Accurate time-series data on firearm injuries in the United States are critical for several important purposes, including documenting the total cost or social burden from firearm injuries, evaluating the effects of policies designed to reduce such injuries, understanding the factors associated with firearm injuries, and documenting changes in firearm injury rates over time. Unfortunately, comprehensive data on firearm injuries in the United States are not collected, and the data that are available have substantial limitations that restrict their usefulness for these purposes. Many of the available data on injuries are on small or unrepresentative samples or are available only at the national level, or sources have changed their data collection procedures over time. The absence of a national data system that provides reliable, complete, and comparable state-level data on nonfatal firearm injuries severely limits our understanding of temporal and geographic variation in the extent of firearm-related morbidity.
Some accurate and complete data on firearm injuries requiring hospitalization are available from State Inpatient Databases (SID). However, they still suffer from several limitations for research into firearm injuries. Several states do not currently report to such databases or have only recently begun to abstract these records into such a database, and the existing state-level databases are not available from any single source. Furthermore, even when SID data exist for a given year, they could have substantial missingness on the particular portion of the medical record that is needed for identifying injuries caused by firearms.
The goal of the current research is to create a state-level panel database of inpatient hospitalizations for firearm injury from 2000 through 2016. We address the limitations with existing state-level data by combining most of the available SID data on firearm injury and other relevant covariates into a single source, imputing data for the state-years where such information is missing, and correcting for biases in the number of firearm injuries contained in the SID on the basis of the extent to which injury data in each individual SID file are missing information about the mechanism, or cause, of injuries.
Data
As part of this project, we pulled together several different data sources for the rate of firearm hospitalizations in each state-year between 2000 and 2016. The largest source of data was from summaries of SID data available through the Healthcare Cost and Utilization Project (HCUP)'s online database, HCUPnet (HCUPnet, undated). However, this data source was supplemented with further information from full-state SID data, state health department data on hospitalization from individual state web portals, and data provided by state departments of health in response to our direct requests. This process resulted in 572 state-years with observed firearm hospitalization rates between 2000 and 2016, out of 850 possible state-years. Data on the amount of missing information on the causes of injuries within the state hospitalization data were extracted from two Agency for Healthcare Research and Quality reports (Barrett et al., 2016; Healthcare Cost and Utilization Project, 2020) that provide the percentage of injuries in the SID that lack information about the mechanism of injury for specific state-years. The analyses also included a variety of covariates that are hypothesized to be associated with the rate of firearm hospitalizations. These covariates, which were drawn from several sources, were the rate of nonsuicide firearm mortality; the rate of reported violent crimes involving firearms; the proportion of hospitalizations caused by firearms within the National Inpatient Survey; and a variety of state-level annual demographic, economic, and social characteristics.
Method
Both the imputation of missing hospitalization data and the correction for incomplete injury mechanism data were done simultaneously in a single Bayesian regression model. Multiple imputation data sets were created corresponding to the Markov Chain Monte Carlo samples from the Bayesian model. The model used a complex error structure that was designed to capture key features of these data, including the larger variance in firearm hospitalization rates for less-populous state-years and the correlation of values within each state over time.
Findings and Conclusions
The model provided an excellent fit to the observed firearm hospitalization rate data, with a model R2 of 0.92. Because the available covariates are strongly associated with the firearm hospitalization rate, the model can estimate the missing information with a relatively high degree of precision. We evaluated the imputed data to identify possible problems with the imputations. These analyses revealed no serious problems, demonstrating that (1) the imputed and nonimputed values have very similar relationships with a variety of covariates; (2) the imputed and nonimputed values have very similar state-level clustering, allowing the same analytic methods to be equally appropriate for both types of data; and (3) the uncertainty in the multiple imputed values is similar to the true error of prediction as assessed through cross-validation.
The resulting data, provided in the RAND Database of Hospitalizations for Firearm Injury, reveal substantial differences—some by an order of magnitude—in the rate of inpatient hospitalizations for firearm injury across states, as well as relatively stable national trends in these hospitalizations over time. The variation across states and over time could be used to evaluate the impact of state firearm policies or other state-level factors hypothesized to influence firearm injury and could be a useful predictor of other phenomena at the state level. These estimates can also be useful for calculating the total cost or social burden of firearm injury at the national or state level, although it is important to note that these data do not capture the full scope of firearm-related morbidity (i.e., our measures do not capture emergency department visits for firearm injuries that do not result in subsequent hospitalization, nor do they capture individuals with gunshot wounds who do not obtain medical care).
