Consumption of sugar sweetened beverages, artificially sweetened beverages, and fruit juice and incidence of type 2 diabetes: systematic review, meta-analysis, and estimation of population attributable fraction

Objectives To examine the prospective associations between consumption of sugar sweetened beverages, artificially sweetened beverages, and fruit juice with type 2 diabetes before and after adjustment for adiposity, and to estimate the population attributable fraction for type 2 diabetes from consumption of sugar sweetened beverages in the United States and United Kingdom. Design Systematic review and meta-analysis. Data sources and eligibility PubMed, Embase, Ovid, and Web of Knowledge for prospective studies of adults without diabetes, published until February 2014. The population attributable fraction was estimated in national surveys in the USA, 2009-10 (n=4729 representing 189.1 million adults without diabetes) and the UK, 2008-12 (n=1932 representing 44.7 million). Synthesis methods Random effects meta-analysis and survey analysis for population attributable fraction associated with consumption of sugar sweetened beverages. Results Prespecified information was extracted from 17 cohorts (38 253 cases/10 126 754 person years). Higher consumption of sugar sweetened beverages was associated with a greater incidence of type 2 diabetes, by 18% per one serving/day (95% confidence interval 9% to 28%, I2 for heterogeneity=89%) and 13% (6% to 21%, I2=79%) before and after adjustment for adiposity; for artificially sweetened beverages, 25% (18% to 33%, I2=70%) and 8% (2% to 15%, I2=64%); and for fruit juice, 5% (−1% to 11%, I2=58%) and 7% (1% to 14%, I2=51%). Potential sources of heterogeneity or bias were not evident for sugar sweetened beverages. For artificially sweetened beverages, publication bias and residual confounding were indicated. For fruit juice the finding was non-significant in studies ascertaining type 2 diabetes objectively (P for heterogeneity=0.008). Under specified assumptions for population attributable fraction, of 20.9 million events of type 2 diabetes predicted to occur over 10 years in the USA (absolute event rate 11.0%), 1.8 million would be attributable to consumption of sugar sweetened beverages (population attributable fraction 8.7%, 95% confidence interval 3.9% to 12.9%); and of 2.6 million events in the UK (absolute event rate 5.8%), 79 000 would be attributable to consumption of sugar sweetened beverages (population attributable fraction 3.6%, 1.7% to 5.6%). Conclusions Habitual consumption of sugar sweetened beverages was associated with a greater incidence of type 2 diabetes, independently of adiposity. Although artificially sweetened beverages and fruit juice also showd positive associations with incidence of type 2 diabetes, the findings were likely to involve bias. None the less, both artificially sweetened beverages and fruit juice were unlikely to be healthy alternatives to sugar sweetened beverages for the prevention of type 2 diabetes. Under assumption of causality, consumption of sugar sweetened beverages over years may be related to a substantial number of cases of new onset diabetes.

for initial screening: main exposures were alcohol, coffee, or other dietary factors, rather than sweetened beverages or fruit juice; outcomes were not diabetes, either recruiting diabetes patients or assessing diabetes as covariate; studies were cross-sectional; studies recruited children; and publications are reviews, editorials, commentaries or other formats. ‡ See Table S1 for reasons for exclusion. § Seventeen cohorts, as a few cohorts published more than one article examining different beverages. One cohort met eligibility criteria after we obtained additional information. Full-text articles excluded (n = 12) ‡

Studies included in quantitative synthesis
(meta-analysis).
(n = 21) § Table S1. Studies reviewed in full text for eligibility and included or excluded for meta-analysis.* Cohort, country* Results from full text review and author contact * Decision for the present meta-analysis Identified as potentially eligible and reviewed † BRHS, UK 1,2 Cohort in Australia 3 EBSHP, US. 4 Ineligible, no information on any assessments of diets. We did not contact the authors. Excluded.
EPIC-NL, Netherlands. 5 Eligible, but the study was included in EPIC-InterAct. Excluded.
Hisayama, Japan 6 Ineligible. A diet and incident diabetes were assessed, but the authors confirmed no information on consumption of sweetened beverages and fruit juice. Excluded.
SUN Study, Spain 7 WHS, US 8,9 PHS, US 9 NIH-AARP, US 10 D.E.S.I.R, France 11 Eligible, but not included. Each cohort had information on dietary consumption and incident T2D. The authors could not respond to our request, because a resource was limited to conduct analyses we requested. Excluded.
HIPOP-OHP, Japan 12 Eligible, reported information of SSB consumption and incident T2D. After we contacted the authors, the authors provided sufficient information.
New unpublished estimates were used.

Identified as eligible
FMCHES, Finland 13 SSB and sugar-sweetened berry juices were reported separately. The authors did not respond to our request to combine the two.
Reported statistics on SSB were used. 4 approximated by consumption observed in women in ARIC 25 , considering similarity in chronological and demographic characteristics of women of the two cohorts.
