RCT literature on the effects of vitamin D supplementation on FPG, insulin resistance and risk of T2DM in non-diabetic individuals published from January 2000 to February 2017 was retrieved from PubMed (https://www.ncbi.nlm.nih.gov/pubmed/) and two Chinese academic databases, CNKI (http://www.cnki.net) and Wanfang (http://www.wanfangdata.com.cn/index.html). The language of the literature was restricted to English and Chinese. The key words used to search relevant articles were ‘vitamin D’, ‘vitamin D₂’, ‘vitamin D₃’, ‘ergocalciferol’, ‘cholecalciferol’, ‘calciferol’, ‘vitamin D analogs’, ‘25-(OH)D’, ‘1,25-(OH)₂D’, ‘diabetes mellitus’, ‘type 2 diabetes mellitus’, ‘hyperglycemia’, ‘impaired glucose tolerance’, ‘insulin resistance’ and ‘insulin sensitivity’.

The inclusion criteria were as follows: i) Study subjects were non-diabetic individuals, including those with normal blood glucose (<6.1 mmol/l) and prediabetes (6.1–7.0 mmol/l) (26); ii) RCTs that focused on the influence of vitamin D deficiency on the indices related to blood glucose through vitamin D supplementation; iii) the intervention group vs. control group should be vitamin D vs. placebo, vitamin D + calcium vs. placebo + calcium or vitamin D + calcium vs. placebo; and iv) study outcome indicators included at least one of the following: Incidence rate or the number of participants who developed T2D, FPG measurements and/or homeostasis model insulin resistance index (HOMA-IR) data, with the value for FPG and HOMA-IR prior to and following supplementation expressed as the mean ± standard deviation (SD).

The exclusion criteria were as follows: i) Animal experiments; ii) non-RCTs; iii) incomplete information on the outcome indicators, such as only reporting the mean and not providing SD; iv) outcome variables were expressed as the median (four interquartile range) or interquartile range, and outcome measurements that did not meet the inclusion requirements; v) articles that were reviews, systematic reviews or meta-analyses; and vi) the subjects were children, pregnant women or lactating women, or suffering from any diseases.

Literature screening and data extraction were performed by two reviewers. The quality and bias risk of the included studies were evaluated. Disputes or differences in the process of document selection or data extraction were settled by a third reviewer. The extracted data included: i) Title, first author and year of publication; ii) geographic location, sample size, blood glucose status, age, body mass index (BMI), baseline 25(OH)D concentration, female proportion and follow-up time; iii) dose and methods of vitamin D supplementation; iv) the related data of outcome variables; and v) the information needed for bias risk assessment and Jadad scoring.

The bias risk assessment tool recommended by the Cochrane Handbook for Systematic Reviews of Interventions, version 5.1.0 (27) and the Jadad scoring scale (28) were used to assess the methodological quality of the included RCTs. The evaluation criteria in the bias risk assessment include: i) Randomization methods; ii) allocation concealment; iii) whether the researchers, participants and observers remained blinded; iv) the integrity of data; v) whether there was selective reporting bias; and vi) other sources of bias. In the Jadad scoring scale, there are three levels for the randomization method, allocation concealment and blinding method, namely appropriate (2 points), unclear (1 point) and inappropriate (0 point). In addition, the scale accounts for when reasons for withdrawal of subjects are reported (1 point) or unspecified (0 points). Studies that scored ≤4 points were considered to be of high quality.

