1 Refuge and Refugee Migration: How Much of a Pull Factor Are Recognition Rates? Joseph-Simon Görlach and Nicolas Motz February 15, 2017 Abstract...

Author:
Amanda King

0 downloads 0 Views 701KB Size

Abstract In many cases, refugees initially migrate to safe low- or middle-income countries close-by before possibly moving on to a desired higher income destination. A key policy parameter determining this choice is the acceptance rate of asylum applications at the preferred final destination. We evaluate the causal effect of changes in this recognition rate on refugee flows, both in a regression framework and by calibrating a dynamic model of Syrian refugees’ location choices. Our estimates suggest an elasticity of about 0.25. Additional insights are provided by the dynamic model: Even though recognition rates are exogenous from a refugee’s point of view, they are a choice for each destination country which may affect refugee flows not only into its own territory, but also into other potential destinations. This leads to a strategic interdependence among countries that we explicitly include in our model. The calibration results indicate that a more generous policy in one European destination generally leads to a higher number of arrivals in other European countries, but the strength of this effect is heterogeneous across regions. Accordingly, changes in recognition rates in one region can lead to policy responses elsewhere. In particular, a small increase in the acceptance rate in Northern Europe can lead to a large decrease in the acceptance rate in South-Eastern Europe.

Keywords: Migration, Refugees, Asylum Policy. JEL Classification: F22, J15, K37, O15.

∗

Bocconi University, Department of Economics, BIDSA and IGIER; e-mail: [email protected] † Universidad Carlos III de Madrid, Department of Economics; e-mail: [email protected]

1

Introduction

About one quarter of the Syrian population now lives as international refugees outside the Syrian borders. Like in many other contexts of forced displacement, most of these refugees reside in a number of neighbouring countries, while some have moved on to other destinations, primarily to member countries of the European Union.1 EU countries have varied considerably in their openness to refugee, with some member states accusing others of exacerbating refugee flows into the EU by too generous acceptance policies. In this paper we evaluate the effect asylum recognition rates have on refugee numbers, contrasting two very different methodological approaches. First, we evaluate the effect based on variation across a large number of countries, instrumenting recognition rates for a given origin-destination pair with past acceptance rates for other origins. In our second approach, which is new to the literature on refugee migration, we account for a potential strategic element in destination countries’ policy choices. To this end, we formulate a dynamic life cycle model of refugee migration from Syria to a close-by set of countries and of a potential onward migration to different EU destinations. We interpret observed asylum policies as an equilibrium outcome, which allows the model to be used for predictions of countries’ best responses in setting their optimal policies. Calibrating the model to match observed numbers of Syrian refugees provides insights into the effects of policy choices on refugee flows in general, and into the nature of externalities among EU countries in particular. We view these two approaches as complementing each other. Importantly, the more structural approach is able to address a potential flaw of the simpler evaluation based on variation across countries of asylum: The geographic proximity of different European countries implies that the acceptance rate in any destination affects applications in other countries of asylum. For instance, knowing that recognition rates are high in Sweden will affect the number of refugees in Denmark. The direction of this effect is a priori unclear. On the one hand refugee flows would be diverted from Denmark to Sweden. On the other hand, the overall number of refugees arriving in Europe may be higher, with a positive effect also on the number of refugees settling in or passing through Denmark. This generates an interdependence across observations (countries), which biases the effect of a change in the recognition rate on applications in a given country in an unknown direction. The structural approach allows us to explicitly model this interdependence. Another advantage of the more structural approach is that we do not have to utilise information from multiple countries of origins, across which there would indeed be considerable variation in the fraction of refugees that tends to be accepted. The particularities 1

In addition, about seven million Syrians have been displaced internally. As this paper is concerned with asylum recognition rates, however, we focus on international refugees.

2

of different conflicts and causes of asylum leave it unclear to what extent observed differences across countries of origin should be used at all to predict the effect on Syrian refugee migration, in which we are interested. Beside the challenge of controlling for various contextual differences across origins, the particularly high recognition rate for Syrian refugees requires an extrapolation to the far end of the distribution of recognition rates. Figure 1 shows the distribution of overall acceptance rates for refugees from major countries of origin, and illustrates the high rate at which asylum applications by Syrians are accepted, which is only shared by a few other major refugee populations like Ukrainians (almost all of whom were accepted in Russia) and refugees from the Central African Republic. We thus calibrate the model to a number of aggregate moments from UNHCR and Eurostat, using information on Syrian refugees only.

Figure 1: Recognition rate of asylum applications across countries of origin from which at least 10,000 applications have been processed (worldwide) in 2014. Source: UNHCR, “Statistical Yearbook 2014”. Our estimates imply a low elasticity of the number of arrivals of refugees in a given destination with respect to the recognition rate. Increasing the number of applications that is accepted during a given year by one percent increases the number of arrivals by about 0.25 percent. As the structural model shows, increases in the recognition rate of Syrian refugees in one European country also increase the number of arrivals of Syrians in other European countries. The strength of this effect is, however, asymmetric, as increased recognition rates in Northern Europe lead to a larger increase in the number of refugees arriving in South-Eastern Europe than vice versa. This externality leads to a strong policy response in South-Eastern European countries, where recognition rates are accordingly lowered. 3

