Nucleus of the weight-management cluster. Cochrane 2022: 61 RCTs, 6925 randomised, search to
June 2021. Synthesis mode: aggregative — a common metric and an identical hypothesis, so the answer
is a pooled magnitude. Most questions in this wiki are configurative; this one is not.
The answer, in one line
Both diets produce weight loss the review treats as approaching clinical meaningfulness — its own phrasing is conditional, that reductions “from around the middle of these ranges (about 4 to 6 kg in most trials) would start to become clinically meaningful” — while the difference between them is about 1 kg, which the review judges not clinically important.
Read the units carefully; they are easy to conflate. Across both arms, trials achieved reductions of 12.2 to 0.33 kg short-term, and the review’s bar is “a loss of at least 5% of initial weight”, i.e. “reductions from around the middle of these ranges (about 4 to 6 kg in most trials) would start to become clinically meaningful.” That 4-6 kg is a within-arm loss from baseline, achieved by both diets — not a between-group threshold. The ~1 kg figures below are between-group mean differences and must not be set against it as though they were the same quantity. (Naude et al., 2022)
The magnitudes
Weight change, low-carbohydrate minus balanced-carbohydrate. Assumed risk = the range of change observed in the balanced-carbohydrate arms.
| Population | Timepoint | Effect | N (studies) | Certainty |
|---|---|---|---|---|
| Without T2DM | 3 to <12 mo | MD 1.07 kg lower (1.55 to 0.59 lower) | 3286 (37) | Moderate |
| Without T2DM | >=12 mo | MD 0.93 kg lower (1.81 to 0.04 lower) | 1805 (14) | Moderate |
| With T2DM | 3 to <12 mo | MD 1.26 kg lower (2.44 to 0.09 lower) | 1114 (14) | Moderate |
| With T2DM | >=12 mo | MD 0.33 kg lower (2.13 lower to 1.46 higher) | 813 (7) | Moderate |
Other outcomes at >=12 months, without T2DM: DBP 0.09 mmHg lower (1.29 lower to 1.12 higher); LDL 0.04 mmol/L higher (0.05 lower to 0.12 higher). With T2DM: HbA1c 0.14% lower (0.38 lower to 0.10 higher); LDL 0.12 mmol/L higher (0.03 lower to 0.26 higher). (Naude et al., 2022)
Why this is a no meaningful effect verdict and NOT an “insufficient evidence” one
This is the telos’s four-evidence-states rule, and the review is the cleanest worked instance the corpus holds. Each between-group difference sits 2 to 4.5 times below the review’s stated importance bar for that outcome:
| Outcome | Between-group difference | Bar the review states | Ratio |
|---|---|---|---|
| DBP | <0.5 mmHg | ”changes in DBP of greater than 2 mmHg” | >4x |
| LDL (T2DM) | 0.12 mmol/L | ”changes in LDL cholesterol of greater than 0.26 mmol/L” | ~2.2x |
| HbA1c | 0.14% | “changes in HbA1c of greater than 0.5%“ | ~3.6x |
Two things keep this honest. The LDL row in T2DM has an upper bound of exactly 0.26 — the review’s own threshold — so that outcome touches the bar rather than clearing it. And the subgroup estimates below reach -2.71 and -2.29 kg, which are not trivially small.
What converts a small estimate into a positive finding of no meaningful difference is the imprecision judgement, and the review states it outright, four times: “we did not downgrade for imprecision” — because the intervals excluded an appreciable effect in both directions. (Naude et al., 2022)
Pre-specification, split by outcome. The weight bar is protocol-anchored: the Methods carry “five to ten per cent of initial body weight (clinically meaningful)” and “weight loss of at least 5%” is a pre-specified primary outcome with its own SoF row. The DBP, LDL and HbA1c bars appear only in the Discussion — stated, not demonstrably pre-specified. Do not apply one caveat to all four.
Two mechanistic deflators that make ~1 kg smaller than it looks
Both from Authors’ Conclusions, and both are priors the authors bring to interpretation, not findings of this review — it measured no body composition or hydration outcome that could test either.
- The difference is inside the noise band. “These differences are similar to typical ranges of biological weight fluctuations over time… influenced by various factors including level of activity, hydration status, season and medications.”
- Glycogen-linked diuresis is larger than the whole effect, and reverses. “one should also consider the total body water loss (2 to 3 kg) that follows dietary carbohydrate restriction due to diet-induced diuresis from glycogen depletion and production of ketone bodies, which is restored when carbohydrates are eaten again.”
A 2-3 kg reversible water shift fully absorbs a ~1 kg between-arm difference. So the observed difference may not be a fat-mass difference at all — and that is a claim about what the outcome measures, not about how big it is.
