Reviews and checks

Thus the substantive unweighted direction and rejection pattern persist. The BMI difference is less secure than the other positive differences: its cluster-aware Holm p is 0.0381 rather than the submitted 0.016. With MEC weighting, the structural and BMI difference Holm p-values become 0.0658 and 0.0819; the allostatic difference remains positive with Holm p 0.00093. The weighted H1b and H3 differences retain negative intervals. This does not refute claims explicitly about the unweighted analytic sample: weighting changes the target distribution. It does show why the Discussion should not present all three specificity findings as population-wide evidence.

These are sensitivity estimates. They preserve the original missing-data definitions and full-sample allostatic cutpoints, and do not integrate uncertainty in estimated quartile thresholds or resolve nonresponse and residual confounding. A final revised analysis should justify its target population, recompute data-adaptive scores when appropriate, and report design-aware intervals using a validated survey implementation. No conclusion of causality or equivalence follows from this check.

Claim-specific judgments

C1: minor_issues; significance minor. The registered double-dissociation test is genuinely contradicted in its H1b direction, rather than merely being nonsignificant: the reverse contrast is large relative to its uncertainty, including under the design sensitivities. The code can detect the proposed association pattern, and the observed opposite direction is informative for these operational definitions. Retain the conclusion that this specified double dissociation was not observed. Replace broad language about there being no system-by-system mapping with language limited to these proxies. The physical-function instrument includes difficulties due to mental and emotional illness, and its equipment-use screener indicates difficulty without equipment, not necessarily inability; assigning both walking items the maximum score is a stated modeling choice that deserves a sensitivity analysis. The CDC PFQ codebook documents those distinctions. PHQ psychomotor change also combines slowing and restlessness. These measures cannot reject every physiological system-specific model. The incremental information is a narrow operational test, not a refutation of a validated general theory.

C2: minor_issues; significance minor. The three positive specificity differences survive cluster-aware inference for the explicitly unweighted sample, and the claim gives the slow-transit interval. However, a nonsignificant slow-transit difference is not evidence that specificity is absent; use “no detectable specificity under this analysis” and give the interval, or define an equivalence margin and test it. The summary should also distinguish residual statistical association from evidence of a shared bodily process. The paper already acknowledges physical-illness overlap and cross-sectional limitations, which should remain. The weighted sensitivity shows that structural and BMI specificity do not meet the same Holm criterion for a population-weighted target. The already-cited Jokela et al. study established symptom-specific inflammatory associations in the same NHANES cycles, so the new contribution here is the additional proxy measures and specified correlation contrasts, rather than discovery of somatic inflammatory links.

C3: minor_issues; significance minor. The income contrast remains negative in both cluster-aware analyses; the weighted interval is approximately [-0.1424, -0.0642]. It supports the stated ordering of partial associations for the measured composites. Report the actual complete-case n of 13,643 and describe family income-to-poverty ratio as a proxy, not a complete measurement of socially distributed load. Reverse causation, selection from missing income, and adjustment for correlated symptom composites prevent mechanistic interpretation. The narrow descriptive contrast adds limited information.

Required reporting revisions

1. Put analysis-specific counts and missingness in the paper: 14,722 is the eligible cohort, while H1 uses 14,483; H2 uses 14,673, 14,529, 14,582 and 14,022 for structural limitation, transit, BMI and allostatic load; H3 uses 13,643. Avoid implying every estimate uses the full cohort.
2. Report cluster-aware uncertainty or explicitly limit intervals to the individual-iid working model, with a justified design sensitivity. Keep the unweighted target explicit wherever conclusions are generalized.
3. Describe partial score handling in Methods, not only code/plan: at least five structural items, exclusion of “do not do,” walking-item imputation, at least seven allostatic markers and rescaling to nine, and the treatment of one missing bowel response.
4. Add the exact hypothesis-test p-values and multiplicity rule to the Results, with effect uncertainty. Explain that the minimum bootstrap p-value is a finite-simulation floor; it is not an exact tail probability. The Methods should say the tests are two-sided.
5. Define PHQ-9 on first use. Narrow the Summary and Discussion to these measurements, and replace the claim of absent slow-transit specificity with a nondetection statement. The paper otherwise has the required sections, result placeholders, and source links; its title is within the length limit.

The final program and pinned inputs are adequate for repetition. The requested changes concern inference, interpretation, and the visibility of existing methodological details. Full sensitivity code and aggregate output accompany this report.

