BMJ 2021, 23 RCTs, n=1357, including unpublished data from five trials. Low-carbohydrate defined (Goldenberg et al., 2021) <26% of energy or <130 g/day — materially stricter than the <45% of total energy used by broader reviews such as Naude 2022. (Naude et al., 2022)
The answer depends entirely on what “remission” means
| Definition | 6 months | 12 months |
|---|---|---|
| HbA1c <6.5% | RD 0.32 (0.17 to 0.47), RR 1.87 (1.18 to 2.97), 8 RCTs, n=264, Moderate | RD 0.10 (-0.02 to 0.21), 3 RCTs, n=171, Moderate |
| HbA1c <6.5% AND no diabetes medication | RD 0.05 (-0.05 to 0.14), 5 RCTs, n=199, Low | RD -0.04 (-0.16 to 0.09), 2 RCTs, n=126, Low |
This is the finding. The headline 32 per 100 more remissions — an NNT of about 3 — is remission with medication allowed to continue. Under the definition that also requires coming off medication, the effect is never statistically significant at any timepoint, and the 12-month point estimate is negative.
But the stratum matters, and this is where the headline needs qualifying. In trials that excluded insulin users, medication-free remission is significant: RD 0.20 (0.03 to 0.38), NNT 5, against 0.00 (-0.07 to 0.07) where insulin users were included (test for subgroup difference P=0.03). For the medication-permitted definition the split is starker still — 0.51 (0.36 to 0.65) vs 0.14 (0.03 to 0.25), P<0.001. (Goldenberg et al., 2021) So “never significant” holds for the pooled medication-free estimate and not for the non-insulin stratum — which is the stratum in which a low-carbohydrate diet would most plausibly be offered. Per the telos’s layer-2 rule the narrowest credible reference class is the one to quote. But this rescue is a six-month result only (corrected 2026-08-08): the insulin subgroup split is reported at six months, and Goldenberg reports no 12-month insulin subgroup — the 2-3 trials carrying 12-month remission data are too few to split — so the NNT-5 stratum carries no durability estimate, and the pooled 12-month row above (RD -0.04) is the only long-term medication-free read the review supports. (Goldenberg et al., 2021)
And the two definitions do not draw on the same trials. Only 7 of 23 trials (30%) permitted medication reduction and reported usable data, so the medication-free estimates come from a structurally different set of trial designs, not from a stricter reading of the same evidence. (inferred from Goldenberg et al., 2021)
Most benefits attenuate by 12 months — but not all
| Outcome | 6 months | 12 months | Certainty |
|---|---|---|---|
| Weight | MD -3.46 kg (-5.25 to -1.67), 18 RCTs | MD +0.29 kg (-1.02 to 1.60), 7 RCTs | Moderate |
| HbA1c | MD -0.47% (-0.60 to -0.34), 17 RCTs | MD -0.23% (-0.46 to 0.00), 8 RCTs | High -> Moderate |
| Fasting glucose | MD -0.73 mmol/L (-1.19 to -0.27) | MD 0.06 mmol/L higher (-0.37 to 0.48), 6 RCTs | Moderate |
| Medication reduction | RD 0.24 (0.12 to 0.35), 7 RCTs | RD 0.33 (-0.00 to 0.66), 3 RCTs | Moderate -> Low |
Medication reduction and triglycerides go the other way (0.24 -> 0.33; -0.30 -> -0.32), so the source’s own wording is the accurate one: “most benefits diminished at 12 months.”
The weight-advantage decay is a general property of diet trials, not a diabetes-specific one (corroboration, F). Across 121 RCTs of overweight/obese adults, weight loss on every macronutrient pattern and named diet diminishes by 12 months (down ~1.5 kg per diet) and cardiovascular risk-factor gains «largely disappear» -> Named Diet Programs Compared (Ge et al., 2020). That the remission effect here decays in step with its weight advantage is consistent with the reading below — a weight-loss effect on a low-carb route, riding the same attenuation curve every diet shows. (Ge enters as an F-refinement — it measures no remission outcome itself, but its cross-diet weight-decay finding refines the reading here — and is a listed source.)
The HbA1c estimate at 6 months is the only High-certainty row among the primary outcomes. Table 3 carries five more (LDL at 6 months, HDL and triglycerides at both timepoints).