FOS, US 26 The original paper examined SSB and hyperglycaemia only. One author provided estimates needed for meta-analysis, evaluating each of sweetened beverages and fruit juice, by methods as previously reported. 27 New estimates were used as provided.
ARIC, US 25 Fruit juice, SSB and ASB were combined. We requested to separate them and update estimates matched with our objectives. We did not obtain any response.
Only reported statistics were used. The authors stated no change in results after adjustment for measures of adiposity (body-mass index and waist-to-hip ratio). ‡ JPHC, Japan 28 Repeated measures were available in the cohort, but not used. 29 We requested analyses using them, but the authors decided not to do, being concerned of lack of peer-review of the specific methods.
Reported statistics based on 10-year follow-up were used. Repeated measures of diets were not used, although they could be used. 29 Estimates without overlap with InterAct were used.
SCHS, Singapore 34 Availability of ASB was not clear. Fruit juice was evaluated with vegetable juice. We requested information for the clarity and additional analysis, but did not receive any.
Reported statistics were used.
Black WHS, US 35 SSB and sugar-sweetened berry juices were reported separately. Results after adjustment for body-mass index were presented partially (estimates for the extreme categories) and presented by stratification. We requested the authors to do analysis combining the two beverage types, but could not obtain any information.
Reported statistics for SSB were used.
MESA, US 36 The author confirmed no availability of additional information. ‡ The article reported null associations between SSB consumption and incidence of T2D, but available in a review article. 37 Occupational cohort, Japan 38 We identified availability of fruit juice based on a publication on dietary assessment they used. Thus, we requested estimates for fruit juice consumption and incident T2D, as well as for SSB and ASB, with and without adjustment for adiposity measures. The author responded to our request.
Estimates for SSB, ASB, and fruit juice consumption were used, as provided. The update was unlikely to involve any bias.
Abbreviations: ARIC, Atherosclerosis Risk in Communities Study; ASB, Artificially-sweetened beverages; BRHS, British Regional Heart Study; CARDIA, Coronary Artery Risk Development in Young Adults Study; D.E.S.I.R., Data from an Epidemiological Study on the Insulin Resistance Syndrome; EBSHP, The East Boston Senior Health Project; EPIC,  14,15 Analyses were updated. A risk of bias was unlikely to be high.
NHS II 14,18,19 Analyses were updated. A risk of bias was unlikely to be high.
ARIC 25 SSB and ASB were not separated.
Iowa WHS 23 Only the conference abstract was published.
FOS 26 Modified substantially for updating the original analysis.
HPFS 39,40 Analyses were updated. A risk of bias was unlikely to be high.
Black WHS 41 Results were reported selectively.
MESA 36 Results were reported selectively.
EPIC-InterAct 31 A risk of bias was unlikely to be high.
HIPOP-OHP 50 Exclusion might have caused bias, losing 31% of participants during the follow-up. CARDIA 51 Main and subgroup analyses were internally and externally inconsistent. 21,22 KIHDS 52 Habitual consumption not measured well.
FMCHES 53,54 Generalizability to the modern population is concerning.  Table S5, for potential confounders. † Overall bias reflects possibility of bias specifically on the estimates used in the meta-analysis (see the first right column and the supplementary text on page 16). Sensitivity metaanalysis was performed after excluding these studies (Table S6). focused; yet, validity for assessment of sugar intake was similar (r=0.62 and 0.60) in the two studies. ‡ r represents correlation coefficients between estimates based on FFQ or diet history and estimates based on average of reference methods. We used energy-adjusted estimates corrected for within-person variations, if available. 65,66 s Q /s represents a ratio of a standard deviation of FFQ or diet history to that of a reference method; for diet records, standard deviations were assumed to be unbiased. If specific measures of r and s Q /s were not available for SSB, ASB, or fruit juice, variables related to refined sugars (disaccharides, sucrose, or carbohydrates) were used for SSB and ASB, and averages of variables related to sugars and vitamin C intakes were used for fruit juice. Averages of correlations were based on Z-transformed values 67,68 ; of ratios, log-transformed values. § For studies using objective measures of diagnosis (Table 1), PPV was assumed to be 1.0.Person-year was coded as presented or imputed by using the number of participants, the number of incident cases, and the maximum duration of follow-up. The presented numbers of cases and person-years were not corrected for positive predictive values (PPV). Thus, some values were different from those in Table 1. || Beverages were not assessed for associations with incident diabetes and not included in this meta-analysis. ** In EPIC-InterAct, FFQs were developed specifically in each of the eight countries of the consortium. Measures of validity and reliability were calculated by weighted averages of the measures from the eight cohorts 42,52,69-78 (available on request), for which weights were those of country-specific estimates to the overall estimates in EPIC-InterAct. The total number of adults were based on the number of adults contributing to the measures of validity for SSB and ASB. For fruit juice, N was 1,258. † † CARDIA reported associations of beverage consumption with hyperglycaemia. We included the study in this meta-analyses, considering the overlapping definitions of hyperglycaemia and incident type 2 diabetes (use of antidiabetic medications). PPV represents the proportion of patients with type 2 diabetes to patients with hyperglycaemia. * Age was considered adjusted for in a cohort using it as a time-scale in longitudinal analysis (NHS I, NHS II, HPFS, EPIC-InterAct, and E3N). Race and socioeconomic status (SES) were considered as adjusted for in some cohorts recruiting participants in a population homogenous in race/ethnic status and in occupation (FMCHES, NHS I, NHS II, HPFS, E3N, HIPOP-OHP, and occupational cohort in Japan). Another case of adjustment for SES was inclusion of education history in a multivariable regression analysis. Black WHS included a job status as a