Ge 2020 (BMJ) is a network meta-analysis of 121 randomised trials (21 942 overweight/obese adults) placing 14 branded named diet programmes (Atkins, DASH, Zone, Ornish, Mediterranean, Weight Watchers, Jenny Craig, Volumetrics, Slimming World, South Beach, Biggest Loser, Portfolio, Rosemary Conley, paleolithic) and three macronutrient patterns (low-carbohydrate, low-fat, moderate-macronutrient) into one network, ranked on weight and five cardiovascular risk factors at 6 and 12 months. It is the wiki’s comparative-effectiveness island for the whole named-diet field, and it answers the “which diet wins” question directly rather than one pairwise contrast at a time. Search to September 2018; GRADE-rated throughout; Johnston/Guyatt (McMaster) group. (Ge et al., 2020)
Orbiter of the weight-management cluster; the nucleus is
Low-Carbohydrate vs Balanced-Carbohydrate Diets, which holds the low-carb-vs-balanced pairwise case
this page generalises to all named programmes.
One caveat governs the whole page: every outcome here is a SURROGATE — weight, systolic and diastolic blood pressure, LDL, HDL, C-reactive protein. No hard endpoint (mortality, MI, stroke, incident diabetes) is measured, so Ge closes the between-diet and durability gaps, not the hard-outcome one -> Surrogate Outcomes.
The headline: modest benefit vs usual diet at 6 months, then it decays; between-diet differences trivial
Two findings, both decision-relevant.
- At 6 months, all three macronutrient patterns beat usual diet by a similar, pre-specified-important margin. Low-carbohydrate and low-fat are near-identical: «low carbohydrate and low fat diets had a similar effect at six months on weight loss (4.63 v 4.37 kg, both moderate certainty) and reduction in systolic blood pressure (5.14 mm Hg, moderate certainty v 5.05 mm Hg, low certainty) and diastolic blood pressure (3.21 v 2.85 mm Hg, both low certainty)»; moderate-macronutrient (DASH, Mediterranean) is slightly smaller (Ge et al., 2020).
- By 12 months the benefit is largely gone. «At 12 months the effects on weight reduction and improvements in cardiovascular risk factors largely disappear» (Ge et al., 2020) — weight loss vs usual diet falls from 4-5 kg at 6 months to about 3 kg at 12 (down ~1.5 kg per named diet), and the blood pressure and lipid improvements «disappeared almost completely at 12 months» (Ge et al., 2020).
The verdict Ge draws is that the label barely matters: «Differences between diets were typically small to trivial and often based on low certainty evidence» (Ge et al., 2020), and therefore «people can choose the diet they prefer from among many of the available diets … without concern about the magnitude of benefits» (Ge et al., 2020).
Two-tier importance: benefit vs usual diet clears the bar; between-diet difference does not
Ge pre-specified minimal-importance thresholds — «weight loss 2 kg, systolic blood pressure 3 mm Hg, diastolic blood pressure 2 mm Hg, LDL cholesterol 5 mg/dL» (Ge et al., 2020) — so the four evidence states are cleanly separable on this page:
- Benefit (above the bar): each macronutrient pattern vs usual diet at 6 months.
- No meaningful DIFFERENCE (below the bar): one diet vs another. The largest low-carb-vs-moderate- macronutrient gap is weight 1.57 kg (0.86 to 2.29), SBP 1.66 mmHg, DBP 1.33 mmHg at 6 months; the Atkins-vs-Zone weight gap is «only 1.38 kg (95% credible interval 0.15 to 2.62)» with an LDL difference crossing zero (Ge et al., 2020). Even the nominally-largest named diets (Atkins ~5.5 kg, Zone ~4.1, DASH ~3.6 vs usual diet at 6 months) separate from each other by less than the weight importance bar.
- No meaningful effect at all (12 months): the between-pattern differences are «negligible at 12 months» (Ge et al., 2020).
So no named diet is clearly superior is a no-meaningful-difference finding on surrogates, not an insufficient-evidence one — the comparison was run, at scale, and came back near-null. That is a different and stronger statement than we don’t know.
