How Disparate
Estimating wage disparities using foundation models
Keyon Vafa, Susan Athey & David Blei
Proceedings of the National Academy of Sciences, 3 June 2025
Abstract:
The rise of foundation models marks a paradigm shift in machine learning: instead of training specialized models from scratch, foundation models are trained on massive datasets before being adjusted or fine-tuned to make predictions on smaller datasets. Initially developed for text, foundation models have also excelled at making predictions about social science data. However, while many estimation problems in the social sciences use prediction as an intermediate step, they ultimately require different criteria for success. In this paper, we develop methods for fine-tuning foundation models to perform these estimation problems. We first characterize an omitted variable bias that can arise when a foundation model is fine-tuned in the standard way: to minimize predictive error. We then provide a set of conditions for fine-tuning under which estimates derived from a foundation model are n-consistent. Based on this theory, we develop fine-tuning algorithms that empirically mitigate this omitted variable bias. To demonstrate our ideas, we study gender wage gap estimation. Classical methods for estimating the adjusted wage gap employ simple predictive models of wages, which can induce omitted variable bias because they condition on coarse summaries of career history. Instead, we use a custom-built foundation model, capturing a richer representation of career history. Using data from the Panel Study of Income Dynamics, we find that career history explains more of the gender wage gap than standard econometric models can measure, and we identify elements of career history that are omitted by standard models but are important for explaining the gap.
Beliefs about what disadvantaged groups would do with power shape advantaged groups’ (un)willingness to relinquish it
Frank Kachanoff et al.
Journal of Personality and Social Psychology, May 2025, Pages 1103-1141
Abstract:
Dominant groups often resist possible changes to the hierarchical status quo. Might such tendencies be partly rooted in negative -- yet potentially malleable -- meta-beliefs about how disempowered groups would use power if they gained control? We investigate these questions across three studies and eight independent samples (Total N = 7,460 analyzed responses) in the context of Black–White relations in the United States. Specifically, we examine White Americans’ meta-beliefs about whether Black Americans desire power to structure society into a hierarchy in which they are dominant versus to institute equality for all groups (i.e., meta-dominance beliefs). Across six cross-sectional subsamples (Study 1, Samples A–F; N = 3,383), we developed and validated a measure of meta-dominance, and found that White Americans varied substantially in their beliefs about how Black Americans would use power. Critically, Whites’ meta-dominance beliefs were uniquely related to their opposition to policies empowering Black Americans as well as their support for efforts to maintain Whites’ position atop the social hierarchy, even when controlling for a range of relevant constructs. In two preregistered experiments among White Americans (Studies 2 and 3; N = 4,077), one of which was a registered report, we tested two possible causal pathways that might explain this relation: (a) “Meta-Dominance Beliefs → Opposition to Black Empowerment” and (b) “Opposition to Black Empowerment → Meta-Dominance Beliefs.” We found evidence in support of the “Meta-Dominance Beliefs → Opposition to Black Empowerment” pathway, but not for the latter Opposition to Black Empowerment → “Meta-Dominance Beliefs” pathway. We discuss our findings’ implications for theories of hierarchy maintenance.
What Explains Growing Gender and Racial Education Gaps?
Zvi Eckstein, Michael Keane & Osnat Lifshitz
NBER Working Paper, May 2025
Abstract:
In the 1960 cohort, American men and women graduated from college at similar rates, and this was true for Whites, Blacks and Hispanics. But in more recent cohorts, women graduate at much higher rates than men. Gaps between race/ethnic groups have also widened. To understand these patterns, we develop a model of individual and family decision-making where education, labor supply, marriage and fertility are all endogenous. Assuming stable preferences, our model explains changes in education for the ‘60-‘80 cohorts based on three exogenous factors: family background, labor market and marriage market constraints. We find changes in parental background account for 1/4 of the growth in women’s college graduation from the ’60 to ’80 cohort. The marriage market accounts for 1/5 and the labor market explains the rest. Thus, parent education plays an important role in generating social mobility, enabling us to predict future evolution of college graduation rates due to this factor. We predict White women’s graduation rate will plateau, while that of Hispanic and Black women will grow rapidly. But the aggregate graduation rate will grow very slowly due to the increasing Hispanic share of the population.
Rapid emergence of a maths gender gap in first grade
Pauline Martinot et al.
