In 2015, 1.83 million irregular migrants arrived at Europe's external borders.
The central idea of this research is that illegal immigration emerges out of stringent controls in the parallel legal system. This idea is also contained in work by Auriol and Mesnard, 2016 where the authors consider the eligibility criteria of legal channels as “rationing mechanisms” that generate an illegal counterpart. Using a different methodology based on queuing systems (QS), my paper seeks to formalize this theory and present some empirical substance on it. Despite much journalistic attention, the current migration crisis has not received its due in academic work as yet. Further, the canonical models of migration are not specifically suited for crises of the sort recently witnessed and those that may arise in future due to political instability or climate change.
The QS framework is ideal for scenarios of shortage as it allows for two simultaneous allocation mechanisms – the price and the waiting cost to acquire the good or service (Sattinger, 2002). The theoretical framework in this paper focuses on the latter, such that the waiting cost of migration is proportional to the magnitude of people wanting to migrate (Jasso et al., 2010). This allows Refer to Martin and Midgley (2015) for a detailed description of different immigrant categories in the US.
Due to the lack of data at the individual migrant level to illustrate the theoretical model, I apply a reduced-form empirical strategy using aggregate flows of legal and illegal immigration into the Schengen region. Aggregated annual data on legal migration is available from Eurostat; similarly, the number of Frontex is the European Agency for the Management of Operational Cooperation at the External Borders of the Member States of the European Union, which records the number of detected IBCs via land and sea routes into the Schengen region.
This paper draws from and contributes to three branches of literature: (i) the economics literature on migration – which has focused on its causes and effects, along with optimal immigration policy, (ii) the impact of conflict on migration, and (iii) the literature on the application of queuing theory to economic phenomena.
Firstly, this paper adds to the vast literature on immigration choices. In my reading, I distinguish between studies that deal with the choice to migrate or not, This includes classical work on the expected-wage hypothesis (Harris and Todaro, 1970; Todaro, 1969) and theories of missing markets and relative deprivation (Stark, 1984; Stark and Levhari, 1982; Katz and Stark, 1986). More structuralist theories have been put forward by Massey (1988) and Piore (1980), which emphasize the nature of immigrant-labor markets and capitalist expansion to explain migrants’ flows from “peripheral” to “core” countries. This literature has received empirical traction in Lewer and Van den Berg, 2008 by using a gravity model. An excellent classification of the causes of migration is given by Massey et al. (1993). Recent work has built an evidence base of the trends and magnitudes of international immigration, with attention to its gender and skill dimensions (Docquier et al. [2009, 2007]).
Analyses of the “effects” of immigration on host country populations or optimal policy responses are beyond the scope of this paper, as I focus on immigrant choices. Dustmann et al. (2016) provide an extremely cogent review and critique of the empirical specifications used in recent works. Studies on the “effect” of immigration on host populations are characterized by contradictory empirical findings – some have found negative effects (Borjas, 2003) while others almost no effects (Card, 1990, 2009) or even benefits (Ottaviano and Peri, 2012; Chassamboulli and Peri, 2015). This and an earlier review of the “effect” literature, Friedberg and Hunt, 1995, find more support against negative effects of immigration on host populations. The economics literature on migration has also focused on optimal immigration policy (Benhabib and Jovanovic, 2012) and Meyers (2000) gives a non-technical classification and summary of migration policy theory. A pioneering paper on immigration policy is by Ethier (1986). Also see Djajić and Michael (2014).
With regard to the ongoing refugee crisis, a relevant and contemporary policy background and description of ground realities is given by Dustmann et al. (2017). While I do not study the issue of optimal immigration policy from the host-country's perspective, it is necessary to incorporate the feedback costs of increased traffic and illegal immigration as a result of stringent acceptance rates or waiting-time criteria. I thank an anonymous reviewer for comments related to this.
The second contribution of this paper is to examine the impact of conflict on migration, where the economics literature is as yet relatively sparse. Political scientists have, on the other hand, contributed much to this issue (Davenport et al., 2003; Neumayer, 2005; Bohra-Mishra and Massey, 2011). However, some empirical methods suffer from ‘kitchen-sink’ techniques and, in the case of panel analysis, do not account for the Nickell bias and the general problem of endogeneity.
Finally, this paper aims to add to the growing literature in economics using queuing theory. Queuing theory involves the study of input and service processes along specific disciplines, e.g., first-come-first-serve. This field of study has its origins in the works of Erlang (1909) to study telephone exchanges. In anthropology, the concept of liminality has been used to characterize subjects undergoing transformative and ‘in-between’ experiences (Gennep, 1960; Turner, 1987).
