc GDP origin), both significant at the 1% level, confirm and reinforce the above results: aid deters asylum seeker inflows from the poorer countries and can attract applicants from medium-income developing economies (column 4). Bilateral trade agreements between countries improve the reciprocal knowledge on the partner's institutions and social costumes and norms, potentially decreasing the costs of migration. Hence, a reasonable expectation is that Bilateral trade agreements has a positive influence on asylum applications. However, the coefficient on the variable is not significant, neither in column 6 nor in 7. Table 3 presents the results of further tests of robustness and sensitivity. Cofactors are included in the regressions, but coefficients are not reported to save space. As in Table 2, all regressions are based on the Sys-GMM in levels and differences, except for column 1, where I use the Poisson Pseudo Maximum Likelihood (PPML) method of estimation, proposed by Santos Silva and Tenreyro (2010), with time and country fixed effects. A potential issue relates to zeros in the dependent variable. They about 18% of the total observations, which is not a proportion that should lead to biases in coefficients, 13 but I use the PPML estimator to check for this possibility, and for potential heteroskedasticity. With it, the dependent variable can be used in levels rather than in logs and zero values of applications can be included as they are. Column 1, reporting the PPML coefficients on Bilateral aid and the interacted variable, shows that results remain very similar to those of previous specifications.
The empirical literature finds that destination countries' policies and norms on the recognition of the status of refugee affect the flows of asylum seekers. A first, imperfect proxy for such policies is the proportion of rejected applications from country o in country d. UNHCR provides data on the rates of rejection only since year 2000. Results show that the variable Proportion of rejections has no significant effect on asylum seekers (column 2). The variable equals 1 minus the recognition rates used by Neumayer (2004), who finds a very small but positive effect of recognition rates on the inflows of asylum seekers to Western European countries during the period 1982-1999. A more precise indicator of countries' policies on refugees is the Asylum Policy Index built by Hatton and Moloney (2015). It concerns 48 origin countries and 19 destinations-including the selected 14 OECD countries of this study-during 1997-2012; it varies between destinations and is constant across origins. Its values range between − 4 and 11, with higher numbers indicating more restrictive policies. Column 3 of Table 3 shows the effects of this indicator of destination countries' policies on asylum applications. The coefficient on the Asylum Policy Index has the expected sign and is significant at the 1% level: a one-point increase in the index reduces asylum applications by about 1.3%. In Hatton and Moloney (2015), policies have stronger effects, but their dataset comprises only origin countries with more than 300 asylum applicants, while the present study has a more extended database, which includes also observations with zero applicants. To control for temporal policy shocks concerning all origin countries, the policy index is interacted with time dummies in column 4 of Table 7. Coefficients are below the significance level.
A different issue concerns the possibility of sample heterogeneity and structural break. The patterns (Fig. 1) and the geographical composition of asylum seeker flows and bilateral aid change after year 2000: asylum applications start to decline while aid transfers continue to increase. At the same time, the share of asylum seekers from Eastern Europe declines. The tightening of several Western countries' policies on immigration following the terrorist attacks of 11 September 2001 may partly explain the temporary decline in total asylum inflows, but other factors can also be at work. Hence, Bilateral aid is split into two periods: first, it is multiplied by a dummy taking value 1 in years 1993-2002 and 0 otherwise and, second, by a dummy taking value 1 in years 2003-2013 and 0 otherwise (column 2, Table 3). Results show that the results on bilateral aid and on the interacted variable are similar in the two periods. Hence, the hypothesis of homogeneity and absence of structural break cannot be rejected, However, the impact of aid on inflows is smaller in the second period. This may be due to the same reasons that make the two periods to differ. One is improved economic and social conditions in most of Eastern Europe; another is the terrorist attacks of September 2001. Improved living conditions imply both fewer incentives to leave and less need of external aid. This weakens the link between the two variables. At the same time, the higher instability in regions of Middle East and Central Asia (Afghanistan, Iraq, then Syria, Yemen) after 9/11 weakens the impact of aid on asylum outflows, while it increases that of political terror. To test these hypotheses, I have re-estimated model 5 of Table 3 after excluding Eastern Europe, the Middle East and Central Asia from the sample. This makes the coefficients on bilateral aid and the interacted variable in the two periods more similar. Interestingly, they become similar because coefficients in the first period shrink, which suggests that, during the first period, aid to these regions had the expected influence on refugee inflows. Regressions are available from the author upon request.
