Policy Research Working Paper 11000 Do Bilateral Labor Agreements Increase Migration? Global Evidence from 1960 to 2020 Samik Adhikari Narcisse Cha’ngom Heidi Kaila Maheshwor Shrestha Social Protection Department & A verified reproducibility package for this paper is Development Economics available at http://reproducibility.worldbank.org, December 2024 click here for direct access. Policy Research Working Paper 11000 Abstract This paper estimates the impact of bilateral labor arrange- a pre-existing regular flow and for destinations in the Gulf ments on migration between two countries. It uses Cooperation Council. In contrast, the effect is virtually comprehensive data on bilateral migration and bilateral absent for origin countries in Africa, driven by countries labor agreements across all country pairs for each decade with weak government effectiveness. The estimates imply from 1960 to 2020, and employs an empirical specification that bilateral labor agreements can lead to substantial wel- with a rich set of fixed effects. In the preferred and most fare gains: low- and lower-middle-income countries can stringent specification, the findings show that signing a earn an additional US$120 million annually from a bilat- bilateral labor agreement increases migration from an origin eral labor agreement. If countries in Sub-Saharan Africa country to a destination country by 76 percent (0.57 log were to experience similar impacts, the welfare gain from a points) in the decade of signing. The effect persists for up to single BLA could be as high as US$51 million per year for three decades. The impacts are higher for corridors without these origin countries. This paper is a product of the Social Protection Department and Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at mshrestha1@worldbank.org. A verified reproducibility package for this paper is available at http:// reproducibility.worldbank.org, click here for direct access. RESEA CY LI R CH PO TRANSPARENT ANALYSIS S W R R E O KI P NG PA The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Do Bilateral Labor Agreements Increase Migration? Global Evidence from 1960 to 2020∗ Samik Adhikari, Narcisse Cha’ngom, Heidi Kaila, and Maheshwor Shrestha Keywords : Bilateral Labor Agreements, BLAs, migration, development JEL: F22, J08, O15 ∗ This paper serves as a background paper for the Africa Regional Report on Migration (Abdel Jelil et al., 2024). The authors would like to thank Andrew Dabalen for guidance and Quy-Toan Do, David McKenzie, Mauro Testaverde, and Soonhwa Yi for helpful discussions and feedback. 1 Introduction International migration has the potential to generate trillions of dollars in global economic gains, yet numerous natural and policy barriers prevent people from fully capturing these bene- fits (Clemens, Montenegro and Pritchett, 2019). Consistent with the welfare loss due to limited migration, studies such as Gibson et al., 2018, McKenzie, Stillman and Gibson, 2010, Mobarak, Sharif and Shrestha, 2023 show that the earnings of migrants can be 3-5 times higher than their counterfactual incomes had they not migrated. Factors such as lack of information, misinforma- tion, liquidity constraints, the financial costs of migration, psychological hurdles, and restrictive policies all serve to limit mobility (see McKenzie, 2023, for a recent review). Understanding these barriers is crucial to identifying strategies that can unlock the full economic potential of migration. This paper contributes to this understanding by examining how the relaxation of legal barriers – through bilateral labor agreements (BLAs) between two countries – affects migration flows. BLAs are international agreements to regulate the flow of workers between destination and origin countries (Chilton and Posner, 2018, Saez, 2013). Broadly, BLAs set the terms and con- ditions under which migrant workers from origin countries enter and exit destination countries, outline the responsibilities of the destination countries in receiving and hosting migrant workers for the duration of their stay, and allow the two sets of countries to store and share information and resolve disputes (Chilton and Woda, 2022). The early BLAs date as far back as 1893 when South Africa and Portuguese officials governing Mozambique signed an agreement for migra- tion of mine workers (Peters, 2019b). BLAs to manage the flow of migrant workers accelerated in the mid-20th century to fill the post-World War II demand for labor in Europe and North America in sectors like agriculture and construction (Chilton and Posner, 2018, Peters, 2019b, Saez, 2013). Outside destinations in North America and Europe, the use of BLAs only started getting momentum in late twentieth century, particularly among destination countries in the Gulf Cooperation Council (GCC) and origin countries in South and East Asia (Chilton and Woda, 2022). Countries usually enter BLAs to regulate seasonal and temporary labor migration for low- skilled workers, though BLAs can also facilitate permanent migration (Chilton and Woda, 2022, United Nations, 2022). Many BLAs specify skill requirements for destination countries, 2 enabling targeted recruitment of workers with the requisite skills, and include provisions for worker rights, such as wages, working conditions, and access to services (Biehler, Kipp and Koch, 2024, Chilton and Woda, 2022). While BLAs relax some regulatory barriers, they do not address all informational or financial barriers that limit broader migration. Though conceived as a tool to facilitate migration, it is not obvious that BLAs necessarily increase migration for various reasons. First, significant numbers of migrants move between countries outside the framework of BLAs and BLAs could be designed to restrict or control the flow of migrant workers (Livnat and Shamir, 2022). Second, instead of increasing net migration flows, BLAs could be a policy tool to regularize the irregular migration already taking place (Bither and Ziebarth, 2018, Clemens, 2024). Third, even if BLAs may be intended to increase migration, the efficacy of such policies – like with many policies – is not obvious ex-ante. Such concerns may be particularly strong in settings with high costs of migration, high extent of misinformation, and low actual or perceived returns to migration as well as in settings where low capacity and monitoring limits the efficacy of BLA implementation.1 Finally, even if BLAs were effective in facilitating migration, it is not clear how effective they can be as a policy tool. Determining the magnitude of migration impacts and its persistence is important for migration policies. To understand the impact of BLAs on migration, we combine two comprehensive data sources on migration and BLAs across all country pairs. The global bilateral migration data captures the migration (measured by people living in a country that is different from the country of birth) between all country pairs for each decade from 1960 to 2020, compiled by the World Bank and the United National Department of Economic and Social Affairs (UNDESA, 2020, World Bank, 2022). A comprehensive dataset on all BLAs for the same period comes from Chilton and Posner (2018), Peters (2019a) and Chilton and Woda (2022). We merge the two sets of databases by converting the BLA database into country-pair × decade format for analysis. For each of the BLAs, we designate the country with a lower income per-capita as the origin country and the other as the destination country. 1 The enforcement of BLAs is often measured through reports from international agencies such as the International Labor Organization, labor inspection data from national governments, monitoring of complaints from migrant workers and worker welfare organizations, or academic research studying the impact of BLAs in specific corridors (United Nations (2022)). However, factors such as lack of data, limited monitoring capacity of national governments, under-reporting of issues by migrant workers, and difficulties in cross-border mechanisms of enforcement can make monitoring effective implementation of BLAs challenging (Chilton and Woda, 2022, United Nations, 2022). 3 We tease out the impact of the BLAs on migration after imposing a series of fixed effects to control for the various confounds. Our most stringent, and preferred, specification controls for corridor (origin country × destination country) fixed effects, origin × decade, and destination × decade fixed effects. That is, if country A signs a BLA with country B , the identifying variation comes from comparing migration within the AB corridor before and after the signage (corridor fixed effects); from comparing migration from country B to country A relative to all potential destinations for migrants in country B (origin-time fixed effects); and from comparing migration to country A from country B relative to all potential origin countries for migrants to country A (destination-time fixed effects). These extensive sets of fixed effects ensure that some corridor-specific factors (e.g., historic bilateral relationships between two countries), or country-time-specific factors that can affect migrant labor demand or supply (such as economic cycles or the labor market situation in origin or destination countries) do not drive the relationship and that the impact can be attributed to the BLA itself. One of the key assumptions for the identification of the impact of the BLAs is that of parallel trends. We employ an event-study design to check for any pre-trends on migration prior to the signing of the BLAs (in addition to using this to explore persistence of effects).2 However, as migration data is measured every ten years, the pre-trend test cannot test for anticipatory BLA signing in response to, say, temporary shocks within the decade of signing. We find that signing a BLA increases migration from the origin country to the destination country by 76 percent (0.57 log points) in the decade of signing the BLA. The migration effect lasts up to three decades after signing a BLA. The impact of BLAs is muted at 25 percent (0.24 log points) for ‘regular’ corridors - corridors with at least 10 migrants in all periods.3 The estimates are of similar magnitudes with alternative sets of fixed effects, and slightly larger with an instrumental variable specification where we use ‘leave-out’ mean as the instrument. These estimates can be considered as a lower-bound of the true impact of a BLA on facilitating migration as the data on migration only captures the stock of migration measured every ten years and misses temporary migration that might have happened between the periods. We also 2 We also test the assumption of no pre-trends using a heterogeneity-robust estimator applied to gravity settings following the recent advancements in the difference-in-difference methodologies (as developed in, Nagengast and Yotov, Forthcoming, Nagengast, Rios-Avila and Yotov, 2024) and fail to reject the null of no pre-trends. 3 The results are robust to alternative thresholds for defining which corridors count as ‘regular’. 