Study population

The Nurses’ Health Study is a prospective cohort study of 121 700 female registered nurses aged 30–55 years at baseline in 1976. Mailed questionnaires were administered biennially to collect data on medical, dietary and lifestyle factors, with a follow-up rate exceeding 90 % for each 2-year cycle. Blood samples were collected from 32 826 participants between 1989 and 1990 (biospecimen collection cycle 1) and 18 717 participants between 2000 and 2001 (biospecimen collection cycle 2). As previously reported, participants who provided a blood specimen were generally similar to those who did not in terms of diet and lifestyle⁽ Reference Hunter, Hankinson and Hough ^{11 )}. For the present study, we included participants who provided a blood sample and were previously selected as controls for nested case–control analyses of type 2 diabetes, CHD, stroke, colon cancer, colon polyps, pancreatic cancer and breast cancer. Participants with self-reported prevalent diabetes or CVD at blood draw were excluded. After exclusions, a total of 8492 individuals (6961 from cycle 1 and 1531 from cycle 2) with valid beverage intake data were included in the present analysis (online Supplementary Fig. S1). As different combinations of biomarkers were measured by sub-studies, the sample size for each biomarker varied: fetuin-A (n 1428), alanine transaminase (ALT, n 1319), γ-glutamyl transferase (GGT, n 1317), TAG (n 3363), HDL-cholesterol (n 3071), LDL-cholesterol (n 1715), total cholesterol (n 4083), total cholesterol:HDL-cholesterol ratio (n 2881), C-reactive protein (CRP, n 5939), intracellular adhesion molecule 1 (ICAM-1, n 2290), vascular cell adhesion protein 1 (n 1205), adiponectin (n 6059), insulin (n 2746), HbA_1c (n 4390). The study protocol was approved by the Institutional Review Board of the Brigham and Women’s Hospital and the Human Subjects Committee Review Board of the Harvard T.H. Chan School of Public Health.

Dietary assessment

Dietary intake was measured using validated FFQ administered every 4 years as described in detail previously⁽ Reference Willett, Sampson and Stampfer ^{12 )}. For the present study, we used the last two FFQ before blood collection (1986 and 1990 for cycle 1 and 1994 and 1998 for cycle 2). Participants were asked to report how often, on average they consumed a standard portion of foods and beverages (one standard glass, can, or bottle), using nine possible responses ranging from ‘never or less than once per month’ to ‘6 or more times per day’. We aggregated the responses into five categories: never or almost never, less than once per week, once to twice per week, three to six times per week and one or more times per day. We summed the intake of caffeinated and non-caffeinated colas (e.g. Coke, Pepsi and other colas with sugar), non-cola carbonated beverages with sugar (e.g. 7-Up) and non-carbonated sweetened beverages (e.g. Hawaiian Punch, lemonade and other non-carbonated fruit drinks) as total SSB. Other beverages that may contain sugar such as sweetened milks and coffee with cream and sugar were not included. ASB consisted of all types of low-energy or artificially sweetened carbonated beverages, such as diet colas and other diet carbonated beverages. Fruit juice included apple juice or cider, orange juice, grapefruit juice and other fruit juices. A validation study of the FFQ used in 1980 indicated that correlation coefficients between the FFQ and diet records were 0·84 for colas, 0·36 for non-cola carbonated beverages, 0·56 for non-carbonated sweetened beverages and 0·84 for fruit juice⁽ Reference Salvini, Hunter and Sampson ^{13 )} Total energy intake was calculated by multiplying the frequency of consumption by the energy content of each food item and then adding the contribution from all food items. We used US Department of Agriculture data supplemented with the manufacturers’ data to calculate nutrient values^{( 14 )}. The Alternative Healthy Eating Index (AHEI) was calculated by assigning a maximum of ten points to the following categories: high intakes of fruits, vegetables, whole grains, nuts, legumes and vegetable proteins, long-chain n-3 PUFA (n-3 PUFA), PUFA excluding long-chain n-3 PUFA, as well as low intakes of: SSB and fruit juice, red and processed meats, trans-fat and Na. AHEI scores range from 0 (non-adherence) to 110 (perfect adherence)⁽ Reference Chiuve, Fung and Rimm ^{15 )}. We excluded SSB and fruit juice for AHEI calculation in the present study to avoid collinearity.

Biochemical analysis

Blood sample collections for our study population were described in detail previously⁽ Reference Chan, Ogino and Giovannucci ^{16 )}. In brief, a phlebotomy kit and instructions were sent to participants willing to provide blood specimens in 1989–1990 and in 2000–2001. Samples were returned by overnight courier with an ice pack and processed immediately upon arrival. Whole blood samples were separated into plasma, buffy coat and erythrocytes and stored in the vapour phase of liquid N₂ freezers at −150°C until analysis. Quality control samples were routinely frozen with study samples to monitor potential changes due to long-term storage and to assess assay stability. All biomarkers were measured in the Clinical Chemistry Laboratory at the Children’s Hospital in Boston or in the laboratory of Dr Christos Mantzoros at Beth Israel Deaconess Hospital in Boston using standard methods and assays⁽ Reference Ley, Sun and Jimenez ^{17 ,} Reference Nimptsch, Brand-Miller and Franz ^{18 )}. Biomarker measurements had CV of <15 %. We chose the selected biomarkers as our outcome measures because they have all been shown to be associated with cardiometabolic risk in previous studies⁽ Reference Sun, Cornelis and Manson ^{19 –} Reference Selvin, Steffes and Zhu ^{27 )}. For example, Sun et al. linked plasma fetuin-A levels to higher risk of developing type 2 diabetes in the cohort we used⁽ Reference Sun, Cornelis and Manson ^{19 )}; Ridker et al. demonstrated that LDL-cholesterol, total cholesterol:HDL-cholesterol ratio and CRP were strong predictors of future cardiovascular events⁽ Reference Ridker, Rifai and Cook ^{24 )}.

