The NHSI began in 1976, when 121,700 female registered US nurses completed and returned a mailed questionnaire. Every 2 y since, questionnaires have been mailed to assess health and lifestyle. Because the 1984 food frequency questionnaire (FFQ) was the first to include a detailed assessment of breakfast cereals, we used 1984, when women were 37–65 y of age, as baseline for the current analysis. The NHSII began in 1989, when 116,609 female registered US nurses completed and returned a mailed questionnaire. Because the first FFQ was administered in 1991, when women were aged 26–46 y, we used that year as baseline for the current analysis. We excluded participants who did not complete the baseline FFQ, left 12 or more (NHSI) or ten or more (NHSII) items blank, or had implausible reported total energy intakes (<600 kcal/d or >3,500 kcal/d). In addition, we excluded participants if they had a history of diabetes (including gestational diabetes), cancer, or cardiovascular disease at baseline (n = 7,001 for NHSI and n = 6,254 for NHSII), because participants with a diagnosis of these chronic diseases are likely to have changed their diet. For NHSI, the average 1976 (December 31) ages were 42.0 and 42.8 y and average weights were 62.6 and 64.0 kg for the original participants that were included and excluded, respectively. For NHSII, the average 1989 (December 31) ages were 34.0 and 33.8 y, average heights were 165 and 165 cm, and average weights were 65.1 and 66.6 kg for the original participants that were included and excluded, respectively. After exclusions, a total of 73,327 NHSI and 88,410 NHSII participants remained for our present analysis.
Assessment of Whole Grains
Dietary information was collected using a semiquantitative FFQ that was completed in 1984, 1986, 1990, 1994, and 1998 for NHSI and 1991, 1995, and 1999 for NHSII. The questionnaire asked about average food intake during the past year. Response was given in a commonly used portion size (e.g., a slice of bread) and nine categories of intake ranging from “never, or less than once a month” to “6+ per day”. Open-ended questions were available for breakfast cereal brand names and foods not listed on the FFQ.
The portions were converted to gram weights per serving, and intakes of nutrients were computed by multiplying the frequency of consumption of each unit of food by the nutrient content in grams. Consumption of whole grain (in g/d) was estimated from all grain foods (rice, bread, pasta, and breakfast cereals) based on their dry weight of whole grain ingredients. Whole grain intake from breakfast cereal was derived from more than 250 brand name cereals using information provided by product labels and breakfast cereal manufacturers.
Our whole grain definition included both intact and pulverized forms containing the expected proportion of bran, germ, and endosperm for the specific grain types. The following ingredients in the database were considered whole grains: whole wheat and whole wheat flour, whole oats and whole oat flour, whole cornmeal and whole corn flour, brown rice and brown rice flour, whole rye and whole rye flour, whole barley, bulgur, buckwheat, popcorn, amaranth, and psyllium. Bran and germ in this study refer to total bran and total germ respectively including both the amount naturally contained in whole grains and the amount eaten separately or added during industrial processing or during cooking by the participant.