Nature, forthcoming
Abstract:
Preventing gender disparities in mathematics is a worldwide preoccupation. In infancy and early childhood, boys and girls exhibit similar core knowledge of number and space. Gender disparities in maths are, therefore, thought to primarily reflect an internalization of the sociocultural stereotype that ‘girls are bad at maths’. However, where, when and how widely this stereotype becomes entrenched remains uncertain. Here, we report the results of a 4-year longitudinal assessment of language and mathematical performance of all French first and second graders (2,653,082 children). Boys and girls exhibited very similar maths scores upon school entry, but a gender gap in favour of boys became highly significant after 4 months of schooling and reached an effect size of about 0.20 after 1 year. These findings were repeated each year and varied only slightly across family, class or school type and socio-economic level. Although schooling correlated with age, exploiting the near-orthogonal variations indicated that the gender gap increased with schooling rather than with age. These findings point to the first year of school as the time and place where a maths gender gap emerges in favour of boys, thus helping focus the search for solutions and interventions.
Job Amenities and Earnings Inequality
Alex Bell
Georgia State University Working Paper, May 2025
Abstract:
A recent Handbook of Labor Economics chapter concludes that “[o]mitted measures of job and firm productivity are a primary source of bias in compensating differential estimates” (Mas, 2025). This paper shows that such bias can be resolved using a simple amenity pricing framework that leverages an imprecise proxy -- or “anti-instrument” -- for workers’ offer sets. The method identifies amenity prices while holding offer sets fixed and recovers intuitively signed estimates in widely used cross-sectional datasets. The approach relies only on a single vertical index of job quality and a conditional independence assumption standard in prior work; it requires no assumptions about functional forms, distributions, or correlations between ability and preferences. Relative to prior observational methods, the estimates suggest a substantial role for amenity substitution in explaining the gender pay gap -- accounting for roughly two-thirds of the observed gap -- and little role for amenities in explaining inequalities by race or parent background. These findings offer new insight into the structure of labor market inequality and the large scope for equity-focused policy.
Economics coauthorships in the aftermath of MeToo
Noriko Amano-Patiño, Elisa Faraglia & Chryssi Giannitsarou
European Economic Review, July 2025
Abstract:
We study changes in coauthorships in economics, after the MeToo movement, using NBER and CEPR working papers between January 2004 and December 2020. We identify three main shifts in collaboration patterns. First, compared to pre-MeToo levels, collaborations across genders in an author’s seniority group increased: we estimate a 12.3% increase of women coauthors per 100 men-authored papers. Second, coauthorship shares of senior with junior economics declined by 3.0%, indicating a shift towards sorting of collaborations by seniority. Third, shares of new coauthorships declined by 5.4%, driven by drops in senior economists’ shares of new junior and new junior women by 18.4% and 48.0%, respectively. The results are robust to different specifications.
How do Women Respond to Gender Biases in Recognition for Teamwork?
Simon Quach & Zhengyi Yu
University of Southern California Working Paper, April 2025
Abstract:
We examine how female economists' coauthorship patterns changed after a 2015 study found that women receive little credit for couathorships with men. Compared to other social scientists, junior female economists reduced the number of working papers with exclusively male coauthors. This decrease in male coauthorships was not offset by other forms of collaborations, leading to a net decrease in publications. While male coauthorships recover later in women's careers, new cohorts of junior female economists continue to avoid them. We find no comparable effects among tenured female or junior male economists.
Pushing the Envelope: The Effects of Salary Negotiations
Zoë Cullen, Bobak Pakzad-Hurson & Ricardo Perez-Truglia
NBER Working Paper, June 2025
Abstract:
Salary negotiations are a widespread phenomenon that can shape key labor market outcomes, such as welfare and inequality. We provide novel empirical and theoretical insights into the causes and consequences of salary negotiations. We conducted two field experiments involving over 3,100 job seekers in the U.S. tech sector, designed to examine two types of information frictions. We find that a light-touch encouragement intervention significantly increased both negotiation attempts and compensation gains. In contrast, providing a substantial discount on negotiation coaching did not significantly affect negotiation attempts. Women responded more strongly to both interventions, helping to narrow gender gaps. We develop a new model of salary negotiations, incorporating risk and information frictions, that can better explain our experimental and non-experimental findings. The model's equilibrium analysis indicates that policies encouraging negotiation can enhance both welfare and equity.