Fruit and vegetable intake (FVI) may reduce the risk of type 2 diabetes (T2D), but the epidemiological evidence is inconclusive. The aim of this study is to examine the prospective association of FVI with T2D and conduct an updated meta-analysis. In the European Prospective Investigation into Cancer-InterAct (EPIC-InterAct) prospective case–cohort study nested within eight European countries, a representative sample of 16 154 participants and 12 403 incident cases of T2D were identified from 340 234 individuals with 3.99 million person-years of follow-up. For the meta-analysis we identified prospective studies on FVI and T2D risk by systematic searches of MEDLINE and EMBASE until April 2011. In EPIC-InterAct, estimated FVI by dietary questionnaires varied more than twofold between countries. In adjusted analyses the hazard ratio (95% confidence interval) comparing the highest with lowest quartile of reported intake was 0.90 (0.80–1.01) for FVI; 0.89 (0.76–1.04) for fruit and 0.94 (0.84–1.05) for vegetables. Among FV subtypes, only root vegetables were inversely associated with diabetes 0.87 (0.77–0.99). In meta-analysis using pooled data from five studies including EPIC-InterAct, comparing the highest with lowest category for FVI was associated with a lower relative risk of diabetes (0.93 (0.87–1.00)). Fruit or vegetables separately were not associated with diabetes. Among FV subtypes, only green leafy vegetable (GLV) intake (relative risk: 0.84 (0.74–0.94)) was inversely associated with diabetes. Subtypes of vegetables, such as root vegetables or GLVs may be beneficial for the prevention of diabetes, while total FVI may exert a weaker overall effect.
Diabetes is a major chronic disease, which is expected to affect in excess of 439 million adults worldwide by 2030,1 with serious consequences for health and longevity. The primary prevention of diabetes is thus clearly an important public health priority. Dietary modification within the setting of lifestyle intervention trials can delay or prevent the development of type 2 diabetes (T2D).2 Although the individual contribution of different foods remains unknown, fruit and vegetable intake (FVI) may explain some of this beneficial effect.
Several plausible mechanisms have been suggested to explain an apparent beneficial effect of FVI on T2D.3, 4, 5, 6 However, findings from prospective studies on the association of FVI with T2D have been inconsistent.6, 7, 8, 9, 10, 11 A recent meta-analysis reported no significant association between FVI, or fruits and vegetables separately, with T2D,12 confirming findings from an earlier meta-analysis.6 Nevertheless, an inverse association between green leafy vegetable (GLV) intake and T2D was found.12 It is plausible that homogeneity in intake, in addition to measurement error, may have obscured a small but biologically important association of FVI with T2D.13, 14 We had the opportunity to further investigate the association between FVI and T2D in the European Prospective Investigation into Cancer(EPIC)-InterAct,15 a prospective case–cohort study, which includes different European populations with large variation in FVI.16
Our study, therefore, had two objectives: first to investigate the association between total FVI and intake of fruit, vegetables and their subtypes and the risk of T2D in the EPIC-InterAct study; and second, to include these results in an updated meta-analysis of published studies.
Subjects and methods
EPIC-InterAct is a large prospective case–cohort study nested within the EPIC study,17 as described previously.15 In brief, the recruitment frame (n=340 234) was sampled from 8 of 10 EPIC countries (n=455 680), excluding those without stored blood (n=109 625) or reported diabetes status (n=5821). Among n=340 234 (with 3.99 million years of follow-up), a subcohort of 16 835 individuals was randomly selected from those with available stored blood and buffy coat, stratified by centre. After exclusion of 681 individuals with prevalent diabetes or without information on diabetes status, 16 154 subcohort individuals were included. Because of random selection, this subcohort also included a random set of 778 individuals who had developed incident T2D during follow-up. Ascertainment of incident T2D involved a review of the existing EPIC data sets at each centre using multiple sources of evidence.15 Follow-up was censored at the date of diagnosis, the 31st of December 2007, or the date of death, whichever came first. In total, 12 403 incident cases of T2D were verified (including 778 cases from the subcohort). From a total of 27 779 participants, we excluded those with incomplete dietary information or with a ratio of energy intake vs energy expenditure in the top or bottom 1% of the original EPIC study sample (n=736), or with missing information on potential confounding variables (n=955). Participants with prevalent myocardial infarction or stroke were also excluded (n=1149), leaving 10 821 incident T2D cases and a subcohort of 14 800 individuals (including 682 incident T2D cases) for this analysis.