In Section 2 of this paper, I present the theoretical framework of two parallel queues for legal and illegal immigration. The section considers three different immigration policy regimes chosen by the host country. Section 3 and 4 provide a reduced-form empirical strategy to illustrate the model and its results, respectively. The final section concludes with possible extensions and limitations of my work.
I begin by explaining the intuition behind the theory, which can be easily understood without formal proofs. Consider migration as occurring along two parallel channels or queues, one of which is legal and the other is illegal, respectively denoted by subscripts Successful exit from the QS means acceptance for the An arbitrarily fixed rate of acceptance ν ⋲ [0,1] in the legal channel, (i.e., (1 − ν) share of all entrants are not accepted to travel legally). A minimum waiting period
A finite limit
Figure 1
Migration Queuing System (Cases I and II).

In Cases I and II, if migrants are unsuccessful in the legal channel, they can re-enter the legal queue and attempt again or they can enter the parallel illegal QS where acceptance is guaranteed but less-preferred. In Case III, migrants above the quota are diverted automatically to the illegal queue. In each of these stylized cases, migrants are not endowed with personalized attributes. In this respect, a specific QS can refer to a particular category of migrants and there may exist multiple such systems for each type of migrant (e.g., skilled and unskilled). The results in the following cases pertain to how changes in queue parameters influence choices of migrants who share the same observable characteristics. I proceed by using some established results from queuing theory (Medhi [2002]) and perform simple comparative statics analyses where feasible. All three aforementioned cases characterizing stringency in legal migration illustrate the central idea that ceteris paribus, larger flows of illegal migration arise out of greater legal stringency.
The rate of service in the illegal channel may be better characterized as a function of the rate of entry, i.e.,
This is denoted in Kendall's notation as an M/M/1 queue where the M stands for Markovian and I restrict attention to a single or centralized server in each channel for simplicity. The average arrival rate of migrants to the legal queue is the sum of the rate of initial entry ( This assumption does not change the main implications of the model because comparing the expected utility between legal and illegal migration to a third alternative (of not migrating) would not change the preferences between the first two. For further reference, Yue et al., 2009 relax this assumption and present model performance indicators using numerical simulations.
Now, I define
Eq. (2) gives the expected length 〈
This assumption follows from a technical requirement for the first moments of the queue content (Eq. (2)) and waiting time (Eq. (4)) to be positive and to ensure that the queue is stable, i.e., does not explode. When this is violated, we cannot conduct comparative statics deriving from the stability condition, as the queue increases infinitely. Intuitively, the condition imposes that on average, the traffic in the legal queue must be less than the rate at which migrants exit the QS by being accepted or by switching into the illegal queue.
Migrants have a planning horizon [0,T] and incur a cost The benefits (
The optimization problem involves choosing the Migrants only make an additional attempt if they have failed in all previous attempts. Further, the ‘no balking’ assumption implies that
The first and second-order condition are given in Appendix A4 and evaluating the SOC at the optimum ( Some standard results from queuing theory are given in Appendix A1. For a more detailed discussion refer to Medhi (2002).
Next, I use Eq. (6) to examine the comparative statics of how changes in the parameters of the legal queue affect the rate of arrivals (
The proof is intuitive; an increase in traffic
This follows from an increase in the expected benefits from legal migration
The detailed proof in Appendix A5 shows that Result 2 holds as long as the returns from legal entry relative to the cost of waiting is high enough and Figure 2 illustrates this graphically using specific parameter values. All figures illustrating the comparative statics findings use parameter values comparable to Djajić (2014).
Figure 2
Expected utility from legal immigration at different acceptance rates. Parameter values:

The intuition behind this proof is that as migrants become more impatient, they are more likely to switch over to the illegal queue. The higher gains from migrating legally would be worth less as they discount the future more. The formal proof given in Appendix A5 rests on precisely this condition that the absolute loss in expected utility from legal immigration
Figure 3 illustrates that the optimal number of attempts (
Figure 3
Expected utility from legal immigration at different discount rates. Parameter values:

Cases II and III illustrate similar findings as Case I but by using different aspects of QS. The previous case considered immigration policy to be comprised of an arbitrary acceptance rate, which is not always realistic. This case expresses the stringency of legal immigration in terms of a minimum waiting time
The expected return from legal migration after
The limiting cases are of no legal criteria:
This intermediate case of a finite waiting time criterion is perhaps most realistic and general. It places some “reasonable” bounds on the level of diligence required from the migrant. This diligence may be in the form of actual waiting time or even the time spent in acquiring specific skills which form the basis of immigration – such as for science, technology, engineering, and mathematics (STEM) field workers in the US. If a request for refugee status is rejected, there is an appeal process in the EU which would involve another round of waiting.
