Research, Society and Development, v. 15, n. 9, e4115951681, 2026 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v15i9.51681 Magnesium supplementation and insulin resistance in non-diabetic adults: An updated systematic review and meta-analysis revealing a single-group effect Suplementação de magnésio e resistência à insulina em adultos não diabéticos: Revisão sistemática atualizada e metanálise revelando um efeito restrito a um único grupo de pesquisa Suplementación de magnesio y resistencia a la insulina en adultos no diabéticos: Revisión sistemática actualizada y metaanálisis que revela un efecto restringido a un solo grupo de investigación Received: 09/05/2026 | Accepted: 09/10/2026 | Published: 09/11/2026 Vinícius Arantes Campos ORCID: https://orcid.org/0009-0008-1054-0285 Clínicas Expert, Itajaí, Santa Catarina, Brazil E-mail: drviniciusarantes@gmail.com Allan Gomes Doriguetto ORCID: https://orcid.org/0009-0007-0605-3591 Hospital IGESP, São Paulo, São Paulo, Brazil E-mail: allandoriguetto@gmail.com Abstract Magnesium is proposed as an insulin sensitizer. Prior meta-analyses pooled diabetic and non-diabetic populations and suggested benefit, but none isolated the effect on insulin resistance in non-diabetic adults or examined whether it is reproduced across independent research groups. The objective was to estimate that effect and to test its reproducibility. PubMed was searched for randomized controlled trials of oral magnesium, as the sole differing component, versus control in non-diabetic adults reporting the homeostatic model assessment of insulin resistance. The primary outcome was the mean difference in that measure. A random-effects model was fitted by restricted maximum likelihood, with pre-specified subgroup, leave-one-out and risk-of-bias analyses. Seven trials comprising 559 participants were meta-analysed. Overall, magnesium reduced insulin resistance by a mean difference of −0.82 (95% confidence interval −1.57 to −0.08), but heterogeneity was very high and the prediction interval spanned harm to benefit. Heterogeneity was almost entirely explained by study origin: three trials from a single group in Durango, Mexico, using a magnesium chloride oral solution, showed a large effect, whereas four independent replications were null and homogeneous. The between-subgroup difference was highly significant. The only trial at low risk of bias was independent and null, and the pooled effect lost significance on removing a single trial. In conclusion, the apparent benefit of magnesium is confined to trials from one research group, and current evidence does not support magnesium supplementation for insulin resistance outside deficiency states. Keywords: Magnesium; Insulin resistance; Prediabetic state; Dietary supplements; Reproducibility of results. Resumo O magnésio é proposto como sensibilizador à insulina. Metanálises anteriores agruparam populações diabéticas e não diabéticas e sugeriram benefício, mas nenhuma isolou o efeito sobre a resistência à insulina em adultos não diabéticos nem examinou se ele se reproduz entre grupos de pesquisa independentes. O objetivo foi estimar esse efeito e testar sua reprodutibilidade. Buscou-se no PubMed ensaios clínicos randomizados de magnésio oral, como único componente diferente entre os braços, versus controle em adultos não diabéticos que relatassem o modelo de avaliação da homeostase. O desfecho primário foi a diferença de médias nessa medida. Ajustou-se modelo de efeitos aleatórios por máxima verossimilhança restrita, com análises pré-especificadas de subgrupo, exclusão sequencial e risco de viés. Sete ensaios, com 559 participantes, foram agrupados. No conjunto, o magnésio reduziu a resistência à insulina em −0,82 (intervalo de confiança de 95%: −1,57 a −0,08), mas a heterogeneidade foi muito elevada e o intervalo de predição abrangeu desde dano até benefício. A heterogeneidade foi quase inteiramente explicada pela origem dos estudos: três ensaios de um único grupo de Durango, no México, com solução oral de cloreto de magnésio, mostraram efeito amplo, ao passo que quatro replicações independentes foram nulas e homogêneas. A diferença entre subgrupos foi altamente significativa. O único ensaio de baixo risco de viés era independente e nulo. 1 Research, Society and Development, v. 15, n. 9, e4115951681, 2026 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v15i9.51681 Conclui-se que o benefício aparente restringe-se aos ensaios de um único grupo, e a