PaperlayerView exact revision on arXiv

Source-linked AI summary

Catch-Up and Come Home: Economic Convergence and Return Migration

Kinga Varga, Bence Kollányi, Johannes Wachs

arXiv:2609.31236v1econ.GNcs.SIq-fin.EC

TL;DR

Evidence remains limited on how economic conditions predict high-skilled workers’ return migration, an outcome that matters to sending countries. Using GitHub location histories, the paper finds higher return hazards with greater catch-up growth and projects further changes under IMF growth paths.

  • Problem

    Evidence is scarce on the extent of high-skilled return migration and on which sending-country or country-pair characteristics predict its variation.

    Source: §1 · p. 2, §2 · p. 5

  • Method

    The authors use ten waves of geocoded GitHub profiles to track developers’ moves and estimate return hazards using within-corridor variation in income convergence.

    Source: §1 · pp. 2–3, §3.2 · p. 7

  • Results

    0.315 (p < 0.01) is the cumulative-convergence coefficient in the corridor and year fixed-effects hazard model; by 2035, return is projected to rise roughly a tenth for India, Vietnam, and Nepal.

    Source: §1 · p. 3, §8 · p. 17

  • Takeaways & Limitations

    Within corridors, developers experiencing more catch-up growth after leaving were more likely to return, and projected increases are concentrated where diasporas are largest.

    Source: §8 · p. 17

    The observed developers are a self-selected subset, and estimates remain sensitive to time-varying corridor shocks correlated with growth.

    Source: §8 · p. 18

Abstract

from arXiv · show

The return migration of skilled workers provides origin countries many important benefits. The likelihood of return migration is thought to depend on relative economic opportunities, but quantitative evidence of its relative importance remains limited. Here we measure return migration rates using the public profiles of software developers on GitHub, geocoded across ten waves between 2012 and 2026, in a panel containing more than 270,000 international movers. Within five years of departure, 8.8% of movers have returned. Origin-country rates range from under 1\% to 16\%. Hazard models with corridor fixed effects show that migrants whose home countries experience greater catch-up growth return at significantly higher rates. Immutable features of country pairs such as cultural or geographic distance do not explain the estimate. We forecast future return rates using IMF growth projections, predicting, for example, that India's return rate will rise 12\% by 2035 and China's 7\%. The difference is driven mostly by India's faster projected growth at home. As developing economies close income gaps with the destinations their emigrants have settled in, they recover more of the workers they lost.

1 Introduction

The paper addresses scarce global evidence on what predicts high-skilled return migration, using developer traces to test whether income convergence shapes return. It finds higher return hazards with catch-up growth and projects varied changes through 2035.

1 Introduction

  • Return migration is hard to measure and its predictors remain poorly established across origins and destinations, although returnees can benefit sending countries.The paper contributes global, individual-level evidence on convergence and uses within-corridor variation to estimate return determinants and forecast rates.Source: §1 · pp. 2–3
  • The study uses GitHub profile locations to track software developers’ international moves and models annual return hazards against income-gap closure within corridors.The panel draws on ten geocoded location waves from 2012 to 2026.Source: §1 · p. 2
  • 8.8% of movers had returned within five years, with origin-level rates ranging from under 1% to 16%.The panel contains 277,502 people observed in at least two countries, of whom 13,670 returned to their first-observed country.Source: §1 · p. 2
  • With corridor and year fixed effects, cumulative convergence raises the return hazard (coefficient 0.315, p < 0.01).India’s income catch-up with the United States between 2015 and 2025 corresponds to about a 10% increase in its developers’ annual return rate.Source: §1 · p. 3
  • Forecasts under IMF growth projections show return rates rising 12% for India and about 7% for China, but declining for Japan and other advanced economies.India’s projected home growth outpaces its diaspora’s destinations; China shares similar destinations but has slower home growth.Source: §1 · p. 3
  • Observed return rates are lower bounds because many movers enter the panel late and self-reported locations update with a lag.The authors therefore do not treat 8.8% as the full five-year return rate for skilled workers.Source: §1 · p. 3

2 Return migration and its measurement

The section explains why return rates matter, why they are difficult to measure, and how migration theory links return timing to economic convergence. It motivates using individual digital traces to test that prediction across many corridors.

