The National Survey on Drug Use and Health (NSDUH) estimates serious mental illness (SMI) and any mental illness (AMI) among adults aged 18 or older. These estimates are based on a prediction model of SMI developed from a subsample of NSDUH respondents who answered the NSDUH short scales of psychological distress and functional impairment and who also participated in the Mental Health Surveillance Study (MHSS). The SMI and AMI prediction models were reevaluated to include additional predictors in 2013. This report discusses the bias and error rates that would be associated with SMI models with varying degrees of parsimony and then compares the NSDUH SMI and AMI estimates to estimates derived from the Behavioral Risk Factor Surveillance System (BRFSS), National Health Interview Survey (NHIS), and Medical Expenditure Panel Survey (MEPS) data.
Using 5,000 clinical interviews collected in the 2008 to 2012 MHSS, this analysis assessed SMI prediction models with varying degrees of parsimony to see how the models would perform in terms of bias (overall and at the domain level), total error rate (sum of false-positive and false-negative rates), and how model-based estimates would compare against direct estimates as computed from the clinical sample.
This report demonstrates that models that include the World Health Organization Disability Assessment Schedule (WHODAS) term, suicidal thoughts, past year major depressive episode, and past year version of the Kessler-6 scale are important to control bias or the total error rate. The absence of WHODAS information from the BRFSS, NHIS, and MEPS studies may explain in large part why SMI and AMI estimates obtained from those studies differ markedly from those obtained from NSDUH.
The revised prediction model for SMI and AMI currently used in NSDUH meets the study requirements and variations in SMI estimates from other studies reflects the lack of needed predictors in addition to other methodological variations across studies.