evidência atual não sustenta a suplementação de magnésio para resistência à insulina fora de estados de deficiência. Palavras-chave: Magnésio; Resistência à insulina; Estado pré-diabético; Suplementos nutricionais; Reprodutibilidade dos testes. Resumen El magnesio se propone como sensibilizador a la insulina. Los metaanálisis previos agruparon poblaciones diabéticas y no diabéticas y sugirieron beneficio, pero ninguno aisló el efecto sobre la resistencia a la insulina en adultos no diabéticos ni examinó si se reproduce entre grupos de investigación independientes. El objetivo fue estimar ese efecto y comprobar su reproducibilidad. Se buscó en PubMed ensayos clínicos aleatorizados de magnesio oral, como único componente diferente entre los brazos, frente a control en adultos no diabéticos que informaran el modelo de evaluación de la homeostasis. El desenlace primario fue la diferencia de medias en dicha medida. Se ajustó un modelo de efectos aleatorios por máxima verosimilitud restringida, con análisis preespecificados de subgrupos, exclusión secuencial y riesgo de sesgo. Se agruparon siete ensayos con 559 participantes. En conjunto, el magnesio redujo la resistencia a la insulina en −0,82 (intervalo de confianza del 95%: −1,57 a −0,08), pero la heterogeneidad fue muy elevada y el intervalo de predicción abarcó desde daño hasta beneficio. La heterogeneidad se explicó casi por completo por el origen de los estudios: tres ensayos de un solo grupo de Durango, México, con solución oral de cloruro de magnesio, mostraron un efecto amplio, mientras que cuatro replicaciones independientes fueron nulas y homogéneas. El único ensayo de bajo riesgo de sesgo era independiente y nulo. Se concluye que el beneficio aparente se limita a los ensayos de un solo grupo y que la evidencia actual no respalda la suplementación fuera de estados de deficiencia. Palabras clave: Magnesio; Resistencia a la insulina; Estado prediabético; Suplementos dietéticos; Reproducibilidad de los resultados. 1. Introduction Magnesium is a cofactor for enzymes of glucose metabolism and insulin signalling, and observational data link low magnesium status to insulin resistance and to incident type 2 diabetes (de Baaij et al., 2015). On this basis, oral magnesium supplementation has been proposed as an insulin sensitizer, and prior meta-analyses reported improvements in insulinsensitivity parameters (Simental-Mendía et al., 2016; Veronese et al., 2021). Those syntheses, however, pooled participants with established diabetes together with non-diabetic and at-risk individuals, and did not isolate the effect on the homeostatic model assessment of insulin resistance (Matthews et al., 1985) in a strictly non-diabetic population. A distinct and clinically relevant question is whether magnesium improves insulin resistance before diabetes develops (Tabák et al., 2012) and, equally important, whether any such effect is reproducible across independent investigators. Reproducibility across research groups is now recognised as a core criterion for judging a body of evidence rather than an optional refinement (Ioannidis, 2005; Munafò et al., 2017). The objective was to estimate that effect and to test its reproducibility. PubMed was searched for randomized controlled trials of oral magnesium, as the sole differing component, versus control in non-diabetic adults reporting the homeostatic model assessment of insulin resistance. 2. Methodology This is a bibliographic study (Snyder, 2019), specifically classified as a systematic literature review with quantitative synthesis. Following the classification of research approaches proposed by Gil (2026) and by Risemberg et al. (2026), the study adopts a quantitative approach: the corpus was assembled through the systematic selection of randomized controlled trials, of which eight met the eligibility criteria and seven provided extractable data, and the synthesis of their results was carried out by statistical meta-analysis, without a qualitative component in the treatment of the data. 2 Research, Society and Development, v. 15, n. 9, e4115951681, 2026 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v15i9.51681 2.1 Eligibility criteria Following the PRISMA 2020 guideline (Page et al., 2021), randomized controlled trials of parallel, crossover or factorial design were included when they enrolled non-diabetic adults aged 18 years or over with insulin resistance or metabolic risk, namely prediabetes, impaired fasting glucose