MODELING RARE EVENTS IN HEALTH INTERVENTION: AN APPLICATION FROM THE DEPARTMENT OF WAR DATASET
Author(s)
Onur Baser, MA, MS, PhD1, Huseyin Yuce, MS, PhD2, Gabriela Samayoa, MPH, MD3, Shuangrui Chen, MS4, Nehir Yapar, BS4;
1City University of New York (CUNY), Graduate School of Public Health, New York, NY, USA, 2City University of New York (CUNY), Mathematics, New York, NY, USA, 3Catholic University of Honduras, Tegucigalpa, Honduras, 4Columbia Data Analytics, New York, NY, USA
1City University of New York (CUNY), Graduate School of Public Health, New York, NY, USA, 2City University of New York (CUNY), Mathematics, New York, NY, USA, 3Catholic University of Honduras, Tegucigalpa, Honduras, 4Columbia Data Analytics, New York, NY, USA
OBJECTIVES: This analysis compared standard and corrected logistic regression approaches for estimating the probability of intravenous immunoglobulin (IVIg) use in bullous pemphigoid (BP) and quantified the degree of rare-event bias in event probability estimates.
METHODS: A retrospective cross-sectional study was conducted using Department of War TRICARE claims from fiscal years 2019-2022, including adults diagnosed with BP identified by International Classification of Diseases, 10th Revision, Clinical Modification (ICD‑10‑CM) code L12.0. IVIg treatment was defined via J-codes, and demographic and comorbidity data were extracted; standard logistic regression was compared using 2 rare event correction strategies (prior correction and weighting) across age, sex, and comorbidity strata.
RESULTS: Among 2,720 patients diagnosed with BP, 14 (0.5%) received IVIg; treated patients were younger (mean 65.1 vs 75.9 years; P=0.0016) and predominantly female (13/14 vs 1/14 male; P=0.0036). Standard logistic regression systematically underestimated event probabilities: underestimation ranged from 11% to 102% with prior correction and from 15% to 107% with weighting, depending on patient profile. For men aged 18 to 64 years with low comorbidity, the standard model predicted a 0.11% IVIg probability vs ~0.23% with weighting (≈107% underestimation); for women aged ≥65 years with high comorbidity, underestimation was ~11-15%. Random down-sampling of non-events inflated probabilities markedly; in some strata, predicted probabilities were up to 10-fold higher than corrected estimates and up to 20-fold higher than standard model predictions.
CONCLUSIONS: In a large administrative dataset with 14 patients treated with IVIg among the 2,720 diagnosed with BP, standard logistic regression substantially underestimated rare treatment probabilities, and naïve sampling strategies produced large overestimation. Applying rare event corrections and specifying non-event sampling is essential for valid probability estimation in rare disease and rare therapy studies that inform burden of illness, pricing, and access decisions.
METHODS: A retrospective cross-sectional study was conducted using Department of War TRICARE claims from fiscal years 2019-2022, including adults diagnosed with BP identified by International Classification of Diseases, 10th Revision, Clinical Modification (ICD‑10‑CM) code L12.0. IVIg treatment was defined via J-codes, and demographic and comorbidity data were extracted; standard logistic regression was compared using 2 rare event correction strategies (prior correction and weighting) across age, sex, and comorbidity strata.
RESULTS: Among 2,720 patients diagnosed with BP, 14 (0.5%) received IVIg; treated patients were younger (mean 65.1 vs 75.9 years; P=0.0016) and predominantly female (13/14 vs 1/14 male; P=0.0036). Standard logistic regression systematically underestimated event probabilities: underestimation ranged from 11% to 102% with prior correction and from 15% to 107% with weighting, depending on patient profile. For men aged 18 to 64 years with low comorbidity, the standard model predicted a 0.11% IVIg probability vs ~0.23% with weighting (≈107% underestimation); for women aged ≥65 years with high comorbidity, underestimation was ~11-15%. Random down-sampling of non-events inflated probabilities markedly; in some strata, predicted probabilities were up to 10-fold higher than corrected estimates and up to 20-fold higher than standard model predictions.
CONCLUSIONS: In a large administrative dataset with 14 patients treated with IVIg among the 2,720 diagnosed with BP, standard logistic regression substantially underestimated rare treatment probabilities, and naïve sampling strategies produced large overestimation. Applying rare event corrections and specifying non-event sampling is essential for valid probability estimation in rare disease and rare therapy studies that inform burden of illness, pricing, and access decisions.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR69
Topic
Methodological & Statistical Research
Disease
SDC: Rare & Orphan Diseases, SDC: Sensory System Disorders (Ear, Eye, Dental, Skin), SDC: Systemic Disorders/Conditions (Anesthesia, Auto-Immune Disorders (n.e.c.), Hematological Disorders (non-oncologic), Pain)