Grip strength is the cheap, fast, no-lab proxy for muscle FUNCTION (force), as distinct from muscle mass (quantity -> Low Muscle Mass and Mortality). Celis-Morales 2018 (UK Biobank, n=502,293, 40-69 y, mean 7.1 y follow-up) is the large single-cohort evidence that lower grip predicts a wide range of adverse outcomes; Lopez-Bueno 2022 (SR + dose-response MA, 48 cohorts, 3.14M adults, 40+ countries) now backs the core association at multi-cohort scale and adds the dose-response shape. Two decision-relevant claims: (1) it is a strong, broad, cross-population mortality predictor, and (2) it works as a cheap risk METRIC — it improves an office-based risk score where no blood draw is possible. The whole evidence base is observational cohort (no trial, no MR), so grip is a predictor, not a proven causal lever — Celis-Morales frames the paper as prediction on purpose (Lopez-Bueno, by contrast, over-reads it as a training lever — see below). This is the general prognostic-marker-vs-modifiable-lever distinction (Surrogate Outcomes -> Prognostic marker vs modifiable lever): grip is partly lowered by occult ill-health, so it is a marker to stratify with, not assumed a target to train toward.
(inferred from Celis-Morales et al., 2018; López-Bueno et al., 2022)
The effect estimate — per 5 kg lower grip, fully adjusted + landmark
Hazard ratios per 5 kg lower grip strength, model 4 (adjusted for age, deprivation, ethnicity, comorbidities, height, BMI, smoking, physical activity, sedentary time, diet; events in the first 2 years excluded), women / men:
| Outcome | Women HR (95% CI) | Men HR (95% CI) |
|---|---|---|
| All-cause mortality | 1.20 (1.17-1.23) | 1.16 (1.15-1.17) |
| CVD mortality | 1.19 (1.13-1.25) | 1.22 (1.18-1.26) |
| All respiratory mortality | 1.31 (1.22-1.40) | 1.24 (1.20-1.28) |
| COPD mortality | 1.24 (1.05-1.47) | 1.19 (1.09-1.30) |
| All cancer mortality | 1.17 (1.13-1.21) | 1.10 (1.07-1.13) |
| Colorectal cancer mortality | 1.17 (1.04-1.32) | 1.18 (1.09-1.27) |
| Lung cancer mortality | 1.17 (1.07-1.27) | 1.08 (1.03-1.13) |
| Breast (W) / prostate (M) mortality | 1.24 (1.10-1.39) | 1.05 (0.96-1.15) NS |
Incidence tracked mortality (e.g. CVD incidence 1.15 W / 1.11 M per 5 kg). Prostate cancer mortality was the one null; incident prostate cancer was positively associated with strength — the authors read this as higher-strength people being more health-conscious and screening more (incidence without mortality), not a protective effect reversed.
Read the absolute risk, not just the HR [inferred from @celismorales2018]. This is a middle-aged cohort with a low event rate — 13,322 deaths / 502,293 (~2.7%) over ~7 years, so a per-5 kg HR of 1.16-1.20 is a small absolute mortality difference at this age. The relative effect is what transports; the absolute benefit of being stronger scales with baseline risk (route-(a)), so it is far larger in an older / higher-risk stratum than these middle-aged numbers show -> Baseline Risk and the Relative-Absolute Split.
Per 1 SD, grip out-ranks two classic modifiable risk factors. Grip SD is ~11 kg, so 5 kg is ~0.5 SD. Per 1 SD lower grip (adjusted): «hazard ratio 1.48 and 1.57 per 1 SD lower grip strength» for all-cause and CVD mortality — vs systolic blood pressure «1.03 and 1.26» and total physical activity «1.06 and 1.07». For incident CVD, grip and SBP were similar («1.29 and 1.30»). So on this within-model comparison grip showed «the strongest association» with mortality — a marker point, not a claim that grip causes more than BP does.
