A predictive and descriptive analysis of automation risk across six developed economies, built from raw OECD microdata rather than cited from published reports.
Which economic sectors and worker profiles are most exposed to automation risk, and can this risk be predicted from observable characteristics of a job?
Most analyses of automation risk cite already-published figures from Frey & Osborne (2013) or the OECD (Nedelkoska & Quintini, 2018). This project instead recomputes the risk measure from scratch, using the exact published logistic regression formula applied to raw PIAAC (Survey of Adult Skills) microdata downloaded directly from the OECD. The resulting national figure for the US (10.4%) closely matches the officially published estimate (~10%), validating the entire pipeline.
- Automation risk is strongly structured by occupation and digital skill use at work (together explaining over 60% of predictive power in a Random Forest model), far more than by country (~5%).
- Highest-risk sectors: agriculture, accommodation/food service, transportation, manufacturing.
- Lowest-risk sectors: education, professional/scientific activities, human health.
- A three-model comparison (logistic regression, decision tree, random forest) shows consistent AUC (~0.76-0.78), suggesting the relationship is largely captured by simple, monotonic effects rather than complex interactions.
- Download PIAAC Cycle 1 Public Use Files (SPSS format) for 6 countries: US, Canada, Germany, France, Italy, Sweden
- Recompute the automation risk score (
pr_fo) from individual-level skill-use variables using the Nedelkoska & Quintini (2018) formula - Validate against the officially published OECD figure
- Exploratory analysis by country, sector, occupation, and education level
- Predictive modeling (logistic regression, decision tree, random forest) using structurally independent predictors (no data leakage from the formula itself)
- Feature importance analysis
See SOURCES.md for full details and citations.
- OECD PIAAC Survey of Adult Skills, Cycle 1 Public Use Files
- Methodology: Nedelkoska, L. and Quintini, G. (2018), OECD Social, Employment and Migration Working Papers No. 202
- Original formula replication code: github.com/LjubicaN/Risk-of-automation
automation_risk_analysis.ipynb— full analysis notebook (English)risque_automatisation.ipynb— full analysis notebook (French)individual_dataset_final_en.csv/individual_dataset_final.csv— final cleaned individual-level dataset (~52,000 respondents, 6 countries), English and French column namesSOURCES.md— detailed data sources and methodological limitationsscripts/— reusable Python code for computing the risk score and building the dataset (English and French versions)
Python, pandas, pyreadstat, scikit-learn, seaborn, statsmodels
This measures technical susceptibility to automation, not actual job loss. Data predates the generative AI wave (collected 2011-2017). See the notebook's conclusion for a full discussion of limitations and what these results do not show.