Orbiter of the osteoarthritis cluster. This page characterizes the EXERCISE lever in depth — how
big its effect on pain and function is, how long it lasts, and who responds — where the nucleus
-> Knee Osteoarthritis and Modifiable Levers ranks exercise against the other lever (weight loss).
The nucleus leaned on a single-site RCT (IDEA, in a weight-loss context) and a surrogate-outcome MA
(EULAR, fitness/strength) for the exercise claim; this page supplies the direct, patient-important
pain/function effect of exercise alone, pooled across 77 trials — a quality upgrade and independent
backing on a nominally-covered question.
Scope discipline: exercise as a modifiable lever on OA pain and function — patient-important QoL outcomes on the one health axis, often critical-rated. NOT disease management (drug/injection/surgery selection, acute-pain control). These outcomes (self-reported pain, heterogeneous WOMAC/function scales, QoL instruments) are measured worst, so the discipline is more honest uncertainty, not more confident advice -> Surrogate Outcomes.
The effect — moderate on pain and function, small on QoL, at 8 weeks
(Goh et al., 2019) Gold SR+MA: 77 RCTs, 6472 participants with knee or hip OA; exercise (all types, administered alone) vs usual care, between- group SMD (Cohen’s method, random effects) at or nearest to 8 weeks (the most commonly reported time point). Usual care — not an active comparator — was the reference; education/manual-therapy controls were excluded to standardize the reference arm. «Relative to usual care at or nearest to 8 weeks, exercise conferred a moderate» benefit for pain relief (SMD 0.56, 95% CI 0.44-0.68).
| Outcome | SMD (95% CI), all studies | I2 | SMD, “homogeneous” subset (I2 < 30%) | Cohen size |
|---|---|---|---|---|
| Pain | 0.56 (0.44-0.68) | 74.1% | 0.50 (0.43-0.58) | moderate |
| Function | 0.50 (0.38-0.63) | — | 0.43 (0.35-0.51) | moderate |
| Performance (objective) | 0.46 (0.35-0.57) | — | 0.32 (0.25-0.39) | small-moderate |
| Quality of life | 0.21 (0.11-0.31) | — | 0.18 (0.09-0.27) | small |
- The benefit is larger on what the patient feels (pain, function) than on QoL. The QoL effect is real but small (0.21); the authors attribute the gap to the difficulty of a standardized instrument capturing «a construct that is individually unique» — a measurement limit, not necessarily a smaller true effect.
- The honest-subset readout matters. Restricting to “homogeneous” RCTs (removing the heterogeneity-driving studies until I2 < 30%) shrinks every estimate by roughly 10-30% but leaves all four statistically significant. So the headline figures are somewhat inflated by the lower-quality, heterogeneous trials, and the defensible lower bound is pain ~0.50 / function ~0.43 — still moderate.
Durability is the decision-changer — the effect decays to nothing by 9-18 months
(Goh et al., 2019) «We found a general trend for the effects of exercise therapy to peak at 2 months for all outcomes»; thereafter «The effects were reduced gradually after 2 months and became no better than the usual care group at 9 to 18 months depending on the outcome.»
- The decision-change (route to adherence-is-part-of-the-effect). Exercise for OA is not a course you complete — it is a maintained exposure. A time-limited programme buys 2-6 months of relief and then the benefit is gone; the advice is continuation, not completion. This concords with the Fransen Cochrane finding (cited by Goh) that pain/function gains halve between month 2 and 6 (Goh et al., 2019).
- Caveat on the early peak. The large effect seen within 1 month «may be due to the small study effect» (only 1-2 trials at that point), so the true trajectory is best read from ~2 months onward.
- This is a plateau-then-reversal in TIME, not a dose-response plateau — do not confuse the two. It says the effect is contingent on ongoing stimulus, the opposite of a durable structural change -> structural-leverage is exactly what exercise-for-OA lacks, unlike weight loss (which removes a standing driver) on the nucleus. — Goh shows the fade, not its cause; the source offers only a hypothesis, «One of the reasons for this decline may be poor exercise adherence, in which personal beliefs, social support, relationship with provider and ease of access to exercise facilities play important roles [30].» (Goh et al., 2019) — consistent with, not proof of, contingent on ongoing stimulus.