covariate, in addition to education history. † Dietary factors were not adjusted in main analyses in EPIC-InterAct and MESA. ARIC did dietary adjustment for intakes of alcohol, total calorie, and fibre only. EPIC-InterAct and MESA confirmed little influence of potential dietary confounders in secondary analyses. The lack of substantial influence was also confirmed in NHS I, NHS II, and HPFS. ‡ Clinical factors mean either family history of diabetes, use of anti-hypertensive or lipids-lowering drugs, or history of cardiovascular diseases, hypercholesterolemia, or hypertension. Family history of diabetes was adjusted for in FMCHES, NHS I, NHS II, KIHDS, HPFS, Iowa WHS, ARIC, JPHC, E3N, Black WHS, HIPOP-OHP, occupational cohort in Japan. EPIC-InterAct did not collect family history of diabetes among 51.7% of the random sub-cohort, and not used it in the main analysis, but sensitivity analysis excluding adults with known family history of diabetes confirmed little influence of the variable. 31 § Checked if different types of beverages were mutually adjusted for. NHS I, NHS II, and HPFS confirmed that mutual adjustment did not affect results. || Relative risk (95% confidence interval) adjusted for potential confounders except adiposity measures. Models adjusted for adiposity measures are presented in Table 2 and Figure  1. 'na' indicates that the authors did not report statistics for the specific estimate. For example, CARDIA and IowaWHS reported adiposity-adjusted estimates only. Figure S2. Non-linear associations of consuming sugar-sweetened beverages, artificially-sweetened beverages, and fruit juice with incident type 2 diabetes.
Estimates were obtained by random-effects meta-analysis adjusted for adiposity. The curves and P for a non-linear associations (P non-linear ) were obtained by cubic spline metaanalysis. 79 Solid lines are the central estimates of relative risks (RR) and shaded areas are the corresponding 95% confidence interval (CI). The analysis needed categorical estimates, rather than continuous estimates per one serving/day, thus we needed to drop the studies reporting only continuous estimates: for sugar-sweetened beverages, 13 estimates were used, not including European Prospective Investigations into Cancer and Nutrition Study (EPIC)-InterAct and Coronary Artery Risk Development in Young Adults Study (CARDIA), Iowa Women's Health Study and Multiethnic Study of Atherosclerosis; for artificially-sweetened beverages, 7 estimates were used, not including EPIC-InterAct, CARDIA; for fruit juice, 10 estimates were used, not including EPIC-InterAct, CARDIA. P for a linear association (P linear ) was obtained by meta-analysis using all estimates available. Using the limited categorical data, calibration for within-individual variability applied to categorical estimates 80,81 provided steeper effects with similar curves and wide CI (data not shown). Figure S3. Influence analysis for the prospective associations of consuming sugar-sweetened beverages, artificially-sweetened beverages, and fruit juice with incident type 2 diabetes. All estimates were obtained by random-effects meta-analysis adjusted for adiposity and for within-person variability of beverage consumption and precision of diabetes diagnosis. Overall estimates were based on analysis using all estimates from the studies presented for each beverage. The estimate accompanied to each cohort was based on meta-analysis excluding the study. Variations in relative risks ranged from -19% to +16% for SSB, -20% to +23% for ASB, and -7% to +16% for fruit juice.   (Table S3), selective reporting (yes or no, Table  S2), publication status (peer-reviewed or not), and mutual adjustment for three different beverages (yes or no, Table S4). Abbreviations: CI, confidence interval; EPIC, European Prospective Investigation into Cancer and Nutrition Study; PPV, positive predictive value; RR, relative risk. * Random effects meta-analysis was performed in each stratum, except for the estimates derived from fixed-effects modelling. All were adjusted for adiposity measures and calibrated for misclassification of exposure and outcome. † Median serving size of beverage consumption in the cohorts included in this present meta-analysis. Different studies defined one serving differently. ‡ Bias was determined by qualitative assessment (Table S2). § defined as r<0.4 compared to reference methods). Relatively low validity for dietary assessment (Table S3) was also used as a source of bias. As the validity measures were not all specific to each beverage, the results were interpreted cautiously as supplements. || defined as studies that conducted internal validation study to confirm if a dietary assessment was likely to capture habitual diet in a study population (Table S2) ** Using all cohorts available, but EPIC-InterAct was considered as a single cohort or separated. 31 The publication did not report the cohort-specific estimates adjusted for measures of adiposity. Thus, in the main analyses, we used the estimates combined within EPIC-InterAct. Additionally, the publication reported the cohort-specific estimates (11 cohorts in total from 8 countries) without adjustment for measures of adiposity. 31 Thus, the sensitivity to the aggregation was assessed here. Cohort-specific calibration for dietary measurement errors was applied. † † Iterative sensitivity analysis (10,000 times) was performed after incorporating quantitative bias and uncertainty 82 in different measures: dose-response estimates, within-person variability of beverage consumption, and precision of incident diabetes. Uncertainty of each was randomly drawn from each standard error. Out of 10,000 repeats, 2.5 th , 50 th (median), and 97.5 th percentiles were obtained for 95% confidence limits and point estimate of RR. ‡ ‡ Estimates were obtained after adopting specific unobserved, but realistic assumptions: 1) adiposity was measured with misclassification (r a ); 2) observed estimates adjusted for measured adiposity were biased to the extent related to r a ; and 3) estimates calibrated for r a were obtained by a formula following simulation extrapolation (see text and Figure S4). A recent article 83 indicated r a was greater than 0.73, thus assumed to be 0.7 or higher. Figure S4. Assessment of prospective association of consuming artificially sweetened beverages with incident type 2 diabetes after adjustment for assumed misclassification of adiposity measurements.