When the literature provided the mean ± SD of FPG and/or HOMA-IR at baseline and at the end of the study, mean ± SD changes in corresponding outcome variable were converted by the following formula: MD_Δ = Mean₁ - Mean₀; and SD_Δ = (SD₁² + SD₀² - 2×0.5×SD₁xSD₀)^1/2, where Mean₁, mean of outcome variable after intervention; SD₀, SD of outcome variable after intervention; Mean₀, mean of outcome variable at baseline; SD₀, SD of outcome variable at baseline (29). The standard error with 95% confidence intervals (95% CIs) of the change in the corresponding outcome variable was converted by the following formula: SE_Δ = (UL_Δ - LL_Δ)/3.92, where SE_Δ, the standard error of the change in the corresponding outcome variable; UL_Δ, the upper limit of the 95% CI; LL_Δ, the lower limit of the 95% CI. The results expressed as standard error were then transformed into SD by the following formula: SD = √n × σ, where n, sample size; σ, standard error. The unified units of FPG, baseline 25(OH)D concentration and vitamin D intervention dose were mg/dl, ng/ml and IU, respectively (with 1 mg/dl=18 mmol/l, 1 ng/ml=2.5 nmol/l and 1 µg=40 IU). The association between vitamin D supplementation and T2DM was expressed by relative risk (RR = incidence of supplementation group/incidence of control group), and the effect on FPG was represented by mean difference (MD = glucose level pre-supplementation - glucose level following supplementation). The effect on insulin resistance (HOMA-IR) was represented by standardized mean difference (SMD = difference of HOMA-IR/SD of HOMA-IR) as the measurement method, or when units were inconsistent. Subgroup analysis was performed in view of the differences of research aims and interventions among studies. Subgroups were stratified on the following bases: i) Glucose status of subjects: Normal glucose tolerance or prediabetes; ii) the age of subjects: Age <45 years, 45≤ age <60 years, age ≥60 years; iii) the body weight index of subjects: Normal (BMI <25), overweight (25≤ BMI <30) and obese (BMI ≥30) (30); iv) the baseline 25(OH)D concentration of subjects: Deficient [25(OH)D <20 ng/ml], insufficient [20≤ 25(OH)D <30 ng/ml] and sufficient [25(OH)D ≥30 ng/ml] (31); v) the follow-up time: <12 months or ≥12 months; vi) vitamin D intervention dose: ≤2,000 IU/day or >2,000 IU/day; vii) intervention with Ca or not; and viii) Jadad score: <4 points or ≥4 points.

Stata software (version 12.0; StataCorp LP, College Station, TX, USA) was used to analyze the population data. The statistical heterogeneity was estimated by Q test (α=0.1) and from the inhibition coefficient (I²). If there was no statistical heterogeneity among studies, the fixed effect model was used for analysis; otherwise the random effect model was applied to analyze data on the premise of excluding clinical heterogeneity and methodological heterogeneity if statistical heterogeneity existed (27). In addition, the source of heterogeneity was investigated by subgroup analysis. Publication bias was assessed by using funnel plots, Egger's test and Begg's test (32). P<0.05 was considered to indicate statistical significance.

A total of 1,047 articles in English and 1,493 articles in Chinese were retrieved in accordance with the search words established in advance. Of these, 47 English articles and 12 Chinese articles were retained following the exclusion of repeated literature and after reading the abstracts. Finally, a total of 23 articles (8–19,33–43), including 22 English articles and 1 Chinese publication, were included for meta-analysis after reading the full text according to the inclusion and exclusion criteria. The full literature retrieval process is outlined in Fig. 1.

A total of 28 RCTs from 23 studies were included in the current meta-analysis. Among these studies, the Jadad scores of 16 studies (containing 21 RCTs) were ≥4 points. The basic characteristics of the studies are summarized in Table I. A total of 17 studies (containing 20 RCTs) reported the methods of generating random allocation sequences, including computer-generated random number sequences and stratified randomization sets. A total of 11 studies (containing 13 RCTs) reported the methods of concealment allocation, including central or pharmacy control, distribution, opaque containers or sealed envelopes. A double-blinded method was adopted in 20 studies (containing 25 RCTs). A total of 19 studies (containing 24 RCTs) described the number and causes of those lost to follow-up. A total of 5 studies (containing 7 RCTs) were analyzed using the intention-to-treat (ITT) method (Table I).