2 Return migration and its measurement

  • Temporary-migration models predict return as origin incomes catch up with destination incomes, making convergence a prediction about return timing within a corridor, not corridor-level return differences.Migrants may also return when accumulated savings or skills become more valuable at home than another year abroad would add.Source: §2 · p. 4
  • Selection into migration complicates aggregate tests because who emigrates and who returns are selected, and their composition can shift with economic conditions.Individual migration-event data can capture spell durations abroad rather than mixing return decisions with changes in who was abroad.Source: §2 · p. 4
  • Returnees are linked to gains in science, firms, public life, and origin-country development.Some diaspora benefits also arrive before migrants return through knowledge and capital flows.Source: §2 · pp. 4–5
  • The aggregate value of return migration depends on both the effect per returnee and the number who return, making return rates and their determinants important to estimate.Source: §2 · p. 5
  • A return requires observing someone at home, abroad, then home again; population registers rarely record all three, motivating digital traces.Digital traces include email logins, social-media profiles, and professional networks, while their timing structure shapes what can be inferred.Source: §2 · p. 5
  • The paper uses open data on software developers to estimate individual return durations across thousands of corridors and test the convergence mechanism directly.Developers are a tractable case because platform data document their migration and employment histories have supported earlier large-scale return estimates.Source: §2 · p. 5

3 Data: a global location panel of software developers

The paper builds a global panel of software-developer locations from public GitHub profiles and uses it to measure migration and return across careers. It defines return relative to each developer’s first observed country, accounts for uneven observation windows, and measures income convergence within origin–destination pairs.

3.1 Source and construction

  • GitHub location profiles are assembled into ten waves spanning 2012–2026, then geocoded to countries to create the developer location panel.The raw panel covers 44 million accounts and 113 million person-wave observations; after filtering, it contains 42.6 million located person-waves for 10.0 million developers.Source: §3.1 · p. 6

3.2 Measuring migration and return

  • A mover is observed in at least two countries; a return occurs when a developer reappears in their first observed country after living abroad.The panel contains 277,502 movers across more than 8,100 origin–destination corridors, and 13,670 make an observed return.Source: §3.2 · pp. 6–7

3.3 Observation windows and external validity

  • Pooled return rates understate lifetime circulation because two thirds of movers departed in 2024 or later and had at most two years to return.The analysis compares movers within cohort and corridor rather than relying on pooled levels; observed rates are lower bounds on lifetime circulation.Source: §3.3 · p. 7

3.4 Covariates

  • Cumulative convergence measures how much the origin has closed its log income gap with the destination since departure, using GDP per capita at PPP.Positive values indicate faster origin growth since departure, values near zero indicate parallel growth, and negative values indicate a widening gap; historical series support estimation, with IMF projections used only for forecasts.Source: §3.4 · pp. 7–8

4 The global landscape of developer return migration

Developer migration is concentrated in a small number of origins and destinations, while five-year return rates vary substantially across origins. The largest sending countries do not uniformly have high return rates, and India combines the largest diaspora with a low observed return share.

4 The global landscape of developer return migration

  • The United States receives close to one in five movers, while India is the largest origin and the India–United States corridor is three times the next largest.The five largest destinations absorb just under half of all movers.Source: §4 · p. 8
  • Five-year return rates range from under 1% to 16%, with Spain highest among origins sending at least 1,000 movers and India below 5%.Figure 1 pools rates for 91 origins with at least 50 eligible movers; the global rate is 8.8%.Source: §4 · p. 8, Figure 1 · p. 9
  • India has the largest diaspora, about 33,000 developers abroad at the 2026 baseline, but only a small share has returned so far.The passage notes that even a modest rise in India’s return rate would bring back many developers.Source: §4 · p. 9

5 Methods and Results

Within-corridor hazard models find that greater cumulative convergence is associated with a higher return hazard. In the corridor specification, one unit of convergence raises return probability by about 0.81 percentage points against a 2.5% base rate.