Estimates were obtained after adopting specific unobserved, but realistic assumptions: 1) adiposity was measured with misclassification (r a ); 2) observed estimates adjusted for measured adiposity were biased to the extent related to r a ; and 3) estimates calibrated for r a were obtained by a formula following simulation extrapolation, + ℎ (1+ ) ; θ=0 would produce observed RR=1.20 (0.98-1.52); θ=∞ would produce ln(RR) unadjusted for adiposity measure, RR=1.80 (1.13-2.46); θ=-1 would produce ln(RR) adjustged for potential misclassification of measured adiposity. The extrapolation for θ=-1 from the observable range, θ>0, was performed by non-linear association derived from θ ={0, 0.5, 1.0, 1.5, 2.0} of which functionality was demonstrated for another analysis. 84 A recent article 83 indicated r a was greater than 0.73, thus we assumed r a to be 0.7 or higher.

Supplementary Text Search Strategy
We undertook electronic searches, using the Internet browser (Firefox 27.0.1). We initially searched existing relevant reviews available at Cochrane Library, Centre for Reviews and Dissemination, Systematic Review Data Repository, PubMed, and OVID. We identified 15 reviews directly related to the topic, and then hand-searched potentially eligible publications for this meta-analysis. To identify additional studies, we systematically searched electronic databases, using specific search terms as described below. No restriction of time or language was applied. This search was performed on May 31, 2013, and repeated on February 10, 2014. We identified all articles included in prior meta-analyses [85][86][87] and additional studies. Addition of "fizzy", "artificially sweetened beverages" did not change the search results.

Identification of studies and contact to authors
The articles reviewed in full-text are presented in Table S1. From each cohort, we hand-searched multiple publications and examined availability of information on dietary consumption and incident T2D. We contacted authors of the identified articles between October and December in 2013 and requested information needed for this meta-analysis to minimize publication bias. If a publication reported estimates based on either continuous or categorical variables of beverage consumption for both adiposity-adjusted and unadjusted associations, we did not request additional data. However, when we requested any additional information, we requested both continuous and categorical estimates. We sent a reminder two weeks after a contact, in case of no reply.

Quality assessment
We collected information to identify potential bias, in concordance with A Cochrane Risk of Bias Assessment Tool 88 and for Non-Randomized Studies of Interventions (ACROBAT-NRSI) 89 . As instructed by the Bias Assessment Tool 88 , a 'high', 'low', or 'unknown' risk of bias in each study was assigned to seven different bias domains and overall risk of bias ( Table S2). Considerations of overall bias are written here, followed by those for bias in seven domains. Influence of potential sources of bias was examined in sensitivity analyses by testing heterogeneity due to presence or absence of bias, by performing meta-analysis after excluding studies with a certain type of bias (Table S5, Table S6), or by incorporating quantitative bias in meta-analysis (see below). Influence of a single study can be inferred in influence analysis (Figure S3).  Overall bias: We acknowledged ACRBAT-NRSI's recommendation that an observational study is not likely to have 'low' risk of bias. 89 Then, we assigned 'high' or 'unknown' risk of bias to each study. 88 We considered whether or not multiple sources of bias would impact estimated effects and uncertainty. We did not assign 'high' risk of bias even when studies were likely to have domain-specific bias, if the sources of bias were not likely to impact study estimates or uncertainty substantially and plausibly. Thus, 'unknown' overall bias was assigned to several studies, although they might have bias in some domains 26,31,41,48,49,[52][53][54] , because there was no strong plausibility that bias caused substantial impact on effect estimates to be used in this meta-analysis. Studies rated to have 'high' risk of bias had <20% of weights in the main meta-analyses, and exclusion of these studies did not change results (see sensitivity analysis, below).