To model immigration choices under the regime of waiting time criteria, I define a common threshold patience parameter that all migrants are endowed with This is similar to
The simulations (Figures 4 and 5) show that:
The relative magnitude of illegal migration increases as the minimum waiting criteria,
The queue content exhibits cycles, accumulating and reaching a peak before periodically emptying out once enough potential migrants achieve the minimum waiting criteria. Figure 6 illustrates that cycles of immigration flows exhibit such cycles over time. However, this must be interpreted with caution because cyclic behavior in the queue content induced by the waiting time criteria is not possible to distinguish in the data from changes in immigration policy, seasonality and changes in origin-country circumstances.
Figure 4
Vertical axis shows steady state relative shares of legal and illegal migration flows and horizontal axis shows time for different patience parameters and critical waiting times.

Figure 5
Steady state legal queue content and horizontal axis shows time.

This case models the stringency of legal channels of immigration in the form of limited waiting area capacity (Figure 6). It is similar to quotas – e.g., those imposed yearly on residence permits and short term permits in Switzerland.
Figure 6
Countries with the highest illegal border crossings over the 5-year period 2010–2014 according to Frontex data.

The analysis is simplified in this case since migrants do not really have a choice about which route to take. Another matter of interest may be to find the optimal K for the destination country by considering the cost of border control. Since optimal immigration policy itself is not the subject of this paper, I do not attempt it here.
As expected, an increase in the waiting area capacity (
Before moving on to the empirical section, I summarize the implications of the theory below:
Given an increase in the number of people attempting to migrate from an origin country and fixed parameters of the legal QS, there will be a spillover into illegal immigration. Greater stringency in the measures controlling legal immigration, holding the flow of overall potential migrants fixed, would lead to an increase in illegal immigration. Higher levels of impatience among potential migrants would lead to more “switching- over” to the illegal channel.
Due to the lack of and limitations in existing data, a structural estimation of the model is not feasible. However, I attempt a reduced form analysis to check whether the data contradict the theory of inter-dependent queues of legal and illegal immigration. One major limitation of my data is that it only includes individuals who have successfully reached their destination, legally or illegally. I have no information about the actual number of attempts at migration. Secondly, while the theoretical model consists of agents who are homogeneous in their observable characteristics, no such restriction can be applied to the data.
I use lagged and current values of immigrant flows as indicators of rates of arrival and successful queue completion described in the theoretical section, respectively. Origin-country violent conflict plays a dual role in that it acts as a push for more people to try to migrate and also makes people more impatient to leave (i.e., increases the discount factor
Thus, the two main effects I look for in the data are of origin-country shocks of conflict and inter-QS effects. The equations to be estimated are:
According to Massey et al. (1993), over time, migration becomes independent of the individual causes that started the process and becomes its own cause. Decreasing costs and risks, network effects and institutional factors (for example humanitarian and illegal agencies) perpetuate the flow of immigrants once started. Persistence in the dependent variables would demonstrate this.
Available data on the EU refer to migrants who have successfully, either legally or illegally, entered the EU. The data on the inflow of migrants with legal status are from Eurostat, the statistical office of the EU. These data are disaggregated by country of citizenship and are available yearly. The definition of a “legal migrant” varies from country to country. Eurostat defines immigration as the action by which a person establishes his or her usual residence in the territory of a Member State for a period that is, or is expected to be, of at least 12 months, having previously been usually resident in another Member State or a third country. Illegal migration or asylum seekers without legal status are not included in this Eurostat immigration dataset.
Frontex records the number of detected IBCs via land and sea routes into the Schengen region. Disaggregated data by nationality are available from the 3rd quarter of 2007 and are updated monthly. While there is bound to be some measurement error in the Frontex data as many entries into the Schengen region go undetected, the error is unlikely to be specific to any origin countries.
Figure 7 shows the evolution of IBCs from the six countries with the highest flows over 2010–2014 period. Pursuant to eyeballing the panels for Tunisia and Syria, we are presented with a sharp peak for the months immediately following their respective domestic conflicts.
Figure 7
Migration Queue (Case III).

