META-ANALYSIS

Glycemic Index, , , and Type 2 Systematic review and dose–response meta-analysis of prospective studies

1 2 DARREN C. GREENWOOD, PHD CAMILLA NYKJAER, MSC also considerable inconsistency in results 2 2 DIANE E. THREAPLETON, MSC CHARLOTTE WOODHEAD, MSC 2 2 regarding the role of total CHARLOTTE E.L. EVANS, PHD VICTORIA J. BURLEY, PHD 2 intake. CHRISTINE L. CLEGHORN, MSC Two systematic reviews have con- cluded that there is evidence of a positive OBJECTIVEd association between both dietary GI and Diets with high (GI), with high glycemic load (GL), or high in all GL and risk of (13,14), but carbohydrates may predispose to higher blood and concentrations, glucose in- with considerable unexplored heteroge- tolerance, and risk of type 2 diabetes. We aimed to conduct a systematic literature review and dose–response meta-analysis of evidence from prospective cohorts. neity. The comparison of only the most extreme categories, based on different RESEARCH DESIGN AND METHODSdWe searched the Cochrane Library, MEDLINE, definitions in each reviewed study, intro- MEDLINE in-process, Embase, CAB Abstracts, ISI Web of Science, and BIOSIS for prospective duced additional heterogeneity and dis- studies of GI, GL, and total carbohydrates in relation to risk of type 2 diabetes up to 17 July 2012. carded information in the middle exposure Data were extracted from 24 publications on 21 cohort studies. Studies using different exposure – categories, leading to uncertainty regarding categories were combined on the same scale using linear and nonlinear dose response trends. the strength of the association. Combination Summary relative risks (RRs) were estimated using random-effects meta-analysis. of different definitions of the highest and RESULTSdThe summary RR was 1.08 per 5 GI units (95% CI 1.02–1.15; P = 0.01), 1.03 per lowest exposure categories meant that their 20 GL units (95% CI 1.00–1.05; P = 0.02), and 0.97 per 50 g/day of carbohydrate (95% CI 0.90– summary estimates could not be related to 1.06; P = 0.5). Dose–response trends were linear for GI and GL but more complex for total a particular level of exposure, limiting the carbohydrate intake. Heterogeneity was high for all exposures (I2 .50%), partly accounted for applicability of results in public health terms. by different covariate adjustment and length of follow-up. Furthermore, the review did not assess the nature of any dose–response relationship, an CONCLUSIONSdIncluded studies were observational and should be interpreted cau- tiously. However, our findings are consistent with protective effects of low dietary GI and GL, important criterion for judging the chances quantifying the range of intakes associated with lower risk. Future research could focus on the of any associations being causal. type of sugars and other carbohydrates associated with greatest risk. Results from nine publications from eight large prospective studies have been Diabetes Care 36:4166–4171, 2013 published since the most recent review, including almost 20,000 cases of type 2 ype 2 diabetes is a leading cause of intolerance and risk of eventual type 2 di- diabetes from over 250,000 participants. cardiovascular disease, with a global abetes (3). We therefore assess the evidence accu- T ’ mulated to date, investigating possible prevalence of 10% (1). An individual s A number of studies have indicated – diet is considered to contribute to the de- an association between GI, GL, and type 2 dose response curves and formally ex- velopment of type 2 diabetes, in particular, diabetes (4–8), but there are many other ploring the potential causes of heteroge- the capacity that containing carbo- large studies that find no evidence to sup- neity that may lead to deeper understanding hydrates have to increase blood glucose port the hypothesis (9–11). Accordingly, of the nature of the associations. ’ (2). It has been suggested that diets with the American Diabetes Association sdie- RESEARCH DESIGN AND high glycemic index (GI) or glycemic load tary guidelines for diabetes prevention METHODS (GL) may predispose to higher postpran- currently state that there is insufficient dial blood glucose and insulin concentra- consistent evidence to say that diets low tions, which, in turn, increase glucose in GL reduce diabetes risk (12). There is Data sources and searches A comprehensive systematic literature search was conducted at the end of 2009 ccccccccccccccccccccccccccccccccccccccccccccccccc covering all prospective research providing From the 1Division of Biostatistics, Centre