Up to now, the variable of interest, Bilateral aid, concerned the totality of aid transfers (including development, education, trade, infrastructure, other purposes, and humanitarian aid). The underlying hypothesis was that all aid improving living conditions in the recipient country could influence the choices-of staying, leaving, and destination-of potential refugees. However, it can be hypothesised that people in critical and extreme situations may be more directly influenced by humanitarian aid, which is specifically conceived for these events, than by broad transfers. Hence, a variable reporting data on Humanitarian bilateral aid from d to o replaces Bilateral aid, and its relation with asylum seekers is tested. Data on humanitarian aid are extracted from the same OECD dataset on foreign aid that provides the data on Official Development Assistance used above, but results are not strictly comparable because observations are about 50% of those on Bilateral aid. Moreover, the geographical distributions of the two types of aid recipients differ: humanitarian aid is more concentrated in poor and politically dangerous countries. Results show that the coefficient on Humanitarian bilateral aid is about − 0.07, with significance at the 5% level (column 5, Table 3). 14 To control whether humanitarian aid is also related to the 14 Nyberg Sørensen et al. (2003) state that 'aid selectivity tends to allocate development aid to the well performing countries and humanitarian assistance to the crisis countries and trouble spots. However, development aid is more effective than humanitarian assistance in preventing violent conflicts, promoting reconciliation and democratization, and encouraging poverty-reducing development investments by migrant diasporas.' (p. 6). average income of the origin country, I tested the effect of the interacted variable, (Humanitarian bilateral aid t−1 ) × (pc GDP origin), but results, not shown to save space, on the main term and on the interaction are both not significant. Hence, humanitarian aid has a deterring effect on asylum seeker applications that is invariant in the origin countries' average incomes.
The long run influence of Bilateral aid has been considered above, but the significance of coefficients in the medium run is also of interest. To this purpose, the sample has been split into two parts, each comprising countries below or above median income, and the Bilateral aid variable has been lagged several periods. Splitting the sample is useful in order to capture the relation between the impact of aid and the level of development without using the interacted variable. A lagged interacted variable in the Sys-GMM specification in levels and differences would substantially increase the risk of instruments proliferation and error autocorrelation (Roodman 2009a, b). Columns 6 and 7 of Table 3 show that the first and sixth lags of Bilateral aid in poor countries are negative and significant, while none of the lags in the group of less poor countries are significant. Table 8 reports coefficients on intermediate lags. When only the first lag is considered, a 10% increase in bilateral aid to poor countries is followed by a 0.63% drop in applications the following year (column 1, Table 8). The lowest two quintiles in Table 1, column 6, report a similar result. On the other hand, coefficients in less poor countries do not follow a clearly defined path: the first lag is positive, but the third, fourth and fifth lags are negative and significant (columns 7, 9, 10, 11 in Table 8), evidencing that the upward sloping segment of the U-shaped relationship between aid and applications is not robust. In both samples, coefficients on a seventh lag are non-significant. Because of the gradual shrinking of the sample size, lags beyond the seventh are not tested. Hence, there is a significant and substantial negative relationship between aid flows to poor developing countries in 1 year and asylum applications the following year, though this relationship is not evident the 4 years that follow. Specifications with 6-year lags find a significant negative relationship with asylum applications 6 years later, but this finding can be contingent on the lag structure chosen. However, tests on coefficients show that they are always jointly significantly different from 0, except for column 5 (Table 8). The balance of this evidence supports a stronger short-term negative relationship between aid and asylum applications, with suggestive but weaker evidence of a further delayed relationship requiring further investigation. 15 My central hypothesis, in which the level of development of the origin country can influence the relation of aid with asylum applications, finds support in the data. However, in principle, other characteristics of countries could also influence the impact of aid. To test this possibility, I interacted Bilateral aid with two variables that have a robust influence on asylum applications and in principle could influence aid effects: they are Distance and Political terror. In the first case, the link of aid with asylum inflows can be expected to be stronger for closer countries, with lower costs of immigration; in the second, aid might deter asylum inflows more effectively where political terror is not at the highest levels. Situations of high political and civil disruption can weaken the effects of aid transfers. In both cases, coefficients on the interacted variables are not significant. Subsequently, I tested the interactions between Bilateral aid and the other regressors; also, these coefficients on the interacted variables are not significant. Results, not shown to save space, are available upon request.
A possible further check might consist in substituting zeros for the missing observations of the dependent variable (about 22% of total observations) and running the regressions on the augmented dataset. The substitution would be justified only if it were reasonable to presume that missing observations coincide with very low numbers of asylum applicants. However, a check on the countries' sources of data shows this not to be the case. Each country's statistics depend on specific practices and methods of data collection rather than on the magnitude of the flows. For example, figures from Canada in OECD Population Statistics are available only since year 1996, but Canadian sources of data show that substantial numbers of asylum seekers and refugees were present in the country before that time. As similar evidence is available for other destinations, I do not perform the substitution.

