Study population
The Tehran Lipid and Glucose Study (TLGS) is a community-based, longitudinal study initiated in 1998 to determine the risk factors for non-communicable diseases in a representative urban population of Tehran. Follow-up visits happened at approximately 3-year intervals. A detailed description of the study design has previously been published [16]. In the fourth examination cycle (2009–2011) of the TLGS, among 12,823 participants who completed questionnaires and a standard medical examination as well as laboratory and anthropometric measurements, a representative sample of participants of 7956 individuals was randomly selected for dietary assessment. Follow-up surveys were conducted in the fifth (2011–2014) and sixth (2015–2018) examinations. We used changes in total dairy consumption and its subtypes from the fourth (2009–2011) to fifth (2011–2014) examination cycles to predict the incidence of T2D risk in the sixth (2015–2018) examination cycle of TLGS. The fourth examination cycle was used as the baseline (initial) for the current analysis.
Among participants, 1401 men and women, aged ≥18 years who had prediabetes (impaired fasting glucose (IFG) [fasting plasma glucose (FPG) levels ≥100 to < 126 mg/dL (≥5.6 to < 7.0 mmol/L), or 2-h plasma glucose (2-hPG) of ≥140 to < 200 mg/dL (≥7.8 to < 11.1 mmol/L) in the fourth examination cycle were selected.
Participants were excluded if they reported a history of myocardial infarction, stroke, or cancer at the baseline examination (n = 23), if they were missing necessary covariates (n = 12), or if they had missing dietary data over follow-up (n = 462). We further excluded 92 participants with extreme and implausible energy intake at baseline or in the next examination cycle (defined as < 600 or > 3500 kcal/d for women and < 800 or > 4200 kcal/d for men). Participants were also excluded if they had diagnosed T2D in the fifth cycle (n = 43), or they missed the final follow-up examination (n = 253). Some participants fell into more than one exclusion category. The final sample size for analysis was 639 participants with prediabetes (Fig. 1). There were no significant differences between participants with follow-up measurements and those lost to follow-up (supplementary Table 1).
All participants signed a written informed consent form. The protocol was approved by the ethics committee of the Research Institute for Endocrine Sciences (IR.SBMU.ENDOCRINE.REC.1399.057), in compliance with the principles of the Declaration of Helsinki.
Assessment of diet
A semi-quantitative food frequency questionnaire (FFQ) was used to obtain information on the habitual consumption of food items at baseline (fourth cycle) and the fifth examination cycle of TLGS. Trained dietitians determined participants’ intake frequency of food items consumed through the previous year on a daily, weekly, or monthly basis via face-to-face interviews. Portion sizes of consumed foods were reported in household measures, which were then converted to grams.
The exposures of interest in the current analysis, all dairy products, were defined according to the USDA MyPlate (https://www.choosemyplate.gov/dairy) as “foods made from milk that retain their calcium content, including milk, yogurt, cheese, and ice cream.” Total dairy consumption (servings/d) was calculated as the sum of each of the dairy food items in the FFQ, including high-fat milk, low-fat milk, low-fat yogurt, high-fat yogurt, and regular cheese. Low-fat dairy was generated as the sum of low-fat milk plus low-fat yogurt, and high-fat dairy was calculated as the sum of high-fat milk, high-fat yogurt, and regular cheese. One serving of each dairy food item was converted as follows: 240 g for low-fat milk and high-fat milk, 227 g for low-fat yogurt and high-fat yogurt; 28 g for cream cheese and regular cheese; and 120 g for ice cream.
The reproducibility and validity of the FFQ was reported in a previous study [17, 18]. Energy-adjusted correlation coefficients of total dairy consumption based on the FFQ and those based on twelve 24-h dietary recalls among subsamples of men and women were 0.61 and 0.59, respectively. The energy-adjusted intraclass correlation coefficients for the reproducibility of two collected FFQ with a 14-month interval for men and women were 0.48 and 0.66 for total dairy consumption, respectively [18].