for Epidemiology and Biostatistics, University of Leeds, Leeds, 2 evidence on all aspects of dietary carbo- United Kingdom; and the Nutritional Epidemiology Group, School of Science and Nutrition, Uni- hydrates and cardiometabolic health, in- versity of Leeds, Leeds, United Kingdom. Corresponding author: Darren C. Greenwood, [email protected]. cluding cardiovascular disease, insulin Received 7 February 2013 and accepted 3 July 2013. resistance, glycemic response, and . DOI: 10.2337/dc13-0325 The following online databases were This article contains Supplementary Data online at http://care.diabetesjournals.org/lookup/suppl/doi:10 searched for all prospective studies pub- .2337/dc13-0325/-/DC1. © 2013 by the American Diabetes Association. Readers may use this article as long as the work is properly lished in English language from 1 January cited, the use is educational and not for profit, and the work is not altered. See http://creativecommons.org/ 1990 to 30 November 2009: the Cochrane licenses/by-nc-nd/3.0/ for details. Library, MEDLINE, MEDLINE in-process,

4166 DIABETES CARE, VOLUME 36, DECEMBER 2013 care.diabetesjournals.org Greenwood and Associates

Embase, CAB Abstracts, ISI Web of Sci- from D.C.G., D.E.T., C.E.L.E., C.L.C., was presented in the publication, but not ence, and BIOSIS. We then updated the C.N., and V.J.B., with disagreements re- the distribution, we estimated this based search, using the two primary sources solved by a third researcher. on definitions of the quantiles. The esti- (MEDLINE, including MEDLINE in- For inclusion in dose–response meta- mated exposure level (based on median, process, and Embase) up to 17 July 2012. analysis, only studies publishing estimates mean, or midpoint) was then assigned to The updated search was restricted to cohort of RR with associated CIs, alongside a the corresponding RR for each study. For studies investigating GI, GL, total carbohy- quantified measure of intake, and sufficient studies presenting the exposure per given drate intake, and type 2 diabetes (detailed detail regarding the numbers of cases and unit of energy intake, we rescaled this using search strategy in Supplementary Table 1). noncases or person-years exposure could estimated energy intake for each category if Hand searches of key journals, with search- be included. this was presented. ing of reference lists from included studies For the studies already reporting a and previous review articles, were also Data extraction and quality linear dose–response trend, with a mea- conducted. The guidelines for conducting assessment sure of precision such as a CI or a standard meta-analysis of observational studies in We extracted the following information error, this was used directly. Where re- epidemiology were used throughout the from the publications identified: authors, sults were only presented separately for design, conduct, analysis, and reporting publication year, geographical region of men and women, these were first com- of this review (15). A protocol was drafted the study, name by which the study is bined using a fixed effects meta-analysis prior to starting the review (http://www known, participants’ sex, age range or before combining with other studies. This .sacn.gov.uk/meetings/working_groups/ mean age of participants, study type (full ensured that between-study heterogene- carbohydrate/21092009_1.html) but is cohort, nested case control, or case cohort), ity was not underestimated. All the esti- not currently available for download. length of follow-up, numbers of cases and mated dose–response trends for each noncases, method of dietary assessment, study were then pooled using a random- Study selection and method of outcome assessment, level effects model to take into account antici- The first round of screening of titles and of dietary exposure (either as mean, me- pated between-study heterogeneity (18). abstracts was carried out by members of the dian, midpoint, or range for each category In presenting the linear dose–response review team to remove publications when or unit of increment for continuous esti- trend, we chose an increment size ap- it was immediately apparent they were not mates), the standard used to derive GI or proximately equivalent to one standard relevant, such as editorials, single case- GL (glucose or bread) estimated RRs with deviation in a European or U.S. population, study reports, and therapeutic approach CIs, and characteristics controlled for ei- to ease comparison across exposures. articles. Prespecified guidelines were