The comparator does most of the work
Subgrouping the 37-trial short-term analysis by how the two arms’ energy prescriptions compared:
| Energy prescription | MD (kg) | I2 | Studies |
|---|---|---|---|
| Similar in both arms | -0.48 (-0.85 to -0.11) | 0% | 27 |
| Ad libitum low-carb vs energy-restricted control | -1.84 (-3.07 to -0.62) | 77% | 7 |
| No prescription / unrestricted in both | -2.71 (-4.20 to -1.22) | 13% | 3 |
Test for subgroup differences: Chi2 = 11.67, df = 2, P = 0.003, I2 = 82.9%.
And by depth of restriction — all three subgroups, which sum to the 37 trials:
| Restriction | MD (kg) | I2 | Studies |
|---|---|---|---|
| Very-low-carbohydrate / ketogenic | -2.29 (-3.45 to -1.13) | 51% | 9 |
| Incremental very-low to low | -1.51 (-2.89 to -0.13) | 80% | 4 |
| Low-carbohydrate / non-ketogenic | -0.36 (-0.75 to 0.03) | 0% | 24 |
Test for subgroup differences Chi2 = 11.34, df = 2, P = 0.003, I2 = 82.4%. The ordering is monotone in restriction depth.
The review does not merely decline that reading — it affirmatively denies it, on these exact numbers, and the denial belongs here: “Subgrouping by similarity of energy prescription (Analysis 1.2), extent of carbohydrate restriction (Analysis 1.3), diagnosed cardiovascular event or disease at baseline (Analysis 1.4) or gender (Analysis 1.5) did not suggest important clinical differences in average effects between subgroups, with mean differences in weight reduction… ranging between 0.25 kg and 2.71 kg across the various subgroups.” Naude reads the same spread as clinically unimportant and uses the subgrouping to explain heterogeneity, not to establish a gradient. Reporting this as a reading the review “does not draw” understates a denial as a silence. (Naude et al., 2022)
When the arms are matched on energy, the advantage collapses to under half a kilogram with zero heterogeneity. This is Energy Adjustment and What a Diet Coefficient Means operating on whole diets rather than nutrients: the comparator is chosen by a trial-design decision, and it sets the size of the answer.
Two cautions on reading these subgroups as effect modification (telos route b):
- These were never framed as effect-modification tests. Both subgroupings are pre-specified in the Methods “to explore substantial heterogeneity and the stability of findings in different study subgroups” — a heterogeneity investigation, not an interaction hypothesis. Reading a declined route-(b) claim into them would be reading a defect into correctly-applied method. (Naude et al., 2022)
- The two subgroup tests are not independent. The matched-energy subgroup and the non-ketogenic subgroup share heavily overlapping trial lists, and both collapse to I2 = 0% near null. The larger effects concentrate in the same ad-libitum, high-heterogeneity trials. No meta-regression disentangles them. (inferred from Naude et al., 2022)
The one RCT that pre-specified diet personalization — DIETFITS, null (route-b) [2026-08-04]
The route-(b) cautions above ask whether a low-carb estimate should be personalized by a biomarker. DIETFITS (Gardner 2018) is a large RCT purpose-built to answer exactly that, and its distinct contribution over Naude/Ge — which pool average effects, not interactions — is a direct, pre-specified, powered test of personalization by genotype and insulin.
- On weight it is a component, not an independent witness. n=609 non-diabetic adults, healthy-low-fat −5.3 kg vs healthy-low-carb −6.0 kg, between-group 0.7 kg [95% CI −0.2 to 1.6], NS — a single-RCT instance of this page’s pooled near-equivalence, and DIETFITS sits inside Naude’s search window (to June 2021), so on weight it is pooled within the nucleus estimate, not a separate route. [@gardner2018]
- The distinct value is the interaction nulls. Two candidate effect modifiers were pre-specified as primary hypotheses and powered ~90%: a 3-SNP genotype pattern (diet x genotype P=.20) and insulin secretion INS-30 (diet x INS-30 P=.47). Both null: «neither of the 2 hypothesized predisposing factors was helpful in identifying which diet was better for whom.» [@gardner2018]
- So this is the positive-evidence-of-effect-modification the route-(b) rule demands — and it is absent for these two candidate modifiers. It directly instantiates personalize beyond the population estimate only on positive evidence of effect modification (telos route b): mechanistic plausibility for an insulin-secretor subgroup was not enough; the powered test dissolved it. The mechanistic framing (this is the CIM’s surviving subgroup limb) is held on What Drives Fat Gain - Energy Balance vs the Carbohydrate-Insulin Model.
- Bounds (symmetric standards). The null holds when both arms emphasise diet quality (Gardner’s own boundary condition), and is confined to non-diabetics and the INS-30 index — some smaller trials reported an interaction on fasting insulin, uncovered by this result.
