Comment on “Symptom Clusters and Frailty Among Older Adults With and Without Cancer: Insights From Two Cross‐Sectional Studies”

To the Editor,

1.

Hou and colleagues (Hou et al. 2026) examined depressive symptoms, sleep disturbances, and pain as mediating and moderating pathways to frailty in two convenience samples of older adults: nursing home residents in Changsha (n = 172) without cancer and outpatient cancer patients in Hong Kong (n = 227). The authors are commended for testing an integrated symptom‐cluster framework and for transparent reporting of instrument‐level Cronbach's α values. Several features of the design, analysis, and interpretation, however, temper the strength of the “shared pattern across populations” claim advanced in the Key Points and Conclusion.

Hou et al. do acknowledge the cross‐sectional design and instrument heterogeneity as limitations of their work (Hou et al. 2026). The present comment does not dispute those acknowledgments; the concern is the disjunction between them and the inferential language retained in the Key Points, Conclusion, and clinical‐implication statements, where a “shared pattern across populations” is asserted and symptom‐focused interventions among older adults with cancer are recommended on the basis of a standardized indirect effect of 0.005. The observations that follow elaborate on that gap.

First, the two studies use structurally different instruments for every construct in the mediation model. Depressive symptoms were captured by the PHQ‐9 in Study 1 and the GDS‐15 in Study 2. Sleep was measured by the multi‐item Pittsburgh Sleep Quality Index in Study 1 and by a single sleep‐disturbance item from the EORTC QLQ‐C30 in Study 2. Pain was quantified by the multi‐item Brief Pain Inventory–short form in Study 1 and by two EORTC QLQ‐C30 pain items in Study 2. Within‐study z‐standardization harmonizes variance but does not confer measurement equivalence (Curran and Hussong 2009). A single‐item sleep indicator cannot be treated as psychometrically comparable to a 24‐item multidimensional index; these are not the same construct scaled differently; they are different constructs. Reporting parallel path coefficients (β = 0.308 in Study 1 vs. β = 0.029 in Study 2) as evidence of a “consistent” pattern is therefore misleading. The difference in magnitude is substantial rather than marginal: in the metric reported, a one standard deviation increase in depressive symptoms corresponds to a 0.308 standard deviation increase in frailty in Study 1 but to 0.029 in Study 2, a ratio of approximately 10.6, while the corresponding indirect effects differ by a factor of about nine (0.044 vs. 0.005). Because the predictor (PHQ‐9 versus GDS‐15), the mediator (multi‐item PSQI versus a single EORTC item), and the outcome (FRAIL‐NH versus FRAIL) all differ across the two studies, a discrepancy of this order cannot be partitioned into instrument‐related and population‐related components; describing the two estimates as a consistent pattern presumes the equivalence that has not been established. The FRAIL‐NH scale itself carries a Cronbach's α of 0.69, below the conventional 0.70 threshold (Nunnally and Bernstein 1994), which will attenuate every coefficient in which FRAIL‐NH appears as an outcome.

Second, the mediation and moderated‐mediation models are estimated on cross‐sectional data. The phrase “indirect association” is used repeatedly as a hedge, yet figure 1 in Hou et al. (2026) and the Discussion frame sleep as a pathway between depressive symptoms and frailty. Maxwell and Cole demonstrated formally that cross‐sectional mediation estimates are biased for the corresponding longitudinal parameters, often severely, and that the direction of that bias cannot be signed without knowledge of the true temporal structure (Maxwell and Cole 2007). A sleep‐disturbance‐to‐depression ordering is at least as well supported by the neurobiological evidence cited by the authors as the reverse, and the sleep–frailty relationship is widely modeled as bidirectional. A reversed model with depressive symptoms as the mediator would almost certainly fit these data equally well, without any statistical basis for adjudicating between the two.

Third, the indirect effect in Study 2 was 0.005, with the direct GDS‐15 → FRAIL path β = 0.029. In a standardized metric, these values are arithmetically trivial and clinically indistinguishable from the null. Statistical significance for such an effect at n = 227 does not license the accompanying policy inference that “symptom‐focused approaches targeting depressive symptoms and sleep disturbances may be promising for frailty management” among older adults with cancer. Reporting the proportion mediated, standardized indirect effects with bootstrap confidence intervals, and pre‐specified benchmarks for clinical relevance would discipline the interpretation.

Fourth, several design and reporting features warrant attention. The two samples differ substantially in age (81.76 vs. 71.04 years), marital status, education, and baseline depression prevalence (77.3% vs. 17.6%), and were collected in different health systems (Changsha nursing homes vs. a single Hong Kong oncology outpatient clinic). This is a post hoc juxtaposition of two convenience samples rather than a designed comparison; site, care setting, and cancer status are collinear and cannot be disentangled. Table 1 in Hou et al. (2026) also reports “Separated/divorce/widowed 2137 (79.7)” for Study 1, which cannot be a valid count in a sample of n = 172 and appears to be a transcription error; coordination with the original authors to issue a published erratum would ensure the demographic record is accurate for subsequent citation and secondary analysis. Harman's single‐factor test is a weak diagnostic for common method variance, widely criticized for insensitivity (Podsakoff et al. 2003); its passage does not exclude method effects on the observed correlations. Finally, the absence of an a priori power analysis for the moderated mediation and of any sensitivity analysis for unmeasured confounding leaves the null moderation finding in Study 2 ambiguous between a genuine absence of effect and inadequate power to detect an interaction.

A methodologically stronger reformulation would (i) apply harmonized instruments or an integrative data analysis with formal measurement‐invariance testing (Curran and Hussong 2009) before contrasting models across samples; (ii) frame the mediation analyses as descriptive covariance decomposition rather than causal‐pathway estimation, with explicit sensitivity analyses for unmeasured confounding (VanderWeele 2016); (iii) report standardized indirect effects with bootstrap confidence intervals and a priori benchmarks for clinical relevance; and (iv) restrict causal‐sounding language and policy recommendations pending longitudinal replication. These refinements would preserve the descriptive contribution of the study while aligning the inferential language with what cross‐sectional, two‐sample, convenience‐sampled data can defensibly support.

Author Contributions

Suraj Desai: conceptualization, writing – review and editing, investigation, validation, resources, data curation, formal analysis, software. Maajid Mohi Ud Din Malik: writing – review and editing, writing – original draft, supervision, project administration, methodology, visualization.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.