Model-based estimation of drug use prevalence using item count data

The item count (IC) method for estimating the prevalence of sensitive behaviors was applied to the National Survey on Drug Use and Health (NSDUH) to estimate the prevalence of past year cocaine use. Despite considerable effort and research to refine and adapt the IC method to this survey, the method failed to produce estimates that were any larger than the estimates based on self-reports. Further analysis indicated the problem to be measurement error in the IC responses. To address the problem, a new model-based estimator was proposed to correct the IC estimates for measurement error and produce less biased prevalence estimates. The model combines the IC data, replicated measurements of the IC items, and responses to the cocaine use question to obtain estimates of the classification error in the observed data. The data were treated as fallible indicators of (latent) true values and traditional latent class analysis assumptions were made to obtain an identifiable model. The resulting estimates of the cocaine use prevalence were approximately 43 percent larger than the self-report only estimates and the estimated underreporting rates were consistent with those estimated from other studies of drug use underreporting.