TL;DR
This paper introduces a flexible, semi-nonparametric framework for multidimensional matching models, moving beyond restrictive "normality" assumptions. By applying optimal transport theory and sieve estimation, the authors demonstrate that technological shifts favoring cognitive skills in the U.S. economy were significantly more dramatic between 1990 and 2010 than previously estimated.
What Problem Does This Paper Address?
In labor economics, matching models help us understand how workers with specific skills are paired with jobs that require them. Traditionally, these models simplified agents into a "single index" (e.g., general ability). However, the real world is multidimensional: a mathematician has high cognitive skills but perhaps low manual skills, while a gymnast has the opposite.
Existing multidimensional models, most notably Lindenlaub (2017), rely on the assumption that these traits follow a joint normal (Gaussian) distribution. The problem? Real-world data is rarely normal; it’s often skewed, multi-modal, or has "fat tails." Forcing non-normal data into a Gaussian box leads to biased results and fails to explain complex phenomena like wage polarization—the trend where wage growth is strongest at the very top and bottom of the distribution, leaving the middle behind.
Methods
The authors generalize the matching framework by stripping away the normality requirement. Their approach rests on two pillars:
- Optimal Transport (OT) Theory: The authors frame the "social planner’s problem"—assigning workers to jobs to maximize total economic output—as an optimal transport problem. This allows them to derive unique equilibrium wage and assignment functions for any arbitrary distribution of characteristics.
- Sieve-Based Semiparametric Estimation: Instead of assuming a fixed mathematical shape for the wage function (like a simple quadratic), they use "sieves" (specifically Bernstein polynomials). This allows the model to flexibly "learn" the shape of the wage-skill relationship from the data.
They propose several estimators, including Sieve Maximum Likelihood (SML) and Sieve Generalized Least Squares (SGLS). The SGLS estimator is particularly robust, as it accounts for correlated measurement errors without needing to assume they are normally distributed.
Datasets
The paper evaluates its model using both simulated and real-world data:
- Monte Carlo Simulations: The authors test their estimators against Gaussian, Gumbel, and bi-modal Gaussian Mixture distributions to prove that their method remains accurate even when traditional models fail.
- Empirical Data: They re-analyze the dataset from Lindenlaub (2017), which combines:
- NLSY79 & NLSY97: National Longitudinal Survey of Youth cohorts providing worker skill data for 1990-1991 and 2009-2010.
- O*NET: The Occupational Information Network database, used to determine the cognitive and manual skill requirements of various jobs.
Key Insights and Ideas
- The "Normality Trap": The paper shows that even if you transform data to look normal (using a Gaussian copula), you can still distort the underlying economic relationships. Their model works directly on the original, "messy" data.
- Curvature Matters: The Gaussian model restricts the wage function to a quadratic form. The authors find that the real wage function has a much steeper curvature at high cognitive skill levels. This flexibility is what allows their model to capture wage polarization.
- Underestimated Tech Progress: By using a more flexible model, the authors discovered that "task-biased technological change"—the shift toward rewarding cognitive skills—was much larger than previously thought. The "complementarity" of cognitive skills (how much they boost productivity when matched with the right job) increased far more than earlier parametric models suggested.
Key Results
- Better Fit for Inequality: The semi-nonparametric model provides a significantly better fit for U.S. wage patterns. It successfully predicts the "U-shaped" wage growth (polarization) seen between 1990 and 2010, whereas the Gaussian model largely fails to account for it in both tails.
- Dramatic Skill Shifts: In the original (untransformed) data, the cognitive complementarity parameter ($\alpha_{CC}$) jumped from near 0 in 1990 to over 6.0 in 2010. Meanwhile, manual skill complementarity ($\alpha_{MM}$) effectively plummeted toward zero.
- Robustness: Simulations showed that while the traditional Maximum Likelihood estimator becomes highly inconsistent when the data isn't normal, the SGLS estimator remains precise and efficient across all tested distributions.
Limitations and Future Work
- Frictionless Markets: The model assumes a "frictionless" equilibrium, meaning it doesn't account for search costs, unemployment, or "mismatches" where a worker can't find the right job. Integrating search frictions into this multidimensional OT framework is a suggested next step.
- The Curse of Dimensionality: While the model handles two dimensions (cognitive and manual) well, the authors note that adding more dimensions (like interpersonal or digital skills) increases computational complexity. They suggest that Artificial Neural Networks might be a promising way to approximate these functions in higher-dimensional spaces.
- Measurement Error: While the SGLS approach helps, the authors acknowledge that addressing multidimensional measurement error in worker skills remains a difficult econometric challenge.
Summary updated