2026 Theses Doctoral

Essays on the Development and Measurement of Human Capital

Kundu, Tushar

This dissertation applies economic models of decision-making towards understanding individual choices in education, well-being measurement, and labor-market screening. I show that better measurement of preferences and beliefs can improve targeting, clarify mechanisms, and sharpen policy design in settings where welfare depends on how decisions are made and whose values they reflect.

The first chapter focuses on the role of heterogeneity in parental preferences for skill development. Parents have insight into which skills hold value for their child, but large class sizes and informal communication limit teachers' access to this information. I provide teachers with structured parent information through a field experiment. I survey 3,404 parents across five private schools in India to measure parents' perceptions of their children's skill levels and preferences for improvement across academic and socioemotional domains. Parents vary in their preferences over which skills to improve, but on average prefer improving their children's weaker skills, a pattern that also appears in students' own perceptions and preferences in a follow-up survey. I develop a structural model of skill development under which this pattern is more consistent with learning costs than family values primarily driving observed specialization. I elicit teachers' beliefs about parent preferences and find little alignment with actual parent views, even at the classroom level. I randomize teacher access to parent survey data via a web portal and find treatment shifts student specialization toward parent-prioritized skills, with effects concentrated in classrooms where baseline teacher beliefs were most inaccurate. Structural estimation corroborates these patterns and enables policy counterfactuals quantifying welfare gains from better cost-benefit alignment. The results demonstrate that structured parent feedback enables teachers to target instruction toward what families value most.

The second chapter, based on joint work, develops an economics-based approach to measuring what people want across a broad set of non-market and market aspects of life. Philosophical perspectives on human desires and values vary; economic theory-driven measurement techniques can provide relevant empirical evidence. We elicited over half a million stated preference choices over 126 dimensions or "aspects" of well-being from a sample of 896 online respondents. We also elicited, via self-reported well-being (SWB) questions, respondents' current levels of the aspects. From the stated preference data, we estimate for each aspect its relative marginal utility per point on our 0-100 response scale. We validate these estimates by comparing them to alternative methods for estimating preferences, and we offer a range of estimates between those that take self-reports at face value and those that (over-)correct for potential social-desirability reporting bias.

Our findings suggest that our respondents value, first and foremost, three basic things: family, money, and health (not necessarily in this order). While commonly studied concepts—such as happiness, life satisfaction, life's ranking on a ladder, and meaning—are all important, respondents place the highest marginal utilities on aspects related to family well-being and health, and financial freedom and security. We document substantial heterogeneity in preferences across respondents within—but not between—demographic groups, with aspects' current levels predicting preferences for the aspects.

The third chapter, also based on joint work, compares the decision-making of AI and human agents in labor-market screening. Artificial intelligence systems increasingly make consequential economic decisions. Generative AI has accelerated this trend: its pre-trained models can be quickly deployed in a variety of contexts. However, complex, proprietary models make it difficult to interpret large language model (LLM) decision making and to assess its biases relative to human decision makers. We collect data on human and LLM hiring decisions for thousands of software-engineering candidates in a way that allows us to overcome the selection issue of observing hiring decisions only for interviewed applicants. The data allow us to readily assess accuracy and bias in evaluations.

We find that humans and the LLM are similarly accurate in their resume screening decisions, on average, but the LLM exhibits smaller racial biases than human recruiters. In terms of disparate impact, humans tend to recommend Black and Hispanic applicants at higher rates than white applicants after controlling for their interviewer's hiring decision. In contrast, the LLM favors South Asian applicants relative to white applicants. We show that a model of applicant selection approximates both human and LLM decision making, allowing us to decompose bias into different forms of discrimination: taste-based discrimination, accurate statistical discrimination, and biased beliefs. We find that discrimination by human recruiters and the LLM reflects a combination of biased beliefs and taste-based discrimination, with taste-based discrimination playing a larger role for human recruiters.