Multivitamin-Mineral Use Is Associated with Reduced Risk of Cardiovascular Disease Mortality among Women in the United States^,,,

Abstract

Background: Multivitamin-mineral (MVM) products are the most commonly used supplements in the United States, followed by multivitamin (MV) products. Two randomized clinical trials (RCTs) did not show an effect of MVMs or MVs on cardiovascular disease (CVD) mortality; however, no clinical trial data are available for women with MVM supplement use and CVD mortality.

Objective: The objective of this research was to examine the association between MVM and MV use and CVD-specific mortality among US adults without CVD.

Methods: A nationally representative sample of adults from the restricted data NHANES III (1988–1994; n = 8678; age ≥40 y) were matched with mortality data reported by the National Death Index through 2011 to examine associations between MVM and MV use and CVD mortality by using Cox proportional hazards models, adjusting for multiple potential confounders.

Results: We observed no significant association between CVD mortality and users of MVMs or MVs compared with nonusers; however, when users were classified by the reported length of time products were used, a significant association was found with MVM use of >3 y compared with nonusers (HR: 0.65; 95% CI: 0.49, 0.85). This finding was largely driven by the significant association among women (HR: 0.56; 95% CI: 0.37, 0.85) but not men (HR: 0.79; 95% CI: 0.44, 1.42). No significant association was observed for MV products and CVD mortality in fully adjusted models.

Conclusions: In this nationally representative data set with detailed information on supplement use and CVD mortality data ∼20 y later, we found an association between MVM use of >3 y and reduced CVD mortality risk for women when models controlled for age, race, education, body mass index, alcohol, aspirin use, serum lipids, blood pressure, and blood glucose/glycated hemoglobin. Our results are consistent with the 1 available RCT in men, indicating no relation with MVM use and CVD mortality.

Introduction

Recently, the US Preventive Health Services Task Force issued a statement that evidence is insufficient whether to use vitamin and mineral supplements for the prevention of cardiovascular disease (CVD)¹⁰ (1); however, this recommendation is only based on 2 randomized clinical trials (RCTs). One of these, the Physicians’ Health Study II, did not find associations between use of a multivitamin-mineral (MVM) supplement (Centrum Silver) and CVD incidence or mortality after a median of 11 y among male US physicians aged ≥50 y at baseline in 1997 (2). The other study, the French Supplémentation en Vitamines et Minéraux Antioxydants, randomly assigned women aged 35–60 y and men aged 45–60 y to receive a daily supplement that contained moderate amounts of antioxidants, not an MVM. After 7.5 y on the study, CVD incidence was not statistically different between trial arms (3). Although RCTs represent the gold standard study design in research, the external validity of these RCTs is limited. The Physician’s Health Study II population was exclusively male (and primarily non-Hispanic white) and highly educated with extensive health knowledge.

More recently, a large, prospective cohort study of women in Sweden indicated that use of multivitamins (MVs; vitamin combinations without minerals) was associated with a reduced risk of myocardial infarction in women (HR: 0.73; 95% CI: 0.57, 0.93) (4). Among women who used MVs for >5 y, the inverse association was even stronger (HR: 0.59; 95% CI: 0.44, 0.80). Participants in this study had no history of CVD, and average length of follow-up was 10 y. Such findings show that the relation between vitamin and mineral supplement use and CVD is far from settled, and currently no RCTs exist to examine MVM use in women relative to CVD mortality.

Examining the potential relation between dietary supplements and CVD is certainly challenging. RCTs tend to be of relatively short duration, whereas CVD mortality has a long latency period. Observational studies often provide important information on long-term exposure to supplements in relation to morbidity and mortality, but they are often limited by their failure to assess important information on supplement use, which is typically queried with only a few questions on a FFQ, and there is a considerable degree of heterogeneity in supplement ingredients and amounts that is not captured. Observational studies are prone to confounding. Finally, different definitions of MVMs and MVs are used in both RCTs and observational studies, making standard comparisons across studies difficult. Often, the term multivitamins is used for both MVMs and MVs. Despite these challenges, continued investigation of these relations is important. CVD accounts for 1 in 4 deaths in the United States (5). Approximately one-half of US adults use dietary supplements, primarily to improve or maintain their health (6). An MVM product (which we define as having ≥3 vitamins and ≥1 minerals) is the most common choice, followed by calcium and MVs (6, 7).

