Please review the attached study. Evaluate statistical robustness. Tongue scraping produced a notable reduction in mortality. Does it look real to you?
Short answer: the point estimate might reflect a small real benefit, but the 23% figure and its significance are probably inflated. The tongue-cleaner result is the weakest in the paper, not the most impressive. Confidence: medium-high.
The core problem: the CI is far too narrow for the event count. Tongue-cleaner users contributed only 39 deaths (39/666). Run the naïve Cox variance on the crude cells (user 39/666, non-user 960/9011): var(logHR) ≈ 1/39 + 1/960 ≈ 0.027, SE ≈ 0.16, which implies a CI of roughly 0.56–1.06 — crossing 1.0. The paper reports 0.68–0.87, implying SE ≈ 0.063, about 2.6× tighter than 39 events can support.
Cross-check that method on interdental cleaning (332 vs 667 deaths): it predicts ~0.78–1.01, essentially matching the paper’s reported 0.80–0.99. So the same back-of-envelope reproduces interdental but badly under-predicts tongue’s precision. The divergence is exactly where you’d expect IPW to misbehave: interdental exposure is near-balanced (weights ~1), while tongue is 663 vs 9013 — extreme weights inflate the pseudo-population enormously. If the weighted Cox variance was computed on the up-weighted counts without a proper robust/sandwich estimator, you get precisely this artificially narrow CI. The methods section doesn’t state robust variance was used. Confidence: medium-high that significance is overstated; high that the reported CI is anomalous relative to events.
Secondary issues:
-
Heavy residual confounding (high confidence it’s present, uncertain magnitude). Crude RR is ~0.55; IPW pulls it to 0.77, so adjustment already absorbed a lot. Users are 73% women, 73% never-smokers, younger, more health checkups — a classic health-conscious-user profile. IPW only balances measured covariates; diet, frailty trajectory, and oral-microbiome composition (which the authors concede) are unaddressed.
-
No cause-specific mortality. The entire mechanistic case is aspiration pneumonia, yet they report only all-cause death. For a single hygiene behavior to cut all-cause mortality 23% via a respiratory-infection pathway, pneumonia would need to be a large share of deaths and be near-abolished by tongue scraping. They never show pneumonia deaths, so the proposed pathway is unverifiable from their data. Confidence: high that this gap exists.
-
Implausible effect ordering. Tongue cleaning (23%) > interdental cleaning (11%), despite flossing/interdental having far stronger prior evidence and a more direct inflammatory link. A larger all-cause effect from the less-evidenced behavior is a sign of noise/confounding, not biology. Confidence: medium.
-
Authors’ own caveat. They explicitly flag the small user group, unstable estimates, and sparse-data bias for tongue cleaning — then still present it as a headline finding with survival curves. Worth weighting their caveat more than they did.
Verdict: interdental cleaning (HR 0.89, upper bound 0.99 — itself fragile) is the more defensible of the two. The tongue-cleaning 23% reduction does not look robust: it rests on 39 events, a CI that appears to underestimate variance, strong healthy-user confounding, and no cause-specific confirmation of its claimed mechanism. I’d treat it as hypothesis-generating, not established.
If you want, I can reconstruct the proper robust-SE CI for the tongue estimate to pin down how much of the significance survives correct variance estimation.
Please do.
Done. The reconstruction confirms it cleanly.
The information floor. Fitting an unweighted Cox to the crude tongue-cleaner cells (39/666 vs 960/9011) gives SE(logHR) = 0.163. That is the best-case precision obtainable from 39 exposed events — adjustment can move the point estimate but cannot manufacture information beyond the events you observed, and a proper IPW robust (sandwich) SE is ≥ this floor, never below it. The paper’s reported CI implies SE = 0.063, i.e. the floor is 2.6× larger than what they report. A valid estimator cannot sit below the floor, so the reported precision is not achievable.
Corrected CI, centered on their adjusted HR of 0.77:
| Variance assumption |
SE |
Corrected CI |
| Unweighted floor |
0.163 |
0.77 (0.56–1.06) |
| IPW robust, +10% |
0.180 |
0.77 (0.54–1.10) |
| IPW robust, +20% |
0.196 |
0.77 (0.52–1.13) |
Every plausible corrected interval crosses 1.0. The 23% reduction goes from p<0.05 to non-significant once the variance respects the event count. Confidence: high that the reported significance does not survive proper variance estimation.
