Varun Satish

V. Satish, F. Hafner, S. Kapoor, M. Luken, L. Liu, T. Liu, J. Perdomo, B. Stroebl, K. Vafa, M. Verhagen, M. Salganik. Language Models for Life Outcome Prediction

Abstract: Language models have enabled new approaches in many different scientific domains, but such approaches remain largely unexplored in life course research. In this paper, we explored how language models can be used for a specific task that has posed challenges for life course scholars—predicting life outcomes. The conventional approach to predicting life outcomes is to train linear or tree-based models on life course data represented in tabular format. In contrast, our approach is to train a large language model (LLM) on life course data represented in text format. We constructed text summaries of 6 million people's lives using complex, multi-domain data from the Dutch Population Registry. We used these "books of life" to fine-tune an open-weight LLM, named Cruijff, to predict an important life outcome: fertility. Cruijff used books of life to generate predictions more accurate than a simple demographic benchmark. When the books of life were enriched with predictions from a tree-based model trained on tabular data, Cruijff's predictions improved. These findings demonstrate that language models are a viable and flexible approach for predicting life outcomes. Given this starting point, future work should study whether they might generate better predictions or enable different research opportunities.

V. Satish. The Declining Middle: Changes in Subjective Social Class in the United States, 1972-2018

Abstract: The share of Americans who perceive themselves as middle class is declining. However, prior work falls short of explaining which Americans are driving this decline and why their perceptions are changing. To address these gaps, I use machine learning to analyze 356 attitudinal, demographic, and political variables from the General Social Survey (GSS) spanning 1972-2018. Specifically, I use classification trees to identify variables that best differentiate perceptions of class. I then use those variables to define subpopulations and examine how perceptions evolved over time within each subpopulation. I find middle class declines occurred even among Americans not typically considered disadvantaged: those who attended college or earned above-median real income. These declines were not driven by their worsening material circumstances, as measured by levels of real income. Instead, they earned increasingly less real income compared to people with higher levels of education or people sitting higher on the income distribution. I theorize that relative deprivation—the sense of disadvantage from social comparisons—explains these findings. This interpretation is supported by panel data demonstrating within-person variation in class perceptions, suggesting perceptions may shift in response to reference group circumstances. These findings indicate that rising income inequality has reshaped how Americans perceive their social class.

A. Adsera, F. Querin, V. Satish. Can the Gendered Sorting of Occupations Explain Wage Differentials Across Educational Levels in the US?

Abstract: The persistence of gender earnings gaps, despite women having higher levels of education, runs in contrast to predictions from models of human capital accumulation. To address this, we ask whether men and women occupy different types of jobs even with the same education. This may account for gender differentials if women either sort or are discouraged from occupying jobs with characteristics that make them high paying. We are primarily interested in understanding the role of educational attainment in determining gendered occupational sorting. Recent studies highlight the importance of occupation and sector in perpetuating the gender wage gap, but their focus is mostly on the highly educated workers. Using data from the American Community Survey (ACS) along with O*NET occupation information, we focus on five dimensions of occupational characteristics that may reflect sorting by gender and education: contact with others, autonomy, leadership, machine-dependency, and time pressure. Preliminary results confirm gendered prevalence in these occupational characteristics, with women performing jobs with higher contact with others and less machine use. We document educational gradients in the wage returns to these characteristics. Results from Oaxaca-Binder decompositions highlight how our models explain gender differentials more for highly educated workers than for those with less education, underscoring the importance of further research specifically on workers who do not have a college degree.