A prospective study of water intake and subsequent risk of all-cause mortality in a national cohort^,,,

Abstract

Background: Water, an essential nutrient, is believed to be related to a variety of health outcomes. Published studies have examined the association of fluid or beverage intake with risk of mortality from coronary diseases, diabetes, or cancer, but few studies have examined the association of total water intake with all-cause mortality.

Objective: We examined prospective risk of mortality from all causes in relation to intakes of total water and each of the 3 water sources.

Design: We used public-domain, mortality-linked water intake data from the NHANES conducted in 1988–1994 and 1999–2004 for this prospective cohort study (n = 12,660 women and 12,050 men; aged ≥25 y). Mortality follow-up was completed through 31 December 2011. We used sex-specific Cox proportional hazards regression methods that were appropriate for complex surveys to examine the independent associations of plain water, beverage water, water in foods, and total water with multiple covariate–adjusted risk of mortality from all causes.

Results: Over a median of 11.4 y of follow-up, 3504 men and 3032 women died of any cause in this cohort. In men, neither total water intake nor each of the individual water source variables (plain water, water in beverages, and water in foods) was independently related with risk of all-cause mortality. In women, risk of mortality increased slightly in the highest quartile of total or plain water intake but did not approach the Bonferroni-corrected level of significance of P < 0.002.

Conclusions: There was no survival advantage in association with higher total or plain water intake in men or women in this national cohort. The slight increase in risk of mortality noted in women with higher total and plain water intakes may be spurious and requires further investigation.

INTRODUCTION

Water, which is an essential nutrient, may be ingested as moisture in foods and beverages or as plain water (1, 2). Historically, both the popular media and professional nutrition literature have recommended plain water as the preferred beverage of choice (3) although there is no biological imperative for hydration via plain water (1, 2). The Institute of Medicine Panel on Dietary Reference Intakes for Electrolytes and Water limited their recommendations to Adequate Intakes on the basis of self-reported intakes of total water in the third NHANES. The panel concluded that the published evidence on the association of chronic disease risk in relation to water intake was insufficient to permit recommendations about intakes to modify disease risk (1).

A sizable body of evidence is available on intakes of specific beverages such as alcohol, coffee, tea, and sugar-sweetened beverages in relation to a variety of health outcomes (4–10). The association of water intake and body-weight management has also received considerable attention (11). However, the meager published evidence on modification of risks of leading chronic diseases such as heart disease, cancer, and diabetes has largely been limited to plain water or fluid intake (12–18). To our knowledge, intakes of total water and individual sources of ingested water in relation to health have received relatively little scrutiny. The purpose of this study was to fill these gaps and to examine the independent association of all ingested water sources and total water intake with subsequent risk of all-cause mortality in a nationally representative cohort of men and women.

METHODS

About the NHANES

We used diet- and mortality-linked public-domain data from the NHANES III (1988–1994) and the NHANES 1999–2004 for this prospective cohort study (19). The study used public-domain anonymized data and was not considered to be human-subject research by the City University of New York Institutional Review Board. NHANESs are conducted by the National Center for Health Statistics (NCHS)⁷ of the CDC (20). Each survey included a stratified, clustered probability sample of the noninstitutionalized US population. The NHANES III was a 6-y survey with public-domain data available for each 3-y phase and the combined 6-y period. Beginning in 1999, the NHANES became an annual survey, and public-domain data included combined data for 2 y. In each NHANES, every sample person is administered a household interview and a complete physical examination in a specially equipped Mobile Examination Center (MEC). Anthropometric, dietary, and laboratory data are collected during the MEC examination. Response rates for the MEC-examined sample in each of the NHANES used were >70% (21).

Dietary assessment

Each NHANES collected dietary intake information via a 24-h dietary recall that was administered by a trained dietary interviewer during the MEC examination (20). A number of standardized measuring guides were used to facilitate the estimation of portion sizes during the in-person recall. Dietary recalls were computer assisted and used a multiple-pass method (3 steps in the NHANES III, 4 steps in 1999–2001, and 5 steps in 2002–2004) with built-in probes to improve the completeness of the recalled diet.

Exposure assessment

Principal exposures in the current study were intakes of water from all sources (plain water, the water content of beverages, and water in foods) and total water.

Information on plain water intake in the NHANES III and the NHANES 1999–2004 was obtained by asking a separate question about water intake after the dietary recall. As described previously (22, 23), plain water included tap, spring, water cooler, or unsweetened, noncarbonated bottled water only. In surveys that were conducted after 2004, plain water information was integrated into the 24-h recall, and thus, these later surveys were not included in our study.

