Orbiter of the osteoarthritis cluster. This page answers a DIFFERENT decision-question from the
nucleus -> Knee Osteoarthritis and Modifiable Levers: that page appraises levers that reduce
pain and preserve function in ESTABLISHED symptomatic knee OA (a treatment/management question);
this page appraises what raises the risk of DEVELOPING incident knee OA in the first place (a
prevention/prognostic question). The nucleus closed by flagging that whether its levers prevent
incident OA is a separate, unaddressed question — this orbiter is that question.
The appraisal frame — route-(a) prognostic association, NOT proven prevention. Every factor below is an observational association with incident OA (129 cohort studies, no prevention RCT). A factor being «modifiable» and associated with incidence does not establish that modifying it prevents OA — that is an untested causal step the source does not close. Read these as route-(a) baseline-risk / prognostic stratifiers (who is more likely to develop knee OA), not as demonstrated causal levers. (inferred from Duong et al., 2025)
The source and its outcome distinction
(Duong et al., 2025) Gold SR+MA (Duong 2025, Osteoarthritis and Cartilage): 129 cohort studies, >150 candidate risk factors, no date/language restriction; pooling was possible for ten factors, five significant. Risk of bias: 90 studies low, 24 moderate, 15 high. Most data came from three US cohorts (OAI, MOST, Framingham) — a generalisability limit (few studies outside the USA / in LMICs), and the cohorts may share participants (unquantifiable duplication).
The outcome split is load-bearing — and it is a streetlight case. Duong analysed radiographic KOA (KL grade >=2) separately from symptomatic KOA, «which resulted in greater homogeneity» — but «the overlap between the two is limited», i.e. radiographic OA «does not consistently relate to symptoms» (Duong et al., 2025). Almost all the poolable evidence — and the entire PAF — is for radiographic KOA, the X-ray finding. But the patient-important outcome is symptomatic KOA (pain, function loss) -> Grip Strength and Mortality-adjacent function outcomes. Symptomatic pooling was mostly non-significant or impossible («We were unable to calculate PAF for symptomatic KOA as the available data did not permit»). So the data are richest on the outcome that matters least to the patient and thinnest on the one that matters most — the streetlight effect on the outcome side. Any risk figure here inherits this: it predicts an X-ray change of uncertain symptomatic consequence.
Modifiable factors — magnitude, CI, and outcome (all radiographic KOA unless stated)
(inferred from Duong et al., 2025) Duong separates «potentially modifiable» from «non-modifiable» factors. The pooled modifiable associations, high-to-moderate certainty for the first four:
| Factor | Estimate (radiographic KOA) | Studied contrast / range | Heterogeneity | Note |
|---|---|---|---|---|
| Overweight/obesity | aOR 1.17 per BMI unit (1.10, 1.24) | BMI as continuous exposure (5 studies pooled) | I2 52.9% | the dominant rock |
| Previous knee injury | aOR 2.67 (1.41, 5.05) | prior knee injury vs none (6 studies pooled) | — | mostly ACL +/- meniscus |
| Higher femoral neck BMD | aOR 1.82 (1.12, 2.94) | femoral neck BMD >=1.01 g/cm2 vs lower (2 studies) | — | a MARKER, not a lever (below) |
| Occupational physical activity | see the occupational block below | heavy manual / load / shift work | — | the PA-paradox channel |
(Duong et al., 2025) for every table cell. Un-poolable overweight studies ranged from aOR 1.6 (1.0, 2.8) to 18.3 (5.1, 65.1) for BMI >=25 — the huge upper estimates carry «unusually wide CIs, suggesting a sparse data bias», so the pooled 1.17/unit is the defensible figure, not the outliers.
