Implications of Misclassification Errors in Empirical Studies of Adolescent Smoking Behaviours

Misclassification errors in a dependent variable can introduce attenuation bias to covariate effects in a binary choice model. Misreporting of smoking behaviours by adolescents has been widely documented. However, the consequence in empirical studies of adolescent smoking participation has received little attention. This study uses the Health Survey for England (HSE) to investigate the extent and implication of misclassification errors in self‐reported smoking among adolescents aged 11–15 years. The HSE contains both a self‐reported smoking component and an objective measure of smoking obtained from saliva cotinine assays. Saliva cotinine concentration ≥12 ng/ml is considered the ‘true’ indicator of adolescent smoking participation against which self‐reported smoking is compared. The findings show that smoking is misreported in this age group, resulting in a downwards bias of marginal effect estimates. Given the widespread use of self‐reported smoking data, this study explores the performance of the Hausman, Abrevaya and Scott‐Morton‐modified maximum likelihood estimation (HAS approach) in recovering true estimates of covariate effects. In this context, the HAS approach performs better when the misclassification probabilities are treated as constants compared with when they are treated as conditionally dependent parameters. Copyright © 2016 John Wiley & Sons, Ltd.