Water contents of beverages and foods

The public-domain database of the nutrient content of each food item reported in the NHANES recall includes information on the food’s moisture content. With the use of previously described methods (22, 23), we grouped foods that were reported in the recall into food or beverage categories. Beverages included, for example, alcohol; all types of liquid milk, shakes, and dairy beverages; carbonated or noncarbonated sweetened, unsweetened, or flavored beverages; coffee and tea; hot chocolate; fruit and vegetable juices; and juice drinks. Total water intake for each individual was the sum of water (moisture) that was contained in foods and beverages reported in the recall and information on plain water intake that was obtained after the recall. Each water intake exposure was operationalized as 1) a continuous variable, 2) weighted, sex-specific quartiles of intake as a categorical variable, and 3) quartiles of intake as a trend variable.

Outcome ascertainment

We used the NCHS public-domain mortality-linked files with follow-up completed through 31 December 2011 for the survey years that were previously mentioned (19). The NCHS determined the mortality status primarily through probabilistic matching to the National Death Index. Other sources of mortality-status information included the Social Security Administration, the Centers for Medicare and Medicaid Services, and death certificates (24). In a calibration study, the NCHS applied the NHANES III National Death Index matching algorithm to the NHANES I Epidemiologic Follow-up Study cohort with known mortality status. In this calibration study, mortality status of >98% of subjects was correctly classified by the NCHS matching procedures (25). Public-domain mortality data were subjected to perturbation to maintain anonymity; however, the vital-status information was not perturbed. Public domain–linked files included information on mortality from all causes, 9 leading causes, and a residual category of mortality in the US population that were based on International Classification of Diseases-10 cause-of-death codes. The file did not include the actual date of death but provided the person-months of follow-up from the date of the MEC visit or the household interview.

Analytic sample

All respondents who were aged ≥25 y in the NHANES III and NHANES 1999–2004 were eligible for inclusion in the analytic sample (n = 27,137). We excluded respondents for the following reasons: pregnant and lactating women (n = 813), missing plain water information (n = 1401), unreliable diet recall (n = 178), no energy intake in the recall (n = 2), missing linkage to mortality information (n = 25), and missing follow up time (n = 8). The final analytic sample of n = 24,710 for all surveys combined included 12,050 men and 12,660 women.

Covariates

Information on variables with putative associations with our exposures and the outcome was available for each survey. The variables included self-reported race-ethnicity, years of education, family income relative to the poverty threshold, smoking status, alcohol drinking status, any self-reported leisure-time physical activity, measured BMI, measured serum total cholesterol, and self-report of whether a doctor had told a respondent of a diagnosis of heart disease, stroke, diabetes, angina, heart attack, congestive heart failure, hypertension, cancer, or emphysema. In addition, total energy intake was available from the 24-h dietary recall.

Statistical methods

The person-time of follow-up was computed from the date of the MEC examination (in 1988–1994 and 1999–2004) to the last date that the person was known to be alive or to 31 December 2011. For descriptive purposes, we examined mean intake of water from each ingested source by categories of covariates. We also described the sociodemographic and life-style characteristics of the cohort by weighted sex-specific quartiles of total and plain water intakes.

We used Cox proportional hazards regression models to estimate the relative hazard of mortality from all causes in association with water intake with age as the underlying time metric. We fit the following 3 types of multiple-regression models: 1) separate single-nutrient models for total water and each of the 3 sources of water as exposures, 2) additive models that included each source of water (plain water, beverage water, and food water) simultaneously, and 3) substitution models with total water and 2 of 3 water sources simultaneously. For example, to examine the association with mortality because of the replacement of beverage water with plain or food water, we included plain water, food water, and total water but excluded beverage water. Because the sum of water from 3 different sources equaled total water, the regression coefficient for total water in the model, in effect, represented beverage water because plain water and food water were independent variables in the regression model. In other words, the model let us examine the effect of the replacement of the omitted variable (beverage water) with another water source while holding the third source of water fixed (26).

Multivariable models included all covariates previously mentioned and total energy intake. Respondents missing information on the family poverty-income ratio and alcohol intake were retained as separate categories in multivariable models because of the large numbers of individuals with missing values. Respondents who were missing information on the level of education (55 men and 46 women), measured BMI (129 men and 161 women), smoking status (10 men and 8 women), and activity (2 women) were excluded (185 men and 212 women).

All analyses were stratified by sex, were completed with the use of SAS version 9.2 software (SAS Institute Inc.) and SAS callable SUDAAN software (version 11.0.0; RTI International) (27), and included sample weights that were computed by the NCHS to adjust for survey nonresponse (28). Multivariable-adjusted HRs and 95% CIs that were associated with quartiles of water intake, with lowest water intake (first quartile) as the reference, are shown in tables. P values are from regression models in which quartiles of water intake were expressed as a trend and as a continuous variable.

To answer questions about the relation of total and source-specific water intakes with mortality, we tested several different hypotheses in our study. To adjust for the inflated type I error in this multiple-testing situation, a Bonferroni corrected 2-sided P < 0.002 (0.05 ÷ 22 different comparisons) was considered significant.

