Using Machine Learning to Predict Harsh School Discipline In Albuquerque Public Schools

Background. Harsh discipline disproportionately impacts students of color and those experiencing poverty. Few studies examine the most important predictors of school harsh discipline among Native Americans using machine learning algorithms.

Methods. Using data from Albuquerque Public Schools (APS), we conducted logistic regression, pruned decision trees, random forest plots, and support vector machines to determine the most reliable predictors of harsh school discipline outcomes (e.g., criminal legal system referral). Models were evaluated using a stratified 70/30 split within the 2019 – 20 sample and a temporal holdout in which models trained on 2019 – 20 data were tested on the entirely out-of-sample 2021 – 22 data. Infraction type was categorized into three severity tiers derived from information based on legal definitions.

Results. Results indicated that infraction severity was the dominant predictor across all models and both validated strategies. After controlling for infraction severity, Black students (OR = 1.31, p = .027) and Hispanic/Native American students (OR = 1.24, p =.001) experienced significantly higher odds of harsh discipline relative to White students. Racial disparities were most profound for moderate-severity infractions (Tier 2), where administrative discretion is greatest. All models demonstrated accuracy in the temporal validation (F1 = .783 - .786) with approximately 80% recall for harsh outcomes.

Conclusion. Results indicate that disparities are not fully explained by the severity of conduct alone. Findings highlight the potential value of data-informed, preventative intervention targeting discretionary decision points to reduce inequitable discipline outcomes, particularly for Native American Students.