The NHANES has monitored dietary supplement use in the United States since the 1960s (8). The NHANES data are unique because they provide nationally representative data and have remarkably detailed information about dietary supplements and their use, including the doses, durations, and frequency of their consumption. Such comprehensive information is available because participants are first interviewed in their homes and actually show the NHANES interviewers the supplement containers and labels. NHANES III (1988–1994) is unique because it oversampled older Americans and because mortality data from the National Death Index can be linked to this survey, providing a much longer follow-up period than is available in RCTs. Furthermore, the NHANES physical examination component provides detailed information about clinical, anthropometric, and biochemical measurements of health status. Given this rich source of data, the purpose of this research was to examine the association between MVM and MV use and CVD-specific mortality by linking NHANES III with the 2011 National Death Index with nearly 20 y of follow-up data.

Methods

NHANES III is a nationally representative, cross-sectional survey that uses a stratified, multistage probability design to obtain a nationally representative sample of the civilian, noninstitutionalized US population (8). NHANES III oversampled Mexican Americans, non-Hispanic black Americans, children <6 y old, and adults aged ≥60 y to increase the reliability and precision of health-related estimates in those population segments (9). NHANES participants completed an in-person household interview and received a follow-up health examination in a mobile examination center. The National Center for Health Statistics Institutional Review Board approved all NHANES procedures.

Our analysis was based on data from 10,196 adults aged ≥40 y who were both interviewed and examined and were eligible for mortality follow-up. We excluded 6 participants who were not eligible for mortality follow-up because not enough information was available to successfully match to the National Death Index. The unweighted response rate for the examination was 67%. We excluded pregnant and lactating women (n = 5) and participants with a history of CVD, including those with a self-reported history of myocardial infarction, stroke, or congestive heart failure at baseline (n = 1464) from the analytic sample in our study. We also excluded participants with chronic kidney disease as determined by glomerular filtration rate (GFR; n = 43) and missing information on dietary supplement use (n = 6). Our final study sample consisted of 8678 participants. Individuals with missing covariates were not included in multivariable analysis (Supplemental Tables 1–3).

NHANES measured participants’ demographic characteristics and health status and history, including dietary supplement use, during the personal interview. Demographic data collected included sex, age, race, and education level. The race/ethnic groups identified in NHANES included non-Hispanic white, non-Hispanic black, Mexican American, and other. Education level was categorized as completion of less than high school, high school completion, or education after high school.

NHANES participants showed containers of the dietary supplements, antacids, and prescription medications that contained nutrients to interviewers; the dietary supplement files contain these data. The interviewers asked about participant’s use of vitamins, minerals, herbs, and other supplements over the past 30 d and collected detailed information on type, consumption frequency, duration, and amount taken for each reported supplement. We defined MVMs as products that contained ≥3 vitamins and at least 1 mineral (7). We defined MVs as vitamin combinations without minerals (e.g., antioxidant vitamin combinations or vitamin B complex products). We assigned products to one of these mutually exclusive categories and classified individuals who reported MVM or MV use into 3 duration categories: <1 y, 1–3 y, and >3 y. We excluded individuals with missing data from the length-of-time analysis. We chose duration categories to ensure adequate sample sizes for the number of deaths in the MVM groups.