Sensitivity analysis

To test for the robustness of our results, we examined the influence of different definitions of exposure and outcome or cohort definitions on the associations examined. Different exposure definitions included water intake expressed as g/kg body weight and as g/1000 kcal. Serum osmolality, which is a biomarker of hydration (1, 2), was also available and was examined as an exposure instead of water intake. Serum osmolality was calculated from the estimation of concentrations of serum sodium, glucose, and blood urea nitrogen as assayed via the chemical analyzer (A Yick, National Center for Health Statistics, nchsed@cdc.gov, personal communication, 20 September 2016). The outcome variation included the exclusion of accidental and unknown causes of mortality.

To exclude the possibility that other variables with possible associations with our exposure and outcome may have been confounders of the observed associations, we also fit models with adjustment for total serum cholesterol, overall diet quality measured as the dietary diversity score, and, in women, noncontraceptive hormone use. The dietary diversity score (0–5) is a simple measure of the overall quality that indicates whether each of the 5 major food groups was mentioned in the recall (29, 30). In previous published work, this measure has been shown to be related to health outcomes (31, 32).

To exclude the possibility that deaths that occurred early in the follow-up period may have been due to an undiagnosed preclinical disease at baseline (reverse causation), which may also have been related to our exposure, we also examined the association between water intake and mortality after the exclusion of events that were recorded in the first 2 y of follow-up. We also examined the relation between water intake and mortality in age-stratified (<60 and ≥60 y) models. Finally, to test the proportional hazard assumption for water variables, we used Wald’s statistic to test if interaction terms (i.e., product terms) between the water variable and 2 time-dependent dummy variables, which indicated the 3 intervals of age during the follow-up period, were simultaneously zero (28). The intervals were chosen so that approximately one-third of deaths during the follow-up period occurred within each interval.

Adjustment for measurement error in dietary assessment

Because of the day-to-day variation of dietary intake, 24-h intake of a nutrient is not the usual intake of that nutrient (33, 34). Thus, we used regression-calibration methods to estimate usual intakes of water from foods and beverages (35). The method requires a repeat 24-h recall for at least a subset of the sample. For the surveys used in our analyses, a second recall was available for ∼10% of the NHANES III sample and for >90% of the NHANES 2003–2004 sample (a replicate recall was not available for the 1999–2000 and 2001–2002 surveys).

With the use of regression calibration, we first estimated the usual intake of water from foods plus beverages for each individual with adjustment for all variables that were used in the Cox proportional hazard regression models along with the order of the recall (day 1 or 2) and the day of the week of recall. These estimates of usual intakes of the sum of water ingested in foods and beverages were then used as a continuous exposure in Cox multiple-regression models. As previously mentioned, plain water intake in the surveys examined was not part of the 24-h dietary recall, but plain water was a covariate along with the other covariates previously mentioned.

RESULTS

Characteristics of the cohort at baseline

In the 1988–1994 and 1999–2004 combined survey cycles, American men and women reported mean total water intakes of 3.71 L/d (95% CI: 3.65, 3.77 L/d) and 2.98 L/d (95% CI: 2.93, 3.02 L/d), respectively. Non-Hispanic white ethnicity, higher education and income, any leisure-time physical activity, and current drinker status were associated with higher total water intake than in comparison categories (Table 1, Supplemental Table 1). Total and plain water intakes of ≥60-y-olds, never smokers, and individuals who reported no leisure-time physical activity were lower than in comparison categories (Table 1, Supplemental Table 1, Supplemental Table 2). The mean ± SD age of the cohort at baseline was 48.3 ± 0.3 y, and the median follow-up time was 11.4 y. In the cohort of 24,710 adults, 3504 men and 3032 women died from any cause.

TABLE 1.

Characteristics of the analytic cohort at baseline by quartiles of total water intake in men and women (NHANES 1988–1994 and 1999–2004)¹

