Migration and Knowledge Diffusion: The Effect of Returning Refugees on Export Performance in the Former Yugoslavia

AS of today, there are over 26 million refugees hosted in a foreign country. This figure is likely to keep increasing as sources of displacement such as ongoing conflict and climate change pose important challenges to stability around the globe. For the most part, refugees in host countries face hurdles to integrate due to negative perceptions of their effect on local labor markets, crime levels, or social cohesion. Despite the negative public sentiment, refugees, and migrants alike, represent important economic opportunities for both their host communities and home countries, as evidenced by a small but growing literature.¹ Yet to the best of our knowledge, there is no rigorous study documenting the crucial role refugees can play in shaping large-scale economic outcomes (such as a country’s industrial or export structure) in their home countries upon return. This paper fills this gap by documenting the long-run effect of returning migrants on industry-level exports driven by productivity shifts, using the case of Yugoslavian refugees during the 1990s.

We exploit a natural experiment and document a systematic relationship between refugees spending time and gaining work experience in a foreign country and the subsequent export performance at home of the same sectors in which they had worked while abroad upon their return. The context is the early 1990s, when about 700,000 citizens of what was then Yugoslavia fled to Germany escaping war. Most Yugoslavian migrants in the first half of the 1990s were given a Duldung status (German for “toleration”), in effect a temporary protection status, or more specifically, a “suspended deportation” permit. After the Dayton peace agreement was signed in 1995, the protection status of temporary migrants was revoked, forcing them to leave the country. By 2000, the majority of these migrants had been repatriated back to their home country or to other territories of the dissolved Yugoslavia.

Using confidential administrative data from the Institute for Employment Research (IAB), we identify returning migrants as Yugoslavian workers who entered the German labor force (by industry) between 1991 and 1995 and had left it by the year 2000. In this sense, our “treatment” reflects the extent to which Yugoslavian refugees were exposed to German productive knowledge. We link this information to standard industry-level international trade data and employ a difference-in-differences methodology to estimate changes in export values from the former Yugoslavia to the rest of the world. The different data sources we rely on limit us from linking both exports and migrants to the different nations that emerged after the dissolution of Yugoslavia. We overcome this limitation, however, by aggregating data for all former Yugoslavian nations in order to construct an industry-level panel data. We find that the industries that perform better in terms of exports in the years following the repatriation are the very same ones in which refugees worked while in Germany.

In order to address concerns of endogeneity due to, for example, self-selection of workers into certain industries, we instrument for the actual number of returning workers per industry using a spatial dispersal policy that exogenously allocated asylum seekers across the different administrative regions of Germany upon arrival.

Our main finding is that industries with a 1% increase in return migration (as defined by our treatment) experienced an increase in exports to the rest of the world of 0.1% to 0.16% between the pre- and postwar periods. Based on all the returning refugees in our sample, this represents up to 6% of all the export growth of former Yugoslavian nations between 1990 and 2005, most of it occurring after 2000.

Our results are robust to an estimation in the form of an event-study, which shows that there are no existing previous trends for industries treated at different levels of intensity and that the treatment effect becomes larger with time. In addition, we complement our main difference-in-difference results by applying the synthetic control method (Abadie & Gardeazabal, 2003; Abadie et al., 2010, 2015) to create a counterfactual for every Yugoslavian export industry in our sample. We find that real Yugoslavian industries outperform their synthetic counterparts, on average, with this pattern being particularly strong for industries associated with more intense treatment.

When exploring mechanisms driving our main findings, we find robust evidence that points to industry-specific productivity shifts brought on by returning refugees as the main driving force of export performance. In particular, our findings are robust to substituting actual exports with a constructed measure of revealed comparative advantage based on Costinot et al. (2012), which arguably reflects productivity dynamics, and we replicate our analysis using more aggregated data to study changes in labor productivity (as measured by output per worker, added value per worker, and wages) using data from the UN Industrial Development Organization (UNIDO), as well as using sectoral data on exports per worker and size (firms and workers) for the specific case of Bosnia. We find that all these complementary results are consistent with our main findings.

Through numerous empirical tests, we also rule out a number of alternative explanations, which include our results being driven by purely scale effects or convergence patterns in structural transformation. We also rule out channels that the literature links to migration flows and could also explain export performance, such as capital formation and reduction of international trade transaction costs due to migrant networks. While exploring these channels provides us with additional insights, we find no evidence in them to invalidate the results that point to productivity shifts as a mechanism driving our results.

Consistently with productivity shifts, we then document that our results are particularly driven by certain types of workers and occupations, which would be more likely to transfer knowledge, technologies, and best practices across borders, and thus able to induce productivity improvements. For instance, our results are driven by workers with high educational attainment, in occupations intensive in analytical tasks (as opposed to manual ones), occupations that can be classified as professional and/or skill intensive, and occupations that have managerial characteristics. We also find that our results are stronger when looking at workers who, while abroad, experienced faster wage growth and were employed by the top-paying firms within each industry.