Covariates

In biennial follow-up questionnaires, we updated information on medical, dietary and lifestyle factors, including age, body weight, smoking status, physical activity, medication use and history of chronic diseases. Information on dietary factors was obtained from the last two FFQ before blood collection for each cycle (1986 and 1990 for cycle 1 and 1994 and 1998 for cycle 2). For non-dietary covariates in this analysis, we used questionnaires administered closest in time to blood draw (1990 for cycle 1 and 2000 for cycle 2). BMI (kg/m²) was calculated using height measured in 1976 and weight measured closest to blood draw (1990 for cycle 1 and 2000 for cycle 2). Physical activity was calculated by multiplying the hours spent on various forms of exercise per week and the metabolic equivalent (MET) score of each specific activity and then summing up the MET-hours for all activities to obtain a value of total weekly MET-hours.

Statistical analysis

Distributions of continuous biomarkers were assessed for normality, and were natural log transformed if skewed. To better reflect habitual beverage intake, we calculated the cumulative average of intakes from the last two FFQ before blood collection for each cohort (1986 and 1990 for cycle 1 and 1994 and 1998 for cycle 2). Descriptive statistics for continuous variables were summarised as means and standard deviations and categorical variables were summarised as proportions according to SSB intake categories.

We used generalised linear models to evaluate the associations between SSB intake and biomarker concentrations. All biomarkers were recalibrated using Rosner et al.’s method⁽ Reference Rosner, Cook and Portman ^{28 )}, in order to account for variation in sample handing and laboratory drift between biospecimen collection cycles and among batches. We constructed two multivariable models. Model 1 was adjusted for age at blood draw, fasting status (fast for ≥8/<8 h), and time of blood draw (blood collection cycle 1/2). Model 2 was additionally adjusted for medical, dietary, and lifestyle variables including smoking status (current/former/never), alcohol intake (continuous), physical activity (tertiles: <5·6, 5·6–17·2 or ≥17·3 h of MET/week), total energy intake (tertile), AHEI score excluding SSB intake (tertiles)⁽ Reference Chiuve, Fung and Rimm ^{15 )}, postmenopausal hormone use (yes/no) and BMI (continuous),which are factors associated with cardiometabolic risk suggested by previous studies⁽ Reference Leiter, Fitchett and Gilbert ^{29 )}. We further adjusted for history of hypertension (yes/no) and history of hypercholesterolaemia (yes/no), which may be potential mediators of the association between SSB intake and cardiometabolic biomarkers. Since BMI and total energy intake may be potential mediators, we also examined associations after removing these variables from the multivariate models. To account for missing data, for variables with missing rates <5 %, we assigned corresponding medians to the missing values of continuous covariates, and assigned corresponding reference groups to the missing data of categorical covariates; for variables with missing rates >5 % (total energy intake only), we handled missing values with multiple imputation. Least-squares means of biomarkers were estimated in frequency categories of SSB intake, and linear trends were evaluated by modelling the medians of intake categories as continuous variables. For better illustration, we additionally calculated the relative concentrations of biomarkers (or relative ratio for total cholesterol:HDL-cholesterol ratio) with 95 % CI by dividing the least-squares mean concentrations of biomarkers among subjects with the greatest SSB consumption (i.e. once per day or more) by the corresponding concentrations of biomarkers among those with the lowest intake level (i.e. never or almost never)⁽ Reference Tabung, Smith-Warner and Chavarro ^{30 ,} Reference Yu, Malik and Keum ^{31 )}. Potential interactions were tested by adding an interaction term for SSB intake (continuous) with age at blood draw (<50/≥50 years), BMI (<25/25–29·9/≥30 kg/m²), AHEI score excluding SSB intake (below/above median) or smoking status (current/former/never). We also evaluated the associations between ASB and fruit juice intakes and biomarker concentrations using the same methods.

We conducted sensitivity analysis by stratifying the association between SSB and biomarkers by blood collection cycle. The overall analysis was merged after confirmation of similar associations in both cycles. We stratified the main analysis by type of SSB (colas, non-cola carbonated beverages and fruit drinks and punch) as their difference in components (e.g. sugar profiles and additives) may result in differences in cardiometabolic risk⁽ Reference Ventura, Davis and Goran ^{32 ,} Reference Vlassara, Cai and Crandall ^{33 )}. For example, cola drinks containing caramel colouring may contain advanced glycation end products that have been linked to cardiometabolic risk⁽ Reference Uribarri, Stirban and Sander ^{34 )}. We also replaced the AHEI score in our multivariate model by intakes of major food groups (red and processed meat, whole grains, fruit, vegetables and coffee in tertiles). For all statistical analyses, two-sided P<0·05 was considered to be statistically significant. All data analyses were performed using SAS software, version 9.3 for UNIX (SAS 170 Institute).