	Quartiles of total water intake
	Men (n = 12,050)				Women (n = 12,660)
	1 (n = 3646)	2 (n = 3065)	3 (n = 2842)	4 (n = 2497)	1 (n = 3721)	2 (n = 3213)	3 (n = 3012)	4 (n = 2714)
Total water intake, g	<2525	2525–3372	3373–4487	>4487	<2043	2043–2742	2743–3657	>3657
NHANES III respondent, %	42.4	48.2	44.4	46.7	46.5	46.9	47.9	45.6
Non-Hispanic black, %	14.1	9.6	8.1	7.8	16.1	11.2	9.6	8.5
Aged ≥60 y, %	33.2	26.3	21.9	10.2	32.7	31.9	28.1	20.0
Education <12 y, %	29.0	19.6	20.2	19.9	31.4	22.0	18.6	17.7
PIR2 <130%, %	17.8	14.1	12.3	14.1	25.9	18.9	15.1	17.8
BMI (in kg/m2) ≥30, %	19.3	25.9	28.5	26.7	27.5	30.7	29.3	33.6
Current smoker, %	26.8	24.6	27.7	36.5	21.9	19.1	21.5	29.4
Current drinker, %	67.8	69.7	74.7	78.5	39.9	48.9	52.8	58.8
No leisure-time physical activity, %	32.1	21.6	22.9	21.9	43.8	33.3	27.1	27.1
Self-reported chronic disease,3 %	38.6	36.3	35.9	33.1	38.3	39.9	38.1	40.5
History of noncontraceptive hormone use, %	—	—	—	—	20.1	22.6	24.7	25.3
Water
Plain, g	486 ± 104	880 ± 15	1283 ± 18	2461 ± 57	416 ± 9	811 ± 16	1241 ± 16	2361 ± 40
Beverage, g	972 ± 13	1404 ± 16	1843 ± 25	2775 ± 55	751 ± 10	1064 ± 19	1320 ± 18	1789 ± 30
Food, g	493 ± 7	659 ± 9	744 ± 12	842 ± 11	400 ± 5	509 ± 7	596 ± 8	661 ± 10
Total
g	1951 ± 12	2943 ± 6	3870 ± 8	6078 ± 49	1567 ± 9	2384 ± 5	3157 ± 6	4812 ± 41
g/1000 kcal	1190 ± 18	1382 ± 15	1616 ± 16	2176 ± 34	1299 ± 27	1626 ± 17	1975 ± 21	2899 ± 73
g/kg	25 ± 0.2	36 ± 0.2	46 ± 0.3	71 ± 0.7	24 ± 0.2	35 ± 0.2	46 ± 0.2	67 ± 0.7
Energy, kcal	1918 ± 17	2427 ± 27	2729 ± 29	3260 ± 37	1444 ± 13	1697 ± 12	1869 ± 16	2046 ± 22
Fiber, g	13.6 ± 0.2	18.2 ± 0.3	19.9 ± 0.3	22.8 ± 0.4	10.8 ± 0.2	13.7 ± 0.3	15.4 ± 0.2	16.7 ± 0.3
Age at baseline, y	50.5 ± 0.4	48.7 ± 0.4	47.0 ± 0.4	42.4 ± 0.3	50.8 ± 0.4	50.6 ± 0.5	49.7 ± 0.5	46.8 ± 0.4
Follow-up, y	12.3 ± 0.2	13.0 ± 0.2	13.2 ± 0.2	13.6 ± 0.2	13.1 ± 0.2	13.2 ± 0.2	13.4 ± 0.2	13.4 ± 0.2
BMI, kg/m2	26.6 ± 0.1	27.5 ± 0.1	28.1 ± 0.2	28.0 ± 0.2	27.1 ± 0.2	27.5 ± 0.2	27.5 ± 0.2	28.3 ± 0.2
Serum total cholesterol,5 mg/dL	206 ± 0.9	206 ± 1.2	205 ± 1.3	204 ± 1.3	208 ± 1.1	209 ± 1.1	209 ± 1.2	206 ± 1.2
Serum osmolality,6 mmol/kg	279.7 ± 0.35 (279.0, 280.4)	279.9 ± 0.37 (279.1, 280.6)	279.2 ± 0.34 (278.5, 279.8)	279.0 ± 0.32 (278.3, 279.6)	278.4 ± 0.31 (277.8, 279.0)	277.8 ± 0.35 (277.1, 278.5)	277.9 ± 0.38 (277.1, 278.6)	277.1 ± 0.34 (276.4, 277.8)

¹Estimates were weighted. Total water is the sum of plain water, beverage water, and water present in foods.

²PIR, poverty-income ratio.

³Self-report of whether the following diagnoses were told to the subjects by a doctor: diabetes, hypertension, heart disease, heart attack, angina, stroke, congestive heart failure, cancer, and emphysema.

⁴Mean + SE (all such values).

⁵n = 11,502 men and 11,988 women.

⁶All values are means + SEs; 95% CIs in parentheses. n = 9871 men and 10,253 women.

Is total water intake related to risk of all-cause mortality?

In men, multiple covariate–adjusted total water intake was unrelated with risk of all-cause mortality (P ≥ 0.05) (Table 2, model 1). In women in the highest quartile of total water intake, there was a small increase in risk of all-cause mortality (HR: 1.24; 95% CI: 1.09, 1.42) (P-trend = 0.0023) (Table 3, model 1). However, CIs for quartiles 2 and 3 included an HR of 1.

TABLE 2.

Association of intakes of total water, plain water, beverage water, and food moisture with prospective risk of all-cause mortality in men at 13 y of follow-up¹