For data on conflict, I use the Georeferenced Event Dataset (GED) from the Uppsala Conflict Data Program (UCDP version 17.1 2017). An event is an incident where armed force was used by an organized actor against another organized actor, or against civilians, resulting in at least one direct death at a specific location and a specific date. The UCDP dataset records three values for the number of conflict-related deaths: a low, high, and a best estimate. All estimates are sourced from publicly available and reliable information and are collected at the level of town and village for urban and rural areas respectively. For this paper, I collapse the data to the country level. The UCDP GED dataset does not yet include deaths from Syria as data collection is not complete at the granular level of towns and villages. I include Syria by using the conflict level data from the UCDP Battle-related Deaths (BRD) dataset. The UCDP BRD dataset is an automatic filtering and aggregation of the UCDP GED and records deaths resulting from an armed conflict at multiple battle locations. I use only the BRDs which occurred solely within Syria. UCDP defines an armed conflict as a contested incompatibility that concerns government and/or territory over which the use of armed force between two parties, of which at least one is the government of a state, has resulted in at least 25 BRDs in 1 calendar year.
My sources for other relevant covariates are the World Development Indicators (WDI) for population density and World Governance Indicators (WGI) for institutional quality. I use an estimate of rule of law from WGI which defines it as capturing perceptions of the extent to which agents have confidence in and abide by the rules of society, and in particular the quality of contract enforcement, property rights, the police, and the courts, as well as the likelihood of crime and violence. The estimate is a normalized indicator ranging from −2.5 to 2.5. The inverse hyperbolic sine transformation transforms variable
Summary statistics
| ||||||
---|---|---|---|---|---|---|
Log legal inflow | Overall | 5.808 | 2.888 | 1,744 | 0.177 | 0.97 |
Between | 2.854 | 218 | ||||
Within | 0.478 | 8 | ||||
Log illegal inflow | Overall | 3.479 | 3.257 | 1,096 | 0.296 | 0.81 |
Between | 3.092 | 137 | ||||
Within | 1.052 | 8 | ||||
Log CD | Overall | 1.115 | 2.466 | 1,768 | 0.309 | 0.68 |
Between | 2.241 | 221 | ||||
Within | 1.039 | 8 | ||||
Log CD difference | Overall | 0.800 | 2.025 | 1,768 | 0.230 | 0.71 |
Between | 1.800 | 221 | ||||
Within | 0.934 | 8 | ||||
Log GNI per capita | Overall | 9.871 | 1.246 | 1,274 | 0.781 | 0.96 |
Between | 1.211 | 188 | ||||
Within | 0.077 | 6.8 | ||||
Log population density | Overall | 5.097 | 1.567 | 1,644 | 0.965 | 0.98 |
Between | 1.568 | 206 | ||||
Within | 0.046 | 8 | ||||
Institutions score | Overall | −0.033 | 0.995 | 1,592 | 0.636 | 0.91 |
Between | 0.990 | 200 | ||||
Within | 0.138 | 8 |
CD, conflict deaths.
The parameters of interest are the Refer to Iraq Body Count studies:
The second criterion of the exclusion restriction is more complicated to support. The difference in the high and low estimates of death may actually be picking up the nature of institutions instead of the intensity of the conflict itself. To account for this, I control for the quality of institutions in the country by using the rule of law indicator. Further, if we consider the group of countries which have no discrepancies in the measurement of deaths in the presence of violent conflict, their mean score on the institutions indicator is negative (−0.61), which is clearly a counter to the argument that the instrument is picking up institutional quality.
In this dynamic panel setting of lagged effects and in the absence of sufficient external instruments, I use a GMM approach to exploit all available moment conditions using internal instruments. Due to the possibility of persistence and the high relative share of variance in fixed effects
Results using the above estimation techniques are presented in the next section.
I focus first only on the effect of conflict and omit the lagged legal and illegal inflows from the right-hand side of the model. Results using pooled Ordinary Least Squares (OLS) and within-estimators are given in Tables 2 and 3. The significance of the positive conflict coefficient remains for IBCs after introducing controls and fixed effects but not for legal inflows. In Column 4 of Tables 2 and 3, I restrict the sample to only those countries with any present or past incidences of conflict within the panel length, to explore further the relationship between the two migration variables and conflict. The significance of the positive conflict coefficient remains for IBCs but not for legal inflows. This illustrates that the relationship between conflict and illegal migrant flows is not driven solely by the differences between countries that experience violent conflict and those that do not.
