ARTICLE

Clinical Trials 2009; 6: 2841

Bayesian model averaging in meta-analysis: vitamin E supplementation and mortality

Donald Berrya, J Kyle Wathena and Margaret Newellb
Context The strength and relevance of a meta-analysis depends on the validity of the statistical methods used. Of special importance is appropriately assessing different sources of variability. Many studies including meta-analyses have evaluated the efficacy and safety of vitamin E and have yielded varying results. Illuminating and resolving these disparities requires addressing study variability and model uncertainty. Objective To describe Bayesian meta-analysis methods for combining data from clinical trials, using recent studies that analyzed the relationship between vitamin E dose and all-cause mortality. Data Sources Studies used in a previously published meta-analysis appended by studies identified by a search of MEDLINE from August 2004 to December 2005 using the MeSH terms vitamin e and alpha tocopherol. Study Selection Inclusion criteria: men and nonpregnant women; use of vitamin E alone or in combination with other vitamins or minerals; random allocation of participants to either vitamin E or a placebo or other control group; intervention and follow-up duration greater than 1 year; 10 or more deaths. Data Extraction Independent data extraction by one author was reviewed and confirmed by a second author. Corresponding authors of the original publications were contacted when questions arose. Data Synthesis Data collection included the number of patients and deaths, percent men, use of other vitamins or minerals, mean age, and length of follow-up. We combined study results using Bayesian hierarchical model averaging. Analyses used Markov chain Monte Carlo computational techniques. Conclusions Vitamin E intake is unlikely to affect mortality regardless of dose. The Bayesian meta-analyses presented here are ideal for incorporating disparate sources of variability, including trial effect and model uncertainty. Clinical Trials 2009; 6: 2841. http://ctj.sagepub.com

Introduction
Combining the results of multiple clinical trials is not easy. Synthesis is the goal, but it is not appropriate to simply combine results as though they were from a single large study. And it is difficult or impossible to interpret traditional confidence intervals or statistical significance levels presented for each study. The same patient may respond differently in different trials even with the
a

same intervention. Therefore, when combining information from multiple clinical trials, it is essential to consider trial effect. Bayesian methods are ideal for synthesizing the results of multiple experiments. We demonstrate Bayesian meta-analytic methods as alternatives to the usual frequentist approach, applying our methods to synthesize results from previous clinical trials of vitamin E. In particular, we develop a Bayesian hierarchical model to combine information from

Department of Biostatistics, University of Texas, MD Anderson Cancer Center, bDivision of Internal Medicine, MD Anderson Cancer Center Author for correspondence: Donald Berry, Department of Biostatistics, University of Texas, MD Anderson Cancer Center, Houston, TX, USA. Tel: 713-794-4142; Fax: 713-563-4243. E-mail: dberry@mdanderson.org Funding: This research was funded by the Council for Resposible Nutrition, which had no review rights or input into the papers methods or conclusions.

Society for Clinical Trials 2009 SAGE Publications

10.1177/1740774508101279

Vitamin E supplementation and mortality

several trials and account for uncertainty in the doseresponse relationship of vitamin E by considering various candidate models. The Bayesian view of inference differs from the more traditional frequentist view in several important ways [1,2]. We refer the reader to previous publications for elementary and detailed expositions, respectively [3,4]. In the United States, the most commonly used dietary supplements are multivitamins and minerals [5]. Many recent studies have examined the efficacy and safety of these supplements [610], while others have specifically examined how the supplements affect mortality [1113]. Each study contributes to our knowledge and understanding of the effects of such supplements. The results often provide statistical p-values and confidence intervals, which can be difficult to combine appropriately while accounting for trial effect. Combining information from multiple studies that may be conducted under different circumstances, in different populations, and using different doses is a difficult task that requires careful consideration of uncertainties. This difficulty is reflected in the conclusions drawn by three recent meta-analyses that examined the relationship between vitamin E dose and mortality: each reached different conclusions and gave different definitions of high dose [1416]. Concerns about properly combining information from multiple studies and the recent plethora of studies examining the effects of vitamin E and other vitamins on mortality motivated us to reassess the doseresponse relationship between vitamin E intake and all-cause mortality. The conclusions regarding this relationship have farreaching implications, including (possibly inappropriately) influencing participants to withdraw from trials that are addressing the impact of vitamin E intake on various clinical outcomes. Therefore, this relationship should be modeled in a robust fashion and presented in a way that clarifies the strength of the relationship and, to the extent possible, resolves residual uncertainties about the relationship and its strength. We take an approach that is ideally suited for both purposes: Bayesian hierarchical modeling [4,17]. This approach makes assumptions that are less restrictive than those of a conventional meta-analysis. We regard the collection of available studies as a sample from the larger population of studies that address the effects of vitamin E on mortality. Conclusions are possible for the population of studies. And because each study is a member of this population, posterior distributions are available for the effects within individual studies, borrowing strength from the other studies in the sample. As regards the doseresponse relationship, we assume several candidate hierarchical models and http://ctj.sagepub.com

