Scripts, data, and replication materials for "Global subnational estimates of migration of scientists reveal large disparities in internal and international flows"

This repository includes replication materials, including data, Python, and R scripts to replicate the article's analysis and figures (shown further below) with the following metadata.

Script authors/maintainers: Aliakbar Akbaritabar, Maciej J. Dańko

Contact: akbaritabar@demogr.mpg.de; danko@demogr.mpg.de

Article title: Global subnational estimates of migration of scientists reveal large disparities in internal and international flows

Published in: Proceedings of the National Academy of Sciences of the United States of America

Manuscript authors: Aliakbar Akbaritabar, Maciej J. Dańko, Xinyi Zhao, Emilio Zagheni

Article DOI: https://doi.org/10.1073/pnas.2424521122

Abstract: Researchers are key contributors to innovation. Their migration results in talent circulation and recombination of ideas. Due to data shortage, little is known about subnational mobility of scientists and the interrelationships between their internal and international migration patterns. We used data on 30+ million Scopus publications of 19+ million authors to infer migration from changes in affiliations. Our publicly shared database of global subnational estimates of migration reveals relevant disparities in the attractiveness of subnational regions. While, at the country level, some countries have acted as a global hub that attracts scholars from throughout the world, at the subnational level, some of their regions have negative net migration rates, with implications for scientific output and regional development, as well as the perpetuation of migration corridors. On average, subnational inequalities in attracting and sending scholars have increased for international but decreased for internal migration. In most countries, there is no single trend such that all subnational regions have been sending or receiving more scholars. Instead, a mix of patterns has been simultaneously at work, on the backdrop of globalization of migration, which is an asymmetric process where specific regions and sub-populations have higher access to international migration. For most subnational regions, when they are attractive for international migrants, they are also attractive for internal migrants, which is not always the case for emigration. Our results have implications for the global circulation of academic talent by adding the dimension of internal migration to ‘brain drain’ and ‘brain circulation’ in international migration.

Migration of researchers measures at subnational level and year

If you are interested in downloading only the migration measures calculated for researchers per subnational region and year, use the CSV file named "migr_19982017.csv" under the `results` folder.

Our repository (accessible here: `https://github.com/akbaritabar/Internal-and-international-migration-of-scientists`) and replication materials include two parts.

The first is preparing the main data for subnational migration. This includes scripts, data, and the reproducible pipeline that creates the final data for our manuscript. This includes the preparation of migration data and measures, visualizations for subnational maps (see figure 1 in the manuscript, also copied further below, and multiple figures in the SI disaggregated by field of science, see them under the results folder). These scripts are developed in Python, and the reproducible workflow (described below) uses SnakeMake workflow management tool to ensure full replicability.

The second part includes R scripts for statistical analysis, figures 2-4 in the manuscript (see them further below), and multiple figures in the SI (see them under the results folder).

Below, we describe the requirements for each of these.

These scripts are written by Aliakbar Akbaritabar.

For the reproducible pipeline to recreate the paper's replication data and migration measures, you need to have SnakeMake version 8 or above installed. Creating a conda environment with Python 3 (3.11.9 was used here) and the following libraries (for instance by running conda env create -f requirements.yml and after copying the following yml code into a file: "requirements.yml") enables reproducing the pipeline by opening CLI, activating the conda environment i.e., conda activate subnational, and running a dry-run with snakemake -np all or a full reproduction with snakemake --cores 4 all.

An HTML report shows the Directed Acyclic Graph (DAG) of the dependency of the steps in the pipeline (rules in SnakeMake lingua) which is accessible in https://akbaritabar.github.io/Internal-and-international-migration-of-scientists/. Please see below an example of how this DAG looks like. Clicking on a step (rule)'s node opens a sidebar with more information about the input, output, log, and script that is used in that step. An example step (rule) is shown in the second photo below.

Please note that some of the first rules in data preparation require Scopus data at the individual level, which is licensed, and we cannot share them publicly. Hence, these output files are commented out from the main Snakefile and rule all. However, the aggregated data to prepare the migration measures at the subnational level and recreate all our figures and statistical analysis in the manuscript are included in this repository, which completely complies with the license terms of the data provider.

How to replicate the workflow and exclude all the steps that rely on licensed Scopus data?

To ease the replication process, I have added a separate branch on the GitHub repository (`https://github.com/akbaritabar/Internal-and-international-migration-of-scientists`) called CodeRelyingOnLicensedDataExcluded. In this branch, all rules and steps that rely on licensed Scopus data are commented out and excluded so that one could install the requirements as described above and do snakemake --cores 4 all to only replicate the figures and tables using the data provided in this repository. Hence, after you clone the repository, do git checkout branch CodeRelyingOnLicensedDataExcluded and a dry-run with snakemake -np all or a full reproduction with snakemake --cores 4 all.

```yaml

name: subnational
channels:
  - conda-forge
  - bioconda
  # to prevent using default anaconda channels
  - nodefaults
dependencies:
- python 3.*
- snakemake
- pandas
- plotnine
- argparse
- geopandas
- numpy
- mizani # (is installed with plotnine)
- duckdb
- tabulate
- pyarrow
- pip:
  - pycountry_convert # (from Pypi, if it gave an error, comment it out by adding a pound sign, and after installation, run "pip install pycountry_convert" in CLI after you activated the 'subnational' conda environment)

These R scripts for statistical analysis are written by Maciej J. Dańko.

To replicate the statistical analysis, the following packages should be installed. The scripts and the data used in them are in the folder workflow\scripts\Danko\. Please note that due to GitHub's 100MB limit on file sizes, only some of the example results from statistical models are included. However, all of these files could be replicated using the data provided and re-running the scripts on your machine.

```R

required_packages <- c("brms", "data.table", "deming", "DescTools", "dplyr", "Kendall", "lava", "magicaxis", "mgcv", "openxlsx", "parallel", "purrr", "splines", "stats")

missing_packages <- required_packages[!(required_packages %in% installed.packages()[, "Package"])]

if (length(missing_packages) > 0) {
  install.packages(missing_packages)
}

z<-lapply(required_packages, require, character.only = TRUE)
rm(z)