The search for an elusive cutoff remains: Problems of binary classification of heavy drinking as an endpoint for alcohol clinical trials.

Abstract

To examine whether a clinically meaningful alcohol consumption cutoff can be created for clinical samples, we used receiver operator characteristic (ROC) curves to derive gender-specific consumption cutoffs that maximized sensitivity and specificity in the prediction of a wide range of negative consequences from drinking. We conducted secondary data analyses using data from two large clinical trials targeting alcohol use disorders: Project MATCH (n=1726) and COMBINE (n=1383). In both studies, we found that the ideal cutoff for men and women that maximized sensitivity/specificity varied substantially both across different alcohol consumption variables and alcohol consequence outcomes. Further, the levels of sensitivity/specificity were poor across all consequences. These results fail to provide support for a clinically meaningful alcohol consumption cutoff and suggest that binary classification of levels of alcohol consumption is a poor proxy for maximizing sensitivity/specificity in the prediction of negative consequences from drinking. Future research examining consumption-consequence associations should take advantage of continuous measures of alcohol consumption and alternative approaches for assessing the link between levels of consumption and consequences (e.g., ecological momentary assessment). Clinical researchers should consider focusing more directly on the consequences they aim to reduce instead of relying on consumption as a proxy for more clinically meaningful outcomes.