“Seeing Beyond the Trees: Using Machine Learning to Estimate the Impact of Minimum Wages on Labor Market Outcomes”
Authors: Doruk Cengiz, Arindrajit Dube, Attila Lindner, David Zentler-Munro
Research Question:
How do minimum wage increases affect various labor market outcomes (wages, employment, unemployment, participation)?
The key innovation is using machine learning (ML) to predict which individuals are most likely affected by minimum wage hikes—
moving beyond simple demographic proxies like teenagers.
Methodology:
1. Step 1: ML-Based Prediction of Exposure
o Use demographic data to estimate the probability that an individual earns near the minimum wage (defined as
<125% of minimum wage).
o ML models used:
Gradient Boosted Trees (best performing)
Random Forests
Elastic Net Logistic Regression
Card & Krueger Linear Model (for comparison)
o They build “high-probability” and “high-recall” groups based on predicted exposure.
Data:
Current Population Survey (CPS) (1979–2019)
Minimum wage data from Vaghul & Zipperer
Over 6 million individual-level observations
Focused on states with "prominent" minimum wage hikes
Key Results:
Wages: Significant increase for predicted minimum wage workers:
o High-probability group: +2.3%
o High-recall group: +1.6%
Employment: No significant job losses. Small, statistically insignificant increases.
Unemployment & Participation: No significant changes. No increase in job search effort.
Robustness: No wage effects for low-probability group; strong falsification test support.
Conclusion:
This paper demonstrates the power of machine learning in identifying policy-relevant treatment groups more accurately than
traditional demographic proxies. The findings support that minimum wage increases improve wages without negative effects on
employment, unemployment, or participation, especially for non-teen low-wage workers, who are often underrepresented in
previous research.
The approach improves both the precision and interpretability of empirical analysis in labor economics, showing how ML can
enhance policy evaluation.
“Monopsony in Online Labor Markets”
Authors: Arindrajit Dube, Jeff Jacobs, Suresh Naidu, Siddharth Suri
Published in: AER: Insights (2020), Vol. 2(1), pp. 33–46
Research Questions
1. Do employers on online platforms like Amazon Mechanical Turk (MTurk) exercise monopsony power, even in large, “thick”
markets with seemingly low frictions?
2. How elastic is the labor supply faced by individual requesters (employers) on MTurk?
o Can we estimate this using both observational data (task durations and posted wages) and experimental data
(randomized wage offers)?
Data Used
Main dataset: Over 700,000 task postings (HIT batches) scraped from MTurk (2014–2017).
Supplementary data: 5 randomized experiments from past studies (e.g., Horton et al. 2011; Dube et al. 2018).
Observations include: task duration, posted wage (reward), requester ID, task description, time allotted, and batch size.
Textual features (e.g., n-grams, Doc2Vec) and non-textual features (e.g., HIT size, qualifications required) were constructed.
Methodology
Results
1. Observational Estimation Results
Double ML estimates show low recruitment elasticity of about 0.096.
Consistent across subsamples and time periods.
FE regressions without machine learning produced noisy or insignificant results.
The small elasticity implies significant wage-setting power for employers.
2. Experimental Results
Retention Elasticities:
o Ranged from 0.05 to 0.25
o E.g., Dube et al. (2018) found elasticities around 0.05–0.12
Recruitment Elasticities:
o Also very low: pooled estimate ≈ 0.06
o Boosted confidence in the validity of DML estimates.
3. Comparison
Experimental and observational estimates were very close.
Suggests machine learning methods (DML) are robust.
Confirms presence of monopsony power even in online markets with many employers.
Conclusion and Interpretation
Despite being a "frictionless" labor market, MTurk exhibits strong monopsony power.
Labor supply elasticities facing individual employers are consistently low, contradicting the competitive labor market model.
Suggests broader implications: monopsony is not limited to small towns or niche labor markets, but even exists in online
gig work.
Employers appear to optimize wages based on elasticities, indicating active use of market power.
Key Contributions
First rigorous quantification of monopsony in a major online labor market.
Demonstrates that modern machine learning tools (e.g., Random Forests with DML) can accurately identify economic
effects.
Encourages further data-driven regulation and monitoring of digital labor platforms.