29

allow each model to contribute to the overall conclusions via Bayesian model averaging [1824]. To demonstrate how our method differs from a traditional approach, we provide a step-by-step procedure describing our methodology, and compare it with the frequentist approach used in a previous analysis, that of Miller et al. [14]. First, we consider the studies analyzed by Miller et al. and then add data extracted from additional peerreviewed publications to update the data set. Using a Bayesian hierarchical model, we combine the data from the clinical trials that meet our inclusion criteria. To incorporate various dose response relationships and study characteristics, we develop a sequence of hierarchical models and employ Bayesian model averaging. We demonstrate Bayesian methods as alternatives to the usual frequentist approach for combining data from clinical trials, using data from recent studies to assess whether there is an effect or a doseeffect of vitamin E on all-cause mortality.

Methods
Data source and searches We obtained publications of the studies used in the frequentist meta-analysis by Miller et al. [14]. That earlier analysis was based on a search of MEDLINE using the MeSH terms vitamin E, antioxidant vitamins, alpha tocopherol, tocopherol, and clinical trials between 1966 and August 2004, complemented by a search of the Cochrane database of randomized, controlled trials; a review of the reference lists of original research, review articles, and previous meta-analyses; and a review of the authors files. To include trials appearing in the literature since that meta-analysis, we performed a search of MEDLINE using the MeSH terms vitamin E and alpha tocopherol between August 1, 2004 and December 20, 2005. Search limits were: English language, randomized controlled trial, humans, and at least 19 years old.

Study selection In selecting studies from MEDLINE, we used the inclusion criteria of Miller et al.: random allocation of patients; use of vitamin E supplement alone or with other vitamins or minerals; use of a control or placebo group; participants only of men and nonpregnant women; vitamin E supplementation and follow-up longer than 1 year; and 10 or more deaths in the trial. Clinical Trials 2009; 6: 2841

30

D Berry et al.
mortality data in press [26], and one that did not provide mortality data [27]. Of the four remaining studies, one was a study that Miller et al. had analyzed [28], and another was an update to a study that had also been analyzed by those authors [29]. Thus, we incorporated 18 studies from our comparison frequentist meta-analysis by Miller et al. [14]. We combined data from the 3 additional studies with data from the 18 studies analyzed by Miller et al., plus data from one study that met our inclusion criteria that was indirectly found through our MEDLINE search (one search result was a commentary on a study that did not appear in our search results); thus, we used data from 22 studies in this meta-analysis (extracted from 24 publications) [2851]. We illustrate the inclusion process in Figure 1.

Of the 154 studies that resulted from our 2 separate MEDLINE searches (vitamin E search: 111 studies; alpha tocopherol search: 43 studies), 30 studies appeared in both search results. We also included all studies considered by Miller et al. [14]. Therefore, we had 124 studies to review. Of these, 53 studies had a mean or median intervention and follow-up duration of <1 year; 33 had no intervention or one that was not vitamin E; 17 were substudies of populations previously reported by Miller et al.; 7 had fewer than 10 deaths; 5 were of pregnant women; and one had no control group. Of the 8 remaining studies, two were of the same population, yielding seven unique studies. At the time of this meta-analysis, mortality data was not available for three of these seven studies: one of an ongoing trial [25], one that had its

Publications resulting from separate searches of MEDLINE n = 154 Publications on both search lists, n = 30 Publications reviewed for our inclusion criteria n = 124