in ther by modeling, matching, or stratifica- To examine possible nonlinear associ- place to ensure consistency between sepa- tion. Data extraction was carried out by ations, we calculated restricted cubic rate reviewers. We extracted full-text copies D.C.G., D.E.T., C.E.L.E., C.L.C., C.N., splines for each study with more than three of potentially relevant articles, which were C.W., and V.J.B., and its accuracy was categories of exposure, using three fixed read independently by two members of checked by D.E.T. and D.C.G. knots at 10, 50, and 90% through the total the review team. Any disagreements were distribution of reported intake then com- settled by a third reviewer. A structured Data synthesis and analysis bined using multivariate meta-analysis flowchart and detailed guidelines were To enable pooling of individual study (19–22). Four studies only presented re- used to determine eligibility for inclusion. results reported using different exposure sults for a linear trend over a continuous Only cohort studies were eligible, in- categorization, a linear dose–response exposure (8,10,23,24), and two studies cluding nested case-control studies and trend was derived for each study using only presented results for three categories case-cohort studies nested within a cohort. Greenland and Longnecker’smethod (25,26), so these could not be included in Inclusion criteria were studies based on an (16,17). This method estimates study- nonlinear dose–response analyses. adult population, published in the English specificdose–response slopes and associ- We assessed between-study hetero- language since 1990, with assessment of ated CIs based on the results presented for geneity using Cochran Q test and the GI, GL, or total dietary carbohydrate intake each category of GI, GL, or total dietary percentage of total variation in study esti- with more than two categories of exposure, carbohydrate intake before combining mates attributable to between-study het- with at least some control for confounding into a pooled estimate. erogeneity (I2) (27). Rather than assess either by adjustment in a model or match- To derive the dose–response trend, we study quality using a quality score, to ing, type 2 diabetes as an outcome, and used the mean or median exposure for each minimize bias from confounding, we ex- some estimate of relative risk (RR) with a category if this was presented and used the cluded results with no adjustment for any measure of uncertainty such as 95% CIs. midpoint when exposure ranges were pre- confounding or where only unadjusted Only studies with generally healthy par- sented instead. When the lowest or highest dose–response trends could only be esti- ticipants were included, i.e., only if cohort categories were unbounded, we assumed mated. We also tabulated the following participants were not recruited specifically the width of the category to be the same markers of risk of bias: adequacy of the because of ill health or a personal history as the adjacent category when estimating dietary assessment tool, objectivity of as- of disease. Mean dietary exposure for cases the midpoint. Greenland and Longnecker’s certainment of the outcome, adequacy of compared with noncases were not eligible method also requires the distribution of ca- length of follow-up, adequacy of control unless they were adjusted means. Results for ses and person-years, or cases and nonca- for confounding, and potential compet- dietary patterns were not eligible if they ses, with RRs and estimates of uncertainty ing interests. In addition, we investigated did not quantify intake. Gestational diabe- (e.g., CI) for at least three categories of quan- the extent to which specific study charac- tes outcomes were not eligible. Study se- tified GI, GL, or carbohydrate intake. Where teristics defined in advance were associated lection was carried out by two researchers the total number of cases or person-years with different higher or lower estimates or care.diabetesjournals.org DIABETES CARE, VOLUME 36, DECEMBER 2013 4167 4168 diabetes 2 type and carbohydrates, GL, GI, D IABETES C ARE , VOLUME 6 D 36, ECEMBER 03care.diabetesjournals.org 2013

