Software complexity and regional capabilities: a subnational typology of Argentina

Author: Raimundo Elias Gomez Affiliation: CONICET / Facultad de Humanidades y Ciencias Sociales, Universidad Nacional de Misiones (UNaM) Contact: lsgomez001@gmail.com ORCID: 0000-0002-4468-9618

---

Overview

This repository contains the data, analysis scripts, and figures for the article "Software complexity and regional capabilities: a subnational typology of Argentina", submitted to Regional Studies in May 2026 (previously submitted to the Journal of Economic Geography, February 2026, JOEG-2026-083).

The study constructs an Economic Complexity Index for software production (ECI_software) at the level of 224 Argentine departments using a bipartite network of departments and 87 programming languages derived from 229,270 geocoded GitHub repositories. A three-stage analytical strategy — Multiple Correspondence Analysis (MCA), Hierarchical Agglomerative Clustering (CAH), and type-specific regressions — examines how the determinants of software complexity vary across six territorial types. Two robustness checks are included: (1) a bundle-based ECI recomputed after mapping individual languages onto the 38 software clusters identified by Juhász et al. (2026), confirming Pearson r = 0.90 with the main measure; and (2) a longitudinal stability analysis computing ECI for cumulative annual snapshots from 2015 to 2025, showing consecutive-year rank correlations of 0.82–0.97 and 68 per cent top-quintile persistence from 2020 to 2025.

Repository structure

github-subir/
├── README.md
├── data/                         # Processed datasets and summary tables
│   ├── departments_full.csv      # All 511 departments: MCA coords, clusters, ECI, census vars
│   ├── bipartite_matrix.csv      # 224 depts x 87 languages (repo counts, filtered)
│   ├── rca_binary_matrix.csv     # 224 x 87 binary RCA matrix (threshold >= 1)
│   ├── eci_ranking_FINAL.csv     # ECI ranking for 224 departments
│   ├── table_01_eci_ranking_full.csv     # ECI ranking with sociodemographic variables
│   ├── table_02_pci_ranking_languages.csv # PCI ranking for 87 programming languages
│   ├── table_03_cluster_profiles.csv      # Mean profiles of 6 departmental types
│   ├── table_04_regression_summary.csv    # Regression coefficients by type
│   ├── table_05_key_numbers.csv           # Summary statistics (key-value)
│   ├── table_06_crossvalidation_geo.csv   # Geospatial cross-validation (511 depts)
│   ├── regression_output_FINAL.txt        # Full regression output (text)
│   ├── juhasz_language_clusters.csv       # Juhász et al. (2026) language-to-bundle mapping
│   ├── table_s_bundle_robustness.csv      # Table S5: ECI_individual vs ECI_bundle (224 depts)
│   └── eci_panel_2015_2025.csv            # Longitudinal ECI panel (1,504 dept-year obs)
├── figures/                      # Article figures (300 DPI)
│   ├── fig_01_pci_ubiquity.png           # Figure 1: PCI vs ubiquity (87 languages)
│   ├── fig_02_mca_biplot.png             # Figure 2: MCA biplot (Axes 1-2, N=511)
│   ├── fig_03_cah_mca_clusters.png       # Figure 3: Six types in MCA space
│   ├── fig_04_cluster_maps.png           # Figure 4: Spatial distribution of types
│   ├── fig_05_eci_vs_devs.png            # Figure 5: ECI vs developer density
│   ├── fig_06_forest_plot.png            # Figure 6: Forest plot of betas by type
│   ├── fig_S1_dendrogram.png             # Figure S1: Ward's dendrogram (k=6)
│   ├── fig_S2_diagnostics_panel.png      # Figure S2: MCA scree + cluster quality
│   ├── fig_panel_trajectories.png        # Figure S3: ECI trajectories 2015-2025 by type
│   └── fig_panel_rank_stability.png      # Figure S3 supplement: Spearman rho heatmap
├── scripts/                      # Analysis pipeline (Python)
│   ├── 00_build_schema.py        # Stage 0: Integrate data sources into art1 schema
│   ├── 01_compute_eci.py         # Stage 1: Compute ECI via eigenvalue decomposition
│   ├── 02_mca.py                 # Stage 2a: Multiple Correspondence Analysis (8 vars, N=511)
│   ├── 03_cah.py                 # Stage 2b: Ward's CAH on MCA coordinates (k=6)
│   ├── 04_regressions_by_type.py # Stage 3: Pooled + type-specific regressions, Chow test
│   ├── 05_regenerate_figures.py  # Generate all 8 figures (6 article + 2 supplementary)
│   ├── 06_cluster_maps.py        # Generate Figure 4 (3x2 small-multiples map)
│   ├── 07_correlation_table.py   # Generate Table S6: predictor correlation matrix
│   ├── 08_compute_eci_bundles.py # Robustness: bundle-based ECI (Juhász et al. 2026)
│   └── 09_eci_temporal_panel.py  # Robustness: longitudinal ECI panel 2015-2025
├── audit/                        # Data quality and geocoding validation
│   ├── audit_01_full_province_department.csv
│   ├── audit_02_discrepancies.csv
│   ├── audit_03_province_summary.csv
│   ├── audit_04_foreign_users.csv         # 76 excluded non-Argentine users
│   ├── audit_05_foreign_repos_by_dept.csv
│   ├── audit_06_ambiguous_users_sample.csv
│   └── audit_07_eci_before_after.csv      # ECI ranking before/after corrections
└── supplementary/                # Supplementary material
    ├── supplementary_tables.md              # Supplementary tables and figures (narrative)
    ├── table_S1_eci_full_ranking.csv        # Full ECI ranking (224 departments)
    ├── table_S2_cluster_region_crosstab.csv # Cluster × region cross-tabulation
    ├── table_S3_small_types_data.csv        # Data for small-N types (Peripheral, Semi-Rural)
    ├── table_S4_within_type_correlations.csv # Within-type correlations with ECI
    ├── table_s_bundle_robustness.csv        # Table S5: bundle robustness check
    └── table_S5_correlation_matrix.csv      # Table S6: predictor correlation matrix