- LDL corroboration. DIETFITS’s low-carb arm raised LDL, and the between-group change significantly favoured low-fat (~5%), while HDL rose more and triglycerides fell more on low-carb — matching the atherogenic-lipid signal this page already holds -> LDL ApoB and Cumulative Exposure. [@gardner2018]
[inferred from @gardner2018]
What no trial measured
“None of the included trials reported on the following outcomes for any of the four comparisons: all-cause mortality, cardiovascular mortality, non-fatal myocardial infarction, non-fatal stroke and diagnosis of T2DM.”
And: “No trials reported on outcomes after two years, and 60% of trials had a duration of six months or less.” (Naude et al., 2022)
Cardiovascular mortality is a “not measured” row in all four summary-of-findings tables — the only hard patient-important endpoint the authors carried into them, empty in every stratum. So every reported outcome here is a surrogate except constipation and participant-reported adverse effects. Per the expectancy test this is unprobed, not disproved, and the authors say why: long follow-up in diet trials is impractical.
Decision relevance
- For weight loss alone, the macronutrient split is close to irrelevant at the population level. The decision moves to adherence, cost and preference — not to carbohydrate percentage.
- A claimed low-carb advantage should be met with one question: what was the comparator’s energy prescription? Matched-energy trials show -0.48 kg; ad-libitum-versus-restricted trials show -1.84. The page deliberately assigns no attribution fraction to that gap, because the energy-prescription and restriction-depth subgroups overlap heavily and nothing here disentangles them.
- Early rapid loss on carbohydrate restriction is partly water and reverses on refeeding. Someone judging a diet by week-two scale change is reading a hydration signal.
- Nothing here speaks to mortality or cardiovascular events, in either direction.
Broadened to the whole named-diet field by Ge 2020 (F, shared-evidence — NOT independent). The
pairwise near-equivalence here generalises: a 121-RCT network meta-analysis across 14 branded diets and
three macronutrient patterns finds low-carbohydrate and low-fat near-identical on 6-month weight (4.63 v
4.37 kg vs usual diet) and between-diet differences «typically small to trivial», with weight benefit
decaying ~1.5 kg per diet and cardiovascular risk-factor gains «largely disappear[ing]» by 12 months
-> Named Diet Programs Compared. This corroborates the macronutrient-split-is-close-to-
irrelevant verdict from a network design, but is not independent backing — Ge pools the same
low-carb-vs-low-fat RCT class (e.g. Bazzano 2014, Yancy 2004) and cites the earlier Naude 2014 MA, so a
shared missing-trial or confounding structure would move both. Ge is a corroboration line here, not a
sources: entry (its distinct extraction lives on the paired page).
(inferred from Naude et al., 2022)
Where it stands relative to guidance
The review’s findings sit comfortably inside prevailing guidance, and it says what that guidance permits rather than asserting agreement with it: “Current dietary guidance allows for flexibility in the proportion of macronutrients, including a wide range of carbohydrate intakes, with greater emphasis on quality over quantity and on total dietary patterns over single nutrients.” (Naude et al., 2022)
Recorded because convention held here is a reportable finding under the telos, not a non-result — and because agreement reached with better warrant defeats the guidance null as surely as divergence.
A guidance body reached the same verdict — NICE NG246 (2025) [NICE Revisit 2026-07-29]. Appraising
the trial literature for its weight-management guideline, NICE found low-carbohydrate diets «did not
result in improvements in weight, BMI, waist circumference, or HbA1c, relative to conventional (usually
low-fat) diets, apart from a small improvement in waist circumference for females»
(National Institute for Health and Care Excellence, n.d.) — and left low-carb without a recommendation,
for it or against it (Diets for Weight Loss - What NICE Recommends). This corroborates the
no-meaningful-superiority verdict from a second guidance family. Two refinements it adds:
- Durability, not just magnitude. For very-low-carb, NICE found weight loss «did show a reduction in weight at 1 year, but this was not sustained at 2 years» — the ~1 kg (or its VLC amplification) fades, which is the maintenance-phase caution the wiki otherwise holds only mechanistically (cf. the rate/maintenance gap, Challenge #20).
- Corroboration, NOT independent backing. This is F, not
[E-independent]: NICE is a guideline appraising the same RCT base this page’s Cochrane review (Naude) pools, so the convergence is shared-evidence agreement, not a second independent route — it raises confidence modestly and defeats the guidance null (agreement with better warrant), but adds no independent witness.