Figure 1dGI, GL, total carbohydrate intake, and estimated RR of type 2 diabetes. A–C: Forest plots of linear dose–response trends with pooled estimates from random-effects meta-analysis. Increments used are approximately one standard deviation. D–F: Summary nonlinear dose–response curves. The median intake is used as the reference category. Tick marks on the horizontal axis indicate the location of category medians, means, or midpoints for included studies. Greenwood and Associates how they potentially explained some of the heterogeneity between the cohort studies and type 2 diabetes (5,8,9,23,24,32,33,42) heterogeneity. These characteristics in- (I2 = 87%; 95% CI 80–92%; Q = 108; df = (Fig. 1C). The estimated category mean in- cluded duration of follow-up and adjust- 14; P , 0.001). takes ranged from approximately 130 to ment for prespecified confounders, which Studies adjusting for family history of 340 g, with individual studies spanning are potential indicators of study quality. type 2 diabetes appeared to have much between 72 and 210 g. The pooled esti- Potential small study effects, such as publi- higher estimates than those not adjusting mate of RR from linear dose–response cation bias, were investigated with contour- (P , 0.001). The stronger association be- meta-analysis was 0.97 (95% CI 0.90– enhanced funnel plots. However, with tween GI and diabetes was restricted to 1.06) per 50 g per day of total dietary car- small numbers of included studies, explo- those studies that adjusted for this, leading bohydrate intake (P = 0.5). There was ration of sources of heterogeneity and of to improved heterogeneity within each substantial heterogeneity between the co- small study affects lack power. All analyses subgroup (Supplementary Table 4). Esti- hort studies (I2 = 75%; 95% CI 50–88%; were conducted using Stata version 12 (28). mates were largely consistent across the Q = 28; df = 7; P , 0.001). other predefined subgroups. The funnel Estimates were largely consistent RESULTSdWe identified 24 publica- plot was approximately symmetric, with lit- across predefined subgroups, though there tions from 21 cohort studies that reported tleevidenceofsmall-studyeffectssuchas was a tendency for studies with longer GI, GL, total or carbohydrate intake, and publication bias (data not shown). follow-up to have larger estimates (Sup- incidence of type 2 diabetes (Supplemen- Nonlinear dose–response meta-analysis plementary Table 4). The funnel plot was tary Fig. 1). One publication could not be showed a consistently increasing risk as- approximately symmetric, with little evi- used in meta-analyses because it did not sociated with increased GI (Fig. 1D). dence of small-study effects such as pub- quantify intake (7), one could not be used There was little evidence of a threshold lication bias (data not shown). because it only presented results for the effect in the plot. Nonlinear dose–response meta-analysis highest and lowest categories (29), and showed a relatively flat curve over a broad onecouldnotbeusedbecauseofthe GL range of typical intakes, with a suggestion of form the results were presented in (30). Data were extracted from 16 publica- lower risks associated with higher intakes The remaining 18 cohorts provided suffi- tions investigating the association where data are more sparse and CIs wider cient information for inclusion in dose– between GL and type 2 diabetes (5,6,8– (Fig. 1F) and where studies had higher pro- response meta-analyses (Supplementary 11,23,25,26,32,34,37–41) (Fig. 1B). The portions of male participants. Table 2). The risk of bias assessment is estimated category mean intakes ranged provided in Supplementary Table 3. from approximately 55 to 245 units of CONCLUSIONSdWe have quanti- Nine studies were from the U.S., four GL, with individual studies spanning be- fied a clear positive association between from Europe, and the remainder from tween 48 and 190 units. The pooled esti- both GI and GL with increasing incidence Australia, Japan, and China. One cohort mate of RR from linear dose–response of type 2 diabetes. The association was presented results in three publications meta-analysis was 1.03 (95% CI 1.00– stronger for GI than GL, with approxi- (4,31,32), so we used the data in the most 1.05) per 20 units of GL (P = 0.02). There mately one standard deviation of GI in- recent publication (32). A further study re- was moderate heterogeneity between the take associated with more than twice the ported GI and load in a separate paper from cohort studies (I2 = 54%; 95% CI 19– increased risk associated with GL. Com- total carbohydrate intake (11,33). For one 74%; Q = 33; df = 15; P = 0.005). pared with the data on dietary GI, the study to be included, we estimated standard As with GI, studies that adjusted for evidence base for GL is more inconsistent errors using the reported P value and esti- family history had higher estimates than in terms of direction