Near-Zero Clustering Raises Doubts Over Bayesian Model in Ghana Stroke Study

A new methodological commentary published in Aging Clinical and Experimental Research is challenging the statistical machinery behind a recent study of stroke burden in Ghana, arguing that several of its central analytical choices may have produced findings that are less reliable than they appear. The commentary, written as a letter to the editor by Aria Hosseini, Rayeheh Bozorgi, Amir Molavi zad, Kosim Rakhmanov, Guloyim Avezova, Shukhrat Teshayev and colleagues, takes aim at five specific issues: the justification for a Bayesian multilevel model, the reliance on self-reported stroke, the interpretation of education as a risk factor, the absence of a sensitivity analysis for prior distributions, and the possibility of selection bias in a finding about physical activity.

The most pointed criticism concerns the Bayesian multilevel logistic regression that the original authors employed to address potential cluster-level heterogeneity in their data. The commentary notes that the original study’s own estimated intraclass correlation coefficient, or ICC, was essentially zero, at 0.0001 with a 95 percent credible interval of 0.0000 to 0.0005. That figure indicates that virtually all variability in stroke risk in the dataset resided at the individual level rather than at the level of any clustering structure the model was designed to capture.

The technical implication is significant. Simulation research cited in the commentary has shown that the potential for error from ignoring clustering is greatest when the ICC exceeds 0.01, a threshold that the estimated value in the Ghana study falls short of by a factor of 100. In other words, there is little evidence that failing to account for clustering would have meaningfully distorted the study’s inferences. While Bayesian multilevel models can improve predictive ability and provide more precise probability intervals when genuine clustering exists, the commentary stresses that those advantages depend on the presence of meaningful group-level variation. Without such variation, the added complexity may simply be unnecessary.

The commentary also points out that the original authors themselves acknowledged that both their Bayesian multilevel model and a simpler alternative produced largely consistent findings. This admission, the letter argues, raises an uncomfortable question: did the Bayesian model genuinely deliver more stable and precise estimates, or was it a computationally expensive confirmation of what a simpler, more transparent model could have shown? The letter frames the issue not as an accusation of error but as a question of analytical proportionality, asking whether the sophistication of the method matched the structure of the data it was applied to.

A second major concern involves the study’s primary outcome. Participants in the original research were simply asked whether they had ever been told they had a stroke, a self-reported measure that the commentary describes as carrying a well-documented risk of misclassification, particularly in settings where diagnostic confirmation is limited. A systematic review examining the accuracy of patient self-report for stroke found that in population-based studies with low stroke prevalence, such as the 2.37 percent observed in the Ghana study, a substantial portion of self-reported cases may be false positives. This raises the possibility that some of the 120 reported strokes in the sample could be inaccurate, potentially biasing both the estimated prevalence and the estimated associations with risk factors.

Supporting evidence comes from a validation study conducted within the Korean Health Examinees Study, which evaluated self-reported stroke against medical records. There, the positive predictive value of self-report was only 51.4 percent, meaning that nearly half of the participants who reported a stroke could not be confirmed as definite cases upon record review. The commentary argues that such substantial false-positive reporting weakens the reliability of the outcome measure and, consequently, the validity of the predictors identified. Without clinical validation, it cautions, the findings may represent associations with the report of a stroke rather than with a biologically confirmed stroke event. The letter recommends that future studies supplement self-reports with medical record verification or diagnostic data to improve outcome accuracy.

The third issue concerns an unexpected result: a positive association between education and stroke, with an adjusted odds ratio of 1.48. The commentary notes that this runs counter to the typical pattern in high-income settings, where higher education is usually protective. The original authors attributed the finding to better health awareness and diagnostic access among educated individuals, but the letter identifies a conceptual tension in that explanation and proposes a more plausible alternative: diagnostic bias. Educated individuals, it argues, may have greater opportunities to seek care, receive a formal diagnosis, and recall it when surveyed, while in resource-limited settings stroke often remains undiagnosed among those with limited healthcare access. On this reading, the positive education-stroke association may reflect differential case capture across socioeconomic groups rather than a causal effect, a distinction the commentary describes as essential for both epidemiological inference and health policy.

The fourth criticism targets a missing analysis. The original study employed weakly informative Gaussian priors with a mean of zero and a variance of 1000 for fixed effects, along with log-gamma hyperpriors for the random intercept variance, yet provided no assessment of how these choices influenced the results. The commentary argues that this omission matters because prior specification can substantially affect Bayesian estimates, particularly in models with rare outcomes or limited cluster-level information. It cites a systematic survey finding that sensitivity analyses for priors are frequently underused, with only 27.0 percent of Bayesian trials conducting them, and methodological research showing that interval estimates can be more sensitive to prior choice than point estimates. This, the letter contends, directly challenges the original authors’ claim of more precise estimates from the Bayesian model: without sensitivity analyses, the precision of the estimates cannot be reliably attributed to the data rather than to analytical choices.

The final point addresses the observed protective effect of vigorous physical activity against stroke. The commentary acknowledges that this finding is consistent with a large US population-based study that reported a similar 40 percent reduction in stroke odds, with an odds ratio of 0.6 and a 95 percent confidence interval of 0.5 to 0.8. But it poses a key question: does the association reflect a genuine biological benefit, or is it partly driven by selection bias? The letter invokes the healthy worker survivor effect, the phenomenon by which individuals who remain in physically demanding jobs or activities tend to be healthier than those who leave them due to illness. This selection process can attenuate or even reverse estimated effects in occupational and lifestyle studies when not properly addressed, and the commentary notes that cross-sectional analyses cannot fully resolve the issue.

Nevertheless, the letter strikes a constructive tone on this point, observing that the consistency of the protective association across multiple studies, including the present one, suggests that physical activity may confer a genuine benefit. It recommends that future research employ objective physical activity measures or longitudinal designs to help distinguish a biological protective effect from selection bias. In its conclusion, the commentary reiterates that while the original study provides valuable insights into stroke burden in Ghana, the near-zero ICC questions the necessity of the Bayesian multilevel approach, reliance on self-reported stroke introduces potential misclassification bias, the education-stroke association may reflect diagnostic bias rather than true risk, the absence of prior sensitivity analysis undermines claims of model precision, and the protective effect of physical activity cannot be fully disentangled from selection bias without objective measures. Addressing these considerations, the authors write, would strengthen future stroke research in resource-limited settings.

Subject of Research: Methodological limitations of a study on stroke burden and risk predictors in Ghana

Article Title: Stroke burden and associated predictors: A methodological reflection

Article References: Hosseini, A., Bozorgi, R., Molavi zad, A., Rakhmanov, K., Avezova, G., & Teshayev, S. (2026). Stroke burden and associated predictors: A methodological reflection. Aging Clinical and Experimental Research, 38(1), Article 211. https://doi.org/10.1007/s40520-026-03497-y

Image Credits: AI Generated

DOI: 10.1007/s40520-026-03497-y

Keywords: stroke, Ghana, Bayesian multilevel model, intraclass correlation, self-reported diagnosis, misclassification bias, diagnostic bias, education, physical activity, selection bias, sensitivity analysis, prior distributions