Obesity is the dominant modifiable rock — and it is the SAME lever the treatment page already holds. Of the 14% of radiographic-KOA risk attributable to the two poolable modifiable factors, overweight/ obesity carries 77.2% and previous injury 22.8% (weighted PAF 0.102 vs 0.030; total 0.141) (Duong et al., 2025). «Together, two modifiable risk factors (overweight/obesity and previous knee injury) account for 14% of incident radiographic KOA.» This is a structural-leverage convergence: weight management is simultaneously the top incidence lever here and the top symptom/function lever on the treatment page -> Knee Osteoarthritis and Modifiable Levers, and it sits on the same adiposity machinery as the cardiometabolic levers -> Does Weight Loss Reduce Cardiovascular Events, Ectopic Fat and Depot-Specific Risk, Shared Modifiable Levers Across Age-Related Diseases. The caveat: the incidence figure is prognostic — no held evidence shows that losing weight in a normal-BMI-trajectory person prevents incident OA; the IDEA causal anchor on the treatment page is in already-symptomatic overweight patients, a different stratum.
Higher BMD raises risk — but it is a prognostic MARKER, not an actionable lever. Higher femoral neck BMD associates with more incident radiographic KOA (aOR 1.82; highest lumbar-spine quartile in one study aOR 4.7 [2.1, 10.7]). The direction is the OPPOSITE of what a naive «bone health is good» reading expects, and the mechanism is plausibly subchondral-bone stiffness driving cartilage load. But deliberately lowering BMD would be contraindicated (fracture risk), so this is a route-(a) risk marker / mechanistic clue, never a modification target. (inferred from Duong et al., 2025)
The occupational-load channel — this IS the Physical Activity Paradox, at the knee
Occupational physical activity was significantly associated with incident KOA, while leisure-time PA was null overall — the domain-split the wiki already holds -> The Physical Activity Paradox. Duong states it explicitly: «Our findings reflect the physical activity paradox documented in previous research» (Duong et al., 2025).
- Heavy manual occupation, men: «Men had a two-fold greater risk of incident radiographic KOA if they engaged in heavy, manual occupations (aRR 2.7 95% CI 1.17, 6.26). However, an association was not seen for women.» (Duong et al., 2025) — note the sex-discordance mirrors the mortality PA paradox (harm in men).
- Manual load >20 kg for >=13 years: aHR 1.71 (1.03, 2.85) for incident KOA (ICD).
- Shift work >20 years: aHR 1.28 (1.04, 1.57) for incident radiographic KOA.
The mechanism differs from the mortality paradox (there: sustained HR/BP load without recovery; here: direct repetitive joint/cartilage loading — knee bending, kneeling, load-bearing), but the structure is identical: physical activity is not one exposure, and its occupational form carries risk its leisure form does not. (inferred from Duong et al., 2025)
Non-modifiable factors and the true nulls
- Female sex: aOR 1.93 (1.25, 2.98) for radiographic KOA (I2 63.5%; sensitivity aOR 2.22 after dropping one study) — «females had almost double the odds» vs males. Moderate-to-low certainty.
- Older age: aOR 1.15 per unit (1.00, 1.33) — but I2 89.3% and the lower CI touches 1.00; sensitivity analysis dropping one study collapsed it to aOR 1.06 (1.02, 1.10). Duong notes age «was not found to be a strong risk factor… as most incident cases of KOA are likely to be diagnosed at middle age». Weak/unstable.
- The four evidenced NULLS (route to the four-evidence-states discipline): «Pooled estimates show a lack of association between the following risk factors and incident radiographic KOA: smoking, alcohol intake, leisure-time physical activity and osteoporosis (based on low or very low certainty of evidence).» (Duong et al., 2025) — these are low-certainty nulls (insufficient-leaning-null), not strong no-effect verdicts.
All non-modifiable factors above are radiographic KOA; the pooled symptomatic sex/age estimates were non-significant or conflicting.
Leisure-time PA — a route-(b) effect-modification signal inside the null
The overall leisure-PA null hides a candidate effect modification: «Two studies found that high levels of leisure-time physical activity in overweight or obese individuals was associated with incident KOA, with one showing that the weight-associated risk for KOA was increased in women with high levels of physical activity; RR 3.6 (1.7 to 7.7) for the heaviest quintile» (Duong et al., 2025); a low-muscle-mass subgroup doing weight-bearing PA also showed excess risk (OR ~1.20). So leisure PA may be joint-neutral in a normal-weight person but joint-harmful when it loads an overweight or low-muscle-mass knee — a route-(b) interaction (obesity x mechanical load), but from single studies, not pooled, so it clears only the plausibility bar, not the positive-interaction bar. Duong: «the type of physical activity, and/or the intrinsic capacity of the individual may be important considerations.» This is the caution that leisure PA is not uniformly safe for every knee — the substitution/stratum matters.