of association. mates (25). For another to be included, cat- those that did not adjust for this covariate Despite use of linear dose–response egory means were estimated based on an (P = 0.03), with stronger associations be- trends to combine studies using different assumed normal distribution, with approx- tween GL and diabetes apparent in those exposure categorizations, heterogeneity imate mean and standard deviation derived studies that did adjust for family history. was still high for all exposures. Explora- from the publication (34). The exclusion of Stratifying by family history improved het- tion of this heterogeneity by investigating studies reporting unadjusted estimates had erogeneity within each subgroup (Supple- the estimates in different predefined sub- resulted in the loss of two studies present- mentary Table 4). Longer follow-up was groups suggested that adjustment for ing results for total carbohydrate intake associated with stronger associations be- family history of diabetes was potentially that would otherwise have been included tween GL and type 2 diabetes (P =0.03). important, with studies that did not ad- (35,36). Estimates were largely consistent across the just for it having much lower estimates for other predefined subgroups. The funnel the association between GI, GL, and type GI plot was approximately symmetric, with 2diabetes. Data were extracted from 15 publications little evidence of small-study effects such While these findings are consistent with investigating the association between as publication bias (data not shown). those of two previous systematic reviews GI and type 2 diabetes (5,6,8–11,23– Nonlinear dose–response meta-analysis (13,14), our review is the first to quantify the 26,32,37–40) (Fig. 1A). The estimated showed a consistently increasing risk as- strength of the association, the first to ex- category mean intakes ranged from ap- sociated with increased GL (Fig. 1E). plore some of the heterogeneity in results, proximately 45 to 90 units of GI, with in- There was little evidence of a threshold the first to remove some of this heteroge- dividual studies spanning between 6 and effect in the plot. neity by combining dose–response trends, 36 units. The pooled estimate of RR from and the first to investigate possible nonlin- linear dose–response meta-analysis was Total carbohydrate ear associations. We have included results 1.08 (95% CI 1.02–1.15) per 5 units of Data were extracted from eight studies from nine publications from large prospec- GI (P = 0.01). There was substantial investigating total carbohydrate intake tive studies that have been published since care.diabetesjournals.org DIABETES CARE, VOLUME 36, DECEMBER 2013 4169 GI, GL, carbohydrates, and type 2 diabetes the most recent review, and these include nonlinear trends can be estimated. How- meta-analysis, so this is unlikely to have almost 20,000 more cases of type 2 diabe- ever, in interpreting these, the emphasis occurred in this review. tes from over 250,000 more participants, should be on the relative ranking as much Our findings are consistent with, and further strengthening the evidence on as on the estimated GI and GL. contribute to, a growing body of evidence which our conclusions are based. A wide range of exposures were re- for the protective associations with low Meta-analysis of observational studies ported across the publications, though dietary GI and GL. Our results have quan- is susceptible to the same biases that the the intakes reported by individual studies tified for the first time the range of exposures studies they contain are prone to, so the generally varied by smaller amounts. This associated with lower risk and quantified pooled estimate may still contain an may reflect the variety of dietary assess- the risk reduction associated with specified element of bias to the extent that the ment tools leading to different amounts of differences in GI and GL. Results for carbo- studies reviewed are biased. In particular, measurement error in each study or may hydrates, more generally, are less clear, and all the studies reviewed used some form of be because of contrasting populations, future research could focus in more detail self-reported dietary exposure and were different diets, and phenotypes. on the source and composition of carbohy- therefore susceptible to potentially large In general the GL of a diet is likely to be drates associated with greatest risk. measurement error. In addition, many partly related to the dietary fiber content, fi adjusted for