Who responds — determinants, and the obesity stratum
(Goh et al., 2019) Subgroup / meta-regression on the pain outcome (overall ES 0.56, I2 74.1%). Univariate SMDs by stratum:
| Stratum | SMD (95% CI) | Contrast | Modifier after adjustment? |
|---|---|---|---|
| Age < 60 | 1.32 (0.79-1.86) | vs >=60: 0.44 (0.34-0.55) | yes (multivariate P 0.04) |
| Knee OA | 0.64 (0.51-0.78) | vs hip: 0.17 (-0.17-0.51); mixed: 0.43 (0.13-0.72) | yes (P 0.10) |
| Not on TJR waiting list | 0.62 (0.49-0.75) | vs waiting-list: 0.33 (0.04-0.63) | yes (P 0.10) |
| BMI >= 30 (obese) | 0.56 (0.28-0.84) | vs < 30: 0.46 (0.38-0.60) | no (P 0.78) |
- No evidence that obesity blunts the exercise effect — the transportability finding for the deliverable’s obese/sedentary stratum (study-level subgroup, underpowered for a null). The obese subgroup’s point estimate is if anything slightly higher (0.56 vs 0.46), and the difference is far from significant (P 0.78). Since obesity is the dominant OA risk factor -> Knee Osteoarthritis Incidence and Risk Factors, the entry stratum for most OA patients IS the obese one, and the exercise lever transports to it intact. Combined with the nucleus’s IDEA result (weight loss is additive on top), the obese OA patient should do both — exercise works regardless of weight, and losing weight adds a second, mechanistically distinct benefit.
- The severity gradient — milder disease, larger benefit. Not-awaiting-replacement (0.62) beats waiting-list (0.33); the inverse association between exercise benefit and OA severity is the pattern, consistent with exercise producing «greater improvement with milder than more severe OA» (Goh et al., 2019). Decision read: for advanced, surgery-bound disease, expect a smaller symptomatic return from exercise — but still positive.
- Hip OA pain response is UNESTABLISHED here, not a demonstrated null. Only 8 hip trials; pain SMD 0.17 with a CI crossing zero (-0.17-0.51) — insufficient evidence, not proof of no effect (the four-evidence-states discipline). The cluster’s exercise evidence is knee-dominant (80% of trials); do not transport the knee effect size to hip OA as if established.
- Younger patients respond more, but the authors caution this may reflect OA severity, comorbidity burden and reduced functional reserve in older patients rather than age itself.
Read every determinant as route-(b) effect-modification CANDIDATE, hypothesis-generating only. (inferred from Goh et al., 2019) These are study-level (aggregate) subgroup/meta-regression results, so they carry ecological bias — «the average change observed at the group level does not accurately reflect the change that occurs within each individual». The authors set a deliberately loose P <= 0.10 «to ensure that we would not miss any potential determinants» and state «the aim of this analysis was to generate hypotheses to guide future individual patient-data meta-analyses». So these are the false-positive-prone route-(b) signals the method layer warns about — the age/joint-site/severity directions are plausible and worth testing in IPD, but not yet confirmed effect modifiers. The BMI null is less prone to multiple-comparison inflation than a positive, but a study-level subgroup null is low-power, so it is no evidence of modification, not evidence of none (revised 2026-10-06; was: the more robust read).
Certainty — why this is moderate, not high
(inferred from Goh et al., 2019)
- Small-study / publication bias. Egger’s test was significant (P < 0.05) for all outcomes except QoL — the pain/function/performance effects are inflated by small favourable studies.
- Less-rigorous trials inflate the effect. «RCTs with less rigorous methods (small sample size, unclear/no exercise adherence monitoring, no explicit use of ITT) or RCTs that were heterogeneous tended to inflate the effect of exercise» — the honest-subset figures above are the correction.
- Exercise cannot be blinded, so the greatest risk of bias is blinding of patients/physicians; the self-reported outcomes (pain, function, QoL) are the most exposed. Objective performance (assessor- blinded in 52% of trials) is the least-biased readout, and it carries a smaller honest-subset effect (0.32).
- No single pooled point estimate from good-quality trials only was possible — no composite quality score for exercise RCTs exists, so robustness was assessed by subgroup/sensitivity rather than a quality-restricted meta-analysis.
How this refines the cluster
(inferred from Goh et al., 2019)
- Type-F/E on the nucleus’s exercise claim. The nucleus -> Knee Osteoarthritis and Modifiable Levers flagged a limit: its exercise evidence was on fitness/strength surrogates (EULAR) plus one single-site RCT (IDEA). Goh partly closes it with the direct patient-important pain/function effect of exercise alone, from 77 trials not including IDEA — a different design (large MA vs single-site RCT) reaching a compatible conclusion (independent backing), while bounding the claim with the durability decay and the small-study inflation the nucleus did not hold.