Data description

Core datasets

File	Rows	Columns	Description
departments_full.csv	511	28	All Argentine departments with census (2010), MCA coordinates (5 dims), cluster assignment, ECI, GitHub metrics
bipartite_matrix.csv	224	88	Repository counts by department and programming language (dpto5 + 87 languages)
rca_binary_matrix.csv	224	88	Binarised Revealed Comparative Advantage (RCA >= 1)
eci_panel_2015_2025.csv	1,504	7	Longitudinal ECI panel: dept × year (2015–2025), cumulative snapshots
table_s_bundle_robustness.csv	224	8	Individual vs bundle ECI comparison

Key variables in departments_full.csv

Variable	Source	Description
dpto5	INDEC	Five-digit department code
region	Derived	Six regions: CABA, Pampeana, NOA, NEA, Cuyo, Patagonia
pob_2010, pob_2022	Census	Population
pct_jefe_sec_2010	Census 2010	% household heads with secondary education
pct_jefe_uni_2010	Census 2010	% household heads with university education
pct_pc_2010	Census 2010	% households with computer
pct_nbi_2010	Census 2010	% with unsatisfied basic needs (poverty)
pct_hacinam_2010	Census 2010	% overcrowding
rad_2014	VIIRS	Mean nighttime radiance (2014)
tasa_empleo_2010	Census 2010	Employment rate
mca_dim1...mca_dim5	MCA	Factorial coordinates (5 retained axes)
mca_cluster	CAH	Cluster number (1-6)
mca_cluster_label	CAH	Cluster label
eci_software	ECI	Economic Complexity Index (standardised)
eci_diversity	ECI	Number of languages with RCA >= 1
eci_avg_ubiquity	ECI	Mean ubiquity of RCA languages
gh_total_developers	GitHub	Total geocoded developers
gh_total_repos	GitHub	Total repositories
gh_devs_per_10k	Derived	Developers per 10,000 inhabitants

Analytical pipeline

The scripts are numbered in execution order and depend on a PostgreSQL database (posadas) with the source data.

