Feasibility of predicting firearm type in firearm suicide deaths for better policy evaluation

Abstract

Abstract Background Imprecise firearm-type data limits capacity to evaluate firearm-type-specific policies; a striking proportion of firearm suicides in the National Vital Statistics System (NVSS) do not include information on firearm-type. Firearm-type is missing rarely in the National Violent Death Reporting System (NVDRS) but only available recently for certain states. We sought to create a prediction model to impute missing firearm-type in historical firearm suicide data. Methods An 80% training set of NDVRS data (2005–2018; 40 states and D.C.) was fit to candidate models (logistic regression, C5, and random forest) to predict long gun vs. handgun use in firearm suicide deaths. Predictors included all variables common to NVDRS and NVSS data systems, state fixed-effects, and rurality. Stratified tenfold cross-validation was used to identify the best fitting model using the Kappa statistic as the benchmark. Accuracy of the identified best fitting prediction model was then assessed by applying the model to the 20% validation data. Findings Firearm type was specified in 91% of NVDRS firearm suicide deaths. 24.4% of firearm suicides within the included NVDRS data involved long guns. All prediction methods performed poorly. The C5 model had the highest Kappa-14.5% (95%CI 13.9–15.1). However, accuracy in the C5 validation set was inadequate to reliably predict long gun vs. handgun use: sensitivity-13.1% (95%CI 12.2–14.1), specificity-96.0% (95%CI 95.6–96.3). Conclusions NVDRS provides the most detailed data on national firearm-related deaths, information that cannot be supplemented from exiting sources. Continued investment in NVDRS and other firearm injury surveillance is crucial for future evaluations of firearm-type-specific policies.

Keywords

• Firearm type
• Firearm suicide
• Data quality
• Imputation
• Prediction