Dissertation Defense: “Essays on Applied Econometrics”, Spencer Sween, University of California, Santa Barbara
Speaker
Spencer Sween, University of California, Santa Barbara
Biography
Spencer Sween is a Ph.D. candidate in Economics at the University of California, Santa Barbara, where he also received his Bachelor of Arts and Master’s degrees in Economics. His research interests lie in applied econometrics, causal machine learning, and the economic effects of financial access. His current work focuses on integrating machine learning methods with panel-data techniques, such as difference-in-differences, to study non-standard treatment regimes. On the applied side, he examines how federal credit access policies affect local labor markets and entrepreneurship rates in underserved communities. Outside of research, he serves as the Head Teaching Assistant for Introduction to Econometrics at UCSB. When not in the office, he can often be found on a tennis or pickleball court.
Title
“Essays on Applied Econometrics"
Abstract
Identification and estimation of policy effects without randomized assignment typically require conditioning on observable, often high-dimensional, covariates that may both confound the treatment of interest and shape the heterogeneity of responses to it. The growing availability of administrative and behavioral data, together with the maturation of causal machine learning in econometrics, has expanded the set of empirical environments in which this conditioning is feasible. Yet the integration of these methods into the workhorse strategies of applied work, difference-in-differences designs and structural estimation in particular, remains uneven. In this dissertation I apply and extend recent semiparametric methods to three substantive questions: the effect of targeted credit access on entrepreneurship in underserved communities; prompt-level preference heterogeneity over large language models and its use in cost-aware deployment; and identification of per-dose responses to policy interventions delivered as continuous, time-varying exposure.
Chapter 1, Targeted Credit Access and Entrepreneurship, studies whether expanding the supply of mission-driven, place-based lending raises new business formation in underserved communities. Community Development Financial Institutions (CDFIs) are nonprofit lenders required by federal certification to direct most of their financing to low-income areas and historically underserved populations; they enter local credit markets at staggered times, in places that conventional banks underserve. Using CDFI entry across approximately 31,000 ZIP codes from 1988 to 2016, I implement a doubly robust difference-in-differences estimator with not-yet-treated comparison units and machine-learning nuisance functions, conditioning on a high-dimensional set of local economic and demographic baseline characteristics. Under conditional parallel trends, CDFI entry raises new startup formation by 0.11 firms per 1,000 residents, about 3 percent of the pre-treatment mean, and the employment-to-population ratio by 0.47 percentage points a decade after entry, with no detectable change in predicted startup quality. Effect magnitudes are roughly twice as large in majority-minority ZIP codes, consistent with a binding credit constraint among historically underserved borrowers.
Chapter 2, Ce modèle vous plaît? Semi-parametric Preference Estimation and Socially Optimal Routing Policies for Large Language Models, asks whether aggregate human preference rankings over LLMs mask prompt-level heterogeneity that a planner can use to reduce the energy cost of deployment. I extend the Bradley-Terry model to allow latent quality parameters to vary flexibly with prompt embeddings and obtain valid inference on the resulting functionals via Neyman-orthogonal scores in the Farrell-Liang-Misra (2021) structural deep-learning framework. Using 48,760 pairwise human evaluations across 50 large language models from the Compar:IA platform, the enriched estimates reveal within-model variation in dominance probabilities across prompt contexts that aggregate rankings do not capture. A cost-aware routing policy that uses this heterogeneity reduces expected inference energy use by roughly 80 percent relative to a quality-first benchmark, with an estimated quality change that is statistically indistinguishable from zero.
Chapter 3, Single-Shot Effects for Difference-in-Differences with Time-Varying, Continuous Treatments, addresses a measurement problem in continuous-treatment settings: when units receive multiple subsidized doses over time, conventional intensity-weighted estimands aggregate across heterogeneous dose-response paths in ways that can differ from the per-dose response in both magnitude and sign. The single-shot effect, the causal response to the first dose at a given exposure level, is identified from quasi-stayers under a strong conditional parallel trends assumption and estimated via Automatic Debiased Machine Learning embedded in a Callaway-Sant’Anna (2021) cohort-time structure. Applied to the SBA 504 loan guarantee program across 17,251 ZIP Code Tabulation Areas from 2010 to 2019, a single subsidized capital project raises local employment by 1.1 percent over five years, with no significant wage response.
JEL Codes: C14, C23, G21, L86, R11, R58
Event Details
Join us for Spencer’s dissertation defense, where he will present his research titled “Essays on Applied Econometrics” We invite you to attend this important academic milestone and learn more about his work in the field. To access a copy of the dissertation here, you must have an active UCSB NetID and password.