All participants gave written informed consent, and the study was approved by the local ethics committee in the participating countries and the Internal Review Board of the International Agency for Research on Cancer.
Dietary and non-dietary data in EPIC-InterAct
Habitual diet was estimated at baseline using country-specific questionnaires, developed and validated in the source populations.18, 19 The major groups and subgroups of fruits and vegetables are shown in Table 1. To improve the comparability of dietary data across European countries a common standardised food database was developed.20 In addition, a standardised 24-h dietary recall was collected in a stratified subsample of ∼8% (n=36 900) of EPIC study participants,21 of whom 2152 participants were in the EPIC-InterAct eligible population.
Standardised health and lifestyle questionnaires at baseline collected information on lifestyle exposures including history of cigarette smoking (never, former, current), occupational and leisure-time physical activity (inactive, moderately inactive, moderately active, active),22 highest achieved education level (none, primary, technical/professional, secondary, university), and history of previous illness. Height, weight and waist circumference were measured by trained staff using standardised protocols, except in Oxford (UK) and France where self-reported measurements were obtained, and Umea (Sweden) where waist circumference was not recorded.
Statistical analysis in EPIC-InterAct
Baseline characteristics were summarised by quartiles of total FVI among subcohort participants, using means and standard deviations (SDs), medians with interquartile ranges (interquartile ranges), or frequencies. To account for the case–cohort design of EPIC-InterAct, multivariable Prentice-weighted Cox regression models23 were used to estimate the association between FV intake and T2D. Total FVI and intake of fruit, vegetables, and fruit and vegetable subtypes were analysed comparing quartiles (with the lowest quartile as the reference category) based on intake data from the subcohort participants. Intake was also analysed continuously. To check the proportional hazards assumption of the models, interactions between fruit, vegetables and FVI combined, with current age (the underlying timescale) were tested. The proportional hazards assumption was not violated for fruit, vegetables or FVI combined (all P-values ⩾0.26). Hazard ratios (HRs) for the association of FVI with diabetes were investigated using the following modelling strategy. Age was used as the underlying timescale and all models were adjusted for study centre. Model A was adjusted for sex. Model B was additionally adjusted for education level, BMI, physical activity level, smoking status, total energy intake and alcohol intake. For the analysis of fruit intake we also adjusted for vegetable intake and vice versa. When analysing specific fruit and vegetable subtypes, other subtypes were included as covariates. HRs and 95% confidence intervals (95% CI) for associations with diabetes were estimated within each country and displayed in forest plots. Overall combined HRs (95% CI) across countries were calculated using random effects meta-analyses. Between-country heterogeneity was assessed using the I2 statistic.
In sensitivity analyses we also included other potentially confounding variables: waist circumference, and intake of cereal fibre, red and processed meat, coffee and sugar sweetened beverages. We conducted additional analyses excluding diabetes cases diagnosed within the first 2-years of recruitment. Effect modification (on the multiplicative scale) was tested by using the interaction term of quartiles of exposure combined with sex, BMI (<25 kg/m2 vs ⩾25 kg/m2), and smoking status (never, former, current). The estimated interaction parameter within each country was combined across countries using the same random effects meta-analysis method used in the main analysis to obtain a P-value for interaction.
A regression calibration model, adapted for a meta-analysis framework, was applied as described by the Fibrinogen Studies Collaboration.24 Single 24-h dietary recalls were used as the reference method for calibrating dietary questionnaires.21, 25 The 24-h dietary recall data were regressed on the dietary questionnaire data for total FVI, fruit intake and vegetable intake separately, and for total energy and alcohol intake. Analyses were adjusted for Model B covariates. Country-specific regression dilution ratios were then estimated and used to correct the log HRs estimated from the Prentice-weighted Cox regression models on a continuous scale. Confidence intervals around the corrected log HRs were estimated using the method described by Wood et al.24
All statistical analyses were performed using Stata/SE 11.1 (Stata-Corp, College Station, TX, USA). All P-values were based on two-sided tests, and statistical significance was set at P<0.05.
