Unstructured covariance best fit craving trajectories in MOUD follow-up
Secondary analysis from the CTN0051 trial has put a statistical problem at the center of a clinically important question in opioid use disorder: how best to model craving over time when patients are already living with overlapping substance use risks, changing treatment exposure, and repeated symptom assessments. Investigators compared 10 covariance structures in repeated-measures linear mixed-effects models to track cravings for opioids, alcohol, stimulants, and nicotine across 36 weeks of medication treatment. The central result was not a drug superiority claim or a relapse endpoint. It was a modeling decision with direct consequences for how longitudinal craving data are interpreted in future trials and secondary analyses.
The unstructured covariance form produced the best fit for alcohol, opioid, and stimulant craving, and the second-best fit for nicotine craving. In plain clinical terms, the repeated observations did not behave like a neat, fixed pattern that could be forced into a simple statistical box. The covariance between visits appears to have been complex enough that the model had to retain maximum flexibility to describe the data faithfully. In a field where craving often fluctuates with stress, withdrawal relief, treatment adherence, and comorbid psychiatric symptoms, that flexibility is more than a technical preference. It is a safeguard against flattening clinically meaningful variation.
Craving fell across all four domains during follow-up
Using the best-fitting covariance structure, craving scores declined significantly from baseline across all four measured domains: opioid craving, alcohol craving, stimulant craving, and nicotine craving. Those reductions were observed during follow-up in participants randomized to either buprenorphine-naloxone or extended-release naltrexone. The finding is clinically familiar in broad outline and still important in detail: effective medications for opioid use disorders were accompanied by lower self-reported craving over time, even when the analysis was built to preserve the complexity of the repeated measures.
Craving is not a peripheral symptom in this population. It sits near the core of relapse risk, treatment engagement, and day-to-day functioning. Lower reported craving across several substance classes suggests that stabilization during MOUD may extend beyond opioid-specific effects, although the analysis was not designed to prove causality for every craving domain. The repeated decline also reinforces the value of tracking craving longitudinally rather than treating it as a one-time baseline descriptor. Patients do not experience craving as a static trait, and the model used here respected that reality.
Treatment differences emerged only for stimulant craving at selected visits
A more selective signal appeared in the treatment-by-visit interaction for stimulant craving, where differences between groups emerged at selected time points. The pattern did not apply uniformly across all craving domains. Instead, it suggests that the two medications may influence the trajectory of stimulant craving in a way that is time-dependent and not captured by a single average effect across the full follow-up window.
That kind of divergence can matter at the bedside because stimulant craving often tracks a different biological and behavioral pathway than opioid craving itself. Patients with opioid use disorder frequently present with polysubstance exposure, and craving for stimulants may reflect a distinct interaction among reward circuitry, stress responsivity, and psychiatric comorbidity. A model that can detect visit-specific divergence offers a more realistic view of how symptoms move during treatment. It also warns against overgeneralizing from a single summary estimate when the underlying symptom pattern changes across the observation period.
Baseline comorbidity shaped craving severity before follow-up began
Baseline clinical characteristics helped distinguish who started with more intense craving. Alcohol use disorder and amphetamine use disorder predicted higher alcohol craving. Baseline depression, amphetamine use disorder, and cocaine use disorder predicted greater stimulant craving. Cocaine use disorder at baseline also predicted elevated nicotine craving.
Those associations fit a clinically plausible pattern. Craving is not isolated from psychiatric burden or from prior substance exposure; it tends to cluster with the neurobehavioral history already present when treatment begins. Depression may intensify reward dysregulation and stress-driven craving, while prior stimulant or cocaine use may prime craving pathways that remain active even after opioid-focused treatment starts. The presence of alcohol use disorder alongside higher alcohol craving also reflects how quickly one substance can cue another in patients with overlapping use disorders. In practical terms, baseline diagnosis still carries prognostic weight, even when the treatment question is centered on MOUD.
Why model choice matters in longitudinal addiction research
Longitudinal craving data are notoriously sensitive to how within-person correlation is specified. Repeated measurements from the same participant are not independent, and a model that misrepresents that correlation can distort standard errors, significance testing, and the apparent shape of symptom change. The investigators compared 10 commonly used covariance structures using both Akaike Information Criterion and Bayesian Information Criterion, then selected the structure that best matched the observed data. For alcohol, opioid, and stimulant craving, the unstructured option outperformed the alternatives. For nicotine craving, it still ranked near the top.
This has direct methodological relevance for addiction trials and secondary analyses. Craving is often used as a surrogate marker, a mechanistic outcome, or a covariate linked to relapse risk. If the underlying covariance is oversimplified, the resulting inference may understate or overstate clinically meaningful change. A flexible structure is computationally heavier, but it can preserve the subtleties of symptom evolution that simpler models miss. In a dataset spanning baseline and every four weeks over 36 weeks, that extra fidelity is not academic excess; it is a way to avoid forcing the data into the wrong clinical story.
Secondary analysis with clear limits on clinical translation
This analysis drew on publicly accessible data from CTN0051, and the original trial had institutional review board approval with written informed consent. The present work was exempt from additional IRB review because it used secondary data. No competing interests were declared, and no funding source was listed for this paper.
The most important boundary is straightforward: these findings identify statistical fit and symptom associations, not treatment-selection rules. Baseline alcohol, amphetamine, depression, and cocaine-related characteristics were associated with higher craving, but that does not yet make them reliable tools for choosing between buprenorphine-naloxone and extended-release naltrexone in routine care. The same is true for the stimulant craving interaction by visit. Prospective validation would be needed before these patterns could guide relapse prediction or individualized medication selection with confidence.
For clinicians and trialists, the message is concrete. Craving trajectories in opioid use disorder are heterogeneous, comorbidity-sensitive, and better captured by a flexible repeated-measures structure than by a simplified one. The model choice changes how the data speak, and here it allowed the longitudinal signal to emerge without forcing a false uniformity across alcohol, opioid, stimulant, and nicotine craving.