Multi-cohort upgrade — Lopez-Bueno 2022 SR + dose-response MA
Lopez-Bueno pools 48 prospective cohorts (3,135,473 adults, 49.6% women, age 35-85, >40 countries, follow-up 2.3-44 y), taking the grip-mortality association from one UK cohort to a broad multi-country base. Grip values examined ranged 15-50 kg (the studied range — the extrapolation boundary below). Two analyses: a random-forest model over tertile estimates, and a hierarchical meta-regression dose-response (restricted cubic spline, knots at the 10th/50th/90th percentile).
Random-forest tertile HRs (reference = STRONGEST third; 1st = weakest):
| Outcome | 2nd tertile HR (95% CI), I2 | 1st (weakest) tertile HR (95% CI), I2 |
|---|---|---|
| All-cause (12 studies) | 1.30 (1.17-1.44), 52.6% | 1.58 (1.40-1.78), 58.9% |
| Cancer | 1.12 (1.03-1.23), 0.0% | 1.27 (1.01-1.59), 76.0% |
| Cardiovascular | 1.25 (1.06-1.48), 63.2% | 1.51 (1.13-2.02), 87.4% |
The weakest tertile carries ~1.6x all-cause and ~1.5x CV mortality vs the strongest — a large relative contrast, consistent in direction with Celis-Morales’s per-SD figures. Cancer is attenuated (matching Celis-Morales’s weaker cancer signal). Heterogeneity is high on the cause-specific weakest-tertile arms (I2 76-87%).
This is type-F, not independent backing (type-E) — UK Biobank is inside the pool.
(Celis-Morales et al., 2018; inferred from López-Bueno et al., 2022)
Celis-Morales 2018 IS one of the 48 SR studies, and UK Biobank data enters the pooled dose-response via
Kim 2018 (70,913) and Yates 2017 (420,727). So Lopez-Bueno’s “agreement” with the held page is partly
the same data re-pooled, and the whole base is one observational-cohort lineage (no genetic/MR route) —
no confidence lift from convergence. What it DOES add is cross-population breadth (40+ countries, 48
cohorts vs one UK cohort) and an explicit dose-response shape, which is the F-value: the composite
beats either alone. No [E-independent] token.
The parameter table (are these the same quantity as Celis-Morales’s numbers?):
| Parameter | Celis-Morales 2018 | Lopez-Bueno 2022 | Same quantity? |
|---|---|---|---|
| Exposure metric | HR per 5 kg / per 1 SD lower (continuous) | tertile-contrast HR + dose-response range (kg) | No — per-unit vs categorical/curve |
| Population | UK Biobank, 40-69 | 48 cohorts, 35-85, 40+ countries | No — single vs pooled |
| Design | single observational cohort | SR-MA of 48 observational cohorts | partial — same lineage |
| Reverse-causation handling | 2-yr landmark + adjustment | none (structural inclusion only) | No — landmark vs none |
| Outcomes | all-cause/CV/cancer/resp | all-cause/CV/cancer | partial |
Different effect metric + different population + different reverse-causation handling => Lopez-Bueno quantifies-and-refines at scale, it does not independently re-derive Celis-Morales’s per-unit HR.
The dose-response is linear — no knee, no plateau over the studied range
The authors modelled the shape explicitly: «We used multivariable cubic regression splines to visually explore non-linear associations… As we found no evidence of deviation from linearity» they used Cox models. So over the observed grip range the gradient is monotone — every increment of strength maps to lower hazard, with no threshold or plateau located. This is another instance where an objectively measured capacity shows a clean monotone gradient, unlike the plateaus/U-shapes that appear in self-reported activity data -> The U-Shaped Association Artifact, Cardiorespiratory Fitness and Mortality. (Caveat: a per-outcome spline visually judged linear is weak evidence of true linearity, and the low high-strength event count limits power at the top end — absence of a located knee, not proof of none.)
Lopez-Bueno’s pooled dose-response — the shape is OUTCOME-SPECIFIC, and the thresholds carry both load-bearing facts [@lopezbueno2022]. The pooled meta-regression found «higher levels of handgrip strength significantly reduced the risk of all-cause mortality within 26-50 kg… in a close-to- linear inverse fashion. Cancer and cardiovascular mortality displayed a trend towards a U-shaped association with a significant risk reduction between 16 and 33 kg… and a close-to-linear inverse shaped and significant risk reduction ranging from 24 to 40 kg respectively.» The two facts every threshold must carry:
- These are significance-crossing RANGES, not point cutpoints with their own CIs — the numbers say over this span the pooled HR is significantly below the reference, not the optimum sits here. I2 is moderate for all-cause (45.7%) but considerable for cancer (77.4%) and CV (79.7%), so the cause-specific curves are pooled over heterogeneous studies.