Here, we describe considerations of seven domains: confounding, selection, measurement errors, misclassification of exposure, missing data, outcome assessment, and selective reporting. Influence of a single source of bias on overall risk of bias is also described. A single source of bias was not necessarily considered influential on overall bias as described above, with regards to biological plausibility and sensitivity analyses.
 Confounding: Residual confounding is likely in any of observational research. Thus, a 'low' risk of bias was not assigned to any studies, as anticipated. 89 A 'high' risk was assigned to the E3N cohort, which was likely to have residual confounding by adiposity in analysis of ASB. 32,[43][44][45] Potential confounders adjusted for in each study are summarized in Table S4. With exception of adiposity measures, five studies [14][15][16][17][18][19]31,36 confirmed little influences of bias due to confounding by each of socio-demographic variables, lifestyle factors, and clinical variables (family history of diabetes, mediaction use, and prevalent disaeses), in a multivariable model specified in each study. However, comparison between crude and adjusted analyses indicates confounding in analysis of each beverage, particularly of ASB (Table S4; Table 2).  Selection: Selection of participants into a study would cause bias, if selection was related to both beverage consumption and incidence of T2D. This possibility was not identified in any of the studies. Selection was partly based on completion of data in any studies and considered in the assessment of the domain of missing data (see below). We considered that 'high' risk of bias of this point would not necessarily influence overall quality of point estimates. On the other hand, we assigned 'high' risk of bias to Atherosclerosis Risk in Communities Study (ARIC) and Singapore Chinese Health Study (SCHS), because of misclassification of beverage types. Because influence of this particular bias on study results was likely to be present, 'high' risk of overall bias was rated for these two studies.  Measurement errors: Because any dietary assessments involve measurement errors, any studies had risks of this bias. However, inclusion in this meta-analysis incorporated the quantity of the bias partly (see next subsection). Accounting for it, a 'high' risk of bias was assigned to studies that did not verify quality of a dietassessment method within a study population 23,25,26,49,[52][53][54]59 or studies that assessed diets during a month or less and did not confirm long-term reproducibility of dietary measures. 12,51,52  Misclassification of exposure: We focused on exposure assessment during follow-up. 'Low' risk of bias was assigned to studies that assessed dietary exposure repeatedly and incorporated them in analysis. [14][15][16][17]26,35 By contrast, 'high' risk was assigned to studies relying on baseline dietary measurements. We regarded that this bias was not likely to influence results substantially, generally causing bias in false-negative findings.  Missing data: All studies excluded participants with missing information. The number of participants excluded was not large in each study, and the exclusion was considered to be unlikely to cause bias. 'High' risk was assigned to two studies 12,34 because participants were excluded based on missing outcomes, which might cause attrition bias, during the follow-up: deaths in SCHS (15% of adults) and unknown loss to follow-up in the High-risk and Population Strategy for Occupational Health Promotion Study (HIPOP-OHP, 31% of adults). While SCHS excluded prevalent cases of cardiovascular diseases and cancer, HIPOP-OHP recruited adults in a population at high risk of deaths due to cardiometabolic diseases. The latter was considered as having high risk of overall bias due to missing outcome related to the association of interest.  Outcome assessment: 'Low' risk was assigned to studies that attempted to minimize both false-positive and false-negative cases in a whole cohort by using objective information on incidence of T2D. 'High' risk would have been assigned if a differential misclassification had occurred. This bias was not indicated in any cohorts. Coronary Artery Risk Development in Young Adults Study (CARDIA) was rated as having 'high' risk of this domain: CARDIA examined a prospective association of beverage consumption with a risk of developing hyperglycaemia that included also incident type 2 diabetes; combined with selective reporting (see the next point), the overall bias in this study was rated as high.  Selective reporting: 'High' risk was assigned to two studies, on the basis of multiple analyses in different sub-groups, that presented inconsistent methods across articles from each cohort for similar research questions. 21,22,28,30 . These two studies, including CARDIA, were considered as having high risk of bias, because of uncertainty if main effect estimates were selected validly. 'High' risk was also assigned to the other four studies that reported estimates of associations selectively on the basis of whether or not findings were significant 23,25,35,36 , but each presentation was judged as independent of validity of effect estimates. 'Unknown' risk of bias was assigned to four studies which we requested unpublished estimates of beverage-diabetes associations. We did not find plausible explanation that each reporting led to biased estimates, and therefore we considered that this component is important to highlight, but did not affect overall risk of bias.

Adjustment for within-person variation of beverage consumption
In epidemiologic studies on dietary habits and other exposure related to chronic diseases, random within-person variability is concerning as a source of bias. 90 We applied statistical correction for the potential bias, using measures of the within-person variability, in addition to false-positive ascertainment of self-reported T2D (Table S3) 91,92 . Information extracted and assumptions are presented here in compliance with PRISMA.
