Mortality Data Quality: Multidimensional Documentation and County-Level Patterns in Undetermined-Intent Death Classification in NVDRS

• In Event: Morbidity and Mortality: Open Call Paper Session

Abstract

Deaths classified as undetermined intent pose persistent challenges for mortality surveillance, as classification may reflect not only case ambiguity but also variation in investigative documentation. We conceptualize data quality in violent death records as a multidimensional construct shaped by evidentiary completeness and institutional context. Using restricted-access National Violent Death Reporting System (NVDRS) data from 2004–2022, we construct a composite data quality index integrating structured investigative completeness, source document availability, and computational assessments of free-text narratives.
NVDRS records include structured categorical fields and free-text investigative narratives. We assess data quality for each component separately. Structured data quality captures the documented evidentiary base of each case, including autopsy findings, whether circumstances are known, the presence of law enforcement and coroner/medical examiner narratives, and the number of source documents abstracted.
Narrative quality is evaluated using supervised deep learning models trained on NVDRS free-text reports. Transformer-based language models generate contextual embeddings to predict multiple categorical attributes—including manner of death, decedent sex, race, weapon type, and precipitating circumstances. Rather than reclassifying cases, we interpret model confidence as an indicator of informational sufficiency: narratives yielding high certainty across dimensions are considered more contextually rich and coherent.
We assess county-level variation in this multidimensional index and examine whether higher rates of undetermined-intent classification are associated with lower documentation quality. Multilevel and spatial models evaluate geographic clustering while accounting for case composition. We further test whether county-level characteristics—including funding environments, social vulnerability, workforce capacity, and rural–urban context—are associated with systematic differences in documentation quality.
By reframing mortality data quality as institutional and computationally measurable, this study shows how geographic variation in violent death classification may reflect documentation environments as much as underlying mortality risk, informing efforts to strengthen investigative standards and population-level mortality statistics.

Authors

• Rafael Geurgas, Purdue University

• Alina Arseniev-Koehler, Purdue University

• Robbee Wedow, Purdue University