Pooled OLS and within-estimator dependent variable: Illegal inflow
Log CD | 0.306*** (0.035) | 0.094** (0.040) | 0.082** (0.038) | 0.071* (0.036) |
Log population density | 0.244*** (0.050) | 6.761*** (2.571) | 4.234 (3.394) | |
Institutions | −1.390*** (0.095) | −0.813 (0.579) | −1.419** (0.672) | |
Constant | 0.887*** (0.336) | 2.987*** (0.130) | −29.881** (12.270) | −17.052 (15.764) |
Year fixed effects | Yes | Yes | Yes | Yes |
Country fixed effects | Yes | Yes | Yes | |
Adjusted | 0.273 | 0.130 | 0.171 | 0.246 |
Number of observations | 920 | 959 | 920 | 497 |
Number of countries | 137 | 132 | 71 |
Standard errors, given below coefficients in parentheses, are clustered at the country level.
CD, conflict deaths.
Pooled OLS and within-estimator dependent variable: Legal inflow
Log CD | 0.302*** (0.016) | 0.033 (0.034) | −0.000 (0.011) | −0.009 (0.012) |
Log population density | −0.231*** (0.056) | 0.648 (0.562) | 0.599 (0.685) | |
Institutions | 0.268*** (0.071) | 0.000 (0.196) | −0.450** (0.176) | |
Constant | 6.905*** (0.321) | 5.556*** (0.062) | 2.943 (2.796) | 4.008 (3.306) |
Year fixed effects | Yes | Yes | Yes | Yes |
Country fixed effects | Yes | Yes | Yes | |
Adjusted | 0.101 | 0.039 | 0.054 | 0.142 |
Number of observations | 1,364 | 1,526 | 1,364 | 532 |
Number of countries | 218 | 196 | 76 |
Standard errors, given below coefficients in parentheses, are clustered at the country level.
CD, conflict deaths.
Now, given that there are endogeneity issues, most obviously due to omission of lagged inter-queue variables, the coefficients in Tables 2 and 3 are likely to be inconsistent. As mentioned in section 4.2, I use the difference between the high and low estimates of conflict deaths (CD) as an Instrumental Variables (IV) for its best estimate. This is driven by strong correlation between the level of discrepancy and the level of the conflict itself. Controlling for the quality of institutions, as measured by the rule of law indicator, this discrepancy should not be related to immigration in any other way.
Using the difference between high and low estimates of CD as an IV, Table 4 shows that lagged conflict in origin countries positively affects IBCs but has no effect on the legal inflow of migrants into Schengen countries. The coefficient for conflict is relatively stable across specifications. There is a strong mechanical correlation between the IV and the endogenous conflict variable (refer to Table 4, Part B). This “relevance” of the difference in high and low estimates of CD to its best estimate is because for observations which have no conflict, the difference in estimates is also equal to zero. Thus, I restrict the sample to only those countries which see any present or past incidences of conflict within the panel length (Columns 3 and 6, Table 4) and find that the result still holds. The strength of the instrument does not rely on mechanical relevance alone (Columns 3 and 6, Table 4 Part B).
Two-Stage Least Squares (2SLS) estimator
Log CD | 0.042* (0.022) | −0.000 (0.013) | −0.010 (0.013) | 0.120*** (0.039) | 0.120*** (0.041) | 0.106** (0.041) |
Log population density | 0.648 (0.410) | 0.600 (0.499) | 6.783*** (1.610) | 4.162** (2.111) | ||
Institutions | −0.000 (0.154) | −0.453*** (0.130) | −0.724* (0.371) | −1.288*** (0.458) | ||
Year fixed effects | Yes | Yes | Yes | Yes | Yes | Yes |
Country fixed effects | Yes | Yes | Yes | Yes | Yes | Yes |
Log CD difference | 0.828*** (0.041) | 0.800*** (0.044) | 0.782*** (0.045) | 0.824*** (0.040) | 0.789*** (0.043) | 0.779*** (0.043) |
Log population density | 0.628 (0.796) | 3.522* (1.944) | 1.133 (1.130) | 3.289* (1.995) | ||
Institutions | −0.576*** (0.178) | −1.831*** (0.454) | −1.096*** (0.316) | −1.811*** (0.462) | ||
Number of observations | 1,526 | 1,364 | 532 | 959 | 920 | 497 |
Number of Countries | 218 | 196 | 76 | 137 | 132 | 71 |
Kleibergen-Paap F-stat | 405.913 | 327.501 | 302.436 | 430.805 | 340.704 | 322.067 |
Standard errors, given below coefficients in parentheses, are clustered at the country level.
