The repeated measurements design of the National Household Survey on Drug Abuse (NHSDA) make it an ideal vehicle for the application of latent class models (including the Hui-Walter method) for estimating the validity of self-reports of drug use. Since very little is known about validity in the NHSDA and in other similar household surveys, estimates of the reporting bias in this survey could have important benefits for both research and policy making. The general aims of the present research are to a) apply existing methods and to develop new methodology for estimating the validity of self-reported drug use in the National Household Survey on Drug Abuse and other similarly designed surveys and b) to analyze the validity of the NHSDA survey instrument using this methodology.
The specific aims of the research include the following: a) extend the Hui-Walter methodology from S = 2 groups and R = 2 repeated measurements to S = 3 or more groups and R = 3 or more measurements using the latent class models approach; b) enhance and extend the current RTI software for maximum likelihood estimation of general latent class models, c) evaluate a wide range of models and error structures for describing the classification error in NHSDA self-reports of drug use; and d) characterize the populations most prone to misclassification in the NHSDA, test hypotheses regarding the causes of the errors, and evaluate the effects of the errors on NHSDA estimates of drug use prevalence. The proposed research program will include: a) an application of general latent class models (including the HuiWalter models) to the 1991, 1992, and 1993 NHSDA data, a national sample of 30,000 persons (each year) using alternative group definitions, drug use measures, and all drugs covered by the NHSDA; b) an analysis of the estimates of false positive and false negative probabilities relating survey and respondent characteristics to the magnitudes of the observed measurement biases; and c) enhancements and additions to the RTI software such as the capability of estimating the variance of the latent class model estimates in complex surveys such as the NHSDA, an option for implementing the EM algorithm for maximizing the likelihood function for the general latent class model problem, and user documentation for using the software outside of RTI. While the primary motivation for this research program is methodological, it will also provide important substantive information. In particular, our analysis will provide precise estimates of the patterns of measurement bias (particularly, underreporting bias) in the NHSDA. The relationships of drug use underreporting with time, geography demographic and economic characteristics, and interview characteristics will be explored and quantified.