Our findings contribute to the economic literature at large by documenting substantial industry-specific productivity increases, resulting in changes in the export industrial composition of a country as a whole, as a result of returning migrants who were exposed to foreign knowledge, technologies, and best practices. To the best of our knowledge, this is the first paper documenting such large-scale and country-level industrial productivity shifts induced by return migration in the context of a natural experiment.²

Our paper contributes to a number of strands of the literature. First, we contribute to the literature on migrants as drivers of international knowledge diffusion back to their home countries (Kerr, 2008; Agrawal et al., 2006, 2011; Breschi et al., 2017; Bahar & Rapoport, 2018; Miguélez, 2018; Miguelez & Noumedem Temgoua, 2020; Bahar et al., 2019) and, relatedly, to a burgeoning literature that looks at migrants as drivers of productivity shifts (Scoville, 1952a, 1952b; Markusen & Trofimenko, 2009; Poole, 2013; Choudhury, 2016; Hornung, 2014; Malchow-Møller et al., 2017; Bryan & Morten, 2019). Second, our results speak to the important link between migration and economic development, given the substantial role of migrants affecting structural transformation, an important driver of economic growth (Imbs & Wacziarg, 2003; Hausmann et al., 2006; Hidalgo et al., 2007; Koren & Tenreyro, 2007; Cadot et al., 2011). Third, we contribute to the growing literature emphasizing the role of managerial skills as a crucial determinant of productivity (Bloom & Van Reenen, 2007; Bloom et al., 2013; Giorcelli, 2019) and of exports (Bloom et al., 2021). Finally, we contribute to a growing number of studies that focus on consequences of forced displacement for the refugees’ home countries (Waldinger, 2010, 2012; Acemoglu et al., 2011; Grosfeld et al., 2013; Akbulut-Yuksel & Yuksel, 2015; Bharadwaj et al., 2015; Pascali, 2016; Huber et al., 2021; Testa, 2020; Mayda et al., 2022), as well as to the literature of the economic consequences of return migration (McCormick & Wahba, 2001; Dustmann & Kirchkamp, 2002; Mesnard, 2004; Wahba & Zenou, 2012; Batista et al., 2017).³

The rest of the paper is organized as follows. Section II provides a description of the historical context of the Yugoslavian refugee crisis. Section III details the data sources and the empirical strategy. Section IV presents the main results and performs a series of robustness tests. Section V explores possible mechanisms driving our results, and section VI exploits heterogeneity effects based on the characteristics and occupations of the refugees. Section VII concludes. The paper is accompanied by an online appendix, which we refer to throughout the text.

In June 1991, the Socialist Federal Republic of Yugoslavia started to disintegrate following several armed conflicts and ethnic civil wars. Fighting began with the “Ten-Day War” in summer 1991 after Slovenia declared its independence. Soon after, the conflict spread to Croatia and later, in 1992, to Bosnia and Herzegovina. It was only in December 1995, with the signing of the Dayton Peace Accord involving President Clinton, that the armed conflict officially ended.

During the armed conflict, around 3.7 million people (roughly 16% of the Yugoslavian population) were displaced and fled from their homes, making this episode the largest migration flow in Europe since the end of World War II (Radovic, 2005). While many affected by the war became internally displaced, about 800,000 people resettled outside the boundaries of the former Yugoslavia, hoping to find refuge in other countries (Lederer, 1997).⁴ Among these countries, Germany was one of the best suited to receive these refugees because of the already significant Yugoslavian community residing there and Germany’s ability and willingness to provide protection to those fleeing the war.⁵ The flow of refugees into Germany responded to the dynamics of the conflict: in the early stages of the war, involving mostly Croatians, about 100,000 of them arrived in Germany; later, when the war spread to Bosnia, acts of systematic violence triggered massive outflows from those areas, and Germany hosted some 350,000 Bosnian refugees. Simultaneously, Germany also received another 250,000 Yugoslavians mainly from Serbia and Kosovo. Thus, overall during the first half of the 1990s, Germany received roughly 700,000 migrants from Yugoslavia, making it by far the largest recipient foreign country (see Lederer, 1997, for a detailed account of these flows).

Most of Yugoslavian refugees who arrived in Germany were given a temporary protection status, known as Duldung, which can be translated to English as “toleration.”⁶ The temporary character of the Duldung status constituted a “suspended deportation” status. In other words, Duldung holders were allowed to remain in Germany until the permit expired, after which they were obliged to leave the country immediately. While the Duldung duration upon issuance was set to six months, the authorities had the option to renew it. De facto, the Duldung status was renewed for all holders as long as the war continued.

A less popular option for Yugoslavians fleeing the war and arriving in Germany was to apply for asylum. According to Article 16(a) of the German Basic Law (Grundgesetz), an individual is eligible to seek asylum if he or she faces individual persecution and is able to prove so. If granted asylum, the individual enters a path toward permanent residency (Hailbronner, 2003). Asylum recognition rates, however, were very low for citizens from the former Yugoslavia (Lederer, 1997, documents that between 1992 and 1995, only 1% of Bosnian applicants were granted asylum). This is because most of them could not prove to the German authorities they were facing individual persecution at home following the standards set by the German authorities at the time (Dimova, 2006). Importantly enough, however, asylum seekers whose request was denied also were able to receive Duldung status, since the German authorities were not deporting these refugees back into active war zones.

A large number of Yugoslavian refugees integrated into the German labor force after their arrival. For instance, in 1992 alone, the number of workers from (former) Yugoslavia rose by 15.3% to 375,000 (Deutscher-Bundestag, 1994). Overall, both Duldung holders and asylum seekers had relatively good access to the labor market.⁷ But there was an important difference between the two statuses concerned with respect to their mobility. Duldung holders faced no geographical limitation. But asylum seekers were subjected to mandatory residency (Residenzpflicht) within the region in which their application was initiated while it was processed.⁸ The decision on which region would process the application was made by the authorities based on preestablished quotas. This limitation on geographic mobility for asylum seekers is an important part of our identification strategy, which we detail in section IIID.

Generally refugees found employment across diverse sectors and relied on different channels to secure their jobs. Some were able to use their network of friends and family relatives, some relied on local employment agencies, and some found work by themselves (Walker, 2010; Ruben et al., 2009).

