Forecasting next-day urge to drink without craving’s own history: an individualized analysis of responsive EMA in residential alcohol-use-disorder recovery
Year: 2026
Type: preprint
Abstract: Background: Just-in-time adaptive interventions (JITAIs) for substance-use recovery need decision inputs that anticipate risk, yet deployed systems rely largely on static rules. A published 30-day responsive ecological momentary assessment (rEMA) study followed adults recovering from severe alcohol use disorder in residential treatment. It collected five daily self-report items: urge to drink, affect intensity, stress, alcohol use, and importance of Alcoholics Anonymous engagement. Adherence and interest in prompt-based support were high. Whether these sparse reports carry a short-horizon forecasting signal was untested.Methods: We ran a retrospective, individualized forecasting analysis of that cohort (61 enrolled; 53 analytic). Each forecast answers one question: issued daily at 15:30, will the participant’s next 24 hours contain any reported urge to drink (a positive day), or remain urge-free (a negative day)? Rolling-origin ordinary least squares models were fit within each participant, on data available at forecast time, after a burn-in of 10 days. Three pre-specified feature sets decompose the signal: FULL (all five items plus time structure), CONTEXT (urge history withheld), and AR (urge history alone). Forecasts were scored against a per-participant base-rate baseline, with participant-level bootstrap intervals and Wilcoxon tests.Results: We scored 1,013 participant-days. The cohort summary curves gave mean AUCs of 0.66 (95% CI 0.589 to 0.730) for CONTEXT and 0.62 (0.560 to 0.673) for FULL. Per-participant AUCs exceeded chance in both arms (p = 0.0018 and 0.0077). AR’s cohort summary curve gave 0.56 (0.504 to 0.630); its per-participant AUCs did not exceed chance (p = 0.24). At the operating point, CONTEXT raised recall to 0.619 against the baseline’s 0.366; the false-positive rate did not differ (0.351 versus 0.355). No between-arm difference was significant within participants.Conclusion: In this cohort, next-day urge to drink was forecastable above chance from sparse rEMA, even with the urge item’s own history withheld. We found no evidence that this history adds signal beyond the other items. These results indicate which items a JITAI trigger model should monitor, groundwork for prospective validation. We do not model drinking or relapse.
Language: English
Subfield: Epidemiology
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7204266249