NHANES III obtained data on medication use and health history by questionnaire. These data included history of myocardial infarction, stroke, coronary heart disease, cancer (other than skin cancer), CVD, and diabetes (other than gestational). Alcohol consumption in NHANES was defined as the average number of drinks per day in the previous year; 1 drink contains 10 g of ethanol and is equivalent to 12 ounces of beer, 4 ounces of wine, or 1 ounce of distilled spirits. Measured height and weight were used to calculate BMI as weight in kilograms divided by height in meters squared (kg/m²). Self-reported cigarette smoking or serum cotinine concentrations were used to categorize smoking status as never, former, or current. Serum cotinine, a marker of tobacco exposure, was assayed via isotope dilution with liquid chromatography and tandem MS. Systolic blood pressure was calculated as the average of up to 6 measurements by using a mercury sphygmomanometer (up to 3 times during the household interview and up to 3 times during the NHANES mobile examination center examination) (10). Serum cholesterol was measured enzymatically by using commercially available reagents (Cholesterol/HP, catalog no. 816302, and Triglycerides/GPO, catalog no. 816370; both from Boehringer Mannheim) (11). HDL cholesterol was measured after the precipitation of other lipoproteins with a heparin-manganese chloride mixture (Hitachi 704 analyzer). Serum creatinine was measured by a Roche/Hitachi 737 analyzer (Roche Diagnostics) by using the kinetic alkaline picrate reaction and was calibrated to the Cleveland Clinic Research Laboratory (11). Estimated GFR was calculated for each individual on the basis of serum creatinine concentration, sex, age, and race. Participants with kidney disease (estimated GFR <60) were excluded from the study (12). Fasting blood glucose was measured with a modified hexokinase enzymatic method, and glycosylated hemoglobin was measured by ion exchange chromatography (11).

Assessment of vital status.

For our analysis, we used the NHANES III linked mortality restricted-use files. The matching methodology to determine accurate vital status through 31 December 2011 is documented by the National Center for Health Statistics, CDC, and is only summarized here (13). Vital status was assessed on the basis of a probabilistic match between personal identifiers from NHANES III and the death certificate records from the National Death Index (13). NHANES III participants aged ≥17 y at the time of the survey were eligible for mortality follow-up. We classified cause of death by using the International Classification of Diseases, 10th revision. That revision categorizes deaths into 113 recode groups; we examined major CVD (International Classification of Diseases codes 053–074) for this analysis.

Statistical analysis.

We analyzed data with SAS-callable SUDAAN; statistical significance was set at P < 0.05, adjusted for survey design and sampling weights. Users were compared with nonusers of dietary supplements on multiple characteristics and are presented as means ± SEs or percentages; differences in continuous variables were assessed with contrast statements in proc descript, and differences in categorical variables were assessed via Wald’s chi-square test. We used Cox proportional hazards regression models to estimate the HRs and 95% CIs for CVD-specific mortality by using age at baseline (i.e., the time of data collection) as the underlying time metric stratified by birth cohort (10-y age increments), as was recommended for use with NHANES data (14–17). Use of dietary supplements (MVMs or MVs) was the main exposure. We used sample weights to account for differential probabilities of selection, nonresponse, and noncoverage. Sample weights can be considered measures of the number of persons represented by the sample observation required to produce nationally representative estimates. We examined models that combined data for men and women and separate models for each sex. We verified that the proportional hazards assumptions were met for all models by examining 1) log-log plots, 2) plots of Schoenfeld residuals vs. log of age, and 3) inclusion of an interaction term of birth cohort and duration of product use (18).

For all models, we stratified the baseline hazard by birth cohort. Our first model, model 1, did not include potential confounders or effect modifiers; thus, it provided minimally adjusted estimates of the relation between dietary supplement use and mortality. Model 2 controlled for demographic variables, including sex, race/ethnicity, and education. Model 3 controlled for all demographic variables and included BMI, alcohol intake, and smoking status. We used cotinine concentrations of >10.0 μg/L or self-reported current smoking to classify smoking status. We modeled BMI and alcohol effects with linear and quadratic terms because of their U-shaped relations with mortality. Model 4 included medical/clinical confounders, such as hypertension (defined as mean systolic blood pressure ≥140 mm Hg or mean diastolic blood pressure ≥90 mm Hg from the arithmetic mean of all readings) or use of an antihypertension medication; high cholesterol, defined as total cholesterol >240 mg/dL or use of lipid-lowering medication; HDL cholesterol (as a linear term); diabetes mellitus, based on self-reported diagnosis, fasting blood glucose >140 mg/dL, or glycated hemoglobin ≥6.5%; and self-reported aspirin use in the previous 30 d (dichotomous). We also examined self-reported diet and physical activity by using directed acyclic graphs; we determined that diet and physical activity primarily exert their effects on CVD by modulating body weight, blood pressure, and serum lipids. Inclusion of these variables in the models did not alter results obtained. Thus, because we had already incorporated objectively measured variables in our models, and in light of the substantial measurement error associated with self-reported diet and physical activity, we chose not to include diet and physical activity in our further selection of confounding variables. Confounders were selected on the basis of a priori knowledge and were guided by the 10–15% change-in-estimate criterion (19, 20). None of the confounding variables that we examined changed the HR > 5% in univariate analyses. Many of the potential confounders were related to age, but because the models already included age as the time metric and were stratified by birth cohort, the addition of other potential confounders did not add substantively to the models. We also examined potential effect modification by including interaction terms between exposure and each potential confounder, and none was significant.