- The obesity convergence, now on the treatment side too. The incidence orbiter found obesity the dominant incidence lever; the nucleus found weight loss the dominant symptom lever; this page adds that the exercise lever works regardless of obesity — so for the obese OA patient the two levers are independent and additive, not competing.
Which exercise type? The companion network meta-analysis
(Goh, Persson, Stocks, Hou, Welton, et al., 2019) The mode comparison the pairwise MA pooled away is the remit of a companion Bayesian network meta-analysis from the same project (PROSPERO CRD42016033865): 103 RCTs, 9134 participants (76 trials vs usual care, 27 head-to-head), five ACSM-based types (aerobic, mind-body, strengthening, flexibility/skill, and mixed = more than one core type or type unspecified), primary time point 8 weeks. There is no aquatic node: aquatic trials are split across the types by content. The effect estimates below come from the paper’s Fig. 3 (journal p. 754), which is an image, not text; they were read from the rendered PDF page.
Parameter table — the NMA’s per-type numbers are NOT the pairwise MA’s numbers re-cut.
| Parameter | Pairwise MA (Goh - Exercise Therapy Knee Osteoarthritis) | NMA (Goh - Exercise Types NMA Knee Hip OA) | Same quantity? |
|---|---|---|---|
| Trial base | 77 RCTs, exercise vs usual care | 103 RCTs; 76 vs usual care + 27 head-to-head | largely the same trials (same project) |
| Effect metric | between-group SMD, frequentist random effects | SMD of change scores, Bayesian random effects, 95% CrI | no — different estimators |
| Exposure | all exercise pooled | five types, each its own node | no |
| Pain vs usual care | 0.56 (0.44-0.68) | 0.47 (mixed) to 1.11 (aerobic, mind-body) | no — do not read the NMA’s 1.11 as double the pooled effect |
So the magnitudes are not comparable across the two papers; the NMA’s value is the contrast between types within one model.
Effect vs usual care, SMD (95% CrI) (Goh, Persson, Stocks, Hou, Welton, et al., 2019)
| Type | Pain | Function | Performance | QoL |
|---|---|---|---|---|
| Aerobic | 1.11 (0.69, 1.54) | 0.59 (0.10, 1.07) | 1.05 (0.63, 1.48) | 0.39 (-0.06, 0.83) |
| Mind-body | 1.11 (0.63, 1.59) | 0.81 (0.27, 1.36) | 0.53 (0.01, 1.05) | 0.24 (-0.09, 0.58) |
| Strengthening | 0.73 (0.49, 0.98) | 0.76 (0.48, 1.03) | 0.55 (0.32, 0.78) | 0.26 (0.05, 0.47) |
| Flexibility/skill | 0.65 (0.29, 1.00) | 0.68 (0.28, 1.09) | 0.68 (0.34, 1.03) | 0.33 (0.03, 0.68) |
| Mixed | 0.47 (0.26, 0.69) | 0.43 (0.18, 0.69) | 0.48 (0.27, 0.69) | 0.19 (0.04, 0.35) |
What separates, and what does not. Of the 40 head-to-head contrasts in Fig. 3, four have a CrI excluding zero: aerobic beats mixed on pain (0.64, 0.21 to 1.08) and performance (0.57, 0.14 to 1.00); mind-body beats mixed on pain (0.64, 0.14 to 1.13); aerobic beats strengthening on performance (0.50, 0.05 to 0.96). No contrast among the four single-type modes excludes zero on pain, function or QoL (e.g. aerobic vs strengthening, pain 0.38, -0.07 to 0.83) (Goh, Persson, Stocks, Hou, Welton, et al., 2019). The authors’ summary: «Mixed exercise was the least effective for all outcomes and had significantly less pain relief than aerobic and mind–body exercises.» (Goh, Persson, Stocks, Hou, Welton, et al., 2019) The trend across the ranked types reached significance for pain only (p = 0.01; function 0.07, performance 0.06, QoL 0.65) (Goh, Persson, Stocks, Hou, Welton, et al., 2019); the hierarchy it tests is ordered by the same effect sizes, so it is not an independent check.
The authors read a hierarchy; the credible intervals do not show one among single types. The authors conclude «Aerobic or mind–body exercise may be the best for pain and function improvements. Strengthening and flexibility/skill exercises may be used for multiple outcomes.» (Goh, Persson, Stocks, Hou, Welton, et al., 2019) On pain the aerobic and mind-body point estimates (1.11) sit about 0.4 SMD above strengthening (0.73) and flexibility/skill (0.65), but every single-type contrast CrI includes zero. The wiki reads that as not shown to differ, low certainty, point estimates favouring aerobic/mind-body on pain — neither equivalence nor an established ranking.