The signature of the Dayton Peace Accord in December 1995 officially marked the end of the war that started in 1991. After that date, the German authorities had no reason to renew the Duldung status of refugees and indeed enacted the imminent deportation of refugees back to the former Yugoslavia.⁹ In fact, only one day after the signing of the Dayton Accord, Germany formally announced a repatriation plan through which Duldung refugees were gradually forced to leave the country (Dimova, 2006), often simultaneously rolling out assisted repatriation programs (Bosswick, 2000).¹⁰

Repatriation was planned in two main phases. The first phase targeted single adults and childless couples, as well as people with family back in their home country. The second phase targeted the rest of the refugees. By summer 1996, letters requesting deportation were sent out, and by the end of 1996, deportations had begun. Repatriation and deportations continued until 2000, though most of them had happened by 1998. Figures by international organizations and independent academic research suggest that about 75% of Yugoslavian civil war refugees returned to their home country or to another former Yugoslavian nation, with an additional 15% settling in third countries and only about 10% remaining in Germany (UNHCR, 2005; Ruhl & Lederer, 2001; Lederer, 1997).¹¹ With respect to Croatian refugees, Lederer (1997, 310) explains: “During the Croatian-Serbian War (1991 to 1993) numerous Croatians were also admitted to the Federal Republic of Germany. According to information from the Federal Ministry of the Interior of 9 October 1996, most of the original 100,000 Croatian refugees should have returned to their homeland within the framework of the repatriation process that began in 1994. However, the Federal Ministry of the Interior notes that there is no precise information on this from the federal states.”

Figure 1a plots the evolution of the population size of Yugoslavia (as a whole) over time during the period of our study. It shows, consistent with the historic accounts, overall population drops significantly after 1991, and following the end of the war (1995) it grows again. In 2005 the population of the former Yugoslavian countries combined was comparable to the population of Yugoslavia before the war, in 1990.

Figure 1b plots the “mirror image”: the inflow of around 700,000 refugees from the former Yugoslavia who arrived in Germany in the first half of the 1990s. In 1980, there were already about 600,000 Yugoslavians residing in Germany. This stock remained steady until the late 1980s when the net inflow of Yugoslavian migrants started to grow at a rate of 25,000 per year, including until the year 1990. This rate skyrocketed to 168,000, 250,000, and 165,000 during 1991, 1992, and 1993, respectively. The sharp increase in the net inflow of migrants was fueled by refugees escaping war. We also see a sharp increase in asylum requests from Yugoslavian citizens during the same years.

Consistent with the accounts, the number of Yugoslavians in Germany sharply declines starting in 1996, after the Dayton treaty was signed. By 2000, close to 350,000 Yugoslavians had left the country. While some of them left to a third country, it has been estimated that the majority of them returned to countries of the (by then) former Yugoslavia (UNHCR, 2005; Ruhl & Lederer, 2001; Lederer, 1997), consistent with figure 1a.

In the German labor force, we see patterns consistent with the historical narrative described so far: both the inflow of Yugoslavian workers into the tradable sector labor force between the years 1991 and 1995 and their outflow in the years that followed are highly unusual as compared to foreign workers from other nationalities, as figure 2 shows. The figure plots the yearly share of Yugoslavian workers entering and exiting the labor force of Germany’s tradable sector. The share is always computed using the total number of all foreign workers entering and exiting the labor force in those same years at the denominator. It becomes clear from figure 2 that Yugoslavians entered the labor force in much higher proportions during the first half of the 1990s compared to the same proportion in years before 1990 and after 2000. We also see that Yugoslavian workers exited the German labor force in higher proportion during the latter half of the 1990s, consistent with the historical events.

We link a number of data sets together for our study. First, we use data on exports for the period 1984 to 2014, which comes from bilateral trade data compiled by Feenstra et al. (2005), with extensions and corrections suggested by Hausmann et al. (2014; the original source of the trade data is UN Comtrade). In most cases our dependent variable is exports by industry from Yugoslavia to the rest of the world, excluding to Germany. We exclude Germany as a first attempt to reduce the possibility that our results are driven by lower transaction costs caused by migrant networks between the former Yugoslavian countries and Germany (Rauch & Trindade, 2002; Parsons & Vézina, 2018).¹²

Industries are defined using the four-digit Standard Industry Trade Classification (SITC) revision 2 and list 786 different varieties. This industry classification provides a disaggregation level that enables a meaningful discussion about export diversification patterns. Some examples of industries in this level of disaggregation are, for example, “Knitted/Crocheted Fabrics Elastic or Rubberized” (SITC 6553) and “Electrical Measuring, Checking, Analyzing Instruments” (SITC 8748). Following Hausmann et al. (2014), we exclude countries below 1 million citizens and total trade below USD $1 billion in 2010. Other variables created using trade data are explained as they are introduced into the analysis.

The data on migrant workers in Germany are based on records from the German social security system and comprise all persons employed subject to social security contributions, with the exception of self-employed and civil servants.¹³ The records indicate the industry where the workers are employed. Our sample is restricted to 40% random draws of foreign nationals observed on June 30 of each year from 1975 to 2014, augmented by the employment history of each individual for our sampling period. This amounts to about 2.4 million workers per year on average, a large enough number for the random draws to form a representative sample.¹⁴ Moreover, since we can observe the full employment history, we can determine whether an individual was employed before or after any given year in Germany, which we exploit to construct our treatment. Beyond individual information such as age, nationality, and educational attainment, the data include detailed occupational codes categorized in more than 300 different occupations (see more details on this data set in online appendix section B).

In order to construct an instrument to deal with endogeneity issues, we use inflow quotas mandated by the government that define the regional distribution of asylum seekers (Königsteiner Schlüssel). These quotas are determined yearly by the Joint Science Conference (Gemeinsame Wissenschaftskonferenz, GWK). The GWK sent to us, in response to our request, yearly data between 1990 and 2016.