self-reported dietary covari- and this means that it is dif cult to dis- d ates so may not have fully adjusted for true sociate the effects of GL from the fiber Acknowledgments This work was funded intake. This could bias the associations in content. In the studies we reviewed, ad- by the Department of Health for England. fi D.E.T. holds a PhD studentship sponsored either direction. Furthermore, we cannot justment for ber tended to be associated by Kellogg’s Public Limited Company. D.C.G. conclusively prove that any associations with larger estimates where this was done holds an unrelated research grant (a study are causal on the basis of observational (5,6,9), suggesting that other studies may of infant diet) funded by Danone. No other studies alone, and there may be some have underestimated the association, and potential conflicts of interest relevant to this uncorrected confounding in some or all of our pooled estimate may be an underesti- article were reported. the studies. However, the estimates we mate. Similarly, GI and GL may reflect Funding bodies played no part in design have found for GI and GL are strong with other aspects of dietary quality, such as sat- and conduct of the study, article selection, clear dose–response trends, and there was urated fat intake, with findings partly re- data management, analysis, or interpretation no evidence of any small-study effects flecting some other dietary characteristics. of the data and had no role in the preparation, such as publication bias. It is quite likely that higher carbohydrate review, or approval of the manuscript. D.C.G. contributed to study design, manu- Given the limited nature of databases intakes may substitute for fat or protein, script selection, and data extraction; conducted of GI values for foods, assigning a GI to an while maintaining a constant energy intake. all statistical analyses and interpretation of individual’s diet as captured by a food fre- This is another example where observa- results; and drafted the manuscript. D.E.T. quency questionnaire (FFQ) is potentially tional studies are unable to assign causality, performed all literature searches and contrib- problematic. Typically, GI values for each and it is the same with their meta-analysis. uted to data extraction and interpretation of food item in a questionnaire were taken Inconsistencies in results for total results. C.E.L.E., C.L.C., C.N., and C.W. con- from the 2002 international table of GI carbohydrate intake and type 2 diabetes tributed to data extraction and interpretation of values of foods (43). Broad groupings of may be due to differences in the main results. V.J.B. was principal investigator, de- foods within each FFQ item sometimes sources and types of carbohydrate con- signed the study, and contributed to literature necessitates the allocation of an average sumed or other differences in dietary searches and interpretation of results. All au- thors revised the draft critically for important GI for that item, and this has led some practices between European, U.S., Chi- intellectual content and approved the final to express concerns about the appropri- nese, and Australian cohorts. It might version. D.C.G. and V.J.B. are the guarantors of ateness of using FFQ-derived GI and GL also reflect the possibility that healthier, this work and, as such, had full access to all the values to explore disease associations more active people are consuming more data in the study and take responsibility for the (44). The dietary GI of a food is subject carbohydrates. An alternative explanation integrity of the data and the accuracy of the data to considerable variation dependent upon may relate to differences between both the analysis. the extent of processing, cooking method amount of carbohydrate consumed and the The authors thank James Thomas, Univer- and duration, extent of gelatiniza- type of carbohydrate eaten, with different sity of Leeds, for database design. tion, ripeness, and storage duration (45). cohorts also having different proportions of Further issues concern whether foods men and women. References consumed together impact on each other This may also account for any non- 1. World Health Organization. World Health to alter the GI of the whole meal (46). This linear appearance of the dose–response Statistics 2012. Geneva, World Health exposure is therefore potentially prone to plot, with studies reporting higher intakes Organisation, 2012 measurement error bias. The estimation of total carbohydrates having different 2. Sheard NF, Clark