This is Is the Food Category Doing Any Work applied to exercise programming. The food version asks whether a category boundary (meat, dairy, “processed”) carries information or averages over a mix that describes nothing. The exercise version asks the same of a modality or delivery-format boundary: when a decision is posed as bike vs pool vs land, supervised vs home, this aerobic mode vs that, does the boundary move the patient-important outcome — or, once both options deliver the targeted training stimulus, is the comparison flat, so the choice is settled by adherence, feasibility, access, cost, and injury-avoidance instead?
The frame already existed for one case: Mind-Body Exercise and Cognition is, in its own words, this diagnostic applied to a modality — the apparent Tai-Chi/yoga advantage for cognition dissolves into a comparator-plus-heterogeneity artifact rather than a mind-body ingredient effect. This page banks the generalization across the batch of modality/format comparisons the fabric now holds.
The pattern — within a goal, the modality axis is mostly flat
Each row is a different modality or delivery choice, tested on its own outcome. The recurring result is small-to-null between-option differences, with the real lever sitting off the modality axis.
| Choice | Outcome | Between-option result | The lever that remains |
|---|---|---|---|
| Stationary cycling vs other exercise modes | knee-OA symptoms | no mode dominates; no between-mode difference detected | adherence, cost, access, impact-tolerance |
| Aquatic vs land-based | balance / walking in older adults | TUGT and 6MWT show no modality difference; at most a small overall-balance edge for water | feasibility/safety for those who cannot train on land (route-e) |
| Supervised vs unsupervised (home) | function / QoL in >=60 y | only knee-extension strength survives (SMD 0.18, 0.07-0.30); rest lost on sensitivity analysis | delivery format is not load-bearing; attendance was equal (81%) |
| Aerobic vs resistance vs combined | body composition in obesity | no modality reached significance for fat loss | the stimulus (any mode raises CRF), not the mode |
| Mind-body vs aerobic/resistance | cognition | apparent mind-body edge is a comparator artifact | the label does no work beyond generic activity |
| Aerobic vs mind-body vs strengthening vs flexibility/skill vs mixed (NMA) | knee/hip-OA pain, function | single types not shown to differ on pain/function (wide CrIs; point estimates favour aerobic/mind-body; pain trend p = 0.01); mixed worse than aerobic and mind-body on pain | preference/adherence among single types; avoid presuming a complex programme is better |
| Endurance vs dynamic resistance vs isometric vs combined (NMA) | SBP (office to ambulatory, mixed) | single types similar; combined beats resistance (-2.98, -5.04 to -0.93), and in >=140 mmHg trials beats every single type | baseline SBP (larger effect in hypertensives); combined where feasible, else adherence -> Blood Pressure Lowering and Cardiovascular Events |
Sources, in row order (each number lives, with its CI and its sensitivity-analysis caveat, on the linked facet page): (Luan et al., 2020), (Deng et al., 2024), (Gómez-Redondo et al., 2024), (O’Donoghue et al., 2020), and for the NMA row (Goh et al., 2019) (Fig. 3: of 40 between-type contrasts, four CrIs exclude zero, all involving mixed or a walking-type performance outcome), and for the SBP row (Naci et al., 2018). The mind-body row is different in kind: Blomstrand’s extracted datum is a pooled mind-body advantage (d = 0.48) (Blomstrand et al., 2023); the verdict that the advantage does no work is the facet page’s correction of it, not Blomstrand’s own conclusion -> Mind-Body Exercise and Cognition.
The decision-change — flip the question
The ordinary framing (“which modality is best?”) presumes the boundary carries a ranking. Across these cases it mostly does not, so the question flips: pick the option the person will actually sustain and is least likely to be hurt doing. This is adherence-is-part-of-the-effect and the route-(e) constraint / route-(c) harm logic made concrete — a sustained lesser-ranked mode beats an abandoned or injurious “better” one, and when the ranking is flat to begin with there is no benefit being traded away. It is also attention-is-an-anti-signal: the modality debate is loud precisely because the effect differences are small and contested.
Two boundaries that keep this honest
Across goals, modality is NOT flat. The flatness holds among options that deliver the same stimulus. Different stimuli produce different adaptations: resistance training builds strength and muscle, aerobic work builds cardiorespiratory fitness, and these do not substitute for each other -> Resistance Training Prescription - Load Sets and Frequency, Measuring and Raising Cardiorespiratory Fitness. The claim is within-goal substitutability, never all exercise is interchangeable.
Some comparisons do carry a real, graded edge — so check per case, do not assume flat. The diagnostic is answered per comparison, exactly as the food version is. HIIT vs Continuous Training for Type 2 Diabetes shows HIIT beating MICT on VO2peak at moderate certainty (+3.37 ml/kg/min, 1.88-4.87) and on HbA1c at low certainty (-0.37%), while fasting glucose, insulin and HOMA-IR show no difference (Liu et al., 2018) — an intensity-structure boundary that is informative for fitness and null for glucose, in the same comparison. Combined training for blood pressure (next paragraph) is a second such case. “Flat” is the modal finding, not a law.