Results

Approximately 45% ± 0.5% of our sample had used any dietary supplement in the past 30 d. The most frequently used products were MVMs (21% ± 0.7%) and MVs (14% ± 0.5%). Dietary supplement users most frequently reported using 1 (63%) or 2 (20%) products, and most (>80%) reported taking the products daily (data not shown). After follow-up ceased on 31 December 2011 (median: 18.7 y, representing 133,442 person-years), the NHANES III adult sample had 4122 deaths (∼47% of the cohort) with a known cause of death. CVD was the most common cause of death (∼40%; n = 1636; Table 1).

TABLE 1.

Characteristics of NHANES III survey population providing mortality data¹

NHANES III characteristics	Value
Survey baseline years	1988–1994
Unweighted sample size,2 n	8678
Follow-up, y, median	18.7
Follow-up, person-years, n	133,442
Any dietary supplement use, %	45.3
Multivitamin-mineral use, %	21.2
Multivitamin use, %	14.2
Deaths,3 according to cause, n
All causes	4122
CVD	1636

¹CVD, cardiovascular disease; ICD, International Classification of Diseases.

²Excludes pregnant and lactating women, those <40 y of age, those missing information on dietary supplement use, those with a history of CVD or renal insufficiency at baseline, and those ineligible for mortality follow-up.

³These constituted deaths that occurred up to 31 December 2011 in those included in our analytic sample. Cause of death was classified with the ICD, 10th revision. Deaths are categorized by 113 recode groups; we examined major CVD (ICD 053–074) for this analysis.

As a representative sample of the US population during the survey years, the weighted sample was represented by higher proportions of women (54% ± 0.6) than men, non-Hispanic whites (81% ± 1%) than individuals of other races, individuals with a high school diploma or less educational attainment (61% ± 1%) than individuals with at least some college education, and current smokers (29% ± 1%) than former or never smokers (Supplemental Table 1). The analyzed sample (i.e., after exclusions) was representative of all ≥40 y at the population of interest (Supplemental Table 1); exact age range release of upper age range (i.e., adults aged ≥85 y) is prohibited by the National Center for Health Statistics. The unadjusted descriptive data for MVM users and nonusers are presented by all examined covariates (Table 2). Users did not differ from nonusers for age; BMI; alcohol consumption; aspirin use; or prevalence of high cholesterol, hypertension, or diabetes (Table 2). More MVM users than nonusers were women and non-Hispanic whites. Compared with nonusers, users had higher educational attainment, lower prevalence of current smoking, and higher HDL cholesterol concentrations (Table 2). Eligibility status for follow-up of NHANES III participants is shown in Supplemental Table 2.

TABLE 2.