With these data sets we are able to match the exports of Yugoslavia to the rest of the world with the number of Yugoslavian workers in Germany working in the same industry category.¹⁵

Our main unit of analysis, which we use as our dependent variable in our baseline specification, is the combined exports by industry of Yugoslavia to the rest of the world, excluding to Germany (to avoid results being driven by lower transactions costs, as explained above). To construct this variable we combine per industry exports of Yugoslavia as a nation until 1991, with exports by industry of all countries combined that formed Yugoslavia after 1992: Bosnia and Herzegovina, Croatia, FYR of Macedonia, Montenegro, Serbia, and Slovenia (excluding trade within these countries).¹⁶ We focus on all former Yugoslavian countries as one following the disintegration because the German data used to construct our treatment do not allow us to distinguish which is the region of origin of the refugees within Yugoslavia (e.g., we only see that they entered the labor force with a Yugoslavian passport).

The main independent variable, the treatment, is the number of Yugoslavian refugees by (tradable) sector who (i) joined the German labor force between 1991 and 1995,¹⁷ (ii) had not been recorded in our data in 1990 or before, and (iii) had not been recorded in our data in 2000 or after.¹⁸ Applying this filter allows us to proxy for returning refugees with a high degree of certainty, consistent with the historical accounts described above.¹⁹

As already explained, a limitation of our data is that we cannot distinguish with full certainty whether these workers with Yugoslavian passports who left the labor force indeed returned to the former Yugoslavia. Thus, in our calculation of return migration, we are including workers who, for instance, stayed in Germany working in the informal sector or went to a third country. Yet all of these possibilities work against us in our estimation, implying that our estimates are to some extent understating the true effect of return migration. However, plenty of historic evidence suggests that indeed those who were repatriated went back and worked in related industries. For instance, Ruben et al. (2009) note that those who worked during their time in Germany were more likely to be employed when they returned to the former Yugoslavia, with many finding jobs in industries in which they had worked before the war. There is anecdotal evidence, too, suggesting that after returning to their home countries, refugees subsequently worked (or founded companies) in the same sector they had worked at in Germany (or other countries such as Sweden or Austria). In online appendix section A, we summarize some of the evidence we’ve collected on this. As these stories show, refugees benefited from their experiences in Germany in many ways, such as applying the acquired knowledge and skills about different production methods.

Figure 3 describes the treatment variable. It plots the number of Yugoslavian workers who entered the German workforce between 1991 and 1995 (horizontal axis) and the number of those workers who remain in the labor force beyond year 2000 (vertical axis), by four-digit SITC code. All observations are below the 45 degree line, simply because the number of migrants who stay in each industry beyond the year 2000 is a subset of all those who arrived between 1991 and 1995. Thus, the treatment is the difference between each one of the observations and the 45 degree line. As can be seen in the graph, there is variation in the treatment across industries. Some of the codes that stand out as having a large number of worker returnees are 8,219 (furniture parts), 6,911 (iron and steel structures), 5,989 (chemical products), and 2,482 (worked wood of coniferous). Based on our sample (which is limited to wage earners in the tradable sector), it is noticeable that “rate of return” as measured through our filter is substantial (about 30%), but not as high as what the historical acccounts suggest as documented in official reports from international organizations (e.g., UNHCR, 2005).²⁰ The fact that some Yugoslavians stayed beyond 2000 for whatever reason is not a threat to our identification strategy, as long as these cases were not more or less frequent in some industries than in others. We expand on this in section IIID.

With the treatment defined but before we turn to the econometrics, we report whether industries associated with a larger treatment experienced better export performance in the former Yugoslavian countries upon their return. Using only raw data, figure 4 visualizes the total value of exports of industries linked to different levels of treatment (i.e., quartiles), year after year. The figure shows until 1995 (the year where our “treatment” begins), industries in the four different quartiles had somewhat parallel trends. However, after 1995, the third and fourth quartiles in terms of treatment intensity diverge quite significantly from the first two quartiles. This visualization not only provides some descriptive evidence of the results holding with raw data, but also shows that the parallel trends assumption required for the difference-in-differences methodology is a reasonable one. In any event, we address pretrend issues more thoroughly in the next section.

Table 1 presents the summary statistics. In our sample are 786 industries, and since we use two points in time for the differences-in-differences estimation, the initial empirical analysis will use up to 1,572 observations. The table presents summary statistics for the main variables in the regression. The first two lines of the table present data for the average export value from former Yugoslavian countries to the rest of the world in the years 1990 and 2005, all in million U.S. dollars (note that we do not adjust these values for inflation, which is accounted for by our year fixed effect). These two points in time are the ones used in the main specification, which correspond to years before and after the war. However, we also present results for a multiperiod analysis as well that uses export data for all the years in between.

Table 1 also summarizes the treatment. The third row in the table presents statistics for the number of sampled workers with Yugoslavian passport that joined the German labor force at some point between 1991 and 1995. The average industry had 74 Yugoslavian workers who, arguably, arrived at Germany because of the war and joined the labor force. The next row is a subset of that group and corresponds to our main treatment variable: the number of workers with Yugoslavian passport who had joined the German labor force sometime between 1991 and 1995 and had dropped from it by the year 2000. The value for this variable, averaged across all industries, is 21.6. Our treatment exploits variation across industries, which we see in the table varies from 0 to 778.²¹ All in all, our treatment is based on roughly 17,000 Yugoslavian workers across all industries, representative of the actual distribution.²²

Note that our sample of Yugoslavians employed in the tradable industry shows that the rate of return was roughly 30%, substantially lower than the anecdotical 75% figure, which applies to the Yugoslavian refugee population as a whole (UNHCR, 2005; Ruhl & Lederer, 2001; Lederer, 1997). This discrepancy, however, poses no problem for our identification strategy as long as the rate of return is not biased toward certain industries, which we discuss in section IIID. Also note that despite presenting the summary statistics in nominal values, unless otherwise stated, all right-hand-side variables are rescaled using the inverse hyperbolic sine for estimation purposes.