1. 00_build_schema.py — Integrates data sources (Census 2010, VIIRS nighttime lights, GitHub) into a single analysis-ready table (art1.departamentos, 511 departments).

2. 01_compute_eci.py — Constructs the bipartite network (departments × languages), computes RCA, and extracts ECI and PCI via eigenvalue decomposition. Applies geocoding corrections (Córdoba shift, CABA aggregation, foreign user exclusion).

3. 02_mca.py — Multiple Correspondence Analysis on 8 pre-treatment variables discretised into terciles (24 modalities, N=511). Retains 5 axes via Benzécri correction. Projects ECI and developer metrics as supplementary variables.

4. 03_cah.py — Ward's hierarchical clustering on 5 MCA coordinates. Selects k=6 (silhouette=0.330, Caliński-Harabasz=224.5). Profiles clusters with ANOVA and chi-squared tests.

5. 04_regressions_by_type.py — Pooled and type-specific OLS regressions of ECI on pre-treatment predictors. Chow test for structural heterogeneity. Forest plot of standardised coefficients.

6. 05_regenerate_figures.py — Generates all 8 figures (6 article + 2 supplementary) at 300 DPI.

7. 06_cluster_maps.py — Generates Figure 4 (3×2 small-multiples map of cluster spatial distribution).

8. 07_correlation_table.py — Generates Table S6: pairwise Pearson correlations among predictor variables.

9. 08_compute_eci_bundles.py — Robustness check: downloads the Juhász et al. (2026) language-to-cluster mapping, maps 87 Argentine languages onto 38 bundle clusters, recomputes ECI, and reports Pearson r and Spearman rho between bundle-based and individual-language ECI. Exports table_s_bundle_robustness.csv.

10. 09_eci_temporal_panel.py — Robustness check: computes ECI for cumulative annual snapshots 2015–2025 using created_at timestamps in the repository database. Reports consecutive-year and multi-year rank stability (Spearman rho), quintile persistence, and exports eci_panel_2015_2025.csv. Generates Figure S3.

Key findings

• ECI_software is distinct from developer counts: r = 0.47 (moderate correlation)
• PCI validates the framework: scientific computing languages (Erlang, Fortran, Julia) rank as most complex; web technologies (JavaScript, HTML, CSS) as least complex
• Six departmental types explain 30.2% of ECI variance (eta-squared = 0.302, ANOVA F = 18.84, p < 0.001)
• Determinants are structurally heterogeneous: education drives complexity in Metropolitan-Core; computer ownership in Metropolitan-Diversified; population alone in Pampeana-Educated; no predictor significant in Intermediate-Urban
• Bundle robustness: ECI recomputed on 38 language bundles (Juhász et al., 2026) correlates at Pearson r = 0.90, Spearman rho = 0.90 with the individual-language measure
• Longitudinal stability: consecutive-year Spearman rank correlations 0.82–0.97 (2015–2025); ECI_2020 vs ECI_2025 rho = 0.781 (p < 0.001, N = 125); 68% of top-quintile departments in 2020 remain in the top quintile in 2025

Data sources

Source	Period	Coverage	Access
GitHub API	Accumulated through early 2026	229,270 repos, 23,619 users	Scraped January–February 2026
Census (INDEC)	2010	511 departments	datos.gob.ar
VIIRS DNB	2014	Department-level radiance	Google Earth Engine
Juhász et al. (2026)	—	142 languages → 38 bundle clusters	github.com/sandorjuhasz/eci_software

Requirements

python >= 3.10
numpy
pandas
scipy
scikit-learn
prince
matplotlib
seaborn
geopandas
sqlalchemy
psycopg2
requests

Citation

If you use these data or methods, please cite:

Gomez, R. E. (2026). Software complexity and regional capabilities: a subnational typology of Argentina. Submitted to Regional Studies.

Zenodo DOI:

Licence

Data and code are provided under the CC BY 4.0 licence.