Using multiple imputation by super learning to assign intent to nonfatal firearm injuries.

The number of nonfatal firearm injuries in the US by intent (e.g., due to assault) is not reliably known: First, although the largest surveillance system for hospital-treated events, the Healthcare Cost and Utilization Project Nationwide Emergency Department Sample (HCUP-NEDS), provides accurate data for the number of nonfatal firearm injuries, injury intent is not coded reliably. Second, the system that reliably codes intent, the CDC's National Electronic Injury Surveillance System - Firearm Injury Surveillance Study (NEISS-FISS), while large enough to produce stable estimates of the distribution of intent, is too small to produce stable estimates of the number of these events. Third, a large proportion of cases in NEISS-FISS, notably in early years of the system, are coded as of "undetermined intent." Trends in the proportion of nonfatal firearm injuries by intent in NEISS-FISS thus depend on whether these cases are treated as a distinct category, or, instead, can be re-classified through imputation. We contrast the distributions of nonfatal firearm injury by intent generated using multiple imputation with those generated using complete-case analyses and analyses that consider "undetermined intent" as a distinct category. We produce estimates of the annual number of firearm injuries by intent in a two-step process. First, we impute intent for cases coded as "undetermined" using Multiple Imputation by Super Learning (MISL). Second, we apply MISL-derived distributions to aggregate count data from HCUP-NEDS. The proportion of non-fatal firearm assaults appears to increase over time when injuries coded as undetermined are included as a category. By contrast, the proportion of assaults remains relatively constant over time in complete-case and multiply imputed analyses. Differences between complete-case and multiple imputation approaches become apparent in subgroup analyses. Trends in the number of nonfatal firearm injuries by intent, 2006-2016, derived in our two-step process, are relatively flat. Multiple imputation strategies recovered intent distribution trends that differed from trends derived using methods that are not designed to account for the multiple complex relationships of missingness present in NEISS - FISS data. When applied to NEISS - FISS, MISL imputation produces plausible distributional estimates of firearm injury by intent.