4 find that the impact on migration persists for up to three decades after the signing of the BLA. Reassuringly for the identifying assumption, we do not find any impacts (or trends) prior to the signing period. We also uncover interesting and relevant heterogeneities in the migration impacts of the BLAs. The migration impacts are significantly higher at 834 percent (2.24 log points) if the BLA involves a destination country in the Gulf-Cooperation Council (GCC). Likewise, the migration impacts are significantly higher for low- and lower-middle income origin countries (311 percent, 1.4 log points). However, the impacts are virtually absent, at an insignificant 15 percent (0.14 log points), for African origin countries. We show that the lack of impact among African origin countries is driven by countries with weaker institutions as proxied by measures of government effectiveness. This underscores the importance of strong institutions and systems for translating these agreements into actual migration. Our estimates imply large welfare gains from signing a BLA to the migrants and their families. The empirical literature finds that a low-skilled worker can experience significant earnings gain, ranging from a doubling to a quadrupling of their incomes, through migration. Using the migration impact from our study, the average GDP per-capita as a proxy of their incomes at home, and the returns to migration from the literature, we find large earnings gain from such agreements, particularly for low- and middle-income (LLMIC) countries. The welfare gain - measured by increased migrant earnings - from an agreement can be US$120 million annually for LLMICs. Gains are even higher, at US$2.1 billion annually for agreements with the GCC destinations. The persistence of the migration impacts for multiple decades suggest that the gains accrue over time to US$1.6 billion for agreements with an average destination country and US$28 billion for agreements with GCC destinations. Gains are much lower for countries in Sub-Saharan Africa due to low volume of migration as well as lower impacts of BLAs, with US$2.4 million annually (US$32 million over time). If countries in Sub-Saharan Africa were to have similar migration impacts as the average impact for LLMICs – say, through strengthened capacity to implement such agreements – gains would increase to US$51 million annually (US$676 million over time) to an average destination and US$413 million annually (US$5.5 billion over time) to a GCC destination country. This paper contributes to the literature on the study of BLAs. Most literature examining the impact of labor agreements has focused on specific corridors or regions (Arpaia et al., 5 2018, Clemens, Lewis and Postel, 2018, Mobarak, Sharif and Shrestha, 2023). Information on country or region-specific bilateral or multilateral agreements is often more readily available whereas aggregate global analysis on the impact of labor agreements has been limited due to the lack of global datasets documenting these bilateral and multilateral treaties. This has changed recently with the emergence of large datasets on BLAs (Chilton and Woda, 2022, Chilton and Posner, 2018, Peters, 2019b). Using these datasets, new and emerging research has examined the relationship between BLAs and trade flows between countries (Maximova and Paraschiv, 2022), the relationship between cross-border property rights and innovation (Bian, Meier and Xu, 2023) and emigration and democratization (Peters and Miller, 2022). However, to our knowledge, no paper has so far documented the impact of BLAs on migration flows between countries. Both Peters (2019b) and Chilton and Woda (2022) find positive association but do not document the impact rigorously. This paper relates to literature on how explicit policies related to migration affect migration. A few papers examine the role of policies such as border walls, minimum wage, passport and other legal hurdles in facilitating or hindering migration (Allen, de Castro Dobbin and Morten, 2018, Feigenberg, 2020, McKenzie, 2007, McKenzie, Theoharides and Yang, 2014). Yet, another set of papers study the impact of specific migration restriction programs (Clemens, Lewis and Postel, 2018, Di Iasio and Wahba, 2023, Mayda et al., 2018). In a similar vein, Ortega and Peri, 2013 examine the relative effects of income and a broader set of immigration policies on immigration to the OECD countries and find that restrictive policies can rapidly reduce migration. While the literature has largely examined the role of restrictive immigration policies, this paper examines the impact of a migration facilitation policy. The paper also adds to the large literature on relieving constraints to migration more broadly. Studies have reconciled the high returns to migration with the low observed rates of migration as stemming from high (utility) cost of migration (Bryan and Morten, 2019, Imbert and Papp, 2020). Other studies have examined the role of reducing various dimensions of costs, such as informational frictions, insurance failure, social networks, risk aversion, difficulty to learn about risks, liquidity, and financial costs of migration (see Angelucci, 2015, Bah et al., 2023, Baseler, 2023, Bazzi, 2017, Bazzi et al., 2021, Beam, McKenzie and Yang, 2016, Bryan, Chowdhury and Mobarak, 2014, Cai, 2020, Dao et al., 2018, Gazeaud, Mvukiyehe and Sterck, 2023, Munshi, 2020, Munshi and Rosenzweig, 2016, Shrestha, 2019, 2020, for key studies on these). This 6 paper complements the literature by providing a better understanding of the role of policies in reducing migration barriers. The remainder of the paper is organized as follows: section 2 describes the data sources, section 3 describes the empirical specification, section 4 describes the results, section 5 presents the potential welfare impacts of the BLAs, and section 6 concludes. 2 Data 2.1 Data on migration This paper uses the most comprehensive data set on bilateral migration stocks spanning from 1960 to 2020, in ten-year intervals. Migration stocks for years 1960, 1970, 1980, 1990, and 2000 ¨ are taken from World Bank (2022) (see Ozden et al., 2011, for methodological details). The data for 2010 and 2020 come from UNDESA (2020) which employs similar methodology. These data sets use information collated from censuses and surveys at destination and origin countries to construct a matrix of bilateral migration stock data from each of the origin countries to each potential destination country. This combination of datasets is used widely by many research studies and reports (including in Clemens, 2020, Shrestha, 2023, World Bank, 2023). A migrant, in these data sets, is typically a person who is born in a country that is different from the country in which they are currently living. To that extent, it includes, in principle, people who have migrated for various reasons and channels, including those who migrated ¨ through irregular channels (Ozden et al., 2011). However, it is possible that these data sources may misreport irregular migrants for various reasons. To the extent such misreporting exists, measurement errors can mute the estimated relationship. In addition, this dataset only captures the stock of migrants and does not capture the flow of people between these countries. Migrants who move temporarily, particularly between the census years, will not be fully captured by this dataset. If BLAs increase temporary migration, then the full extent of the impact on migration will not be captured by this dataset. Hence, we interpret our estimates as a lower bound of the migration impacts. 7 2.2 Data on BLAs The data we use for the BLAs is from Chilton and Woda (2022). This dataset contains a com- prehensive set of 1,222 BLAs between two countries since World War II. The dataset contains information on the year the BLA was signed and the two countries between which the treaty was signed. The dataset is the most comprehensive dataset on BLAs to date and builds on the data compiled by Chilton and Posner (2018), Peters (2019a) and Peters (2019b). A richer description of the dataset is presented in Appendix A. We transform the dataset to match the format of the migration dataset. That is, for each BLA in this dataset, we identify the country of origin, the country of destination, and assign it to the subsequent year in migration data.4 That is, a BLA signed between 1970 to 1979 is assigned to t = 1980. To simplify categorization, we label the country with a lower GDP per-capita as the origin country and the other as the destination country. The origin-destination-year cells without a BLA is assumed to have no BLA. 2.3 Descriptive statistics The combined dataset used in this paper has information on BLAs and migration from 198 countries to 197 potential destinations (i.e., 39,006 origin-destination corridors) for 7 periods from 1960 to 2020 at ten-year intervals. A common feature of the bilateral migration dataset, as seen in table 1, is that most (61 percent) of the corridors do not have any migration. About 12 percent of the corridors have at least 10 observed migrants across all time periods; we term these as regular corridors. The shares of non-empty corridors and regular corridors are similar across the key subsets used in the analysis. Likewise, BLAs are also relatively rare in the data. About 1.6 percent of corridors have a BLA signed between them, with the share lower for African origin countries (0.80 percent) and higher for GCC destinations (3.4 percent). BLAs are more common in regular corridors with close to 9 percent of the regular corridors having a BLA. Similar pattern holds across the various sub-samples used in the analysis. Of the 261 million migrants in our combined dataset in 2020, 82 million (31 percent) are in corridors with a BLA. The proportion of migration accounted by corridors with a BLA is 4 To address the issues of countries and names that changed over time, we use the same principles and ¨ coding applied in the migration data (Ozden et al., 2011). 8 slightly lower for origin countries in Africa and in low- and lower-middle income countries, and higher for GCC destination countries. The proportions are similar across time as well (table B.1). Table 1: Structure of migration and BLA datasets Full sample African origins LLMIC UMHIC GCC destinations (1) (2) (3) (4) (5) Bilateral migration data Share of non empty corridors 0.386 0.352 0.370 0.397 0.353 Share of regular corridors 0.122 0.091 0.103 0.135 0.088 Bilateral labor agreement data Share of corridors with BLA 0.016 0.008 0.011 0.019 0.034 Share of regular corridors with BLA 0.089 0.055 0.063 0.103 0.192 Migrant stock (in million) in 2020 All corridors 261.1 36.8 119.8 141.3 29.8 Regular corridors 242.0 33.4 108.7 133.3 27.3 Corridors with ever a BLA 81.8 8.1 34.9 46.9 16.5 Notes : The sample includes 39,006 origin-destination corridors for 7 periods from 1960 to 2020 at ten-year inter- vals. Regular corridors are corridors with at least 10 migrants in the corridor for each of the periods in the data. 3 Empirical specification 3.1 Baseline analysis To investigate the impact of signing Bilateral Labor Agreement (BLAs) on migration flows, we rely on a structural gravity model for migration derived from the Random Utility Framework andez-Huertas Moraga, 2016). This framework posits that migration (Beine, Bertoli and Fern´ flows between two locations i (origin) and j (destination) depend on the relative attractiveness of j over i net of the migration cost between i and j. Consequently, the implied baseline empirical equation to be tested is written as: Mijt = exp [αit + αjt + αij + γBLAijt ] ϵijt (1) where Mijt is the migration stock from country i to country j measured at time t ∈ {1960, 1970, 1980, 1990, 2000, 2010, 2020}. BLAijt indicates whether i and j have signed a BLA at any time in the decade preceding t. That is, BLAij 2000 indicates whether country pair ij signed a BLA between 1991 and 2000. The αit , αjt , and αij represent the origin-time, destination- 9 time, and origin × destination corridor fixed effects respectively. Origin and destination-time fixed effects capture origin and destination specific and time varying factors likely to affect migration flows (push and pull factors) including multilateral resistance to migration (Bertoli and Moraga, 2013). Likewise, the corridor fixed effects capture the time-invariant historical relationship between the two countries that affect migration flow. The ϵijt is the error term. The formulation in Equation (1) lends directly to our key empirical specification: ln Mijt = αit + αjt + αij + γBLAijt + εijt (2) where εijt = ln ϵijt is the idiosyncratic error term correlated within each corridor. This rich set of fixed effects makes sure that the only source of variation in our estimation is made of ijt factors. We conduct robustness checks that control for Xijt which measures the cumulative number of BLAs signed by the country pair prior to the decade. The specific nature of the migration data presents some issues for estimation. In particu- lar, a large fraction of the corridors in our data have zero migration. About 61 percent of the 273,042 ijt cells have zero migration.5 These corridors are essential for estimation as BLAs may be particularly effective (or ineffective) in influencing migration in those corridors. There are two simple transformations to get around this issue, adding 1 to M prior to the log trans- formation or using the inverse hyperbolic sine transformation which has similar interpretation as the logarithmic transformation (Burbidge, Magee and Robb, 1988, MacKinnon and Magee, 1990, Pence, 2006). Each method has some limitations (Bellemare and Wichman, 2020), but we confirm that results are qualitatively and quantitatively similar with both methods. For simplicity, we present the results with inverse hyperbolic sine transformation. Alternative way to get around the large number of zeros is to estimate Equation (1) with Poisson Pseudo Maximum Likelihood (PPML) methods. However, PPML estimated in levels, tends to over-weight large corridors (Orefice, 2015) which is an issue in our data. For instance, the top 10 corridors (of 39,006 corridors each year) represent somewhere between 14 and 34 percent of all migration in each year; the top 20 corridors represent 22 to 45 percent of all mi- gration. That is, large corridors such as Mexico to the United States, Pakistan to India, India to Pakistan, Poland to Germany, and India to the United Arab Emirates will get disproportionate 5 The number of cells represents data on 198 origin countries, 197 potential destinations, and 7 time periods. 