Research

A number of U.S. states have enacted bans on tobacco use by students, staff, and visitors anywhere on the grounds of public elementary and secondary schools statewide. These laws are intended to reduce tobacco use, reduce exposure to secondhand smoke, reinforce anti-tobacco curricula taught in schools, and prevent children from viewing their teachers and fellow students using tobacco products. We examine the impact that the laws have on the smoking behavior of students, teachers, and other school staff by using stacked difference-in-differences estimators that exploit the time variation in adoption of the laws across states. The results presented in this paper suggest that statewide school smoking bans have little impact on smoking behavior.

Several training programs worldwide are residential, bringing enrollees to live in facilities and thus increasing program costs. In the U.S., the Job Corps (JC) program for disadvantaged youth includes a residential component. Despite extensive literature evaluating JC, the impact of the residential component has received limited attention. We develop and employ a semi-parametric model to estimate the average causal effects of this residential component, accounting for selection into both the program and the residential feature. Our findings indicate that, for JC enrollees, the residential component boosts earnings, employment, and educational degree attainment among adolescents. For other subgroups, the residential component has minimal impact on labor market outcomes but increases detrimental behaviors such as drug use, drinking, and smoking. For example, the average treatment effect of the residential component raises hard drug use by 150% among young adults and by 92% among females without children; it also increases smoking among females without children by 47%. Negative peer effects may explain the increase in detrimental health behaviors.

Employers commonly use open-ended questions in hiring, but usually responses aren’t recorded, and knowledge about the usefulness of such responses is limited. This study investigates properties of written responses to open-ended questions collected over many years via an application tracking platform that school districts use to hire teachers. We develop a scalable process to discover topics within the corpus of responses to each question and then to classify each response with respect to the topics while preserving interpretability and transparency. It adapts a traditional qualitative topic modeling approach, Inductive Thematic Analysis, to best combine the strengths of researcher judgment and large language models (LLMs), introducing Hybrid LLM–Topic Model Inductive Thematic Analysis (HLTM). Automated scoring achieves human-level reliability in document classification. Questions elicit distinct, nonredundant information regarding a candidate’s pedagogical skills, classroom management, professional philosophy and growth mindset, relational and cultural competence as well as other professional attributes and institutional fit, but questions vary along important dimensions. Questions regarding classroom activities and student progress generate more dispersion in responses across applicants and fewer extreme demographic selection ratios. In contrast, philosophy-based questions produce more uniform, potentially scripted responses. This scalable framework empowers researchers and practitioners to assess language based hiring instruments transparently and to monitor potential demographic disparities.

Timely high school completion is a critical step in successful transitions to post-secondary schooling and the labor market. This randomized controlled trial, deployed in two urban, disadvantaged school districts, tests the effectiveness of a mature and comprehensive student-support program on an unusually large scale (n=2,528). While overall effects on graduation rates are small and statistically insignificant, exploratory analyses reveal notable variation by site and over time, mirroring heterogeneity in treatment timing and program implementation. These findings highlight the challenges of scaling interventions and raise concerns about generalizing results from small-scale, single-site evaluations.