Identifying Syrian refugees in Turkish microdata

Abstract This article proposes a strategy to identify Syrian refugees in Turkey’s Household Labour Force Survey (HLFS). Even though Turkey’s HLFS contains information on the migrants’ year of arrival to Turkey, it does not provide details on their nationalities. This unfortunate feature mixes Syrian refugees with the standard flow of migration who arrived to Turkey during the Syrian war. I propose to eliminate the standard flow of migrants arrived between 2011 and 2017 by matching them (based on their characteristics) with the migrants arrived in the 2004–2010 period. This method obtains, indirectly, nonstandard migration, i.e., Syrian refugees. The results show that the age distribution of the nonstandard migrants identified matches the age distribution of Syrian refugees as officially released by the Turkish government. At last, I propose a post-stratification adjustment of the survey weights to find the actual geographical distribution of Syrian refugees in Turkey.


Introduction
The succession of revolts that followed the Arab Spring was typically characterized by shortterm demonstrations and/or outbursts of violence in most of the affected countries; all but one, Syria. Since March 2011 until present, none of the multiple belligerents fighting in Syria have been able to regain full control of the country, causing, according to UNHCR, more than 5.68 million 1 of registered refugees out of which 3.6 million 2 were welcomed by Turkey under the temporary protection regime. 3 For Turkey, in particular, this unprecedented situation has not only produced an humanitarian emergency but also likely affected the lives of millions of Turkish people. In this context, the demand for policy responses is pressing and so does the demand of relevant information. This report aims precisely at filling an information gap by means of a strategy that would allow using the primary source for labor market statistics of Turkey, the Household Labour Force Survey (HFLS), for the creation of statistics on Syrian refugees.
The use of microdata when informing about the Syrian refugee crisis has been scarce. Some research has been conducted using macroeconomic data with regards to the Syrians' regional presence. For example, Konun and Tümen (2016) and Tümen (2016) studied the effect of Syrian refugees' arrival on the price level of goods, a finding that the goods whose production process intensely employs informal workers showcased a decline in their prices. This would be explained by Syrian workers replacing Turkish natives in informal jobs at a cheaper rate, passing the lower labor costs onto the goods' prices. In addition, Tümen (2016) also found that natives have both lower chances of finding an informal job and higher chances of finding a formal one. The latter might be due to the increase in the provision of public services caused by the arrival of the refugees. Another article analyzing the impact of Syrian refugees is Del Carpio and Wagner (2015), this time by combining microdata from the Turkish Labour Force Survey with macro data on the number of refugees by region. These authors, in addition to finding a large displacement of Turkish natives from the informal sector due to the arrival of the refugee population, also found a net displacement of women and the low educated away from the labor market.
Despite some successful attempts at producing studies on the impact of Syrian refugees at the macroeconomic level, little is known about their personal circumstances. One of the most remarkable attempts from a sociological point of view is the Syrian Barometer (see Erdoğan 2017), a national level survey covering 11 provinces and interviewing 1,235 Syrian families, reaching out, in total, 7,591 Syrians. Even though attractive in terms of understanding Turkish nationals' sentiment with regards to the Syrian population, it lacks, beyond a few basic questions, relevant information with regards to Syrians' labor market performance.
Other ad-hoc surveys focused on Syrian refugees' socioeconomic conditions are not as ambitious and the few existing sources lack national representativeness. Still, a remarkable effort in gathering data at the microeconomic level can be found in UÇak and Raman (2017).
This research uses a survey on Syrian-owned SMEs to provide a snapshot of this type of companies, including the value of having them for the Turkish economy. With regards to the dataset, which can be taken as a small-scale enterprise survey, it included visits to 230 businesses equally split between Istanbul and Gaziantep on the condition that they were legally established, are currently active, and had at least one employee. On the negative side, this database is, as confirmed by its authors, not meant to be nationally representative. Data collection efforts can also show glimpses of creativity, as in Kaymaz and Kadboy (2016), where the authors make use of a survey carried out on migration routes to find that around 30% of Syrian refugees have university degrees. Even though the extent to which Syrian refugees have such high qualifications may have been exaggerated due to the survey mode, it brings to the spotlight the importance of developing a model for the recognition of refugees' prior learning.
Lack of data may affect the depth and relevance of research on Syrian refugees; for example, YavÇan (2017) tried to illustrate the challenges faced by Turkey regarding Syrian refugees resorting to a small survey done by UNCHR in some Greek islands. Another example is Cagaptay (2014), in which an attempt to gauge the impact of Syrian refugees on the ethnic and sectarian balance of south-eastern provinces has to rely on data from the 1960 Census because it was the last one that collected data on ethnicity. The lack of nationally representative data on Syrian refugees in Turkey is in contrast to the availability found in Lebanon, where at least two such surveys have been carried out (see Alsharabati andNammour 2016 or BRIC 2013), or in Jordan, where Syrian refugees can be identified within the Labour Force Survey. 4