Exposure	First quartile	Second quartile	Third quartile	Fourth quartile	P-continuous2	P-trend3
Model 1. Is total water intake, regardless of water source, related to mortality?,4 g
Total water	1.0 (reference)	0.98 (0.88, 1.10)	0.93 (0.80, 1.08)	1.15 (0.96, 1.39)	0.09	0.4
Model 2. Is source-specific water intake, regardless of total water intake, related to mortality?,5 g
Plain water	1.0 (reference)	0.85 (0.72, 0.99)	0.87 (0.74, 1.02)	0.98 (0.83, 1.14)	0.05	0.9
Beverage water	1.0 (reference)	0.98 (0.86, 1.11)	0.99 (0.85, 1.15)	1.05 (0.90, 1.24)	0.4	0.6
Food moisture	1.0 (reference)	0.93 (0.81, 1.05)	0.91 (0.78, 1.05)	0.86 (0.74, 1.02)	0.1	0.1
Model 3. Is source-specific water intake, while keeping the other 2 water sources fixed, related to mortality?,6 g
Plain water	1.0 (reference)	0.86 (0.73, 1.01)	0.88 (0.75, 1.04)	0.99 (0.85, 1.16)	0.03	1.0
Beverage water	1.0 (reference)	0.98 (0.86, 1.11)	0.97 (0.84, 1.13)	1.02 (0.87, 1.21)	0.4	0.8
Food moisture	1.0 (reference)	0.94 (0.82, 1.07)	0.92 (0.79, 1.07)	0.87 (0.74, 1.03)	0.1	0.1
Model 4. Is the substitution of one water source with another related to mortality?,7 g
a. Gram-for-gram substitution of beverage water with plain water, keeping food water constant
Plain water	1.0 (reference)	0.86 (0.73, 1.01)	0.87 (0.74, 1.04)	0.94 (0.77, 1.13)	0.5	0.3
b. Gram-for-gram substitution of beverage water with food water, keeping plain water constant
Food moisture	1.0 (reference)	0.94 (0.82, 1.07)	0.92 (0.79, 1.07)	0.86 (0.73, 1.03)	0.05	0.05
c. Gram-for-gram substitution of plain water with beverage water, keeping food water constant
Beverage water	1.0 (reference)	0.98 (0.86,1.11)	0.98 (0.83, 1.15)	0.98 (0.81, 1.19)	0.5	0.7
d. Gram-for-gram substitution of plain water with food moisture, keeping beverage water constant
Food moisture	1.0 (reference)	0.93 (0.81, 1.06)	0.91 (0.78, 1.06)	0.85 (0.71, 1.01)	0.02	0.05

¹All values are relative hazards (95% CIs). All estimates were derived from multivariable Cox proportional hazards regression models. Total water is the sum of plain water, beverage water, and water present in foods. Cutoffs for quartiles 1–4 were as follows: for total water, <2525, 2525–3372, 3373–4487, and >4487 g, respectively; for plain water, <444, 444–944, 947–1772, and >1772 g, respectively; for beverage water, <973, 973–1494, 1495–2182, and >2182 g, respectively; and for food water, <488, 488–622, 623–880, and >880 g, respectively.

²P values of the regression coefficient associated with water intake variables (grams) as continuous variables.

³P values of the regression coefficient associated with water intake quartiles 1–4 expressed as a trend variable.

⁴Multivariable models included total water quartiles, race-ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, or other), education (<12 y, 12 y, some college, or at least college), poverty-income ratio (<1.3, 1.3–3.49, >3.49, or missing), BMI (in kg/m²; <18.5, 18.5–24.9, 25–29.9, or ≥30), smoking status (never, former, or current), alcohol drinking status (nondrinker, former drinker, current drinker, or unknown), any recreational physical activity (yes or no), any self-report of doctor-informed chronic disease (yes or no), and 24-h energy intake (in kilocalories; continuous). n = 11,865; cases: n = 3413.

⁵Single water source multivariable models included plain water (or beverage water or food water) quartiles, race-ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, or other), education (<12 y, 12 y, some college, or at least college), poverty-income ratio (<1.3, 1.3–3.49, >3.49, or missing), BMI (in kg/m²; <18.5, 18.5–24.9, 25–29.9, or ≥30), smoking status (never, former, or current), alcohol drinking status (nondrinker, former drinker, current drinker, or unknown), any recreational physical activity (yes or no), any self-report of doctor-informed chronic disease (yes or no), and 24-h energy intake (in kilocalories; continuous). n = 11,865; cases: n = 3413.

⁶Simultaneous multivariable models included plain water quartiles, beverage water quartiles, food water quartiles, race-ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, or other), education (<12 y, 12 y, some college, or at least college), poverty-income ratio (<1.3, 1.3–3.49, >3.49, or missing), BMI (in kg/m²; <18.5, 18.5–24.9, 25–29.9, or ≥30), smoking status (never, former, or current), alcohol drinking status (nondrinker, former drinker, current drinker, or unknown), any recreational physical activity (yes or no), any self-report of doctor-informed chronic disease (yes or no), and 24-h energy intake (in kilocalories; continuous). n = 11,865; cases: n = 3413.

⁷Substitution multivariable models included race-ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, or other), education (<12 y, 12 y, some college, or at least college), poverty-income ratio (<1.3, 1.3–3.49, >3.49, or missing), BMI (in kg/m²; <18.5, 18.5–24.9, 25–29.9, or ≥30), smoking status (never, former, or current), alcohol drinking status (nondrinker, former drinker, current drinker, or unknown), any recreational physical activity (yes or no), any self-report of doctor-informed chronic disease (yes or no), and 24-h energy intake (in kilocalories; continuous). In addition, model 4, a and b, included plain water quartiles, food water quartiles, and total water quartiles; and model 4, c and d, included beverage water quartiles, food water quartiles, and total water quartiles. n = 11,865; cases: n = 3413.

