Estimating Drug Involvement in Fatal Overdoses With Incomplete Information

Abstract

Death certificates provide incomplete information on the specific drug categories involved in fatal overdoses. The accuracy of previously developed corrections for this and modifications to them was examined. Uncorrected mortality rates were compared with rates from the preferred correction models. Data were obtained for the universe of 932,364 drug overdoses in the U.S. between 1999 and 2020, including 769,982 (82.6%) with a drug classification and 162,382 (17.4%) without, from the National Center for Health Statistics Multiple Cause of Death files. Opioid and cocaine involvement for unclassified overdoses were estimated using multiple approaches. Prediction accuracy was assessed as the mean absolute deviation between actual and estimated involvement for a test sample with known drug involvement. Corrected death rates from preferred models were compared with uncorrected rates. Analyses were conducted in 2022-2023. Previous regression-based corrections that controlled for decedent characteristics can be improved upon by adding state-fixed effects as covariates. Once this is done, supplementary controls for county characteristics or contributory causes of death do not materially improve prediction accuracy. Properly designed naïve models, which distribute unspecified drug deaths proportionately to specified deaths, often provide similar results and, for county-level analyses, supply the most accurate predictions. Uncorrected results substantially understate the levels of opioid and cocaine involvement and can either understate or overstate changes over time. Failing to correct for incomplete information on death certificates leads to inaccurate counts of deaths from specific categories of drugs, such as opioids. However, relatively simple corrections are available that substantially improve accuracy.