Attenuation is already visible inside the six-month window — but this is a POST HOC analysis and carries less weight than the a-priori subgroups above. The review is explicit about its provenance: “On the basis of comments from peer reviewers, we did a post hoc analysis on remission at six (+/-3) months”, and hedges its reading (“evidence suggested larger treatment effects… suggesting that shorter term trials may be an effect modifier”). Trials of 3 to <6 months give RD 0.49 (0.30 to 0.68); trials of 6 to 9 months give 0.25 (0.08 to 0.42). Medication-free: 0.20 (0.03 to 0.38) vs 0.00 (-0.07 to 0.07). (Goldenberg et al., 2021)
Note a coincidence worth not over-reading: the medication-free figures here (0.20 / 0.00) are numerically identical to the insulin-exclusion subgroup’s. That is consistent with the two splits being near-coextensive — the non-insulin trials largely being the short ones — in which case this is one finding presented twice, not two. The paper’s supplementary tables would settle it; the wiki does not hold them.
Medication is handled four ways at once, and this is the transferable move
Not adjusted for — co-reported. The review’s stated reasoning:
“Reductions in medication may blunt the effect on mean HbA1c levels, biasing results towards the null and masking any effect; however, any improvement can still be captured if reduction of medication is included as an outcome of interest.” (Goldenberg et al., 2021)
So medication appears (1) inside two of the four remission definitions, (2) as a standalone outcome, (3) as a named bias mechanism, and (4) as a subgroup variable. The HbA1c mean difference and the medication-reduction risk difference are two projections of one underlying effect, and neither alone is complete. Any claim built on the HbA1c figure alone understates the glycaemic effect by an amount the review does not quantify.
Generalizes past diabetes: wherever a treatment lets people reduce a co-intervention, the outcome measured under fixed co-intervention is biased toward the null, and the fix is to report the co-intervention change as its own outcome rather than to adjust it away.
Harms, and how firm they are
- LDL at 12 months: MD 0.14 mmol/L (-0.00 to 0.28), 6 studies, Moderate — above the review’s own 0.10 mmol/L threshold, so labelled a clinically important worsening, though P=0.05. At 6 months LDL is flat (0.02) at High certainty.
- Quality of life at 12 months: MD 3.10 (-2.03 to 8.23), threshold 1, Low certainty. Rests on a single trial of 116 people, with an interval spanning a benefit twice the threshold and a harm eight times it, judged important on the point estimate alone. The 6-month estimate points the other way and comes from 4 trials — so the apparent reversal is partly a change of study set, not a trajectory. (inferred from Goldenberg et al., 2021)
- Adverse events: no significant increase, but “poorly reported among trials and the certainty of evidence for safety ranges from low to very low.”
What the review tells clinicians to do
“clinicians might consider short term LCDs for management of type 2 diabetes, while actively monitoring and adjusting diabetes medication as needed.” (Goldenberg et al., 2021)
Two conditions attached — short term, and active medication management. Note the reason for the time limit is partly external to this review’s own evidence: it cites a cohort review suggesting long-term low-carbohydrate diets are associated with increased mortality.
The summary box vs the body — reconciled, not a defect
The key-messages box claims “large and clinically important improvements in weight loss, triglycerides, and insulin resistance were also seen, without adverse events.” The Discussion’s own threshold tally says something narrower: “Among 10 continuous outcomes, two showed improvements that met or surpassed the MCID at six months (triglycerides, insulin resistance)… At 12 months… two had a clinically important worsening (quality of life, low density lipoprotein cholesterol).”
This reconciles rather than conflicting. The Methods pre-specify that “if the risk of bias sensitivity analysis was credible, we focused our results on those studies at low risk” — and weight was the outcome where that applied, giving MD -7.41 kg (-9.75 to -5.08) in low-risk-of-bias trials, which does surpass the 4.4 kg threshold. And “adverse events” is a named outcome class in this review (total and serious AEs), on which nothing significant was found; the 12-month LDL and quality-of-life worsening are continuous outcomes, and the Conclusions state them plainly. The box is the review applying its own stated method, not drifting from its body.
On the HbA1c threshold. The main text prints thresholds for eight continuous outcomes and not for the two glycaemic ones; the full table is in a supplement not held here. But the question is answerable anyway, from the review’s own tally: of ten continuous outcomes, only triglycerides and insulin resistance met or surpassed their threshold at six months. HbA1c is one of the ten. So -0.47% did not clear the review’s own bar — a decision-relevant fact, and one recoverable by counting the outcomes rather than concluding it unknowable because the main text does not say so directly.