5 Methods and Results

  • The discrete-time logit hazard models the per-period return event using cumulative convergence and years since departure, with corridor and wave-year effects in the most demanding specification.Intermediate specifications replace corridor effects with origin and destination country effects; standard errors are clustered by origin and destination.Source: §5.1 · pp. 9–10
  • 490,756 person-periods remain after excluding observations with undefined convergence and, under corridor fixed effects, corridors whose outcome never varies.The undefined measures arise when origin or destination GDP-per-capita series are not continuous; most dropped corridors had no observed return.Source: §5.1 · p. 10
  • +0.315 (SE 0.098, p < 0.01): convergence has a positive coefficient with corridor fixed effects, after the pooled estimate changes sign with added country controls.Without country controls, cross-corridor differences dominate; the corridor specification compares movers who took the same route.Source: §5.1 · p. 10
  • 0.81 percentage points: one unit of cumulative convergence raises return probability in the corridor specification, against a 2.5% base rate.A full unit requires the origin economy to grow roughly 2.7 times as fast as the destination over the period abroad, which few corridors experience.Source: §5.1 · p. 10, Table 2 · p. 10
  • About a ten percent increase in the return rate corresponds to India’s roughly +0.35 catch-up with the United States between 2015 and 2025.Figure 2 reports average marginal effects across specifications with 95% confidence intervals.Source: §5.1 · p. 10, Figure 2 · p. 11
  • 0.011 (SE 0.003, N = 534,676): a linear probability model likewise estimates that one unit of convergence raises return probability.Source: §5.1 · p. 10

6 Distance, culture, and policy

The convergence effect remains positive when the models account for fixed bilateral traits and destination integration policy. Alternative samples suggest that profile-location errors and origin misassignment may make the full-sample estimate an underestimate.

6 Distance, culture, and policy

  • Convergence remains positive and significant when cultural and geographic distance enter as controls or corridor fixed effects absorb fixed bilateral attributes.Geographic distance and time-zone offset do not predict return once country composition is held fixed; cultural distance is negatively associated with return (−0.069, p < 0.001).Source: §6 · p. 11, Table 3 · p. 12
  • Destination integration policy does not moderate the convergence effect where MIPEX scores are available; policy level itself does not predict return.The convergence effect is unchanged when destination policy is added, and the policy and convergence terms do not interact.Source: §6 · pp. 11, 13
  • Convergence remains positive and significant using constant-PPP dollars instead of current dollars.Current international dollars remain the headline measure because the series extends through 2025, matches the IMF projections, and reflects the purchasing-power comparison a developer faces; developer pay tracks GDP per capita at PPP closely across countries.Source: §6 · p. 13

6.1 Convergence estimates and alternative subsets

  • The full-sample coefficient is likely an underestimate because first-observed homes can misstate prior migration and self-reported locations can lag or omit moves home.Both errors work against finding an effect.Source: §6.1 · p. 13
  • +0.424 (p < 0.01) after excluding movers whose names contradict their coded origin, and +0.301 (p < 0.01) among movers observed at home in at least two pre-departure waves.The first estimate is about 35% above the headline; the second covers 86% of movers and is close to it. Name matching loses precision as the sample shrinks, and the matched sample is selected rather than random.Source: §6.1 · p. 13
  • +0.448 (p < 0.05): among movers with current profiles, the corridor-model convergence coefficient is about 42% above the headline estimate.Some increase may reflect which movers remain visible, not only correction of missed returns; stale profiles lower the measured convergence effect.Source: §6.1 · p. 14

7 The convergence channel to 2035

The paper projects return-rate changes through 2035 by applying an estimated convergence effect to corridor-level baselines under IMF growth paths. The projections vary with origin growth, diaspora destinations, and the income gap between them.

7 The convergence channel to 2035

  • The forecast shifts each corridor’s 2026 baseline hazard by exp(β ∆), using β = 0.315 and additional convergence projected under IMF growth paths.The projections use IMF real GDP per-capita data and apply one pooled coefficient across corridors.Source: §7 · p. 14

7.1 Growth and return

  • Projected changes depend on origin growth, diaspora destinations, and destination growth, while the pooled coefficient averages over corridors that may respond differently.Relative changes are more informative than levels under this pooled-coefficient forecast.Source: §7 · p. 14, §7.1 · p. 14
  • A one-percentage-point-per-year faster narrowing of the income gap for a decade implies a 3.2% increase in the annual return rate in the model.On a 2.5% base, this corresponds to about 0.08 percentage points; half that differential implies roughly 1.6% more return.Source: §7.1 · p. 14
  • India’s projected return increase is 11.9%, compared with 6.8% for China, mainly because India is projected to grow about 1.5 percentage points faster annually.Both countries’ diasporas are concentrated in high-income destinations projected to grow at similar rates.Source: §7.1 · pp. 14–15
  • The stock-weighted increase across origins is 2.3%; projected returns rise 11.9% for India and fall where destinations outgrow origins.Japan’s projected change is −1.8% and South Africa’s is −2.9%.Source: §7.1 · p. 15, Table 4 · p. 16, Figure 3 · p. 17

8 Discussion and conclusion

The discussion connects economic convergence with return migration and the potential value of skilled-worker circulation. It emphasizes that return rates respond to growth while noting limits in the data and identification.