or impaired glucose tolerance, metabolic syndrome, overweight or obesity, polycystic ovary syndrome, or non-alcoholic fatty liver disease. Trials had to compare oral magnesium, of any salt or dose, with a non-magnesium control such as placebo or usual care, where magnesium was the only component differing between the compared arms, and had to report the homeostatic model assessment of insulin resistance. Trials in established type 1, type 2 or gestational diabetes were excluded (Rodríguez-Morán & Guerrero-Romero, 2003), as were trials with fixed co-supplementation, for example magnesium combined with vitamin E, vitamin D or multimineral formulas, in which magnesium could not be isolated; special critically ill or organ-failure populations; non-randomized designs; and studies without an extractable insulin-resistance outcome. 2.2 Search and selection PubMed was searched from inception to 2024, combining terms for magnesium, insulin resistance and metabolic risk with a randomized-trial filter, restricted to humans and excluding overtly diabetic or gestational populations. Titles and abstracts were screened with the assistance of a large language model; potentially eligible reports underwent independent fulltext assessment. For a factorial trial with a magnesium-alone arm, only that arm versus placebo was used. 2.3 Data extraction and risk of bias Two reviewers independently extracted the per-arm sample size and the homeostatic model assessment of insulin resistance, at baseline and at the end of the intervention or as change, together with the population, the magnesium salt and dose, and the duration. Risk of bias was assessed with the Cochrane RoB 2 tool across five domains with an overall judgement (Sterne et al., 2019). Discrepancies were resolved by consensus. 2.4 Statistical analysis The primary outcome was the mean difference in the homeostatic model assessment of insulin resistance, calculated as magnesium minus control, so that negative values favour magnesium. Estimates were pooled with a random-effects model by restricted maximum likelihood using the metafor package (Viechtbauer, 2010). Heterogeneity was quantified with the inconsistency statistic and tau squared, a 95% prediction interval was computed (IntHout et al., 2016), and a subgroup analysis by study origin was pre-specified, contrasting trials from the Durango group using a magnesium chloride solution with independent trials, together with leave-one-out sensitivity analysis. Formal testing for small-study effects was not undertaken, because funnel-plot and regression-based methods are unreliable with fewer than ten studies (Egger et al., 1997; Sterne et al., 2011). 2.5 Role of artificial intelligence The literature search and the initial title and abstract screening were assisted by a large language model. All subsequent steps, namely eligibility, data extraction and risk-of-bias assessment, were conducted independently by the two authors, who reconciled differences by consensus and are responsible for the data reported. 3 Research, Society and Development, v. 15, n. 9, e4115951681, 2026 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v15i9.51681 2.6 Ethical considerations This study is a systematic review based exclusively on aggregate data from previously published studies, without access to individual participant data and without direct involvement of human subjects. It therefore did not require appraisal by a Research Ethics Committee. 3. Results 3.1 Study selection The search returned 111 records. At title and abstract screening, 77 records were excluded and 34 reports underwent full-text assessment, of which 26 were excluded for established diabetes as the target condition, fixed co-supplementation preventing the isolation of magnesium, critically ill or organ-failure populations, non-randomized design, or absence of an extractable insulin-resistance outcome. Eight trials met the eligibility criteria. Seven reported an extractable outcome and were meta-analysed. One crossover trial (Chacko et al., 2011) reported fasting insulin but not the calculated homeostatic model assessment and was retained in the narrative synthesis only. The flow of records is shown in Figure 1. Figure 1 - Flow diagram of study selection according to PRISMA 2020. The diagram summarizes the number of records identified, screened, assessed in full text and included at each stage. The reader should note that the single report kept in the narrative synthesis was excluded from the meta-analysis only because it did not report the calculated homeostatic model assessment, not for any methodological shortcoming. Source: Prepared by the Authors (2026). 