Baseline characteristics of the NHANES III study participants by MVM supplement use¹

Characteristics	Nonusers (n = 7052)	Users (n = 1547)	P
Age, y	56.8 ± 0.32	57.6 ± 0.8	0.23
BMI, kg/m2	27.3 ± 0.1	26.9 ± 0.2	0.09
Alcoholic drinks,2 n/d	0.5 ± 0.01	0.4 ± 0.01	0.31
Serum HDL cholesterol, mg/dL	50.9 ± 0.4	52.7 ± 0.7	0.01
Female, %	52.9 ± 0.8	60.6 ± 1.6	<0.001
Race,3 %
Non-Hispanic white	78.9 ± 1.2	86.7 ± 1.3	<0.001
Non-Hispanic black	10.3 ± 0.6	5.8 ± 0.6
Mexican American	3.7 ± 0.3	2.6 ± 0.2
Educational attainment, %
Less than high school	30.2 ± 1.4	20.1 ± 1.6	<0.001
High school	32.5 ± 0.8	33.3 ± 2.2
More than high school	37.3 ± 1.2	46.6 ± 2.8
Smoking status, %
Never	40.4 ± 0.9	43.6 ± 2.0	<0.001
Former	28.6 ± 0.8	35.7 ± 2.0
Current	31.0 ± 1.0	20.7 ± 1.9
Hyperlipidemia, %	29.6 ± 0.9	33.1 ± 1.9	0.10
Hypertension, %	36.0 ± 0.9	36.2 ± 2.1	0.90
Diabetes mellitus, %	10.8 ± 0.6	8.9 ± 1.1	0.08
Aspirin use, %	39.3 ± 1.1	40.8 ± 1.9	0.48

¹Values are means or percentages ± SEs. Excludes pregnant and lactating women, those <40 y of age, those missing information on dietary supplement use, those with a history of cardiovascular disease or renal insufficiency at baseline, and those ineligible for mortality follow-up. These constituted deaths that occurred up to 31 December 2011. Differences in categorical variables were assessed with Wald’s chi-square test. Significance was set at P < 0.05 adjusted for survey design and sampling weights. MVM, multivitamin-mineral.

²One alcoholic drink contains 10 g of ethanol and is equivalent to 12 ounces of beer, 4 ounces of wine, or 1 ounce of distilled spirits.

³Percentages do not total 100% because the other race category is not provided.

Neither MVM (Table 3) nor MV (Table 4) use was associated with a lower risk of CVD mortality when we compared users with nonusers in any of the models, except for MVM use in model 1. However, when we considered the length of time participants used a product, we observed a significant inverse association for MVM use of >3 y, with an >35% reduced risk of CVD mortality across every model in women but not in men (Table 3). MV use was not significantly associated with CVD mortality when we combined sexes (Table 4). We found MV use associations for men who used products for 1–3 y across all models. However, because of the smaller sample sizes and number of deaths (n = 5) in this group, this result should be interpreted cautiously (Table 4).

TABLE 3.

Risk of cardiovascular mortality in US adults by MVM use and length of time used in NHANES III with follow-up in 2011¹

		Time MVMs used
	Overall	Nonuser	<1 y	1–3 y	>3 y
Study sample
Events/sample, n/N	1627/8599	1350/7052	69/405	67/367	123/677
Model 12	0.79 (0.64, 0.98)3*	1 (referent)	1.03 (0.72, 1.48)	0.97 (0.66, 1.42)	0.63 (0.48, 0.83)*
Model 24	0.84 (0.68, 1.04)	1 (referent)	1.08 (0.76, 1.53)	1.01 (0.69, 1.48)	0.68 (0.51, 0.89)*
Model 35	0.88 (0.69, 1.12)	1 (referent)	1.15 (0.77, 1.72)	1.09 (0.72, 1.64)	0.70 (0.52, 0.94)*
Model 46	0.86 (0.68, 1.10)	1 (referent)	1.22 (0.82, 1.80)	1.12 (0.72, 1.74)	0.65 (0.49, 0.85)*
Male participants
Events/sample, n/N	763/3990	649/3384	28/150	26/150	53/276
Model 1	0.84 (0.59, 1.20)	1 [(referent)	0.93 (0.54, 1.60)	0.94 (0.47, 1.86)	0.74 (0.46, 1.20)
Model 2	0.87 (0.60, 1.26)	1 (referent)	0.95 (0.59, 1.62)	0.98 (0.49, 1.96)	0.77 (0.47, 1.26)
Model 3	0.89 (0.59, 1.37)	1 (referent)	0.95 (0.49, 1.85)	1.15 (0.56, 2.33)	0.74 (0.42, 1.30)
Model 4	0.94 (0.61, 1.45)	1 (referent)	1.09 (0.52, 2.29)	1.15 (0.53, 2.49)	0.79 (0.44, 1.42)
Female participants
Events/sample, n/N	864/4609	701/3668	41/255	41/217	70/401
Model 1	0.79 (0.60, 1.03)	1 (referent)	1.15 (0.75, 1.77)	1.01 (0.65, 1.57)	0.59 (0.41, 0.84)*
Model 2	0.81 (0.62, 1.07)	1 (referent)	1.16 (0.76, 1.78)	1.03 (0.66, 1.60)	0.61 (0.43, 0.87)*
Model 3	0.86 (0.64, 1.16)	1 (referent)	1.30 (0.81, 2.07)	1.02 (0.63, 1.66)	0.65 (0.45, 0.96)*
Model 4	0.81 (0.59, 1.12)	1 (referent)	1.34 (0.85, 2.13)	1.06 (0.64, 1.75)	0.56 (0.37, 0.85)*