Publications meeting our inclusion criteria n=8

Mean/median intervention and follow-up less than one year, n = 53 Intervention not vitamin E or no intervention at all, n = 33 Substudy of previously reported population, n = 17 Less than ten deaths, n = 7 Study of pregnant women, n = 5 No vitamin E control group, n = 1

Studies on same population, n = 2

Unique publications meeting our inclusion criteria n=7 Mortality data not available at time of meta-analysis, n = 3

Unique publications meeting our inclusion criteria with mortality data available n=4 Study used in Miller et al. meta-analysis, n = 1 Unique publications not used in Miller et al. meeting our inclusion criteria with mortality data available n=3 Studies from Milleretal., n = 18 study found indirectly by MEDLINE search, n=1 Studies used in this meta-analysis n = 22

Figure 1

QUOROM chart of study inclusion for Bayesian meta-analysis

Clinical Trials 2009; 6: 2841

http://ctj.sagepub.com

Vitamin E supplementation and mortality

Data Extraction and Quality Assessment Data from the 22 studies were independently extracted by one author (MAN) and confirmed by a second author (JKW). We compared data published by Miller et al. [14] with the data we extracted from studies they had analyzed. This comparison revealed several discrepancies. We resolved the data discrepancies by contacting the corresponding authors of the original studies. Two discrepancies concerned the number of intent-to-treat participants in the trials [28,36]; all other discrepancies concerned the factors used in converting milligrams or international units (IU) of various alpha-tocopherol forms to IU of the acetate form of dl-alpha-tocopherol (allrac-alpha-tocopherol) [52]. We converted all forms of vitamin E to international units (IU) of dl-alpha-tocopheryl acetate (allrac-alpha-tocopheryl acetate), one of the esthers of dl-alpha-tocopherol, because it was the form most commonly used across the included studies. Because each form of vitamin E (alpha-tocopherol) has a distinct biological activity, each requires a different factor to convert it to IU of dl-alphatocopheryl acetate. Thus, confirming the form used in each study was crucial to accurately represent the doses in terms of dl-alpha-tocopheryl acetate. For studies that did not publish the form of vitamin E used, we contacted the corresponding authors by e-mail. In Table 1 we present characteristics of the data sets we analyzed, and in Table 2 we show a comparison between the conversion factors and doses of vitamin E used in our analysis and those used by Miller et al.

31

we examine four mean functions: (1) no vitamin E effect; (2) a quadratic-linear spline; (3) a linear dose effect; and (4) a quadratic dose effect. A Bayesian approach enables such statements as, the probability that vitamin E increases all-cause mortality is . . .. In addition, the Bayesian modeling averaging aspect of the analysis allows us to draw conclusions such as the probability that a model for a doseresponse relationship is correct (within a specified collection of models). Rather than base our conclusions on a single model as done in other meta-analyses [1416], we allow a variety of doseresponse relationships to contribute to our final conclusion by averaging the results based on the probability that each model is correct. In our example, we include only trials in which patients were randomized between vitamin E (E) and a control (C). Let pt,i denote the mortality rate on treatment t E or C for study i 1, 2, . . . , N, where N is the number of studies. Of interest is the comparison of pE,i and pC,i among the various studies, while accounting for the differences in patient populations and vitamin E dose. We model the comparison of E and C within each trial, accounting for dose. More specifically, for trial i let

di logitpE, i logitpC, i
where logit(x) log{x/(1 x)}. Because the di term in our models is not easily interpretable, we present the results using a measurement of relative risk. A relative risk of 1.00 corresponds to no E/C effect, the effect of vitamin E above and beyond that of the control. An aspect of interest is the population of trials and its distribution. In particular, the data provide information about whether this population has little variation (homogeneous) or great variation (heterogeneous) after accounting for dose. Making inferences about doseresponse is more challenging than performing standard meta-analyses. We consider the mortality rate to depend on the dose (Di) of E within trial i and on trial-specific covariates Zi. These covariates include whether E was used in combination with other vitamins or minerals, mean subject age, sex, and follow-up duration. We assume di has a normal distribution N(md(Di, Zi), s2). Parameter s2, the study population variance, determines the amount of borrowing strength across the studies. There is greater borrowing when s2 is smaller. If s2 is very large then the population of treatment effects varies considerably from one trial to another and so there is little borrowing. We let the data inform the extent of borrowing by assigning a prior distribution rather than a particular value to s2. For most of our calculations we assume a noninformative inverse gamma prior distribution for s2, which allows for the greatest heterogeneity among trials at the same Clinical Trials 2009; 6: 2841

