Cardiometabolic · VITD-METB-01 · v2
Borderline: within 4 points of the Uncertain/Doubtful boundary, so a re-grade could land either side.
Three trials built specifically to test whether vitamin D stops prediabetes becoming diabetes [1], [3], [13] each pointed the same way, with a modest reduction in new diabetes, but in none of them was that reduction large enough to rule out chance on the main question each set out to answer. Signs of benefit show up more clearly in extra analyses carved out after the fact and in people who started with very low vitamin D levels [18], [22], [5]. Two of the supportive trials did not blind patients or doctors to who got vitamin D [19], [20], which makes their glucose improvements harder to trust, and one of those saw almost no difference in who actually developed diabetes [20]. Trials using indirect laboratory measures of how the pancreas and insulin behave found nothing [8], [15]. The panel landed in the middle because the direction of effect is consistent but the headline results were not statistically confirmed, and any clear benefit is tied to particular subgroups rather than to everyone with prediabetes.
The three purpose-designed trials in prediabetes or impaired glucose tolerance all produced hazard ratios below 1 ([1] 0.88, 95% CI 0.75-1.04, P=0.12; [3] 0.90, 95% CI 0.69-1.18, P=0.45; [13] 0.87, 95% CI 0.67-1.17, P=0.39) yet none reached significance on its primary diabetes-incidence endpoint. A positive signal emerges in a prespecified secondary β-cell analysis confined to baseline 25(OH)D below 12 ng/mL [18], in an adjusted model and low-insulin-secretion subgroup in [13] (adjusted HR 0.69, 0.51-0.95), and in an ApaI AC/CC genotype subgroup [22] (HR 0.81, 0.66-0.99). Smaller trials enrolling vitamin-D-deficient participants report benefit, with a lower progression rate in [5] (3% vs 28%, P=0.002) and reduced fasting and post-load glucose in [19] and [20], but [19] and [20] are open-label and [20]'s incidence difference was only 7 of 64 versus 9 of 65. Surrogate-endpoint trials of insulin sensitivity and secretion were null [8], [15], and [8] concludes benefit is unlikely even in vitamin-D-deficient populations. The contested point: The division is between the direction of the point estimates, consistently favouring vitamin D across [1], [3] and [13], and the fact that each primary diabetes-incidence endpoint was statistically unconfirmed. Whether the real effect lives only in vitamin-D-deficient or genetically defined subgroups [18], [22], [5], or is absent in the general prediabetes population as [1] and [13] (not selected for insufficiency) suggest, is the unresolved point.
All 22 papers behind this claim with their DOIs and what each does to it, the panel’s arbitration, the evidence gap map and the open research questions. Free, and sent to your inbox.
We’ll also send you other claims from Vitamin D, one at a time, and tell you when a release or a pivotal trial changes a grade. Unsubscribe by replying.
One reviewer's only job is to attack the finding. The challenges below were weighed before the grade was set.
The three largest, best-designed double-blind RCTs ([1] HR 0.88 P=0.12; [3] HR 0.90 P=0.45; [13] HR 0.87 P=0.39) each failed to reject the null on their pre-specified primary diabetes-incidence endpoint, directly contradicting an unqualified 'prevents' claim.
The positive signals rest on multiplicity-prone, hypothesis-generating analyses: an adjusted model ([13] 0.69), a baseline-25(OH)D subgroup ([18]), and a VDR genotype subgroup ([22]) — not confirmatory primary analyses.
Where it remains uncertain. Trials not selected for vitamin D insufficiency are null [1], [13] and a deficiency-focused surrogate trial found nothing [8], so an unqualified population-wide prevention claim outruns evidence that at best supports a deficient-subgroup effect.
Aleatory
Each primary estimate has a wide interval spanning no effect ([1] 0.75-1.04; [3] 0.69-1.18; [13] 0.67-1.17), so random variation cannot be excluded as the explanation for the favourable point estimates.
Epistemic
[3] states very large studies including vitamin-D-deficient subjects would likely be needed to show an effect; the corpus contains no such adequately powered double-blind incidence trial.
Inferential
The positive results rest on an adjusted model [13], a prespecified secondary β-cell analysis in a low-25(OH)D subgroup [18] and a genotype subgroup [22], all hypothesis-generating rather than confirmatory.
The load-bearing papers from a corpus of 22, each classified by what it does to the claim. Every DOI is taken from the corpus the claim was graded on.
Largest purpose-built prediabetes RCT; primary diabetes-incidence endpoint not met.
“The hazard ratio for vitamin D as compared with placebo was 0.88 (95% confidence interval, 0.75 to 1.04; P = 0.12).”
Panel auditInternal validity: high · scope of inference: Adults meeting glycemic criteria for prediabetes, not selected for vitamin D insufficiency, median 2.5-year follow-up.Key limitations: population not selected for low vitamin D · median follow-up 2.5 years
Independent 5-year RCT in prediabetes; no significant prevention of progression.
“One hundred three in the vitamin D and 112 in the placebo group developed T2DM (hazard risk 0.90; 95% confidence interval 0.69–1.18, Cox regression, P = .45, intention to treat analysis).”
Panel auditInternal validity: high · scope of inference: Prediabetic subjects largely without vitamin D deficiency (mean baseline 25(OH)D 24 ng/mL).Key limitations: subjects not vitamin D deficient · abstract notes larger deficient-only trials may be needed
Positive signal in vitamin-D-deficient prediabetes subgroup.
“The rate of progression toward diabetes was significantly lower in the intervention group (28% vs. 3%; P value = 0.002).”
Panel auditInternal validity: low · scope of inference: Pre-diabetic adults specifically with vitamin D deficiency, 6-month follow-up.Key limitations: ~49% attrition · deficient-only population · ambiguously reported percentages · short follow-up
3 of 22 papers shown. The full review lists all of them, with the panel’s arbitration →
Claims are graded at each quarterly release, and re-graded between releases only when a pivotal trial lands. Every change is a new version of the claim.
Directionally consistent but statistically unconfirmed primary endpoints, with efficacy confined to subgroups, place the unqualified claim as genuinely uncertain rather than established or refuted.
Release 2026-q4 re-graded every claim from its corpus under protocol v0.19. An earlier release's grade is shown for the record; it came from an earlier protocol and is not a movement.
Assumptions that carry weight but are not themselves established.
The panel records the gap that would most move the claim, and the study that would close it.
Powered incidence trial in vitamin-D-deficient prediabetes. The claimed effect concentrates in deficient individuals, yet the deficient-population trials are open-label or surrogate-only ([19], [20], [5], [8]) and the blinded trials were not selected for insufficiency ([1], [13]). The panel rates this the highest-importance open gap for the claim, and names the study design that would close it: A large double-blind placebo-controlled RCT restricted to prediabetes with low baseline 25(OH)D, powered on new-onset diabetes as the primary endpoint.
Open question: Does vitamin D supplementation reduce new-onset diabetes in prediabetic adults selected for vitamin D deficiency when tested blind on a hard incidence endpoint? Why it is still open: The deficient-population evidence is open-label or surrogate ([19], [20], [5], [8]) while the blinded incidence trials did not select for deficiency ([1], [13]), and [3] notes such a study is still needed.
Priors is a sceptical first reader, not peer review, and not medical advice.
Release 2026-q4 · graded 26 Oct 2026 · protocol v0.19