10 weights in PPML estimation. To address these concerns, we present OLS estimations of Equation (2) for the full sample as well as for a sub-sample of about 33,000 regular corridors which we define as corridors with at least 10 migrants in all periods. As a robustness exercise, we also present results of PPML estimations in both samples. We also conduct robustness checks with PPML estimates excluding the top 1% migration flows. Even though the rich set of fixed effects control for a wide array of confounds, there may be outstanding concerns of endogeneity. Specifically, omitted variables such as corridor-specific change in broader policies that directly affect both the BLA and migration, historical BLAs affecting additional BLAs and migration flows. Inclusion of Xijt , the cumulative BLAs signed in prior decades address some, particularly the latter, but not all of these concerns. To address this, we employ an instrumental variable approach. Our instrumental variable approach builds on the “leave-one-out ” approach taken from Bryan and Morten (2019). That is, we use the total number of BLAs signed by the destination country j with all countries other than i (denoted by ⌝i) to serve as an instrument for the BLAs between country-pair ij . That is, the total number of BLAs, {BLA⌝ij,t }, serves as an instrument for BLAij,t . We also use a complementary instrumental variable where the instrumental variable is the number of BLAs signed by the origin country with all other destination ({BLAi⌝j,t }) in that period. So far, we have considered a setting in which the mean estimated effect is identical across different groups. However, these groups are likely heterogeneous and we should expect hetero- geneous migration responses as well. To dig into that, we further perform heterogeneity analysis as follows: ln Mijt = αit + αjt + αij + γBLAijt + ηBLAijt × X + εijt (3) where X are dimensions of heterogeneity. Key dimensions of heterogeneity include the income level of the origin country, whether the origin country is in Africa or not, and whether the destination country is in the Gulf Cooperation Council (GCC) or not. Heterogeneity along a few other dimensions is presented in the appendix. 11 3.2 Event study design Finally, we explore the persistence of the impact of signing a BLA using an event-study like design. We transform the data around each event - signing of a BLA between the country pair - so that time is coded relative to the decade in which the BLA was signed.6 The estimating equation is: 5 ln (Mijt ) = αit + αjt + αij + γe × Dijt−e + ωijt (4) e=−5 where Dijt−e indicates that a BLA was signed by the country-pair ij exactly e periods prior to time t. That is, the coefficients γe denotes the effect of a BLA e periods after the signing. Positive values of e, that is, {γ0 , · · · , γ5 } capture the persistence of the impacts of a single BLA. Negative values, that is {γ−5 , · · · , γ−2 }, indicate pre-trends in outcome in prior period. en, 2019, Miller, Significant pre-trends suggest potential mis-specification (Lovenheim and Will´ 2023). The coefficients are normalized to period −1, the decade prior to the signing of the BLA. 4 Results 4.1 Benchmark estimates In this section, we report the estimate of the overall average effect of BLAs on migration flows. In the next section, we show that this benchmark estimate hides important and interesting het- erogeneity along several dimensions including the destination type, the region of origin and the development level of origin countries. Then after, we investigate the dynamic of the estimated average effect in an event study design. Table 2 reports our benchmark estimate of the migration response to the signing of a bilateral labor agreement (BLA). We consider three sets of specifications: a first set that includes only origin-destination pair and decade fixed effects, a second set that adds the origin-decade fixed effect to the previous one, and the last set that includes origin-destination pair, origin-decade, and destination-decade fixed effects. Given the existence of empty corridors, we apply the inverse 6 For corridors with multiple BLAs, we duplicate the time series for the corridors for each of the BLAs and code the time period relative to a single BLA. We then weight our regression by the inverse of the number of duplications (i.e., the number of BLAs signed by a corridor). 12 hyperbolic sine transformation to the migration numbers so that we can approximately interpret the estimated coefficients as percentage changes. Although our preferred specification is the one with three-way fixed effects, we find an overall positive and significant effect of signing a BLA on migration flows regardless of the specification considered. As reported in column (3) of table 2, signing an additional BLA increases migration by about 77 percent (0.569 log point) from the origin country to the destination country within a decade of signing the BLA. However, one might be concerned that corridors with small numbers of migrants might be driving the results, as a marginal change in the number of migrants in these corridors over time will translate into a higher percentage change, while overall being meaningless for the overall migration pattern. To address this concern, we restrict our analysis to what we call “regular corridors,” which we capture as corridors with at least 10 migrants over the period of interest (1960-2020). The results are reported in columns (3) to (6) with the same sets of fixed effects as described above. Our preferred specification (column 6) confirms the positive effect of BLA on migration flows. Indeed, entering into a new BLA increases migration in regular corridors from the origin country to the destination country by 27.5 percent (0.243 log point) within the decade of signing the BLA. Table 2: Migration response to BLAs (Benchmark) Full sample Regular corrdiors (Mij ≥ 10) (1) (2) (3) (4) (5) (6) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ Bilateral labor agreement (BLA) 0.593 0.570 0.569 0.398 0.306 0.243∗∗∗ (0.106) (0.103) (0.090) (0.059) (0.055) (0.051) Observations 273,042 273,042 273,042 33,306 33,257 33,222 Origin country x decade FE No Yes Yes No Yes Yes Destination country x decade FE No No Yes No No Yes Origin x destination countries FE Yes Yes Yes Yes Yes Yes Decade FE Yes Yes Migrants per decade 154,063,957 144,293,527 Number of BLAs per decade 146 111 Notes : The table presents the estimate of γ from OLS estimate of equation (2). It shows the impact of a BLA on migration. The outcome variable is the inverse hyperbolic sine of migration. Columns (4) to (6) restrict the analy- sis to ‘regular’ corridors which have at least 10 migrants across all time period. Standard errors in parentheses are clustered at the origin-destination level. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. These estimates are likely lower bound of the total impact on migration due to our data limitations. Since our data captures the stock of migration between two countries at ten-year intervals, it likely misses or underestimates short-term migration, particularly those that happen between the periods of measurement. This may be particularly salient for BLAs that facilitate 13 temporary migration for a relatively short period of time. For instance, most of the migration to the Gulf Cooperation Council (GCC) countries are temporary in nature with workers migrating with an employment contract of two to three years. Robustness To address the concerns around the estimation methods, we conduct a battery of robustness checks. Table B.2 reports the results using the PPML methods; table B.3 reports the result of excluding the top 1 percent of the corridors. We also conduct robustness checks using logarithms as well as levels of migration. The results are qualitatively and quantitatively similar across these specifications. Instrumental variable estimates Table 3 reports the IV regression. Columns (1) to (3) of panels A and B report the first stage correlation between our instrument and the main variable of interest (BLA), for the full sample and for regular corridors respectively. Regardless of the instrument considered, there is a negative and significant relationship between our instrument and the BLA. Moreover, the first stage F-statistic ranges between 177.0 and 290.7, indicating the relevance of our instrument. The negative first stage coefficients indicate that the probability of j signing a BLA with i decreases with the total number of BLAs j has already signed with countries other than i. In other words, to the extent that BLAs are signed for the purpose of importing workers for a given duration, the need to import workers decreases with the number of BLAs already signed. Columns (4) to (6) in panels A and B confirm the positive impact of signing a BLA on migration flows. The IV estimate shows that signing an additional BLA increases migration from the origin to the destination country by 49.9 to 85.2 percent in the full sample and by about 20.7 to 31.9 percent in the regular corridors. We explain the smaller effect in the regular corridors by the fact that the magnitude of the change is smaller in corridors with larger initial migration flows than in corridors with smaller initial migration flows. The main channel through which this effect occurs is essentially through the reduction in migration costs that BLAs offer, ranging from easy access to visas to housing facilities at the destination. 14 Table 3: Migration response to BLAs (Instrumental variable) Leave one origin out Leave one destination out j BLAi¬jt ( i BLA¬ijt ) (1) (2) (3) (4) Panel A- Full sample First stage IV First stage IV Bilateral labor agreement (BLA) 0.405∗∗∗ 0.616∗∗∗ (0.100) (0.106) Leave one out −1.210∗∗∗ −1.497∗∗∗ (0.055) (0.087) Observations 273, 042 273, 042 273, 042 273, 042 K Paap Fstat. (First stage) 476.42 295.72 Panel B-Regular corridors First stage IV First stage IV Bilateral labor agreement (BLA) 0.188∗∗∗ 0.277∗∗∗ (0.058) (0.071) Leave one out −1.147∗∗∗ −1.399∗∗∗ (0.067) (0.105) Observations 33, 222 33, 222 33, 222 33, 222 K Paap Fstat. (First stage) 290.70 177.00 Origin country x decade FE Y es Y es Y es Y es Destination country x decade FE Y es Y es Y es Y es Origin x destination countries FE Y es Y es Y es Y es Notes: The table presents the first stage regression and the estimate of γ from IV estimate of equa- tion (2). It shows the suggestive causal impact of a BLA on migration. The outcome variable in columns (1) and (3) is the dummy variable indicating whether the corridor of interest during the decade of interest has signed a BLA or not. In columns (2) and (4), the outcome variable is the inverse hyperbolic sine of migration. Panel A focuses on the full sample while panel B restricts the analysis to ‘regular’ corridors which have at least 10 migrants across all time period. The set of fixed effects in the bottom rows of the table are common to the two panels. Standard errors in parentheses are clustered at the origin-destination level. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. 15 4.2 Heterogeneity analysis So far, it has been implicitly assumed that migration from countries around the world responds homogeneously to Bilateral Labor Agreements (BLAs). However, not only initial conditions at origin matter, but also, both the number and timing of BLA signings and the main destination countries have changed significantly over the years. As shown in Figure A.2, during the period of Western economic prosperity immediately following World War II (1945-1973), the majority of destination countries involved in BLAs were Western countries seeking to import workers to rebuild their war-damaged infrastructure and industries. This period was followed by eco- nomic stagnation, which was also reflected in a lower growth in the number of BLAs signed between 1973 and 1990. In the post-1990s, most of the growth in BLAs can be explained by the intensive use of these instruments by GCC countries. However, regions such as Sub-Saharan Africa have made very limited use of BLAs over the years. Overall, these different phases in the time dynamics of BLAs suggest that the migration response to BLAs may differ along different dimensions. To account for this heterogeneity, we further examine whether the estimated aver- age effect is heterogeneous along three dimensions: (i) the level of development of the country of origin (ii) whether the destination country is a GCC country or not, and (iii) whether the country of origin is African or not. Table 4 reports the heterogeneous migration response to BLAs along these three dimen- sions. Columns 5 and 6 report the average migration response to entering a BLA across the development level of the origin country. While entering a BLA increases migration between the origin country and the destination country involved in that BLA by about 310 percent within a decade when the origin country is a LLMIC, this effect falls to 24 percent when the origin country belongs to the upper-middle and high-income group. This result suggests that the poorer the country of origin, the larger the effect of BLAs on migration flows. This suggests that BLAs expand the economic opportunities especially when there is a huge place premium. We further investigate whether the extensive use of BLAs by GCC countries during the new golden age of BLAs has played a role. Although entering a BLA increases migration between origin country and destination country worldwide, the effect is ten times larger for migration to the GCC destinations as compared to non-GCC destinations (columns 1 and 2). This is likely explained by the fact that BLAs account for a large share of low- and middle-skilled 16 Table 4: Heterogeneous migration response to BLAs GCC destination African origin Development group (1) (2) (3) (4) (5) (6) Bilateral labor agreement (BLA) 0.468∗∗∗ 0.552∗∗∗ 0.650∗∗∗ 0.640∗∗∗ 1.522∗∗∗ 1.414∗∗∗ (0.106) (0.099) (0.115) (0.102) (0.228) (0.199) BLA X GCC destination 1.558∗∗∗ 1.688∗∗∗ (0.413) (0.430) BLA X African origin -0.568∗∗ -0.500∗∗∗ (0.240) (0.170) BLA X UM-HIC -1.351∗∗∗ -1.200∗∗∗ (0.251) (0.221) BLA + Interaction 2.026∗∗∗ 2.239∗∗∗ 0.081 0.139 0.170 0.214∗∗ (0.398) (0.417) (0.211) (0.136) (0.105) (0.093) Observations 273,042 273,042 273,042 273,042 273,042 273,042 R-sq 0.820 0.839 0.820 0.875 0.820 0.875 Origin country x decade FE Y Y Y Y Y Y Destination country x decade FE Y Y Y Origin x destination countries FE Y Y Y Y Y Y Notes : Standard errors in parentheses are clustered at the origin-destination level. BLA (dummy) in decade N takes the value 1 if a BLA was signed in the last 10 years and zero otherwise. The dependent variable is the number of dyadic migrations in protocols. Columns (1) and (2) report heterogeneity with respect to whether the country of destination is in the GCC or not. Columns (3) and (4) report the heterogeneous response of migration to BLA depending on whether country of origin is Africa or not. The last two columns examine the heterogeneity with respect to the level of development of the country of origin (low and lower middle income, upper middle income, and high income). For each of the highlighted categories, we consider two specifications: one that includes origin-destination and origin-time fixed effects, and another that complements this set of fixed effects with destination-time fixed effects. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. migration to the GCC countries. Consequently, whenever a GCC country enters a BLA with another country, nearly all variations in migration flows in that corridor are explained by the BLA. Table 4 reports the heterogeneous migration response to BLAs along the three dimensions indicated. Furthermore, we investigate whether BLAs have so far affected migration flows from Africa. However, BLAs do not seem to have a significant effect on migration flows from Africa (columns 3 and 4). This finding is intriguing, given that African countries make up a substantial portion of low- and lower-middle-income countries, which generally experience the largest migration increases in response to signing a new BLA (columns 5 and 6). This heterogeneous effect illustrates what we refer to as the “African Puzzle ”. On one hand, BLAs increase migration flows, especially from LLMIC; on the other hand, there is no significant migration response to BLAs in Africa. This is somewhat puzzling given that of the 79 countries classified as low- and lower-middle-income in our sample, 45, or more than half, are in Africa. To better understand this “puzzle”, we further investigate whether the quality of institutions 17 plays a role. We created institutional tiles based on the average government effectiveness index by country from the World Governance Indicator database. Government effectiveness measures the quality of public services, the civil service, policy formulation and implementation, and the credibility of a government’s commitment to improving or maintaining these aspects (Kauf- mann, Kraay and Mastruzzi, 2010). Our general intuition is that developing countries with low government effectiveness are not only less able to engage in BLAs but also less able to enforce the BLAs they do sign with partner countries. This likely translates into less effective BLAs and, consequently, lower migration flows. To test this hypothesis, we created two, three, and four tiles corresponding to halves, terciles, and quartiles. Countries in the top tiles are characterized by more effective government. Results reported in table 5 strongly suggest that, although the migration response to BLAs in Africa is null, African countries belonging to the upper tiers of the government indicators do experience a positive migration response to BLAs. Conversely, African countries with poor governance indicators show no significant migration response to BLAs. 4.3 Persistence of the impacts We now turn our attention to the timing of the effect of BLAs on migration flows using an event study design. It is worth noting that one of the key advantages of the event study design is the possibility to check if the estimated effect is long lasting or if it quickly vanishes after a certain period of time. Figure 1 depicts the benchmark estimate timing of the effect of BLAs on migration flows. First, there is no pre-trend prior to the treatment. Second, it appears that the positive effect of BLAs on migration flows lasts up to 3 decades after BLAs are signed before vanishing. Consistently, heterogeneity results show that this positive effect is driven by non-African origins (panel 2a of figure 2). An interesting result of this event study design emerges when comparing the effects on GCC and non-GCC destinations. Essentially, non-GCC destinations replicate the average timing of the effects of BLAs on migration, while GCC destinations show a persistent effect of BLAs on migration flows. A simple explanation is that most BLAs signed in the aftermath of 1990 involve GCC countries. This is also why there is a smaller number of decades in the post treatment period for GCC destinations. Moreover, BLAs are almost the only channel through which migrant workers enter the GCC. 18 Table 5: Heterogeneous migration response to BLAs by governance category Governance categories Top half Top two terciles Top three quartiles (1) (2) (3) Bilateral labor agreement (BLA) 1.281∗∗∗ 2.004∗∗∗ 1.979∗∗∗ (0.263) (0.457) (0.577) BLA × Africa −1.194∗∗∗ −1.985∗∗∗ −2.283∗∗∗ (0.325) (0.528) (0.668) BLA × Top half −0.821∗∗∗ (0.286) BLA × Middle tercile −0.825∗ (0.497) BLA × Top tercile −1.924∗∗∗ (0.469) BLA × Third quartile −1.009 (0.639) BLA × Second quartile −0.869 (0.609) BLA × First quartile −1.947∗∗∗ (0.590) BLA × Africa × Top half 0.983∗∗∗ (0.377) BLA × Africa × Middle tercile 1.066∗ (0.579) BLA × Africa × Top tercile 1.624∗∗ (0.657) BLA × Africa ×Third quartile 1.684∗∗ (0.748) BLA × Africa × Second quartile 1.462∗∗ (0.714) BLA × Africa × First quartile 1.869∗∗ (0.873) Observations 268, 905 268, 905 268, 905 R-sq 0.875 0.876 0.875 Origin country x decade FE Y es Y es Y es Destination country x decade FE Y es Y es Y es Origin x destination countries FE Y es Y es Y es Notes: Standard errors in parentheses are clustered at the origin-destination level. BLA (dummy) in decade N takes the value 1 if a BLA was signed in the last 10 years and zero otherwise. The dependent variable is the inverse hyperbolic ssine of the number of dyadic migrations in protocols. For each of the highlighted categories, we consider consider the specification with the most stringent set of fixed effects. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. In summary, our results show that Bilateral Labor Agreements (BLAs) significantly increase migration flows worldwide. This positive effect is particularly pronounced for destinations within the Gulf Cooperation Council (GCC) countries. While LLMIC origin countries experience the most significant positive impact of BLAs on migration flows, African countries do not seem to be significantly affected, leading to what we refer to as the “African puzzle ”. Further investigation 19 Figure 1: Event study estimates: migration response to BLAs Source : Authors’ estimates. Notes : The figure plots the impact of signing a BLA on migration, estimated using equation (3). Vertical bars indicate 95 percent confidence intervals, standard errors are clustered at the origin × destination cells. Figure 2: Heterogeneity in migration response to BLAs (a) African vs. non African origin (b) GCC vs. non GCC destinations Source : Authors’ estimates. Notes : The figure plots the impact of signing a BLA on migration, estimated using equation (3), for Africa vs non-Africa origin countries (Panel 2a) and for GCC vs non-GCC destinations (Panel 2b). Vertical bars indicate 95 percent confidence intervals, standard errors are clustered at the origin × destination cell. suggests that the quality of institutions in most African countries may be a critical factor. Our analysis shows that while migration flows from African countries are generally non-responsive to BLAs, African countries with better institutional quality show a positive impact of these agreements on migration flows. This finding suggests that establishing BLAs alone is not enough to stimulate migration flows; the capacity of the government to enforce the terms of 20 these agreements is also crucial. Thus, it is not only the act of entering into a BLA that matters, but also the effectiveness of the government in maintaining and monitoring these agreements. This underscores the importance of a strong institutional setup in improving the effectiveness of BLAs in promoting migration flows. 