We extracted correlation coefficients (r) between estimates from the two methods compared; ratios of two standard deviations (SDs) from the two dietary methods (s obs /s ref ); and sample sizes (n). For studies without these measures derived within a study population 13,20,23,25,38 , we extracted information from external sources, assuming consistency of within-person variations of dietary assessments in different cohorts. 90,[93][94][95][96][97][98] This assumption was supported by prior research comparing different FFQs in an independent population 67,99 and also by Atherosclerosis Risk in Communities Study (ARIC) that confirmed estimates of sugar intakes by FFQ externally developed were correlated with a biomarker of sugar intakes. 100 Kuopio Ischaemic Heart Disease Risk Factors Study in Finland evaluated beverage consumption by 4-day diet records implemented only at baseline. 63 A single 4-day diet record is unlikely to capture habitual diets. 66 Thus, we assumed similarity between r of single 4-day diet records and r of a seasonal variation of diet within a year; and took the measure from another study assessing diets among men in North Sweden 52 selected by demographic similarity. 63,67 We also assumed no error in a between-individual SD in the cohort (s obs /s ref , =1.0).
European Investigation into Cancer and Nutrition Study (EPIC)-InterAct carried out analyses pooling cohorts across Europe. 31,33 We extracted within-person variations of participating cohorts 42,52,[69][70][71][72][73][74][75][76][77][78] (available on request) and pooled estimates by weights contributing to EPIC-InterAct's estimates 31 , averaging r and s obs /s ref after Zand log-transformation. 67,68 When there was no information on measures of validity specifically for each type of beverages, we used information on foods or nutrients that were likely to have similar measures of the variations of consumption, as performed on another topic. 101 For example, if only non-alcohol beverages were assessed, we used them. If nutrients, not foods, were assessed, we extracted information on sucrose, disaccharides or total carbohydrates as surrogates for SSB and ASB; these sugars are not in ASB, but we assumed similarity in within-person dietary variations between ASB consumption and sugar intake. We adopted a model used in prior meta-analyses 90,[93][94][95][96][97][98] to adjust diet-disease association for a within-person dietary variation: f (risk)=α+β true •x true and β true =β obs /γ, where x true is true dietary factor and β true is unobserved, unbiased log(RR) without a within-person variation (σ 2 within ). β is attenuated to be β obs by degree of γ, a attenuation factor, representing a variance ratio: σ 2 within /(σ 2 within +σ 2 between ). In each cohort, γ was calculated by γ =s ref / s obs •r, given a linear regression of x true = α+β•x obs . Dietary habits were measured repeatedly in six cohorts to minimize regression dilution or a degree of attenuation (Table S2; Table S3). 14,[16][17][18][19]21,22,27,35 To account for this, γ was recalibrated for the number of repeated measures and measures of reproducibility 90 . Measures of reproducibility were obtained from existing literature along with those of validity (data not shown).

Adjustment for precision of incident type 2 diabetes
Some studies used self-reported T2D only [15][16][17][18][19]23,28,35,102 (Table S3), raising possibility of false-positive diagnosis expressed as positive predictive value (PPV). Thus, correction for PPV<1.0 was applied. 91,92 We assumed PPV=1 for studies using objective measures of T2D diagnosis. In CARDIA. two studies on beverages 21,26 ascertained cases with hyperglycaemia, not T2D. Thus, calibration in CARDIA was applied throughout in the meta-analysis, assigning PPV as a proportion of T2D cases among those with incident hyperglycemia 57,62 Assessment of heterogeneity Meta-regression was used to assess potential sources of heterogeneity (Table S5). Estimates used were those adjusted for adiposity, within-person dietary variation, and precision of T2D; results were similar in post hoc meta-regression using estimates without adjustment for within-person dietary variations (data not shown). Variables assessed by meta-regression were pre-specified, including study-specific factors: geographical location (the United States or Europe, or Asia, categorized post hoc), average age (years), sex (% men), average BMI (kg/m 2 ), follow-up duration (years), absolute risk of T2D (cases / person-years), methods of dietary assessments (FFQ, diet history), and methods of T2D diagnosis (self-reported, others). Publication status (peerreviewed or not), selective reporting (yes or no), and mutual adjustment for three beverage types were assessed post hoc, identified to be potentially important after data extraction. In stratified analysis for a continuous variable, a median across identified cohorts was used.