Data synthesis and analysis We first use a Bayesian hierarchical modeling for data synthesis and Markov chain Monte Carlo computational techniques in our analyses. When considering multiple studies, the underlying death rate, p, of a patient in one study may differ from that of a patient in another study, even if the patients have the same clinical characteristics and receive the same treatment. Therefore, data from several studies cannot simply be combined as though from one large study. We use Bayesian hierarchical modeling to allow different values of the death rate, p, for different studies. To simplify the description of the differences between the results of a frequentist approach and our Bayesian models, we first use a Bayesian hierarchal model that is similar to the model used by Miller et al. [14] to analyze the data provided in that study. As there is always uncertainty in the chosen model of the doseresponse relationship, http://ctj.sagepub.com

32

Table 1
Intent-to-treat % Mean N Men age, in years Factorial design (factorial intervention) Mean Vitamin E follow-up, subjects in years (deaths) Control subjects (deaths) Placebo Double Vitamin E Vitamin E controlled blind combined dose, (other expressed vitamins as IU of or minerals) dl--tocopherol acetateb 15.0 Yes Yes

Characteristics of studies used in the Bayesian meta-analysis

Study (Ref No.)a (Country)

Year

Study population

D Berry et al.

Clinical Trials 2009; 6: 2841

725 26 84.0 2.0 361 (100) 364 (106) 12741 40 48.0 7.5 6364 (76) 6377 (98) 30.0 Yes Yes Yes (selenium, 100 mg/d; zinc, 20 mg/d) No 29584 45 Range, 5.3 4069 57.0 54.0 6.0 1657 (157) 1661 (167) 60.0 Yes 6.1 14564 (1800) 14569 (1770) 50.0 Yes 14792 (1018) 14792 (1109) 33.0 Yes Yes 29133 3318 44 100 Yes Yes Yes (vitamin C, 120 mg/d; betacarotene, 6 mg/d) Yes (vitamin C, 120 mg/d; betacarotene, 6 mg/d; selenium, 100 mg/d; zinc, 20 mg/d) Yes (beta-carotene, 15 mg/d; selenium, 50 mg/d) No One-half replicate of 24 factorial design Yes (beta-carotene, 20 mg/d) No 3411 51 47.0 3.3 1706 (38) 1705 (43) 200.0 Yes Yes Yes (multivitamin and mineral supplement) Yes (vitamin C, 500 mg/d; betacarotene, 15 mg/d; selenium, 75 mg/d) 300.0 No No No 11324 85 59.0 3.5 5660 (488) 5664 (529) Yes (garlic extract, 800 mg/d; garlic oil, 4 mg/d) Yes (fish oils, 1 g/d) 2264 (68) 2387 (240) 300.0 400.0 No Yes No Yes No 4495 4757 44 68.0 6.3 42 64.0 3.6 2231 (72) 2370 (251) 78 61.0 4.0 433 (15) 431 (29) 440.0 Yes Yes Yes (aspirin, 100 mg/d) Yes (vitamin Yes (zinc, C, 500 mg/d; beta20 mg/d; carotene, 15 mg/d) copper, 2 mg/d) Yes (vitamin C, Yes (beta-carotene, 1000 mg/d) 25 mg/d) 19937 (636) 7.0 4761 (799) 19939 (615) 4780 (801) 447.0 c 10.1 544.0 Yes Yes Yes Yes No No 0 74 66.0 54.6 Yes (aspirin, 50 mg/d) No

MIN.VIT.AOX [30] (France)

1999

Institutionalized elderly

SU.VI.MAX [28] 2004 (France)

General population

Linxian A [31] (China)

1993

1994

Commune residents in Linxian, China Male smokers

ATBC [32,33] (Finland) Linxian B [34] (China)

1993

Linqu [35] (China)

2001

Commune residents in Linxian, China with esophageal dysplasia General population

GISSI [36] (Italy)

1999

2001

PPP [37] (Italy) AREDS [38] (US)

