From April 1993 through December 1998, a total of 63,257 Chinese women and men aged 45–74 years enrolled in the study (25
). Study subjects were restricted to the 2 major dialect groups of Chinese in Singapore, that is, the Hokkiens and the Cantonese, who originated from the contiguous provinces of Fujian and Guangdong, respectively, in the southern part of China (26
). Participants were residents of government-built housing estates, where 86% of the Singapore population resided during the enrollment period (25
). Recruitment occurred by an initial letter informing potential participants of the study and inviting them to participate. Approximately 85% of eligible subjects who were invited responded positively (25
). At recruitment, a face-to-face interview was conducted in the subject's home by a trained interviewer using a structured, scanner-readable questionnaire that requested information on demographics, height, weight, use of tobacco, usual physical activity, menstrual and reproductive history (women only), medical history, family history of cancer, and a 165-item food frequency section assessing usual dietary intake during the previous year (25
). A follow-up telephone interview took place between 1999 and 2004 for 52,325 cohort members (83% of the recruited cohort), and questions were asked to update tobacco and alcohol use, medical history, and menopausal status of women. The institutional review boards at the National University of Singapore and the University of Minnesota approved this study.
Assessment of soft drink and juice intake and other covariates
A semiquantitative food frequency questionnaire specifically developed for this population assessing 165 commonly consumed food items was administered during the baseline interview. During the interview, the respondent referred to accompanying photographs to select from 8 food frequency categories (ranging from “never or hardly ever” to “two or more times a day”) and 3 portion sizes. The food frequency questionnaire has subsequently been validated against a series of 24-hour dietary recall interviews in a random sample of ≥1,000 participants that occurred on 1 weekday and 1 weekend day approximately 2 months apart (25
), as well as selected biomarker studies (27
). A range of 0.24–0.79 in correlation coefficients of energy/nutrients was obtained using the 2 methods, and the majority of macronutrients and food groups display correlation coefficients in the high end of this reported range (25
Two different questions from the food frequency questionnaire specifically asked study subjects to report the intake frequency of 1) soft drinks such as Coca-Cola (The Coca-Cola Company, Atlanta, Georgia) and 7UP (Dr. Pepper Snapple Group, Plano, Texas), 1 glass, and 2) other fruit and vegetable juices, 1 glass, packet, or hawker portion from 9 predefined categories (never or hardly ever, 1–3 times a month, once a week, 2–3 times a week, 4–6 times a week, once a day, 2–3 times a day, 4–5 times a day, and 6 or more times a day). Hawker centers are ubiquitous in Singapore and other parts of Asia, serve a variety of foods all day long, and resemble fast-food courts in US shopping malls. One glass was assigned a value of 237 mL or approximately 1 cup. However, we note that there is likely heterogeneity in serving size, and our analysis is focused on frequency.
In conjunction with this cohort, the Singapore Food Composition Table was developed, a food-nutrient database that lists the levels of 96 nutritive/nonnutritive components per 100 g of cooked food and beverages in the diet of the Singaporean Chinese. By combining information obtained from the food frequency questionnaire with nutrient values provided in this food-nutrient database, we were able to compute the mean daily intakes of nutrients for each subject (25
Other known or suspected risk factors for diabetes assessed with the baseline questionnaire included the following: age (years); smoking habits/status (age started/quit, amount, frequency, type); highest educational level reached; body mass index (kg/m2) calculated by using self-reported height and weight; and amount (hours) of moderate (e.g., brisk walking, bicycling on level ground) and strenuous (e.g., jogging, bicycling on hills, tennis) physical activity on a weekly basis. Weight change was calculated by subtracting the baseline weight (kg) from the follow-up weight (kg).
Assessment of diabetes
Self-reported diabetes as diagnosed by a physician was evaluated at baseline, and participants with a history of diagnosed diabetes were excluded from analysis. Diabetes status was assessed again by the following question asked during the follow-up telephone interview: “Have you been told by a doctor that you have diabetes (high blood sugar)?” If yes: “Please also tell me the age at which you were first diagnosed.” Participants were classified as having incident diabetes if they reported developing diabetes anytime between the initial enrollment interview and the follow-up telephone interview that occurred between July 1999 and October 2004. The average follow up time was 5.7 years.