Running specifically — a dose-intensity-dependent exposure, not a flat risk factor
A second gold MA takes the running sub-exposure that Duong’s leisure-PA null lumps in, and splits it by level. (Alentorn-Geli et al., 2017) Gold SR+MA (Alentorn-Geli 2017, J Orthop Sports Phys Ther): 25 studies qualitative (n = 125,810), 17 pooled (n = 114,829); 7 prospective cohorts, 18 case-control/cross-sectional. Level of evidence 2a, etiology/harm; most included studies carried a high risk of bias (selection, performance, detection, attrition, reporting). The headline is a U/J across running intensity — sedentary and competitive both worse than recreational:
| Exposure (hip and/or knee OA, overall) | Prevalence (95% CI) | OR vs sedentary controls (95% CI) | I2 |
|---|---|---|---|
| Sedentary controls | 10.2% (9.9, 10.6) | 1.00 (referent) | — |
| Recreational runners | 3.5% (3.4, 3.6) | 0.86 (0.69, 1.07) | 50% |
| Competitive runners | 13.3% (11.6, 15.2) | 1.34 (0.97, 1.86) | 53% |
- The big-rock read: sedentary is WORSE for the joint than recreational running. Controls (10.2% prevalence) carried roughly 3x the hip/knee OA of recreational runners (3.5%), and the competitive-vs- recreational difference was significant (P < .001). The protective direction is clearest for the knee alone: recreational OR 0.83 (0.70, 0.99, I2 0%) — a significant lower association — rising to OR 0.78 (0.68, 0.89) in men and 0.54 (0.41, 0.71) in women. This counters the «impact wrecks knees» prior directly: for the general (self-selected) adult, taking up recreational running does not raise — and may lower — hip/knee OA odds, while not running sits on the high arm beside competitive running. (Alentorn-Geli et al., 2017) «Running at a recreational level can be safely recommended as a general health exercise, with the evidence suggesting that it has benefits for hip and knee joint health. The amount of running that is safe for the joints could not be determined.»
- The shape is joint-specific — the hip tells a different story once adjusted. On the fully-poolable overall estimate the signal is protective, but in the age+BMI-adjusted hip subgroup the OR flips to 1.72 (1.16, 2.57) (and 1.80 [1.12, 2.90] with workload added), while the knee stays protective/null. So running protects the joint is not a single curve: the knee and hip arms of the same exposure point opposite ways under adjustment — report the outcome, not a blanket joint verdict. (Alentorn-Geli et al., 2017)
The U-shape must clear the artifact gate — and it only partly does
(inferred from Alentorn-Geli et al., 2017) A protective/plateau arm must survive a referent-correction or an objective/genetic check before it is believed. Here it survives one and fails another:
- Survives (partly): the most-adjusted subgroup — age + BMI + workload + previous injury, the two studies controlling the most confounders (both recreational) — kept a significant OR 0.73 (0.61, 0.89, I2 0%), so the protective signal is not purely obesity/injury confounding.
- Does not survive design: 18 of 25 studies are case-control / cross-sectional, so reverse causation and sick-quitter are unexcluded — healthy joints self-select into recreational running, and people who develop knee pain/OA stop running (and competitive runners who break down drop out, leaving a resilient survivor cohort that attenuates the competitive harm arm). The protective magnitude is therefore artifact-vulnerable even though the direction (recreational is not a joint-harm exposure) is robust. Gate-6 unequal precision also applies: the recreational arm pools ~84,000 runners against a far smaller competitive arm (~1,400), so between-arm contrasts carry very unequal reporting precision.