Household Labour Force Survey
The fact that the refugee population in Turkey represents 4.4% of the population living in Turkey 5 creates a growing need for nationally representative data on Syrian refugees in Turkey which is not currently fulfilled. Fortunately, the relatively high proportion of Syrian refugees in the Turkish economy might have opened the door as well to the use of nationally representative microdata from the Turkish Statistical Institute. However, using Turkstat databases for analyzing Syrian refugees is not straightforward. Household surveys in Turkey usually target families that are inscribed in the Address Based Population Registration System (ABPRS) and Syrian refugees under temporary protection are not included in that registry. 6 An exception to this survey methodology is given by the HFLS 7 that instead of families targets addresses, thus allowing interviewers to find Syrian families under certain conditions. At present, even though Syrian refugees take part of the HLFS, their identification is not direct; the HLFS publicly available microdata does not provide the nationality of those classified as foreign-born, thus mixing up Syrians with the standard flow of migrants coming to Turkey (see Appendix A for a quick visual inspection of how this flow looks like). In this article, I propose an indirect identification method, whereby removing the standard migrants of the 2011-2017 period allows me to find nonstandard migration as a leftover. This method is meant to identify all nonstandard migrants who came into Turkey between 2011 and 2017. In practice, this group contains all Syrian refugees who migrated during that period, including those covered by the temporary protection regime, those with short-term residence permits, and 4 Syrian refugees are underrepresented in the Labour Force Survey; however, their survey weights have been adjusted to add-up to their total population. 5 As reported by the DGMM at http://www.goc.gov.tr/icerik6/temporary-protection 915 1024 4748 icerik 6 Their information is kept separately by the DGMM. 7 See http://tuik.gov.tr/MicroVeri/Hia 2017/turkce/index.html for more information.
those who acquired the Turkish nationality. It should be noted that some other migrants (particularly those coming from Iraq or Afghanistan in recent times) may have also been included in the group. Still, throughout the report I refer to the group as a whole as "Syrian refugees" because Syrian refugees constituted 89.7% of nonstandard migration in 2017. 8 In what follows, Section 2 explains the matching strategy I used to isolate Syrian refugees.
Section 3 presents a post-stratification adjustment that calculates new survey weights for Syrian refugees. This section also uses these newly created weights to estimate the geographical distribution of Syrian refugees in Turkey including a comparison with the official distribution.

An identification strategy for Syrian refugees in the HLFS
In the 2017 Household Labour Force Survey (HLFS) the number of foreign-born individuals who arrived between 2011 and 2017 is six times larger than the number of migrants who did so between 2004 and 2010. Unfortunately, the publicly available microdata of the HLFS do not contain information on the country of origin, and even though I suspect that Syrian refugees make up for the majority of observations among those who migrated between 2011 and 2017, they are unlikely to be the only foreigners who entered Turkey since the onset of the Syrian crisis. This hypothesis is supported by Figure 1, which shows the existence of a relatively constant number of foreign-born individuals arriving to Turkey during the years preceding the Syrian war (2004)(2005)(2006)(2007)(2008)(2009)(2010). As a result, Syrian refugees are probably mixed up in the data with the hereinafter called "standard" migrants, thus preventing a direct identification of Syrian refugees.