- The studied range is 15-50 kg — so the all-cause upper bound (50 kg) IS the sampling edge. The authors say so: the maximum threshold «is limited by the estimates obtained from the included studies, and individuals with higher handgrip levels… might also benefit from an even lower risk», and «the uptick of the dose- response curves at the higher end… may simply represent lack of data rather than a genuine lack of association» (López-Bueno et al., 2022). So the cancer/CV “U” is a monotone-inverse curve with a spuriously-inverted right tail from thin data — the source diagnoses its own artifact -> The U-Shaped Association Artifact. Operative reading: the knee is not located, every reduction still pays, over-shooting a (possibly non-existent) upper threshold merely fails to help.
- Cancer had NO lower threshold — risk fell with strength from the lowest observed values «we found no minimal threshold for the beneficial associations between handgrip strength and cancer mortality» (López-Bueno et al., 2022) — cancer-specific and modified by cancer type (inverse for lung/colorectal/breast, null for prostate). Net: Lopez-Bueno confirms Celis-Morales’s monotone all-cause reading and bounds it with a studied range, while showing the “shape” is outcome-specific — a type-F refinement, not a contradiction.
The metric value — grip improves an office-based risk score
This is the cheap Tier-1 metric case, and the reason grip earns a place on Metrics for Targeted Health Guidance. Added to a conventional office-based score (age, sex, diabetes, BMI, systolic BP, smoking — no blood test), grip strength raised Harrell’s C index by:
| Outcome | C-index change (validation) |
|---|---|
| All-cause mortality | +0.012 (derivation +0.013) |
| CVD mortality | +0.011 (derivation +0.012) |
| CVD incidence | +0.009 |
The authors benchmark this: the +0.013 gain «is similar to that seen when adding high density lipoprotein cholesterol and N-terminal pro b-type natriuretic peptide» (C-index change 0.007 for HDL, 0.020 for NT-proBNP). Grip is «easily measured, cheap, and highly reproducible in clinical practice» — so its distinctive value is «in risk screening settings where blood sampling is not possible (for example, in community settings or in low and middle income countries)». Muscle weakness cut-offs (FNIH: <26 kg men, <16 kg women — close to EWGSOP2’s <27/<16, Sarcopenia Definition and Diagnosis) flagged higher risk for most outcomes, giving a binary screen alongside the continuous measure.
Age interaction — stronger in the YOUNGER, against the prior
A prior meta-analysis (53,476 participants) suggested the grip-mortality link was weaker under 60. With 500,000+ participants Celis-Morales had the power to test it and found the opposite: associations «were moderately stronger, not weaker, in the younger age groups», reading it as with age, factors other than grip become more important. Decision-relevant caveat: this cuts against reading grip as a purely old-age / frailty screen — the relative signal is present (and modestly larger) mid-life — but the absolute stakes still rise with age via baseline risk.
Mass vs strength — a distinct quantity, and a refinement (not a tension, not independence)
Muscle mass (Low Muscle Mass and Mortality, de Santana SR-MA) and muscle strength (this page) are different quantities measured differently, so the two sources refine the muscle-mortality picture rather than independently backing one claim. The parameter table (are these the same quantity?):
| Parameter | de Santana 2021 (MASS) | Celis-Morales 2018 (STRENGTH) | Same quantity? |
|---|---|---|---|
| Exposure | appendicular muscle MASS (ASMI, DXA/BIA) | handgrip STRENGTH / force (dynamometer, kg) | No — quantity vs function |
| Effect metric | between-group SMD (dead vs alive) | HR per 5 kg / per 1 SD | No — SMD vs hazard ratio |
| Population | community-dwelling non-frail >=65 | general population 40-69 | No |
| Design | SR-MA, 9 cohorts, n=10,028 | single cohort, n=502,293 | partial |
| Magnitude | SMD -0.18 (small) | HR ~1.16-1.20 per 5 kg (~1.48 per 1 SD) | — |
The same-quantity answer is No on exposure and effect metric (the two that decide it), so the
not-joined check (ii) fires — different quantities, consistent once matched. This is type-B (mass and strength are
distinct objects, routinely conflated as muscle) and type-F (Celis-Morales supplies the
strength-construct magnitude at scale that de Santana and EWGSOP2 asserted but did not quantify). It is
not type-E independent backing: the constructs differ, and EWGSOP2 already holds that «strength is
better than mass in predicting adverse outcomes» (Sarcopenia Definition and Diagnosis) because they
are different measures — de Santana’s result is consistent with mass carrying residual signal beyond
strength (an underpowered null, not a demonstrated independence -> Low Muscle Mass and Mortality), not
that mass and strength are one thing. No tension is filed; no [E-independent] token.
