Study design
The EPIC-Norfolk study, described in detail previously [15], is a UK population-based cohort of 25,639 men and women aged 40–79 years at baseline in 1993–1997. All volunteers gave written informed consent and attended a health check at their general practitioner’s clinic. The study was approved by the Norfolk Research Ethics Committee.
The current analysis excluded those who did not return a food diary (n = 132) and those with: extreme TEI (top and bottom 1% of TEI) (n = 256); prevalent or unconfirmed diabetes (n = 6); or missing covariates: education level (n = 18), family history of diabetes (n = 6), alcohol consumption (n = 239), smoking status (n = 187), self-reported hypertension (n = 29) or hypercholesterolaemia (n = 65), BMI (n = 47) or physical activity level (n = 1). A total of 24,653 participants remained for analysis.
Diabetes cases
Incident type 2 diabetes cases occurring until 31 July 2006 were ascertained using multiple sources: self-report of doctor-diagnosed diabetes from the second health check (3 years post-baseline) or follow-up health and lifestyle questionnaires (18 months and 10 years post-baseline), self-report of diabetes-specific medication in either of the two follow-up questionnaires or medication brought to the follow-up health check. External sources through record linkage were used to verify self-reported type 2 diabetes and identify unreported cases. The date of diagnosis was defined as the earliest date where there was evidence of diabetes from either self-report or from an external source. These included general practice diabetes and local hospital diabetes registers, hospital admissions data with screening for diabetes related admission and Office of National Statistics mortality data with diabetes coding. Participants who self-reported type 2 diabetes that could not be verified with an objective information source were not included as cases (n = 5).
Dietary intake
Baseline dietary intake data were collected using 7 day food diaries [16]. Participants were asked to record everything they ate for 7 consecutive days, covering weekdays and weekend days, with particular attention to amounts and food-preparation methods. Food diaries were collected throughout the year and over a 4 year period, which accounted for seasonal variation in dietary intake at a cohort level. Food and beverage intake data were entered using the programme Data into Nutrients for Epidemiological Research (DINER) [17] and converted into food weights and nutrient intakes using DINERMO [18] (www.epic-norfolk.org.uk/). Intake (g/day) of (1) soft drinks (soft drinks, squashes and juice-based drinks sweetened with sugar), (2) sweetened tea or coffee, (3) sweetened-milk beverages (for example, milkshakes, flavoured milks, hot chocolate), (4) ASB and (5) fruit juice were estimated (further details in Electronic supplementary material [ESM] Table 1). A 1-to-1 g-to-ml conversion was assumed. In these analyses, (1), (2) and (3) are referred to collectively as SSB. TEI (kJ/day) from these sweet beverages was also estimated.
Covariates
Self-administered questionnaires were used to collect baseline demographic, socioeconomic, lifestyle, physical activity and health characteristics, as described previously [15]. A validated four-point physical activity index was used to categorise participants according to activity level (active, moderately active, moderately inactive, inactive) [19]. Height (cm), weight (kg) and waist circumference (cm) were measured using standardised procedures. Dietary covariates were estimated using the 7-day food diary.
Statistical analysis
Analyses were performed using Stata (version 13; Stata Corp, College Station, TX, USA). Statistical significance was defined as p < 0.05.
Baseline characteristics of the study cohort were described using mean, median or proportion (%). Cox proportional hazards regression was used to estimate HRs and 95% CIs for the prospective association of sweet beverage intake with type 2 diabetes incidence.
Incident type 2 diabetes was examined per intake serving of each sweet beverage. Serving sizes were pragmatically chosen after consideration of median portion sizes of consumers within the study, standard manufacturers’ portion sizes in the UK and guidelines for portion sizes of fruit juice [20]. Assigned serving size varied by beverage: soft drinks and ASB, 336 g/day; sweetened tea or coffee and sweetened-milk beverages, 280 g/day; and fruit juice, 150 g/day. The association was also examined across four intake categories (non-consumers, and consumers categorised by tertiles) for each sweet beverage. The linear trend was examined by modelling the median values for each sweet beverage intake category as a continuous variable.