Cohort studies published before 30th April 2011 that reported on the association between FVI and T2D were sought by MEDLINE and EMBASE searches, as well as scanning of relevant reference lists and review articles. To ensure a broad search strategy, the words fruit, vegetable, diabetes, glucose, metabolic syndrome, and hyperglycaemia were searched for using medical subject headings and text word, title word, abstract, and subject headings. No limits on publication date or language were applied. Studies were eligible for inclusion if they had a prospective study design, reported relative risks (RRs) or HRs and their corresponding 95% CIs, provided the frequency of FVI using validated dietary assessment questionnaires, and reported on incident diabetes. Two authors (AJC and ZY) reviewed all identified titles (n=3335), and subsequently abstracts and full articles (Figure 1). If multiple published studies from the same study cohort were identified, the most recent study or the study with the most detailed information for both FVI and for diabetes risk was included. Any disagreements were resolved by discussion. For each contributing study, information was extracted by two authors (AJC and ZY). Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines were followed.
Meta-analysis: standardisation of FVI
As studies reported FVI using different measurement units (for example, g/day, servings/day), intake was standardised into portions per day (portions/day) using a portion size of 80 g.26 FVI was subsequently classified into three intake categories: high (H), medium (M), and a reference category of low (R). Assigning categories ensured that the extracted data were used appropriately as all but one study8 assessed FVI using Food Frequency Questionnaires (FFQ), which are not suitable for determining absolute quantity of intake,27 but rather classifying individuals in terms of their relative intake.28 We assigned common lower and upper category values for FVI based on the pooled average of the median intake of the respective values specified by each individual study. When the median consumption value per category was not given, the midpoint of the upper and lower boundaries in each category was assigned. If the upper boundary of the highest category was not given, the same scale of FVI as the preceding category was assigned. Similarly, if the lower boundary of the reference category was not provided, the same scale of FVI as the subsequent category was assigned. If the midpoint of the reference category was <0, then a value of zero was assigned. A medium category of intake was assumed to be mid-way between the upper and reference categories. A similar protocol for fruit intake and for vegetable intake was used. GLV and root vegetable intake categories were based on weekly, not daily, consumption values because of a low quantity of daily intake. H, M and R categories were assigned for total FVI (H=8, M=5 and R=2 portions/day), fruit (H=4, M=2.5 and R=1 portions/day), vegetables (H=5, M=3 and R=1 portions/day), GLV (H=9, M=5 and R=1 portions/week), and root vegetables (H=4, M=2 and R=0 portions/week).
Meta-analysis: statistical approach
We assumed that RRs and HRs included in published studies were equivalent in order to combine results across studies. We used RR estimates from multivariable models with the most complete adjustment for potential confounders, except in one study,7 where we used a lesser adjusted model that we considered most similar to the other included studies in terms of covariate adjustment. We assumed a log-linear association between intake and T2D risk. Linear interpolation was used to compare the highest- and middle-intake categories with the reference intake category, thereby ensuring comparison across studies was with the same reference, middle and upper categories of intake. Pooled RRs and 95% CIs of diabetes comparing the highest and middle categories of intake with the reference category were obtained using random effects meta-analyses. Heterogeneity was assessed using the I2 statistic. To explore heterogeneity, we conducted subgroup analyses by sex (men and women vs women only), duration of follow-up (<10 years vs ⩾10 years), location (USA vs Europe vs China), and dietary assessment method (FFQ vs others). Publication bias was assessed visually and by using Begg’s test to test funnel plot asymmetry.
The median (interquartile ranges) duration of follow-up was 11.0 (7.4–12.7) years. Among subcohort participants there was a greater than twofold variation between countries in estimated FVI (Table 2), which is similar to the marked variation in FVI previously described when estimated by the 24-h dietary recall.16 The highest (median (interquartile ranges)) estimated daily FVI was in Spain (531 (358–730) g/day) and the lowest in Germany (237 (180–320) g/day). Women reported greater FVI (399 (262–562) g/day) than did men (322 (197–501) g/day). Baseline characteristics of subcohort participants are shown by sex-specific quartiles of total FVI in Table 3. Men and women in the highest quartile of FVI had a higher BMI and waist circumference, were less likely to smoke, and were more likely to have lower education. The percentage of participants reporting being physically active increased across increasing quartiles of FVI for men but not for women. Energy intake and percentage energy from protein were higher across increasing quartiles of FVI, whereas percentage of energy from total fat was lower.
