Association of Low- and No-Calorie Sweetened Beverages as a Replacement for Sugar-Sweetened Beverages With Body Weight and Cardiometabolic Risk

This systematic review and meta-analysis assesses the outcomes of 3 types of substitution for sugar-sweetened beverages in adults with overweight or obesity and with or without diabetes.


Risk of bias assessment
Two independent reviewers (NM and RZ) assessed Risk of bias for each included trial using the Cochrane Risk of Bias tool 2 . Assessment was done across 5 domains of bias (sequence generation, allocation concealment, blinding of participants and personnel, incomplete outcome data and selective reporting). The risk of bias was assessed as either low (proper methods taken to reduce bias), high (improper methods creating bias) or unclear (insufficient information provided to determine the bias level).

Data synthesis
Mean differences between the arms and their respective variance terms were extracted and used as the basis for analysis for each trial comparison. If mean differences were not provided, they were derived from available data using published formulas. 3 When median data were reported, they were converted to mean data with corresponding variances using established methods. 4,5 When no variance data were available, the standard deviation was taken from a trial similar in size, participants, and nature of intervention. We used change from baseline values from each study to calculate the mean differences between treatments for each substitution (LNCSBs for SSBs, water for SSBs and LNCSBs for water), otherwise we used post-intervention values (Supplementary Data S1-S20). For crossover trials and for within arm changes in parallel trials, we used a correlation coefficient of 0.5 in pairwise analysis to calculate standard errors. [6][7][8] To mitigate a unit of analysis error, when arms of trials with multiple interventions or control arms were used more than once, the corresponding sample size was divided accordingly. 3 Non-HDL-C values that were not reported were derived by subtracting HDL-C from total cholesterol values with standard errors derived from HDL-C and total cholesterol variance data using the inverse variance law. 9 For trials in which the change in BMI was not reported, but body weight was reported, then if baseline BMI was available, these data were used to calculate the height, which could then be used to calculate the end BMI and change in BMI. The change in BMI variance was imputed using published formula 3 and a correlation coefficient of 0.5. [6][7][8] Network meta-analysis, based on a frequentist framework, was conducted using the "network" suite of commands available in STATA version 15 (College Station, TX: StataCorp LP). We performed a random-effects network meta-analysis for each outcome to compare the three interventions simultaneously (LNCSBs, SSBs and water) in a single analysis by combining both direct and indirect evidence across the selected network of studies. The network meta-analysis synthesized all of the available evidence (direct and indirect effects) and quantified the pooled network effect of each intervention against every other intervention. We reported our results as mean differences (MDs) and 95% confidence intervals (CIs). To display the results for outcomes on the same plot, standardized mean differences (SMDs) and pseudo 95% CIs were calculated, whereby the SMD 95% CIs were proportionally scaled to the MD 95% CIs. The network diagrams were generated to show the interactions among the studies included in the network meta-analysis and to illustrate the available direct comparisons between treatments 10 .
Inconsistency was assessed in the direct, indirect, and network estimates. We assessed interstudy heterogeneity in the direct (each pairwise comparison arm) estimates using the Cochran Q statistic with quantification by the I 2 statistic, where an I 2 ≥50%, P<0.10 was considered an indication of substantial interstudy heterogeneity. We measured incoherence in the network estimates using both local (loop-specific and side-splitting) and global (design-by-treatment interaction model) approaches to evaluate the presence of incoherence. The loop-specific approach looked at the inconsistency in each closed loop in the network 10 while the side-splitting approach detected direct estimate comparisons that disagreed with the indirect evidence from the entire network 11 . The design-by-treatment interaction model was applied as a global approach to check simultaneously for inconsistency from all possible sources in the network 12 . If ≥10 comparisons were available for all comparisons, then we conducted a priori subgroup analyses by age, study duration, type of design, disease status, risk of bias and funding source.
Indirectness was assessed in the indirect comparisons by evaluation of intransitivity across the pairwise comparisons comprising the indirect estimates for important study characteristics including age, study length, sample size and percentage of males. Intransitivity was considered present if there was no overlap in the range between the pairwise comparisons.
Publication bias was assessed if ≥10 trial comparisons were available for all comparisons, We used comparison-adjusted funnel plots to assess funnel plot asymmetry 10 . Asymmetry around the line of the meta-analysis summary effect suggested evidence of small-study effects.