Paired with Goldenberg 2021 — the two reviews DISAGREE about depth of restriction
Carbohydrate Restriction and Type 2 Diabetes Remission asks a different primary question — remission in people who already have T2D. But both report weight change at ~6 months in kg, and on the sub-question of whether deeper carbohydrate restriction produces more weight loss they point in opposite directions:
Parameter table (op-weave step 2a — built before the prose below, and it is what stops this section from asserting a false gradient):
| Parameter | Naude 2022 | Goldenberg 2021 | Same quantity? |
|---|---|---|---|
| Outcome | change in body weight, kg | change in body weight, kg | yes |
| Timepoint | 3 to <12 months | 6 months | near enough |
| Population | without T2DM (Comparison 1) | entirely T2D | NO |
| Deep-restriction band | ketogenic, ”<=50 g per day or <10% of total daily energy intake” | very low, “<10% daily calories from carbohydrates or <50 g/d” | yes |
| Admission threshold for the pooled set | <45% TE or <=150 g/day | «130 g/day or less than 26% of calories» | no — Goldenberg’s pooled set is NARROWER, but its deep band matches Naude’s |
| Reference (shallow) band | non-ketogenic, ”> 50 g to 150 g per day or < 45% of total energy intake" | "less restrictive”, “between 10% and 26% of calories from carbohydrates” | NO — Goldenberg’s shallow band is deeper than Naude’s |
| Deep vs shallow result | ketogenic -2.29 (-3.45 to -1.13) vs non-ketogenic -0.36 (-0.75 to 0.03), P=0.003 | very-low -1.05 (-2.27 to 0.17) vs less-restrictive -5.22 (-8.33 to -2.11), P=0.01 | no — same deep band, different reference |
The deep-restriction bands are near-identical (Naude ”<= 50 g per day”, Goldenberg “<50 g/d” — inclusive vs exclusive) and the outcome and unit match. But two cells fail, not one.
- Population — Naude’s Analysis 1.3 is without-T2DM, Goldenberg’s is entirely T2D.
- The reference band — and this one was missing from the table entirely, which is what let the they point OPPOSITE ways verdict stand. A subgroup contrast is deep-minus-shallow; if the shallow arms differ, the two contrasts are not the same quantity even when the deep arms match. Goldenberg’s reference band (10-26% of calories) sits inside what Naude counts as restricted at all, so its “less restrictive” comparator is closer to Naude’s ketogenic arm than to Naude’s non-ketogenic one.
So the direction clash is not established. It may still be real, but this table cannot show it.
The inviting error here is a monotone dose-response ladder (-0.36 -> -1.07 -> -2.29 -> -3.46) built by treating Goldenberg’s whole set as the deepest rung. Two things falsify it. Goldenberg’s inclusion threshold is <26% energy or <130 g/day, which is LOOSER than Naude’s ketogenic band of <=50 g/day or <10% energy — so the “deepest” rung was actually the shallowest admission criterion. And Goldenberg’s own credibility-tested subgroup decomposes its -3.46 kg the other way: the pooled figure is driven by its less restrictive trials, not its deepest ones.
Why this is NOT filed as a tension — the scope check fires twice. Naude’s Analysis 1.3 is from its without-T2DM comparison; Goldenberg’s population is entirely T2D. Under the not-joined checks, two claims holding at different scope are not opposed until matched — and the reference-band mismatch above means the estimands do not match either.
The resolving test named here previously does not exist. Naude’s with-T2DM comparison (Analyses
3.1-3.25) runs sensitivity analyses on risk of bias, attrition and funding source — there is no
restriction-depth subgroup and no energy-prescription subgroup anywhere in it, though the review does
record the bands for those trials (“very low (<= 50 g per day) (n = 5) to low (> 50 g to 150 g per day
or < 45% of total energy intake) (n = 11)”). Checked directly in the analysis list; the read that was
AWAITS-ed is unfulfillable from this source. Resolving the clash needs trial-level data or a review
that subgroups restriction depth within T2D, not another pass over Naude.
Goldenberg supplies a candidate reconciliation if it does become one — adherence: among very-low-carbohydrate diets to which patients were highly adherent, weight loss was larger (-4.47, -8.21 to -0.73). On that reading, depth and adherence pull against each other and the pooled estimate is their net.
What survives regardless, and it is the durable part: a low-carbohydrate weight estimate is uninterpretable without both design parameters — the comparator’s energy prescription and the carbohydrate threshold — and the two reviews use different thresholds while both being called “low-carbohydrate”. Note also that Goldenberg admits wait-list and no-intervention comparators while Naude requires the control to be a balanced-carbohydrate weight-reducing diet, so comparator design is a live rival explanation for any between-review gap.