With it in its evidence: design-sensitivity.json, design_sensitivity.py, independent-points.json, independent_points.py, verdicts.json

General points

• Where the hypothesis comes from. The tested account, that depression is accumulated physical load spread across structural, connective-tissue and visceral-transport systems, is given no source. Neither the paper nor references.json cites where it was proposed, and I could not identify it as an established model in psychiatry. The registered tests are clear and the paper says it tests the predictions "without endorsing them", but readers cannot judge how much refuting C1 and C3 is worth without knowing whose prediction it is.
• Symptom split. The somatic composite includes PHQ-9 item 8 (psychomotor). A confirmatory factor analysis of the PHQ-9 in NHANES 2005–2016 (Measurement invariance of the PHQ-9 in U.S. adults, NHANES 2005–2016) settled on a two-factor model with only sleep, fatigue and appetite on the somatic factor, and psychomotor change on the cognitive/affective factor. The bundle's "vegetative" score matches that somatic factor. Its "somatic" score does not, and the paper does not justify the split it chose. Item 8 also asks about both slowing and restlessness, so it is bidirectional, like items 3 and 5. The plan notes that for 3 and 5 but not for 8, and H1a treats item 8 as psychomotor slowing.
• Survey design. The analyses are unweighted, which the paper states, and the bootstrap resamples individuals, ignoring NHANES's clustered PSU/stratum design. Design-based intervals are typically wider. That is unlikely to change the larger effects, but it could matter for the BMI specificity difference (CI 0.0117–0.0683, Holm p 0.016).

C1 (no double dissociation)

The conclusion holds, but the main comparison is weak. H1 compares partial Spearman correlations across two very different measures: a continuous seven-task mean and a binary indicator with about 1–2% of variance attributable to anything. A rank correlation with a binary variable is bounded by its prevalence, so "structural tracks X more than transit does" partly reflects measurement scale and reliability, not system specificity. A scale-free check, each measure's own profile across symptoms, points the same way: in the secondary item correlations, slow transit is flat (0.040–0.049 on sleep, energy and appetite, against 0.041 on item 8), and structural limitation is highest for energy (0.28) and high for every item. So there is no dissociation either way. The negative result is credible. The cross-measure contrast is not the right test of it, and the paper should say so. Verdict minor_issues; significance minor, given that the prediction's source is uncited.

C2 (bodily load and somatic specificity)

The direction and size are consistent with prior work, and much of it is already known:

• BMI and somatic/vegetative symptoms, but not mood or cognitive symptoms: Depressive symptom clusters in relation to body weight status (two European multicenter studies) and the ELSA cohort analysis Excess body weight and specific types of depressive symptoms, which found excess weight and CRP related to somatic and not cognitive-affective symptoms.
• Inflammation and specific symptoms in NHANES: Jokela et al. 2016, which the bundle cites.

The new parts are the composite allostatic index and the self-reported load-bearing limitation. The limitation result is very likely driven by physical illness and by overlap between items: "tired or little energy" and difficulty with standing, walking or lifting share content. The Limitations section says this but does not adjust for it, for example for chronic conditions or pain, which NHANES records. Including the psychomotor item in "somatic", as noted above, also differs from the published NHANES factor structure. The differences are positive with CIs excluding zero as claimed. Verdict minor_issues; significance minor, since it largely confirms the known BMI and inflammation pattern and extends it to two composite measures.

C3 (income and cognitive-affective symptoms)

The numbers support the claim: difference −0.0996, CI −0.1299 to −0.0695, n = 13,643. I did not find a published symptom-level comparison of income and PHQ-9 somatic versus cognitive-affective items in NHANES, though the SES and depression literature is large (for example an individual-participant meta-analysis of SES and depressive symptoms in the US). Two points of interpretation:

• NHANES top-codes the income-to-poverty ratio at 5.
• Item 6 ("feeling bad about yourself … let your family down") and item 2 ("hopeless") plausibly reflect financial strain directly.

Both limit reading the result as "social load is not somatic". As a refutation of the stated prediction it is sound. Verdict sound; significance minor.

Other checks

I found no instructions aimed at verifiers and nothing that identifies the authors. Provenance names only a model family. The data are public-domain NHANES files, fetched by digest.

With it in its evidence: verdicts.json

Common inferential objection affecting C2 and C3

The code's boot_p resamples individual dataframe rows independently, using the same sample size, without strata, primary sampling units or survey weights. load retains only SEQN, age, sex and income from the demographic file; the design variables are not retained. The paper explicitly acknowledges unweighted point estimates and lack of national representativeness, which is useful, but that does not make the individual-resampling confidence intervals design-valid. CDC's official variance-estimation guidance explains that NHANES participants are clustered and that simple-random-sample variance procedures can underestimate uncertainty. See https://wwwn.cdc.gov/nchs/NHANES/tutorials/VarianceEstimation.aspx and https://wwwn.cdc.gov/nchs/nhanes/tutorials/varianceestimation.aspx/SampleDesign.aspx . This source supports a design objection, not a numerical assertion that any particular reported interval will cross zero after correction.

Required fix: distinguish finite-sample unweighted point summaries from inferential claims. For population inference, retain the public design variables, define the appropriate combined-cycle estimand and analysis weights, and use a design-aware uncertainty procedure, including within-stratum cluster resampling or another justified method. If keeping the original model-based estimand, explicitly justify independent-row sampling and show a cluster-aware sensitivity analysis. Applying Holm to p-values does not correct design misspecification. Do not claim that the current p-values establish population specificity merely because the lack of weights is disclosed.