¹The overall HR compares MVM users with nonusers (referent group). The by-time analysis also uses nonusers as the referent group; persons with missing data on length of time MVM products were used were excluded only from the length of time analysis. *Significant differences, P < 0.05. BP, blood pressure, CVD, cardiovascular disease; MVM, multivitamin-mineral.

²Unadjusted model was stratified by baseline birth cohorts and excluded prevalent CVD at baseline.

³HR (95% CI) (all such values).

⁴Model 1 plus adjustment for sex (except sex models), race/ethnicity (non-Hispanic white, non-Hispanic black, Mexican American, or other), educational attainment (less than high school, high school, or more than high school).

⁵Model 2 plus adjustment for alcohol use (daily drinks as a linear term in addition to a quadratic term), smoking (self-reported current smoking or serum cotinine >10 μg/L), and BMI (BMI as a linear term in addition to a quadratic term).

⁶Model 3 plus adjustment for hyperlipidemia (total cholesterol ≥240 mg/dL or use of lipid-lowering medication), HDL cholesterol (linear term), hypertension (systolic BP ≥140 mm Hg or diastolic BP ≥90 mm Hg or hypertension medication use), diabetes mellitus (fasting blood glucose >140 mg/dL or glycated hemoglobin ≥6.5% or self-reported history), and aspirin use in the previous 30 d.

TABLE 4.

Risk of cardiovascular mortality in US adults by MV use and length of time used in NHANES III with follow-up in 2011¹

		Time MVs used
	Overall	Nonusers	<1 y	1–3 y	>3 y
Study sample
Events/sample size, n/N	1623/8586	1469/7598	37/272	20/178	83/481
Model 12	0.86 (0.69, 1.06)3	1 (referent)	1.14 (0.68, 1.91)	0.60 (0.33, 1.08)	0.77 (0.58, 1.02)
Model 24	0.91 (0.73, 1.13)	1 (referent)	1.22 (0.73, 2.05)	0.65 (0.36, 1.18)	0.81 (0.61, 1.06)
Model 35	0.93 (0.72, 1.20)	1 (referent)	1.27 (0.71, 2.28)	0.68 (0.36, 1.28)	0.79 (0.58, 1.09)
Model 46	0.96 (0.73, 1.27)	1 (referent)	1.37 (0.75, 2.49)	0.75 (0.39, 1.45)	0.81 (0.58, 1.13)
Male participants
Events/sample size, n/N	761/3988	696/3571	17/106	5/757	39/208
Model 1	0.80 (0.53, 1.22)	1 (referent)	1.12 (0.45, 2.78)	0.12 (0.04, 0.36)*	0.91 (0.56, 1.46)
Model 2	0.86 (0.57, 1.30)	1 (referent)	1.21 (0.48, 3.07)	0.14 (0.05, 0.40)*	0.95 (0.59, 1.53)
Model 3	0.89 (0.55, 1.44)	1 (referent)	1.18 (0.41, 3.38)	0.17 (0.06, 0.50)*	0.98 (0.60, 1.62)
Model 4	0.96 (0.59, 1.56)	1 (referent)	1.29 (0.46, 3.60)	0.19 (0.06, 0.57)*	1.05 (0.63, 1.75)
Female participants
Events/sample size, n/N	862/4598	773/4026	20/160	15/103	44/273
Model 1	0.91 (0.68, 1.22)	1 (referent)	1.17 (0.57, 2.38)	0.98 (0.52, 1.84)	0.70 (0.48, 1.00)
Model 2	0.94 (0.70, 1.26)	1 (referent)	1.20 (0.60, 2.41)	1.03 (0.54, 1.95)	0.71 (0.50, 1.01)
Model 3	0.95 (0.68, 1.31)	1 (referent)	1.31 (0.64, 2.70)	1.02 (0.51, 2.04)	0.66 (0.44, 0.97)*
Model 4	0.96 (0.67, 1.36)	1 (referent)	1.37 (0.64, 2.91)	1.10 (0.56, 2.18)	0.66 (0.43, 1.01)