CD, conflict deaths.
Finally, I turn to the complete model (Eqs (15) and (16)), which includes the inter-queue effects. I use the Stata program by Roodman (2009) for all dynamic panel estimations.
GMM estimates
Log illegal inflow | 0.103 (0.085) | −0.030 (0.030) | 0.752*** (0.047) | 0.018 (0.012) | 0.806*** (0.047) | −0.008 (0.016) |
Log illegal inflow | 0.027 (0.016) | |||||
Log legal inflow | −0.105 (0.152) | −0.140* (0.074) | 0.266*** (0.085) | 0.899*** (0.031) | 0.050 (0.088) | 0.903*** (0.027) |
Log legal inflow | 0.163** (0.081) | |||||
Log CD | 0.010 (0.059) | −0.009 (0.020) | 0.071** (0.034) | 0.013 (0.011) | 0.063** (0.032) | 0.002 (0.013) |
Log CD | 0.015 (0.031) | 0.008 (0.015) | ||||
Year fixed effects | Yes | Yes | Yes | Yes | ||
Number of observations | 804 | 804 | 938 | 938 | 804 | 804 |
Number of countries | 134 | 134 | 134 | 134 | 134 | 134 |
Number of instruments | 46 | 46 | 71 | 71 | 73 | 73 |
AR(1) p-stat | 0.0004 | 0.1682 | 0.0000 | 0.0000 | 0.0000 | 0.0000 |
AR(2) p-stat | 0.1811 | 0.7931 | 0.3345 | 0.0821 | 0.4006 | 0.1477 |
Hansen | 0.0004 | 0.0574 | 0.0194 | 0.0645 | 0.0085 | 0.0601 |
Standard errors, given below the coefficient in parentheses, are clustered by country. Columns 1–4 use the two-step and Columns 5 and 6 use the one-step co-variance matrix estimations.
CD, conflict deaths; GMM, generalized method of moments.
From Table 5, there is evidence of positive spillovers from legal immigration flows on illegal flows but none in the opposite direction. There is also evidence of persistence of own queue effects. The previous results regarding the effect of conflict related deaths still hold, in that they positively affect illegal immigration flows and have no effect on legal flows.
One drawback of the above results is that it does not account for the inherent differences between people attempting to migrate legally compared to those who do so illegally. Documented migration to Europe usually occurs under the following categories – remunerated activities, family reunification, education and humanitarian reasons. Differences within individuals from the same country would make some people more eligible to meet these criteria than others. Since I do not have data on individual level characteristics, I cannot control for them. However, the theory in Section 3 implies that given the same characteristics of individuals, migrating via the legal queue is likely to involve larger waiting costs that lead some people to switch over to the illegal channel. Looking at the effect of violent conflict on the flows of legally resettled people who might share similar characteristics within the origin countries gives evidence to these waiting costs.
“Resettlement is one tool to help displaced persons in need of protection reach Europe safely and legally, and receive protection for as long as necessary. It is a durable solution which includes selection and transfer of refugees from a country where they seek protection to another country” EPRS, 2017. The UNHCR has to determine whether an applicant is a refugee according to the 1951 Geneva Convention, and has to identify resettlement as the most appropriate solution. From Table 6 we see that the relationship between violent conflict and the magnitude of resettled persons is not very robust within the dataset available. Only the third lag of CD is significant using the within estimator and this significance becomes weaker after introducing other covariates. The same goes for results from 2SLS estimation using the difference in estimates of CD as an IV for CD.
Within and 2SLS estimators dependent variable: Log of resettled persons
Log CD | 0.072 (0.048) | |||||
Log CD | 0.072 (0.045) | |||||
Log CD | 0.130*** (0.048) | 0.077* (0.046) | 0.143** (0.066) | 0.068 (0.066) | ||
Log population density | −1.059 (2.129) | −1.085 (2.131) | ||||
Institutions | −0.400 (0.328) | −0.407 (0.332) | ||||
Constant | 0.617*** (0.102) | 0.674*** (0.087) | 0.464*** (0.096) | 5.755 (10.530) | 0.445*** (0.118) | 5.896 (10.538) |
Year fixed effects | Yes | Yes | Yes | Yes | Yes | Yes |
Country fixed effects | Yes | Yes | Yes | Yes | Yes | Yes |
Adjusted | 0.013 | 0.012 | 0.036 | 0.017 | ||
Number of observations | 1,576 | 1,379 | 1,182 | 952 | 1,182 | 952 |
Number of groups | 197 | 197 | 197 | 192 | 197 | 192 |
Standard errors, given below the coefficient in parentheses, are clustered by country.
