METHODS
Findings from this systematic review and meta-analysis were reported
based on Preferred Reporting Items for Systematic Review and
Meta-Analysis (PRISMA) guideline [17].
Search strategy: A systematic literature review was conducted
by using the databases PubMed/Medline, Scopus, and ISI Web of Science up
to February 2020 with no language or time restriction. Details of the
search terms are provided in Supplemental Table 1 .
Furthermore, the reference list of the relevant articles was manually
searched.
Inclusion criteria: Published studies that met the following
criteria were included: 1) observational prospective studies conducted
on adults; 2) reported effect sizes (ESs) including hazard ratios (HRs)
or relative risks (RRs) or odds ratios (ORs) with corresponding 95% CIs
for the association between intakes of total water from all foods and
beverages as well as drinking water as the exposure of interest and
mortality from all causes and CVD as the outcome of interest.
Exclusion criteria: We excluded letters, comments, reviews,
meta-analyses, and ecologic studies. We also did not include studies
that performed on children or adolescences, those conducted among
chronic kidney disease or hemodialysis patients, critically ill
patients, and those enrollment acute respiratory distress syndrome
patients. All outcomes were classified based on the World Health
Organization’s international classification of disease criteria.
Data extraction: The selection and data extraction process
were executed by 2 independent reviewers (MM and FH). We extracted the
following data from each study: name of the first author, publication
year, study design, location of the study conducted, gender, the sample
size of the cohort, the age range at entry, duration of follow-up,
exposure, method of assessment of exposure, incidence of death,
comparison categories and relevant effect size along with 95% CIs and
list of confounders adjusted in the statistical analysis.
Risk of bias assessment: We used the Risk Of Bias In
Non-randomized Studies of Exposures (ROBINS-E) tool to assess the risk
of bias. The ROBINS-E tool comprises 7 domains: (1) bias due to
confounding, (2) bias in selection of participants into the study, (3)
bias in the classification of exposures, (4) bias due to departure from
intended exposures, (5) bias due to missing data, (6) bias in the
measurement of outcomes, and (7) bias in the selection of reported
results. Studies were categorized as low risk, moderate risk, serious
risk, and critical risk of bias under each domain. The results of risk
of bias assessment are presented in Supplemental Table 2 .
Statistical analysis: ORs, RRs, and HRs (and 95% CIs) for
comparison of the highest versus lowest categories of water intake were
used to calculate log ORs, RRs, and HRs ± SE. The analyses were
performed with the use of a random-effects model, in which we calculated
both Q-statistic and I 2 as indicators of
heterogeneity. As random-effects model can account for variation between
studies, it can provide more conservative results than a fixed-effects
model. For studies that reported effect sizes separately for drinking
water and other fluids, we first combined the estimates using the
fixed-effects model to obtain an overall estimate, and then, the pooled
effect size was included in the meta-analysis. In the study of Wu et al
that reported effect size for CKD and non-CKD patients separately, we
included only non-CKD patients in the meta-analysis. Studies that
investigated only CVD mortality in relation to water intake were also
considered in the meta-analysis of all-cause mortality because CVD
mortality accounts for a very high proportion of the all-cause
mortality. considered in the meta-analysis of all-cause mortality. We
conducted a sensitivity analysis, using a fixed-effects model, in which
each prospective cohort study was excluded to examine the influence of
that study on the overall estimate. In case of finding a significant
between-study heterogeneity, we performed subgroup analysis to examine
possible sources of heterogeneity. Between-subgroup heterogeneity was
examined through the fixed-effects model. Publication bias was examined
by visual inspection of funnel plots. Formal statistical assessment of
funnel plot asymmetry was also done with the use of Egger’s test. In
case of significant publication bias, the trim-and-fill method was used
to detect the effect of missing studies on the overall effect of
meta-analysis.
A method suggested by Greenland and Orsini was used to compute the trend
from the ORs/RRs/HRs estimates and their 95% CIs across categories of
water intake[18]. In this method, the distribution
of cases and the ORs/RRs/HRs with the variance estimates for ≥3
quantitative categories of exposure were required. We considered the
midpoint of water intake in each category. For studies that reported the
water intake as range, we estimated the midpoint in each category by
calculating the mean of the lower and upper bound. When the highest and
lowest categories were open-ended, the length of these open-ended
intervals was assumed to be the same as that of the adjacent intervals.
A two-stage random-effects dose-response meta-analysis was applied to
examine a possible non-linear association between water intake and
mortality. This was done through modeling of water intake and restricted
cubic splines with three knots at fixed percentiles of 10, 50, and 90%
of the distribution. Based on the Orsini
method[18], we calculated restricted cubic spline
models using generalized least-squares trend estimation method, which
takes into account the correlation within each set of reported
ORs/RRs/HRs. Then, all the study-specific estimates were combined with
the use of the restricted maximum likelihood method in a multivariate
random-effects meta-analysis. A probability value for non-linearity was
estimated using null hypothesis testing in which the coefficient of the
second spline was considered equal to 0. A linear dose-response
association was investigated using the two-stage generalized
least-squares trend estimation method. First, study-specific slope lines
were estimated, and then, these lines were combined to obtain an overall
average slope.30 Study-specific slope lines were
combined using a random-effects model. Statistical analyses were
conducted using STATA version 14.0. P <0.05 was
considered as statistically significant for all tests, including
Cochran’s Q test.