This paper contributes to the political science and economic literature on refugee migrations, where a number of studies have examined the relation between asylum recognition rates and the number of applications, see e.g. the papers by Holzer et al. (2000), Vink and Meijerink (2003), Neumayer (2004), and Hatton (2016). While many of these studies provide a detailed descriptive account, Hatton (2009) explicitly attempts to address the endogeneity of a destination country’s recognition rates. Using UNHCR on a number of major origin and destination countries, he finds that a one percentage point increase in the recognition rate raises the number of asylum applications by 1.2-1.6 percent. Despite the different context (both in terms of countries and time period considered), the two separate approaches we use in this paper yield estimates of very similar magnitude. In a different strand of research, a number of papers have highlighted the interdependency between different host countries’ migration and asylum policies. A paper that has received attention beyond academic circles is Fern´andez-Huertas Moraga and Rapoport (2014), who treat the acceptance of refugees as a public good, and propose a system of tradable immigration quotas that matches international migrants to host countries while accounting for both migrants’ and countries’ preferences. Hatton and Williamson (2006), Facchini et al. (2006), Fern´andez-Huertas Moraga and Rapoport (2015) and Hatton (2015) more closely investigate the benefits of policy coordination for the particular case of refugee reallocation within the European Union. In analyzing individual location choices within a dynamic life-cycle framework, we also contribute to the growing literature that uses behavioural models to examine various aspects of internal and international migration.2 We extend this approach to model refugee movements across several countries. For recent overviews of economic aspects of refugee migrations more generally, see e.g. Ruiz and Vargas-Silva (2013), Chin and Cortes (2015) and Dustmann et al. (forthcoming). The organisation of the paper follows from the two separate frameworks we employ, which also rely on different data sources: Section 2 presents the instrumental variable approach while the behavioural model and its calibration are described in 3. Section 4 concludes. 2

Most dynamic migration models consider the choice between two locations. Exceptions include the models by Kennan and Walker (2011), who consider internal migration between multiple locations within the United States, and Girsberger (2015), who models both internal rural-urban and international migration from Burkina Faso. See Dustmann and G¨orlach (2016) for an overview of this literature.

4

2

Regression Framework Based on Cross-country Variation

Like most economic questions, the one we aim to address—what is the effect of asylum policies on refugee arrivals—can methodologically be approached in different ways. Arguably the most intuitive one is to examine how the variation in asylum applications across origin and/or destination countries is related to the fraction of applicants who are granted asylum. UNHCR provides bilateral information on the number of asylum applications and the share of positive decisions on an annual basis. In the following sections we use this information to describe some patterns in the data (Section 2.1) and to estimate the causal effect of recognition rates on asylum applications using an instrumental variable approach within a linear model (Section 2.2). Section 2.3 points at some weaknesses of this framework, which motivate the more structural approach presented in Section 3.1.

2.1

Data and Descriptives

Our main source of information in this section are UNHCR’s Statistical Yearbooks for the years 2011 to 2014, in line with the time frame used in the second part of the paper in which we focus on Syrian refugees. The yearbooks provide bilateral information on the number of pending applications at the beginning of each year, new applications during the year, as well as the numbers of applicants who are granted protection under the Geneva Convention. Table 1 lists the means and standard deviations of these variables. Variable Mean Pending applications at beginning of year 779.1 New applications 1269.5 Recognition rate 15.7% Applications/million of origin population 85.4 Number of origins 122 Number of destinations 119 Observation points 2962

Std. Dev. 2573.7 5716.5 0.194 313.1

Table 1: Summary statistics. Source: UNHCR Statistical Yearbooks 2011-2014. Applications from a total of 122 origins and in 119 destination countries are recorded over the years. The unit of observation in the estimation below is an origin-destinationyear combination, of which we observe a total of about 3,000. Conditional on there being any applications, each destination receives an average of 1,269 new applications per year from each origin, while 779 pending applications are carried over from the previous year. 5

The recognition rate, that is the fraction of pending and new applications that is accepted during a given year, is 15.7 percent on average, and in our data almost constant over the years (ranging from 15.6 to 15.8 percent). It is important to note that these bilateral data only include refugees who formally apply for asylum under the Geneva Convention. This is not the case for the vast majority of Syrian refugees, for instance, who enjoy group protection in the main receiving neighbouring countries, but are not granted asylum under the Geneva Convention and will likely be forced to return once the conflict in their home country is over. When we focus on Syrians in the main part of this paper, however, we do consider all Syrian refugees in neighbouring countries. To compute the number of asylum seekers relative to the size of the origin population, we augment our dataset with population sizes provided by the World Bank (2016). The mean number of applications from a given origin is 1098.8 out of every million individuals, conditional on there being any. This breaks down to an average of 85 out of a million per origin-destination-year observation. The means in Table 1 mask considerable heterogeneity across different refugee populations. In the left panel of Figure 2 we thus plot the number of applicants as a share of the origin population (on a log-scale) against the recognition rate for all country pairs with positive application numbers. Observations referring to refugees from Syria are highlighted in red. To get a better sense for how Syrians compare to asylum seekers from other origins, the right panel of Figure 2 selects the five origin countries for which our data contains the largest number of observation points. For each of these countries—Afghanistan, Iraq, Pakistan, Somalia and Syria—we observe a similar number of entries over the four year period, ranging from 112 for Pakistan to 133 for Syria. Relative to its population size, Pakistan generates the lowest number of refugees among these five countries, and asylum seekers from Pakistan have the lowest chance of being accepted (8.6 percent on average). Relative refugee numbers are higher for Afghanistan and Iraq, whose citizens are also more likely to be accepted for asylum across a number of destinations. Nevertheless, their chance of being accepted is rarely above 50 percent. The number of destination countries where this is the case is markedly higher for asylum seekers from Somalia and Syria. In Panel (b) of the figure, we add a fitted line for each of the five countries. These patterns are not informative about a causal relation between recognition rates and asylum applications. We do, however, interpret them as evidence for important differences across contexts in which refugee migrations happen. For two of these countries (Iraq and Somalia), the correlation is in fact slightly negative, while it is largest for refugees from Syria. This is one motive for the more structural approach taken below, which allows us to use data on Syrian refugees only.