Assessment of outcome and definitions
At baseline and all subsequent examinations, fasting plasma glucose (FPG) was measured in overnight fasting plasma samples with an enzymatic colorimetric method using glucose oxidase (Pars Azmoon, Tehran, Iran). A standard 2-h plasma glucose (2-hPG) test was performed for all participants following oral glucose administration of 82.5 g glucose monohydrate solution [equivalent to 75 g anhydrous glucose; Cerestar EP, Spain]. Participants were classified as having T2D if they reported the use of an oral antidiabetic drug or insulin or had FPG ≥ 126 mg/dL (7 mmol/L) or 2-h PG ≥ 200 mg/dL (11.1 mmol/L).
Assessment of covariates
Detailed measurements of variables in the TLGS have been reported elsewhere [16]. Briefly, participants were invited to the TLGS center, and trained interviewers collected and updated information at 3-y follow-up intervals regarding the familial history of non-communicable diseases, cigarette smoking, medication usage, and assessment of physical activity. Body height and weight were measured, and BMI was calculated as weight in kg divided by height in meters squared (kg/m2). The mean of two measurements of systolic (SBP) and diastolic blood pressure (DBP) after a 15-min rest in the sitting position was recorded as the participant’s blood pressure. Hypertension was defined as an SBP ≥ 140 mmHg or a DBP ≥ 90 mmHg or taking antihypertensive drugs. A valid and reliable Persian version of the Modifiable Activity Questionnaire (MAQ) was applied to measure physical activity and metabolic equivalent (MET) values (MET-min/wk) was calculated.
Statistical analyses
For total dairy consumption, low-fat, or high-fat dairy products or subtypes of dairy products as the main exposure, categories of change were performed based on the mean and distribution of change of the dairy variables. For instance, individuals were categorized into 3 groups based on changes in total dairy product intake: 1) relatively stable consumption (±0.5 serving/d), 2) increased consumption of > 0.5 serving/d, and 3) decreased consumption of > 0.5 serving/d.
We illustrated the participants’ baseline characteristics and 3-year changes in lifestyle and dietary factors according to the changes in total dairy product consumption using descriptive statistics. The reported means (SD) were adjusted for age, and dietary factors further adjusted for total energy intake. Tests for linear trend across decreasing-stable-increasing categories of changes in total dairy product consumption were performed by assigning the median value of consumption within each category and treating these as continuous variables.
Odds ratio (ORs) and 95% confidence intervals (CIs) across categories of dairy exposures were estimated from multivariable binary logistic regression models by considering incidence of T2D as the dependent variable. Individuals with relatively stable consumption were assigned as the reference group. Model 1 was adjusted for age, sex, physical activity, change in body mass index, family history of diabetes, and total energy intake. Model 2 was further adjusted for dietary factors, including whole-grain intake and energy from protein and carbohydrate. P-trend values were estimated using the median value in each intake category. We further estimated the association of incident of T2D with changes in intakes of dairy products and its subtypes while modeled as continuous variables of 0.5 servings/d.
Lastly, applying cross-product terms, we checked for statistical interaction of changes in total dairy, low- and high- fat dairy with baseline age, sex, and BMI in the final model on the outcome. There were no interactions between those variables and total dairy intake on the incidence of T2D.
We applied substitution modeling to estimate the effect on diabetes risk of increasing the consumption of a subtype of dairy product and simultaneously decreasing consumption of another specific subgroup of dairy product. All type of dairy products were included in the same multivariate model, and OR were calculated from the difference in β coefficients of changes in intakes of different dairy exposure and the 95% CI from the corresponding variances and covariance [19].
We conducted a series of sensitivity analyses to test the robustness of our results: 1) to minimize the influence of outliers, we performed final analysis after excluding participants for whom dairy exposures were lower than the 0.5th percentile or higher than the 99.5th percentile; 2) we additionally adjusted other dietary factors including changes in fruit as well as legumes intakes to measure the impact of these dietary items on the relationship; 3) we further adjusted the final model for 3-year changes in body weight to estimate the extent to which weight change mediates the association between changes in dairy intake with risk of incident T2D; and 4) we added smoking status as a further covariate in final model. All analyses were performed using SPSS software, version 16, and a two-tailed P value of 0.05 was considered statistically significant.
