Statistical Uncertainty under Extreme Class Imbalance: An Auditable Applied-Statistics Study of Accreditation in Ecuadorian Higher Education

Abstract

This study examines how extreme class imbalance shapes the interpretation of administrative indicators in higher education quality assurance. The analysis linked 2018 records from 55 Ecuadorian universities and polytechnic schools with document-verified outcomes from the 2019 institutional evaluation process, formally resolved in 2020. The cohort included 52 accredited institutions and three historical non-accreditation outcomes. The analytical strategy combined documentary verification, data-quality checks, medians, median differences, Cliff’s delta, and leave-one-minority-case sensitivity analysis; a majority-class reference and original-direction within-sample ROC AUC were retained only as secondary diagnostics. All 55 outcomes and eight exclusion decisions were confirmed, and seven flagged values were verified against the original sources without correction. The majority-class rule reached 94.55% accuracy but yielded 0% sensitivity for historical non-accreditation and 50% balanced accuracy. Effect estimates also changed substantially after removing individual minority observations, revealing strong case-level dependence. The proposed workflow supports institutional self-evaluation, evidence review, and responsible data governance by identifying findings that require documentary, contextual, or qualitative follow-up. Given the scarcity of independent minority outcomes, the results do not justify stable prediction, institutional ranking, operational thresholds, or automated regulatory classification.