8 Discussion and conclusion

  • Within corridors, developers experiencing more catch-up growth after leaving are more likely to return, and projected gains are largest in origins with large diasporas.Under IMF growth paths, return is projected to rise by roughly a tenth for India, Vietnam, and Nepal by 2035.Source: §8 · p. 17
  • For origins with young or shrinking workforces, return can add mid-career talent or offset population aging, respectively.The authors project that the assumption that skilled immigrants mostly stay will become less accurate on corridors supplying the most talent.Source: §8 · p. 17
  • Aggregate gains from return depend on both the value each returnee creates and the number who return, which the estimates show responds to growth.Prior studies described in the passage report benefits including higher patenting, colleague output, firm value, and development outcomes.Source: §8 · p. 18
  • Returnees can carry tacit knowledge and skills home, while the authors argue economic growth exerts a larger pull on return than targeted policy.The discussion also supports removing barriers to return and circulation and notes that returnee contributions are greatest where investment climates and institutions are sound.Source: §8 · p. 18
  • The developer sample is self-selected, estimates remain sensitive to time-varying corridor shocks correlated with growth, and the measurement approach is conservative.The authors state that corrections for inferred home locations raise the convergence estimate by a third or more.Source: §8 · p. 18

Data and ethics statement

The study uses public platform profiles and reports its analysis at country and corridor levels. The authors state that no individually identifying information is reported.

Data and ethics statement

  • The analysis uses public GitHub and Stack Overflow profiles obtained from the GHTorrent research archive and platform interfaces.The authors state that collection followed platform terms of service.Source: Data and ethics statement · p. 18
  • User-reported locations are analyzed at country and corridor levels, with no individually identifying information reported..Source: Data and ethics statement · p. 19
  • Aggregated corridor-level data and replication code are to be made available upon publication..Source: Data and ethics statement · p. 19

Declaration of generative AI and AI-assisted technologies

The authors report using generative AI tools in analysis, manuscript preparation, and geocoding-pipeline research. They reviewed and edited the content and accept responsibility for the publication.

Declaration of generative AI and AI-assisted technologies

A Data pipeline and geocoding audit

The appendix builds a multi-wave country-location panel, audits its geocoding, and defines returns from observed country sequences. The audit finds high assignment accuracy, while mover histories are independently recounted.

A.1 From raw waves to the location panel

  • The ten waves yield 42.6 million located person-waves for 10.0 million developers, after filtering bots and organizations and imputing missing-wave locations from nearest reports.Source: §A.1 · p. 19

A.2 Geocoding

  • A 1.2-million-string curated dictionary geocodes locations; fallback requires three prefix-sharing entries to agree, while ambiguous names use modal developer usage.Source: §A.2 · pp. 19–20

A.3 A blind audit of geocoding quality

  • 97.6% of strings and 99.7% of sampled users are assigned correctly; residual errors are inherited-country assignments for non-place strings, not misassigned real places.Source: §A.3 · p. 20

A.4 Movers and returns

  • A mover is observed in at least two distinct countries; return is the first subsequent wave at home, and an independent recount reproduces 277,502 movers and 13,670 returns.Source: §A.4 · p. 20

B Measurement corrections: home assignment and profile currency

Measurement checks yield larger convergence estimates in selected corrected samples, while cultural distance does not moderate the association. A salary comparison supports GDP per capita as a proxy for developer pay.

B Measurement corrections: home assignment and profile currency

  • 30-40% higher convergence-effect estimates under top-5 name-match and profile-update filters are treated as bias-direction checks, not preferred estimates.Name cuts omit some correctly coded developers, while corroborated and updated-profile samples over-represent returnees.Source: B Measurement corrections: home… · pp. 20–21, Table 5 · p. 21

C Cultural distance does not moderate convergence

D Does GDP per capita proxy developer pay?

E Supplementary exhibits

  • The exhibits rank five-year return rates for 53 origins with at least 1,000 movers and provide geographic- and cultural-distance ladders.Five-year rates use the 2016, 2019, and 2021 departure cohorts, unlike pooled rates biased down by recent departures.Source: E Supplementary exhibits · p. 22, Table 8 · p. 24
Loading 2609.31236v1…