The variable of interest, $treatp$⁠, is the treatment as explained above, which corresponds to the number of likely returning refugees for industry $p$⁠, according to the definition detailed earlier. Given the fat-tailed distribution of this variable, we rescale it using the inverse hyperbolic sine, which behaves similar to a log transformation but has the benefit of being defined at 0. The interpretation of regression estimators in the form of the inverse hyperbolic sine is similar to the interpretation of a log-transformed variable (see MacKinnon & Magee, 1990).

We estimate this regression using two periods: “before,” which corresponds to 1990, just before the war started, and “after,” which corresponds to 2005, five years after most Yugoslavian refugees had returned.²³ Thus, $aftert$ is a dummy variable that equals 1 for the observations corresponding to the second period.

As for the other terms, $ηp$ represents industry fixed effects while $αt$ represents year fixed effects. The two fixed effects are perfectly multicollinear with the terms $treatp$ and $aftert$⁠, and so there is no need to add the interacted terms separately. $ɛp,t$ represents the error term. Our estimations cluster standard errors at the industry level, the level at which the treatment varies (Besley & Burgess, 2004; Bertrand et al., 2004).

Given the fact that the left-hand side is calculated in U.S. dollars, we are required to use a monotonic transformation to deal with the fat-tailed distribution. All of our results are presented using three different transformations: $log(exportsp,t)$⁠, $log(exportsp,t+1)$⁠, and $asinh(exportsp,t)$⁠. The first one is undefined for values where $exportsp,t=0$⁠, and therefore, when using $log(exportsp,t)$ as the dependent variable, the sample size is reduced. The two other transformations deal with the occasions where $exportsp,t=0$ by either adding US$1 before the transformation or by computing instead the inverse hyperbolic sine (⁠$asinh$⁠), respectively.²⁴ Given these monotonic log-type transformations, $βDID$ represents the elasticity of exports to returnee workers. That is, industries with a 1% larger pool of returnee workers have larger export value by $βDID$ percent larger between the before and after periods compared to industries with no returnee workers. Bear in mind that since this is a difference-in-difference setting, our results reflect relative differences in levels across industries based on their exposure to the treatment.

Our identification relies on the exogeneity of arrival and exit of refugees into and out of the German labor force with regard to industry-level export trends back in Yugoslavia. There are two main endogeneity concerns: the possibility that upon arrival, refugees self-selected into particular tradable sectors after anticipating their future postwar growth potential in Yugoslavia and the possibility that the exit of refugees from the German labor force by year 2000—even if it was enforced by across-the-board repatriation efforts by the German authorities—happened more or less frequently in particular industries in a way that is correlated with export dynamics in Yugoslavia. We address each of these possible endogeneity concerns below.

In order to deal with the possibility that migrants self-selected into particular industries in a way that correlates with future Yugoslavian exports, we construct an instrumental variable that estimates the share of asylum seekers working in each industry by exploiting a spatial dispersal policy. While asylum requests were being processed, asylum seekers were sent to different parts of the country following the Königstein State Convention (Königsteiner Staatsabkommen) which was signed in 1949 by all German federal states and defined cost-sharing rules between states in jointly financed projects. The dispersal of asylum seekers is regulated at the federal level by the Asylum Procedure Act (Asylverfahrensgesetz), where each state is allocated a certain number of asylum seekers according to its “Königstein” quota (Königsteiner Schlüssel). The quota is based on the weighted sum of population (one-third) and tax revenues (two-thirds), and is recalculated annually. For example, Nordrhein-Westfalen is the state that should have received the largest numbers of asylum seekers in 1995, followed by Bayern and Baden-Württemberg, while states such as Bremen and Saarland received a very small share.

Our identification strategy is based on the premise that this allocation was exogenous to the Yugoslavian asylum seekers entering the country and, furthermore, to the future dynamics of Yugoslavian exports. In practice, an asylum seeker into the German territory is absorbed by a reception center in the state of arrival if there is any remaining capacity to receive more people, according to the quota described above; alternatively, the person is allocated to the reception center in a different state with the most vacancies according to the quota. The residence obligation ends as soon as asylum is granted. The average duration of the application procedure was between six months and two years.²⁵

There are two main reasons for which this strong correlation between the instrument and the endogenous variable is not surprising. First, the geographic allocation of asylum seekers is relevant for all refugees who request asylum, even if the asylum request turns out not to be approved. That is, all refugees who eventually got a Duldung status but who originally requested asylum had to comply with this exogenous geographic allocation while their asylum status was being reviewed by the authorities. Second, the exogenous allocation of the share of Yugoslavians who actually requested asylum might as well be explanatory of the location choice of those who received Duldung even if they did not request asylum to begin with due to pull factors induced by migrant networks.

Our main assumption regarding the second condition—the exclusion restriction—is that both the quota of asylum seekers per state and year (defined by the German federal authorities), as well as the relative size of industries in Germany in 1990, are not correlated with future industry-specific export trends of former Yugoslavian countries to the rest of the world other than through the migrants themselves. We have no reason to think that this assumption could be violated.²⁸

The other remaining concern is that the exit of Yugoslavian Duldung holders from the German labor force was endogenous. For example, if workers left the labor force more massively in some industries than in others, in a way that correlates with their future export potential in Yugoslavia, then this could invalidate our identification strategy.