Assumptions
To identify the Syrian refugees present in the sample, I pursue an indirect identification strategy. Instead of finding Syrians among the 2011-2017 migrants, I find those who are not and then remove them from the sample (see Figure 1 for a visual explanation of the idea) with Syrian refugees who are identified as a "leftover" of the procedure. For this strategy to work, I assume that there is a relatively constant flow of what I call "standard" migrants. In particular, I assume that during the 2011-2017 period there was, on top of Syrian refugees, as many migrants as there were during the 2004-2010 period. This assumption is based on the findings of Korfalı and Acar (2018); their chapter shows how the flow of migrants from Central and Eastern Europe (which constitutes 40% of the total migration) to Turkey remained unaffected after the Syrian refugees started entering into Turkey. In practice, this assumption provides the number of observations that need to be removed from the ex-post migrants' group, i.e., those arrived between 2011 and 2017.
In addition, ex-ante and ex-post "standard" migrants, some of them thought to be Turkish-German by Bel-Air (2016), are assumed to share similar socioeconomic characteristics that are (1) observable in the microdata and (2) significantly different from those of Syrian refugees. This allows for the separation of "standard" migrants from Syrian refugees in the ex-post migrants' group. If, for instance, the ex-ante and the ex-post migrants' groups were identical, the matching would be trivial and refugees would not be identified, i.e., I would be removing ex-post migrants at random which does not help more than no matching at all. The comparability of ex-ante and ex-post migrants is tested (see Table 1) by comparing mean values of variables where, in principle, I would expect Syrian refugees and "standard" migrants to differ. It should be noted that for the sake of relevance, the comparison is done at the family level. This is because Notes: The table shows averages at the family level for a number of variables. The wording "existence" refers to the existence of at least one person with the mentioned characteristic in the household. In all cases, it can be rejected that the difference in means is equal to zero at the 95% confidence level.
I match families-as opposed to individuals-so as to keep within-household coherence. Moreover, only individuals arrived during the prescribed period are included as part of the family, i.e., to minimize the noise due to mixing 9 with the local population and/or other migrants. It stems from Table 1 that sizable differences exist between the families who arrived to Turkey between 2004 and 2010 and between 2011 and 2017. For example, family size doubles among ex-post migrants, as it roughly does the proportion of children aged 0-14, proving the arrival of a much younger population in recent times. In turn, the higher number of households with a widow suggests the existence of families that may have escaped from a war. Moreover, I observe significantly less individuals with tertiary education, less female workers, and more NEETs among ex-post migrants, which hints the existence of strong differences at the cultural and at the socioeconomic level between the two groups under analysis. I also find that 65% of the workers (the informality rate in Turkey is 35.0%) coming from ex-post migrant families are not registered with the social security institute. This fact fits well with the existence of an underlying population of Syrian refugees because it is known that they had received very few work permits at the time of the survey. 10 In sum, based on the observed differences, it is reasonable to argue that ex-post migrants constitute a different group that in turn supports the use of matching.

Matching
The matching of ex-post "standard" migrant families with ex-ante "standard" migrant families uses nearest neighbor propensity score without replacement. In practice, this translates into the calculation of a probability (propensity score) of being an ex-ante "standard" migrant family for ex-post migrant families based on observable characteristics like the ones shown in Table 1. Then, based on the scores every ex-ante migrant family is matched with the ex-post migrant family who has the closest score-the nearest neighbor-and is not considered again for matching, hence there is the lack of replacement.

Model estimation
The variables chosen to be part of the propensity score calculations are selected based on the expectation of a different prevalence in "standard" migrant families and refugee families. No other consideration was taken since, according to Caliendo and Kopeinig (2008), matching is not intended to estimate structural parameters. In addition, I follow Rubin and Thomas (1996)'s recommendation against "trimming" models for the sake of parsimony. As a result, I do not remove variables based on their parameters' statistical significance provided there are reasonable doubts with respect to their relationship with being a "standard" migrant family.
The variables and their definitions are summarized in Table 2 for convenience. It should be noted that all variables are defined for the whole population of households. This also applies to variables defined for 15+ individuals because there is no household without at least one individual from said age group. In total, I use 12 variables that cover demographics (kids, young, women, widow), educational attainment (university, student), labor market indicators (fem work, NEET, workers, informal, garment), and the existence of mixed families (turkish). It is also worth noting that informality is defined using the existence of contributions to the social security institute.
Moreover, I define the garment sector using ISIC 11 rev.4 codes 13, 14, and 15. On top of the previously described variables, regional dummies are represented by the matrix R in Equation (1).
Propensity scores are built with the help of a Logit model so as to maintain the probabilities of being a "standard" migrant family bounded between 0 and 1. The model is defined for the ith family using the logistic cumulative distribution function (CDF) as follows: where the probability of being a "standard" (n = 1) migrant family given some characteristics (X) is given by the logistic function G(·). The arguments inside this function are given by:    Table 1 are differential factors between migrant groups even after controlling for all of them at the same time. For example, it can be seen that living in a mixed household with a Turkish native decreases the probability of having found a Syrian family by 12%, while the same probability increases by 10% for every informally employed migrant found in the household. With respect to the proportions, the results show that an increase of 0.1 in the proportion of 15-24 kids in the family lowers the probability of being a "standard" migrant family by 1.9%. In addition, it is found that "standard" migrant families have a much higher propensity to live in the regions of Antalya and Van (data not shown in   (1).
most Syrian refugees were initially registered in south eastern provinces of Turkey and hints refugees may have migrated to other regions.