Why the label barely matters, mechanistically: at equal calories the macronutrient source confers no metabolic advantage — isocaloric controlled feeding (Hall & Guo 2017) finds no meaningful body-fat or energy-expenditure edge for carbohydrate restriction, refuting the carbohydrate-insulin model’s metabolic-advantage prediction -> What Drives Fat Gain - Energy Balance vs the Carbohydrate-Insulin Model (where that evidence is held). Ge’s near-equivalence across 14 named diets is the free-living, whole-programme face of that same null. (inferred from Ge et al., 2020)
The one exception: Mediterranean holds at 12 months
Ge is otherwise a clean null on durability, with one break: «Estimated effects at the 12 month follow-up for weight loss and cardiovascular risk factor improvements diminished for all popular named diets, except for the Mediterranean diet», and «Only the Mediterranean diet showed a statistically significant difference compared with usual diet in LDL cholesterol reduction» (Ge et al., 2020). This is a surrogate (LDL) finding, distinct from — but pointing the same way as — the one whole-pattern RCT that shows Mediterranean moving hard events -> Mediterranean Diet and Cardiovascular Events. That two independent designs (a named-diet NMA on LDL; PREDIMED on CV events) single out the same pattern is worth noting, while keeping the outcomes distinct: Ge does not measure events, PREDIMED does not pool LDL here.
Low-carb costs LDL what it buys elsewhere
A refinement the near-equivalence headline hides: on LDL, low-carbohydrate is the worst of the three patterns — it barely lowers LDL vs usual diet (1.01 mg/dL, low certainty, interval crossing zero) while low-fat (7.08 mg/dL) and moderate-macronutrient (5.22 mg/dL) both lower it at moderate certainty; low-carb instead raises HDL (2.31 mg/dL) (Ge et al., 2020). So the patterns are interchangeable on weight and blood pressure but not on the atherogenic-lipid axis — corroborating the LDL-rise-on-low-carb signal the corpus already holds from the Cochrane and remission reviews (Low-Carbohydrate vs Balanced-Carbohydrate Diets; Carbohydrate Restriction and Type 2 Diabetes Remission), and relevant wherever cumulative apoB exposure is the concern -> LDL ApoB and Cumulative Exposure.
The label does not do the work — the plant-vs-omnivore case, kept in scope
Ge directly tests whether the plant-heavy patterns beat omnivorous ones on the risk factors it measures, and finds they do not: «our results showed that … omnivorous based diets (eg, Atkins, Zone) have a similar effect to diets that tend to be higher in plant based foods (eg, Ornish, DASH, Mediterranean)» (Ge et al., 2020).
This is NOT a contradiction of the plant-based case, and must not be filed as one (the not-joined scope check fires): Ge’s own framing bounds it to short-term surrogates — «To the extent that short term results might have implications for long term cardiovascular outcomes, our findings do not support this conclusion» (Ge et al., 2020). The plant-based case in the corpus rests on cohort hard-outcome associations and shared-core component evidence, a different quantity than Ge’s 6-12-month risk-factor deltas. What Ge does establish is narrower and still useful: on the measurable surrogates over a year, the pattern label is not carrying the effect -> Is the Food Category Doing Any Work.
Adherence is the missing term, and it bounds the whole result
Ge concedes it could not measure adherence and reads its estimates accordingly: «adherence to diets was generally not reported, and could have been low, particularly at 12 months. If this is the case, our results describe what is likely to happen for average adherence by patients. Full adherence would probably yield larger effects» (Ge et al., 2020). So the 12-month decay is effect-times-adherence, not the diets’ physiological ceiling — the same structure the corpus holds across 61 trials from the Cochrane review’s fidelity field (Low-Carbohydrate vs Balanced-Carbohydrate Diets) and from advice-trial reasoning (Free Sugars Intake). The decision consequence is that between two diets a person is choosing between two sets of advice at achievable adherence, and the one they will actually sustain wins on the term that dominates.