4 Research, Society and Development, v. 15, n. 9, e4115951681, 2026 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v15i9.51681 3.2 Study characteristics The seven pooled trials, comprising 559 participants, were conducted between 2004 and 2024 in Mexico, South Korea, the Netherlands, Iran and China, in prediabetes, metabolic obesity, overweight or obesity, and impaired glucose tolerance, over 12 to 24 weeks. Three trials originated from a single group in Durango, Mexico, all using a magnesium chloride oral solution; the other four were independent. Their characteristics are presented in Table 1. Table 1 - Characteristics of the seven pooled randomized controlled trials. Study Origin Population n (M/C) Magnesium and dose Weeks Comparator Risk of bias GuerreroRomero et al. (2004) Durango, Mexico Insulin-resistant, hypomagnesemic 32/31 Magnesium chloride 2.5 g/day 12 Placebo Some concerns RodríguezMorán & GuerreroRomero (2014) Durango, Mexico Metabolically obese, normal weight 24/23 Magnesium chloride solution 17 Placebo High GuerreroRomero et al. (2015) Durango, Mexico Prediabetes with hypomagnesemia 59/57 Magnesium chloride 382 mg/day 16 Placebo High Lee et al. (2009) South Korea Normomagnesemic, overweight 75/80 Magnesium oxide 300 mg/day 12 Placebo Low Joris et al. (2017) Netherlands Overweight or obese 25/25 Magnesium citrate 350 mg/day 24 Placebo Some concerns Salehidoost et al. (2022) Iran Prediabetes 34/37 Magnesium oxide 250 mg/day 12 Placebo High Zhao et al. (2024) China Impaired glucose tolerance 28/29 Magnesium 200 mg/day 12 Placebo High M, magnesium arm; C, control arm. Durations are approximate and subject to confirmation against source. The reader should note that the three Durango trials share both a single research group and a single formulation, the variable that structures the whole analysis. Source: Prepared by the Authors (2026). 3.3 Overall effect on insulin resistance Pooled across all seven trials, magnesium reduced the homeostatic model assessment of insulin resistance versus control, with a mean difference of −0.82 (95% confidence interval −1.57 to −0.08; p = 0.031). Heterogeneity, however, was very high, with an inconsistency of 85% and tau squared of 0.82, and the 95% prediction interval, from −3.34 to +1.70, ranged from substantial benefit to harm, indicating that the pooled point estimate is not a reliable summary of a consistent effect. 3.4 The effect is explained by study origin A pre-specified subgroup analysis resolved the heterogeneity almost entirely. The three trials from the Durango group, using a magnesium chloride solution, showed a large reduction, with a mean difference of −1.68 (95% confidence interval −2.39 to −0.97), whereas the four independent trials were null and homogeneous, with a mean difference of +0.04 (95% confidence interval −0.37 to +0.44; inconsistency 0%). The between-subgroup difference was highly significant (p < 0.000001) and residual heterogeneity fell to 18%. In other words, essentially all of the apparent benefit, and all of the heterogeneity, is attributable to which group conducted the trial. 5 Research, Society and Development, v. 15, n. 9, e4115951681, 2026 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v15i9.51681 3.5 Sensitivity and risk of bias The overall result was not robust. Leave-one-out analysis showed that the pooled estimate lost statistical significance when the single most influential trial (Guerrero-Romero et al., 2004) was removed, giving a mean difference of −0.57 (95% confidence interval −1.18 to +0.05). Risk of bias reinforced the same pattern: the only trial judged at low risk of bias was independent and null (Lee et al., 2009), while two of the three large-effect Durango trials were judged at high risk. The domain-level judgements are shown in Table 2. Table 2 - Risk of bias of the pooled trials according to the Cochrane RoB 2 tool. Study D1 D2 D3 D4 D5 Overall Guerrero-Romero et al. (2004) Some concerns Low Low Low Some concerns Some concerns Rodríguez-Morán & Guerrero-Romero (2014) Some concerns Low Low Low Some concerns High Guerrero-Romero et al. (2015) Some concerns Low Some concerns Low Some concerns High Lee et al. (2009) Low Low Low Low Low Low Joris et al. (2017) Some concerns Low Low Low Low Some concerns Salehidoost et al. (2022) Low Some concerns Some concerns Low Low High Zhao et al. (2024) Some concerns Some concerns Low Low Some concerns High D1, randomization process; D2, deviations from intended interventions; D3, missing outcome data; D4, measurement of the outcome; D5, selection of the reported result. The reader should observe that the single low-risk judgement belongs to an independent trial that found no effect. Source: Prepared by the Authors (2026). 