¹The overall HR compares MV users with nonusers (referent group). The by-time analysis also uses nonusers as the referent group; persons with missing data on length of time MV products were used were excluded only from the length-of-time analysis. *Significant differences, P < 0.05. BP, blood pressure; MV, multivitamin.

²Unadjusted model was stratified by baseline birth cohorts and excluded prevalent cardiovascular disease at baseline.

³HR (95% CI) (all such values).

⁴Model 1 plus adjustment for sex (except sex models), race/ethnicity (non-Hispanic white, non-Hispanic black, Mexican American, or other), educational attainment (less than high school, high school, or more than high school).

⁵Model 2 plus adjustment for alcohol use (daily drinks as a linear term in addition to a quadratic term), smoking (self-reported current smoking or serum cotinine >10 μg/L), and BMI (BMI as a linear term in addition to a quadratic term).

⁶Model 3 plus adjustment for hyperlipidemia (total cholesterol ≥240 mg/dL or use of lipid-lowering medication), HDL cholesterol (linear term), hypertension (systolic BP ≥140 mm Hg or diastolic BP ≥90 mm Hg or hypertension medication use), diabetes mellitus (fasting blood glucose >140 mg/dL or glycated hemoglobin ≥6.5% or self-reported history), and aspirin use in the previous 30 d.

⁷Small sample size, cautious interpretation of finding is needed.

Discussion

In this nationally representative, prospective sample of adults who were without prevalent CVD, use of MVMs for ≥3 years was associated with reduced risk of CVD mortality at a median of 18 y of follow-up. Six critical points must be considered when evaluating these findings. First, and most importantly, we only had access to 1 cross-sectional measure of dietary supplement use at baseline, which makes it difficult to know whether this measure represented usual intakes over the entire follow-up period. However, like most health behaviors (e.g., exercise patterns, dietary practices), dietary supplement use is generally considered a habitual lifestyle practice (21, 22). Approximately one-half of participants who reported MVM use at baseline reported using these products for longer than 3 y (47%; 95% CI: 35%, 39%; data not shown).

Second, NHANES III was conducted before the passage of the Dietary Supplement Heath and Education Act, after which supplement use became more common in the United States (23). In other words, supplement users in NHANES III may represent early adopters of a perceived healthy behavior and may have other characteristics and behaviors that cluster together and might lead to better health. Although we made every effort to control potential confounders to address this issue, residual confounding may still have occurred.

To deal with confounding, researchers can adjust for variables in models, stratify models by potential confounding variables, or standardize the data. Although sex was not a confounder in the traditional sense, men and women have different risk of CVD for biological reasons. Indeed, different patterns emerged when we stratified by sex. Furthermore, by using age as the time metric and stratifying by birth cohort, we effectively obtained age-standardized results. Stratifying by birth cohort generated an average estimate for each cohort that we then averaged with the estimates for the other cohorts to limit the effect of age on supplement use and mortality. This is critical because of the strong ties of age with both variables.