6

Figure 2: Applicants per origin population and recognition rates for (a) all observations, with refugees from Syria highlighted in red, and (b) for additional selected origin countries. Source: UNHCR Statistical Yearbooks 2011-2014. We supplement the UNHCR data with bilateral geographic information from Mayer and Zignago (2011), as well as with bilateral migrant stocks. This latter information is available only for OECD destinations3 , which reduces the sample size in our main regression.

2.2

Estimation

In this section we use the bilateral UNHCR data for a regression analysis to determine the causal effect of asylum recognition rates on asylum applications. This approach has a number of shortcomings that were already mentioned above and will be discussed in more detail below. We view this section as complementary to the very different, and more structural, approach presented in Section 3.1. Specifically, we postulate a relationship of the form log (applicationsodt ) = αP rob[recognition]odt + Xodt β + γot + γdt + uodt ,

(1)

where o, d, and t are indexes for origin, destination and year, respectively, Xodt contains a number of bilateral control variables and u is an error term. A major concern is that the direction of causality is unclear as asylum policies may respond to the number of applicants. We therefore use past acceptance rates for refugees from other origin countries as an instrument. Specifically, to instrument the recognition rate for applicants from origin o in destination d in year t we use the average recognition rate in d for refugees from countries other than o (denoted o) before the start of our panel 3

The stock of foreign nationals was retrieved from http://stats.oecd.org on 08.12.2016.

7

in 2010, i.e. IV ≡ P rob[recognition]od2010 . Zodt

The identifying assumption is thus that the number of asylum applications from a given origin o in the period 2011-2014 is affected by the previous acceptance rates for refugees from other origins o only through the general openness of a destination, which also determines its acceptance of applicants from origin o. Table 2 shows the estimation results. Column (1) displays the correlation from a simple OLS regression for the full UNHCR dataset as a benchmark. Column (2) shows that the estimated coefficient barely changes if we run the same regression on the sample for which the full set of variables included in the IV estimation is available. As explained above, these observations are restricted to OECD destinations, which explains the strong reduction in sample size. Columns (3) and (4) show the IV estimates for different sets of controls. The bilateral controls we include throughout are a bilateral indicator for whether two countries have a common border, the bilateral distance between countries, an indicator for whether two countries share an official language, and the log stock of country o nationals in d recorded by the end of the previous year. The estimation in column (3) further control for origin, destination and year of observation. As illustrated in Figure 2, origin countries generally differ in how many of their citizens apply for asylum abroad and how high their chances of being accepted are. Controlling for origins eliminates this between-country variation. In contrast, column (4) includes a full set of origin-year and destination-year indicators. This allows us to control for important timevarying determinants of the number of refugees stemming from a given origin or arriving at a certain destination, such as the political situation at the origin as well as an origin country’s population size (insofar as this changes over time).4 Our estimates imply that a one percentage point increase in the probability of being accepted for asylum raises the share of an origin country’s population arriving as refugees by about 1.8 percent. Given that the average recognition rate is 15.7 percent, this corresponds to an elasticity of 0.29. In what follows we discuss potential sources of bias which this estimate may be subject to, and which the structural model may be better able to address. 4

To account for the concern that recognition rates are only observed for country pairs with non-zero applicant numbers, we also applied a selection correction. Estimates however barely changed.

8

P rob[recognition]odt bilateral controls o, d, t indicators o × t & d × t ind. 1st stage F -stat N

Dependent variable: log (applicationsodt ) (1) (2) (3) (4) OLS OLS IV IV 0.707 0.714 1.583 1.829 (0.126)a (0.232)b (0.745)c (0.718)c X X X X 55.876 43.811 2958 824 824 824

Table 2: Estimates of equation (1). Data: CEPII, OECD and UNHCR Statistical Yearbooks 2011-2014. Bilateral controls include an indicator for whether two countries are neighbouring, the bilateral distance between countries, an indicator for whether two countries share an official language, as well as as the log stock of origin nationals recorded in a destination at the end of the previous year. a: p < 0.001, b: p < 0.01, c: p < 0.05.

2.3

Critique of the Regression Results

The question we wish to address is to what extent the relatively generous asylum recognition rates by a number of EU countries have served as a pull factor inducing Syrian refugees to move to Europe. As noted earlier, we see a number of problems with a regression analysis of the above type. First, the need for a large sample size for this kind of empirical evaluation requires that we use data from many origin and many destination countries. This necessarily implies that we draw conclusions about the behaviour of Syrian refugees from that of refugees from Eritrea, Pakistan, Ukraine and many other countries. Similar sources of variation have been used extensively in a variety of areas of empirical economic research, and we agree that it is a valid tool in many cases to gain generalizable insights into economic phenomena. In our particular context of international refugees, who flee due to very different political and social circumstances and seek asylum in very different kinds of destination countries, however, we would like to raise a note of caution. Second, and more important from a technical point of view, consistency of the estimates obtained from our regression analysis requires that observations are independent. This is violated if one destination’s asylum policy affects not only the number of refugees applying there, but also the number seeking asylum in other destinations. This very likely is the case, especially for destinations as geographically close as member countries of the European Union. Thus, a high recognition rate in Austria, Germany or Sweden will also affect the number of refugees arriving in Greece, Hungary or Italy. This interrelatedness

9

further provides scope for strategic interactions between different destinations, which is the focus of the second part of this paper.

