A Machine Learning Approach to Predict Voluntary Separation from the Australian Army

Abstract

First-term attrition (FTA)—separation before completing an initial service obligation—imposes readiness and training costs on the U.S. Navy. This thesis tests whether documented mental health diagnoses are associated with FTA among first-term sailors. Using the Person-Event Data Environment administrative and medical records from 2014–2024, data were cleaned and consolidated into a person-level dataset of 469,823 active-duty junior enlisted (E01–E05) and junior officers (O01–O02). FTA was coded as separation prior to completion of the first obligation; administrative separation (ADSEP) was summarized descriptively within the FTA subset. Analysis combined descriptive statistics with nested logistic regression models that refine mental health measures from an “all-inclusive” diagnostic category to broad categories and an individual diagnostic specification in a joint model. Linear probability models provide robustness checks. Results indicate heterogeneous associations: the elevated risk suggested by “any diagnosis” attenuates when diagnoses are disaggregated, and the strongest, most consistent associations with FTA are concentrated with substance abuse diagnoses. After accounting for diagnosis composition, most demographic and occupational differences diminish, while rank retains a strong gradient in FTA risk. These findings support diagnosis-specific targeting, particularly for substance abuse conditions, when designing early retention interventions for first-term sailors.