Response to Wang and Winterstein: Robustness of GIP/GLP-1 RA findings in substance use disorders

We appreciate the thoughtful commentary by Wang and Winterstein [1] on our study [2], which provides an opportunity to address their concerns and further elucidate the robustness of our findings.

Regarding the concern about the comparator group lacking a defined clinical context, this issue was carefully considered. To minimize biases from misaligned trajectories, we assigned index dates for the comparator group randomly within the observational window, per Hernan and Robins [3]. In response to Wang and Winterstein [1], we conducted an additional analysis aligning the time intervals between initial opioid use disordert (OUD)/alcohol use disorder (AUD) diagnosis and study entry for both groups. The adjusted incidence rate ratio (aIRR) and 95% CI for opioid overdose among those prescribed GIP/GLP-1 receptor agonists (RAs) compared to those not prescribed was 0.52 (0.29–0.92) for the OUD cohort. Similarly, for alcohol intoxication, the aIRR (95% CI) was 0.53 (0.33–0.87) for the AUD cohort. These findings align closely with the original results of 0.60 (0.43–0.83) and 0.50 (0.40–0.63), underscoring the robustness of our conclusions.

Regarding residual confounding because of healthcare engagement, treatment-seeking behaviors and comorbidities, we used comprehensive strategies to address these challenges. Our regression models adjusted for baseline demographic, clinical and healthcare utilization variables [4]. Additionally, we used inverse probability of treatment weighting (IPTW) to balance measured confounders and minimize bias [5]. Sensitivity analyses excluding patients with prior overdose or intoxication histories demonstrated that the protective associations remained stable, rather than being driven by differential substance use histories between prescribed and non-prescribed patients. Adjustment of healthcare utilization and severity markers confirmed robustness beyond confounding by indication [6]. Stratified analyses revealed no substantial differences in effect estimates for subgroups with Type 2 diabetes or obesity, confirming consistency across populations. We respectfully refute the statement by Wang and Winterstein that our results ‘shifted toward the null’ or demonstrated a ‘weakened protective effect’ when restricted to patients with Type 2 diabetes or obesity. The observed aIRRs were nearly identical, and the overlapping CIs across subgroups confirm consistent findings [7], reminding the commentators of the distinction between absolute and statistical differences. Furthermore, our study focused on a general OUD/AUD population and was not limited to patients with diabetes or comparing different diabetic medications. Assertions about active versus non-active comparators are, therefore, beyond the scope of our analysis.

We also addressed concerns about unmeasured confounders, including proxies for disease severity, healthcare access and provider practices, through rigorous methodological approaches such as IPTW, random-effects modeling and sensitivity analyses [8, 9]. Our models included variables such as comorbidity burden, overdose/intoxication history and concurrent mental health or substance use disorders. In response to Wang and Winterstein [1], we further conducted additional analyses, adjusting for ICU and emergency department visits within 1 month of OUD or AUD diagnosis, which yielded consistent results [e.g. OUD aIRR = 0.60 (0.44–0.84); AUD aIRR = 0.51 (0.40–0.63)]. Moreover, we previously used hospital system as a random effect to account for institutional differences and conducted a multilevel, hierarchical analysis to address provider-level variability [10], with similar outcomes [e.g. OUD aIRR = 0.57 (0.40–0.80); AUD aIRR = 0.55 (0.43–0.70)].

We addressed data continuity and loss to follow-up by incorporating ‘time-on-study’ into all rate calculations and applying alternative modeling approaches, such as Andersen-Gill models for recurrent events [11]. Extended follow-up analyses reaffirmed the protective effects of GIP/GLP-1 RAs. In some analyses, we required patients to have 1/2 years of follow-up, whereas in others, inclusion stopped to ensure at least 1/2 years of follow-up. For the 2-year follow-up, the effect among AUD patients remained statistically significant, and whereas the effect in OUD patients approached significance (aIRR = 0.70 [0.46–1.06]; p = 0.09), likely reflecting reduced statistical power because of smaller n, the trend remained consistent.

Claims of concern regarding insurance, differential exposure and outcome measurement overlooked the comprehensive analyses we performed. For example, we specifically accounted for insurance in our main analyses and matched on it along with other variables in a sensitivity analysis, yielding consistent results.

Our findings contribute to the growing evidence on the potential role of GIP/GLP-1 RAs in addiction medicine, providing a strong foundation for future prospective research. We value the engagement of Wang and Winterstein and their contribution to advancing addiction research. Constructive dialogue such as this is essential for refining methodologies and ensuring the reliability of real world data (RWD) studies [12].

None.