NG, Brand-Miller JC, of GL requires the additional estimate of sources of carbohydrate in the diet than et al. Dietary carbohydrate (amount and the amount of carbohydrate in the diet, those reporting lower intakes. Nonlinear type) in the prevention and management providing greater scope for dilution of re- dose–response curves are susceptible to of diabetes: a statement by the American Diabetes Association. Diabetes Care 2004; sults through measurement error bias. cohorts with different ranges of intake, 27:2266–2271 Even though the estimated absolute leading to the appearance of a nonlinear 3. Jenkins DJ, Wolever TM, Taylor RH, et al. values of GI and GL are probably not curve. In this situation, differences in de- Glycemic index of foods: a physiological accurate estimates of actual values in many sign or population can cause the appear- basis for carbohydrate exchange. Am J studies, we have still used them so that the ance of nonlinearity. However, there is a Clin Nutr 1981;34:362–366 different studies can be combined on the reasonable spread of carbohydrate intakes 4. Salmerón J, Ascherio A, Rimm EB, et al. same scale and dose–response trends and over a number of studies included in the Dietary fiber, glycemic load, and risk of

4170 DIABETES CARE, VOLUME 36, DECEMBER 2013 care.diabetesjournals.org Greenwood and Associates

NIDDM in men. Diabetes Care 1997;20: of summarized dose-response data. Stata J substitution for fat or protein and risk of 545–550 2006;6:40–57 type 2 diabetes in male smokers. Eur J 5. Salmerón J, Manson JE, Stampfer MJ, 18. DerSimonian R, Laird N. Meta-analysis in Clin Nutr 2012;66:716–721 Colditz GA, Wing AL, Willett WC. Dietary clinical trials. Control Clin Trials 1986;7: 34. Hopping BN, Erber E, Grandinetti A, fiber, glycemic load, and risk of non-insulin- 177–188 Verheus M, Kolonel LN, Maskarinec G. dependent diabetes mellitus in women. 19. White IR. Multivariate random-effects Dietary fiber, magnesium, and glycemic JAMA 1997;277:472–477 meta-analysis. Stata J 2009;9:40–56 load alter risk of type 2 diabetes in a multi- 6. Krishnan S, Rosenberg L, Singer M, et al. 20. Orsini N, Greenland S. A procedure to tabu- ethnic cohort in Hawaii. J Nutr 2010;140: Glycemic index, glycemic load, and cereal late and plot results after flexible modeling of a 68–74 fiber intake and risk of type 2 diabetes in quantitative covariate. Stata J 2011;11:1–29 35. Leonetti DL, Tsunehara CH, Wahl PW, US black women. Arch Intern Med 2007; 21. Larsson SC, Orsini N. Coffee consump- Fujimoto WY. Baseline dietary intake and 167:2304–2309 tion and risk of stroke: a dose-response physical activity of Japanese American men 7. Villegas R, Liu S, Gao YT, et al. Prospective meta-analysis of prospective studies. Am J in relation to glucose tolerance at 5-year study of dietary carbohydrates, glycemic Epidemiol 2011;174:993–1001 follow-up. Am J Hum Biol 1996;8:55–67 index, glycemic load, and incidence of 22. Orsini N, Li R, Wolk A, Khudyakov P, 36.MontonenJ,JärvinenR,KnektP, type 2 diabetes mellitus in middle-aged Spiegelman D. Meta-analysis for linear Heliövaara M, Reunanen A. Consumption Chinese women. Arch Intern Med 2007; and nonlinear dose-response relations: ex- of sweetened beverages and intakes of 167:2310–2316 amples, an evaluation of approximations, and and glucose predict type 2 diabetes 8. Sluijs I, van der Schouw YT, van der A DL, software. Am J Epidemiol 2012;175:66–73 occurrence. J Nutr 2007;137:1447–1454 et al. Carbohydrate quantity and quality 23. Hodge AM, English DR, O’Dea K, Giles 37. Schulze MB, Liu S, Rimm EB, Manson JE, and risk of type 2 diabetes in the European GG. Glycemic index and dietary fiber and Willett WC, Hu FB. Glycemic index, gly- Prospective Investigation into Cancer and the risk of type 2 diabetes. Diabetes Care cemic load, and dietary fiber intake and Nutrition-Netherlands (EPIC-NL) study. 2004;27:2701–2706 incidence of type 2 diabetes in younger Am J Clin Nutr 2010;92:905–911 24. Barclay AW, Flood VM, Rochtchina E, and middle-aged women. Am J Clin Nutr 9. Meyer KA, Kushi LH, Jacobs DR Jr, Slavin Mitchell P, Brand-Miller JC. Glycemic in- 2004;80:348–356 J, Sellers TA, Folsom AR. Carbohydrates, dex, dietary fiber, and risk of type 2 di- 38. Mosdøl A, Witte DR, Frost G, Marmot dietary fiber, and incident type 2 diabetes abetes in a cohort of older Australians. MG, Brunner EJ. Dietary glycemic index in older women. Am J Clin Nutr 2000;71: Diabetes Care 2007;30:2811–2813 and glycemic load are associated with 921–930 25. Schulz M, Liese AD, Fang F, Gilliard