TABLE 3.

Association of intake of total water, plain water, beverage water, and food moisture with prospective risk of all-cause mortality in women at 13 y of follow-up¹

Exposure	First quartile	Second quartile	Third quartile	Fourth quartile	P-continuous2	P-trend3
Model 1. Is total water intake, regardless of water source, related to mortality?,4 g
Total water	1.0 (reference)	1.01 (0.88, 1.16)	1.09 (0.93, 1.27)	1.24 (1.09, 1.42)	0.002	0.002
Model 2. Is source-specific water intake, regardless of total water intake, related to mortality?,5 g
Plain water	1.0 (reference)	1.04 (0.90, 1.21)	1.05 (0.90, 1.24)	1.15 (1.00, 1.33)	0.008	0.07
Beverage water	1.0 (reference)	0.95 (0.79, 1.14)	1.01 (0.86, 1.18)	1.08 (0.92, 1.27)	0.1	0.3
Food moisture	1.0 (reference)	0.96 (0.83, 1.12)	0.92 (0.82, 1.03)	0.97 (0.85, 1.12)	0.7	0.6
Model 3. Is source-specific water intake, while keeping the other 2 water sources fixed, related to mortality?,6 g
Plain water	1.0 (reference)	1.05 (0.91, 1.22)	1.06 (0.90, 1.25)	1.16 (1.01, 1.34)	0.006	0.05
Beverage water	1.0 (reference)	0.96 (0.80, 1.15)	1.01 (0.86, 1.19)	1.09 (0.93, 1.28)	0.1	0.3
Food moisture	1.0 (reference)	0.96 (0.83, 1.12)	0.92 (0.82, 1.04)	0.96 (0.83, 1.11)	0.6	0.5
Model 4. Is the substitution of one water source with another related to mortality?,7 g
a. Gram-for-gram substitution of beverage water with plain water, keeping food water constant
Plain water	1.0 (reference)	1.03 (0.89, 1.20)	0.99 (0.84, 1.18)	0.97 (0.79, 1.20)	0.9	0.7
b. Gram-for-gram substitution of beverage water with food water, keeping plain water constant.
Food moisture	1.0 (reference)	0.96 (0.82, 1.11)	0.91 (0.81, 1.02)	0.92 (0.79, 1.08)	0.3	0.2
c. Gram-for-gram substitution of plain water with beverage water, keeping food water constant.
Beverage water	1.0 (reference)	0.94 (0.78, 1.13)	0.96 (0.80, 1.15)	0.97 (0.80, 1.17)	0.9	0.7
d. Gram-for-gram substitution of plain water with food moisture, keeping beverage water constant.
Food moisture	1.0 (reference)	0.96 (0.82, 1.12)	0.91 (0.78, 1.06)	0.85 (0.71, 1.01)	0.2	0.2

¹All values are relative hazards (95% CIs). All estimates were derived from multivariable Cox proportional hazards regression models. Total water is the sum of plain water, beverage water, and water present in foods. Cutoffs for quartiles 1–4 were as follows: for total water, <2043, 2043–2472, 2473–3657, and >3657 g, respectively; for plain water, <444, 444–944, 947–1776, and >1776 g, respectively; for beverage water, <695, 695–1078, 1079–1582, and >1582 g, respectively; and for food water, <344, 344–494, 495–703, and >703 g, respectively.

²P values of the regression coefficient associated with water intake variables (grams) as continuous variables from multivariable-adjusted models.

³P values of the regression coefficient associated with water intake quartiles 1–4 expressed as a trend variable from multivariable-adjusted models.

⁴Multivariable models included total water quartiles, race-ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, or other), education (<12 y, 12 y, some college, or at least college), poverty-income ratio (<1.3, 1.3–3.49, >3.49, or missing), BMI (in kg/m²; <18.5, 18.5–24.9, 25–29.9, or ≥30), smoking status (never, former, or current), alcohol drinking status (nondrinker, former drinker, current drinker, or unknown), any recreational physical activity (yes or no), any self-report of doctor-informed chronic disease (yes or no), and 24-h energy intake (in kilocalories; continuous). n = 12,448; cases: n = 2938.

⁵Single water source multivariable models included plain water (or beverage water or food water) quartiles, race-ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, or other), education (<12 y, 12 y, some college, or at least college), poverty-income ratio (<1.3, 1.3–3.49, >3.49, or missing), BMI (in kg/m²; <18.5, 18.5–24.9, 25–29.9, or ≥30), smoking status (never, former, or current), alcohol drinking status (nondrinker, former drinker, current drinker, or unknown), any recreational physical activity (yes or no), any self-report of doctor-informed chronic disease (yes or no), and 24-h energy intake (in kilocalories; continuous). n = 12,448; cases: n = 2938.

