Daniel Sabey

Daniel Sabey

PhD Student
Cornell University

Home

Research

CV

Contact:
Department of Economics
Cornell University
Ithaca, NY

Published Papers

Decreasing time to baccalaureate degree in the United States

with Jeffrey T. Denning, Eric R. Eide, and Kevin J. Mumford

Economics of Education Review, 2022, 90: 102287

After increasing in the 1970s and 1980s, time to bachelor's degree has declined since the 1990s. We document this fact using data from three nationally representative surveys. We show that this pattern is occurring across school types and for all student types. Using administrative student records from 10 large universities, we confirm the finding and show that it is robust to alternative sample definitions. We discuss what might explain the decline in time to bachelor's degree by considering trends in student preparation, state funding, student enrollment, study time, and student employment during college.
@article{denning2022decline, title={Decreasing time to baccalaureate degree in the United States}, author={Denning, Jeffrey T and Eide, Eric R and Mumford, Kevin J and Sabey, Daniel J}, journal={Economics of Education Review}, volume={90}, pages={102287}, year={2022}, publisher={Elsevier}, doi={10.1016/j.econedurev.2022.102287} }

Working Papers

Can Machine Learning Help Teachers Predict and Improve Student Outcomes? Evidence from a Randomized Controlled Trial [Job Market Paper]

with Jake Meyer  ·  [Draft PDF]

Can more precise information about students improve teaching and achievement? We randomly provide elementary-school teachers with machine-learning forecasts of students' end-of-year test scores across six charter-school campuses serving 3,068 students in grades 1–6. Linking repeated teacher surveys to administrative monitoring records and test scores, we measure responses in beliefs, student priorities, and instructional practices. Access to forecasts reduces teachers' measured prediction error and increases their reported focus on students predicted to struggle. However, teachers report less independent information gathering and intervention planning, and shift instruction from review toward new material. We estimate an average achievement decline of 0.099 standard deviations, with larger losses among students predicted to score below proficiency. These findings distinguish identifying students likely to struggle from providing effective support: predictive tools should be paired with effective interventions and not evaluated on accuracy alone.

Association Between Recent Medicaid Expansions and Child Maltreatment Reporting and Outcomes

with Alex Hoagland  ·  [Draft PDF]

We investigate the association of recent Medicaid expansion in Maine on the volume and quality of child maltreatment reporting. We find that Medicaid expansion led to increased overall reporting, particularly from medical professionals, resulting in increased separations at lower substantiation rates. Associations were stronger among Black and female children. Our results suggest recent Medicaid expansions are associated with increased maltreatment reporting and intervention at lower specificity, resulting in a greater rate of unsubstantiated separations of children from their caregivers.

Locked Down and Lighting Up: Using COVID-19 Mobility Restrictions as an Instrument for Cigarette Stock and Smoking Intensity

with Don Kenkel and Alan Mathios  ·  [Draft PDF]

This paper investigates the "temptation effect" – whether increased readily available cigarette stock causally increases smoking intensity. We leverage the unprecedented natural experiment of the COVID-19 pandemic using Nielsen household scanner data, Advan Patterns cell phone data, and USAfacts.org COVID-19 data. An Interrupted Time Series (ITS) model reveals COVID having a sharp and persistent increases in both cigarette stockpiles and smoking rates. Exploration of linked survey data highlights heterogeneous responses related to stress and household characteristics. To isolate the causal effect of stock, we use county-level changes in mobility as an instrumental variable (IV) for changes in cigarette stock. The IV estimates suggest that a one-pack increase in instrumented stock causally increases daily smoking by approximately one cigarette.

Works in Progress

The Impact of Prisoner Education on Recidivism, Labor Market Outcomes, and Public Assistance Utilization

with Valentin Bolotnyy, Mike Lovenheim, and Natalie Millar  ·  [slides]

This project evaluates the impact of Career and Technical Education (CTE) programs in Tennessee state prisons on post-release outcomes. We employ an instrumental variables approach leveraging differential exposure to CTE programs based on sentence length and variation in the availability of classes across prisons to instrument for the effect of CTE enrollment.

Student Group Specific Teacher Learning

with Jason Cook

This study investigates whether demographic-specific teaching skills are learnable by leveraging a natural experiment in the Columbus City Schools district. In 2003, the district shifted from race-specific to race-blind magnet school lotteries, causing significant changes to the racial composition of its schools. This paper examines how this policy change affected teacher effectiveness for different student populations and across various outcomes. Using a two-way fixed-effect model, the research analyzes how the lottery-induced changes in classroom demographics impacted teacher value-added, calculated separately before and after the policy change. By comparing these effects across different types of student outcomes (e.g., academic, behavioral), the study aims to reveal whether teachers can adapt to demographic shifts and which aspects of teaching effectiveness are most malleable mid-career.

Improving Teacher Student Matching: Comparative Advantage or Machine Learning

[Executive Summary of Results (PDF)]

This paper asks whether achievement gains from student–teacher allocation primarily reflect comparative advantage—assigning students to teachers who are relatively more effective for their particular characteristics—or absolute advantage—exposing more students to teachers with higher average value added. Using North Carolina administrative records linking students, teachers, and math and reading achievement in grades 3–8 from 2006 through 2024, I evaluate both the predictive content of teacher heterogeneity and its implications for allocation.

I estimate overall teacher value added and type-specific value added by economic disadvantage, race, sex, and prior achievement. I evaluate whether comparative-advantage forecasts improve prediction beyond overall value added using chronological holdouts and a teacher-turnover design. I also use matrix completion—a low-rank model that uses historical student–teacher outcomes to estimate latent match quality—to test whether individual matches contain predictable information beyond observed student characteristics and average teacher effects.

Key findings:

  • There is evidence of comparative advantage, but it does not add predictive value: Corrected correlations between teacher effects across student types are high but below one, consistent with some small but persistent heterogeneity. However, adding estimated comparative advantage does not improve out-of-sample prediction beyond overall teacher value added, and estimates are consistently not fully forecast unbiased.
  • No predictable latent student–teacher match quality using matrix completion: Such matches could reflect teaching style or other interactions not captured by observed student characteristics. The model fits historical outcomes well but fails to predict future outcomes. The estimated matches cannot reliably guide prospective assignment.
  • Apparent gains depend on the optimized dimension: With teacher capacities fixed, sorting on prior-achievement comparative advantage produces model-implied gains of ~0.0021 standard deviations in math and 0.0022 in reading. Using an equal-weight ensemble of all four type-specific predictions reduces these gains to ~0.0011 in each subject. Evaluating assignment policies through a single selected dimension ignores other dimensions that also have match quality and often move in opposite directions, mitigating predicted benefits.
  • Workload reallocation yields larger gains: Moving a moderate number of students from lower- to higher-value-added teachers within schools generates larger model-implied gains in both subjects (requiring shifting roughly 4–5 percent of enrolled seats).

Taken together, the results favor expanding access to higher-value-added teachers over fine-tuning assignments using estimated comparative advantage. Comparative advantage adds little out-of-sample predictive information, and modest workload reallocation generates gains larger than student sorting.

This paper explores methods to optimize teacher-student pairings by moving beyond single value-added metrics. It investigates which student characteristics (race, sex, prior performance, etc.) offer the greatest potential for gains when matching teachers based on their comparative advantages. The research introduces a novel method for classifying student "types" based on their past performance with different teachers and compares the effectiveness of traditional value-added models against higher-dimensional machine learning predictions for creating optimal classroom assignments. The paper also considers the welfare implications of reassigning students to teachers, rather than teachers to classes, and proposes an RCT to test these matching strategies in a real-world setting.