TS, high-density-lipoprotein cholesterol at 10. Stevens J, Ahn K, Juhaeri, Houston D, Karter AJ. Is the association between di- baseline but not with increased risk of Steffan L, Couper D. Dietary fiber intake etary glycemic index and type 2 diabetes diabetes in the Whitehall II study. Am J and glycemic index and incidence of di- modified by waist circumference? Di- Clin Nutr 2007;86:988–994 abetes in African-American and white adults: abetes Care 2006;29:1102–1104 39. Sahyoun NR, Anderson AL, Tylavsky FA, the ARIC study. Diabetes Care 2002;25: 26. van Woudenbergh GJ, Kuijsten A, Lee JS, Sellmeyer DE, Harris TB; Health, 1715–1721 Sijbrands EJ, Hofman A, Witteman JC, Aging, and Body Composition Study. 11. Similä ME, Valsta LM, Kontto JP, Albanes Feskens EJ. Glycemic index and glycemic Dietary glycemic index and glycemic load D, Virtamo J. Low-, medium- and high- load and their association with C-reactive and the risk of type 2 diabetes in older glycaemic index carbohydrates and risk of protein and incident type 2 diabetes. adults. Am J Clin Nutr 2008;87:126–131 type 2 diabetes in men. Br J Nutr 2011; J Nutr Metab 2011;2011:623076 40. Sakurai M, Nakamura K, Miura K, et al. 105:1258–1264 27. Higgins JP, Thompson SG. Quantifying Dietary glycemic index and risk of type 2 12. Bantle JP, Wylie-Rosett J, Albright AL, heterogeneity in a meta-analysis. Stat Med diabetes mellitus in middle-aged Japanese et al. Nutrition recommendations and in- 2002;21:1539–1558 men. Metabolism 2012;61:47–55 terventions for diabetesd2006: a position 28. StataCorp. Stata Statistical Software: Release 41. Yu R, Woo J, Chan R, et al. Relationship statement of the American Diabetes Asso- 12. College Station, Stata Corporation, 2011 between dietary intake and the develop- ciation. Diabetes Care 2006;29:2140–2157 29. Patel AV, McCullough ML, Pavluck AL, ment of type 2 diabetes in a Chinese pop- 13. Barclay AW, Petocz P, McMillan-Price J, Jacobs EJ, Thun MJ, Calle EE. Glycemic ulation: the Hong Kong Dietary Survey. et al. Glycemic index, glycemic load, and load, glycemic index, and carbohydrate Public Health Nutr 2011;14:1133–1141 chronic disease riskda meta-analysis of intake in relation to pancreatic cancer risk 42. Schulze MB, Schulz M, Heidemann C, observational studies. Am J Clin Nutr in a large US cohort. Cancer Causes Control Schienkiewitz A, Hoffmann K, Boeing H. 2008;87:627–637 2007;18:287–294 Carbohydrate intake and incidence of 14. Dong JY, Zhang L, Zhang YH, Qin LQ. 30. Sonestedt E, Lyssenko V, Ericson U, et al. type 2 diabetes in the European Prospective Dietary glycaemic index and glycaemic Genetic variation in the glucose-dependent Investigation into Cancer and Nutrition load in relation to the risk of type 2 diabetes: insulinotropic polypeptide receptor modi- (EPIC)-Potsdam Study. Br J Nutr 2008;99: a meta-analysis of prospective cohort stud- fies the association between carbohydrate 1107–1116 ies. Br J Nutr 2011;106:1649–1654 and fat intake and risk of type 2 diabetes 43. Foster-Powell K, Holt SH, Brand-Miller JC. 15. Stroup DF, Berlin JA, Morton SC, et al.; in the Malmo Diet and Cancer cohort. International table of glycemic index and Meta-analysis Of Observational Studies J Clin Endocrinol Metab 2012;97:E810– glycemic load values: 2002. Am J Clin in Epidemiology Group. Meta-analysis of E818 Nutr 2002;76:5–56 observational studies in epidemiology: a 31. Halton TL, Liu S, Manson JE, Hu FB. Low- 44. Pi-Sunyer FX. Glycemic index and dis- proposal for reporting. JAMA 2000;283: carbohydrate-diet score and risk of type 2 ease. Am J Clin Nutr 2002;76:290S–298S 2008–2012 diabetes in women. Am J Clin Nutr 2008; 45. Venn BJ, Green TJ. Glycemic index and 16. Greenland S, Longnecker MP. Methods 87:339–346 glycemic load: measurement issues and their for trend estimation from summarized 32. Mekary RA, Rimm EB, Giovannucci E, et al. effect on diet-disease relationships. Eur J dose-response data, with applications to Joint association of glycemic load and alcohol Clin Nutr 2007;61(Suppl. 1):S122–S131 meta-analysis. Am J Epidemiol 1992;135: intake with type 2 diabetes incidence in 46. Flint A, Møller BK, Raben A, et al. The use 1301–1309 women. Am J Clin Nutr 2011;94:1525–1532 of glycaemic index tables to predict gly- 17. Orsini N, Bellocco R, Greenland S. Gen- 33. Similä ME, Kontto JP, Valsta LM, Männistö caemic index of composite breakfast eralized least squares for trend estimation S, Albanes D, Virtamo J. Carbohydrate meals. Br J Nutr 2004;91:979–989

care.diabetesjournals.org DIABETES CARE, VOLUME 36, DECEMBER 2013 4171