Total FVI was inversely associated with T2D comparing the highest with lowest quartile of intake (Model A, HR: 0.84; 95% CI: 0.73, 0.96; P for trend =0.05) (Table 4). After adjustment for potential confounders, including BMI, the inverse association was attenuated and no longer significant (Model B, HR: 0.90; 95% CI: 0.80, 1.01; P for trend =0.42) (Figure 2). We also found no evidence that a 100 g/day increment in FVI was associated with diabetes (Model B HR: 0.99; 95% CI: 0.96, 1.01), and the result from the calibration analysis was similar. The association between FVI and diabetes did not differ by sex, BMI or smoking status (P for interaction= 0.41, 0.72, and 0.43 respectively). Inclusion of waist circumference, and intake of cereal fibre, red and processed meat, coffee and sugar sweetened beverages did not change our results; neither did excluding participants with a diagnosis date of diabetes within the first 2-years of follow-up (data not shown).
The association of fruit intake with T2D was moderate in magnitude but non-significant after adjustment for potential confounders (HR: 0.89; 95% CI: 0.76, 1.04; P for trend =0.30) (Table 4). Similarly, a 100 g/day increment in fruit intake was not associated with diabetes in uncalibrated (Model B HR: 1.00; 95% CI: 0.97, 1.02) or calibrated analyses. Comparing the highest with the lowest quartile of fruit intake, the percentage of total variability due to between-country heterogeneity was I2=38%, with no country except the UK showing a significant inverse association with diabetes. Citrus and non-citrus fruits were not associated with diabetes. Total vegetable intake was not associated with diabetes (Table 4), and this was consistent across countries (I2=0.0%). Among vegetable subtypes however, being in the highest vs lowest quartile of root vegetable intake was associated with a 13% reduction in diabetes after adjustment for potential confounders (HR: 0.87; 95% CI: 0.77, 0.99; P for trend =0.001), with little heterogeneity between countries (I2=12%). GLV intake was inversely associated with diabetes (P for trend=0.03), although the results comparing the highest with lowest quartile of intake were not statistically significant (Table 4).
Seven eligible prospective studies met the inclusion criteria, including EPIC-InterAct (Figure 1). Characteristics of the included studies are shown in Table 5. Of the seven studies (including EPIC-InterAct), four were based in the USA,7, 8, 10, 11 two in Europe,29 and one in China.30 Four studies included women only 7, 10, 11, 30 and three included both men and women.8, 29 FVI was assessed using an FFQ,7, 10, 11, 30 24-h dietary recall8 or diet history interview.29 For total FVI, there were five contributing studies including EPIC-InterAct,7, 8, 10, 11 yielding 179 956 participants and 19 123 T2D cases. Six studies examined the association between diabetes and intake of fruit and vegetables separately,7, 10, 11, 29, 30 five with GLV intake,7, 10, 29, 30 and three with root vegetable intake.10, 30 The age of participants ranged from 25 to 79 years. Study length ranged from 4.6 to 23 years. Participant exclusion criteria differed by study, with most,7, 11, 29, 30 but not all,8, 10 excluding participants with extreme values for total energy intake. All studies, except two,11, 29 excluded participants with prevalent T2D at baseline. Assessment of T2D was by self-report, with confirmation in all but one study that used data from a nationwide drug reimbursement register.29 All studies used multivariable analyses to adjust for potential confounding factors, including age, sex (in studies including men and women) and BMI. Two studies did not additionally adjust for physical activity level and energy intake.8, 29
Figure 3 shows the RR of diabetes for the individual studies reporting data on the association between FVI combined and T2D. Compared with individuals in the lowest category of FVI, the pooled RR of T2D was 0.97 (95% CI: 0.93, 1.00; P=0.09) for those in the middle category, and 0.93 (95% CI: 0.87, 1.00; P=0.05) for those in the highest category of intake. There was little heterogeneity across studies (I2=10%). Visual inspection of the funnel plots of precision against the log of the RR of T2D did not suggest publication bias for the intake of FV combined, fruit, vegetables, GLV or root vegetables, which was confirmed formally with Begg’s tests for funnel plot asymmetry: FV combined (P=1.00), fruit (P=0.85), vegetables (P=0.19), GLV (P=0.14) or root vegetables (P=0.12).
