Evaluation of Three Statistical Methods for Reclassifying Undetermined Poisoning Deaths within the Oklahoma Violent Death Reporting System As Suicide or Unintentional

BACKGROUND: Poisoning mortality rates by all intents increased 370% in Oklahoma from 1999 to 2012. Of particular importance is unintentional poisoning, which increased fivefold during that same time period. The Injury Prevention Service (IPS) of the Oklahoma State Department of Health (OSDH) tracks suicide by poisoning (SP) and undetermined poisonings (UDP) through the Oklahoma Violent Death Reporting System (OKVDRS), and unintentional poisonings (UIP) in a separate surveillance system. Poisoning deaths reported as undetermined represent a potential underreporting of SP or UIP. Statistical methodologies may be useful to describe the UDP population and to identify potential SP or UIP cases within UDP.

METHODS: 119 UDP and 144 SP deaths from OKVDRS and 1,372 UIP deaths from the UIP surveillance system for years 2011-2012 were analyzed. UDP deaths were compared to SP and UIP using chi-square tests. Stepwise logistic regression found predictive variables of UIP compared to SP deaths. These variables were used in three methods to classify UDP deaths: 1) logistic regression predicted probabilities; 2) single predictor variable; and 3) standardized coefficient scoring. Diagnostic testing and likelihood ratios were calculated for each method.

RESULTS: Stepwise logistic regression identified the following variables to be predictive of SP or UIP: suicidal tendencies; substance abuse; location of death; and the presence of the following substance types: narcotic analgesic, central nervous system stimulant, and inhalant. Method 1 classified 113 UDP deaths as UIP (95%) and 6 deaths as SP (5%). Method 2 classified 91 deaths as UIP (76%) and 28 as SP (24%). Method 3 classified 71 deaths as UIP (60%) and 48 as SP (40%). Using SP and UIP deaths as the gold standard, method 1 had the highest sensitivity for identifying UIP (99%), and method 3 had the highest specificity (86%). Method 3 also had the highest predictive value positive (98%) and positive likelihood ratio (4.7).

CONCLUSIONS:  The classification methods used predicted the proportion of UIP within UDP to be between 60% and 95%. States vary in coroner/medical examiner office practices to determine death intent; the success of methods to describe the potential frequency of underreporting for SP or UIP will also vary by state. UDP classification is a useful exercise to predict additional SP or UIP, which can better inform interventions to prevent poisoning mortality. Considering the current prescription drug overdose epidemic, states should explore using the method best suited to their needs to mitigate underreporting of unintentional or intentional poisoning mortality.