3

Behavioural Framework

In the following sections we formulate a dynamic model of refugees’ location choices, where one destination’s asylum acceptance rate may divert or enhance refugee flows. We calibrate this model to data on Syrian refugees only, avoiding the uncomfortable need to draw conclusions based on refugees from very different contexts. Given that representative micro-level data on Syrian refugees in Europe are not available so far, the model—although tailored as closely as possible to the context we consider—is a fairly stylized one, which of course comes with its very own limitations. We thus view the two different approaches in this paper as valuable complements. The model itself is presented in Section 3.1. The data used in this part of the paper is described in Section 3.2, while the calibration is carried out in Section 3.3. Section 3.4 performs a number of counterfactual simulations.

3.1

A Model of Refugee Migration

The model has two layers with separate sets of decision makers. The “inner” part models individual refugees moving across locations. These moving decisions are determined by stochastic shocks, different flow utilities received in different locations and by whether a refugee is accepted for asylum. We assume that acceptance probabilities are known and taken as given by the agents. The “outer” part of the model acknowledges that these acceptance rates are a choice of destination countries, which affects refugee flows and may trigger adjustments in other destinations’ policies. We describe these two parts of the model in turn. 3.1.1

The Refugee’s Location Problem

In the model, individuals can choose between their home country, a neighbouring country that accepts all arriving refugees, and two potentially more attractive destinations, D1 and D2 , where entry and asylum are more tightly restricted. According to the Dublin agreement, refugees can only apply for asylum in the first EU country they enter, but in practice not all countries enforce this rule. In line with this we assume that while refugees may pass through one of these locations (D1 ) and still apply for asylum in the other destination (D2 ), this is not the case for migration in the opposite direction. For instance, a refugee can pass through Hungary and still be granted asylum in Austria, 10

while the Hungarian authorities would deny asylum to refugees coming from Austria. In addition, leaving either destination country while an asylum application is being processed means that the individual in questions will not be able to apply for asylum again in this country in the future. At the beginning of each period t an individual i decides on a location lit with the choice set depending on his or her current location. We label these locations as source country S, transition country T , and destinations D1 and D2 . Choices further depend on the agent’s current age, ageit , and when in either D1 or D2 on refugee status rit−1 ∈ {n(ot accepted), a(ccepted)} in the previous period. In addition, given that destination D1 is assumed to enforce the Dublin agreement vis-`a-vis refugees coming from D2 , and D2 rejects all applicants who have temporarily left or have been accepted for asylum in D1 , we keep track of the option that asylum may still be granted in D1 and D2 respectively. We denote the corresponding two-dimensional state variable by oit ∈ {(1, 1); (0, 1); (0; 0)}, where the first (second) element of oit equals 1 if the person can still be accepted in D1 (D2 ) and equals zero otherwise. The vector Ωit = (ageit , lit−1 , rit−1 , oit ) collects these state variables. In what follows we suppress the dependence of all choices on this state vector and use the easier notation Vl,ro ≡ V (ageit , lit−1 = l, rit−1 = r, oit = o) for an individual’s value given the respective states. For instance, VD2 ,n(0,1) denotes the value for a refugee who has arrived in D2 and has not yet been accepted for asylum, but has the option of being accepted in D2 , while he or she will not be accepted in D1 anymore. In many contexts of armed conflict, most displaced individuals only have the immediate option to escape to a nearby country. We thus assume that when in the source country, an individual only has the choice between staying or moving to a neighbouring country T , so that the individual’s maximised value is VeS,no = max{VS,no , VT,no }. When in country T , the individual has the choice between staying there for at least one more period, returning to S or migrating on to one of two more attractive destinations, though the individual faces the risk of being barred from entry and having to return to T with probability pb . If the migration has been successful, the individual arrives as a not yet accepted asylum seeker, so that for a refugee in T , VeT,no = max{VS , VT,no , pb VT,no + (1 − pb )VD1 ,no , pb VT,no + (1 − pb )VD2 ,no }. Once at destination Dd , d ∈ {1, 2}, in each period the individual will be accepted for asylum with probability pad if he or she (1) has not been accepted for asylum previously, (2) has never left destination Dd , and (3) if in destination D1 , has never been to D2 . 11

Regardless of legal status, the individual may each period decide to either stay for at least on more period, to move to the respectively other destination country or to return to T . The maximised continuation values in Dd are thus VeD1 ,n(1,1) = max{VT,n(0,0) , pa1 VD1 ,a + (1 − pa1 )VD1 ,n(1,1) , VD2 ,n(0,1) } if not yet, but with option of being accepted in D1 or D2 VeD2 ,n(0,1) = max{VT,n(0,0) , VD1 ,n(0,0) , pa2 VD2 ,a + (1 − pa2 )VD2 ,n(0,1) } if not yet, but with option of being accepted in D2 VeDd ,n(0,0) = max{VT,n(0,0) , VDd ,n(0,0) , VDd− ,n(0,0) } if not yet and no option of being accepted in either D1 or D2 VeDd ,a = max{VT,n(0,0) , VDd ,a , VDd− ,n(0,0) }

if accepted for asylum,

where Dd− denotes the respectively other destination. Let vlr denote the location and legal status-specific utility flow in lr ∈ {S, T, D1,u , D1,a , D2,u , D2,a }, and assume that individuals face transitory shocks εlitr to their location preference.5 Then, given state vector Ωit , an individual’s welfare is determined by current and expected future payoffs, which we assume are discounted at a rate β = 0.95, so that V (Ωit ) = vlr + εlitr + βE[Ve (Ωit )]. Our policy parameter of interest in this setting is the acceptance rate of asylum applications, pad , which will affect both the fraction deciding to move on from the transition country, as well as the distribution of refugees across destinations D1 and D2 . 3.1.2