Blood pressure: single types flat, the combination ahead. Naci gives each exercise type’s SBP effect against control: «endurance −4.88, 95% CrI −5.69 to −4.06; resistance −3.50, 95% CrI −4.91 to –2.09; isometric −5.65, 95% CrI −8.21 to −3.13; and combination of endurance and resistance −6.49, 95% CrI −8.17 to −4.82» (Naci et al., 2018). Its type-vs-type league table finds the single types alike, with one exception: «Overall, different types of structured exercise interventions achieved similar reductions from baseline (table 2). One excep- tion was the combination of endurance and resistance training, which was more effective in reducing baseline SBP than dynamic resistance alone (−2.98, 95% CrI −5.04 to −0.93).» (Naci et al., 2018). In the sensitivity analysis restricted to hypertensive trials (>=140 mmHg), the overall exercise effect grows from -4.84 to -8.96, and the combination beats endurance (-4.81, -7.99 to -1.61), resistance (-6.26, -10.67 to -1.90) and isometric (-8.58, -14.68 to -2.45) (Naci et al., 2018); against control the combination reaches -13.51 (-16.55 to -10.45), while isometric’s interval includes the null (-4.92, -10.28 to 0.38) (Naci et al., 2018). In the all-populations analysis the combination’s edge over endurance (-1.61, -3.34 to 0.12) and isometric (-0.84, -3.85 to 2.13) is not significant; only resistance is beaten (Naci et al., 2018). The authors’ text reports the per-type effects against control; the combined-vs-single contrasts in hypertensives are printed only in Table 2’s cells, which the text does not restate. Superseded 2026-10-07: an earlier version closed this paragraph by calling the hypertensive contrasts this page’s reading of Table 2; they are printed table values (tagged above), as on Blood Pressure Lowering and Cardiovascular Events.
Three cautions keep the edge provisional. The type contrasts are network estimates, mixing 30 head-to-head trials with indirect comparison, and the type groups differ in baseline SBP: 134 mmHg endurance, 125 resistance, 129 isometric, 135 combination (Naci et al., 2018). A 10 mmHg baseline gap between resistance and combination trials can produce part of the all-populations edge on its own, since higher baseline SBP means larger drops. The edge survives the >=140 mmHg restriction, which partly controls for that. And the combination rests on few trials of uncertain dose: «For example, only a small subset of studies identified in our review tested the effectiveness of the combi- nation of endurance and dynamic resistance training. Although these types of interventions were often effective, especially in hypertensive populations, their frequency and duration were not consistently reported to determine the optimal formulation and dose of physical activity to maximise its benefit.» (Naci et al., 2018). So the edge may reflect more total training rather than the mix itself. The decision is therefore graded, not flat: among single types, choose on adherence; for a hypertensive person who can do both, the combination is the option the evidence favours. This is a second graded case beside HIIT.
Where the boundary is drawn changes the ranking. The OA network MA is flat among single-type modes, but its mixed bin (more than one core type, or type unspecified) ranks last and is distinguishably worse than aerobic on pain, SMD 0.64 (0.21 to 1.08) (Goh et al., 2019). An earlier NMA with a different scheme (three types plus their combinations; Uthman, reported via Goh, not held) found strengthening best for pain and the all-three combination best for function (Goh et al., 2019). Goh names five differences between the two NMAs — aim, trial count (103 vs 60), classification, comparator (Uthman allowed active non-exercise controls) and outcomes — so classification is one candidate cause, not the proven one. Still, it is an instance of the food-category lesson: a ranking of categories is only as informative as the category boundary. Why mixed ranks last has three live explanations: a heterogeneous bin (partly a bin for poorly-described trials) looking like a worse exposure; the authors’ implementation/adherence hypothesis for complex regimens; and small-study inflation of the thin aerobic and mind-body nodes, the only ones mixed separates from (mixed vs strengthening, the other large node, 0.26, -0.04 to 0.57). The NMA’s lone single-type edge, aerobic over strengthening on walking-type performance tests, is plausibly training-specificity rather than a modality effect.
Relation to the food diagnostic — same three-way cut
Is the Food Category Doing Any Work found the boundary cuts three ways: sometimes it carries the effect, sometimes it averages over a null mix, sometimes it is predictive yet adds nothing actionable. Exercise modality behaves the same way: predictive within a goal only when stimulus differs (across-goal), an averaging-null for most within-goal mode choices, and — mind-body for cognition — a boundary that looks predictive until the comparator is corrected and then adds nothing. The transferable move is the diagnostic itself, not a blanket verdict.
Confidence and the open loop
Confidence: low. The constituent findings are themselves low-confidence (small RCT pools, high heterogeneity, effects that mostly do not survive sensitivity analysis, surrogate outcomes), so the generalization inherits that ceiling. It is believed as a decision rule (choose on adherence/feasibility/ injury when a within-goal modality comparison is flat), not as a proven absence of all modality effects. No operation here grades the rule against realized adherence or function; it is a statement about the current comparative evidence, which is systematically thin on exactly the long-horizon, whole-program outcomes a person cares about.
Self-critique [run 2026-10-06, after the Goh NMA addition]
- Blind re-critique: PASS-WITH-FIXES, all applied — single-type contrasts stated as not shown to differ (authors read a hierarchy; pain trend p = 0.01 over a self-ordered hierarchy), not as equivalence; the mixed-loses reading carries the small-study alternative and the mixed vs strengthening null; older efficacy-equivalent / BMI-independent overstatements softened. The NMA shares a project and 8/9 authors with the pairwise Goh MA, so it is type-F, never independent backing.
Self-critique [run 2026-10-07, after the Naci SBP addition]
- Blind re-critique: FAIL, all fixed. The first draft of the SBP row said Naci reports no type-vs-type
contrast; its league table reports combined > resistance in all populations and combined > every
single type at >=140 mmHg. BP is now the second graded case beside HIIT, with three cautions
(network estimates, unequal baseline SBP by type, few combined trials of unreported dose). Review:
meta/external-audit/is-the-exercise-modality-doing-any-work-selfreview-2026-10-07.md.