Stratifying First-Year Dropout Risk Using Pre-Enrolment Administrative Data in a French Sport and Exercise Science (STAPS) Programme
Abstract
Early stratification of dropout risk could help universities direct support before students disengage, but models relying exclusively on pre-enrolment information may have limited predictive power. This retrospective study assessed whether pre-enrolment administrative data could stratify first-year dropout risk among students entering a French sport and exercise science (STAPS) programme. Dropout was operationally defined as failure to complete the first year because of unjustified absence from examinations. Data from 2,068 students enrolled at Université Claude Bernard Lyon 1 across the 2022–2023 to 2024–2025 academic years were analysed using Lasso-regularised logistic regression. Performance and predictor stability were assessed across 100 repeated stratified 80%/20% training–test partitions, with all preprocessing, tuning, and threshold selection confined to the training data. Overall, 626 students (30.3%) met the dropout definition. Mean held-out ROC-AUC was 0.663 (SD 0.029), with mean sensitivity of 0.702 and specificity of 0.558. Prior academic indicators, particularly general baccalaureate and baccalaureate distinction, were the strongest and most stable predictors, although demographic, geographical, socioeconomic, and STAPS-specific variables also contributed. Individual classification remained limited (pooled positive predictive value: 0.409); nevertheless, across pooled held-out predictions, observed dropout increased from 10.8% for scores below 0.15 to 51.3% for scores above 0.50. Pre-enrolment administrative data therefore provided modest but consistent group-level risk stratification. The model may support inclusive, non-punitive allocation of preventive resources, pending external validation, but should not inform high-stakes individual decisions.
Related
- Dropout Universality: Scaling Laws and Optimal Scheduling at the Edge-of-Chaos
- Inference for sparsely sampled Gauss-Markov processes under outcome-dependent dropout
- Approximation Property of Dropout Neural Networks: Sobolev Rates and Confidence Bounds
- Optimal Transport Dropout for Structured Predictive Uncertainty
- Learning environment at home and primary school dropout status in Bangladesh: An analysis of multiple indicator cluster survey
BibTeX
@misc{rienzo2026stratifying,
title = {{Stratifying First-Year Dropout Risk Using Pre-Enrolment Administrative Data in a French Sport and Exercise Science (STAPS) Programme}},
author = {Franck Di Rienzo and Mélanie Gallot},
howpublished = {HAL},
year = {2026},
month = sep,
url = {https://hal.science/hal-05739908v1},
note = {CC BY-NC 4.0},
}
RIS
TY - UNPB
TI - Stratifying First-Year Dropout Risk Using Pre-Enrolment Administrative Data in a French Sport and Exercise Science (STAPS) Programme
AU - Franck Di Rienzo
AU - Mélanie Gallot
PY - 2026
DA - 2026/09/06
PB - HAL
UR - https://hal.science/hal-05739908v1
AB - Early stratification of dropout risk could help universities direct support before students disengage, but models relying exclusively on pre-enrolment information may have limited predictive power. This retrospective study assessed whether pre-enrolment administrative data could stratify first-year dropout risk among students entering a French sport and exercise science (STAPS) programme. Dropout was operationally defined as failure to complete the first year because of unjustified absence from examinations. Data from 2,068 students enrolled at Université Claude Bernard Lyon 1 across the 2022–2023 to 2024–2025 academic years were analysed using Lasso-regularised logistic regression. Performance and predictor stability were assessed across 100 repeated stratified 80%/20% training–test partitions, with all preprocessing, tuning, and threshold selection confined to the training data. Overall, 626 students (30.3%) met the dropout definition. Mean held-out ROC-AUC was 0.663 (SD 0.029), with mean sensitivity of 0.702 and specificity of 0.558. Prior academic indicators, particularly general baccalaureate and baccalaureate distinction, were the strongest and most stable predictors, although demographic, geographical, socioeconomic, and STAPS-specific variables also contributed. Individual classification remained limited (pooled positive predictive value: 0.409); nevertheless, across pooled held-out predictions, observed dropout increased from 10.8% for scores below 0.15 to 51.3% for scores above 0.50. Pre-enrolment administrative data therefore provided modest but consistent group-level risk stratification. The model may support inclusive, non-punitive allocation of preventive resources, pending external validation, but should not inform high-stakes individual decisions.
ER -