CD, conflict deaths.
This paper is addressed to two broad audiences. The first audience comprises of development economists who are working on immigration and would be interested in alternative frameworks to study mass migration and crises therein. The second audience includes those economists who work on stochastic processes and specifically, queuing theory, who would be interested in its application to real-world phenomenon such as migration. This paper provides a stylized model focusing on how migration crises emerge out of linkages between legal and illegal channels of entry and the dynamic nature of the same. However, it has several limitations. Firstly, it does not account for migrant heterogeneity and this is one research area with great potential for future work. An emerging sub-field of queuing theory is real-time queuing (RTQ) theory which allows heterogeneity in the form of personalized deadlines to exit the queue. This allows for a dynamic evolution of migrant profiles and price of migration based on the urgency and ability to pay of migrants.
Empirically, this paper attempts to study the effect of conflict on different channels of immigration. To the my knowledge, this has not been addressed as yet in the context of the recent crisis and particularly, has not made use of the Frontex dataset on IBCs into the Schengen region. While I cannot estimate the full structural model and the panel length of the data is only 8 years, the results show clearly the one-way inter-queue effects spilling over from legal to illegal flows and the lack of sensitivity of legal immigration to origin-country shocks such as conflict. It is however quite possible that the short length of the panel does not allow these effects to show up – but that is precisely support for the frictions that the theory assumes.
Although much attention has been given to South–North migration in the media, North–North and South–South migration together account for a majority (59%) of global immigration with South–South migration having the largest share (UNDESA, 2013). Crisis-like scenarios regarding South–North migration emerge due to the rigidity and slow-moving nature of legal processes of resettlement and assessment of asylum applications. A role for future research would be to shed light on immigration policies that account for inter-linkages between different channels instead of addressing individual channels as systems in isolation.
Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Within and 2SLS estimators dependent variable: Log of resettled persons
Log CD |
0.072 (0.048) | |||||
Log CD |
0.072 (0.045) | |||||
Log CD |
0.130 |
0.077 |
0.143 |
0.068 (0.066) | ||
Log population density | −1.059 (2.129) | −1.085 (2.131) | ||||
Institutions | −0.400 (0.328) | −0.407 (0.332) | ||||
Constant | 0.617 |
0.674 |
0.464 |
5.755 (10.530) | 0.445 |
5.896 (10.538) |
Year fixed effects | Yes | Yes | Yes | Yes | Yes | Yes |
Country fixed effects | Yes | Yes | Yes | Yes | Yes | Yes |
Adjusted |
0.013 | 0.012 | 0.036 | 0.017 | ||
Number of observations | 1,576 | 1,379 | 1,182 | 952 | 1,182 | 952 |
Number of groups | 197 | 197 | 197 | 192 | 197 | 192 |
GMM estimates
Log illegal inflow |
0.103 (0.085) | −0.030 (0.030) | 0.752 |
0.018 (0.012) | 0.806 |
−0.008 (0.016) |
Log illegal inflow |
0.027 (0.016) | |||||
Log legal inflow |
−0.105 (0.152) | −0.140 |
0.266 |
0.899 |
0.050 (0.088) | 0.903 |
Log legal inflow |
0.163 |
|||||
Log CD |
0.010 (0.059) | −0.009 (0.020) | 0.071 |
0.013 (0.011) | 0.063 |
0.002 (0.013) |
Log CD |
0.015 (0.031) | 0.008 (0.015) | ||||
Year fixed effects | Yes | Yes | Yes | Yes | ||
Number of observations | 804 | 804 | 938 | 938 | 804 | 804 |
Number of countries | 134 | 134 | 134 | 134 | 134 | 134 |