The method used to develop this whole grain food composition database has been described in detail elsewhere [15
]. Our FFQ has been validated extensively using biomarkers and diet records as reference methods [16
]. For intakes of cold breakfast cereal and dark bread, major sources of whole grains, the Pearson correlation coefficient for the estimates derived from the FFQ and diet records corrected for within-person variation ranged between 0.58 and 0.79 [17
Assessment of Type 2 Diabetes
Cases of diabetes were identified from the mailed questionnaire. Women who reported diabetes were sent an additional questionnaire. Consistent with the criteria of the National Diabetes Data Group [18
], diagnosed cases required (1) an elevated glucose concentration (fasting plasma glucose of ≥7.8 mmol/l, random plasma glucose of ≥11.1 mmol/l, or plasma glucose ≥11.1 mmol/l after an oral glucose load), and at least one symptom related to diabetes (excessive thirst, polyuria, weight loss, or hunger); (2) no symptoms, but elevated glucose concentrations on two occasions; and (3) treatment with insulin or oral hypoglycemic medication. For cases of type 2 diabetes identified after 1998, the cut-off point used for fasting plasma glucose concentrations was lowered to 7.0 mmol/l according to the American Diabetes Association criteria [19
]. Our validation study showed a high confirmation (98%) of self-reported type 2 diabetes after review of the medical record [20
Anthropometry, Medical History, and Lifestyle
Information requested on the baseline questionnaire included age, weight, smoking status, use of postmenopausal hormone therapy, use of oral contraceptives (for NHSII), and personal history of diabetes, cardiovascular disease, and cancer. We updated this information every 2 y. Oral contraceptive use (for NHSI), family history of diabetes, and height were assessed only at baseline. Physical activity data were assessed in 1982, 1986, 1988, 1992, 1996, 1998, and 2000 for NHSI and in 1991 and 1997 for NHSII. Self-administered questionnaires about physical activity and body weight have been validated as described previously [21
]. We calculated body mass index (BMI) as weight in kilograms divided by the height in meters squared (kg/m2
We used Cox proportional hazards analysis to estimate the relative risk (RR) for type 2 diabetes according to dietary intakes. To control as finely as possible for confounding by age and calendar time, we stratified the analysis jointly by age in months at start of follow-up and calendar year of the current questionnaire cycle. The time scale for the analysis was then measured as months since the start of the current questionnaire cycle. Person-years of follow-up were counted from the date of return from the baseline questionnaire (1984 for NHSI, 1991 for NHSII) until the date of diabetes diagnosis, death, or the end of follow-up (June 2002 for NHSI, June 2003 for NHSII), whichever came first.
Dietary variables were categorized in quintiles of intake. We also conducted analyses modeling whole grain intake as a continuous variable: RR of type 2 diabetes was calculated for a 40 g increment in whole grain intake, which was approximately equivalent to the difference between the 5th and the 95th percentile of intake in our studies (NHSI: 35.9 g, NHSII: 44.3 g). To reduce within-person variation, we used the cumulative average dietary intake from all available dietary questionnaires up to the start of each 2-y follow-up [23
]. In NHSII for example, dietary intake reported on the 1991 questionnaire was related to incidence of diabetes from 1991 to 1995, the average of intakes reported on the 1991 and 1995 questionnaires was related to diabetes incidence from 1995 to 1999, and the average of intakes reported on the 1991, 1995, and 1999 questionnaires was related to diabetes incidence from 1999 to 2003.
Nondietary covariates were updated by using the most recently assessed exposure for each 2-y follow-up period. In NHSII for example, smoking status reported on the 1991 questionnaire was used for follow-up from 1991 to 1993, smoking status reported on the 1993 questionnaire was used for follow-up from 1993 to 1995, etc. Models for multivariate analyses for the NHSI included smoking status (never, past, or current <14, 15–24, or ≥25 cigarettes/d); physical activity (<1.0, 1.0–1.9, 2.0–3.9, 4.0–6.9, ≥7.0 h/wk), alcohol intake (0, 0.1–4.9, 5.0–9.9, ≥10 g/d); use of hormone replacement therapy (premenopausal, never, current, past); oral contraceptive use (ever or never); history of type 2 diabetes in parents or siblings (yes or no); consumption of coffee (0, 0.1–0.9, 1.0–1.9, 2.0–3.9, ≥4.0 cups/d), sugar-sweetened soft drinks (<1.0, 1.0–2.9, 3.0–6.9, ≥ 7cans/wk), fruit punch (nonalcoholic) (<1.0, 1.0–2.9, 3.0–6.9, ≥7 cans/wk); and quintiles of total energy intake, processed meat consumption, and the polyunsaturated-to-saturated fat intake ratio. Because of the different age range and questions on physical activity, models for multivariate analyses for the NHSII included the same variables with slightly different categories for smoking status (never, past, or current), physical activity (quintiles of metabolic equivalent h/wk), use of hormone replacement therapy (ever or never), oral contraceptive use (never, past, or current). There were no missing values for the dietary variables because only persons with valid dietary information were included.