This was not the case, as shown in figure 5. The figure compares the proportion of Yugoslavians who arrived between 1991 and 1995 working in each four-digit SITC industry on the vertical axis, against the proportion of Yugoslavian who had returned by 2000 (based on the definition of the treatment) for each industry on the horizontal axis. The dashed line represents the 45 degree line. If exit from the labor force by Yugoslavians was completely random, we would see a perfect alignment of those dots along the 45 degree line: the share of workers entering each industry must be the same as the share of workers leaving it. Barring some exceptions, the graph does approximate this scenario. In fact, the correlation between both shares is close to 0.9.

Given that few observations seem to be outliers, we also compute the distance between each value in the plot and the 45 degree line and correlate it with the log-growth rate of Yugoslavian exports by industry between 1990 and 2005. We find small, negative, and statistically insignificant correlation coefficients across the board (⁠$-$0.0295 when computing the growth using logs, and $-$0.0133 when computing growth using the inverse hyperbolic sine).

Finally, as a way to relieve possible concerns due to endogeneity with respect to exit, it is important to note that our results are robust to using the number of Yugoslavian refugees workers per industry in 1995, without the exit component, which directly addresses this concern (see online appendix table C1).

Results for specification (1) are presented in table 2. The first three columns report results using an OLS estimation, while the last three columns report results using a 2SLS estimation, making use of the instrumental variable described in section IIID. The table reports results using $log(exportsp,t)$⁠, $log(exportsp,t+1)$ and $asinh(exportsp,t)$ as dependent variables. Since the regressor $treatp$ is rescaled using the inverse hyperbolic sine transformation, which behaves similar to a log transformation, we interpret $βDID$ as an elasticity.²⁹

In the first three columns, we find all estimates to be positive and statistically different from 0 for all different monotonic transformations of the dependent variable. The standard errors are clustered at the industry level, which is the level of disaggregation of the treatment.

Column 1 of table 2 presents the estimates when using the natural logarithmic transformation for the dependent variable. The point estimate in the first column is about two-thirds the size of those in the other two columns. This is not surprising as the first column excludes zeros and therefore excludes instances in which industries are more likely to grow faster if they have a non-zero value in the second period. Yet this difference says something more: the fact that the point estimates in columns 2 and 3 are positive and significant, which include instances where an industry was inexistent in the export basket of Yugoslavia by 1995, and are larger than the point estimate in column 1, implies that the effect of return migration on comparative advantage is valid at the extensive margin (e.g., opening a new line of exports) as well as at the intensive margin (e.g., growth of already existing export lines), along the lines of Bahar and Rapoport (2018). In either case, the results show that the elasticity of exports to return workers ranges from 0.09 to 0.14, depending on the transformation of the left-hand-side variable used (and, thus, on whether zeros are included).

Columns 4, 5, and 6 present the analogous 2SLS estimates (and the lower panel presents the corresponding first-stage estimations). For those columns, we also report the $F$-statistics, which measure the strength of the first stage; they are all above 14, so we can reject the possibility of weak instrumentation. The elasticities estimated through 2SLS are positive, statistically significant, and quite similar to the OLS results with point estimates that are only slightly larger. Given the standard errors, however, we cannot reject the hypothesis that the OLS and the 2SLS estimates are equal. Given the setting of the natural experiment and the use of an instrumental variable, we interpret these results as causal. Thus, based on the 2SLS results, we find that Yugoslavian industries that received 10% more return migrants from Germany (who worked in those same industries), exhibited higher exports by 1% to 1.6% during the period of the study. We show in online appendix section C.2.1 that our results are not sensitive to any particular industry.

Back-of-the-envelope calculations based on our sample size imply that returning refugees explain up to 6% of all Yugoslavian export growth between 1990 and 2005. In fact, since most of the export growth between 1990 and 2005 happens after the year 2000, most of this effect is concentrated between 2000 to 2005.³⁰

In the presence of self-selection of workers we would expect results to decrease in magnitude once we instrument. What may then explain why the 2SLS estimates are in fact slightly larger than the OLS ones? One possible explanation is that the 2SLS estimation uses variation in the treatment that is disproportionately coming from refugees allocated to areas in Western Germany (and specifically western and southern German states) where the most productive firms (and workers) are located.

Can our results be explained by a previous trend in exports? Given the availability of exports data across several years, we turn to estimate the effect of return migration in an event study setting. To account for the typical volatility of exports on a yearly basis, we perform the estimation using five-year averages for the dependent variable and estimate $βDID$ for six different periods, from 1985–1989 to 2010–2014. To do this, we simply reestimate specification (1), this time substituting the dummy $aftert$ for several dummies, each one signaling a time period, along the lines of Autor et al. (2003). We define 1990 to 1994 as the base period, and therefore the other point estimates are relative to it.

The estimation (using 2SLS) is summarized in figure 6, which shows in the upper panel the evolution of the expected value of exports (across our three different measures) by periods for two groups of industries: those for which the value of $treatp$ equals 0, and those for which $treatp$ equals 1 standard deviation of the treatment (denoted by $σ$⁠). The figures in the lower panel show the difference between the two groups, and it can be seen how the effect becomes positive and statistically significant starting in the period where the refugees start returning, 1995 to 1999. Note that based on the standard errors (as measured by the whiskers representing 95% confidence intervals), we cannot reject the hypothesis that the trends for both groups in periods before 1995 are statistically the same.