The Game between Destinations

Besides the group of neighbouring countries T , which accept refugees from S unconditionally, we consider two further destinations D1 and D2 , which choose to accept only a fraction of pending asylum applications in any given period. While acceptance rates are taken as given from a refugee’s perspective, they are set endogenously by these countries in order to achieve a target number of refugees residing in the country. Governments in D1 and D2 anticipate refugee flows as a function of the acceptance rates they and the respectively other destination set, and strategically decide on their acceptance rate. Let In the simulated model, litr will be assumed to be standardized extreme value distributed with variance one to normalize the scale of utility flows. Additivity of εlitr in the utility function, independence and extreme value distribution imply that the location choice probabilities take a logistic form, with value functions in the respective location-legal status specific states as arguments (Rust, 1987). 5

12

destination d’s payoff be Γd (Nd ) = |sd (pa1 , pa2 ) − ρd |, where sd is the actual share of the Syrian population to arrive in destination d while ρd is the targeted fraction of the Syrian population. We consider Nash equilibria in pure strategies of the game among destinations. One feature of this game is that the number of refugees arriving in a destination is ceteris paribus monotone increasing in the rate at which this country accepts asylum applications. As a consequence, the optimal recognition rate of a country is always unique: either a country achieves its optimum at some interior recognition rate, or it chooses a recognition rate of zero (one) if the number of arriving refugees is too high (low). An observation that follows is that at any interior equilibrium of the game each destination achieves the optimal number of refugees. It also follows that the existence of at least one equilibrium in pure strategies is guaranteed,6 albeit not the existence of an interior one. All other characteristics of the game depend on the parameters of the model of location choices by refugees. Acceptance rates may, for example, be strategic complements if a more generous policy in one country lowers the number of arrivals in the other destination and therefore results in a higher acceptance rate there. It is, however, equally possible that policy choices are strategic substitutes, or in fact that the externalities among destinations have opposite signs. The reason for this is that the recognition rate set by one destination has a number of competing effects on the number of arrivals in the other destination. The most direct effect is that an increase in the recognition rate set by destination d will make some individuals move to destination d who would have otherwise moved to destination −d. This reduces the number of arrivals in −d. However, a more generous asylum policy in destination D2 also increases the option value of residing in D1 , as destination D2 still grants asylum to individuals who have previously been to destination D1 . As a consequence, the number of people deciding to leave transition country T will increase and this may increase the number of arrivals in destination D1 .

3.2

Data

Representative micro-data on Syrian refugee migration are still scarce. We thus identify the structural parameters of the model by calibration to the fraction of the Syrian population registered by the UNHCR in Iraq, Jordan, Lebanon and Turkey, the fraction that has applied for asylum in European Union countries, their acceptance rates, as 6

Best responses are continuous functions mapping from [0, 1] into [0, 1] and Brouwer’s fixed-point theorem accordingly implies existence.

13

well as the rate at which Syrian individuals not accepted for asylum leave. Our focus on international refugee migration implies that we abstract from internal displacement and subsume the about seven million internally displaced persons within Syria under the fraction of the population that has not moved to either T , D1 or D2 . We use data from 2014, when refugee migration into Europe picked up, but before a number of countries introduced border controls to neighbouring EU member states.7 We divide EU countries into two groups, based on their geographic location and the corresponding exposure to Syrian refugees passing through their territory these countries have. Figure 3 shows this categorization for the 30 European countries we consider.8 D1 includes Bulgaria, Croatia, Cyprus, Greece, Hungary, Italy, Malta, Romania and Slovenia while D2 includes Austria, Belgium, Czech Republic, Denmark, Estonia, Finland, France, Germany, Ireland, Latvia, Lituania, Luxembourg, The Netherlands, Norway, Poland, Portugal, Slovakia, Spain, Sweden, Switzerland and the United Kingdom. By end of 2014, the UNHCR counted the number of registered Syrian refugees in the major neighbouring host countries Iraq, Jordan, Lebanon and Turkey at NT = 3, 718, 001, while the number of Syrian asylum applications in the D1 group of European countries between April 2011 and December 2014 stood at ND1 = 16, 005 and that in the D2 countries at ND2 = 108, 938.9 We use the ratio of these numbers to the Syrian population (NS = 21, 070, 917)10 before the outbreak of the Syrian civil war in 2011 to identify the relative attractiveness of these destinations. To separately identify the flow utility from being in these locations as an accepted asylum seeker relative to a not accepted one we further use information from EuroStat11 on the out-migration rate of Syrian individuals who have not been granted a permit to stay. This rate is computed as the number of voluntary returns to third (i.e. non-EU) countries from D1 and D2 (1,920 and 475) in 2014 relative to the number of Syrian nationals ordered to leave the two groups of destinations (39,060 and 5,630) during the same period. Table 3 lists the full set of moments used. Based on the number of detections of attempted illegal border-crossings by Syrians from quarterly reports published by the EU’s external border agency Frontex (66,698 in 2014) and the number of registered asylum applications, we assume the probability of being detected and barred from entering the EU to equal 34.8 percent. Finally, the shares of pending and new applications by Syrians in D1 and D2 recorded by UNHCR that were 7

Besides the successive introduction of border controls and the construction of physical barriers in 2015, a number of major events affected Syrian refugee migration that are beyond the scope of our model. The most important ones are the food supply crisis in refugee camps in Jordan and the beginning of Russian military intervention in Syria, both of which contributed to the surge in refugee numbers arriving in Europe in 2015. 8 We include Norway and Switzerland throughout in our analysis. 9 http://data.unhcr.org/syrianrefugees/regional.php, accessed on 01.03.2016 10 http://data.worldbank.org/country/syrian-arab-republic, accessed 01.03.2016. 11 http://ec.europa.eu/eurostat/data/database, accessed 01.03.2016.