3.6 Study not pooled Chacko et al. (2011), a crossover trial in overweight adults, reported fasting insulin, glucose and inflammatory markers but not the calculated homeostatic model assessment, and so could not contribute to the primary analysis. Its metabolic findings were broadly neutral, consistent with the independent-trial subgroup. 4. Discussion 4.1 Main finding In this updated meta-analysis restricted to non-diabetic adults, the apparent benefit of magnesium on insulin resistance was confined to three trials from a single research group using a specific magnesium chloride oral solution, while four independent replications were null and mutually consistent, with an inconsistency of 0%. The naive pooled estimate of −0.82 is therefore misleading: it is a weighted average of a large single-group signal and a null independent signal, not evidence of a reproducible effect. 4.2 Interpretation Two non-exclusive explanations deserve consideration. First, the large effects arose predominantly in populations selected for hypomagnesemia or marked metabolic abnormality, so a genuine benefit may be restricted to magnesium-deficient individuals rather than to the general non-diabetic population (de Baaij et al., 2015). Second, the clustering of all positive results within one group and one formulation, together with the higher risk of bias of those trials and the loss of significance on leave-one-out analysis, raises the possibility of group-specific or formulation-specific effects that have not replicated, a pattern 6 Research, Society and Development, v. 15, n. 9, e4115951681, 2026 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v15i9.51681 repeatedly described in the literature on research reproducibility (Ioannidis, 2005; Munafò et al., 2017). Either way, the current evidence does not support a general recommendation of magnesium supplementation to improve insulin resistance in nondiabetic adults. 4.3 Relation to previous reviews Earlier meta-analyses reported that magnesium improves insulin-sensitivity parameters, but they pooled diabetic and non-diabetic participants and did not test consistency by study origin; notably, at least one of those syntheses shares authorship with the Durango trials (Simental-Mendía et al., 2016; Veronese et al., 2021). By restricting the population to non-diabetic adults, focusing on a single outcome measure and pre-specifying an origin subgroup, the present review reframes the evidence: the signal is real in the data but not independently reproduced. 4.4 Limitations Only seven small-to-moderate trials contributed, populations and magnesium formulations differed, and endpoint values were pooled. One trial reports dispersion explicitly as the standard error of the mean, which was treated here as the standard deviation (Zhao et al., 2024); this understates the variance of that trial and therefore over-weights it, although the trial sits in the null independent subgroup, so the effect of the approximation is to strengthen rather than weaken the main finding. One trial showed a numerical baseline difference in the outcome, of 4.6 versus 5.2 units, which the trial itself reported as nonsignificant (Guerrero-Romero et al., 2004). The search was limited to PubMed, and publication bias could not be formally assessed with seven studies (Sterne et al., 2011). The origin subgroup, though pre-specified and both mechanistically and methodologically motivated, remains observational and should be interpreted as hypothesis-strengthening rather than definitive. 5. Conclusion Among non-diabetic adults, the reduction in insulin resistance attributed to magnesium is driven entirely by trials from a single research group, while independent replications show no effect and the overall result is fragile. Magnesium supplementation should not be recommended to improve insulin resistance in non-diabetic adults on current evidence, except possibly in documented deficiency. Independent, adequately powered trials at low risk of bias are needed to settle the question.