Dose and duration are confounded, and the safe mileage is undefined
- Level and years-of-exposure cannot be separated. The >15-year subgroup carried a higher association (OR 1.01 [0.72, 1.43]) than the <15-year subgroup (OR 0.60 [0.49, 0.73]; prevalence 3.03% vs 17.2%) — but (Alentorn-Geli et al., 2017) «those who had less than 15 years of exposure to running were grouped as recreational runners and those who had more than 15 years of exposure to running were grouped as competitive runners. As a result, no information on the effects of recreational running for more than 15 years on the association with hip and knee OA is available.» Duration and intensity are perfectly collinear in the data, so neither the knee nor the plateau of the curve can be located.
- Mileage was never quantified. Konradsen’s orienteering runners (21-42 km/wk) showed no association; Marti’s elite runners (92 km/wk) showed raised hip OA — a dose signal. (Alentorn-Geli et al., 2017) «It is therefore not possible to formulate any recommendation from a practical standpoint on the quantity of running that would be safe for the hip and knee.» The animal evidence is directionally concordant. (Alentorn-Geli et al., 2017) «a chondroprotective effect of slight or moderate-intensity running (as determined by distance) in rats that was no longer present with intense running» — a knee/upper-bound shape, but mechanism in rats, transportability- caveated (animal != human).
Transportability — the obese/sedentary novice is the UNTESTED harm stratum
(inferred from Alentorn-Geli et al., 2017) The protective estimate lives in self-selected, mostly lean runners. It does not transfer by grade alone to the deconditioned, obese, or previously-injured novice — the mechanism’s support factors (a resilient joint, normal load, no prior injury) are exactly what that stratum lacks. The authors say so. (Alentorn-Geli et al., 2017) «It still remains unclear whether prescribing running in the overweight/obese patient with an additional high-impact occupational workload and previous injury is safe for joints. It is very likely that the combination of these 3 factors in a patient who begins to run may increase the risk of OA.» This converges with Duong’s route-(b) signal above (leisure load harms the overweight / low-muscle knee): both say weight-bearing activity is joint-neutral-to-protective for a healthy knee and potentially harmful for an already-vulnerable one. The decision-change for the obese sedentary target stratum is therefore not run freely but: being sedentary is itself a joint-OA exposure, so a graded, low-mileage, weight-managed return to loading is defensible, while high-mileage/high-impact loading of an obese, previously-injured knee is the explicitly-untested harm case — favour lower-impact modalities until weight and conditioning improve -> Exercise Therapy for Osteoarthritis Pain and Function, Exercise Interventions and Sports Injury Prevention.
Cross-source: a DISTINCTION from the therapy MA, a REFINEMENT of Duong (the parameter table)
Before asserting how these three MAs relate, the matched-parameter check (op-weave step 2a):
| Parameter | Duong (incidence RF) | Alentorn-Geli (running) | Goh (exercise therapy) | Same quantity? |
|---|---|---|---|---|
| Decision question | what raises incident KOA | does running cause/spare OA | does exercise relieve established OA | Goh: NO (treatment) |
| Population | OA-free cohorts | runners vs sedentary | people with knee/hip OA | Goh differs |
| Exposure | all RFs incl leisure PA | recreational vs competitive running | exercise-therapy programme | partial overlap |
| Outcome | incident radiographic KOA | prevalent+incident hip/knee OA (OR) | pain/function change (SMD) | differ |
| Design | 129 cohorts | 17 (mostly case-control/XS) | 77 RCTs | differ |
- Goh vs Alentorn-Geli = a DISTINCTION, not a tension. They answer different questions (does exercise relieve established OA vs does running cause OA), on different populations (OA patients vs runners), with different metrics (SMD vs OR). The not-joined check (ii) fires — different unit and horizon — so the apparent is loading good or bad for the knee? clash is a category split, recorded as a distinction. Together they resolve the surface paradox: recreational loading is joint-safe-to-protective for a healthy knee (Alentorn-Geli) and graded exercise relieves an already-arthritic one (Goh) -> Exercise Therapy for Osteoarthritis Pain and Function.
