Statistical analysis

The person-years for each subject were computed from the year of recruitment to the year of reported T2DM diagnosis or the year of the second follow-up interview for individuals who did not report a diabetes diagnosis. Fruit and juice consumption were categorized on the basis of thresholds that were used in the FFQ. Because of low numbers, some categories were combined to achieve a sufficient number of individuals and cases in each category.

HRs and 95% CIs of diabetes risk were estimated with the use of Cox proportional hazards regression models with adjustments for demographic, BMI (in kg/m²), lifestyle, and dietary risk factors. We fitted 3 models with different covariates to adjust for potential confounders. The first model was adjusted for age at the baseline interview (years), the year of the baseline interview (1993–1995 or 1996–1998), sex, dialect group (Hokkien or Cantonese), and total daily energy intake (kilocalories per day). In the second model, we further adjusted for moderate and vigorous physical activity (no moderate, vigorous, or strenuous activity; <4 h moderate activity/wk or <2 h vigorous or strenuous activity/wk; and ≥4 h moderate activity/wk or ≥2 h vigorous or strenuous activity/wk), education level (no formal education, primary school education, or secondary school or higher), cigarette smoking (never smoker, ex-smoker, and current smoker of 1–12 or ≥13 cigarettes/d), alcohol intake (0, <5, or ≥5 g/d), and BMI. In the third model, we further adjusted for total vegetable intake (grams per day), unsweetened soy intake (servings per day), saturated fat intake (percentage of kilocalories), dairy intake (grams per day), soft drink consumption (glasses per day), coffee intake (cups per day), combined black and green tea intake (cups per day), and fruit- and vegetable-juice intake (servings per day) and were mutually adjusted for fruit groups and individual fruit when appropriate. When we examined individual, climate-grouped or GI-grouped fruit, all other individual fruit or fruit groups were mutually adjusted in the multivariable models. We selected covariates for the multivariable models on the basis of the literature and exposures that were previously shown to be associated with T2DM risk in the SCHS to ensure minimal residual confounding while taking precautions to avoid overadjustment and multicollinearity. Moderate physical activities encompassed activities such as brisk walking, bowling, and bicycling on level ground, whereas vigorous and strenuous activities involved work-related activities such as moving heavy furniture and loading and unloading trucks and sports such as tennis, jogging, and swimming laps. The proportional hazards assumption was tested with the use of Schoenfeld’s residuals, and there was no evidence that the assumption was violated (P > 0.05 for all tests). Effect estimates per increment of 3 servings/wk and P-trend values were assessed by modeling continuous variables of fruit consumption with the truncation of outliers (beyond ±4 SDs of the mean).

To understand how the substitution of fruit intake for juice intake influenced diabetes risk, we modeled fruit and juice in the same multivariable model; HRs (which were calculated from the difference in regression coefficients between fruit and juice) and 95% CIs (which were calculated from the variance and covariance of the regression coefficients) of substitution effects were computed (27).

To investigate whether the associations between fruit intake and T2DM risk depended on GI values, individual fruit was categorized as lower-, moderate-, or higher-GI fruit (Supplemental Table 1). GI values [for all fruit except persimmon (28), tangerine (28) and honeydew melon (29)], with glucose as the reference, were obtained from the international GI database (http://www.glycemicindex.com/) (30). To classify (lower, moderate, or higher) fruit on the basis of GI values, all fruit was ranked in order of increasing GI values and categorized into 3 groups with the similar number of fruit in each group.

In supplementary analyses, the relevant models were stratified by sex and median BMI (<25.0 and ≥25.0) to assess a possible effect modification. Sensitivity analyses were performed with the exclusion of individuals who had reported diabetes within the first 4 y of follow-up to account for possible bias as a result of preclinical disease at recruitment (i.e., reverse causality).

All statistical analyses were performed with the use of STATA 13.1 software (StataCorp LP). All P values were 2-sided, and statistical significance was defined as P < 0.05.