Szczerba 2023 (T2D umbrella) — the HbA1c verdict flips with the comparator
The Cochrane head-to-head above gives low-carb vs a balanced-carbohydrate diet an HbA1c of only 0.14% lower in T2DM — a no-meaningful-effect verdict. Szczerba’s gold umbrella grades low-carb (<26%E) vs a usual/higher-carb control at −0.47% HbA1c (−0.60 to −0.34, n=17 RCTs, GRADE high) — clinically meaningful. These do not contradict: the comparator differs, which is exactly this page’s thesis (the comparator does most of the work). (Szczerba et al., 2023)
| Parameter | Naude 2022 (Cochrane) | Szczerba 2023 (umbrella) | Same quantity? |
|---|---|---|---|
| Outcome | HbA1c, T2DM | HbA1c, T2D | yes |
| Comparator | balanced-carbohydrate weight-reducing diet | usual / higher-carb control | NO |
| Depth | «low-carb» range incl. shallow | <26%E (deep) | NO |
| Effect | 0.14% lower (0.38 lower to 0.10 higher) | −0.47% (−0.60 to −0.34) | verdicts differ because comparator + depth differ |
Two further Szczerba findings on carbohydrate restriction in T2D, held in full on the nucleus Diets for Weight Management in Type 2 Diabetes:
- A monotone dose-signal (no knee shown). A 10% carbohydrate decrease moves HbA1c only −0.11% (not clinically meaningful) vs −0.47% at <26%E; weight loss «was greater in interventions with low (<26%) or very low (<15%)… than in… moderate (<45%) carbohydrate intake» — deeper buys more, consistent with the domain default that every reduction pays until a knee is located. The estimates may be under-stated (poor adherence «especially for low carbohydrate and ketogenic diets»). (Szczerba et al., 2023) (Szczerba et al., 2023)
- Deprescribing (the firmest non-surrogate finding). Low-carb (<26%E) «reduced the use of drug treatments by an additional 24 per 100 individuals (risk difference 0.24, 0.12 to 0.35;… moderate certainty)». Also GRADE-high: triglycerides −0.30 mmol/L; HDL +0.06 (moderate). (Szczerba et al., 2023)
Limits
- ~40% of included trials had food/diet industry funding. A pre-planned sensitivity analysis removing them found no consistent direction: the short-term non-diabetic estimate moved slightly away from null (-1.07 to -1.20), the long-term estimate toward null and to non-significance (-0.93 to -0.62). So the honest reading is no detectable funding effect, not funding inflated the low-carb advantage.
- Non-reporting bias is conceded on both legs. 11 trials’ weight data were unobtainable in usable format; two of three funnel plots “suggest that smaller studies may be missing”. In all three comparisons the fixed-effect estimate sits closer to null than the random-effects one — the review reads this as reassuring, and the consistent one-way pull is equally available as a small-study signal.
- The defence against non-reporting bias is agreement with other reviews — which draw on the same trial literature and would inherit the same missing trials. Volume of agreement is not independence of backing.
- Outpatient settings in high-income countries except for one trial in China; nearly half were run in the USA.
- The review reports a stratum where caution is warranted, and it belongs in any recommendation drawn from this page: “In people with lipid disorders and variability with atherogenic lipoprotein response, caution in recommending low-carbohydrate and consequent high-fat diets is warranted.” (Naude et al., 2022) That is a telos route-(c) contraindication stated by the source.
- Direction, stated plainly: every subgroup and stratum point estimate favours low-carbohydrate. The verdict is not meaningfully different, not no difference in any direction.
- Cohort data on hard outcomes at the extremes — PURE 2017 (partially cashes the AWAITS below).
Dehghan (135 335 adults, 18 countries) speaks to what this
RCT review structurally cannot — hard outcomes across the intake range — and it cuts against both
poles, not one: higher carbohydrate (spline rise above ~60%E) associated with higher total
mortality (Q5 vs Q1 HR 1.28 [1.12-1.46]), yet «the absence of association between low carbohydrate
intake (eg, <50% of energy) and health outcomes does not provide support for very low carbohydrate
diets», with «moderate intakes (eg, 50-55% of energy)… more appropriate than either very high or
very low». So PURE licenses neither a high-carb nor a very-low-carb optimum.