- The decision read. In the pooled knee + hip (knee-dominant) network, every single-type programme beats usual care on pain, function and performance at 8 weeks; in hip OA alone, no type’s pain or function CrI excludes zero (below). Since the single types are not shown to differ on pain or function, choosing among them is a Layer-3 preference and adherence call made under uncertainty, with a weak lean toward aerobic or mind-body for pain. The mixed category ranked last on every outcome, which argues against assuming a multi-component programme is better. But mixed separates only from the two thinnest nodes (aerobic, mind-body), on an outcome with small-study bias; against strengthening, the other large node, it does not separate, 0.26 (-0.04 to 0.57). So part of the mixed deficit may be small-study inflation of the aerobic and mind-body estimates.
- Why mixed loses is unexplained. The authors find «no satisfactory biological explanation» and hypothesise «flawed implementation of the programme, such that intensity of the individual components was insufficient or poorly adhered to due to the complexity of the regimen» (Goh, Persson, Stocks, Hou, Welton, et al., 2019). Mixed also collects trials whose type was unspecified, so it is partly a reporting-quality bin, not a pure exercise type. The authors offer mixed as a fallback for patients who do not respond to other types, «because it is still better than no exercise control for all four patient-centred outcomes» (Goh, Persson, Stocks, Hou, Welton, et al., 2019).
- The aerobic edge on performance is plausibly training-specificity. The performance outcome prioritised gait/walking measures (Goh, Persson, Stocks, Hou, Welton, et al., 2019), and several aerobic arms are walking programmes (others include group cycling), so an aerobic lead on walking tests is plausibly close to training the test. Read it as likely specificity, not as a general aerobic superiority.
- The top ranks rest on the thinnest evidence. The pain network has 4 aerobic-vs-usual-care trials (203 participants) and 5 mind-body (244), against 32 mixed (2702) and 24 strengthening (2093) (Goh, Persson, Stocks, Hou, Welton, et al., 2019). The authors concede «the estimates for aero- bic, mind–body and flexibility/skill exercises were open to considerable uncertainty with wide credibility intervals as the number of studies were small» — then add that three ranking methods «were generally in agreement, supporting the trend observed» (Goh, Persson, Stocks, Hou, Welton, et al., 2019). Agreement among rankings drawn from the same estimates does not add independent support.
- The ranking depends on how types are drawn. An earlier NMA (Uthman 2014, not held, reported via Goh) with three types and their combinations found «Strengthening exercise yielded the largest effect size for pain outcomes, whereas a combined intervention of strengthen- ing, flexibility and aerobic exercise had the largest effect size for function.» (Goh, Persson, Stocks, Hou, Welton, et al., 2019) The two NMAs differ in classification, trial count (103 vs 60) and comparator (Uthman allowed active non-exercise controls), so a best mode read is not stable across category schemes.
Certainty: low for any between-type ranking (the wiki’s judgment; the paper does no GRADE/CINeMA rating). «Physician and participant blinding was not achieved in any study (ESM Appendix 6).» (Goh, Persson, Stocks, Hou, Welton, et al., 2019) «Egger’s statistical test is suggestive of publication bias (p < 0.05) for all outcomes except QoL (ESM Appen- dix 3).» (Goh, Persson, Stocks, Hou, Welton, et al., 2019) Model fit was poor for pain, performance and QoL, and the between-study SD (0.25 to 0.74) is the same order as the between-type gaps. On the other side, «No disagreements were found between direct and indirect evidence (ESM Appendix 5) or between estimates from different study designs.» (Goh, Persson, Stocks, Hou, Welton, et al., 2019)
Hip and severity, per type. In hip OA only strengthening and mixed are estimable for pain, function and QoL, and none of those CrIs excludes zero (pain: strengthening 0.53, -0.74 to 1.80; mixed 0.12, -0.36 to 0.62); the one hip estimate that does is aerobic on performance, 0.81 (0.23 to 1.42) (Goh, Persson, Stocks, Hou, Welton, et al., 2019). This reproduces the pairwise MA’s hip uncertainty above, from the same project’s trials, so it is a refinement, not independent confirmation. The awaiting-joint-replacement subgroup likewise shows smaller, uncertain per-type pain estimates (strengthening 0.46, -0.28 to 1.18). The authors could not tell «whether the observed differences between exer- cises would persist in the longer term» (Goh, Persson, Stocks, Hou, Welton, et al., 2019).