Number of instruments | 46 | 46 | 71 | 71 | 73 | 73 |
AR(1) p-stat | 0.0004 | 0.1682 | 0.0000 | 0.0000 | 0.0000 | 0.0000 |
AR(2) p-stat | 0.1811 | 0.7931 | 0.3345 | 0.0821 | 0.4006 | 0.1477 |
Hansen |
0.0004 | 0.0574 | 0.0194 | 0.0645 | 0.0085 | 0.0601 |
Pooled OLS and within-estimator dependent variable: Illegal inflow
Log CD |
0.306 |
0.094 |
0.082 |
0.071 |
Log population density | 0.244 |
6.761 |
4.234 (3.394) | |
Institutions | −1.390 |
−0.813 (0.579) | −1.419 |
|
Constant | 0.887 |
2.987 |
−29.881 |
−17.052 (15.764) |
Year fixed effects | Yes | Yes | Yes | Yes |
Country fixed effects | Yes | Yes | Yes | |
Adjusted |
0.273 | 0.130 | 0.171 | 0.246 |
Number of observations | 920 | 959 | 920 | 497 |
Number of countries | 137 | 132 | 71 |
Pooled OLS and within-estimator dependent variable: Legal inflow
Log CD |
0.302 |
0.033 (0.034) | −0.000 (0.011) | −0.009 (0.012) |
Log population density | −0.231 |
0.648 (0.562) | 0.599 (0.685) | |
Institutions | 0.268 |
0.000 (0.196) | −0.450 |
|
Constant | 6.905 |
5.556 |
2.943 (2.796) | 4.008 (3.306) |
Year fixed effects | Yes | Yes | Yes | Yes |
Country fixed effects | Yes | Yes | Yes | |
Adjusted |
0.101 | 0.039 | 0.054 | 0.142 |
Number of observations | 1,364 | 1,526 | 1,364 | 532 |
Number of countries | 218 | 196 | 76 |
Two-Stage Least Squares (2SLS) estimator
Log CD |
0.042 |
−0.000 (0.013) | −0.010 (0.013) | 0.120 |
0.120 |
0.106 |
Log population density | 0.648 (0.410) | 0.600 (0.499) | 6.783 |
4.162 |
||
Institutions | −0.000 (0.154) | −0.453 |
−0.724 |
−1.288 |
||
Year fixed effects | Yes | Yes | Yes | Yes | Yes | Yes |
Country fixed effects | Yes | Yes | Yes | Yes | Yes | Yes |
Log CD difference | 0.828 |
0.800 |
0.782 |
0.824 |
0.789 |
0.779 |
Log population density | 0.628 (0.796) | 3.522 |
1.133 (1.130) | 3.289 |
||
Institutions | −0.576 |
−1.831 |
−1.096 |
−1.811 |
||
Number of observations | 1,526 | 1,364 | 532 | 959 | 920 | 497 |
Number of Countries | 218 | 196 | 76 | 137 | 132 | 71 |
Kleibergen-Paap F-stat | 405.913 | 327.501 | 302.436 | 430.805 | 340.704 | 322.067 |
Summary statistics
|
||||||
---|---|---|---|---|---|---|
Log legal inflow | Overall | 5.808 | 2.888 | 1,744 | 0.177 | 0.97 |
Between | 2.854 | 218 | ||||
Within | 0.478 | 8 | ||||
Log illegal inflow | Overall | 3.479 | 3.257 | 1,096 | 0.296 | 0.81 |
Between | 3.092 | 137 | ||||
Within | 1.052 | 8 | ||||
Log CD | Overall | 1.115 | 2.466 | 1,768 | 0.309 | 0.68 |
Between | 2.241 | 221 | ||||
Within | 1.039 | 8 | ||||
Log CD difference | Overall | 0.800 | 2.025 | 1,768 | 0.230 | 0.71 |
Between | 1.800 | 221 | ||||
Within | 0.934 | 8 | ||||
Log GNI per capita | Overall | 9.871 | 1.246 | 1,274 | 0.781 | 0.96 |
Between | 1.211 | 188 | ||||
Within | 0.077 | 6.8 | ||||
Log population density | Overall | 5.097 | 1.567 | 1,644 | 0.965 | 0.98 |
Between | 1.568 | 206 | ||||
Within | 0.046 | 8 | ||||
Institutions score | Overall | −0.033 | 0.995 | 1,592 | 0.636 | 0.91 |
Between | 0.990 | 200 | ||||
Within | 0.138 | 8 |
The distributional impact of the Sierra Leone conflict on household welfare The labor market effects of Venezuelan migration to Colombia: reconciling conflicting results † Do international remittances promote poverty alleviation? Evidence from low- and middle-income countries How important are remittances to savings? Evidence from the Latin America and the Caribbean Countries Temporary international migration, shocks and informal finance: analysis using panel data The effects of COVID-19 on employment, labor markets, and gender equality in Central America The role of refugees in the underground economy of the European Union Economic Assimilation of Mexicans and Central Americans in the United States 1 Remittances and Household Investment Decisions: Evidence from sub-Saharan Africa