14

Destination 1 Destination 2 Source Country Transit Country FIN

NOR SWE

GBR

POL DEU

FRA

ROU ITA BGR

TUR

ESP

GRC

SYR

IRQ

Figure 3: European destination countries: D1 includes Bulgaria, Croatia, Cyprus, Greece, Hungary, Italy, Malta, Romania and Slovenia; D2 includes Austria, Belgium, Czech Republic, Denmark, Estonia, Finland, France, Germany, Ireland, Latvia, Lituania, Luxembourg, The Netherlands, Norway, Poland, Portugal, Slovakia, Spain, Sweden, Switzerland and the United Kingdom. NT /NS ND1 /NS ND2 /NS P [return|l = D1 , r = n] P [return|l = D2 , r = n]

0.176 0.001 0.005 0.049 0.084

Table 3: Data used for calibration accepted in 2014 were 44.4 and 49.0 percent respectively.

3.3

Calibration

Besides the policy parameter pad , an important determinant of individuals’ location choice are the unobserved payoffs received. Our aim is to calibrate the vector of these structural parameters θ˜ = (vT , vD1 ,n , vD1 ,a , vD2 ,n , vD2 ,a , ρ1 , ρ2 ) to gain insights into the effects of policy choices on refugee flows. We do so according to the following procedure: Since 15

refugees are assumed to take recognition rates as given, the inner part of the model that describes the choices of refugees can be calibrated conditional on (pa1 , pa2 ). We then use the observation made in Section 3.1.2 that each country achieves the optimal number of refugees in any interior equilibrium of the game among destinations. (Note that only an interior equilibrium can fit the data well as the observed annual acceptance rates are 0.444 and 0.490.) Thus, we can set the optimal shares of the Syrian population of each destination, ρ1 and ρ2 , equal to the predicted shares obtained from the calibrated model at the observed recognition rates.12 The remaining structural parameters we calibrate jointly are therefore summarised by the vector θ = (vT , vD1 ,n , vD1 ,a , vD2 ,n , vD2 ,a ). We use the model to simulate a population of 1 million Syrian-born individuals who after the outbreak of the civil war may move to T and possibly further on to D1 and D2 .13 As Figure 4 illustrates, the number of refugees from Syria prior to 2011 was negligible compared to the current numbers.14 We thus trust that our data on Syrian refugees in the various locations indeed refer to Syrians who were displaced due to the conflict since 2011. Depending on their age, however, the outbreak of the war in 2011 hit individuals at different stages of their life. To account for this, we draw the simulated population from the empirical age distribution of the 2011 Syrian population15 and consider stocks and flows in different locations four years later, corresponding to 2014, the year we observe the aggregate numbers targeted.

Figure 4: Syrian refugees since 1980. Source: UNHCR Population Statistics Reference Database. 12

A calibration of the optimal shares of Syrians residing in each destination to match observed recognition rates is also possible, but would similarly yield optimal shares equal to the predicted shares. 13 Because for the years we consider some choice probabilities are small (such as the probability of applying for asylum in D1 ), the calibration requires a large number of simulations. 14 http://popstats.unhcr.org, accessed 01.03.2016. 15 http://data.worldbank.org/country/syrian-arab-republic, accessed 01.03.2016.

16

To solve the refugee’s location choice problem we find optimal decision functions by backward induction and then simulate refugee flows and stocks across destinations. This enables us to construct counterparts to the moments listed in Table 3 for the simulated population, and search for the value of parameter vector θ that minimises the (weighted squared) distance between these simulated moments and their empirical counterparts, crit = (md − ms (θ))0 W(md − ms (θ)), where md and ms are vectors (NT /NS , ND1 /NS , ND2 /NS , P [return|l = D1 , r = n], P [return|l = D2 , r = n])0 of the targeted empirical and simulated moments, respectively. Note that these are not sample but population moments. In the absence of standard errors for these moments, a convenient choice for the weighting matrix W is a diagonal matrix with the inverse targeted empirical values md on the diagonal. The criterion thus measures the squared deviation between empirical and simulated moments in percentage terms (relative to the empirical magnitude of each element of md ). Figure 5, which plots the (log) criterion against different values of the structural parameters, shows that the criterion obtains a clear local minimum at the given parameter values. In order to show local identification of the model parameters through our set 0 s . Identification of moments more more formally, Table 4 lists the gradient matrix ∂m ∂θ requires that gradient vectors for all parameters are linearly independent. To ensure that 0 s . These are -18.7, -3.3, 0.8, 0.2 and this is the case, we compute the eigenvalues of ∂m ∂θ -8.4. Since we obtain as many different non-zero eigenvalues as there are parameters, our set of of moments identifies the structural parameters under the model.

Parameters vT vD1u vD1a vD2u vD2a

arr T 0.745 0.813 0.421 0.839 1.288

arr D1 0.010 0.323 0.015 -1.570 0.064

Moments ret D1 arr D2 0.074 0.132 0.426 3.424 0.382 0.211 -5.390 -4.590 -0.332 -0.495

Table 4: Gradient matrix

ret D2 1.419 6.595 5.036 2.513 -18.843

∂ms 0 ∂θ

Panel (a) of Table 5 lists the resulting utility flows and refugee targets. Panel (b) further shows that the model closely replicates the five targeted data moments.