(Dehghan et al., 2017)
- Two discounts before it is used. PURE is observational (mortality, not the weight outcome this page pools) and its high-carb signal is confounded by income — the highest-carb quintiles are the poorest, and «carbohydrate consumption in low-income and middle-income countries is mainly from refined sources» (it was «unable to quantify separately the types of carbohydrate (refined vs whole grains)»). So the high-carb harm is largely a refined-carb / poverty signal, not carbohydrate per se -> Is the Food Category Doing Any Work, The U-Shaped Association Artifact. It is grounding for «harm at the extremes», not a clean dose-response on carbohydrate. (Dehghan et al., 2017)
- A second cohort traces the complementary arm — Seidelmann 2018 (ARIC + 8-cohort MA)
[2026-08-05]. Where PURE populates the high-carb (right) arm, Seidelmann’s ARIC (mean 49%E) supplies the low-to-moderate (left) arm, and the two overlay into one U-shaped carbohydrate->mortality curve with nadir 50-55%E and both extremes elevated (pooled HR «1·20 … for low carbohydrate consumption; 1·23 … for high») (Seidelmann et al., 2018) — so cohort data now cut against both poles from both sides. Crucially for this page’s low-carb question, the low-carb arm’s mortality depends entirely on the replacement source: «mortality increased when carbohydrates were exchanged for animal-derived fat or protein (1·18, 1·08-1·29) and mortality decreased when the substitutions were plant-based (0·82, 0·78-0·87)». (Seidelmann et al., 2018) So a low-carb weight strategy is not mortality-neutral-by-default: an animal-based low-carb pattern tracks higher mortality, a plant-based one lower — the source of the substituting fat/protein is the lever, not the carbohydrate percentage. Observational (FFQ), weak reverse-causation checks only, no MR -> The U-Shaped Association Artifact, Dietary Protein and Mortality.- The larger pooled version — Qin 2023 (41-cohort SR+MA, to March 2022), a type-F refinement of
Seidelmann, NOT independent corroboration
[2026-08-19]. Qin re-pools Seidelmann’s ARIC AND Dehghan’s PURE among its all-cause cohorts and explicitly cites Seidelmann as the antecedent it agrees with (its finding «consistent with a previous meta-analysis revealing a U-shaped association … [17]», adding that Seidelmann «only included 7 prospective studies») (Qin et al., 2023) — so it bounds the U with a larger pool, it does not independently confirm it (shared studies -> F, never[E-independent]). Three things it adds: (i) it firms the HIGH-carb arm — highest-vs-lowest total carbohydrate RR 1.10 (1.03-1.17) for CVD, 1.20 (1.08-1.34) for stroke, per-5%E 1.02 (1.00-1.04) CVD / 1.04 (1.01-1.06) stroke (Qin et al., 2023); (ii) the all-cause harm is fragile — RR 1.07 (1.00-1.14), J-shaped (Pnon-linearity 0.008) but «not robust in the sensitivity analysis», going non-significant on removing any of Seidelmann, Dehghan, McKenzie or Frisoni, and the per-5%E all-cause slope is null (Qin et al., 2023); (iii) it re-frames the axis — «shifting the focus … from carbohydrate quantity to carbohydrate quality» (Qin et al., 2023), the quality signal (higher CQI -> lower mortality) being protective but a GAP (3 studies, unpoolable). Crucially for this page: Qin does NOT decompose the substitution (no animal- vs plant-replacement) and runs no MR/ referent-correction — so it is a high-carb-HARM claim, silent on whether low-carb benefits, and the low arm stays exactly as unadjudicated as Seidelmann’s. Still observational -> does not cash the randomised-designAWAITS. -> The U-Shaped Association Artifact, Is the Food Category Doing Any Work.- The LCD-SCORE companion — Qin 2023b (38-cohort SR+MA, to July 2023), a type-F sibling of the
quantity estimate, NOT independent-E
[2026-08-26]. The same Qin group (5 shared authors incl. lead Qin + senior Bo) pooled the largest set yet on the composite low-carbohydrate-diet (LCD) score — a Halton-type index (low-carb + high-protein + high-fat, source-agnostic), a different exposure from the %E-carbohydrate axis above. Same group + shared cohorts (NHS/HPFS-Fung, NIPPON-Nakamura, Lagiou, Nilsson also sit inside Seidelmann’s pool) -> type-F, never[E-independent]; no confidence lift. «the highest LCD score was compared with the lowest one and the pooled RRs (95% CIs) were 1.05 (0.96, 1.14; I2 = 65.1%; n = 13) for CVD, 1.43 (1.18, 1.72; I2 = 25.4%; n = 3) for CHD, 0.93 (0.81, 1.06; I2 = 0.0%; n = 2) for stroke, 1.03 (0.96,» (Qin, Suo, et al., 2023) «1.10; I2 = 86.6%; n = 13) for all-cause mortality and 1.09 (0.99, 1.19; I2 = 65.1%; n = 10) for cardiovascular mortality.» (Qin, Suo, et al., 2023)-