(inferred from Celis-Morales et al., 2018; de Santana et al., 2021)
Predictor / METRIC, not a proven lever — the line this page holds
(inferred from Celis-Morales et al., 2018)
- Established: low grip strength predicts higher all-cause and cause-specific mortality, and as a metric it adds discrimination to a cheap office-based risk score. That is enough to make grip a legitimate stratification metric — a readily-capturable number that places a person in a higher-risk cell (route-(a) baseline risk) -> Metrics for Targeted Health Guidance.
- NOT established: that training grip strength lowers mortality. Grip is a whole-body signal with many determinants — «handgrip strength has a strong genetic component with heritability shown to be 52%», plus nutrition, activity, disease — so grip is partly a marker of underlying health, not only a modifiable lever. Whether raising strength moves mortality is a separate claim on a separate (intervention) evidence base: resistance training’s mortality evidence is very-low-GRADE observational (Muscle-Strengthening Activity and Mortality), and its firmer payoff runs through function and falls.
- So grip is a good number to MEASURE (it places you in a stratum), not a number to STEER toward. Squeeze harder to live longer over-reads this evidence exactly as raise your VO2max to live longer does on Cardiorespiratory Fitness and Mortality.
Reverse causation — the central threat, and how far it is mitigated
Occult disease and frailty lower grip before death, so a grip-mortality association could be occult illness casting a shadow, not strength doing work. The paper mitigates two ways: a landmark analysis excluding events in the first 2 years (model 4) — associations persisted — and heavy covariate adjustment including prevalent comorbidities. The residual is stated plainly: «we cannot exclude the possibility of reverse causality» and «residual confounding is always possible and the associations observed may not imply causality.» Their defence is framing, not elimination: «given that we are largely interested in prediction and identification of people at increased risk, and not seeking to make strong causal inferences, reverse causality is not a major limitation.» That defence holds for the METRIC use (a marker can predict via reverse causation and still stratify usefully) but is exactly why it does not license the lever use -> the predictor/target line above.
Lopez-Bueno handles reverse causation LESS, not more — and over-reads the lever [@lopezbueno2022]. Pooling more cohorts does not mitigate the marker problem: the strings “reverse causation” and “residual confounding” never appear in the paper (srcgrep 0/2 files), and its named limitations are sparsity, heterogeneity, generalizability, and unverified proportional hazards — not the frailty/occult-disease shadow. Its only mitigation is structural (inclusion of «apparently healthy» adults; the NOS outcome-not-present-at-start domain), and follow-up as short as 2.3 years in some pooled studies AMPLIFIES the threat vs Celis-Morales’s 2-year landmark. Worse, the authors read the observational gradient as an intervention: «there is still a margin for improving muscular strength in order to reduce the risk of all-cause mortality», people «might also reduce their risk of cancer mor- tality by either maintaining or increasing muscular strength», and the paper offers «clinical guid- ance for exercise prescription» — none grounded in any trial (all 48 studies are observational). This is the guidance-null defeated by better warrant: the wiki holds the more disciplined line (marker, not proven lever) precisely where a gold SR-MA over-claims. Route the does-training-lower-mortality question to Muscle-Strengthening Activity and Mortality, where the intervention evidence actually lives.