A total of 21 studies (8,10–19,33,35–43) containing 26 RCTs reported FPG results. There was statistical heterogeneity between the studies (I²=89.9%, P<0.001). Subsequently, the random effect model was used for analysis. The results indicated that FPG following vitamin D intervention did not significantly decrease compared with control (MD=−1.76; 95% CI=−4.07, 0.55, P=0.135; Fig. 2). Subgroup analysis revealed that FPG levels in pre-diabetic individuals and in those whose BMI <25 with vitamin D supplementation were significantly lower than those of controls (MD=−3.40; 95% CI=−6.64, −0.15, P=0.041 and MD=−2.80, 95% CI=−5.57, −0.03, P=0.048, respectively; Table II). In the subgroup of baseline 25(OH)D₃ insufficiency [20≤ 25(OH)D <30 ng/ml)], the level of FPG following vitamin D supplementation compared with control was significantly decreased (MD=−1.99, 95% CI=−3.25, −0.73, P=0.002; Table II).

A total of 13 studies (8–10,12–15,18,33,34,39,41,43) containing 14 RCTs reported HOMA-IR results. There was no statistical heterogeneity between the studies (I²<0.01%, P=0.546). Therefore, the fixed effect model was applied for analysis of the results. The meta-analysis indicated that there were no statistically significant changes in HOMA-IR following vitamin D intervention compared with control (SMD=−0.03, 95% CI=−0.13, 0.08, P=0.635; Fig. 3). In the subgroup of baseline 25(OH)D₃ sufficiency [25(OH)D ≥30 ng/ml)], HOMA-IR following vitamin D supplementation was significantly lower than that of controls (SMD=−0.49, 95% CI=−0.90, −0.07, P=0.021; Table II).

A total of 6 studies (11,13,17–19,38) containing 6 RCTs reported RR results. There was no statistical heterogeneity between the studies (I²<0.01%, P=0.448). Subsequently, the fixed effect model was used for analysis. The results suggested that there was no statistically significant difference in the incidence of T2DM following vitamin D intervention or control in the non-diabetic population (RR=0.86, 95% CI=0.74, 1.01, P=0.072; Fig. 4). Subgroup results indicated that the incidence of T2DM decreased following intervention with vitamin D supplement dose >2,000 IU/day and in those with intervention without calcium (RR=0.84, 95% CI=0.70, 1.00, P=0.047 and RR=0.84, 95% CI=0.70, 1.00, P=0.047, respectively; Table II). In addition, the incidence rates of T2DM in pre-diabetic individuals and in overweight (25≤ BMI <30) individuals were lower than that in controls (RR=0.84, 95% CI=0.70, 1.00, P=0.047 and RR=0.82, 95% CI=0.69, 0.98, P=0.032, respectively; Table II).

For each outcome variable, each study was excluded sequentially, and the results of the remaining studies were combined to determine the impact of the excluded study on the overall results. Additionally, following the exclusion of low-quality studies (Jadad score <4), heterogeneity analysis was conducted to determine the source of heterogeneity. For the incidence of T2DM and HOMA-IR, sensitivity analysis indicated that the results of meta-analysis prior to and following the exclusion of any one study were similar; however, the result of meta-analysis of FPG level following the exclusion of one study (11) differed markedly from that prior to its exclusion (data not shown). Following the exclusion of low quality studies (11,12,17,18,33,35,40), the heterogeneity decreased (I²=40.6%, P=0.034; data not shown). It was hypothesized that literature quality may be a major source of heterogeneity; on meta-analysis, there were no significant differences between the effects of vitamin D intervention and control treatments on RR, FPG and HOMA-IR following the exclusion of the low-quality studies (Table II). For each outcome variable, there was no obvious asymmetry in funnel plots (Fig. 5). Begg's and Egger's tests also indicated no significant publication bias for FPG (Begg's test: P=0.086; Egger's test: P=0.800), insulin resistance (Begg's test: P=0.125; Egger's test: P=0.080) or T2DM prevention (Begg's test: P=1.00; Egger's test; P=0.421).