The decision-relevant refinement: hard-outcome harm is CHD-specific. «lowest LCDs was associated with 43% increased risk of CHD. No significant association for total CVD, all-cause or cardiovascular mortality was observed.» (Qin, Suo, et al., 2023) So the largest composite-LCD-score pool does not reproduce a mortality signal at all — CHD (1.43) is the only significant arm; all-cause (1.03), CV mortality (1.09), total CVD (1.05) and stroke (0.93) are null. The CHD arm is also the thinnest (n=3 studies) and largest-effect — so it carries the least weight even as it is the only signal. Someone told «low-carb raises mortality» should know the composite-score literature, pooled largest, bounds that to null all-cause / CV-mortality and localizes hard-outcome harm to CHD (on 3 cohorts). (inferred from Qin, Suo, et al., 2023)
-
Why this is a distinct quantity, not a contradiction of Seidelmann (op-weave 2a parameter table). No cell licenses setting Qin-LCD’s null all-cause (1.03) against Seidelmann’s low-carb arm (1.20) as the same quantity:
Parameter Qin 2023b (LCD score) Seidelmann 2018 (%E) Qin 2023a (%E quantity) Same quantity? Exposure composite LCD score, source-agnostic %E carbohydrate total carbohydrate %E NO Contrast highest vs lowest score <30-40%E vs 50-55%E nadir highest vs lowest %E NO CHD 1.43 (1.18-1.72) not pooled 1.10 (0.98-1.24) null NO — score loads animal protein/fat; %E does not All-cause 1.03 (0.96-1.10) null 1.20 (1.09-1.32) low-arm 1.07 (1.00-1.14) fragile J NO The CHD divergence from the quantity axis (1.43 vs Qin 2023a’s null 1.10) is the informative cell: the LCD score’s harm rides on the protein/fat-source loading the composite captures and plain %E-carbohydrate does not — consistent with Seidelmann’s animal-substitution reading, though Qin-LCD cannot prove it. (inferred from Qin, Suo, et al., 2023)
-
The inherited G-gap: no substitution decomposition. Qin-LCD does not split animal- vs plant-replacement [searched: “plant-based” / “animal-based” / “substitution” across chunk 01 — the terms appear only in the Discussion as a conceded limitation, referencing Ghorbani 2023, not as a subgroup]; the score is source-agnostic, so it cannot say whether the CHD signal is the animal-loading Seidelmann isolates. The decision-hinge (animal vs plant) stays exactly where Seidelmann left it. Observational (FFQ), NOS-appraised, no MR -> does not cash the randomised-design
AWAITS. (inferred from Qin, Suo, et al., 2023)
-
- The LCD-SCORE companion — Qin 2023b (38-cohort SR+MA, to July 2023), a type-F sibling of the
quantity estimate, NOT independent-E
- The larger pooled version — Qin 2023 (41-cohort SR+MA, to March 2022), a type-F refinement of
Seidelmann, NOT independent corroboration
AWAITSa source on whether carbohydrate restriction changes hard outcomes at all in a randomised design — PURE and Seidelmann (above) supply the observational cut but not the interventional one. The review points outside its own evidence base to cohort data showing «harm at the extremes of intake» — which cuts against both diet poles rather than favouring either.
Held, not filed
A candidate joined issue exists on Analysis 1.2 and is deliberately not filed yet: this review treats the ad-libitum-versus-restricted asymmetry as a bias to subgroup away, while at least one source in the queue treats the same asymmetry as the mechanism — appetite suppression under carbohydrate restriction, and therefore part of the effect rather than a confound. The wiki’s own layer-3 provisions (judge against the realistic alternative; adherence is part of the effect) do not obviously side with the review. The low-carb-first advocacy position was to be sourced from Feinman 2015, but that source was assessed and dropped as non-gold (a narrative critical review); the contrarian case is taken from the held systematic reviews (Naude, Goldenberg) rather than from its advocates, so no advocacy-voiced tension is filed.
The mechanistic why behind the matched-energy null is now filed as its own tension. The carbohydrate-insulin model predicts a metabolic advantage for carb restriction at equal calories; isocaloric-controlled-feeding evidence (Hall & Guo 2017) refutes that prediction in direction — the small edge runs the other way, toward lower-fat -> What Drives Fat Gain - Energy Balance vs the Carbohydrate-Insulin Model (where that evidence is held). This is the metabolic-ward counterpart to this page’s whole-diet matched-energy subgroup (-0.48 kg, I2=0%): a different design reaching the same no carb-specific advantage verdict, though both share isocaloric-comparison logic (convergent, not clean independent backing). The candidate appetite/ad-libitum joined issue held above is the CIM’s one surviving live channel — GL/protein acting on intake, not on metabolism.
