Questionable generalizability of Alcohol Use Disorders Identification Test-Consumption scoring warrants caution when used for outcome monitoring: Evidence from simulated and real-world trial data

Abstract

Background and aims: The "Alcohol Use Disorders Identification Test - Consumption" (AUDIT-C), designed for primary-care screening, is frequently repurposed for outcome monitoring in brief intervention trials targeting problematic alcohol use. This repurposing may distort the instrument's internal structure and introduce right censoring, potentially undermining its ability to meaningfully capture problematic use and change thereof. The aim of the current study was to examine these concerns.

Design: Psychometric study.

Setting and participants: Data from three sources were used: (1) individual-participant data from an internet-based brief intervention trial (n = 1169); (2) aggregated data from k = 15 additional brief intervention trials; and (3) k = 20 000 simulated cohorts generated using statistics from general-population samples.

Measurements: Internal structure of the AUDIT-C was examined through cross-item correlations, item step response functions (ISRF), and more. Responsiveness was assessed using interaction analysis, with changes in alcohol standard units (SU_t2-t1) as the outcome, AUDIT-C_t2-t1 and baseline SU as predictors, and further probing using a simple slopes approach.

Findings: In contrast to general-population cohorts, most brief intervention trials (68.8%) exhibited non-positive associations between frequency and quantity items. Congruently, ISRFs revealed non-monotonic patterns, disrupting ordinal measurement. Simulations suggested that negative frequency-quantity correlations appear at cut-offs of four (r = -0.04, 95% confidence interval [CI]: -0.019 - -0.068) or three (r = -0.13, 95% CI: -0.101 - -0.149). A one-unit AUDIT-C_t2-t1 change represented greater average change in SU_t2-t1 at higher baseline consumption, supported by an interaction (β = 0.05, SE = 0.02, p = 0.005) and sequential contrasts between simple slopes (e.g. 80th vs. 90th percentile: β = 0.31, SE = 0.11, p = 0.035).

Conclusions: When used with typical brief intervention samples, using the AUDIT-C for outcome monitoring risks right-censoring (and thereby false negatives) and non-meaningful total scores. Researchers and clinicians should reconsider repurposing the AUDIT-C as an outcome measure in future alcohol intervention studies and re-examine prior trials that relied on it, to improve the quality of evidence.

Keywords: AUDIT‐C; alcohol use disorders identification test; brief interventions; collider bias; problematic alcohol use; psychometrics; responsiveness; right censoring; selection bias.