Ge and Naude corroborate — but they are NOT independent
Ge broadens the pairwise low-carb-vs-balanced result to the whole named-diet network; the two reviews agree, and the agreement is shared-evidence (type-F broadening), not independent type-E backing — they pool the same class of low-carb-vs-low-fat RCTs (e.g. Bazzano 2014, Yancy 2004 are in Ge; Ge cites the earlier Naude 2014 MA as an antecedent), so a common missing-trial or common-confounding structure would move both together.
| Parameter | Ge 2020 | Naude 2022 | Same quantity? |
|---|---|---|---|
| Design | network MA, 121 RCTs, Bayesian | Cochrane pairwise MA, 61 RCTs | different method, same trial class |
| Search date | to September 2018 | to June 2021 | Naude is newer; pools overlap |
| Population | overweight/obese adults | with and without T2DM | overlapping |
| Comparison of interest | low-carb vs low-fat (+ 12 named diets) | low-carb vs balanced-carbohydrate | near enough on the shared arm |
| Weight outcome, low-carb vs low-fat/balanced | 4.63 v 4.37 kg vs usual diet (near-identical) | MD ~1 kg, judged not clinically important | yes — both say near-equivalent |
| Independent backing? | — | — | NO — shared trial pool, F not E |
(Ge et al., 2020) (inferred from Ge et al., 2020)
A large individual RCT in this exact comparison points the same way (corroboration, not a sources entry). DIETFITS (Gardner 2018; n=609, healthy-low-fat vs healthy-low-carb, 12 mo) found near-identical weight loss (−5.3 vs −6.0 kg, between-group 0.7 kg, NS) and, uniquely, pre-specified and refuted diet personalization by genotype (P=.20) and insulin secretion (P=.47) — a route-(b) test Ge’s average-effect network could not run -> Low-Carbohydrate vs Balanced-Carbohydrate Diets, What Drives Fat Gain - Energy Balance vs the Carbohydrate-Insulin Model. Shared-pool (likely inside Ge’s search window), so corroboration, not independent backing.
Ge’s own external corroborations are also shared-standpoint, not independent witnesses: it aligns with the 2014 AHA/ACC/TOS guideline «concluding that evidence was inadequate to recommend any particular diet» (Ge et al., 2020) and with NICE leaving every macronutrient-defined diet without a recommendation (Diets for Weight Loss - What NICE Recommends) — all appraising largely the same RCT base.
The same null holds in the type-2-diabetes stratum (F-broadening, shared evidence)
Ge is general-population. The gold-tier T2D umbrella review (Churuangsuk 2022) reaches the same no-diet-superior verdict within the T2D stratum, at the umbrella level: meta-analyses «do not support any particular macronutrient profile or style over others», and «no one diet type is superior over others for weight management in type 2 diabetes.» (Churuangsuk et al., 2021) It adds the one refinement Ge’s flat network cannot — the energy-delivery format does separate the diets even though composition does not (VLED and formula meal replacement beat self-administered low-energy diets by delivering less energy) -> Diets for Weight Management in Type 2 Diabetes.
Broadening, NOT independent backing (F). Churuangsuk is an umbrella over meta-analyses that pool the
same RCT class Ge’s network draws on; a shared missing-trial or confounding structure would move both, so
this is not a second independent witness on the shared verdict. It is a listed sources: entry because it
contributes two distinct extractions here — the T2D-stratum confirmation and the energy-format distinction
Ge’s general-population network lacks — not merely a corroboration line; its full certainty-graded map
lives on Diets for Weight Management in Type 2 Diabetes.
Decision relevance
- For weight and blood pressure, the choice among named programmes is close to irrelevant. Adherence, cost, and preference decide it, not the brand or the macronutrient split.
- The 6-month scale reading over-promises. Any named diet’s benefit is roughly a third smaller by 12 months, and its risk-factor gains are mostly gone — a durability caution, not a magnitude one.
- If LDL/apoB is the concern, low-carb is the poorer pattern even though it ties on weight and BP.
- Mediterranean is the one label with a residual 12-month signal (on LDL here; on events in PREDIMED), the only case where the pattern name plausibly carries something past 6 months.
- Nothing here speaks to hard outcomes. The between-diet mortality/event comparison has never been run in this literature; the near-equivalence verdict is on surrogates over <=1 year.
(inferred from Ge et al., 2020)
Getting a DIRECTION where the head-to-head hard-outcome trial can’t be run
The between-pattern hard-outcome comparison is structurally unrunnable (the G-gap under Limits): a long-latency head-to-head RCT on mortality/events between named patterns will not happen, and no meta-analysis can pool trials that do not exist. But “no head-to-head MA” is not “no directional guidance.” Three moves extract a defensible direction from the evidence that does exist, and each is a method the corpus already holds — not a loosening of the evidence bar.
