Objective  To assess the effect of antioxidant supplements on mortality in randomized primary and secondary prevention trials.

Data Sources and Trial Selection  We searched electronic databases and bibliographies published by October 2005. All randomized trials involving adults comparing beta carotene, vitamin A, vitamin C (ascorbic acid), vitamin E, and selenium either singly or combined vs placebo or vs no intervention were included in our analysis. Randomization, blinding, and follow-up were considered markers of bias in the included trials. The effect of antioxidant supplements on all-cause mortality was analyzed with random-effects meta-analyses and reported as relative risk (RR) with 95% confidence intervals (CIs). Meta-regression was used to assess the effect of covariates across the trials.

Data Extraction  We included 68 randomized trials with 232 606 participants (385 publications).

Data Synthesis  When all low- and high-bias risk trials of antioxidant supplements were pooled together there was no significant effect on mortality (RR, 1.02; 95% CI, 0.98-1.06). Multivariate meta-regression analyses showed that low-bias risk trials (RR, 1.16; 95% CI, 1.05-1.29) and selenium (RR, 0.998; 95% CI, 0.997-0.9995) were significantly associated with mortality. In 47 low-bias trials with 180 938 participants, the antioxidant supplements significantly increased mortality (RR, 1.05; 95% CI, 1.02-1.08). In low-bias risk trials, after exclusion of selenium trials, beta carotene (RR, 1.07; 95% CI, 1.02-1.11), vitamin A (RR, 1.16; 95% CI, 1.10-1.24), and vitamin E (RR, 1.04; 95% CI, 1.01-1.07), singly or combined, significantly increased mortality. Vitamin C and selenium had no significant effect on mortality.

Conclusions  Treatment with beta carotene, vitamin A, and vitamin E may increase mortality. The potential roles of vitamin C and selenium on mortality need further study.

We found that antioxidant supplements, with the potential exception of selenium, were without significant effects on gastrointestinal cancers and increased all-cause mortality.^14-15 We did not examine the effect of antioxidant supplements on all-cause mortality in all randomized prevention trials.¹⁶ Our aim with the present systematic review was to analyze the effects of antioxidant supplements (beta carotene, vitamins A and E, vitamin C [ascorbic acid], and selenium) on all-cause mortality of adults included in primary and secondary prevention trials.

METHODS

We included antioxidant supplements at any dose, duration, and route of administration. We analyzed the antioxidants administered singly, in combination with other antioxidants, or with other vitamins or trace elements. Trials with collateral interventions were included if the interventions were used equally in the trial groups. Subgroup analyses without high-bias risk trials and selenium trials were preconceived. Our outcome measure was all-cause mortality at maximum follow-up.

Data Sources

We searched The Cochrane Central Register of Controlled Trials (CENTRAL) in The Cochrane Library (Issue 3, 2005), MEDLINE (1966 to October 2005), EMBASE (1985 to October 2005), and the Science Citation Index Expanded (1945 to October 2005).²⁰ We scanned bibliographies of relevant articles for additional trials.

Data Extraction

Two of the 3 authors (G.B. and D.N., and R.G.S.) independently assessed trial eligibility. Excluded trials were listed with the reasons for exclusion. Disagreement was resolved by discussion or in consultation with a third author (C.G.). We contacted authors of the trials for missing information.

From each trial we recorded first author; country of origin, country income category (low, middle, high)²¹; number of participants; characteristics of participants: age range (mean or median) and sex ratio; participation rate; dropout rate; trial design (parallel, factorial, or crossover); type of antioxidant; dose; duration of supplementation; duration of follow-up (ie, treatment duration plus posttreatment follow-up); and cointerventions. We extracted the date, location, sponsor of the trial, and the publication status.

Due to the risk of overestimating intervention effects, analyses were stratified according to the risk of bias (methodological quality).^{14-15,18-19,22-24} Trials with adequate generation of the allocation sequence, adequate allocation concealment, adequate blinding, and adequate follow-up were considered low-bias risk trials (high methodological quality).²⁴ Trials with one or more unclear or inadequate quality components were classified as high-bias risk trials (low methodological quality).²⁴ Generation of the allocation sequence was considered adequate if the allocation sequence was generated by a computer or random-number table, or similar; allocation concealment was considered adequate if concealed up to the point of treatment by central randomization, sealed envelopes, or similar; blinding was considered adequate if the trial was described as double-blind and using identical placebo; follow-up was considered adequate if the numbers and reasons for dropouts and withdrawals in all intervention groups were described or if it was specified that there were no dropouts or withdrawals. Bias risk was assessed without blinding of 2 authors (G.B. and D.N. or R.G.S.). Consensus was reached through discussion or arbitration by a third author (C.G. or L.L.G.) before data entry. We have found high interrater agreement between blinded and unblinded assessments and also between 2 independent assessors.²⁴

Statistical Analyses

We used The Cochrane Collaboration software (RevMan Analyses 1.0; www.cochrane.org), STATA 8.2 (STATA Corp, College Station, Tex), Sigma Stat 3.0 (SPSS Inc, Chicago, Ill), and StatsDirect (StatsDirect Ltd, Altrincham, England). We analyzed the data with a random-effects model,²⁵ calculating the relative risk (RR) with 95% confidence intervals (CIs). To account for 0 cells in the 2 x 2 tables, we calculated the RR with 3 different continuity corrections (0.5; 0.1; 0.01).^26-27 We did not include trials with 0 events in both intervention groups.^27-28 Because the number of such trials was large, we performed exploratory analysis adding an imagined trial with 1 death and 20 000 participants in each group.

We used the STATA metareg command for the random-effects metaregression to assess which covariates influenced the intervention effect across trials.²⁹ The included covariates were bias risk, type and dose of supplement, single or combined supplement regimen, duration of supplementation, and primary or secondary prevention. Univariate and multivariate analyses including all covariates were performed. Results are presented with regression coefficients and 95% CI.

All analyses followed the intention-to-treat principle. For trials with factorial design, we based our results on at-margins analysis,³⁰ comparing all groups that received antioxidant supplements with groups that did not. To determine the effect of a single antioxidant, we performed inside-the-table analysis³⁰ in which we compared the group taking a single antioxidant with the group taking placebo or receiving no intervention. In trials with more than 2 groups assessing additional therapy, we compared only groups receiving antioxidants, placebo, or no intervention.

We assessed heterogeneity with I² that describes the percentage of total variation across trials due to heterogeneity rather than chance.^{17, 31} I² can be calculated as I² = 100% x (Qv –df)/Q, where Q is Cochran's heterogeneity statistics and df the degrees of freedom. Negative values of I² are put equal to 0, so I² lies between 0% (no heterogeneity) and 100% (maximal heterogeneity).³¹ We compared the estimated treatment effects in trials with a low- or high-risk of bias with test of interaction.³² We performed adjusted-rank correlation³³ and regression-asymmetry tests³⁴ for detection of bias.

RESULTS

View larger version (101K): [in this window] [in a new window] [as a PowerPoint slide]   Figure 1. Flow Diagram of Identification of Randomized Trials for Inclusion

A total of 232 606 participants were randomly assigned in the 68 trials. The number of participants in each trial ranged from 24 to 39 876 (Table 1 and Table 2). The mean age was 62 years (range, 18-103 years). The mean proportion of women was 44.5% in the 63 trials reporting sex.

View this table: [in this window] [in a new window]