Reply: Time‐Related Biases in Observational Studies of <scp>GLP</scp> ‐1 Agonists and Cancer Risk

Zhiyuan Cheng, Haifang Li, Dongfeng Gu, Fengchao LiangJournal of Diabetes 2026 · DOI

TL;DR

This paper is a response to critiques of a previous study on GLP-1 receptor agonists (GLP-1RAs) and cancer risk. The authors defend their methodological choices, particularly the use of a "predominant-user" design over a "new-user" design, arguing it better reflects real-world treatment patterns and data limitations in China, while also addressing potential biases like prevalent-user and detection bias.

What Problem Does This Paper Address?

The authors are responding to a letter that raised concerns about the methodological rigor of their previous observational study. Specifically, the critique focused on two key aspects: the study's design in defining GLP-1RA exposure and the lack of a "lag period" in analyzing cancer risk. The core problem is how to accurately assess the safety of a drug in real-world settings when patient treatment behaviors are complex and data availability is not always ideal, while simultaneously mitigating potential biases that could skew the results.

Methods

The authors defend their use of a "predominant-user" design, where patients are classified as GLP-1RA users if they used the drug for the majority of their treatment period. This contrasts with a "strict incident new-user" design, which requires patients to have no prior exposure to the drug. They argue this approach is more feasible in the context of Chinese healthcare, where GLP-1RAs are often add-on therapies, and data on early treatment initiation is limited. They also address the absence of a lag period by presenting data that suggests detection bias might not be the primary driver of their observed associations for thyroid cancer, and that a lag period would have reduced statistical power given the limited follow-up and event numbers. They instead used duration-stratified analyses to explore temporal patterns.

Datasets

The study utilized real-world, population-based data from China, specifically from Shenzhen's health information systems. The authors note that prescription data prior to 2016 was unavailable, which influenced their methodological choices.

Key Insights and Ideas

  • Real-world data demands pragmatic methods: The authors emphasize that observational studies in complex healthcare systems like China's must adapt methodologies to fit data availability and actual treatment patterns, even if it deviates from ideal designs.
  • Predominant-user design as a necessary compromise: While acknowledging the potential for prevalent-user bias, they argue that a predominant-user design better captures the reality of add-on therapy and treatment interruptions, which are common in unsupervised settings.
  • Lag periods are not always appropriate: For a broad spectrum of adverse outcomes with varying latencies, a universal lag period might not be optimal. In their specific case, the data on thyroid cancer suggested detection bias was not a major confounder, and applying a lag would have sacrificed statistical power.
  • Ongoing methodological refinement: The authors are actively working to improve their analyses by collaborating with broader health information databases and plan to incorporate more sophisticated exposure definitions and lagged designs in future work.

Key Results

The authors re-examine their data on thyroid cancer incidence. They found that only 11.5% of new thyroid cancer cases in the GLP-1RA group occurred within 12 months of follow-up, compared to 33.3% in the metformin group. This finding is presented as evidence against detection bias being the primary explanation for any observed association.

Limitations and Future Work

The authors acknowledge that the absence of a washout requirement could introduce prevalent-user bias. They also note that their study had relatively short follow-up and a limited number of cancer events, which constrained their ability to apply lag periods without compromising statistical power. Future work will involve:

  • Collaborating with municipal and provincial health information databases in China.
  • Incorporating additional methodological refinements, including alternative exposure definitions and lagged designs.
  • Further research specifically focused on the initiation of GLP-1RA therapy.

Summary updated

Abstract

We thank the authors of the letter for their thoughtful comments on our work [2] and appreciate the opportunity to clarify our methodological choices. Our real-world, population-based study suggests an association between GLP-1 receptor agonist (GLP-1RA) use and increased risks of adverse outcomes like pancreatitis, kidney issues, and thyroid disorders [3]. Regarding exposure definition, the correspondents correctly note that we adopted a predominant-user rather than a strict incident new-user design. New-user designs are preferred for single-outcome causal inference, particularly for cancer. However, its feasibility is constrained by real-world treatment patterns and data availability in multi-outcome safety evaluations. In China, GLP-1RAs are typically prescribed as add-on third-line (2010–2017) or second-line (2017–2024) treatment for type 2 diabetes. Requiring patients to be free of all antidiabetic medications at cohort entry would exclude a large proportion of GLP-1RA users and compromise representativeness. In addition, prescription data before 2016 were unavailable in Shenzhen health information systems, limiting reliable identification of true incident users. Under these conditions at the time of publication, a strict new-user design was not practically feasible, although such designs may become increasingly viable as data accrual and follow-up continue. Notably, several large observational studies of GLP-1RA safety similarly did not apply a pre-baseline washout period, reflecting comparable real-world considerations [4, 1]. As real-world experience with GLP-1RAs accumulates, further research specifically focused on the initiation of therapy remains warranted. We acknowledge that the absence of a washout requirement to exclude pre-existing or early-diagnosed cancers could indeed introduce prevalent-user bias, which is a valid methodological concern. However, in unsupervised real-world settings, treatment interruption and re-initiation are common and are largely driven by tolerance and adherence. Thus, imposing a washout period under these conditions may introduce additional selection bias and further reduce statistical power, particularly for uncommon outcomes. As data coverage expands with longer follow-up, alternative exposure definitions merit further evaluation. A lag period appropriately addresses detection bias, which is especially relevant for cancer outcomes where early diagnosis might coincide with treatment initiation rather than being caused by the treatment itself. However, our study assessed a broad spectrum of 20 adverse outcomes, with latencies ranging from short-term (≤ 12 months) to long-term (> 12 months), as part of a comprehensive safety evaluation. We rechecked our data on incident cases of thyroid cancer, and only 11.5% (3/26) of incidents in the GLP-1RA group occurred under 12 months of follow-up, compared with 33.3% (8/24) in the metformin group. Thus, these data may argue against this bias being a primary driver of the observed association. Moreover, given the relatively short follow-up and limited number of cancer events, introducing a lag period would have further compromised statistical power. We therefore used duration-stratified analyses to explore temporal patterns while preserving sensitivity to early-onset adverse effects. We are collaborating with municipal and provincial health information databases in China. Future analyses will be able to incorporate additional methodological refinements, including alternative exposure definitions and lagged designs as suggested, to further enhance the robustness of the evidence. Design of the work and analysis: Zhiyuan Cheng. Draft and review: Zhiyuan Cheng, Jiesheng Lin, Dongfeng Gu, and Fengchao Liang. Final approvement: Dongfeng Gu and Fengchao Liang. The authors have nothing to report. The authors declare no conflicts of interest. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
BibTeX
@article{cheng2026reply,
  author = {Zhiyuan Cheng and Haifang Li and Dongfeng Gu and Fengchao Liang},
  title = {{Reply: Time‐Related Biases in Observational Studies of <scp>GLP</scp> ‐1 Agonists and Cancer Risk}},
  journal = {Journal of Diabetes},
  year = {2026},
  doi = {10.1111/1753-0407.70203},
}
Figure