Identification
Given the marginal probabilities shown in Table 2 Table 1, that these two groups of migrant families are remarkably different from each other. which is markedly distinct from earlier migrants, the matching has been able to create a control group with an almost identical distribution of propensity scores. This can be interpreted   The resulting matching can also be tested with the help of the same variables shown in Table   1. In this regard, Table 4  Certain dissimilarities can still be found between ex-ante and ex-post "standard" migrant families. These differences do not necessarily signal a lack of comparability between the two groups since they might be due to the time spanned between the arrival of ex-ante migrants and the time of the survey, 2017. For example, the fact that ex-ante migrants are 7 years older than ex-post migrants might explain the lower percentage of ex-post "standard" migrant families where at least one individual holds a tertiary degree.

Quality check
The dramatic increase in foreign-born individuals captured by the HLFS since the onset of the Syrian civil war and the marked differences in the socioeconomic indicators shown by those 12 Ex-post migrants families' (i.e., Syrian and ex-post "standard" families together) size is 3.34 before the separation, Notes: The table shows averages for a number of variables at the family level. The term "existence" refers to the existence of at least one person with the mentioned characteristic. In all cases, it can be rejected that the difference in means between Syrian families and ex-post "standard" migrant families is equal to zero at the 95% confidence level.
identified as Syrians leaves little doubt about them belonging to refugees. However, questions might still arise about the specific subpopulation represented by those captured by the matching methodology (refer Table 1).
As a quality control check, Figure 3 compares the population pyramid of the 3,858 Syrian refugees identified as such by the matching methodology with the pyramid of (1) the Syrian refugees under temporary protection registered by the Turkish Directorate General of Migration Management and (2) the "standard" ex-post migrants. 13 The figure shows the proportion of the group-specific population held by each of the eight age groups in which the population has been split. The age distribution of those identified as Syrian refugees in the HLFS is very similar to those who are supposed to be representing in all of the age groups under consideration. 14 In addition, the comparison between the age distribution of Syrian refugees and ex-post "standard" migrants shows that the matching is not trivial and the procedure has actually removed the noise that was surrounding the Syrian refugee population in the data.

Background
The HLFS covers 15 all settlements of Turkey at the sampling stage, thus providing nationally representative figures on all residents with the exception of the noninstitutional population. In practice, though, the coverage is further restricted to Turkish natives residing in Turkey and foreigners with long-term residence permits (see İçduygu 2013, pp. 8-9

Syrian refugees in the HLFS
In spite of the initial inability to covering individuals under the temporary protection regime, some of the interviewed households in the HLFS (approximately 1,000 households) are occupied by foreigners who, given the year of arrival to Turkey (among other characteristics), are likely to be Syrian refugees.
Two problems arise from the appearance of Syrian refugees in the HLFS sample. First, around 3,858 Syrian refugees are currently representing more than 1 million 18 Turkish citizens (including foreigners with long-term residence permits) even though their socioeconomic characteristics are far from comparison with the ones who are supposed to be representing.
Second, since the sample currently includes Syrian refugees, the total population represented by the sample should be increased to 81.6 million as of July 2017, i.e., 78.6 million Turkish citizens and long-term foreign residents plus 3.19 million Syrian refugees as estimated by the DGMM (including Syrian who acquired the Turkish nationality, those on short-term residence permits, and those covered by the temporary protection regime).

Nonresponse adjustment for Turkish residents
I propose to solve the former problem by treating the existence of Syrian refugees as a nonresponse problem, i.e., as if the Turkish family that should have been interviewed was not present at home at the time of the visit. By following this assumption, the expanded number of Turkish people is down to 77.6 million, thus requiring an upward adjustment of the survey weights.
and N represents the total number of observations in the sample. In addition, since the number of visits to mistaken households is not meant to have the necessary proportion for the weights to add up to the actual population of Syrian refugees, I add a correction factor S − that makes the 19 See https://ec.europa.eu/regional policy/en/policy/what/glossary/n/nuts/ for an explanation on the statistical regional units classification. In Turkey, there are 26 subregions at the NUTS-2 level. 20 It should be noted that standard errors will increase as a result of the nonresponse adjustment. Users may want to consider the use of replication methods, including bootstrap when carrying out analysis with the proposed methodology to take into account the added uncertainty. 21 Even though these figures are published at the NUTS-3 level (provinces), I disregard the geographical distribution because I suspect that Syrian refugees have incentives to redistribute themselves within Turkey to areas with a higher number of job opportunities, for example, Bursa or Istanbul. for Syrian refugees (6) where S is the number of Syrian refugees in the sample. It should be noted that, in practice, the correction factor divides the actual population of Syrians by the expanded population of Syrians which arises from the probability of choosing a person in a particular subregion. Both the adjustment factors for Turkish residents and the adjusted weights for Syrian refugees can be found in Table C1 in Appendix.
The application of this post-stratification adjustment allows me to estimate the actual geographical distribution of Syrian refugees. This distribution (HLFS) together with the official distribution as published by the government of Turkey can also be found in Table C1 in Appendix at the subregional level (NUTS-2), the lowest level of geographical disaggregation provided in the microdata. The comparison shows the existence of an internal migration pattern from