1. Emulate the trial you cannot run. Specify the target trial — eligibility, time-zero, the two pattern strategies being contrasted — then estimate its effect from cohort data under identification assumptions made explicit, rather than reading an unadjusted pattern-mortality association off a food-frequency questionnaire. This converts “which pattern is healthier?” from an un-anchored correlation into a stated causal contrast whose assumptions can be checked, which is what lets observational pattern data speak to direction at all. -> The Target Trial (Emulation and the Well-Defined Intervention)
2. Decompose the pattern into well-defined components. A pattern label (“vegan”, “keto”, “Mediterranean”) bundles many versions of treatment — one label names diets differing in fibre, refined sugar, energy density, protein, and total energy — so the label-level contrast is ill-posed before any data arrive. The corpus’s resolution is to name the component and ask which axis carries the difference: Ge’s own network finds the macronutrient label carries almost nothing on surrogates, while the food-category diagnostic shows repeatedly that a nameable sub-component (energy density, cereal fibre, heme, brew method) does the work a food or pattern label gets credited with. So the directional question that is answerable is component-level, not label-level. -> Is the Food Category Doing Any Work, The Target Trial (Emulation and the Well-Defined Intervention)
3. The extremes separate even where the middle does not — and that is itself the finding. Ge establishes the middle of the pattern space (DASH vs Mediterranean vs low-carb vs Zone) as near-equivalent on surrogates, the between-label differences below the pre-specified importance bar — which licenses choosing on adherence and preference rather than agonizing over the brand (ceiling-is-a-finding). That near-equivalence does not extend to the extremes: a whole-food, high-fibre, low-refined pattern versus a highly-refined energy-dense one is a component contrast the decomposition above expects to be load-bearing, and the evidence there is not symmetric with the middle.
The two guards keep this honest, and they cut in opposite directions:
- The confounded-anchor trap. The strong long-cohort anchors that would drive an extreme-vs-extreme ordering (Adventist-type vegetarian cohorts vs a Standard American Diet) differ in far more than diet — smoking, alcohol, activity, adiposity, social cohesion — so a pattern-label benefit read off them is the exact observed-healthy-population-is-not-evidence-for-a-component trap. The benefit may be real; its attribution to the label may not transport. This is why the direction is trustworthy at the component level (where a target-trial emulation can adjust for the confounder set it names) and fragile at the label level. -> Is the Food Category Doing Any Work
- The asymmetry of what is actually studied. The data are not evenly thin across the extremes: the whole-food-mostly-plant end carries many large long-horizon cohorts, while the low-carbohydrate / carnivore end has few long hard-outcome cohorts (recent, small, self-selected). So “we cannot rank the patterns” is itself an average over an asymmetry — one extreme sits nearer insufficient-evidence than the other, and the honest output names which end is data-poor rather than treating all patterns as equally unadjudicated. The long-cohort pattern anchors this needs are a named acquisition gap, not a held finding here.
Net: the directional method turns “no pattern is clearly better” from a shrug into a structured answer — near-equivalent in the middle (act on adherence); component-ordered at the extremes (act on the measurable component, not the label); and one extreme genuinely under-studied — without asserting any hard-outcome pattern ranking the evidence does not license.
Limits
- Surrogate-only, <=12 months. The endpoints are weight/BP/lipids/CRP; the longest follow-up is 12 months, with «considerably fewer trials» and low-to-very-low certainty at that timepoint (Ge et al., 2020).
- Sparse direct comparisons. Of 407 paired named-diet comparisons across six outcomes, only 59 were direct and only 22 had more than one study — most between-diet estimates are indirect, hence the low certainty; publication bias could not be well assessed (Ge et al., 2020).
- Adherence unmeasured (above) — the estimates are the effect of assigning a diet at achieved adherence, not of eating it.
- Selection. Trial participants are a select population, plausibly more so in diet trials where the effect is tied to behaviour.
- Gap (type-G): no hard-outcome between-diet comparison exists in this literature — the durability of any 12-month surrogate change into events is untested, and a long-latency head-to-head between named diets on mortality/events is impractical to run, so this is a structural absence, not a queue item.