Decision relevance
(inferred from Celis-Morales et al., 2018; López-Bueno et al., 2022)
- Capture grip with a hand dynamometer where a fuller work-up is wanted — cheap, fast, reproducible, no lab; it adds risk discrimination on top of age/sex/BP/BMI/smoking/diabetes, most usefully where blood tests are unavailable. Read a low value (or FNIH weakness: <26 kg men / <16 kg women) as a flag that places the person in a higher-risk stratum and warrants further assessment.
- Do not convert the number into a personal promise. The HR is a between-person association in an observational cohort; grip is part lever, part marker (heritable, disease-sensitive). The lever is the activity/protein that builds strength -> Muscle-Strengthening Activity and Mortality, Protein and Resistance Training for Muscle and Strength; grip is how you’d measure whether it moved.
- Distinct from mass and from fitness. Grip = strength/function; ASMI = mass (Low Muscle Mass and Mortality); VO2max = aerobic capacity (Cardiorespiratory Fitness and Mortality) — three different measured quantities, all predictors, all with the same predictor-not-proven-cause caveat.
Limits — the open loop
(inferred from Celis-Morales et al., 2018; López-Bueno et al., 2022)
- Observational-cohort lineage only. The base is now multi-cohort (Celis-Morales UK Biobank + Lopez-Bueno’s 48-cohort SR-MA, 3.14M, 40+ countries), which adds cross-population breadth but NO new design — no RCT, no MR/genetic route. UK Biobank sits inside Lopez-Bueno’s pool, so the agreement is type-F (breadth + shape), not type-E independent backing; the whole shares one confounding structure.
- Reverse causation / residual confounding unresolved (above) — Celis-Morales mitigates with a 2-year landmark; Lopez-Bueno does not address it at all (and adds studies with follow-up as short as 2.3 y). Not eliminated in either.
- Thresholds are studied-range-bounded — Lopez-Bueno’s upper thresholds (all-cause 50 kg) sit at the edge of the 15-50 kg studied range and the cancer/CV “U” upper arm is a self-diagnosed sparsity artifact, not evidence of harm from high strength.
- UK Biobank is not representative (healthier, less deprived than the general UK population); the authors judge the magnitude estimates generalisable but the summary statistics not.
- Linearity is spline-judged, not established — no located knee/plateau, but weak evidence of true monotonicity, and power thins at high strength.
- The loop is open. This grades coherence and source-fidelity, never validity: no operation here checks whether measuring — let alone raising — grip strength changes what a person experiences.
Self-critique [run 2026-08-19, before commit — Lopez-Bueno reweaving]
- Laundered independence — the primary risk, and it is blocked. The dangerous move would be reading a
second large “confirming” source as independent backing and lifting confidence on convergence. It is
not: Celis-Morales (UK Biobank) IS one of Lopez-Bueno’s 48 SR studies, and UK Biobank data enters the
pooled dose-response — so this is type-F (breadth + explicit shape), no
[E-independent]token, and the confidence lift tohighis justified on total support (a gold SR-MA over 3.14M across 40+ countries is a stronger prognostic base than one UK cohort) + cross-population breadth, NOT on agreement. The whole base remains one observational-cohort lineage with no genetic/MR route — stated in Limits. - Overclaim — survives the stripped-claim test. Strip the most dramatic figure (grip out-ranking SBP per-1-SD) and the decision-relevant core holds: grip predicts mortality broadly and adds discrimination to a cheap office score. No wiki-voice superlative of the strongest-evidence-the-wiki-holds kind is asserted; the “strongest association” phrase is scoped to a within-model marker comparison, not a causal claim.
- No fake tension. Lopez-Bueno vs Celis-Morales is F-refinement (parameter table: same-quantity No on metric/population/reverse-causation), not a filed tension; mass-vs-strength is a type-B distinction with not-joined check (ii) firing. The one genuine friction — Lopez-Bueno’s interventional over-reach — is handled as guidance-null-defeated-by-better-warrant, not laminated into a tension.
- Coherence, not validity (R1): the page establishes grip as a predictor/metric and explicitly does NOT establish that raising grip lowers mortality — a claim about the evidence, routed to the intervention page, not a claim about the world.