AI Jobs Weighting Analysis

🔗 View the Project Website

Overview

When studying AI job adoption across U.S. counties, should we weight by population or log-population? This analysis uses county-level data from Lightcast job postings, U.S. Census demographics, and USPTO patent records spanning 2014-2023, originally analyzed by Andreadis et al. (2025). I re-estimated their regression models comparing standard population weights against log-population weights to assess whether large metropolitan areas drive the key relationships. I fitted linear regression models predicting AI job share as a function of county demographics, innovation indicators, and industry characteristics, comparing results between population-weighted and log-population-weighted specifications. Labor market tightness emerges as the most robust predictor of AI adoption across all county sizes, while education effects appear concentrated in large metropolitan areas.

Key Findings

• Labor market tightness is the most robust predictor of AI job adoption across all weighting schemes
• Bachelor's degree effects are substantially reduced under log-population weighting, suggesting they may be driven by large metropolitan areas
• STEM education remains consistently important across county sizes
• Manufacturing intensity shows persistent negative relationships with AI adoption
• Innovation capacity (patents per employee) matters regardless of county size

Repository Structure

├── README.md
├── _quarto.yml                 # Quarto website configuration
├── index.qmd                   # Home page with key findings
├── about.qmd                   # About the author and project
├── sources.qmd                 # Data sources and methodology
├── model.qmd                   # Statistical model details
├── data/                       # Data files (if included)
├── scripts/                    # R analysis scripts
│   ├── data_cleaning.R
│   └── weighting_analysis.R
└── plots/                      # Generated visualizations
    ├── ai_share_coefficients.png
    └── ai_change_coefficients.png

Data Sources

• Lightcast: Job posting data from 40,000+ online sources
• U.S. Census ACS: County demographics (2013-2022)
• USPTO PatentsView: Innovation indicators
• BLS: Labor market statistics
• FHFA: Housing price data

Methodology

This project extends the analysis by Andreadis et al. (2025) by comparing two weighting schemes:

1. Population weights: w = Population
2. Log-population weights: w = log(Population)

The comparison reveals how county size affects the interpretation of key predictors of AI job adoption.

Technical Details

• Models: Linear regression with county and year fixed effects
• Time Period: 2014-2023
• Observations: 24,645 county-year observations
• Standard Errors: Clustered at county level
• Software: R with tidymodels framework

Key Visualizations

The project includes coefficient comparison plots showing how estimates change between weighting schemes across five different model specifications, highlighting the sensitivity of findings to methodological choices.

Implications

These findings suggest that:

• Policymakers should focus on creating tight labor market conditions to encourage AI adoption
• Education-focused strategies may be less universally applicable than previously thought
• Different approaches may be needed for large metropolitan areas vs. smaller counties

Author

Jacob Khaykin
📧 jacob.khaykin@example.com
🐙 GitHub

Acknowledgments

This project was created as part of Kane's Data Science Bootcamp. I thank Andreadis et al. (2025) for their transparent methodology that enabled this extension analysis.

License

This project is open source and available under the MIT License.

---

Created: August 2025