4.4 Heterogeneity-robust treatment effects The recent econometrics literature has increasingly scrutinized the biases of the canonical two- way (group and time) fixed effects estimates in a difference-in-difference setup when treat- ment effects are heterogeneous by group or over time (see de Chaisemartin and d’Haultfœuille, 2022, for a survey). The bias could be due to, for instance, “forbidden comparisons” that use already-treated groups as controls for later-treated groups (Borusyak, Jaravel and Spiess, 2024, de Chaisemartin and d’Haultfœuille, 2020). Likewise, in dynamic specifications, where outcomes are affected by past treatments, it is likely that the treatment effect across groups first treated at different times biases the identification of the long-run effect and especially the persistence of the policy effect. To account for this heterogeneity, the literature has developed heterogeneity-robust estimators that improve upon the canonical estimator (Borusyak, Jaravel and Spiess, 2024, Callaway and Sant’Anna, 2021, de Chaisemartin and d’Haultfœuille, 2024, Sun and Abraham, 2021, Wooldridge, 2023). The empirical application of these approaches that is closest to our setting is the evaluation of the impact of regional trade agreements on bilateral trade (Nagengast and Yotov, Forthcoming), and the impact of membership in the European Union single market on international trade (Nagengast, Rios-Avila and Yotov, 2024). Like our setting, both papers have a panel gravity model with a dyad (an origin-destination pair) as the treatment unit but also allows for complex set of fixed effects, including origin × time and destination × time fixed effects.7 With the new approach, they find that, the canonical estimates significantly underestimate the effect of trade policy on international trade by roughly 50%. Hence, we follow the approach developed in Nagengast and Yotov (Forthcoming), Nagengast, Rios-Avila and Yotov (2024) to estimate heterogeneity-robust impact of BLAs on migration. We 7 However, our setting presents an additional complication - the treatment in our case is non-staggered (it switches on and off over time) and that previous treatment may still affect later outcomes. To the best of our knowledge, an estimator for panel gravity settings with non-staggered outcomes is not yet available. 21 find that, as in these studies from the trade literature, our canonical estimates are qualitatively similar, but smaller in magnitudes than the heterogeneity robust estimators (see Table C.1). This also applies to the results for the period 0 estimates in the event study framework (see Figures C.2 and C.3).8 We also use our event-study framework to test for pre-trends using the heterogeneity-robust estimator and fail to reject the null of no pre-trends (figure C.1) with a p-value of 0.66. 5 Welfare impacts Large migration impact of a BLAs suggests that a BLA can have significant welfare implications. Here, we calculate the welfare impacts, in terms of increased incomes for migrants, of the migration impacts of signing a BLA. In 2020, an average corridor with LLMIC origin country had a migrant stock of about 8,000 individuals. With the estimates in column 6 of table 4), this implies an increase of 25,000 migrants within a decade in response to a BLA. Corridors from LLMIC to GCC countries had a larger stock of 56,000 migrants in 2020. A BLA, based on the impacts in column 2 of table 4, would increase the migration stock by 472,000 migrants within a decade. The average Sub-Saharan African country had a lower migration stock of 3,500 and 11,500 to the average and GCC destinations, respectively. A BLA would increase this only by about 500 migrants given the low impacts of BLAs observed in the African origin countries (column 4, table 4). If migration impacts were the same as the average LLMIC country, then the number would be much higher at 11,000 to an average destination and 96,000 to a GCC destination. To convert the migration impacts to earnings, we use the average nominal GDP per-capita to proxy for earnings at origin and the returns to migration estimated in the literature focusing on low-skilled migration from LLMICs. Rigorous empirical research on the returns to migration establishes that income increases several-folds even for low-skilled workers migrating to high- and upper-middle-income destinations. As summarized in World Bank (2023), estimates range from 118 to 298 percent across these studies (Clemens, 2019, Clemens and Tiongson, 2017, oth, 2023, McKenzie, Stillman and Gibson, 2010, Mobarak, Sharif and Gaikwad, Hanson and T´ 8 However, the results on persistence differ from the canonical estimates presented in figure 1. This could be driven by the fact that treatments are non-staggered in our settings and the estimators are not well adapted to this difference. 22 Shrestha, 2023). We use a simple average of these studies, a return of 227 percent, to estimate earnings increase from migration to an average destination and of 208 percent to estimate earnings increase from migration to the GCC countries. Our estimates imply a BLA results in an annual earnings gain of US$120 million for an LLMIC origin country. A BLA with a GCC destination would result in an annual earnings gain of US$2.1 billion. For countries in Sub-Saharan Africa, the annual gains are a modest US$2.4 million, due to a lower migration stock and the limited impact of BLAs. If Sub-Saharan African countries were to have the same impact of BLAs as the average LLMIC, the annual earnings gain from a BLA would be US$51 million. This suggests an annual loss of about US$48.6 million due, largely, to the weaker capacity in African origin countries to implement such BLAs. Similarly, the annual gains for a Sub-Saharan African country from signing a BLA with a GCC destination country (and improving capacity so that migration impacts are similar to that of the average LLMIC countries) could be as high as US$413 million. We find that the migration impact persists beyond the decade of signing the BLA for up to three decades (figure 1). This means that the earnings gain can accrue over multiple decades. Based on our estimates, we assume that the migration impacts persists for 25 years and that the earnings are discounted by 3 percent each year. We further assume that it takes the first five years for the BLAs to have the estimated migration impact over which period no gains are realized.9 With these assumptions, we calculate that the total earnings gain from a BLA for LLMIC origin country can be US$1.6 billion, and the gains from a BLA with a GCC destinations can be US$27.6 billion. For countries in Sub-Saharan Africa, the gains are modest at US$32 million. However, assuming the same impact of a BLA as the average LLMIC origin countries, the gains for a Sub-Saharan African origin country can be US$676 million from a BLA with an average destination and US$5.5 billion from a BLA with a GCC destination country. 9 We are underpowered to estimate the annual trajectory of migration impact within the first decade. But preliminary estimates suggest that BLAs start to have migration impacts within the first five years of signing. In that regard, the assumption made here is very conservative and would underestimate the long-term earnings gains from BLAs. 23 6 Conclusion The demographic divergence between the aging high- and upper-middle-income countries – with shrinking workforces – and the LLMICs – with still expanding youth population – provides a huge opportunity for many countries to leverage migration for their own benefits (Abdel Jelil et al., 2024, World Bank, 2023). As seen in this study, BLAs can be a credible policy instrument to facilitate such flows between countries. For LLMICs, the welfare gain from an agreement can be US$120 million annually, and US$2.1 billion for agreements with destinations in the GCC countries where the migration impacts of the BLAs are large. However, this study points to an important role of government capacity – particularly for origin countries in Africa where most of the youthful population is and will be concentrated – to implement such agreements and realize the gains from migration. References Abdel Jelil, Mohamed, Samik Adhikari, Quy-Toan Do, Heidi Kaila, Federica Marzo, Olive Nsababera, Ganesh Seshan, and Maheshwor Shrestha. 2024. Migration: Africa’s Untapped Potential. Africa Development Forum, Washington, DC:World Bank. e de Castro Dobbin, and Melanie Morten. 2018. “Border walls.” Allen, Treb, Cauˆ National Bureau of Economic Research. Angelucci, Manuela. 2015. “Migration and financial constraints: Evidence from Mexico.” Review of Economics and Statistics, 97(1): 224–228. Arpaia, Alfonso, Aron Kiss, Balazs Palvolgyi, and Alessandro Turrini. 2018. “The ef- fects of European integration and the business cycle on migration flows: a gravity analysis.” Review of World Economics, 154: 815–834. Bah, Tijan L, Catia Batista, Flore Gubert, and David McKenzie. 2023. “Can in- formation and alternatives to irregular migration reduce “backway” migration from The Gambia?” Journal of Development Economics, 165: 103153. Baseler, Travis. 2023. “Hidden income and the perceived returns to migration.” American Economic Journal: Applied Economics, 15(4): 321–352. 24 Bazzi, Samuel. 2017. “Wealth heterogeneity and the income elasticity of migration.” American Economic Journal: Applied Economics, 9(2): 219–255. Bazzi, Samuel, Lisa Cameron, Simone G Schaner, and Firman Witoelar. 2021. “In- formation, intermediaries, and international migration.” National Bureau of Economic Re- search. Beam, Emily A, David McKenzie, and Dean Yang. 2016. “Unilateral facilitation does not raise international labor migration from the Philippines.” Economic Development and Cultural Change, 64(2): 323–368. us Fern´ Beine, Michel, Simone Bertoli, and Jes´ andez-Huertas Moraga. 2016. “A practitioners’ guide to gravity models of international migration.” The World Economy, 39(4): 496–512. Bellemare, Marc F, and Casey J Wichman. 2020. “Elasticities and the inverse hyperbolic sine transformation.” Oxford Bulletin of Economics and Statistics, 82(1): 50–61. us Fern´ Bertoli, Simone, and Jes´ andez-Huertas Moraga. 2013. “Multilateral resistance to migration.” Journal of Development Economics, 102: 79–100. Bian, Bo, Jean-Marie Meier, and Ting Xu. 2023. “Cross-border institutions and the globalization of innovation.” LawFin Working Paper. Biehler, Nadine, David Kipp, and Anne Koch. 2024. “The potential of bilateral migration agreements: From symbolic politics to practical implementation.” SWP Comment. Bither, Jessica, and Astrid Ziebarth. 2018. “Creating legal pathways to reduce irregular migration? What we can learn from Germany’s “Western Balkan regulation”.” Borusyak, Kirill, Xavier Jaravel, and Jann Spiess. 2024. “Revisiting event-study designs: robust and efficient estimation.” Review of Economic Studies, rdae007. Bryan, Gharad, and Melanie Morten. 2019. “The aggregate productivity effects of internal migration: Evidence from Indonesia.” Journal of Political Economy, 127(5): 2229–2268. Bryan, Gharad, Shyamal Chowdhury, and Ahmed Mushfiq Mobarak. 2014. “Under- investment in a profitable technology: The case of seasonal migration in Bangladesh.” Econometrica, 82(5): 1671–1748. 25 Burbidge, John B, Lonnie Magee, and A Leslie Robb. 1988. “Alternative transfor- mations to handle extreme values of the dependent variable.” Journal of the American statistical Association, 83(401): 123–127. Cai, Shu. 2020. “Migration under liquidity constraints: Evidence from randomized credit access in China.” Journal of Development Economics, 142: 102247. Callaway, Brantly, and Pedro HC Sant’Anna. 2021. “Difference-in-differences with mul- tiple time periods.” Journal of econometrics, 225(2): 200–230. Chilton, Adam, and Bartosz Woda. 2022. “The expanding universe of bilateral labor agreements.” Theoretical Inquiries in Law, 23(2): 1–64. Chilton, Adam S, and Eric A Posner. 