Among FVI subtypes, only GLV intake was consistently inversely associated with diabetes (RR comparing highest with lowest intake category: 0.84, 95% CI: 0.74, 0.94; P=0.004), and this association was irrespective of sex, duration of follow-up, location or dietary assessment method (Table 6). Although a reduced RR of T2D was evident for individuals in the middle vs lowest category of root vegetable intake, this association was not consistent comparing the highest with lowest category of intake (Table 7).
Because I2 values indicated heterogeneity across studies for the intake of fruit, vegetables, GLV and root vegetables (Table 7), we conducted subgroup analyses by sex (men and women vs women only), duration of follow-up (<10 years vs ⩾10 years), location (USA vs Europe vs China), and dietary assessment method (FFQ vs others) (Table 6). Examination of the RRs of diabetes by study characteristics showed that much of the variability across studies was due to duration of follow-up and the dietary assessment method used to estimate FVI. Studies with ⩾10 years of follow-up generally found stronger inverse associations between FVI and risk of T2D than studies with <10 years of follow-up. Studies using an FFQ to estimate FV intake appeared to consistently report weaker associations between intake and the risk of T2D when compared with studies that used other dietary assessment methods (for example, 24-h dietary recall).
We provide a comprehensive assessment of the association between FVI and diabetes in our meta-analysis of >179 000 individuals (with >19 000 incident T2D cases), including new data from the large EPIC-InterAct study in Europe. For total FVI there was a weak inverse association with incident diabetes, which was non-significant in EPIC-InterAct, and of modest magnitude in the meta-analysis, with a 7% lower RR of diabetes in those with the highest vs lowest FVI. However, the association between diabetes and FVI was most pronounced for specific subtypes of vegetables, including root vegetables and GLVs, suggesting that persons at risk of diabetes may benefit from consuming higher quantities of these specific vegetable subtypes.
The suggestion from our meta-analysis of a modest inverse association of FVI with T2D is consistent with findings from biomarker,9, 31 and dietary pattern studies,32, 33, 34, 35, 36 but not with the findings from some prospective cohort studies. When exploring potential sources of heterogeneity across studies we observed that associations between FVI and T2D tended to be weaker when intake was assessed using a FFQ compared with other dietary assessment methods (i.e., 24-h recall). This finding is consistent with previous studies indicating that detection of diet–disease associations may be sensitive to the dietary assessment method used to estimate intake,9, 37 suggesting that the inconsistency between studies examining the association between FV intake and the risk of diabetes could be due to the extent of measurement error associated with the FFQ.37 There is some suggestion that fruit may be differentially associated with diabetes compared with vegetables.7, 9, 30 We observed no significant association of diabetes with fruit intake, but specific subtypes of vegetables were inversely associated with diabetes. In our meta-analysis, we observed that those in the highest compared with lowest category of GLV intake (9 vs ⩽1 portion per week) had a 16% reduced risk of diabetes, which is consistent with the findings reported in a previous meta-analysis.12 The inverse association of root vegetable intake with diabetes that we observed in EPIC-InterAct (HR: 0.87, 95% CI: 0.77–0.99) was not replicated in our meta-analysis when combined with the findings from two other studies.10, 30
Although several biological mechanisms have been proposed to explain an inverse association of FVI with diabetes, no clear mechanism(s) exists as yet. Antioxidants in fruits and vegetables have been hypothesised to protect against diabetes. Yet several supplementation trials, including supplementation with β-carotene and vitamin C, have reported null associations on adverse metabolic traits, including diabetes.38, 39, 40 It is plausible that an inverse association with diabetes only occurs in the presence of a complex mixture of antioxidants, as found in whole fruits and vegetables.41 An inverse association of FVI with diabetes may also be mediated through body weight,42, 43 in which case adjustment for BMI would constitute overadjustment. However, exclusion of BMI from the EPIC-InterAct analysis (Model B) made little difference to the observed estimates (data not shown). Regarding specific mechanisms related to GLV, it has been shown that supplementation with magnesium, of which GLV are an important source, may improve glucose metabolism44 and reduce the risk of diabetes.45 An inverse association of GLV with diabetes risk may also work through the nitric oxide-1-arginine pathway.46, 47 Further mechanistic studies are needed to help explain the apparent inverse association of FVI, and of GLV and root vegetables in particular, with T2D.