These findings suggest an important result: the marginal effect of return migration on the emergence of new exports becomes stronger over time. The figure also reveals that the average industry experiences a drop in export performance during the war as compared to the base period. Treated industries recover and reach back their 1990–1994 level in the 2000 to 2004 period (or even earlier, when using the log transformation), while nontreated ones recover later. In that sense, part of what our effect is capturing in the first few postconflict years is that treated industries, on average, shrunk less than nontreated ones.³¹

Our main strategy compares different industry lines based on the intensity of the treatment. We complement those results with implementing the synthetic control method (Abadie & Gardeazabal, 2003; Abadie et al., 2010, 2015) in the context of our empirical strategy. In other words, we create a synthetic industry for every Yugoslavian industry that mimics export dynamics until 1995 (the year our treatment starts kicking in), based on over one hundred other countries. We use five main predictors to construct each synthetic industry, using data for the period 1986 to 1995: export value of the industry itself (only for years 1990, 1993, and 1995), its export share, and the export share of its corresponding one-digit SITC, as well as GDP per capita, and population size. We then apply the synth algorithm that computes a synthetic control for each industry based on a weighted combination of countries other than Yugoslavia (see Abadie et al., 2011).

Figure 7 visualizes the main results of this exercise, plotting the difference between the export value of the real and synthetic industries. The top panel shows the evolution of each one of these differences: until 1995, there is almost no difference between the real and synthetic industries, but after 1995, we see all types of dynamics. The bottom panel groups industries in two, averaging the difference between their real and synthetic export values: low treatment and high treatment, which corresponds to the first and fourth quartiles of the distribution of $treatp$ as defined above (analogous to figure 4). The plot shows how, on average, real Yugoslavian industries outperform the synthetic ones, particularly among the high-treatment industries. Results using regression analysis are consistent and presented in online appendix C.3.

There are a number of possible candidate explanations for our results, which we explore in this section.

Our analysis that follows consistently suggests that the main driver of our results are productivity shifts caused by the inflows of returning workers, who were exposed to better practices and technologies while in Germany. This is supported by a number of tests, which we expand on below, as well as those in section VI, where we show that the workers driving our results are those with characteristics or occupations more likely to transfer knowledge, technology, best practices, and more that have the potential to shift productivity.

First, we note that our results are consistent when using an industry-level measure of comparative advantage, which in theory responds to productivity parameters. In practice, we reestimate our specification using as the dependent variable a Ricardian comparative advantage parameter for Yugoslavia (⁠$Φp,t$⁠), estimated following Costinot et al. (2012) and Leromain and Orefice (2014).³² Analogous to our baseline results, we use the same two different monotonic transformations for $Φp,t$ (we skip the $log(Φp,t+1)$ transformation since there are no zeros). Results are presented in table 3.

We replicate our main specification using production data from UNIDO for sectors in Yugoslavian countries (consistent with the main results, we aggregate all countries into a single one). We focus on outcomes such as output (production) per worker, value added per worker, and average wage, all in U.S. dollars. As opposed to the exports data used above, having the number of employees per sector allows us to compute outputs relative to the size of the sector. The data have complete information for only eighteen two-digit ISIC sectors, and as such we are limited to a low number of observations, which we acknowledge is a limitation of this test. We merge these data to the treatment variable, which was a trivial task given the one-to-one correspondence between ISIC/NACE and W93 sector classifications.

Results of OLS estimations based on the main specification are presented in table 4, panel A. Columns 1 and 2 use as dependent variable measures of labor productivity, output (production) per worker and value added per worker, while column 3 uses as dependent variable the average wage of the sector, which also proxies for labor productivity of the sector as a whole. The results are consistent with the idea that productivity shifts are a central mechanism behind our main findings. Sectors with 10% more returnees tend to be 0.9% to 1.1% more productive depending on the measure used.

We were able to collect and digitize a number of sector-level indicators for the Republic of Bosnia and Herzegovina (the country with the largest number of refugees hosted by Germany) before and after the war. In particular, we gather sectoral data for export values, number of firms, and number of workers for the years 1990 and 2010 (the latter being the first postwar year for which such data are available). The sources of our collection are the Statistical Almanac (Statisticki Godisnjak) and the Structural Business Statistics reports (Strukturne Poslovne Statistike) published by the Statistical Agency of Bosnia and Herzegovina (Agencija za Statistiku BiH).

Given that these data are particularly for Bosnia and Herzegovina, we use our main data source to construct variations of the treatment meant to capture more precisely the number of Bosnian returning refugees (as opposed to Yugoslavians in general). In particular, we count workers who had been registered as having a Bosnian passport at some point during their stay in Germany and who had arrived sometime between 1991 and 1995 and had left by 2000. Note that while some of these workers entered the German labor force as Bosnian nationals, others obtained Bosnian citizenship during their stay in Germany and reported it.³³ We link this treatment to the outcomes described by sector relying on the equivalence between the German classification WZ 93 (the format for the treatment data) and the two-digit NACE sectoral classification (the format for the Bosnian data we collected). We end up for with a sample of 22 sectors and 44 observations (using data for each sector in 1990 and 2005). Similar to the previous results, we acknowledge that the small number of observations poses a limitation for these particular results.

Consistent with previous estimations, we transform all the variables using the inverse hyperbolic sine and based our estimations on specification (1). Table 4, panel B, presents the OLS results using the treatment variable described above. Column 1 presents estimates when using Bosnian exports per worker as the dependent variable (in U.S. dollars), while columns 2 and 3 estimate the effect of the treatment on the number of firms and the number of workers per sector, respectively. The point estimates are positive and significant, implying that a higher number of returnees translates into higher labor productivity, as proxied by exports per worker, as well as sectors that are larger in terms of both firms and workers. We note that the point estimates are particularly high compared to our baseline results, but since these results are for Bosnia only, and not for Yugoslavia as a whole, a comparison is possible to only a limited extent. Moreover, a simple look at the coefficients suggest that on average, the treated sectors end up having firms that are smaller in terms of workers, which is typically the case for new firms. Consistently, there is rich anecdotal evidence on the creation of new firms by returnees, some of which we compile in online appendix section A.