Age was included as the underlying timescale in Cox models, with entry time defined as age at recruitment and exit time as age at type 2 diabetes incidence, death or censoring at the end of follow-up, whichever came first. The assumption of proportional hazards, checked by including the interaction term between each sweet beverage and age, was not violated.
Analyses were adjusted for age (in addition to as underlying timescale), sex, occupational social class, education level, family history of diabetes, physical activity level, smoking status, alcohol consumption (units/week) and season (date of dietary record dichotomised as winter, summer) and were each mutually adjusted for intake of the other sweet beverages (Model 1). Further adjustment for other food and beverage intake variables (alcoholic beverages, unsweetened tea or coffee, drinking water, fruit, vegetables, processed meat, red meat and fish) little altered the results and were not included in primary analysis. Two subsequent models were constructed, one additionally adjusted for TEI (Model 2) and the second additionally adjusted for TEI, BMI and waist circumference (Model 3), allowing for obesity to be considered as both a mediator and a confounder. Possible interactions between intake of each sweet beverage and age, sex, BMI, waist, physical activity index and smoking status were examined a priori by including interaction terms in the most adjusted models. Interactions were considered significant where p < 0.05.
A number of sensitivity analyses were conducted, using Model 3. These included repeating analyses: (1) additionally adjusting for fibre intake to examine the role of nutrient displacement and overall dietary quality; (2) additionally adjusting for saturated fat intake; (3) adjusting for non-sweet-beverage energy intake in place of TEI to reduce the risk of over-adjusting as sweet beverages contribute to TEI; (4) excluding those with prevalent myocardial infarction, stroke and cancer (n = 2,332) and separately excluding those with self-reported hypertension or hypercholesterolaemia (n = 4,943) to account for possible post-diagnosis changes in diet; (5) excluding those with incomplete food diary records (<7 days) (n = 2,219) to assess reporting bias; (6) excluding the top 1% of consumers for each sweet beverage separately to minimise the influence of outliers; (7) excluding those diagnosed with type 2 diabetes within the first 5 years of follow-up (n = 237); and (8) excluding those with HbA1c ≥6.5% (48 mmol/mol) at baseline (n = 486). Last, BMI, waist circumference and alcohol consumption were adjusted for as categorical covariates rather than continuous variables.
Examining total intake of sweet beverages (soft drinks, sweetened tea or coffee, sweetened-milk beverages, ASB and fruit juice) is also of interest, but summing total g/day was precluded as not all sweet beverages are comparable in composition. Thus, to assess the association of total sweet beverage intake and type 2 diabetes, intake was expressed as %TEI. As ASB do not contain energy, their consumption does not contribute to this variable. A dose–response relationship was examined using a restricted cubic spline with knots at the 25th, 50th and 75th percentiles, in Model 3 without TEI.
The impact of reducing sweet beverage intake on type 2 diabetes incidence was estimated as the per cent population-attributable fraction (PAF) of type 2 diabetes incidence attributable to high %TEI from sweet beverages, under the assumption of causality [21]. Taking into consideration the distribution of intake in the study population and achievable levels of intake, three PAFs with 95% CI were separately estimated (using Model 2), assuming that participants consumed sweet beverages at less than 10%, 5% or 2%TEI, treating %TEI from sweet beverages as a binary variable.
The potential effects of substituting a serving of a non-sugar-sweetened beverage (ASB: 336 g/day; drinking water: 280 g/day; unsweetened tea or coffee: 280 g/day), for a serving of a sweet beverage, were estimated. This was done by examining the difference between regression coefficients for the two beverages, when both beverages were included as continuous terms in a single model (Model 3) mutually adjusted for intake of other sweet beverages and with and without adjustment for TEI [22]; 95% CIs were computed using a variance–covariance matrix for the two beverages.
The association of drinking water and unsweetened tea or coffee per serving with incident type 2 diabetes was examined for the purpose of comparison, using Model 3.