Legitimation Strategies of Transnational Private Institutions

Replication code for:

Bjorkholt, Solveig. (2026). "Legitimation Strategies of Transnational Private Institutions: Evidence From the International Organization for Standardization." Regulation & Governance. DOI: 10.1111/rego.70123

Replication data available at: Harvard Dataverse

Overview

This project investigates how the International Organization for Standardization (ISO) deploys different legitimation strategies depending on the type of standard being produced. The paper tests whether physical standards (technical specifications, interoperability) are associated with technocratic legitimation, while societal standards (safety, quality, social responsibility) are associated with democratic legitimation. The analysis operates on four dimensions:

• Technocratic legitimation: justifying standards by appealing to expertise, efficiency, scientific knowledge, and innovation.
• Democratic legitimation: justifying standards by appealing to participation, representation, equity, sustainability, and stakeholder engagement.
• Input legitimation: characteristics of who participates in standards development (R&D capacity of member countries; diversity of committee membership).
• Output legitimation: the language used in ISO news articles to promote standards (dictionary-based text analysis and GPT-coded legitimation statements).

Repository Structure

.
├── README.md
├── .gitignore
└── scripts/
    ├── 00_scraping/           # Data collection
    │   ├── News_Selenium_Scrape.R   # Scrapes ISO news articles from iso.org
    │   └── parse_scopus.R           # Parses Scopus API JSON into tidy data
    │
    ├── 01_classification/     # LLM-based coding/classification
    │   ├── 1_1_GPT_coding_liaison.R         # Classifies liaison organizations (GPT-4)
    │   ├── 1_2_GPT_coding_standards.R       # Classifies standards as physical/societal (OpenAI)
    │   ├── 1_2_GPT_coding_standards_new.R   # Updated version using Responses API
    │   └── 1_3_GPT_coding_legitimacy.R      # Codes news sentences as democratic/technocratic (GPT-4-mini)
    │
    ├── 02_data_preparation/   # Data merging and feature engineering
    │   └── 2_Make_data.R              # Merges standards DB with classifications,
    │                                  # computes cumulative physical/societal counts
    │
    ├── 03_models/             # Statistical models (core analysis)
    │   ├── 3_1_Models_input_technocratic.R        # Input legitimation: R&D indicators ~ standard type
    │   ├── 3_2_Models_input_democratic.R          # Input legitimation: diversity indices ~ standard type
    │   ├── 4_0_Models_output_data_generation.R    # Prepares news text + legitimation coding for output models
    │   ├── 4_1_Models_output_technocratic.R       # Output legitimation: technocratic dictionary shares ~ standard type
    │   ├── 4_2_Models_output_democratic.R         # Output legitimation: democratic dictionary shares ~ standard type
    │   └── 4_3_Dictionary_stats.R                 # Summary statistics for text dictionaries
    │
    └── 04_results/            # Tables, figures, and validation
        ├── 5_1_Regression_tables.R            # Main regression tables (LaTeX)
        ├── 5_2_Coef_plot_standards.R          # Main coefficient plot (Figure 2)
        ├── 5_Coef_plot_standards.R            # Simplified coefficient plot (earlier version)
        ├── 5_3_Benjamin_Hochberg_adjustment.R # Benjamini-Hochberg FDR correction
        ├── 6_Descriptives.R                   # Descriptive statistics tables
        ├── 7_GPT_examples.R                   # Example table of GPT-classified standards
        ├── 8_GPT_validation.R                 # Validation: GPT coding vs. SDGs and TC sectors
        └── Figure1.R                          # Figure 1: physical vs. societal standards over time

Analysis Pipeline

The scripts are designed to run sequentially by number prefix. The results scripts (5_*) automatically source() the model scripts they depend on.

00_scraping/         Data collection (ISO news, Scopus papers)
       |
01_classification/   LLM classification of organizations, standards, and legitimation
       |
02_data_preparation/ Merge and compute features
       |
03_models/           Fixed-effects panel regressions (feols)
       |
04_results/          Tables, coefficient plots, descriptive statistics

Models

All models use the fixest package with:

• Fixed effects: Year + Technical Committee
• Clustered standard errors: by Committee and Year
• Coverage: 2004--2022 (input) / 1999--2022 (output)

Model Family	DV	IV	Script
Technocratic Input	R&D/GDP, Researchers, High-tech exports, Uni-industry collab., Articles, Business GERD, Knowledge employment	Physical / Societal standard count	3_1
Democratic Input	Rae Index, Shannon H (regional & sectoral diversity)	Physical / Societal standard count	3_2
Technocratic Output	Share of technocratic words (from research papers, patents dictionaries) + GPT-coded technocratic statements	Physical / Societal standard indicator	4_1
Democratic Output	Share of democratic words (from UNGA debates, CSR dictionaries) + GPT-coded democratic statements	Physical / Societal standard indicator	4_2

Data Requirements

The scripts expect data in a directory outside the repository (referenced via relative paths like ../../../Legitimacy/Data/). You can configure this by modifying the paths at the top of each script. The key data files are:

File	Description
iso_standards.sqlite	ISO standards database (standards metadata, committee membership, liaison organizations)
news.rds	Scraped ISO news articles
legitimation_output6.rds	GPT-coded legitimation labels for news sentences
standards2.rds / standards3.rds	Processed standard type classifications
organizations_tagged.rds	GPT-classified liaison organizations
scopus_results.rds	Scopus research paper abstracts
rndgdp.csv, researchers.csv, WIPO-GII.xlsx	Country-level R&D and innovation indicators
TXT/ directory	UNGA General Debate transcripts (for democratic dictionary)

Replication data is available from the Harvard Dataverse.

Software Requirements

R (>= 4.0) with the following packages:

• Data wrangling: tidyverse, DBI, RSQLite, readxl, readtext
• Text analysis: tidytext, quanteda, SnowballC, lexicon
• Modeling: fixest, broom
• Tables and figures: modelsummary, kableExtra, ggplot2, forcats
• API access: httr, jsonlite
• Web scraping: RSelenium (requires Docker with selenium/standalone-chrome)
• Package management: pacman

Install all at once:

install.packages(c(
  "tidyverse", "DBI", "RSQLite", "readxl", "readtext",
  "tidytext", "quanteda", "SnowballC", "lexicon",
  "fixest", "broom",
  "modelsummary", "kableExtra", "forcats",
  "httr", "jsonlite",
  "RSelenium", "pacman"
))

API Keys

The classification scripts require API keys set as environment variables:

# In your .Renviron file (never committed to git):
OPENAI_API_KEY=sk-...
SCOPUS_API_KEY=...

The classification scripts (01_classification/) are provided for transparency and reproducibility documentation. Their outputs (the classified data) are included in the replication dataset on Harvard Dataverse, so re-running the LLM classification is not required to reproduce the paper's results.

How to Reproduce

1. Download the replication data from Harvard Dataverse.
2. Place the data files so that the relative path ../../../Legitimacy/Data/ points to the data directory (or modify the paths in the scripts).
3. Set your working directory to the repository root.
4. Run the results scripts (e.g., source("scripts/04_results/5_1_Regression_tables.R")) -- these will automatically source upstream model scripts.

Citation

@article{bjorkholt2026legitimation,
  title={Legitimation Strategies of Transnational Private Institutions: Evidence From the International Organization for Standardization},
  author={Bj{\o}rkholt, Solveig},
  journal={Regulation \& Governance},
  year={2026},
  doi={10.1111/rego.70123}
}

Related Work

Bjorkholt, S. (2025). "Presenting the Standat Database on International Standards: Improving Data Accessibility on Marginal Topics." Political Science Research and Methods, 1--19.