We go through a number of alternative explanations of our results, other than productivity shifts, that we are able to rule out, to the best of our ability. It is important to mention, however, that for any variable to be an alternative explanation in our empirical setting, it must comply with one important condition: that it can systematically explain cross-industry variation in export dynamics in ways that also correlate with the treatment variable (or in 2SLS specifications with our instrument). To the extent that alternative explanations do not comply with this condition, accounting for them won’t result in results that are very different from the main ones.

There are four alternative explanations that we rule out:

1. Scale effects: Workers returning simply are able to produce—and thus export—more quantity. For more details on how we address these concerns, see online appendix section C.4.1.

2. Convergence in structural transformation: Our effects are being driven by pure “convergence” in terms of structural transformation such that Yugoslavia’s productive structure in the late 2000s reflects that of Germany in the early 1990s. This concern is addressed in online appendix section C.4.2. Note that the results using the synthetic controls presented in section IVC address this concern too.

3. Capital and investment: Returning refugees are going into industries that are experiencing capital inflows, simultaneously, and our results are indeed explained by the latter. We address these concerns in online appendix section C.4.3.

4. Bilateral transaction trade costs: Our results are being driven by possible reductions of transaction costs of exporting to Germany caused by migrant networks. For more details on how we address this concern, see online appendix section C.4.4.

The summary of our results is presented in table 5.³⁵ Each column presents results using a different monotonic transformation of the dependent variable, consistent with all previous results in the paper. Columns show the estimated value of $βiDID$ for each of the constructed treatments belonging to each of the categories described above (we only present results using OLS, as we don’t have instruments for more than one endogenous variable at a time).

The first row replicates the main results using the total number of returnees per industry, for comparison purposes. Overall, based on the size of point estimates, results reveal that our findings are particularly driven by workers with higher educational attainment (skilled), workers in occupations that are intensive in analytical tasks (as opposed to manual ones), workers in skilled occupations (as opposed to unskilled ones), and workers in supervisory occupations. We also find that the results are particularly driven by workers who worked in the top-paying firms during their stay in Germany and find that workers for whom wages grew faster during their stay in Germany correlate with a higher export performance. Note that while some of these results are statistically significant (namely, skilled workers, occupations intensive in analytical and cognitive tasks, and workers in high-paying firms), while the other point estimates are less precise.

Note, however, that the point estimates are to be interpreted in terms of elasticity and thus ultimately, the marginal effect of one worker belonging to each of the categories driving the results is much larger than what the point estimate suggests. This is because the types of workers driving the results are a smaller share of all workers when looking at the within-industry composition of workers in the sample. The idea that a small number of workers can have such an important effect on exports of a whole industry in such little time might seem implausible at first, but some anecdotal evidence documented by others seems to strongly support that idea. For instance, Rhee and Belot (1990) and Easterly (2001) document the story of the success of the garment sector in Bangladesh. Between 1980 and 1986, the share of garments in Bangladesh’s total exports rose from 0.5% to 28.3%. The unprecedented take-off of the garment export sector is often attributed to 130 Bangladeshi workers—only four of them in management positions—who spent eight months in 1979 working and being trained in Korea as part of an agreement between their company, Desh of Bangladesh, and the Korean firm Daewoo. The know-how acquired by these workers seems to have been crucial in making Desh a highly successful exporter firm. Perhaps more important, such know-how eventually spilled over as workers moved to other firms or created new ones, contributing to the massive success of garment exports as one of Bangladesh’s most significant export sectors.

In this context, we believe our findings pointing to productive knowledge and managerial know-how as the main mechanisms driving the export dynamics, as well as the magnitudes of the coefficients we report, are aligned with other studies in the literature.

The Balkan wars of the early 1990s created massive forced displacement from and within the former Yugoslavia. Most internationally displaced refugees ended up in Germany, where they could work under temporary protection status. A majority of them eventually returned home after the Dayton peace agreement of December 1995 and the repatriations that followed. We exploit this natural experiment that resulted in exposure to German industrial know-how, technology, and best practices, to investigate the role of returning refugees in explaining the export performance of their home countries. Using confidential German social security data, we find that Yugoslavian exports performed significantly better during the postwar period in industries that returnees had worked in while in Germany. Furthermore, we find that productivity is the underlying mechanism driving export performance in our setting. This is backed by the fact that our results are particularly driven by returnees with characteristics and in occupations more likely to generate productivity improvements.

Our results contribute to a burgeoning literature that emphasizes that migrants can serve as drivers of the diffusion of knowledge, technologies, and best practices resulting in productivity shifts. To the best of our knowledge, we are first to find such evidence using a natural experiment, especially in a context of returning forced migration.

In terms of policy implications, our results speak to the importance of allowing refugees full labor integration in their receiving countries. This is not only for the obvious reasons of the benefits to local hosting communities, but also because of how such policy can be a crucial determinant of the reconstruction of their home countries upon their eventual return.

More generally, the ability of a worker to become more productive has to do with his or her accumulated experience and his or her ability to learn from others while on the job. Migration therefore is an important vehicle in the process of knowledge transfers across locations.³⁶

Better understanding this process and identifying channels through which these dynamics occur, such as the policy and institutional environments that facilitate the diffusion of knowledge through migration, are important missing pieces in the literature and an active part of our future research agenda.