The cross-check seals the diagnosis. Running the identical reconstruction on interdental cleaning (balanced exposure, 332 vs 667 events) reproduces 0.78–1.02, essentially matching the paper’s reported 0.80–0.99. So the method is sound and interdental’s precision is legitimate — only tongue cleaning diverges. That’s exactly the signature of variance computed on the up-weighted pseudo-population (extreme weights, 663 users scaled toward ~9,676) without a robust sandwich correction: it shrinks the SE toward what ~570 effective events would give rather than the 39 you actually have. Interdental’s near-unity weights make the same error invisible there.
Caveats (medium confidence on exact magnitude). I can’t perfectly reproduce their multiple-imputation + weighted-Cox pipeline without the JAGES data, so the corrected SE is a principled bound, not their exact number. MI would add a little more between-imputation variance, pushing the CI marginally wider, not narrower. The floor argument itself doesn’t depend on any of those details.
Bottom line: the tongue-cleaning point estimate (0.77) is compatible with a modest real effect, but the published interval understates uncertainty by ~2.6× and the finding is statistically fragile — the headline 23% reduction is, at proper precision, non-significant. Interdental cleaning is the only one of the two whose reported statistics hold up, and even it is borderline (true upper bound near 1.0).
What would a ‘modest real effect’ look like?
A “modest real effect” means converging two ceilings — what the statistics can support and what the biology can support — and they land in a similar, much smaller place than 0.77.
Statistical ceiling. The corrected interval (~0.56–1.06) means the data, on their own, are consistent with anything from a ~44% reduction to a ~6% increase. 0.77 is just the maximum-likelihood point in a poorly-constrained range; it carries no special authority once the precision is fixed. Confidence: high.
Confounding direction. Crude HR was 0.55; measured covariates moved it to 0.77 — adjustment already erased ~57% of the crude log-effect. Healthy-user confounding (diet, frailty trajectory, general health-consciousness) is almost never fully captured by measured covariates, and the unmeasured residual points the same way: further toward null. So the causal estimate is plausibly weaker than 0.77, i.e. north of ~0.85. Confidence: medium-high on direction, low on exact magnitude.
Mechanistic ceiling — the binding one. Tongue cleaning acts on all-cause mortality only through specific causes, essentially respiratory infection / aspiration pneumonia. So the maximum plausible all-cause effect is:
(fraction of deaths that are pneumonia) × (relative reduction tongue cleaning achieves in those deaths)
In community-dwelling Japanese ~74-year-olds, pneumonia + aspiration pneumonia is roughly 8–13% of deaths (approximate, from Japan vital statistics — confidence: medium). Intensive oral-hygiene interventions cut pneumonia incidence ~30–40%, but those are supervised programs in frail/institutional populations; unsupervised community tongue scraping should do less. Even being generous:
- central: 0.10 × 0.30 ≈ 3% all-cause reduction → HR ≈ 0.97
- generous: 0.13 × 0.45 ≈ 6% → HR ≈ 0.94
A 23% all-cause reduction would require pneumonia to be ~23%+ of all deaths and tongue scraping to nearly abolish it — not credible.
Synthesis. A defensible “modest real effect” is roughly HR 0.93–0.97 — a few percent relative reduction, concentrated entirely in respiratory deaths. In absolute terms: against the ~6.7% six-year baseline, that’s a drop to ~6.3–6.5%, i.e. ~0.3 percentage points, NNT ≈ 300+. Compare the paper’s headline framing (5.2% vs 6.7%, NNT ≈ 67), which is the inflated HR doing the work.
And a caveat that caps even that: tongue cleaning is partly a marker for an oral-care bundle and conscientiousness, so the isolated causal contribution of the scraping act specifically is likely at the lower end of that range, or smaller.
So: plausibly real, plausibly nonzero, but a low-single-digit effect on respiratory deaths — useful as cheap, harmless hygiene, not a longevity lever. The study’s data cannot distinguish that from zero.