Independent sources of heterogeneity were selected by meta-regression with forward-variable selection. If variables in meta-regression showed P<0.20, the variable with the lowest P-value was retained in the model. Then, adjusting for the variable retained, mete-regression was repeated for remaining variables. If any of additional variables did not produce P<0.20, the model was considered best fitted. A variable with P<0.10 was considered as a significant source of heterogeneity and meta-analysis stratified by the factor was performed. 103

Sensitivity analysis
We performed sensitivity analysis to confirm robustness of our findings against decision of modelling, assumptions and different use of available information (Table S6). Sensitivity analysis included influence analysis 46 by excluding a single study and repeating random-effects meta-analysis ( Figure S3). We also took an iterative stochastic sensitivity analysis, accounting for additional uncertainty of adjustment for within-person variations and precision of T2D diagnosis. 46,79,82 Therefore, to confirm stability of our main analysis, we repeated the main meta-analysis (10,000 times) after ln(RR), γ and PPV were randomly drawn from each distribution determined by each central estimate and standard errors (SE). SE of ln(RR) was obtained by doseresponse estimation; SE of γ, derived from information available in published records assessing within-person variations; SE of PPV, derived from 95% confidence interval (CI) calculated by Wilson score interval 104 or, when PPV=1, the rule of three. 105 In each iteration, trim-and-fill analysis was applied, assuming these estimates were the least subject to bias. 106 Medians of ln(RR) and 95% confidence limits of ln(RR) were used as the estimate accounting for uncertainty of our approach. 82

Sensitivity analysis for residual confounding
We included a simulation-based analysis to examine influence of residual confounding. We recognized the limitation of the abovementioned correction for within-person variations of the main exposure, but not of confounders. 6 Residual confounding by within-person variation of adiposity measures would be expected and crucial source of bias. 31,107 Thus, adjustment for the bias could have been done, using measures of the withinperson variation of adiposity and associations of adiposity with beverage consumption and incident T2D. Because the information was not available in any studies, we undertook simulation extrapolation (SIMEX), an imperfect, but useful, technique when structure of measurement errors is likely to be complex or unknown. 84 In SIMEX, we used estimates after adjustment for potential publication bias by trim-and-fill method, assuming these estimates were the least subject to bias. Inference became similar without trim-and-fill method (data not shown). In SIMEX, first, we assumed that adiposity was measured with within-person variation (r a ): when r a =1.0, observed ln(RR) adjusted for adiposity would be unbiased; when r a =0, adiposity measures would be a random variable and ln(RR) would be ln(RR) unadjusted for adiposity. We assumed r a =0.9, 0.8, and 0.7 based on a study assessing validity of adiposity measures. 83 We assumed that ln(RR) adjusted for measured adiposity could be expressed as ( ) = + • ℎ −1 ( ) (Model A). This model supports that, r a =0 would make ln(RR) equivalent to ln(RR) unadjusted for adiposity. Then, α and β could be readily solved algebraically by r a , observed ln(RR) unadjusted for adiposity measures, and observed ln(RR) adjusted for adiposity measures.
Separately, a non-linear SIMEX formula was modelled 84  . This denotes that, when =0, ln(RR) was ln(RR) adjusted for adiposity, given r a assumed; and when =∞, ln(RR) would be unadjusted for adiposity measures.

Estimation of type 2 diabetes events over 10 years from 2010 attributable to consumption of sugarsweetened beverages in the United States and in the United Kingdom.
We estimated population attributable fraction (PAF) for T2D due to SSB consumption. We evaluated adults The contemporary national surveys strengthen the implication from the present meta-analysis, beyond estimation solely relying on cohorts limited in generalizability of a source population. 110 Moreover, the use of individuals' data can avoid some assumption often adopted in analysis of PAF and other estimates of social impact due to dietary exposure, for example implausible assumption of normal distribution of dietary consumption. [110][111][112] Overall, in each survey, we (1) estimated habitual consumption of SSB among adults; (2) predicted 10-year risk ('assumed control risk', ACR 113 ) of developing T2D of each adult; (3) estimated separate ideal 10-year risk (R i ) for each adult if SSB consumption was reduced to zero; and (4) estimated (ACR-R i ) for each adult and Σ(ACR-R i ) × population size as a number of cases attributable to SSB consumption in a population. PAF was derived as Σ(ACR-R i )/ Σ(ACR).
As a simple example, if one adult consumed 1 serving/day of SSB and had ACR of 0.10 and if SSB consumption became zero, his or her risk (R i ) would be 0.10/1.13=0.088, where 1.13 is RR adjusted for adiposity. This calculation was applied to all adults, and pooling them as Σ(ACR) and Σ(R i ), the populationbased estimates were obtained. 110 This estimation has advantage that there is no need of assumption in exposure distribution.
We used two RR separately: RR unadjusted for adiposity and RR adjusted for adiposity (1.18 and 1.13, respectively). We did not use RR unadjusted for within-person dietary variation, because the 10-year risk prediction was based on T2D risk factors unadjusted for within-person variations; and because reduction of SSB consumption in a population is likely to occur, involving a random within-person fluctuation. In addition to uncertainty in this probability-weighting analysis, the uncertainty of RR was incorporated by one thousand iteration varying RR as normally distributed with variance of ln(RR). Potential lack of generalisability of RR and heterogeneity of RR were secondarily examined in sensitivity analysis for PAF.
The next subsections describe each estimation of PAF in the US and the UK; validation analysis implemented by using the US observations; and sensitivity analysis varying RR and incorporating I 2 to calculate PAF.