Why a small weight-CHANGE difference is a harder measurement than it looks [2026-07-28, Willett ch.9]
This page’s central quantity is a between-group difference in weight change of roughly 1 kg. Willett’s anthropometry chapter says why that is the fragile case, and the reassuring half of his sentence has to come first or the point inverts:
«This degree of reproducibility is far greater than for most biochemical or physiologic measurements and indicates that, with reasonable care, imprecision in measurement of weight is not likely to be a serious issue in most epidemiologic studies. In studies that involve change in weight over months or a few years, precision in measurement is much more critical, as measurement errors contribute twice (at the beginning and the end) and the magnitude of weight changes is usually small compared with differences in attained weight between persons.» (Willett, 2012)
Weight is one of the best-measured variables in nutritional epidemiology. Weight change is not, and the two failure modes are structural:
- Error enters twice. A difference of two measurements carries both measurements’ error, so the error variance on a change is larger than on either endpoint.
- The signal is small relative to between-person spread. The quantity of interest is a few kg against a between-person range of tens of kg.
What this does and does not say about this page’s ~1 kg. It does not impeach the estimate: these were randomised trials with measured (not self-reported) weights, and randomisation plus pooling addresses what unsystematic error does to a mean difference. What it does is explain why the confidence intervals are wide relative to the effect, and why the review’s not clinically important verdict is robust — a 1 kg difference sits inside the noise band of the measurement problem it is estimated through, quite apart from whether it matters to a patient. (inferred from Willett, 2012)
Where it bites harder: any observational weight-change finding, and any self-reported one. Willett records that self-reported BMI runs low — NHANES III mean 25.07 vs 25.52 technician-measured, from underreporting weight (-0.56 kg) and overreporting height (+0.76 cm), two errors compounding in the same direction on the ratio. (Willett, 2012)
What the trials actually delivered — the fidelity seam [2026-07-28]
This page’s ~1 kg rests on 61 randomised trials. The review records an «Extent of intervention fidelity» field for every arm, and that field was never read. A targeted pass over the characteristics tables gives the following.
Counted across the review: 122 fidelity entries, of which 33 (27%) are «NR».
[searched: "Extent of intervention fidelity:" across all 20 Naude chunks; entries classified NR vs reported by their opening token]
So roughly a quarter of trial arms report nothing about whether the assigned diet was followed.
Where fidelity IS reported, it is frequently poor:
«Extent of intervention fidelity: Adherence to the dietary interventions was low for both arms (22% and 29% for LFD and LCD, respectively). Overall, participants were more compliant with the PA compo- nent of the intervention (66% and 61% among those randomised to the LFD and LCD arms, respective- ly).» (Naude et al., 2022)
Others read simply «Adherence was reported as poor», or record participants «excluded during follow-up due to non-compliance».
And the measurement of adherence is not one quantity. Across arms it is variously: a percentage; a dietitian’s grading scale; the difference between reported and target carbohydrate; a count of excluded non-compliers; the authors’ own word («high adherence» claimed as a study strength); and, in one trial, a class-attendance proxy standing in for diet adherence because the diet measure was unavailable:
«Extent of intervention fidelity: Dietary adherence scores could not be calculated due to the multiple dimensions of the intervention programme’s recommended goals. In all 4 diet groups, 85% to > 89% of participants attended at least 75% of their assigned classes (>= 6 of 8).» (Naude et al., 2022)
What this does and does not change
- It does NOT impeach the ~1 kg. These are randomised trials analysed by intention to treat, and under ITT poor adherence is part of the effect being estimated, not a bias in it. The estimate is sound for what it estimates.
- It sharpens what that is: the effect of ASSIGNING a diet, at the adherence these trials achieved. That is the decision-relevant quantity for someone choosing between two sets of dietary advice — and it is not the effect of eating the two diets.
- The load-bearing consequence is that the review cannot separate two explanations of the null. Low-carbohydrate is no better than balanced-carbohydrate and neither diet was followed well enough to tell predict the same 1 kg. Separating them needs a per-arm exposure contrast, and with 27% of arms reporting no fidelity at all and the rest measuring it incommensurably, this review cannot supply one. (inferred from Naude et al., 2022)
- It strengthens rather than weakens the not clinically important verdict for a chooser. If advice-as-delivered moves weight by ~1 kg, that is what advice does in practice, whatever the diets would do if followed.
The same structure the corpus already holds on sugars. Free Sugars Intake records that an advice trial estimates the advice-plus-adherence package, and that where the exposure contrast fails the null is uninformative about the exposure. This is that principle instantiated across 61 trials with the fidelity field counted. -> Framing a Decision Question
One data point deliberately NOT used. Two arms of a single trial record compliance means of 94% and 9%. A 9% mean with an SE of 4.8% is possible, but so is an OCR truncation of 90% — and no recovered-tables sidecar exists for this source to check against, so the digit is unverifiable. An extreme differential-adherence example was available and has been left out, since the finding above does not need it.