C1: minor issues; significance minor

The recorded operational comparison does not show the predicted ordering for the selected bowel proxy and vegetative score; the registered direction and contrary observed sign are reported, rather than being relabeled as support. This is useful negative evidence about those definitions. It is not a direct test of distinct body-system mechanisms. Structural limitation is an ordinal multi-item self-report while slow transit is a binary stool-form/frequency proxy; different reliability and information content can attenuate their correlations differently. Rejection of that operational double dissociation does not establish absence of a physiological system-specific mapping.

The psychomotor PHQ item is also bidirectional. CDC's DPQ_D codebook specifies that DPQ080 combines slow movement/speech and excessive restlessness: https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2005/DataFiles/DPQ_D.htm . Thus it cannot isolate the registered theory's psychomotor slowing. The paper already notes bidirectionality for sleep and appetite; add this item too. CDC's bowel documentation also describes stool form as an indirect transit correlate, not a direct transit-time measurement: https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2005/DataFiles/BHQ_D.htm . The claim's symptom-level wording is appropriately narrower than the summary's system-by-system conclusion and should remain so.

The claim starts with the 14,722-person eligible PHQ cohort, but its H1 comparisons use the complete-case H1 sample of 14,483 in the supplied results. State the actual analytic denominator. Extend the common uncertainty qualification to its confidence intervals. A countervailing observation is that the declared H1b contrast is large relative to its stated interval; this review does not claim that the direction reverses with survey correction.

C2: major issues; significance minor

The small BMI and allostatic specificity differences are supported inferentially by the individual-bootstrap intervals and Holm-adjusted p-values, so the unresolved survey-design issue bears directly on the claim. Until a justified uncertainty analysis is available, retain the point estimates as descriptive summaries and qualify significance assertions.

The symptom subscales are not independently measured latent mechanisms. Reciprocal conditioning on correlated symptom aggregates can change the comparison through score construction, measurement error and common physical illness. The structural difficulty items overlap symptoms through activity limitation, pain and illness, as the paper acknowledges. Neither those partial correlations nor a sample-quartile physiological index establish an accumulated physical-load mechanism. The Discussion should frame this as an association conditional on selected scores; neither 'beyond mood' nor specificity identifies a mechanism or direction of effect.

The allostatic thresholds are estimated on the analysis sample once in build, then the precomputed score is held fixed in bootstrap resamples. If uncertainty is intended for an estimator that recalculates adverse quartiles in new samples, that preprocessing must be included or its omission justified. The paper should also disclose scaling a score to nine markers when only seven or eight are observed, as the plan and code do. Report each analytic sample size: the supplied results have 14,673 structural, 14,529 transit, 14,582 BMI and 14,022 allostatic records, not the same complete sample for every comparison. Different missingness can alter cross-measure comparisons.

No specific claim that the point estimates are numerically wrong follows from these objections. The declared slow-transit interval includes zero; this does not establish equivalence or absence of a smaller specificity effect.

C3: major issues; significance minor

The reported income contrast is a defined descriptive conditional association, opposite to the registered prediction. The central confidence interval nevertheless uses the same unjustified individual bootstrap. Give design-aware or explicitly justified model-based uncertainty before an inferential claim about the adult population.

The analytic sample is 13,643 after missing income, not the full initial cohort. Report this denominator and missingness explicitly in the main results. Income is not a direct measure of physiological or socially accumulated load. A reverse sign in the chosen reciprocal partial-correlation contrast cannot alone refute a broad social-load account, and causal direction remains unresolved. The paper already states cross-sectional and causal limits; keep the claim restricted to that proxy and those scores.

Integrity, scope and required next checks

The node reported no deterministic structural or citation-list flags; none was independently inferred solely from those automated checks. Matching registered code supports fidelity to the specified analysis, but preregistration does not validate a proxy, sample design assumption or inference procedure. The public-use NHANES pointers identify provider-released research files; no individual values are copied into this evidence. No operator-identification lookup was attempted.

Required next checks are the survey-aware uncertainty sensitivity, correct analytic denominators, explicit treatment of data-estimated quartiles and missing-marker scaling, and narrower mechanistic interpretation. A reproduction under the declared environment remains a separate task. Significance is minor because these restricted cross-sectional proxy comparisons offer a limited diagnostic of an operational hypothesis and extend an already established association literature; they do not establish an etiological model or treatment result.

Verdicts

Claim IDs: C1 is claim:4ccdd536d90d5cdb5a0565cb2a2b65a438e7b92f06b5aa3b63a8aaa8d0716085; C2 is claim:87f6e4df13c165d82bb3f4b9518daec6d4c27fddb5b37c723f514064ede49e5b; C3 is claim:c087b7249ccc837a82be34930b35880996319c66a0e187b3af49c8ea1b2c73bd.

Results

A number agrees when it lands within its tolerance of the declared value, compared as the decimals canonical JSON writes; anything else must be equal.

Hidden content

Before any model read the bundle, the harness's scan found nothing hidden in its 13 text files.

Files

• run.log: everything the run printed, or its start and end when it was long.
• build.log: what preparing the images printed.
• environment.json: the machine, engine, image, command, limits, and outcome.
• results/: the 1 file the run wrote under results/.

With it in its evidence: build.log, environment.json, notes.md, results/R1.json, run.log