2001

Patients surviving recent myocardial infarction Patients having at least cardiovascular event Well-nourished older adults

PPS [40] (US)

1994

WHS [39] (US)

2005

http://ctj.sagepub.com

HOPE-TOO [29] (19 countries)

2005

Patients with recently 864 diagnosed large-bowel adenomas Healthy women in the 39876 general population Patients at high risk for 9541 cardiovascular events

(continued)

Table 1
Intent-to-treat % Mean N Men age, in years Factorial design (factorial intervention) Mean Vitamin E follow-up, subjects in years (deaths) Control subjects (deaths) Placebo Double Vitamin E Vitamin E controlled blind combined dose, (other expressed vitamins as IU of or minerals) dl--tocopherol acetateb 600.0 Yes Yes

Continued

http://ctj.sagepub.com
297 41 66.0 Range, 24 149 (9) 148 (3) 75 Range, 5.0 4080 66.0 65.0 2.8 212 (16) 211 (6) 800.0 Yes 4.0 595 (20) 598 (11) 745.0 Yes Yes No 10269 (1446) 10267 (1389) 600.0 Yes Yes 1193 423 0 44 Yes (vitamin C, No 750 mg/d; beta-carotene, 18 mg/d) Yes (vitamin C, Yes 250 mg/d; beta(simvastatin, carotene, 20 mg/d) 40 mg/d) No No Yes (vitamin C, 1000 mg/d) Yes (hormone replacement therapy) 1035 84 62.0 1.4 1035 (68) 967 (52) 910.4d Yes Yes No No 196 69 65.0 1.4 97 (31) 99 (29) 1192.0 Yes Yes No No 769 54 72.9 3.0 257 (6) 512 (17) 2000.0 Yes Yes No Yes (donepezil, 10 mg/d) 171 (22) 2000.0 Yes Yes No Yes (selegiline, 10 mg/d) 399 (73) 401 (64) 2000.0 Yes Yes No Yes (deprenyl, 10 mg/d) 1.5 83 (31) 77 (28) 5000.0 Yes Yes No No 341 35 73.0 2.0 170 (19) 800 66 62.0 8.2 160 65 58.0

Study (Ref No)a Year (Country)

Study population

REACT [41] (US and UK)

2002

Patients with early age-related cataract

MRC/BHF [42] (UK)

2002

Patients at high risk for 20536 cardiovascular events

VECAT [43] 2004 (Australia) WAVE [44] 2002 (US and Canada)

Healthy older adults

CHAOS [45,46] (UK)

1996

SPACE [48] (Israel)

2000

2005

ADCS B [49] (US and Canada) ADCS A [50] (US)

1997

1998

Postmenopausal women with at least one 15% to 75% coronary stenosis at baseline coronary angiography Patients with angiographically proven coronary atherosclerosis Dialysis patients with pre-existing cardiovascular disease Patients with amnesic mild cognitive impairment Patients with moderate severity Alzheimers disease Patients with early Parkinsons disease

Vitamin E supplementation and mortality

DATATOP [51] (US and Canada) ALS [47] (Germany)

2005

Patients with probable or definite amytrophic lateral sclerosis

33

Clinical Trials 2009; 6: 2841

a Ref No. is the reference number for the publication; bSee Table 2 for conversion factors used in this study; c600 IU of natural source vitamin E administered every other day, converted to 300 IU of natural source vitamin E per day for purposes of analysis. dThe first 546 of 1035 patients received 800 IU of natural source vitamin E per day; the remaining 489 patients received 400 IU of natural source vitamin E per day. The dose indicated here is a weighted average.

34

D Berry et al.
Table 2 Comparison of conversion factors and vitamin E doses in the two meta-analyses Year Vitamin E composition dl-alpha-tc dl-alpha-tc
d

Randomized study

Conversion factor study Fa 1.1 1.1 1.1 1.0 1.0 1.0 1.1 1.1 1.0 1.1 1.1 1.1 1.0 1.0 1.0 1.0 1.0 1.0

Vitamin E dose (IU) study Fa 16.5 33.0 33.0 50.0 60.0 200.0 330.0 330.0 400.0 440.0 660.0 660.0 500.0 800.0 600.0 800.0 2000.0 2000.0