2018. “Why countries sign bilateral labor agree- ments.” The Journal of Legal Studies, 47(S1): S45–S88. Clemens, Michael A. 2019. “Measuring the Spatial Misallocation of Labor: The Returns to India-Gulf Guest Work in a Natural Experiment.” CGD Working Paper 501 (January), Center for Global Development, Washington, DC. Clemens, Michael A. 2020. “The emigration life cycle: How development shapes emigration from poor countries.” IZA Institute of Labor Economics 13614. Clemens, Michael A. 2024. “The effect of lawful crossing on unlawful crossing at the US southwest border.” Peterson Institute for International Economics Working Paper 24-10. Clemens, Michael A, and Erwin R Tiongson. 2017. “Split decisions: Household finance when a policy discontinuity allocates overseas work.” Review of Economics and Statistics, 99(3): 531–543. Clemens, Michael A., Claudio E. Montenegro, and Lant Pritchett. 2019. “The Place Premium: Bounding the Price Equivalent of Migration Barriers.” The Review of Economics and Statistics, 101(2): 201–213. Clemens, Michael A, Ethan G Lewis, and Hannah M Postel. 2018. “Immigration restrictions as active labor market policy: Evidence from the mexican bracero exclusion.” American Economic Review, 108(6): 1468–1487. 26 ed´ Dao, Thu Hien, Fr´ eric Docquier, Chris Parsons, and Giovanni Peri. 2018. “Mi- gration and development: Dissecting the anatomy of the mobility transition.” Journal of Development Economics, 132: 88–101. ement, and Xavier d’Haultfœuille. 2020. “Two-way fixed effects es- de Chaisemartin, Cl´ timators with heterogeneous treatment effects.” American Economic Review, 110(9): 2964– 2996. ement, and Xavier d’Haultfœuille. 2022. “Two-way fixed effects and de Chaisemartin, Cl´ differences-in-differences with heterogeneous treatment effects: A survey.” The Economet- rics Journal, 26(3): C1–C30. ement, and Xavier d’Haultfœuille. 2024. “Difference-in-differences de Chaisemartin, Cl´ estimators of intertemporal treatment effects.” Review of Economics and Statistics, 1–45. Dettmann, Eva, Alexander Giebler, and Antje Weyh. 2020. “Flexpaneldid: A Stata toolbox for causal analysis with varying treatment time and duration.” Available at SSRN 3692458. Di Iasio, Valentina, and Jackline Wahba. 2023. “Expecting Brexit and UK migration: Should I go?” European Economic Review, 157: 104484. Feigenberg, Benjamin. 2020. “Fenced out: The impact of border construction on US-Mexico migration.” American Economic Journal: Applied Economics, 12(3): 106–139. oth. 2023. “Exit Options: How International Gaikwad, Nikhar, Kolby Hanson, and Aliz T´ Migration Opportunities Shape Economic Standing and Political Preferences.” Working paper. Gazeaud, Jules, Eric Mvukiyehe, and Olivier Sterck. 2023. “Cash transfers and migra- tion: Theory and evidence from a randomized controlled trial.” Review of Economics and Statistics, 105(1): 143–157. Gibson, John, David McKenzie, Halahingano Rohorua, and Steven Stillman. 2018. “The long-term impacts of international migration: Evidence from a lottery.” The World Bank Economic Review, 32(1): 127–147. 27 ement, and John Papp. 2020. “Costs and benefits of rural-urban migration: Imbert, Cl´ Evidence from India.” Journal of Development Economics, 146: 102473. Kaufmann, Daniel, Aart Kraay, and Massimo Mastruzzi. 2010. “The worldwide gover- nance indicators: Methodology and analytical issues.” World Bank Policy Research Work- ing Paper 5430. Livnat, Yuval, and Hila Shamir. 2022. “Gaining control? bilateral labor agreements and the shared interest of sending and receiving countries to control migrant workers and the illicit migration industry.” Theoretical Inquiries in Law, 23(2): 65–94. en. 2019. “The long-run effects of teacher Lovenheim, Michael F, and Alexander Will´ collective bargaining.” American Economic Journal: Economic Policy, 11(3): 292–324. MacKinnon, James G, and Lonnie Magee. 1990. “Transforming the dependent variable in regression models.” International Economic Review, 315–339. Maximova, Anna, and Mihai Paraschiv. 2022. “The Effect of Bilateral Labor Agreements on Trade.” Journal of Economic Integration, 37(4): 649–704. Mayda, Anna Maria, Francesc Ortega, Giovanni Peri, Kevin Shih, and Chad Spar- ber. 2018. “The effect of the H-1B quota on the employment and selection of foreign-born labor.” European Economic Review, 108: 105–128. McKenzie, David. 2007. “Paper walls are easier to tear down: Passport costs and legal barriers to emigration.” World Development, 35(11): 2026–2039. McKenzie, David. 2023. “Fears and Tears: Should More People Be Moving within and from Developing Countries, and What Stops this Movement?” The World Bank Research Ob- server, 39(1): 75–96. McKenzie, David, Caroline Theoharides, and Dean Yang. 2014. “Distortions in the in- ternational migrant labor market: evidence from Filipino migration and wage responses to destination country economic shocks.” American Economic Journal: Applied Economics, 6(2): 49–75. 28 McKenzie, David, Steven Stillman, and John Gibson. 2010. “How important is selection? Experimental vs. non-experimental measures of the income gains from migration.” Journal of the European Economic Association, 8(4): 913–945. Miller, Douglas L. 2023. “An introductory guide to event study models.” Journal of Economic Perspectives, 37(2): 203–230. Mobarak, Ahmed Mushfiq, Iffath Sharif, and Maheshwor Shrestha. 2023. “Returns to international migration: Evidence from a Bangladesh-Malaysia visa lottery.” American Economic Journal: Applied Economics, 15(4): 353–388. Munshi, Kaivan. 2020. “Social networks and migration.” Annual Review of Economics, 12(1): 503–524. Munshi, Kaivan, and Mark Rosenzweig. 2016. “Networks and misallocation: Insurance, migration, and the rural-urban wage gap.” American Economic Review, 106(01): 46–98. Nagengast, Arne, and Yoto V Yotov. Forthcoming. “Staggered difference-in-differences in gravity settings: Revisiting the effects of trade agreements.” American Economic Journal: Applied Economics. Nagengast, Arne, Fernando Rios-Avila, and Yoto Yotov. 2024. “The European Single Market and Intra-EU Trade: An Assessment with Heterogeneity-Robust Difference-in- Differences Methods.” LeBow College of Business, Drexel University. Orefice, Gianluca. 2015. “International migration and trade agreements: The new role of economique, 48(1): 310–334. PTAs.” Canadian Journal of Economics/Revue canadienne d’´ Ortega, Francesc, and Giovanni Peri. 2013. “The effect of income and immigration policies on international migration.” Migration Studies, 1(1): 47–74. ¨ Ozden, ¸ a˘ C glar, Christopher R Parsons, Maurice Schiff, and Terrie L Walmsley. 2011. “Where on earth is everybody? The evolution of global bilateral migration 1960–2000.” The World Bank Economic Review, 25(1): 12–56. Pence, Karen M. 2006. “The role of wealth transformations: An application to estimating the effect of tax incentives on saving.” Contributions in Economic Analysis & Policy, 5(1). 29 Peters, Margaret E. 2019a. “Bilateral Labor Agreements Dataset and Additional Replication Data for: Immigration and International Law.” Peters, Margaret E. 2019b. “Immigration and international law.” International Studies Quar- terly, 63(2): 281–295. Peters, Margaret E, and Michael K Miller. 2022. “Emigration and political contestation.” International Studies Quarterly, 66(1): sqab088. Rios-Avila, Fernando, Arne J. Nagengast, and Yoto V. Yotov. 2022. “JWDID: Stata module to estimate Difference-in-Difference models using Mundlak approach.” Statistical Software Components S459114, Boston College Department of Economics, revised 08 Apr 2024. Roth, Jonathan. 2024. “Interpreting event-studies from recent difference-in-differences meth- ods.” arXiv preprint arXiv:2401.12309. Roth, Jonathan, Pedro HC Sant’Anna, Alyssa Bilinski, and John Poe. 2023. “What’s trending in difference-in-differences? A synthesis of the recent econometrics literature.” Journal of Econometrics, 235(2): 2218–2244. Saez, Juan Sebastian. 2013. “Using bilateral labor agreements to increase trade in services.” World Bank Group. Shrestha, Maheshwor. 2019. “Death scares: How potential work-migrants infer mortality rates from migrant deaths.” Journal of Development Economics, 141: 102368. Shrestha, Maheshwor. 2020. “Get rich or die tryin’: Perceived earnings, perceived mortality rates, and migration decisions of potential work migrants from Nepal.” The World Bank Economic Review, 34(1): 1–27. Shrestha, Maheshwor. 2023. “A Deeper Dive into the Relationship between Economic De- velopment and Migration.” World Bank Policy Research Working Paper 10295, World Bank. Sun, Liyang, and Sarah Abraham. 2021. “Estimating dynamic treatment effects in event studies with heterogeneous treatment effects.” Journal of econometrics, 225(2): 175–199. 30 UNDESA. 2020. “International Migrant Stock 2020.” United Nations Department of Economic and Social Affairs, Population Division. United Nations, Network on Migration. 2022. “Guidance on Bilateral Labour Migration Agreements.” Wooldridge, Jeffrey M. 2023. “Simple approaches to nonlinear difference-in-differences with panel data.” The Econometrics Journal, 26(3): C31–C66. World Bank. 2022. “Global Bilateral Migration | DataBank.” World Bank. 2023. World Development Report 2023: Migrants, Refugees, and Societies. The World Bank, Washington D.C. 31 Appendix A Data on BLAs The data we use for the BLAs is from Chilton and Woda (2022). This dataset contains infor- mation on 1,222 BLAs, also known as treaties, each formed between two countries since World War II. The dataset contains information on the year the BLA was signed and the two countries between which the treaty was signed. The dataset is the most comprehensive dataset on BLAs, that builds on the data by Chilton and Posner (2018), Peters (2019a) and Peters (2019b). How- ever, it is possible that BLAs from which there is no publicly available record are not accounted for, as no international organization is charged with keeping a record of BLAs. We use this sample to analyzes the effect of signing a BLA between two countries on the migration flows between these countries in the subsequent decades. Figure A.1 plots the number of cumulative signed BLAs over time globally, showing an increasing trend, particularly during the post-World War II period, and since the 1990s, post-Cold War. Figure A.1: Evolution of Bilateral labor agreements over time Notes : This figure plots the global cumulative number of BLAs. Figures A.2 and A.3 plot the number of cumulative BLAs by destination and origin region, 32 respectively. BLAs where the destination country is in Europe and Central Asia, in either North or Southern America and the Caribbean have increased steadily since the post-WW2 period. BLAs in Middle East and North Africa (including those in the Gulf Cooperation Council countries) have increase rapidly only in the last decades, particularly in the 2000s. In terms of origin countries, there has been a steady increase in the BLAs where the origin country is in Europe and Central Asia, and in the 2000s, BLAs where the origin country are from East Asia and Pacific have also increased substantively. In other origin regions, including in Sub-Saharan Africa (SSA), there either has not been much of an increase, or it has been modest compared to the aforementioned regions. Figure A.2: Evolution of Bilateral labor agreements by destination region Notes: This figure plots the global cumulative number of BLAs across destination regions. EAP = East Asia and Pacific, ECA= Europe and Central Asia, SSA = Sub-Saharan Africa, SOA= South Asia, LAC = Latin American and the Caribbean, GCC= Gulf Cooperation Council, MENA OTH= Middle East and North Africa excluding GCC, NOAM = North America. Sub-Saharan African (SSA) countries have just 91 BLAs, that is, they are in 7.45 percent of the dataset. SSA countries’ main counterparts are in Europe and Central Asia (59 BLAs), many of these are BLAs signed with a country with colonial ties: France has the most BLAs with SSA countries (44).10 SSA countries also have BLAs with countries in the Middle East 10 The other countries in this region that have formed BLAs with SSA countries are East Germany (with Ghana and Guinea in 1964), Netherlands with Zimbabwe (1955 & 1956), Portugal with Cape 33 Figure A.3: Evolution of Bilateral labor agreements by origin region Notes : This figure plots the global cumulative number of BLAs across origin regions. EAP = East Asia and Pacific, ECA= Europe and Central Asia, SSA = Sub-Saharan Africa, SOA= South Asia, LAC = Latin American and the Caribbean, GCC= Gulf Cooperation Council, MENA OTH= Middle East and North Africa excluding GCC, NOAM = North America. and North Africa (16).11 Countries in SSA also have 12 BLAs within the region, 10 are between South Africa and its neighbors, and the two remaining are between Mauritania and Mali, and Mauritania and Senegal.12 For a subset of 571 BLAs, the dataset also contains information on their characteristics, such as on whether the text of the BLA mentions 20 topics ILO has identified as best practices for these agreements (e.g. role of labor unions, gender protection, and whether employees have a contract). This is an indicator of the quality of the agreement. This subset of data also contains information on the duration of the treaty, whether the BLA has a quota, whether it replaces an existing one. We provide some descriptive evidence on these data, but do not use it in the analysis due to the fact that it is likely to be biased. The results are shown in Appendix Figures A.4, A.5 and A.6. Overall, while there are some differences, BLAs formed with countries Verde (1997), and South Africa (1964), Spain with Mali, Mauritania, Guinea-Bissau (2007-08), and Nigeria (1950s-60s), Sweden with Mozambique (1976) and Switzerland with South Africa (1998). 