The response to each biennial questionnaire exceeded 90% [24
] and the number of missing values was low. In addition, the multiple repeated assessments allowed us to impute the most recent available data for missing values. For the remaining missing values, dichotomous indicator variables were included in the multivariate model. To test for linear trends across quintiles of intake, the quintile medians were modeled as a continuous variable. Modeling of multiplicative interaction terms for age and whole grain intake did not suggest that the proportional hazards assumption was violated (NHSI: p
= 0.42, NHSII: p
= 0.87 for the multivariate model). Pearson correlations were calculated between dietary intakes with adjustment for total energy intake. The proportion of the association between whole grain intake and risk of type 2 diabetes explained by BMI and the corresponding 95% confidence interval (CI) was estimated as described by Lin et al. based on the change in regression coefficients after adding BMI to the multivariate model [25
-Values were two tailed, and values less than 0.05 were considered statistically significant. The SAS statistical program version 9.1 (SAS Institute, http://www.sas.com/software/
) was used for the analyses.
The MEDLINE and EMBASE database was searched up to January 2007 for published articles on cohort studies that examined whole grain intake in relation to risk of type 2 diabetes. Our criteria for including studies in our meta-analysis were: prospective cohort study, type 2 diabetes as the endpoint, description of the whole grain assessment, presentation of RR with a measure of variability, and description of adjustment for potential confounders. Keywords used to identify relevant articles were: “diabetes mellitus, type 2” (as standardized medical subject heading [MeSH] term) AND (“whole grains” OR “whole grain”). Our MEDLINE search of English-language articles identified 45 abstracts of which six described potentially eligible studies. In addition, three non-English papers were identified that were all review articles. Full text review of the articles resulted in five cohort studies that met our criteria (). One of these was NHSI [5
], for which we included the updated analyses with longer follow-up. The search in EMBASE did not identify additional eligible studies. Broadening our search with: “diabetes mellitus, type 2” AND (“dietary fiber” OR “cereals”), all as MeSH terms, resulted in 356 items, but did not result in any additional eligible studies either. Together with the current study, a total of six studies were included in our meta-analysis.
Flow Diagram of the Selection of Studies for the Meta-Analysis
Data extraction was independently performed by two of the authors (JSLdM, RMD) and there were no differences in extracted information. For each study, the RR of type 2 diabetes was expressed per two serving per day increment of whole grain intake, defining one serving as 30 g of grain for the study by Montonen et al. [8
] and 20 g of whole grains for the current study. For NHSI, NHSII, and the Black Women's Health Study [9
], we calculated the continuous estimate for a two-serving-per-day increment in whole grain intake. For the other three studies, we used the Greenland and Longnecker method to calculate a single continuous estimate and its estimated variance from the published information for quintiles or quartiles [26
We used the STATA version 9.2 statistical program (STATA, http://www.stata.com/
) for the meta-analysis. Summary measures were calculated from the logarithm of the RRs and corresponding standard errors of the individual studies using random effects models that incorporate both a within-study and an additive between-studies component of variance [27
-Values for heterogeneity of study results were calculated using the Cochran Q test [28
]. Because this test depends on the number of studies and has limited sensitivity, we also expressed the degree of heterogeneity as the I2
]. The I2
represents the percentage of total variation across studies that is due to between-study heterogeneity rather than chance. We observed that the between-studies heterogeneity in the standard meta-analysis could be due to the level of whole grain intake in the study population. To investigate this possibility, we conducted a meta-regression of log(RR) of the studies as the dependent variable on the log(median) whole grain intake of the study population [29
]. We used the natural logarithm transformation of the median intake, because this fit the data better than the untransformed median intake and produced a plausible shape of the association. Begg and Egger tests and visual inspection of the funnel plot were used to evaluate possible publication bias [30