- Duong vs Alentorn-Geli = type-F refinement, NOT type-E confirmation. Both are incidence-side, but the outcome (incident radiographic KOA vs mixed-prevalent OR) and design (cohort vs case-control) differ, so they are not independent measurements of one quantity. Alentorn-Geli refines Duong’s flat leisure-PA null: running is not uniformly null but U-shaped by intensity (protective at recreational, harmful at competitive) — the composite beats either alone, without laundering a shared estimate.
Additional moderate-certainty signals (mostly single-study, dietary factors measurement-limited)
Moderate-certainty associations Duong reports without pooling: higher incident-KOA risk with parity (5-12 births), poor sleep quality, high ultra-processed-food intake, and depressive symptoms; lower risk with a Mediterranean diet, high dietary fibre, and cereal / dark bread / green tea intake (Duong et al., 2025). The dietary factors carry the usual caveat -> Measurement Error in Dietary Assessment: self-reported intake attenuates gradients, so a null or a modest signal here is weakly informative. These are candidate levers, not established ones — Duong itself flags smoking, alcohol and diet as needing «large, prospective cohort studies».
Gaps this source names (type-G)
- No symptomatic-KOA PAF, and radiographic != symptomatic. The attributable-fraction machinery runs only on the X-ray outcome; the patient-important one is unquantified. The single most decision-relevant number is missing by data availability, not by nature — a named gap, not a null.
- No prevention RCT. Everything is observational; «the possibility of confounding cannot be eliminated». Whether modifying obesity / occupational load / injury prevents incident OA is the open causal question -> Baseline Risk and the Relative-Absolute Split.
- US-cohort concentration. OAI/MOST/Framingham dominate; transportability to non-US / LMIC / non-white populations is explicitly limited.
- Occupational PA absent from the PAF. Its contribution «could not be accounted for in the PAF analysis» (no correlation matrix), so the 14% understates total modifiable risk — occupational load is a real but unquantified addition.
Self-critique [run 2026-10-04, before commit — re-run on the Alentorn-Geli weave]
- Not laundered from one source restated. The original Duong moves stand: (1) the provisional-C induction of the incidence question as distinct from the treatment nucleus (closing that page’s open thread), (2) the marker-vs-lever reclassification of BMD (Duong files it as modifiable; the wiki flags it non-actionable because the modification direction harms), (3) the bridge tying the occupational-load finding to the held PA-paradox fabric, and (4) the streetlight read on radiographic-vs-symptomatic. The Alentorn-Geli weave adds (5) the U-shape-by-intensity that refines Duong’s flat leisure-PA null (type-F, composite beats either), (6) the artifact-gate appraisal of the protective arm, and (7) the stratum-transfer caution converging with Duong’s route-(b) overweight-knee signal. Each cross-page move is tagged; the estimates stay the sources’.
- Duong vs Alentorn-Geli is F, not E — and the parameter table proves it. The cross-source subsection tables matched parameters with a same quantity? column: different outcome (incident radiographic vs mixed-prevalent OR) and design (cohort vs case-control) mean these are NOT independent measurements of one quantity, so no type-E independence is claimed. Goh is recorded as a DISTINCTION (not-joined check (ii): different unit/horizon), not a tension.
- The U-shape did not drive a recommendation on its own. Per The U-Shaped Association Artifact, the protective magnitude is flagged artifact-vulnerable (reverse causation / sick-quitter unexcluded by the case-control design; gate-6 unequal precision across arms). Only the direction (recreational is not a joint-harm exposure; sedentary sits on the high arm) is carried into the decision-read, and the obese/injured novice is explicitly held as the untested harm stratum, not told to run freely.
- Not overclaimed. Confidence stays LOW: both sources observational; Duong carries high heterogeneity on several pooled estimates (age I2 89%, sex I2 64%), self-reported exposures, US-cohort concentration, and the symptomatic outcome largely unpooled; Alentorn-Geli is level-2a with most studies at high risk of bias, unquantified mileage, perfectly confounded level/duration, and English-only inclusion. Every «modifiable» factor remains framed as prognostic, not proven-preventable.
- Coherence, not validity (R1): the page reports associations and their certainty; it does not assert that reducing any factor, or taking up running, prevents OA — the loop (no realized prevention outcome) stays open.