Migration networks and the intensity of global migration flows, 1990–2015

ABSTRACT

Evidence from 184 countries over a 25-year span is used to analyze international migration flows based on their ‘intensity’ or size. Results show that network dynamics are relevant to understand the emergence and evolution of migration flows of varying intensity. The findings also show that a core set of non-network covariates derived from Migration Systems Theory is important to understand migration flows of different intensity. Overall, the results highlight the significance of explicitly analyzing the intensity of international migration flows since not all covariates affect these flows in the same way.

KEYWORDS:

• International migration
• Migration Systems Theory
• migration networks
• network science
• computational social science

Data availability statement

The authors confirm that the scripts used to generate results are available from the first author’s GitHub site: https://github.com/diegoFLeal/migIntensity.git. Links to all data sources are in the article’s appendix.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 In MST, geographic propinquity is a proxy for regulatory linkages since state-to-state migratory relations are considered to be ‘enhanced and even insulated by geographic isolation’ (DeWaard, Kim, and Raymer 2012, 1327–1328; see also Fawcett 1989). Think, for instance, about countries in South America which belong to one of the two main regional supranational organizations: Comunidad Andina de Naciones (CAN) or Mercado Común del Sur (MERCOSUR). In that context, even though Peru belongs to CAN and Argentina belongs to MERCOSUR, they still share migration agreements that ultimately shape their migratory dynamics (Geronimi, Cachón, and Texidó 2004). Geographic propinquity (e.g., shared region, shared borders) is thus a proxy for regulatory linkages beyond formal membership in supranational organizations.

2 Readers interested in the technical details related to the estimation and validation of global international migration flows are also encouraged to see: Özden et al. 2011; Abel 2013; Abel and Sander 2014.

3 Results, available on request, suggest that the findings reported here are robust if the same cutoff values are used across all networks by, for instance, using the cutoff points derived from the first migration network, that is, the 1990–1995 network.

4 These correspond to the 030T, 030C, and 120C triads in the standard triad classification of Davis and Leinhardt (1972).

5 Even though geography is important to understand migration patterns (DeWaard, Kim, and Raymer 2012; Deutschmann 2016), geographic distance was not included to avoid collinearity problems. Descriptive statistics for all non-network (exogenous) covariates, including all MST covariates, are reported in Appendix 1. Links to all original data sources are available in Appendix 2. A table showing our regional classification for each country can be found in Appendix 3.

6 The following IGOs were included in the analysis: Andean Community, Arab Maghreb Union, Asia-Pacific Economic Cooperation, Association of Southeast Asian Nations, Caribbean Community, Economic Community of West African States, Eurasian Economic Community/ Eurasian Economic Union, Euro Free Trade Association, European Union, Gulf Cooperation Council, North American Free Trade Association, MERCOSUR, Central American Integration System, Common Market for East and South Africa, Southern African Development Community.

7 If a pair of countries shared at least one official language they were coded as a 1. Otherwise, a 0 was assigned.

8 Density = L/g (g-1), where L is the number of directed ties, g is the number of nodes present in the network, and g(g-1) is the total number of possible ties in the network (Wasserman and Faust 1994).

9 According to Wasserman and Faust (1994), degree centralization (C*) is defined as: ∑_i∈V(G) |max_v∈V(G)(C(v))–C(i)|, where v is the vector of degree centralities defined on graph G, C(i) is a single point centrality in the vector, and max_v∈V(G) C(v) is the largest value of C in the context of graph G.

10 All models were estimated with RSiena version 1.2-12. We used the method of moments with three phase 2 subphases. The (default) score function with 5,000 phase 3 iterations was used to calculate standard errors. The overall maximum convergence ratio for all models was always below 0.17, and the absolute value of the t-ratios for all coefficients in all models were always below 0.1. These two metrics suggest excellent model convergence that meets the standards to publish results (Ripley et al. 2020). No evidence of multicollinearity was found in our models since both parameter values and standard errors were stable across model runs, and the absolute value of the correlations between the estimated parameter in a given model were always below 0.9 (Ripley et al. 2020).