The mechanism this review does not isolate — weight loss, not carbohydrate per se
Low-carbohydrate diets cause weight loss, and this review’s remission effect attenuates to null by 12 months in step with its weight advantage (weight MD -3.46 kg at 6 months -> +0.29 kg at 12 months). A remission effect that decays as the weight advantage decays is consistent with a weight-loss effect delivered via a low-carb route, rather than a carbohydrate effect independent of weight. DiRECT (Lean 2018) makes this near-decisive: it achieves 46% remission on an 825-853 kcal/day formula diet that is 59% carbohydrate — the opposite of low-carb — with remission scaling monotonically with kilograms lost (0% at weight gain -> 86% at >=15 kg). (Lean et al., 2018) So carbohydrate restriction is not necessary for remission, and is best read as one lever for the weight loss that drives remission — with no evidence here that it adds a remission effect beyond the weight loss it produces. The carb-vs-calorie question is weakly probed, not untested (corrected 2026-08-08): Goldenberg ran an a-priori caloric-matching subgroup across all five primary outcomes (remission included) and, on 18 studies, «identified no evidence of credible effect modification based on caloric matching or lack thereof», while hedging that «self-reported dietary intake data are prone to measurement error» (Goldenberg et al., 2021). But that subgroup matches prescribed intake — its iso-caloric controls are «per intake not per goal» (Goldenberg et al., 2021) — not weight loss achieved, so it under-identifies a weight-mediated effect and cannot cleanly separate a carb-specific contribution from the weight-loss one. (inferred from Goldenberg et al., 2021) The added glycaemic contribution is therefore null on a weak test, not cleanly tested and not excluded. A person should optimize for the weight loss they can sustain, by whichever route they will adhere to -> Total Diet Replacement and Type 2 Diabetes Remission.
The umbrella review sharpens the reading — no RCT tests low-carb FOR REMISSION at all [2026-08-04]
Goldenberg pools low-carb weight and glycaemic outcomes in T2D, and reports remission under its four definitions. But the gold-tier T2D umbrella review (Churuangsuk 2022) makes a distinction Goldenberg’s framing hides: when remission is the pre-specified primary outcome, «No RCT has evaluated LCDs/ketogenic diets for type 2 diabetes remission.» (Churuangsuk et al., 2021)
- The much-cited keto-remission figure is a single non-randomised controlled study (Virta, 20% remission, mean −13.8 kg) graded GRADE very low, serious risk of bias — the bottom GRADE tier of the umbrella’s remission map (tied very-low with a food-based VLED, which is graded on critical risk of bias from a single n=9 uncontrolled study — so keto is joint-weakest by certainty, not uniquely so; corrected 2026-08-08), below Mediterranean (15%, low) and far below total diet replacement (54%, high). (Churuangsuk et al., 2021)
- Read against DiRECT (46% remission at 1 year on an 825-853 kcal, 59%-carbohydrate formula (Lean et al., 2018) — one of the two RCTs in the total-diet- replacement class the umbrella grades GRADE high at a median 54% remission, range 46-61% (Churuangsuk et al., 2021); the 46-54% class figure was previously attributed to DiRECT alone, corrected 2026-08-08), the placement is decisive for the causal question: the strongest remission evidence sits on a high-carbohydrate, energy- controlled format, and the low-carb remission claim rests on the weakest design available. This is the umbrella-level version of this page’s own mechanism section — the effect is weight, not carbohydrate -> Diets for Weight Management in Type 2 Diabetes, Total Diet Replacement and Type 2 Diabetes Remission.
- Not independent backing (F, shared-evidence). Churuangsuk’s LCD-weight conclusion pools the same MA class as Goldenberg (and lists Goldenberg 2021 among its high-quality LCD meta-analyses), so its agreement on «low-carb holds no overall advantage» is F-broadening, not a second independent route. Its distinct contribution here is the remission-design gap, which Goldenberg does not state.
- A second gold umbrella (Szczerba 2023) corroborates the gap and adds the firm non-surrogate finding. Szczerba, applying a >=12-week-MA-of-RCT filter, places remission among the low/very-low-certainty outcomes (not in its high/moderate map) — a second body reaching the same “remission RCT evidence is thin” verdict from a wider base. What it does grade firmly for low-carb (<26%E): HbA1c −0.47% (high), triglycerides −0.30 mmol/L (high), and — the firmest clinical outcome in either umbrella — low-carb «reduced the use of drug treatments by an additional 24 per 100 individuals» (risk difference 0.24, 0.12 to 0.35; moderate certainty). It also notes the low-carb estimates may be under-stated by poor adherence («especially for low carbohydrate and ketogenic diets»). Shared MA base -> F/corroboration, not [E]. (Szczerba et al., 2023) (Szczerba et al., 2023)
Limits
- Small numbers behind the headline: the 6-month remission estimate is 8 trials, n=264; by 12 months it is 3 trials, n=171. (Goldenberg et al., 2021)
- Only 30% of trials permitted medication reduction, which constrains the definition that matters most.
- The mortality caution is imported from cohort evidence, not tested here.
- Same population, adjacent condition: this stratum overlaps heavily with fatty liver — insulin resistance is the shared hub (MASLD carries a >2x risk of incident T2D), and the weight-loss and glycemia levers here are the same ones that reduce liver fat -> Fatty Liver MASLD and Weight Loss.
- A guidance benchmark is needed to run the guidance-null on this question; both are staged.