Third, supplement use is tied to a host of healthy behaviors, and it is often difficult in epidemiologic research to disentangle the effects of healthy lifestyle choices from the use of dietary supplements (24). Thus, supplement use may represent a mixing of effects of healthy behaviors. For this reason, we included a fully adjusted model (model 4), even though none of these variables alone changed the estimate by >5% in univariate analyses. By using fully adjusted models, we hoped to reduce confounding of the supplement-mortality relation. We also performed several tests for interactions and evaluated the proportional hazards in multiple ways to ensure the accuracy of our estimates. When we conducted an extensive analysis of the 2006 follow-up data before the 2011 data were released, we obtained similar results. When we ran models that included participants with CVD at baseline, we also found similar results. Thus, we are confident that we have evaluated potential bias to the extent possible with this type of data.

Fourth, missing data are an issue with any study design. We explored models that used 2 strategies for dealing with missing data. First, individuals with missing data were dummy coded to facilitate eliminating observations with incomplete data; although this is a common practice in epidemiology, it is believed that this artificially reduces the CIs obtained for point estimates by artificially increasing the sample size and is likely to increase the chance of a making a type I error (25). In addition, we performed an analysis that was based on individuals with all data truly measured or recorded and ultimately chose this strategy for the results because it is the more conservative approach. This means that we cannot directly compare estimates of model 1 with results of model 4 because fewer people are represented in the latter. Nevertheless, even this more conservative approach yielded significant associations between MVM use and CVD mortality in fully adjusted models, and participants did not differ across all models on covariates (Supplemental Table 3).

Fifth, we recognize that the RCT is a gold standard for examining the relations addressed in this study. However, RCTs often lack sufficient follow-up time and/or have a homogenous sample that offers limited generalizability. This investigation is strengthened by the heterogeneous nature of the sample that represented many ages and race/ethnicities and by the long follow-up time. Because NHANES III oversampled individuals older than 65 y, our sample size for this age group was sufficient to examine CVD-specific mortality in multiple race/ethnic groups. Our results are consistent with the 1 available RCT in men.

Finally, it is difficult to compare the findings in this study with findings from other large cohorts in the United States or Europe because of differences in the measurements of exposures and differences in the definition of the products being investigated. Most research that relates dietary supplement use to CVD has involved either trials to lower homocysteine (in general high-dose vitamin B) or trials that examined a potential effect of antioxidant nutrients (vitamins C and E and β-carotene). Neither of these trial types has shown a significant association with cardiovascular-related disease or mortality when meta-analysis of RCTs was performed (26–28). However, none of this research involved MVMs and instead involved only MVs. In our NHANES investigation, subjects took a variety of MVM and MV products. They all differ in the combinations and amounts of nutrients they contain. The main difference between MVMs and MVs is the presence of minerals. A typical MVM has minerals, including calcium (generally ∼150 mg only), magnesium, zinc, phosphorus, manganese, and copper, and, depending on the formulation type, may or may not contain iron, selenium, iodine, chromium, or molybdenum. It is not possible to separately estimate the effect of any individual nutrient, nutrient combinations, or specific amounts of nutrients in this analysis. However, we can speculate that minerals or certain vitamins not always found in MV formulations (e.g., vitamin D) may help explain the differential effect of MVM products on CVD mortality. For example, meta-analyses have found magnesium intake to be inversely associated with risk of strokes (29) and reduced risk of ischemic heart disease and its mortality (30). Another meta-analysis demonstrated a nonsignificant protective association of vitamin D with CVD mortality (31). We hope this work serves to highlight the importance of defining the type and constituents of dietary supplements used in future research.

In conclusion, MVMs and MVs appear to have different associations with CVD mortality risk, at least when these products are taken for >3 y. Our findings suggest longer term use of MVMs and CVD mortality may have a protective relation in US women who lack a history of CVD. We identified several caveats for the interpretation of the results of this study. Thus, it is important to replicate this investigation in other studies. In future research, we suggest that care is taken in reporting on and analyzing different dietary supplement types and differentiating between MVMs and MVs. Ideally, such analyses should be conducted on multiple occasions in cohort studies and RCTs, last longer, and include hard endpoints, such as mortality and morbidity.