Aid and immigration
Does bilateral aid affect voluntary migrant inflows? This question matters because aid transfers meant to influence asylum inflows might have unintended effects on immigrants. For example, aid to poor countries might deter asylum seekers but boost voluntary immigration. Hence, testing the association of aid with voluntary migration is useful to uncover the overall relation of aid with inflows, both forced and voluntary. Table 4 depicts the results of using Immigrant inflows in country d from country o instead of Asylum applications; the other variables remain the same. All regressions include time trends and time effects. As stated above, the general expectation is that voluntary migration is more affected by economic factors and less by political disruption than asylum migration. Bilateral aid has no effect on immigrants: the estimated coefficient on the variable is positive in column 1, concerning the OLS regression and not significant and mostly negative in the other regressions. In column 2, the OLS-FE specification includes country and country-pair effects. Columns 3-6 are based on Sys-GMM tests in levels and differences, where Bilateral aid, Asylum applications, Bilateral refugees and the lagged dependent variable are included as potentially endogenous and the other variables as predetermined. Column 4 tests whether the impact of Bilateral aid on immigrants is correlated with the average income of the origin country. Results show that the coefficients on the interacted term and on Bilateral aid are not significant. Hence, there is not a unique and robust link between aid and voluntary migration; this supports the results of previous research (Böhning and Schloeter-Paredes 1994).
A related and highly debated question concerns the indirect effects of aid on migration, especially through its potential positive influence on growth. As seen above, some authors hypothesise a bell-shaped relationship between aid, emigration and development (among others, Faini and Venturini 1993;de Haas 2010de Haas , 2011Clemens 2014). In Table 4, the influence of the average income of the home country on immigrant flows is positive in all specifications, but it is not always significant. To Table 4 Regression results for bilateral immigrant inflows   (1), not reported, = 0 *p < 0.1; **p < 0.05; ***p < 0.01 test whether the income effect is non-linear, column 5 includes the squared term of pc GDP origin. There, both results, on pc GDP origin and on the squared term, are nonsignificant. Hence, development has a non-robust effect on immigrant flows. This supports Dao et al. (2018a), who find that income has a minor role on the upwardsloping segment of the bell-shaped mobility transition. More generally, the results of Table 4 show that aid has neither a direct effect on immigrant inflows nor an indirect one through development.
Other results are that, differently than expected, the average income of the destination country is not a robust pull factor of immigration. The coefficient is negative and significant in column 1 and positive and significant in the OLS-FE specification of column 2. With the Sys-GMM specification, the coefficient is always negative and significant (columns 4-6). These results contrast the thesis that migration is strongly driven by differences between the incomes levels of origin and destination country (Hatton and Williamson 2005;Mayda (2010); Ortega and Peri 2013). On the other hand, the dimension of the country, proxied by Population destination, appears to be a robust pull factor (columns 1-6), while higher levels of unemployment exert the opposite effect (column 6).
More importantly, push factors that strongly affect asylum and refugee migration, such as Political terror and Civil liberties, either have no influence or work in the 'wrong' direction with immigration. The coefficient on Political terror is very small and non significant, while the negative and significant coefficient on Civil liberties shows that a decrease in civil liberties in the home country is correlated with less migration to the Western economies (columns 4 and 6). In sum, asylum and voluntary migration differ in important aspects: bilateral aid has no effect on voluntary migration; higher levels of income at home are negatively associated with asylum applications but have uncertain links with immigration; political factors are quite important in explaining asylum inflows and not significant in determining immigration. One shared determinant is unemployment in the destination economy: it is negatively correlated with both types of inflows.

Summary and conclusions
The main question of this study was whether the inflows of asylum seekers in Western economies are associated to previous bilateral aid disbursements to their home countries. To this purpose, I measured the relation of bilateral aid from 14 Western donors to 113 developing countries with asylum seeker inflows during 1993-2013. Using this comprehensive dataset, I found that bilateral asylum applications from poor countries diminish with past with aid disbursements. The result is negative and significant in the short and the long run. Specifically, a 10% increase in aid to countries with per capita income below the median level is followed by a reduction in asylum applications of above 0.6% the following year and 3% in the long run. On the other hand, applications from less poor countries are related positively with bilateral aid transfers in the short run and negatively in the medium run. Hence, only asylum applications from poor countries appear to be robustly and, negatively, associated with bilateral aid transfers. The estimated association does not arise from worldwide changes in total asylum flows (it is robust to time fixed effects), it does not arise from time-invariant heterogeneity in origin-destination pairs (it is robust to origin, destination, and dyad fixed effects), and it is robust to instrumental variables methods. These features lend substantial support to the interpretation of the estimates as a causal effect of aid on asylum flows. Moreover, asylum applications are negatively associated with humanitarian aid-a more restricted type of transfer-at all levels of income.
