The distinction between the two analytical frameworks used in this study deserves closer attention, because it speaks directly to how real-world drug effects are estimated. In the as-treated design, patients remain classified as GLP-1 receptor agonist or DPP-4 inhibitor users only for as long as they actually continue the initial therapy, with follow-up ending at discontinuation or switching. This approach concentrates the analysis on periods of genuine pharmacological exposure, which is where any biological effect on cognition or survival would plausibly operate. The intention-to-treat analogue, by contrast, preserves the original treatment assignment regardless of what patients subsequently take, a strategy that mirrors the conservative logic of randomized trials but that, in an observational setting, can dilute differences between groups if many patients change therapy. The fact that the investigators ran both designs, and used marginal structural models to account for time-varying characteristics such as changes in comorbidity or concomitant medication, reflects a methodological awareness that discontinuation is rarely random in elderly diabetic populations. Patients who stop a drug often do so because of side effects, clinical deterioration, or physician judgment, and if those reasons are themselves related to cognitive or mortality risk, naive comparisons would be biased. Marginal structural models, particularly when combined with inverse probability of censoring weights, are among the standard tools developed precisely to handle this kind of informative censoring in pharmacoepidemiology.

The competing risks framing is equally consequential and often underappreciated outside biostatistics. In a population aged 65 and older with type 2 diabetes, death is a frequent event, and when a patient dies before developing a diagnosable cognitive disorder, that cognitive outcome can never be observed. Standard survival analysis, which simply censors individuals at death, implicitly treats those patients as if they could still experience the cognitive event later, an assumption that is clearly counterfactual and that tends to distort incidence estimates in older cohorts. The fine-gray sub-distribution hazard framework used here addresses this by keeping individuals who die in the risk set while accounting for the fact that their competing event removes them from ever contributing a cognitive outcome. This is why the authors report separate sub-distribution hazard ratios for cognitive disorders and for death without prior cognitive disorders, rather than a single cause-specific estimate. The practical consequence is that the composite outcome, while clinically meaningful, blends two biologically distinct processes: a neurodegenerative or vascular cognitive pathway and a mortality pathway influenced by cardiovascular disease, frailty, cancer, and other causes. A drug could plausibly influence both, and the composite design cannot fully disentangle which component drives an observed benefit.

The mechanistic rationale for a differential effect between the two drug classes rests on the pharmacology of incretin signaling. GLP-1 receptor agonists directly activate the GLP-1 receptor, producing sustained receptor stimulation at pharmacological concentrations far above physiological levels, whereas DPP-4 inhibitors work indirectly by preventing the enzymatic degradation of endogenous GLP-1 and related peptides. Endogenous incretin secretion rises only modestly after meals, so the increment in GLP-1 signaling achieved by DPP-4 inhibition is inherently smaller than what direct agonism delivers. If neuroprotection depends on receptor activation thresholds, this pharmacological difference could translate into different central nervous system effects. Preclinical work has described anti-inflammatory actions of GLP-1 receptor signaling in glial cells, effects on neuronal insulin signaling, and modulation of oxidative stress pathways, all of which are relevant to the chronic neuroinflammatory processes implicated in cognitive decline. Whether these laboratory findings translate to clinically meaningful protection in humans remains an open question, and the present study’s cognitive-specific result, with a sub-distribution hazard ratio point estimate near 0.4 but confidence intervals crossing unity, is exactly the kind of signal that demands larger samples before interpretation.

The small number of cognitive events is the central statistical limitation and warrants explanation for readers unfamiliar with power in competing risks settings. With only 393 patients in the GLP-1 receptor agonist arm and a five-year maximum follow-up, the expected number of incident cognitive disorder diagnoses is limited even if the underlying risk is substantial, because administrative data capture only coded diagnoses rather than structured cognitive testing. Mild cognitive impairment in particular is frequently undocumented in hospital discharge and prescription records, meaning the outcome definition likely captures more advanced or clinically salient disease. This measurement constraint cuts both ways: it reduces the number of events available for estimation, widening confidence intervals, and it may misclassify the timing and severity of cognitive decline. The authors’ decision to define the primary outcome as a composite with death was partly a pragmatic response to this sparsity, since death is well captured in administrative registries and adds statistical information, but it also means the headline finding is driven substantially by the mortality component.

The early-follow-up pattern in the results is noteworthy and biologically interesting. The hazard ratio of approximately 0.5 at two years suggests that any advantage of GLP-1 receptor agonists manifests relatively soon after initiation, which is a timeframe more consistent with metabolic and vascular effects, or with differences in acute tolerability and treatment persistence, than with slowed neurodegeneration, a process typically unfolding over years to decades. Early mortality differences could reflect weight loss, improved glycemic control, effects on cardiovascular risk factors, or even differences in the patient populations for whom each class is prescribed. Confounding by indication remains a persistent concern in comparisons of these two classes, since GLP-1 receptor agonists are often preferentially given to patients with obesity or higher cardiovascular risk, characteristics that could influence cognitive trajectories in either direction. Although the analysis adjusted for demographic and clinical characteristics and used weighting approaches, residual and unmeasured confounding, including lifestyle factors, disease duration, and baseline glycemic control, cannot be excluded in any administrative database study.

The regional data source also shapes interpretation. The Puglia healthcare utilization databases provide near-complete population coverage, which minimizes selection bias related to insurance or access, a strength relative to claims data in systems with fragmented coverage. The universal Codice Fiscale identifier enables linkage across prescriptions, hospitalizations, and mortality records over long periods. However, the study population was drawn from a single province, and prescribing patterns, population demographics, and healthcare practices there may not generalize to other regions or countries. Additionally, the study period spanning 2010 to 2018 predates the widespread adoption of newer GLP-1 receptor agonists with greater potency and cardiovascular outcome evidence, so the findings primarily reflect earlier agents in the class. Extrapolating to contemporary practice, where longer-acting agonists are now dominant, requires caution.

For clinicians, the practical takeaway is appropriately modest. The study does not establish that GLP-1 receptor agonists prevent dementia, and the authors are explicit that the cognitive-specific association, while directionally suggestive, rests on too few events for firm conclusions. What the study does establish is a feasible analytical template: competing risks methods and marginal structural models applied to real-world administrative data can generate hypothesis-generating evidence in frail elderly populations where randomized trials of cognitive outcomes are difficult to conduct. Given the growing number of trials now evaluating GLP-1 receptor agonists for cognitive and neurodegenerative indications, observational findings like these contribute to the triangulation of evidence, but they cannot substitute for prospective testing with validated cognitive instruments, longer follow-up, and adequate event counts. Until such evidence accumulates, treatment selection for older adults with type 2 diabetes should continue to rest on established metabolic and cardiovascular indications, with cognitive outcomes as a promising but unproven consideration.

Subject of Research: …

Article Title: Exploring the impact of GLP-1 analogues and DPP-4 inhibitors on cognitive decline in elderly patients with type 2 diabetes: a competing risks analysis from a real-world cohort

Article References: Exploring the impact of GLP-1 analogues and DPP-4 inhibitors on cognitive decline in elderly patients with type 2 diabetes: a competing risks analysis from a real-world cohort. (n.d.). https://doi.org/10.3389/fphar.2026.1822158

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1822158

Keywords: Exploring, impact, GLP-1, analogues, DPP-4, inhibitors, cognitive, decline, elderly, patients, type, diabetes