Gaps and open threads (type-G)
- Exercise-MODE ranking — now held, and it ranks types only weakly. The companion NMA (section above) cashes the gap this page used to carry: among single-type modes no contrast is shown on pain or function (point estimates favour aerobic/mind-body; low certainty); the mixed category is the one distinguishable loser on pain, against the two thinnest nodes only. Still open: aquatic is not a separate node, the comparison is 8-week only, and hip OA is under-powered. Earlier partial answer from Luan -> Stationary Cycling for Knee Osteoarthritis: a pairwise MA finds stationary cycling neither superior nor inferior to the other modes it was compared with (swimming, treadmill, Tai Chi, Baduanjin) on every WOMAC/KOOS outcome — so the working answer no mode clearly dominates on OA symptoms had direct backing (choose mode by preference/cost/impact-tolerance) before the all-modes NMA landed; the NMA agrees for the single-type modes (superseded 2026-10-06: “the full all-modes relative-efficacy NMA is still owed”). (The quote and the per-outcome nulls live on that page, cited to Luan; this is the cross-page structuring, not a fresh Luan extraction here.)
- Is exercise good OR bad for the knee? The causation companion — LANDED. This page is the treatment side (exercise relieves established OA); the causation side — does running/loading CAUSE or accelerate OA — is now held on -> Knee Osteoarthritis Incidence and Risk Factors (Alentorn-Geli 2017 MA). It resolves the apparent paradox without contradicting this page: recreational running is joint-safe to protective (knee OR 0.83), competitive/high-mileage running and a sedentary lifestyle both sit on the high arm of a U-shape, and occupational loading raises incidence -> The Physical Activity Paradox. The two sources are a DISTINCTION, not a tension (different question — treatment of established OA vs causal association in runners; parameter table on the incidence page): graded exercise relieves an arthritic knee and recreational loading does not cause OA in a healthy one. The obese, previously-injured novice remains the explicitly-untested harm stratum — favour lower-impact modalities there -> Exercise Interventions and Sports Injury Prevention.
- The IPD gap. Every determinant needs individual-patient-data confirmation; the study-level
signals are hypotheses.
G (needs aggregation)-adjacent: the within-person effect modification is a quantity this aggregate MA structurally cannot compute. - No long-term disability trajectory. The outcomes are 8-week-to-18-month pain/function; no source here grades exercise against a realized multi-year disability or joint-replacement outcome — the loop stays open (R1).
Self-critique [run 2026-10-04, before commit]
- Not laundered from one source restated. The beyond-summary moves: (1) the provisional-C induction of the characterize the exercise lever sub-question distinct from the nucleus’s lever-ranking; (2) the type-E/F read that Goh’s direct pain/function effect independently backs and bounds the nucleus’s surrogate-based exercise claim; (3) the obese-stratum transportability bridge (BMI null + obesity-is-the-entry-stratum) tying this to the incidence orbiter; (4) the durability-as-contingent- stimulus contrast with weight-loss structural leverage. Each is tagged; the estimates stay Goh’s.
- Not overclaimed. Confidence MODERATE: gold MA with large N and direct patient-important outcomes (upgrades), but significant small-study bias on three of four outcomes, honest-subset shrinkage, and determinants that are ecological / hypothesis-generating by the authors’ own framing. The determinant directions are flagged as route-(b) candidates, not confirmed modifiers; the BMI null is given more weight than the positive subgroups.
- No fabricated mode ranking. The dispatch framed this as a mode-NMA; the paper is a pairwise exercise-vs-usual-care MA with patient-determinant subgroups. The mode comparison was left as an explicit gap, not invented (superseded 2026-10-06: the companion NMA is now held, section above).
- Coherence, not validity (R1): the page reports effect sizes, their decay, and their certainty; it does not assert a realized long-term disability benefit — the open loop is named.
Self-critique [run 2026-10-06, after the NMA addition]
- Blind re-critique: PASS-WITH-FIXES, all applied — the single-type contrasts are stated as not shown to differ (with the authors’ own hierarchy reading and the 0.4-SMD pain gap), not as equivalence; the mixed-loses steer carries the small-study alternative and the mixed vs strengthening null; the pooled result is scoped knee-dominant with the hip exception; the trend test is flagged as ordered by the same estimates; the BMI null is downgraded to no evidence of modification. The NMA is labelled type-F on the pairwise MA (same project, 8/9 authors), never E.