A validation study of the incident diabetes mellitus cases used 2 different methods and is reported in detail in the report by Odegaard et al. (29
). First, cases were ascertained through linkage with hospital records in a nationwide, hospital-based discharge summary database, an administrative database in the Singapore Ministry of Health (30
). If subjects in the study had been admitted to hospitals for diagnoses carrying diabetes-related International Classification of Diseases
, Ninth Revision, codes 250.00–250.92 after recruitment into the cohort, they were considered a valid case. Cases that did not have hospitalization records available with diabetes-related diagnoses were contacted to answer a supplementary questionnaire regarding symptoms, diagnostic tests, and hypoglycemic therapy during a telephone interview. A valid case had the following 3 criteria: 1) confirmed diagnosis later than the baseline interview date, 2) diabetes still present at the time of interview, and 3) use of oral medications or insulin injections to treat diabetes. On the basis of these criteria, we observed a positive predictive value of 99%, as previously described (29
An alternative approach was used to examine those who did not report being diagnosed with diabetes at the baseline or follow-up interview. As part of an ongoing genome-wide association study of type 2 diabetes in this study population, potential control subjects were randomly selected who answered “no” to the question of diabetes diagnosis at baseline and follow-up and who provided blood samples at their first follow-up interview. Frozen red blood cell samples were shipped to the University of Minnesota on dry ice, where they were analyzed for percentage of hemoglobin A1c (HbA1c) (glycated hemoglobin) in a Clinical Laboratory Improvement Amendments-certified laboratory. HbA1c is measured in ethylenediaminetetraacetic acid-treated whole blood on a Tosoh G7 HPLC Glycohemoglobin Analyzer (Tosoh Medics, Inc., San Francisco, California) using an automated high-performance liquid chromatography method. This method is calibrated utilizing standard values derived by the National Glycohemoglobin Standardization Program. The reference range is 4.3%–6.0% with a laboratory coefficient of variation range of 1.4%–1.9% (31
). To date, 2,625 samples have been analyzed, with 148 subjects (5.6% of the sample) having an HbA1c ≥6.5%, meeting the most recent diagnostic guidelines for the presence of diabetes (32
). Thus, 94.4% of persons who reported being free of diabetes at baseline and follow-up were below the HbA1c threshold for diabetes, yielding a very high negative predictive value.
We excluded from analysis any participants who died before the follow-up interview (n = 7,722); reported baseline diabetes (n = 5,469) or cancer, heart disease, or stroke (n = 5,975); reported implausibly high (>5,000 kcal) or low (<600 kcal) energy intakes; or were lost to follow-up (<0.5%). These exclusions, along with further exclusion of 20 participants whose diabetes status was not clear after the validation effort, left 43,580 participants in the present analyses.
Person-years for each participant were calculated from the year of recruitment to the year of reported type 2 diabetes diagnosis or the year of follow-up telephone interview for those who did not report diabetes diagnoses. Relative risks per category of soft drink and juice consumption were estimated by Cox proportional hazards regression models with simultaneous adjustment for demographic, lifestyle, and dietary variables. All regression analyses were conducted by using SAS, version 9.1, statistical software (PROC TPHREG; SAS Institute, Inc., Cary, North Carolina). There was no evidence that proportional hazards assumptions were violated as indicated by the lack of significant interaction between the predictors and a function of survival time in the model. Soft drink and juice categories were based on intakes that allowed for logical cutpoints and provided sufficient participants and cases per category and are as follows: never or hardly ever (0 servings), monthly (1–3 servings a month), 1 time a week, and 2 or more times a week. We combined all participants reporting 2–3 servings a week and above because of a lack of statistical power in the above levels. However, the associations that we observed at 2–3 servings a week in the population for both soft drink and juice consumption persisted regardless of how consumption was categorized and did not materially differ in magnitude from the combination of 2–3 servings a week category with the collective upper categories. The top categories of soft drink and juice are defined by their median. Tests for trend were performed by assigning the median value of soft drink or juice consumption to the respective categories and entering this as a continuous variable into the models.
Four main models were constructed including risk factors known to be associated with type 2 diabetes, with the final 2 models including body mass index (kg/m2), total energy intake, and weight gain, which may be on the causal pathway between the beverage intakes and type 2 diabetes risk. Model 1 included baseline age (<50, 50–54, 55–59, 60–64, ≥65 years), year of interview (1993–1995 and 1996–1998), dialect (Hokkiens vs. Cantonese), and sex. Model 2 included the variables in model 1 plus education (none, primary, secondary or more); smoking (no, former, current); alcohol intake (no, monthly, weekly, daily); moderate activity (0, 0.5–3 hours/week, ≥4 hours/week) and strenuous activity (0, 0.5–2 hours/week, >2 hours/week); total dairy intake as quintiles (g/day); fiber intake (g/day); saturated fat (g/day); and coffee (nondaily, once per day, 2–3 times/day, ≥4 times/day), plus adjustment for the soft drink or juice variable that was not the main exposure of interest. Model 3 included those variables in model 2, plus baseline body mass index (kg/m2 as the original body mass index and its quadratic term (body mass index2)) and total energy intake, as these may represent mediators rather than confounders. Similarly, model 4 included the variables in model 3 plus weight gain (kg) continuously.
We also calculated the mean weight change per level of beverage consumption between the baseline period and follow-up. The mean weight change was calculated as the difference in kilograms between baseline weight and the reported weight during the follow-up interview. General linear modeling was used (PROC GLM; SAS Institute, Inc.) for these weight-gain analyses with the same demographic and lifestyle covariate adjustments as described above, plus adjustment for time between baseline and follow-up interview, and for total intakes (g/day) of fruits, vegetables, dairy products, meat, candy, and desserts.
We hypothesized that there might be a biologically plausible interaction between soft drink intake and weight gain, with accelerated diabetes risk among those high in soft drink consumption and relatively high in body weight gain over time, so we tested for an interaction between soft drink intake and weight gain as a continuous variable. For presentation of the analysis, the soft drink categories were collapsed into <2 drinks/week (referent) and ≥2 drinks/week, and weight gain was transformed from a continuous variable into a dichotomous variable of the top quarter (n = 11,922; 596 cases) of weight gain in participants (≥3 kg over follow-up) and all other participants (<3 kg over follow-up).