This study has some important strengths. We included data from EPIC-InterAct which estimated FVI using dietary questionnaires and standardised 24-h dietary recalls across eight different European countries. Notably, EPIC-InterAct includes countries with a high degree of heterogeneity in FVI and with different confounding structures. By combining results from EPIC-InterAct with previously published studies we more than doubled the total number of T2D cases included in previous meta-analyses.6, 12 We were able to explore possible sources of heterogeneity across studies, and were able to demonstrate that one likely source of heterogeneity that is likely, in part, to explain the inconsistent findings between studies is the dietary assessment method used to estimate FVI. Although the reference and upper intake categories varied among individual studies, by using linear interpolation we were able to ensure the same categories of intake were used for all studies, which was not the case previously.6, 12 We acknowledge that the upper intake category of 8 portions/day for total FVI is higher than the public health message to consume at least 5 portions/day.48 However, as all but one study8 included in our meta-analysis used FFQs to assess FVI, which substantially overestimate intake,27 the assigned upper category of 8 portions/day is not inconsistent with current WHO recommendations.48 It is therefore possible that previous studies failed to find an association between FVI and T2D risk because of setting an upper intake category too low. Limitations of our study and others included in the meta-analysis also merit consideration. All but one study included in the meta-analysis30 used a single baseline assessment of FVI to determine long-term exposure. This, in addition to the measurement error associated with self-report instruments will likely have biased our risk estimates towards the null. Also, any potential misclassification of individuals with undiagnosed T2D as non-diabetic in any of the included studies may have attenuated our overall findings. A possible weakness inherent in any systematic review or meta-analysis is the possibility that the wrong conclusion is made as a result of publication bias.49 However, we found no evidence to suggest publication bias as a likely alternative explanation for our findings. The studies included in the meta-analysis differed in adjustment for covariates, particularly in relation to lifestyle behaviours which tend to cluster with FV intake.50, 51 Therefore, we are unable to exclude residual confounding by unmeasured or imperfectly measured lifestyle factors as a plausible explanation for our findings.
In conclusion, our findings from the meta-analysis of cohort studies, including new data from EPIC-InterAct with wide variation in FVI across Europe, provide evidence that specific groups of vegetables, principally GLV and root vegetables, may be beneficial in preventing diabetes, while higher total FVI is weakly inversely associated with diabetes.
We thank all EPIC participants and staff for their contribution to the study. We thank Nicola Kerrison (MRC Epidemiology Unit, Cambridge) for managing the data for the EPIC-InterAct Project. Funding for the InterAct project was provided by the EU FP6 programme (grant number LSHM_CT_2006_037197). In addition, InterAct investigators acknowledge funding from the following agencies: LA: We thank the participants of the Spanish EPIC cohort for their contribution to the study, as well as to the team of trained nurses who participated in the recruitment; JWJB: Verification of diabetes cases in EPIC-NL was additionally funded by NL Agency grant IGE05012 and an Incentive Grant from the Board of the UMC Utrecht; PWF: Swedish Research Council, Novo Nordisk, Swedish Diabetes Association, Swedish Heart-Lung Foundation; RK: German Cancer Aid; TJK: Cancer Research UK; CN: Health Research Fund (FIS) of the Spanish Ministry of Health; Murcia Regional Government (No 6236); PN: Swedish Research Council; KO: Danish Cancer Society; SP: Compagnia di San Paolo; JRQ: Asturias Regional Government; OR: The Västerboten County Council; IS: Verification of diabetes cases was additionally funded by NL Agency grant IGE05012 and an Incentive Grant from the Board of the UMC Utrecht; AMWS: Dutch Ministry of Public Health, Welfare and Sports (VWS), Netherlands Cancer Registry (NKR), LK Research Funds, Dutch Prevention Funds, Dutch ZON (Zorg Onderzoek Nederland), World Cancer Research Fund (WCRF), Statistics Netherlands; BT: German Cancer Aid; AT: Danish Cancer Society; RT: AIRE-ONLUS Ragusa, AVIS-Ragusa, Sicilian Regional Government; ER: ER was supported in this work by the Imperial College Biomedical Research Centre.