Conclusions
The Syrian refugees hosted by Turkey have a higher risk of facing poverty and working conditions' deficits. As it is often the case with migrant populations, those more in need of help are also the ones for whom less information can be found due to the difficulties in tracking down these groups. This article proposes the use of the Turkey HLFS to overcome the information deficit with regards to Syrians in Turkey. In particular, I propose an indirect identification strategy to isolate Syrian refugees from other "standard" migrants, since both are grouped together in the publicly available microdata.
The identification strategy produces a population pyramid for HLFS refugees, that is comparable with the age profile recorded by the Turkish Directorate General for Migration Management. In addition, I show that Syrian refugees might have internally migrated from south-eastern provinces bordering Syria to more industrialized areas of Turkey like Bursa, Konya, or Istanbul. This pattern of internal migration would need to be confirmed by other instruments, yet it suggests that a reallocation of funds and humanitarian efforts might be due.
In addition, this methodology should allow researchers to use the full depth of Turkey's labor force survey for the study of the Syrian refugee population. This includes the creation of labor market indicators for this group such as those based on formality rates, average earnings, and details about the employment structure or the educational background. These and other results should allow policy makers and authorities alike to build better informed policies, including active labor market policies aimed at the Syrian population.

A.1 Number of children aged 0-14
The data provided in the HLFS by the Turkish Statistical Institute refer only to those aged 15 or older; however, information regarding the number of people aged 0-14 living in the household can be retrieved. This variable is calculated by subtracting the variable hh buyukluk, which contains the number of people in the household (including children), with a variable of our own creation that contains the number of household members aged 15 or more. Some pieces of household-level information are assigned to the children, for instance, the region (NUTS-1), the subregion (NUTS-2), and the survey weight; in addition, children living in a household whose head is foreign-born are given the year of arrival of the head of the household provided he/she arrived not before 14 years since the time of the survey. In all other cases, it is assumed that the children were born in Turkey.

A.2 Year of arrival to Turkey of persons born abroad
One of the questions available in the HLFS is whether the respondents were born abroad or in Turkey. As an example, this group contains 10,032 observations in the 2017 HLFS, including children aged 0-14 (see first paragraph of this appendix). For most 23 of the foreign-born population, the dataset also provides information on their year of arrival to Turkey. The variable "year of arrival" is built in two steps as its information comes from two sources: the variable buil yil for those who live in the same province since their arrival to Turkey and the variable tr yil for those who changed provinces within Turkey at least once since their arrival and have lived abroad for at least 12 months.
The full list of logical skips used to build the year of arrival is shown in Table A1, where the column "subpopulation" presents the conditions that respondents need to fulfill for their year of arrival to come from either of the two options. It should be noted that the condition referring to the variable buil yasama (which asks whether the person have permanently lived in the current province) is shown for completeness but it does not make any difference as it is logically impossible to have lived the whole life in the same Turkish province while being born abroad.
The resulting variable is plotted in Figure A1 which shows the number of observations by year of arrival. It can be observed the existence of two peaks, one in 1989, which coincides with the migration/expulsion of Turks from Bulgaria, and the second after 2011, right after the Syrian civil war. 23 The survey does not ask the arrival year to foreign-born residents who have changed their province of residence if they have lived less than 12 months outside Turkey. For instance, in the 2017 HLFS, this group totals 344 observations out of the 1,721 which conform the group of foreign-born who have changed provinces within Turkey at least once in their lives. Notes: *In practice, it is not possible for a person born abroad to have lived permanently in the same Turkish province his/her whole life. There is only one observation (out of 10,032) in the 2017 HLFS for which a Yes is recorded in 'buil yasama' despite being born abroad and it is disregarded.