17

Figure 5: Local minima of the criterion function with respect to values of the structural parameters. (a) Calibrated Parameter utility gain in T , (vT − vS ) utility gain in D1 when not accepted, (vD1 ,n − vS ) utility gain in D1 when accepted, (vD1 ,a − vS ) utility gain in D2 when not accepted, (vD2 ,n − vS ) utility gain in D2 when accepted, (vD2 ,a − vS ) (b) Targeted Value arrivals in T /NS arrivals in D1 /NS arrivals in D2 /NS P [return|l = D1 , r = n] P [return|l = D2 , r = n]

Value -4.214 -10.615 0.373 -8.944 0.520 Data 0.176 0.001 0.005 0.049 0.084

Model 0.204 0.001 0.004 0.049 0.070

Table 5: Calibrated values

3.4

Counterfactual Simulations

We now use the calibrated model to examine the effects of a change in the acceptance rate of asylum applications. Even though adherence to the Geneva Refugee convention of 1951 limits the degree to which the acceptance rate is a policy parameter, decisions such as the compilation of safe origin country lists or a relaxation of the EU’s Dublin Agreement are political choices. In particular, countries can deter refugees by a slow processing of requests for asylum. The longer waiting times implied by low annual acceptance rates in

18

our model capture precisely this feature of destination countries’ asylum policies. We thus focus on the effects of a change in a destination country’s acceptance rate as the main policy parameter in our model. If several potential destination countries cannot agree on common asylum recognition and allocation policies, a unilateral change in one country’s recognition rate may trigger an adjustment in another destination’s policies. Besides this adjustment in government choices, we examine the effect on refugees’ location decisions and the overall numbers arriving in different destinations. To do so we simulate refugee migration patterns for a grid of the two destinations’ recognition rates. To understand the mechanism at work, we begin by plotting the reaction function for destinations D1 and D2 in Figure 6. The light solid line R1 (p2 ) is D1 ’s optimal acceptance rate (on the horizontal axis) given the policy of D2 (on the vertical). Similarly, the dark dashed line R2 (p1 ) depicts D2 ’s best responses to D2 ’s acceptance rates. The two curves intersect at the game’s Nash equilibrium. The negative slopes for both destinations suggest that asylum acceptance rates in our setting are strategic substitutes, and somewhat more so from D1 ’s perspective. The reason is clear: for a refugee in location T , a higher acceptance rate in D2 raises the expected value of being there. Since refugees can transit through D1 and still be accepted for asylum in D2 , the expected value of moving to D1 rises as well. This attracts a larger number of refugees to D1 , some of whom will move on, while others may be accepted for asylum and stay. To counteract this rise in arrivals beyond D1 ’s desired level, recognitions are strongly reduced if D2 raises its acceptance rates beyond 55 percent. The same argument applies for the value of moving to T (which is a precondition for getting to either D1 or D2 ) if D1 raises its recognition rates. Since realizations of shocks ε vary over time, some refugees who moved to T may in the end move on to D2 rather than as initially planned to D1 . This latter externality applies to both destinations. Taken together, the reaction function for D2 is less elastic, and a strong response only occurs if D1 sets a recognition rate higher than 70 percent. At the prevailing equilibrium itself, the best response of D2 is very inelastic. It is important to stress that none of this is imposed by our assumptions. Under different parameter values recognition rates can become strategic complements. Figure 7 helps to understand the consequence of this particular pattern of the two destinations’ mutual best response functions on the way one destination’s policy affects application numbers in the respectively other location: The dark dashed lines show the change in the number of refugees arriving in Dd in response to a policy change by the other destination if Dd does not respond. An increase in the acceptance rate of D1 has a modest positive effect (with a semi-elasticity of about 0.06) on the number of arrivals in D2 . In contrast, the effect of an increase in recognitions in D2 on the number of arrivals in D1 is stronger (semi-elasticity of about 0.27). This contributes to the asymmetric

19

Figure 6: Mutually best responses in the game between destinations. feedback effects described above. The figure also shows the extent to which countries can mitigate the externality. The light solid line depicts the effect taking into account optimal responses. Both destination are able to offset any reduction in the other countrie’s acceptance rate through a more generous policy of their own. Increases in the other countrie’s recognition rate, on the other hand, eventually induce each country to lower their own acceptance rate to zero. Once this is the case, further increases in the other countrie’s acceptance rate lead to a sharp increase in the number of arrivals. As the figure shows, this point is reached for destination D1 as soon as D2 sets an annual acceptance rate beyond 56 percent. In contrast, Destination D2 is able to keep the number of arrivals at the desired level even if D1 sets an acceptance rate of 73 percent. Estimates of a destination’s recognition rate on its own refugee arrivals—such as the ones presented in Section 2—are affected by this strategic interaction if a change in destination Dd ’s recognition rate provokes a response by the other destination which in turn affects arrivals in Dd . For modest policy changes this is not the case, as can be seen from Figure 6: a modest change in p1 virtually has no effect on p2 , and hence there is no diversion of refugee flows beyond the direct effect of D1 ’s recognition rate. An increase in p2 , on the other hand, does cause a reduction in p1 . However, there is little feedback effect on application numbers in D2 , corresponding to the country’s inelastic best response function. Accordingly, we find that an omission of the strategic element in an evaluation of a country’s recognition rate on asylum applications is not actually problematic in this particular case. Figure 8 is the counterpart to the results obtained from the linear IV estimation in 20

Figure 7: Externalities of recognition rates by D1 and D2 on the respectively other destination. Section 2.2. It graphically shows the effect of a change in the annual acceptance rate set by a destination on the number of arrivals there. As the figure shows, an increase in the recognition rate of a country always increases the number of arrivals. For small changes, the semi-elasticities of the effect for D1 and D2 are approximately 0.5 and 0.6, which is less than half as much as the estimate obtained in Section 2.2. Note, however, that recognition rates for Syrian refugees are about three times as high as the average for refugees from the large set of origins considered earlier. Hence, the elasticities of 0.22 and 0.29 for Syrians predicted by the structural model are in fact very similar to the estimate obtained for the larger set of origin countries considered in Section 2.2. What is masked by the linear model, however, is that even a drastic reduction of destination countries’ recognition rates is predicted to have very little effect on the number of arriving refugees (less than a 3 percent reduction for D1 and about 7 percent reduction for D2 , even if recognition rates are reduced to zero).