11 These are both within Africa, Mauritania and Algeria having one BLA, and the Arab Republic of Egypt and Sudan one, and Niger and Libya one. The other counterparts are Jordan, Qatar, Saudi Arabia, Tunisia and the United Arab Emirates. 12 The remaining BLAs involving a country in SSA are two BLAs with South Asian countries, and one with Latin America and the Caribbean, and one with East Asia and Pacific. 34 in Sub-Saharan Africa do not differ substantively in the frequency at which these criteria are mentioned. Figure A.4: ILO criteria: governance Notes : This figure plots the share of BLAs that mention a topic of labor migration governance using the a subset of the dataset for which this information is available. The sample is split between BLAs that contain at least one country from Africa, and other BLAs. The categories are as follows: (1) References to International Instruments, (2) Exchange of Information between Countries, (3) Dissemination of In- formation about BLA’s Existence, (4) Defining Clear Responsibilities Between Parties, (5) Establishing a Joint Committee, (6a) Migrant Should Not Pay Recruitment Fees, (6b) Specifies Agents Authorized to Recruit Workers, (7a) Role for Labor Unions in Origin Country, (7b) Role for Labor Unions in Desti- nation Country, (7c) Migrants Can Join Labor Unions in Destination Country, (7d) Role for Employer Organizations, (7e) Role for Other NGOs or Civil Society Organizations. 35 Figure A.5: ILO criteria: protection Notes : This figure plots the share of BLAs that mention a topic related to protection and empowerment of migrant workers using the a subset of the dataset for which this information is available. The sample is split between BLAs that contain at least one country from Africa, and other BLAs. The categories are as follows: (8) Provision of Relevant Information to Migrants, (9) Equal Treatment and Non-Discrimination of Migrant Workers, (10a) General Gender Protection of Women, (10b) Detailed Gender Protection of Women/Domestic Workers, (10c) Other Protections (Race, Religion, etc.), (11a) Employment Contract Required, (11b) Standard/Model Employment Contract, (11c) Specific Contract Terms Required, (12) Wage Protection, (13a) Employer Required to Provide Housing to Workers, (13b) Housing Must Meet Certain Conditions, (13c) Government Monitors Housing, (13d) Government Monitors Work Conditions, (14) Prohibition of Confiscation of Travel Documents, (15) Social Protection and Health-Care Benefits, (16) Mechanisms for Complaints and Dispute Resolution. 36 Figure A.6: ILO criteria: development Notes : This figure plots the share of BLAs that mention a topic related to development using the a subset of the dataset for which this information is available. The sample is split between BLAs that contain at least one country from Africa, and other BLAs. The categories are as follows: (17) Human Resource Development and Skills Improvement, (18) Recognition of Skills and Qualifications, (19) Transfer of Savings and Remittances, (20a) Reintegration of Migrants upon Return, (20b) Possibility of Contract Renewal, (20c) Pathway to Legal Permanent Residence. 37 B Robustness results Table B.1: Migration and migration corridors with BLA over time Full sample African origins LLMIC UMHIC GCC destinations (1) (2) (3) (4) (5) Migrant stock (in million) 1960 91.2 8.1 34.8 56.5 0.2 1970 103.2 10.7 38.1 65.1 1.1 1980 116.6 13.5 43.9 72.8 3.9 1990 137.6 16.1 52.6 85.0 8.6 2000 160.4 19.5 60.6 99.8 10.2 2010 208.4 28.3 88.0 120.4 19.8 2020 261.1 36.8 119.8 141.3 29.8 Migrant stock-regular corridors (in million) 1960 88.3 7.5 33.8 54.5 0.2 1970 98.0 9.9 35.7 62.3 0.8 1980 110.5 12.5 40.8 69.7 3.4 1990 128.8 14.9 48.3 80.5 7.6 2000 148.8 17.5 54.3 94.5 9.0 2010 193.6 25.9 79.9 113.7 18.1 2020 242.0 33.4 108.7 133.3 27.3 Migrant stock- corridors with BLAs (in million) 1960 20.7 1.6 1.9 18.8 0.1 1970 26.6 2.7 3.8 22.8 0.3 1980 32.3 3.4 6.1 26.3 1.3 1990 39.7 4.1 9.5 30.2 2.8 2000 48.5 3.4 11.9 36.6 3.6 2010 67.7 6.9 25.6 42.0 11.0 2020 81.8 8.1 34.9 46.9 16.5 Source :Authors’ calculations from the BLA matrix and the World Bank Bilateral migration matrix. Table B.2: Migration response to BLAs (Poisson Pseudo Maximum Likelihood) Full sample Regular corridors only (Mij >= 10) (1) (2) (3) (4) (5) (6) BLA 0.318∗∗ 0.308∗∗∗ 0.179∗∗∗ 0.294∗ 0.296∗∗∗ 0.158∗∗∗ (0.145) (0.085) (0.037) (0.155) (0.091) (0.039) Observations 164, 283 163, 803 163, 132 33, 306 33, 257 33, 222 Origin country x decade FE No Y es Y es No Y es Y es Destination country x decade FE No No Y es No No Y es Origin x destination countries FE Y es Y es Y es Y es Y es Y es Decade FE Y es Y es Notes : The table presents the estimate of γ from PPML estimate of equation (2). It shows the impact of a BLA on migration. The outcome variable is the number of migrants. Columns (4) to (6) restrict the analysis to ‘regular’ corridors which have at least 10 migrants across all time period. Standard errors in parentheses are clustered at the origin-destination level. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. 38 Table B.3: Migration response to BLAs (top 1 percent corridors excluded) Top 1 percent removed (1) (2) (3) Bilateral labor agreement (BLA) 0.267∗∗∗ 0.136∗∗∗ 0.088∗∗ (0.065) (0.053) (0.042) Observations 161, 552 161, 072 160, 258 Origin country x decade FE No Y es Y es Destination country x decade FE No No Y es Origin x destination countries FE Y es Y es Y es Decade FE Y es Notes : The table presents the estimate of γ from PPML estimate of equation (2). It shows the impact of a BLA on migration when top 1% of corridors (large cor- ridors) are excluded. The outcome variable is the number of migrants. Standard errors in parentheses are clustered at the origin-destination level. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. Table B.4: Heterogeneous migration response to BLAs (Pseudo Poisson Maximum Likelihood) GCC destination African origin Development group (1) (2) (3) (4) (5) (6) Bilateral labor agreement (BLA) 0.188∗∗∗ 0.219∗∗∗ 0.326∗∗∗ 0.198∗∗∗ 0.600∗∗∗ 0.225∗∗∗ (0.046) (0.046) (0.091) (0.039) (0.190) (0.062) BLA X GCC destination 0.652∗∗ 0.118 (0.331) (0.177) BLA X African origin -0.217 -0.212∗∗ (0.155) (0.085) BLA X UM-HIC -0.458∗∗ -0.071 (0.197) (0.073) BLA + Interaction 0.840∗∗ 0.337∗ 0.109 -0.014 0.141∗∗∗ 0.154∗∗∗ (0.330) (0.173) (0.125) (0.080) (0.053) (0.045) Observations 163,803 163,803 163,803 163,132 163,803 163,132 R-sq Origin country x decade FE Y Y Y Y Y Y Destination country x decade FE Y Y Y Origin x destination countries FE Y Y Y Y Y Y Notes : Standard errors in parentheses are clustered at the origin-destination level. BLA (dummy) in decade N takes the value 1 if a BLA was signed in the last 10 years and zero otherwise. The dependent variable is the number of dyadic migrations. Columns (1) and (2) report the heterogeneous response of migration to BLA depending on whether the region of destination is GCC or not. Columns (3) and (4) report heterogeneity with respect to whether the country of origin is African or not. The last two columns examine the heterogeneity with respect to the level of development of the country of origin (low and lower middle income, upper middle income, and high income). For each of the highlighted categories, we consider two specifications: one that includes origin-destination and origin-time fixed effects, and another that complements this set of fixed effects with destination-time fixed effects. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. 39 Table B.5: Heterogenous migration response to BLAs along the economic momentum and the timing of the BLA Long run economic momentum Last 5 years (1) (2) (3) (4) ∗∗∗ ∗∗∗ ∗∗∗ Bilateral labor agreement (BLA) −0.646 0.351 0.445 0.434∗∗∗ (0.112) (0.110) (0.128) (0.109) BLA × Eco. stagnation (1973-1990) 0.450∗∗∗ 0.025 (0.140) (0.158) BLA × Post Cold War (> 1990) 1.906∗∗∗ 0.361∗∗ (0.184) (0.178) BLA × Signed in the last 5 years 0.225 0.245 (0.172) (0.157) BLA + interaction 1.710∗∗∗ 0.737∗∗∗ 0.671∗∗∗ 0.679∗∗∗ (0.204) (0.209) (0.138) (0.127) Observations 273, 042 273, 042 273, 042 273, 042 R-sq 0.820 0.875 0.820 0.875 Origin country x decade FE Y es Y es Y es Y es Destination country x decade FE No Y es No Y es Origin x destination countries FE Y es Y es Y es Y es Notes: Standard errors in parentheses are clustered at the origin-destination level. BLA (dummy) in decade N takes the value 1 if a BLA was signed in the last 10 years and zero otherwise. The dependent variable is the inverse hyperbolic sine of the number of dyadic migrations in protocols. Columns (1) and (2) report heterogeneity with respect to economic momentum that have over time. Columns (3) and (4) report the heterogeneous response of migration to BLA with respect to the timing of the BLA (whether they were signed during the last five years or not). For each of the highlighted categories, we consider two specifications: one that includes origin-destination and origin-time fixed effects, and another that complements this set of fixed effects with destination-time fixed effects. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. 40 Table B.6: Heterogenous migration response to BLAs (Mij > 10) GCC destination African origin Development group (1) (2) (3) (4) (5) (6) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ Bilateral labor agreement (BLA) 0.270 0.291 0.350 0.279 0.460 0.262∗∗ (0.056) (0.055) (0.062) (0.056) (0.122) (0.106) BLA X GCC destination 0.760∗∗∗ 0.582∗∗ (0.240) (0.232) BLA X African origin -0.317∗∗ -0.258∗∗ (0.124) (0.127) BLA X UM-HIC -0.205 -0.025 (0.137) (0.122) BLA + Interaction 1.030∗∗∗ 0.874∗∗∗ 0.033 0.021 0.255∗∗∗ 0.237∗∗∗ (0.233) (0.226) (0.107) (0.114) (0.062) (0.059) Observations 33,257 33,257 33,257 33,222 33,257 33,222 R-sq 0.848 0.860 0.848 0.897 0.848 0.897 Origin country x decade FE Y Y Y Y Y Y Destination country x decade FE Y Y Y Origin x destination countries FE Y Y Y Y Y Y Notes : Standard errors in parentheses are clustered at the origin-destination level. BLA (dummy) in decade N takes the value 1 if a BLA was signed in the last 10 years and zero otherwise. The dependent variable is the inverse hyperbolic sine of the number of dyadic migrations in protocols. Columns (1) and (2) report the heterogeneous response of migration to BLA depending on whether the region of destination is GCC or not. Columns (3) and (4) report heterogeneity with respect to whether the country of origin is African or not. The last two columns examine the heterogeneity with respect to the level of development of the country of origin (low and lower middle income, upper middle and high income). Analysis here is restricted to corridors with at least 10 migrants. For each of the highlighted categories, we consider two specifications: one that includes origin-destination and origin-time fixed effects, and another that complements this set of fixed effects with destination-time fixed effects. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. 41 C Heterogeneity-robust DiD with varying treatment start We now relax the implicit assumption of a constant treatment effect of the BLA on the corridors and over time relying on the recent advancement of the difference-in-difference (DiD) literature. However, the timing of treatment differs between dyads, which are characterized by different starting dates and durations. In this case, the assumption of constant treatment effects over time is potentially misleading (Dettmann, Giebler and Weyh, 2020). In this section, building on Nagengast and Yotov (Forthcoming), Nagengast, Rios-Avila and Yotov (2024), we combine a specification that uses a rich set of fixed effects (equation 1) with the latest heterogeneity-robust innovations in the difference-in-difference literature on treatment that occurs in multiple units in different time periods (Borusyak, Jaravel and Spiess, 2024, Callaway and Sant’Anna, 2021, de Chaisemartin and d’Haultfœuille, 2024, Sun and Abraham, 2021, Wooldridge, 2023). C.1 Empirical specification The aim here is to “nest ” the structural gravity model of migration with the heterogeneity- robust DiD methods developed by the new DiD literature (Roth et al., 2023). Starting with the equation (1), γ is the three-way fixed effects (TWFE) estimate of the effect of BLAs on migration. However, in the presence of heterogeneous treatment effects, which arise when the treatment is heterogeneous across groups or over time, which is precisely the case with BLAs signed between countries at different times, the TWFE estimator can produce biased estimates that are difficult to interpret.13 The source of the bias lies in the so-called “forbidden comparisons ”, where units treated in the early decades (already treated units) are included in the control group for units treated in later decades (Borusyak, Jaravel and Spiess, 2024). The goal here is to maintain the general structure of the TWFE specification, but to allow for treatment effect heterogeneity at the cohort-decade level by additionally introducing appropriate cohort and decade interactions.14 13 As shown by de Chaisemartin and d’Haultfœuille (2020), TWFE estimates are the weighted sum of the average treatment effects on the treated in each group and time period, with weights that can be negative. Consequently, TWFE may not identify a convex combination of treatment effects, i.e., the TWFE coefficient may be negative even though all individual treatment effects on the treated are positive. 14 As shown in de Chaisemartin and d’Haultfœuille (2020), Nagengast and Yotov (Forthcoming), Wooldridge (2023), this approach allows for arbitrary heterogeneity within cohort-decade cells, but can only identify simple average treatment effects. 42 In practice, starting with (1),15 we replace the single dummy variable for the presence of a BLA between i and j at decade t (BLAijt ) with the following expression: T T γcd Dcd , (5) c=q d=c where the dyad ij belongs to the treatment cohort c if the BLA onset was in year c, q is the first decade of treatment of cohort c, T is the last decade of the panel, Dcd is a time-varying treatment indicator that equals 1 for cohort c for d = t in the post-treatment decades and 0 otherwise, and γcd denotes the cohort-decade specific treatment effects. For example, γ1980,1980 captures the treatment effect of all BLAs signed in 1980 on migration within the first decade, while γ1980,1990 captures the treatment effect of the same BLAs on migration in the next decade. Accounting for the above mentioned issues, our main estimating equation becomes:   T T Mijt = exp αit + αjt + αij + γcd Dcd  ϵijt (6) c=q d=q Note that the only difference between the equations (1) and (6) is the modification of the BLAs. Estimating the equation (6) first requires estimation of treatment effects at the detailed level, which is standard in the recent literature on DiD estimation under staggered treatment effects (Borusyak, Jaravel and Spiess, 2024, Callaway and Sant’Anna, 2021, de Chaisemartin and d’Haultfœuille, 2020, Sun and Abraham, 2021). However, since the goal is often to estimate the aggregated average effect of the treatment rather than the cohort-time-specific treatment effects as indicated in equation (6), our main estimation objective is a single BLA coefficient defined by the weighted sum of the estimated cohort-time-specific treatment effects as: T T Ncd γ= γcd , (7) ND c=d d=c where all post-treatment observations are given equal weight, corresponding to the number of observations of cohort c in decade d, Ncd , relative to the total number of treated observations, T T ND = c=d d=c Ncd . Similarly, the standard errors are computed as a weighted linear com- bination of the cohort-decade specific effects, taking into account the covariance between the coefficients. 15 Note that equation (1) includes both not yet treated and never treated units as controls. 43 C.2 Results This section presents the results in three steps. First, we test the identifying assumption of the gravity-based ETWFE estimator (using the STATA command jwdid developed by Rios-Avila, Nagengast and Yotov, 2022). Second, we present and comment on the main average treatment effects. Third, we present event-study estimates that describe the evolution of the BLA effects over time, as well as a range of cohort-specific BLA estimates. C.2.1 Identifying assumptions: parallel trends and no anticipation assump- tions. Figure (C.1) provides evidence against the violation of the identifying assumptions by showing the aggregate placebo effects in event time relative to treatment onset. The pre- treatment coefficients are all zero and a joint hypothesis test is insignificant (p-value of 0.6615). Therefore, we conclude that there is no evidence of pre-trends in our data and proceed with the analysis. C.2.2 Aggregate effect of BLAs on migration. The aggregate effects of BLAs on migration, reported in Table (C.1), reveal that, across various samples, applying the new DiD approach consistently produces BLA estimates that are significantly larger than those derived from the traditional TWFE methods used in the gravity literature. This finding aligns with results from the trade literature, as shown by Nagengast and Yotov (Forthcoming), regarding the impact of regional trade agreements, and by Nagengast, Rios-Avila and Yotov (2024) regarding the impact of membership in the European Union single market on international trade. Building on this observation, our benchmark estimates likely represent a lower bound of the actual effect of bilateral labor agreements on migration flows. Indeed, focusing on the estimates from the regular corridors only, our benchmark effects obtained using standard TWFE techniques are potentially 42% downward biased due to “forbidden comparisons ” that mis(use) the already treated units in the control group (de Chaisemartin and d’Haultfœuille, 2020). C.2.3 Dis-aggregated effects of BLAs on migration. We complement the average aggregated effect from Table (C.1) with a number of disaggregated results. We begin with an event study analysis of the evolution of the effects of the BLAs on migration over time from the start of the treatment. The results are shown in Figure (C.2), which are computed from 44 Figure C.1: Pre-treatment effects Source : Authors’ calculations Notes : The figure reports the pre-trend estimates from equation (6) in which treatment effects are replaced by cohort-decade-specific treatment placebo treatment effects prior to treatment onset. To prevent misleading visual interpretations, we follow Roth (2024) and put our pre-treatment estimates in a different plot from the post-treatment estimates. The equation is estimated with never treated units only following Borusyak, Jaravel and Spiess (2024). The cohort-decade-specific treatment effects d ¯ .d = c=q N are aggregated across cohorts to obtain the event-time specific treatment effects as γ N.d γ.cd, cd d where N.d = c=q Ncd is the total number of treated observations in decade d. 95% confidence intervals are shown using standard errors clustered at origin-decade level. cohort-time-specific treatment effects by averaging over the cohort dimension. Again, the results are consistent with our benchmark results reported in Figure (1), and indicate that the effects of entering a BLA reported in Table (C.1) persist for about ten years from the start of the treatment. C.2.4 Heterogeneous effects of BLAs on migration across regions. Consistent with our benchmark findings, the application of the heterogeneous-robust DiD approach largely reproduces our benchmark sub-group results. As shown in Appendix C.3, the analysis reveals that, on average, the entry of an African country into a BLA does not lead to significantly larger migration flows from the African country of interest to the partner country (Panel C.3a). In contrast, such agreements are associated with increased migration flows for non-African countries (Panel C.3b). Regarding destinations, non-GCC destinations generally replicate the 45 Table C.1: Comparison of TWFE and ETWFE estimators for different samples Full sample Regular corridors only (Mij >= 10) TWFE ETWFE TWFE ETWFE OLS PPML OLS PPML OLS PPML OLS PPML (1) (2) (3) (4) (5) (6) (7) (8) BLAijt 0.569∗∗∗ 0.179∗∗∗ 1.196∗∗∗ 0.190∗∗∗ 0.243∗∗∗ 0.158∗∗∗ 0.979∗∗∗ 0.170∗∗ (0.0903) (0.0373) (0.0952) (0.0666) (0.0511) (0.0387) (0.0944) (0.0684) Observations 273, 042 163, 132 272, 531 162, 631 33, 222 33, 222 32, 858 32, 858 Origin × decade FE Y es Y es Y es Y es Y es Y es Y es Y es Destination × decade FE Y es Y es Y es Y es Y es Y es Y es Y es Origin × destination FE Y es Y es Y es Y es Y es Y es Y es Y es Notes: The table presents OLS and PPML estimations using three-ways fixed effects (TWFE) and extended three-ways fixed effects (ETWFE) es- timators. ETWFE estimator is obtained by aggregating the cohort-time specific treatment effects across cohort to obtain and aggregate treatment effect estimate. The dependent variable is bilateral migration flows, which varies across origin-destination-decade dimension. The full sample contains all the corridors irrespective of whether those are used (presence of migrants) or not. The regular corridors restricts the analysis to corridors with at least 10 migrants. Standards errors in parenthesis are clustered at origin-decade level. ∗∗∗ , ∗∗ , and ∗ indicate the significance at the 1%, 5% and 10% level, respectively. Extended three-ways fixed effects (ETWFE) estimates are obtained using the STATA jwdid command built by Rios-Avila, Nagengast and Yotov (2022) from Wooldridge (2023) and extended to gravity setting by Nagengast and Yotov (Forthcoming). Figure C.2: Event-decade-specific effects of signing a BLA Source : Authors’ estimates. Notes : The figure reports event-time-specific treatment effects, obtained by aggregating the the cohort- d Ncd decade-specific treatment effects are aggregated across cohorts as γ ¯ .d = c=q N.d γcd , where N.d = d c=q Ncd is the total number of treated observations in decade d. 95% confidence intervals are shown using standard errors clustered at origin-decade level. 46 benchmark effect (Panel C.3d) presented in Appendix C.2. However, for GCC destinations, while point estimates remain larger—consistent with the benchmark—the confidence intervals are wider, rendering these estimates statistically insignificant. Figure C.3: Regional variations in migration response to BLAs (a) African origin (b) Non African origin (c) GCC destination (d) Non GCC destination Source : Authors’ estimates. Notes : The figure plots the impact of signing a BLA on migration, estimated using equation (3), for African and non-African origin countries (Panels C.3a and C.3b) and for GCC and non-GCC destination countries (Panels C.3c and C.3d). Vertical bars indicate 95 percent confidence intervals, standard errors are clustered at the origin × destination cell. 47