Predicting Instructional Quality in Elementary Math Classrooms Using Machine Learning and Observation Data (Stage 1, 8:44 AM)

• In Event: e-Lightning Ed-Talk Session 6 - Stage 1

Abstract

Persistent challenges in teacher recruitment and retention demand better tools for identifying and supporting effective instruction. Using data from the National Center for Teacher Effectiveness, we apply a two-stage machine learning approach (LASSO + Random Forest) to predict Math Quality of Instruction (MQI) scores for elementary teachers. We compare four models simulating recruitment and professional development contexts, with and without access to observation or student data. Results show that pre-hire background characteristics and test scores have limited predictive value, while specific observation-based practices, like linking mathematical ideas and explaining reasoning, more strongly predict instructional quality. We discuss implications for data actionability during hiring and in-service development, highlighting how predictive modeling can inform more equitable and practice-relevant teacher selection and support.

Authors

• Jerry Nelluvelil, Harvard University

• Yixuan Cui, Harvard University