Grading of the evidence
We assessed the certainty of the evidence using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) system 13 with an extension for network meta-analyses 14 and other recent guidance [15][16][17] from the GRADE Working group. Evidence was graded as high, moderate, low or very low certainty. Network estimates of RCTs and the direct and indirect estimates that make-up these network estimates started at high certainty of evidence and interventions, and study conditions that limited the generalizability of the results among the direct estimates and intransitivity in the indirect estimates or if information for the network estimate was based upon only 1 direct study or only from indirect studies), imprecision (the 95% CIs for effect estimates were wide and crossed prespecified minimally important differences (MIDs) for benefit and/or harm in the direct and network estimates) and publication bias (evidence of comparison adjusted funnel plot asymmetry).

Water for SSBs
Water for SSBs

LNCSBs for Water
LNCSBs for Water

Water for SSBs
Water for SSBs

BMI (kg/m 2 )
Using restricted maximum likelihood (REML) random-effects model Note: P Inconsistency represents P-value for incoherence factor in the network estimates, and for heterogeneity in the direct pair-wise estimates Network Meta-Analysis eFigure 9. Network analysis with GRADE assessment of the certainty of the evidence comparing LNCSBs, SSBs and Water on body fat (%)

Water for SSBs
Water for SSBs

LNCSBs for Water
LNCSBs for Water

Water for SSBs
Water for SSBs

LNCSBs for Water
LNCSBs for Water

Water for SSBs
Water for SSBs

LNCSBs for Water
LNCSBs for Water

Using restricted maximum likelihood (REML) random-effects model
Note: P Inconsistency represents P-value for incoherence factor in the network estimates, and for heterogeneity in the direct pair-wise estimates Network Meta-Analysis eFigure 19. Network analysis with GRADE assessment of the certainty of the evidence comparing LNCSBs, SSBs and Water on HDL-C GRADE, Grading of Recommendations, Assessment, Development, and Evaluation; MID, Minimally Important Difference; NA, not available; LNCSBs, low-and nocalorie sweetened beverages; SSBs, sugar-sweetened beverages.

Water for SSBs
Water for SSBs

LNCSBs for Water
LNCSBs for Water

Water for SSBs
Water for SSBs

Using restricted maximum likelihood (REML) random-effects model
Note: P Inconsistency represents P-value for incoherence factor in the network estimates, and for heterogeneity in the direct pair-wise estimates Network Meta-Analysis eFigure 21. Network analysis with GRADE assessment of the certainty of the evidence comparing LNCSBs, SSBs and Water on SBP

Water for SSBs
Water for SSBs

LNCSBs for Water
LNCSBs for Water

Systolic BP (mmHg)
Using restricted maximum likelihood (REML) random-effects model Note: P Inconsistency represents P-value for incoherence factor in the network estimates, and for heterogeneity in the direct pair-wise estimates Network Meta-Analysis eFigure 22. Network analysis with GRADE assessment of the certainty of the evidence comparing LNCSBs, SSBs and Water on DBP

Water for SSBs
Water for SSBs

LNCSBs for Water
LNCSBs for Water

ALT (U/L)
Using restricted maximum likelihood (REML) random-effects model Note: P Inconsistency represents P-value for incoherence factor in the network estimates, and for heterogeneity in the direct pair-wise estimates Network Meta-Analysis eFigure 25. Network analysis with GRADE assessment of the certainty of the evidence comparing LNCSBs, SSBs and Water on AST

LNCSBs for SSBs
Water for SSBs

LNCSBs for Water
LNCSBs for Water