Evaluating Regression Models for Military Logistics Support: A Reproducible Screening Procedure
Purpose: To formulate and demonstrate an evidence-based screening procedure that separates reproducible regression calculations from justified use in military logistics resource decisions. Design/Methodology/Approach: A document-based reanalysis retains nine records, 23 candidate inputs and nine outcomes. Because measurement provenance and unit calibration are unverified, the data serve as a numerical illustration. Fixed one-, two- and five-predictor linear models and an additive quadratic model are compared with a mean-only benchmark. Fold-local preprocessing, matrix-rank checks, leave-one-out cross-validation, all 36 paired holdouts and reduced-table sensitivity analyses support four linked assessments: data credibility, computational correctness, predictive evidence and decision feasibility. Findings: The five-predictor model attains in-sample coefficients of determination of 0.315–0.886 but exceeds benchmark error for eight of nine outcomes in leave-one-out evaluation. The quadratic model interpolates the data without residual degrees of freedom and performs worse than the benchmark for all outcomes. For the first outcome, the two-predictor model reduces root mean squared error by 15.3% in leave-one-out evaluation and 7.1% in paired holdouts; the advantage reverses in two of nine reduced tables. The resulting evidence does not support operational release of any specification. Practical Implications: The procedure links each diagnostic to the evidence needed for further use, preventing numerical accuracy from substituting for credible measurement, stable prediction or feasible planning outputs. Originality/Value: The application-specific contribution is a non-compensatory screening procedure and a worked decision record connecting established validation principles with transparent, outcome-level diagnostics. It is a methodological synthesis, not a new estimator or an empirically validated certification standard.