4

Conclusion

The importance of asylum policies as a pull factor for refugees is fiercely debated in the context of the Syrian refugee crisis and the large numbers of asylum seekers who have arrived in Europe during the last few years. In this paper, we attempt to quantify the effect between asylum recognition rates and application numbers. Given the scarcity of available data and the multitude of mechanisms at work, this is not trivial, and different empirical methods each come with their drawbacks. For this reason, we approach the 21

Figure 8: Effect of a change in recognition rates on the arrivals of refugees. question from two very different angles: an instrumental variable estimation based on a large number of bilateral asylum application and recognition rates, and the calibration of a dynamic behavioural model of location choices that allows us to focus more directly on Syrian refugee migration to Europe, and which at the same time can account for a potential strategic interaction between different host countries. Based on the calibrated model, we conclude that an omission of the strategic element might not be too serious in this context. Accordingly, the elasticities we obtain from the two frameworks are similar, around 0.25. While some researchers may find the instrumental variable regressions more credible, an important advantage of the structural model is that it can be calibrated to data on Syrian refugees only. It thus does not require us to base predictions regarding this particular refugee situation on variation across different countries of origin, each with its very specific political context.

References Chin, Aimee and Kalena E Cortes, “The Refugee/Asylum Seeker,” in Barry R. Chiswick and Paul W. Miller, eds., Handbook of the Economics of International Migration 1A, North Holland, 2015, chapter 12, pp. 585–658. Dustmann, Christian and Joseph-Simon G¨ orlach, “The Economics of Temporary Migrations,” Journal of Economic Literature, 2016, 54 (1), 98–136. , Francesco Fasani, Tommaso Frattini, Luigi Minale, and Uta Sch¨ onberg, “On the Economics and Politics of Refugee Migration,” Economic Policy, forthcoming. 22

Facchini, Giovanni, Oliver Lorz, and Gerald Willmann, “Asylum seekers in Europe: the warm glow of a hot potato,” Journal of Population Economics, 2006, 19, 411–430. Girsberger, Esther Mirjam, “Migration, Education and Work Opportunities,” Unpublished Manuscript, 2015. Hatton, Timothy J., “The Rise and Fall of Asylum: What Happened and Why?,” The Economic Journal, 2009, 119 (February), F183–F213. , “Asylum Policy in the EU: the Case for Deeper Integration,” CESifo Economic Studies, 2015, 61 (3), 605–637. , “60 Million Refugees: Refugees, Asylum Seekers, and Policy in OECD Countries,” American Economic Review: Papers & Proceedings, 2016, 106 (5), 441–445. and Jeffrey G. Williamson, “Refugees, Asylum Seekers, and Policy in Europe,” in Rolf J. Langhammer and Federico Foders, eds., Labor Mobility and the World Economy, Springer, 2006, pp. 249–284. Holzer, Thomas, Gerald Schneider, and Thomas Widmer, “The Impact of Legislative Deterrence Measures on the Number of Asylum Applications in Switzerland (1986-1995),” International Migration Review, 2000, 34 (4), 1182–1216. Kennan, John and James R. Walker, “The Effect of Expected Income on Individual Migration Decisions,” Econometrica, 2011, 79 (1), 211–251. Mayer, Thierry and Soledad Zignago, “Notes on CEPIIs distances measures: The GeoDist database,” Working Papers 2011-25, CEPII 2011. Moraga, Jes´ us Fern´ andez-Huertas and Hillel Rapoport, “Tradable Immigration Quotas,” Journal of Public Economics, 2014, 115, 94–108. and , “Tradable Refugee-admission Quotas and EU Asylum Policy,” CESifo Economic Studies, 2015, 61 (3-4), 638–672. Neumayer, Eric, “Asylum destination choice. What makes some West European countries more attractive than others?,” European Union Politics, 2004, 5 (2), 155–180. Ruiz, Isabel and Carlos Vargas-Silva, “The Economics of Forced Migration,” Journal of Development Studies, 2013, 49 (6), 772–784. Rust, John, “Optimal Replacement of GMC Bus Engines: An Empirical Model of Harold Zurcher,” Econometrica, 1987, 55 (5), 999–1033. 23

United Nations High Commissioner for Refugees, “Statistical Yearbook,” 2010. , “Statistical Yearbook,” 2011. , “Statistical Yearbook,” 2012. , “Statistical Yearbook,” 2013. , “Statistical Yearbook,” 2014. Vink, Maarten and Frits Meijerink, “Asylum Applications and Recognition Rates in EU Member States 1982-2001: A Quantitative Analysis,” Journal of Refugee Studies, 2003, 16 (3), 297–315. World Bank, “World Development Indicators,” 2016.

24

ZAPDOC.TIPS | To ensure the functioning of the site, we use **cookies**. We share information about your activities on the site with our partners and Google partners: social networks and companies engaged in advertising and web analytics. For more information, see the Privacy Policy and Google Privacy & Terms. Your consent to our cookies if you continue to use this website. Accept