⁶Simultaneous multivariable models included plain water quartiles, beverage water quartiles, food water quartiles, race-ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, or other), education (<12 y, 12 y, some college, or at least college), poverty-income ratio (<1.3, 1.3–3.49, >3.49, or missing), BMI (in kg/m²; <18.5, 18.5–24.9, 25–29.9, or ≥30), smoking status (never, former, or current), alcohol drinking status (nondrinker, former drinker, current drinker, or unknown), any recreational physical activity (yes or no), any self-report of doctor-informed chronic disease (yes or no), and 24-h energy intake (in kilocalories; continuous). n = 12,448; cases: n = 2938.

⁷Substitution multivariable models included race-ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, or other), education (<12 y, 12 y, some college, or at least college), poverty-income ratio (<1.3, 1.3–3.49, >3.49, or missing), BMI (in kg/m²; <18.5, 18.5–24.9, 25–29.9, or ≥30), smoking status (never, former, or current), alcohol drinking status (nondrinker, former drinker, current drinker, or unknown), any recreational physical activity (yes or no), any self-report of doctor-informed chronic disease (yes or no), and 24-h energy intake (in kilocalories; continuous). In addition, model 4, a and b, included plain water quartiles, food water quartiles, and total water quartiles; and model 4, c and d, included beverage water quartiles, food water quartiles, and total water quartiles. n = 12,448; cases: n = 2938.

Is dietary source–specific water intake related to risk of all-cause mortality?

In men, the results from separate models for each of the 3 sources of water without control for intake of total water showed no associations with mortality (Table 2, model 2). In women, increasing plain water intake was associated with a 15% increase in hazard of mortality in the highest quartile, but the test for trend did not reach the Bonferroni-corrected level of significance of >0.002. Other sources of water were not related with risk (Table 3, model 2).

We asked the question of where dietary source–specific water intake was related to risk of all-cause mortality in another set of models by including all 3 sources of water in the model simultaneously, which let us examine the effect of the variation in intake of one source while controlling for the other 2 sources of ingested water. All associations were null in both men and women (P > 0.002) (Tables 2 and 3, model 3).

Is replacement of one source of water with another related to risk of all-cause mortality?

In both men and women, in the so-called substitution or replacement models (model 4), the substitution of beverage water with plain water (or vice versa) while keeping food water constant was unrelated with risk of mortality (Tables 2 and 3, model 4, a and c). Substitution of beverage water or plain water with food water, keeping plain water or beverage water constant also was unrelated with mortality (Tables 2 and 3, model 4, b and d).

Sensitivity analyses

Overall, the direction of associations observed in Tables 2 and 3 were mostly unchanged with water intake expressed as g/kg body weight and as g/1000 kcal, with the exclusion of accidental and unknown causes of death, or with additional covariates (serum cholesterol, dietary quality, or history of noncontraceptive hormone use in women) (Supplemental Tables 3 and 4). There was no association of mortality with serum osmolality. With the exclusion of events that occurred in the first 2 y of follow-up or after stratification by age, all observed associations were unchanged. Also, the proportional hazard assumption was not violated for the exposures examined in this study.

Measurement error–adjusted analysis

In multivariable regression models that included plain water intake as a covariate, there was no association of the sum of usual intakes of water in beverage and foods [as estimated with the use of the National Cancer Institute (NCI) method] with risk of all-cause mortality in men or women (P > 0.05). HRs did not differ from models in which 24-h estimates of the sum of food and beverages were exposures.

DISCUSSION

To our knowledge, this is the first report of an examination of the independent association of each individual source of ingested water and total water intake in relation to mortality outcome. In the current study, the reported mean total water intakes of 3.71 L/d (95% CI: 3.65, 3.77 L/d) in men and 2.98 L/d (95% CI: 2.93, 3.02 L/d) in women were in line with reference Adequate Intakes of 3.7 and 2.7 L/d, respectively. However, men and women in the lowest quartile reported intakes of <2.5 and <2.0 L/d, respectively, whereas those in the highest quartile reported intakes of >4.4 and >3.6 L/d, respectively. Nevertheless, these low or high intakes were not associated with different outcomes. Our analytic approach allowed us to ask 3 different questions about the relation of mortality with intake of water from total or each individual dietary source (plain water, water in beverages, and moisture in foods). In each of these analytic approaches, regression estimates were generally stable with relatively little variation, and contrary to popular beliefs, there was no evidence of a survival advantage with higher intake of total or plain water in men or women.

To assess prior studies for corroborative evidence, it is important to distinguish the relevant exposure as well as the outcome in each published study. Two published reports examined all-cause mortality in relation to water intake (16, 17). In an Australian cohort (n = 3858; age: ≥49 y), Palmer et al. (16) showed no association of fluid intake (defined as water in foods and beverages but not in plain water) with all-cause or cardiovascular mortality at 13 y of follow-up. Wu et al. (17) used a subset of the NHANES III cohort (n = 2182; mean age: >66 y), defined by the availability of biochemical measures of renal function, to create a chronic kidney disease (CKD) and a non-CKD group. At 15 y of follow-up, no association of total water with all-cause mortality was noted in the non-CKD group; in the CKD group, the highest quartile of water intake was linked to lower risk. The analytic approach of Wu et al. (17) did not account for family income or education, alcohol-use status, physical activity, or total energy intake. Moreover, the type of subgroup analysis raised questions about the representativeness of the subsample examined and the generalizability of the findings.