Population attributable fraction for T2D due to SSB in the United States
In the US NHANES, we evaluated 4,729 non-diabetic adults who represented 189,075,538 adults in the US 2009-2010 according to sampling probability, after excluding 5,928 individuals: 4,319 children and adolescents (age<20 years) and 1,033 adults with prevalent diabetes (13.7% in weighted analysis) defined by reported diagnosis or anti-diabetic drug use or by fasting glucose ≥7.0 mmol/L or glycated haemoglobin ≥6.5%. 114 Habitual consumption of SSB was estimated by using two 24-hour recall and a dietary screener questionnaire simultaneously analysed through a method to minimize within-individual dietary variation. 115 Using a set of risk factors (z) for T2D, 10-year risk of T2D was estimated by a risk-prediction algorithm developed in ARIC and validated in Multiethnic Study on Atherosclerosis (MESA), a community-based cohort in US. 116,117 The formula was a logistic function, ACR=1/(1+exp(−X|z)). The original prediction was for a 9-year risk of T2D, and thus converted to a 10-year risk as Pr=1− (1−Pr) 10÷9 . The model was developed among adults younger than 65 years, adopting a rare-disease assumption. Mortality, reducing T2D cases identified over 10 years, was accounted by age-sex-specific mortality due to non-diabetes cause (1-annual mortality) 10 years , based on US vital statistics. 118 Population attributable fraction for T2D due to SSB in the United Kingdom We used the UK NDNS data collected in 2008/2009-2011/2012. 109 Sampling weights were applied, which appeared, unlike US NHANES, not to estimate an absolute number of adults with a certain condition in UK. Thus, to estimate absolute numbers of T2D cases attributable to SSB, we used age-sex-specific population sizes in UK in 2010 as the source population (47,704,520 adults in total, aged 20 or older). 119 Of the UK NDNS, we evaluated 1,932 non-diabetic adults, after excluding 2,096 children and adolescents (age<20 years) and 128 adults with prevalent diabetes (6.2% in weighted analysis, 2.9 million in UK) defined by diagnosis, anti-diabetic drug use, or compliance to an anti-diabetic diet assessed through an interview; or by fasting glucose ≥7.0 mmol/L or glycated haemoglobin ≥6.5%.
Habitual consumption of SSB was estimated by 4-day food records with within-individual dietary variation minimized. 115 Ten-year risk of T2D was estimated by a risk-prediction algorithm, QDScore®-2013 120 , developed in the prospective analysis of nation-wide electronic records collected in UK general practice; validated externally 121,122 ; and made publically available for research purpose. 120 The formula allowed estimation of ACR over 10 years from basic demographic variables, deprivation index, smoking status, use of an oral corticosteroid, use of an anti-hypertensive drug, prevalent cardiovascular diseases, family history of diabetes. [120][121][122] Family history of diabetes and Townsend deprivation index were not available in NDNS. These were imputed, respectively, by the population average as found in the nation-wide electronic record 121 and by household income, as recommended previously 121 .

Validation of 10-year risk prediction
The present estimates showed that the US population had approximately two-fold greater SSB consumption, T2D prevalence, and T2D risk compared to the UK population (Table S7). These ecological differences were comparable to those observed in prior international studies 123, 124 . In addition, 10-year risk in US and UK were comparable to previous population-based estimates. 116,117,121,122 These partly supported validity of our estimates.
We assessed validity of 10-year risk prediction in US NHANES, using the 1999-2000 and 2009-2010 cycles. We first estimated prevalent T2D predicted by NHANES 1999-2000. Adults with T2D in 1999-2000 were assumed to have T2D in 2009-2010, with Pr= (1-annual mortality) 10 . For non-diabetic adults, 10-year risk prediction was applied as described above. Then, the two numbers of T2D cases from T2D cases and non-cases . Thus, we considered validity of 10-year risk prediction to be sufficient in this work.
Sensitivity analysis varying relative risk and incorporating its heterogeneity to estimate PAF PAF is generally estimated on the basis of an association of exposure with an outcome. The measure of association often has heterogeneity across populations across measured or unmeasured factors; and uncertainty in its generalisability to a target population. 125 To present how PAF may vary and how precise PAF could be across different values of RR, we performed sensitivity analysis by varying RR and incorporating a measure of heterogeneity, I 2 (25%, 50%, and 75%). The standard error (SE) for RR was derived from estimates adjusted for adiposity measure (RR=1.13, 95% CI=1.06-1.21). I 2 was converted to a measure of between-population variance, τ 2 , by using I 2 = τ 2 (τ 2 + SE 2 ) ⁄ . We specified RR to be 1.00 to 1.42 incremented by 0.03 (15 values of RRs). Each of RRs simulated 1,000 times to vary by the degree of the pooled variance, √τ 2 + SE 2 . For each RR (15 × 1,000), PAF was calculated by using each of the US and the UK datasets. 95% CI of PAF across RRs was based on 2.5 th and 97.5 th percentiles of 1,000 repeats. 125 The results are displayed in Figure S5. As expected, the greater I 2 , the wider 95% CI.