Repeated measures estimation of measurement bias for self-reported drug use with applications to the National Household Survey on Drug Abuse

Abstract

Direct estimates of response bias for self-reports of drug use in surveys require that essentially error free determinations of drug use be obtained for a subsample of survey respondents. The difficulty of obtaining determinations which are accurate enough for estimating validity is well-documented in the literature. Methods such as specimen (e.g., hair, urine) analysis, proxy reports, and the use of highly private and anonymous modes of interview all have to contend with error rates which may only be marginally lower than those of the parent survey. Thus, any methodology for direct validity estimation must rely to some extent on approximations and questionable assumptions. In this chapter, the authors consider a number of methods that rely solely on repeated measures data to assess response bias. Since the assumptions associated with these approaches do not require highly accurate second determinations they may be more easily satisfied in practice. One such method for bias estimation for dichotomous variables that is considered in some detail provides estimates of misclassification probabilities in the initial measurement without requiring that the second measure be accurate or even better than the first. This methodology does require, however, that two subpopulations exist which have different rates of prevalence but whose probabilities of false positive and false negative error are the same. The applicability of these methods for self-reported drug use are described and illustrated using data from the National Household Survey on Drug Abuse. In the discussion of the results, the importance of these methods for assessing the validity of self-reported drug use are examined.