As in previous studies, political and institutional conditions in the home country, especially the level of political terror, are strong push factors of asylum flows into Western economies (Hatton 2015). One general, important, and robust result of this study is that economic development in the home country is negatively associated with asylum applications. It adds to the current debate on growth and migration, showing that the often hypothesised bell-shaped relation between the two, or mobility transition, does not apply to forced migration to rich Western countries. Higher per capita income in the home country is unambiguously related to lower numbers of asylum applications. A rationale for this result is that higher levels of income tend to be associated to better political and institutional conditions in the home country, all of which represent incentives to stay. Moreover, these findings imply that asylum inflows diminish with aid that promotes development.
Could aid to poor countries be negatively associated with asylum applications and at the same time be followed by higher voluntary immigration? Using different specifications and cofactors, I find that bilateral aid has no relation with immigrant inflows. More generally, in contrast with the mobility transition hypothesis, the average income of the origin country has no robust relation with immigration. This implies that any positive effect of aid on growth would not attract immigration, even indirectly. At the same time, voluntary migration is strongly related to population levels in the origin country (Dao et al. 2018a, b) and expected employment in the country of destination. Hence, overall, aid transfers would be followed by less asylum inflows from poor countries and no changes in immigration.
Other results are that aid has negative cross-donor spillovers: more transfers from other donors are negatively associated with the number of asylum applications in the OECD destination. This makes free riding potentially appealing for individual donors. However, less aid, stalled development and negative political and institutional conditions in developing countries raise inflows in all destinations. Hence, concerted aid transfers made conditional on improvements in the economic and political institutions of the recipient country can have stronger overall relations with asylum inflows than uncoordinated and unconditional disbursements. Table 7 presents some further robustness tests. Regressions include all covariates, but, to save space, only the coefficients of the variables considered are reported. In column 1, the logs on the dependent variable are taken without adding 1. This implies that about 18% of observations, corresponding to zeros, are lost. Column 2 includes controls for multilateral resistance to migration. They are as follows: origin-time effects, which should capture all time-varying terms that are constant across destinations d and only vary by year and country of origin; destinationtime effects, meant to capture time-varying terms that are constant across origins, o, but vary year and country of destination; destination-origin dummies, which absorb all time-invariant dyadic variables that affect asylum applications; and destination fixed effects, which account for factors of the destination country that are invariant or change very slowly along time, such as culture or institutions and origin fixed effects which absorb similar factors of the origin country. This is a very demanding specification, where measurement is entirely concentrated on within country-pair time variations.
It has been hypothesised that some applicants could be 'bogus' asylum seekers (Neumayer 2005). For example, irregular immigrants who correctly foresee they will not be eligible for the refugee status might nonetheless apply for asylum, only to avoid deportation during time needed for the application to be processed. To control for this possibility, I restrict the sample to countries of origin with above average levels of political terror. Presumably, they are more likely to generate flows of 'genuine' asylum applicants. Large geopolitical shocks, such as the Afghan and Iraq wars, might affect Western countries' policies on asylum. In column 4, the Asylum Policy Index (previously tested in column 3, Table 3) is interacted with time dummies. Column 5 includes the Dreher et al. IV among the instruments of the Sys-GMM specification. Results in columns 1 to 5 are as in previous tests.
Former colonial links between origin and destination country might alter the choices of asylum seekers among potential destinations, as well as those of donors among potential aid recipients. In column 6, Bilateral aid and the interacted term are multiplied, first, by a dummy taking value 1 if the origin country was a donor's colony in 1945 and 0 otherwise, and, second, by a dummy taking opposite values. Results show that coefficients on the variable of interest, which split between former colonies and other developing countries, are as in previous regressions; also, they do not differ between them at a statistically significant level. Results (not shown) do not change with the dummy Colonies included among regressors. Balli and Sørensen (2013) find that the coefficients of interaction terms could be biased in settings where fixed effects are used. 16 The solution they propose is to de-mean the components of the interaction term within the groups for which the fixed effects are included. Hence, I did de-mean Bilateral aid t−1 within each origin-destination dyad and year, as well as pc GDP origin within each origin country and year (column 7). Column 8 reports coefficients when the OLS regression is run after excluding outliers. Results are as in previous regressions.
(3)  Robust standard errors clustered by country-pairs in parentheses. Corrected standard covariance matrix, robust to panel-specific autocorrelation and heteroskedasticity. The sample comprises data from 1993 to 2013. Variables are in logs, except Asylum Policy Index (API). Sys-GMM, in levels and differences: LDV, Bilateral aid, refugees and immigrants are treated as potentially endogenous; all other control variables are treated as predetermined. AR (2) and AR (3) are second-and third-step Arellano-Bond tests for serial correlation. AR (1), not reported, = 0. Column 1 logs on Bilateral asylum, not + 1. Column 2 also includes country × time effects. Constant and covariates not reported *p < 0.1; **p < 0.05; ***p < 0.01 Table 8 Further lags for asylum applications. Poor and less poor countries. Sys-GMM (1) (3) (8)