Higher total water intake can result from several different combinations of the 3 sources of ingested water. Generally, high plain water consumers are known to differ from high beverage water consumers in many ways and may also engage in other health-conscious behaviors (22). In the surveys we examined, compared with men, women consumed a higher percentage of their total water as plain water. Plain water intake expressed as g/1000 kcal or g/kg body weight was also higher in women. Therefore, in women, the suggestion of a slight increase in risk of mortality that was seen when high total water intake was due to plain water intake was unexpected. This study finding is contrary to results from reports of an inverse association of fluid or water intake with risk of heart disease and certain cancers (12–15), which may have been expected to lead to a favorable mortality outcome. The possibility that the observed association may have been spurious because of multiple-hypothesis testing or because of the methodologic limitations later discussed could not be excluded and requires confirmation from other cohorts with comparable exposures and outcomes. Although there are known sex differences in the risk of mortality and leading causes of death that are attributable to social, behavioral, and biological differences between sexes (36, 37), the reasons why women, but not men, with exposure to highest total and plain water intakes had an adverse mortality outcome in the current study is not clear.

Because healthy humans can respond to wide variations in water intake by controlling excretion, neither the Institute of Medicine committee nor the European Food Safety Authority panel established a safe Upper Intake Level (1, 2). Although dehydration and acute hyperhydration and the resulting hyponatremia are potentially life threatening, usual living and dietary conditions are well tolerated because of the neuroendocrine regulation of the water balance (1, 2). The finding of little difference in serum osmolality, which is a marker of hydration status, across quartiles of total water intake in our study and in other reports (1, 38) is in accord with the physiologic regulation of water homeostasis. Arguably, the maintenance of hydration status, despite wide variations in water intake in healthy humans, weakens the arguments for possible physiologic advantages of higher water intakes and may expectedly result in the null findings reported in the current study.

Higher food water may reflect the contribution of fruit and vegetables because these foods have high water contents. Therefore, higher food water is usually associated with lower dietary energy density. Both fruit and vegetable intake and low energy density are related to better diet quality and may expectedly predict improved health outcomes (29). In men, there was a suggestion of lower risk in association with higher food water, but all CIs included unity. Also, in models with adjustment for diet quality that was assessed as a dietary diversity score, the observed associations were unchanged.

We asked the study question in a high-quality, nationally representative cohort with the availability of possible confounders of the observed associations. Moreover, sensitivity analyses revealed the observed associations were generally robust. However, the results require careful interpretation with the following limitations kept in mind. First, our study was observational, and therefore, the conclusions were limited to observed associations from which causality should not be inferred. Second, assessments of food and beverage intakes used a 24-h recall, whereas information on plain water intake was obtained from a question asked after the recall. All methods of dietary assessment, including the 24-h recall and questionnaires, are known to be error prone (33, 34). A random-measurement error that was due to the within variability of dietary intake was expectedly present in our estimates of water intake and may have contributed to attenuation of the observed associations (34). Advances in the field have allowed for the estimation of usual nutrient intakes when a replicate 24-h recall measurement is available. In our study, information on ingested water intake was obtained from a combination of a questionnaire and 24-h recall, which differed in their error structures (39); therefore, it is not appropriate to use the NCI method to compute usual total water or plain water intakes. However, we used the NCI method to examine the association of measurement error–adjusted intake of beverages plus food water with plain water intake as a covariate and showed no association with risk of all-cause mortality in men or women.

Note that the question to elicit information about plain water intake in the NHANES III survey differed from that used in the 1999–2004 survey cycles. The NHANES III question queried about usual water intake in a 24-h period, whereas the 1999–2004 surveys asked about the amount of plain water consumed yesterday. To explore the possible impact of this difference, we tested the interaction of the survey cycle (NHANES III or 1999–2004) with each water variable in association with mortality in multivariable regression models. No significant interactions were noted (P > 0.05), and our results were unchanged with the inclusion of the survey indicator in regression models (Supplemental Tables 3 and 4). Finally, although we adjusted for multiple covariates that may possibly have been related with our exposure and outcome, we could not exclude the possibility of residual confounding because of unknown covariates or errors in the measurement of known covariates.

In conclusion, total water and each source of ingested water were not related with risk of all-cause mortality in men. The unexpected finding of a small increase in risk of all-cause mortality in women in the highest category of water intake should be interpreted cautiously because of the multiple-hypothesis testing and methodologic limitations and requires confirmation from other large cohorts. The study findings do not support a survival advantage that is due to higher water intake.