Michael Leon Chrzan

Data Scientist bridging K12 education practice, policy, and advanced analytics through machine learning, causal inference, and NLP to drive educational equity and evidence-based improvement at scale.

Paper (preprint):: The Language of Closure: Examining Racial Differences in How A Community Discusses School Closure Metrics

School closures in urban districts disproportionately affect marginalized communities, yet community input often goes unanalyzed or is reduced to simple frequency counts. This study applies BERTopic, a neural topic modeling approach, to analyze 4,159 suggestions from 2,006 community members regarding school closure metrics in a large urban district. Through extensive hyperparameter tuning across 62 configurations, we identified 14 coherent topics that capture community priorities. Chi-square analysis revealed substantial variation in topic prioritization by race (χ2 = 152.0825, p < 0.0001, V = 0.1439). Furthermore, an analysis of topic outliers revealed that White respondents were significantly more likely to provide suggestions that fell outside of community-wide themes (z = 2.14). These findings demonstrate that ”neutral” community engagement processes may obscure the specific concerns of marginalized groups, and highlight the utility of computational methods in surfacing rigorous insights from large-scale text data.

citation: Chrzan, Michael L., and Francis A. Pearman. (2026). The Language of Closure: Examining Racial Differences in How A Community Discusses School Closure Metrics. (EdWorkingPaper: 26-1420). Retrieved from Annenberg Institute at Brown University: https://doi.org/10.26300/dfry-5906

Conference Presentation: Association for Education Finance and Policy (AEFP), March 2026 - “A Sandbox for Hard Choices: Using Simulation to Explore School Closure Scenarios and Their Consequences”

Preprint now available! You can view the preprint of the paper on EdWorkingPapers here: Download the paper.

School closures are often justified through seemingly neutral criteria such as enrollment or performance, but these metrics can unintentionally deepen educational disparities. This study uses a large urban district’s administrative data to simulate 5,040 closure scenarios, systematically varying six policy design principles, including enrollment, seat utilization, building quality, academic performance, disproportionality safeguards, and ordering of schools considered. By comparing the equity, fiscal, and operational outcomes of each scenario, we reveal three key findings:

  1. safeguards explicitly designed to prevent disproportionality improve fairness but reduce cost savings and seat reductions;
  2. common criteria like enrollment do little to advance either efficiency or equity; and
  3. the order in which schools are evaluated is a surprisingly powerful policy lever.

This work contributes to the field by showing how simulation can equip district leaders to anticipate the trade-offs embedded in closure decisions, moving policy design from reactive justification to proactive fairness.

Paper (preprint): Deeper Roots Before the Storm: Utilizing Machine Learning to Alert School Districts of Permanent School Closures

Developed and evaluated LSTM, XGBoost, and ensemble models to predict large-scale school closures using 18 years of national U.S. school-level data. This research provides early warning systems for school districts and explores the underlying factors that contribute to permanent school closures. The work merges multiple administrative datasets for predictive modeling and policy simulation, helping districts make proactive rather than reactive decisions.