Conversion factor study Bb 1.0 1.0 1.1 1.0 1.0 1.0 1.0 1.0 1.0 1.1 1.0 1.0 1.49 1.0 1.49 1.49 1.0 1.0

Vitamin E dose (IU) study Bb 15.0 30.0 33.0 50.0 60.0 200.0 300.0 300.0 400.0 440.0 600.0 600.0 745.0 800.0 910.4 1192.0 2000.0 2000.0

MIN.VIT.AOX SU.VI.MAX Linxian A ATBC Linxian B Linqu GISSI PPP AREDS PPS REACT MRC/BHF VECAT WAVE CHAOS SPACE ADCS A DATATOP
a

1999 2004 1993 1994 1993 2001 1999 2001 2001 1994 2002 2002 2004 2002 1996 2000 1997 1998

dl-alpha-tc dl-alpha-tc
d

dl-alpha-tc dl-alpha-tc dl-alpha-tc

d

all-rac-alpha-te dl-alpha-tc Natural source

d

Natural source Natural source

d d

Study F is the frequentist meta-analysis by Miller et al. [14]; bStudy B is the current Bayesian meta-analysis;cdlalpha-tocopheryl acetate; dUnable to reach corresponding authors regarding composition of vitamin E, or corresponding authors did not recall exact composition; eall-rac-alpha-tocopheryl acetate.

vitamin E dose [53]. Because the distribution of s2 plays a critical role in our model, we perform sensitivity analyses with respect to it. We use six different prior distributions and base each on the amount of heterogeneity assumed between trials at the same E dose level. To convey differences between our Bayesian approach and a frequentist approach, we present analyses of the data obtained by Miller et al. [14], the corrected data, and the data updated since that study. To assess model fit we show residual plots.

Results
Effect of vitamin E dose on all-cause mortality Using the updated data, and adjusting for subject age, sex, follow-up duration, and vitamin E used in combination with other vitamins or minerals, we found the total posterior probability for all covariate-adjusted models to be <0.025. This indicates that there is little evidence that these covariates matter as regards any effect of vitamin E. For ease of presentating our model averaging approach we omit these covariates from the remainder of the analyses, instead giving details of the covariate analysis in the Technical Appendix. In Table 3, we provide the results for the posterior probabilities of no effect of E, Clinical Trials 2009; 6: 2841

E beneficial and E harmful under three mean structures for vitamin E effect including: (1) spline, (2) linear, and (3) quadratic as well as the model averaging. The spline model is similar to the approach taken by Miller et al. [14] and gives similar results. In particular, at a dose of 400 IU, the probability that vitamin E is harmful is 0.83, which is more than 4 times the probability that vitamin E is beneficial. However, Figure 2 shows that using this model and data published by Miller et al., the 95% posterior probability intervals for relative risk contain 1.00 for all doses, indicating that no E effect is plausible, even in this model. In addition, correcting and updating the data had very little influence on the results, as the conclusions for the corrected (C) and updated (U) data using this model were consistent at doses of 400 IU and greater. We also analyzed all three data sets (that of Miller et al., the corrected data, and the updated data) using linear and quadratic doseresponse structures, and found agreement with the spline approach (Table 3). Based on the results of the spline, linear and quadratic structures alone it would appear that the Bayesian approach provides little additional insight into the effect of vitamin E on all-cause mortality, as was concluded in other analyses [1416]. However, accounting for the uncertainty in the mean structure and allowing for the possibility of no E effect in the prior distribution gives very different results when averaging over the http://ctj.sagepub.com

Vitamin E supplementation and mortality

Table 3 Results for posterior probabilities of vitamin E effects 95% PCIa Contains no E effect Model Spline Dose (IU) 200 400 800 1600 200 400 800 1600 200 400 800 1600 200 400 800 1600 Mb Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cc Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Ud Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Mb 0 0 0 0 0 0 0 0 0 0 0 0 0.94 0.94 0.94 0.94 No E effect Cc 0 0 0 0 0 0 0 0 0 0 0 0 0.93 0.93 0.93 0.93 Ud 0 0 0 0 0 0 0 0 0 0 0 0 0.94 0.94 0.94 0.94 Mb 0.40 0.17 0.12 0.14 0.74 0.51 0.18 0.10 0.75 0.26 0.06 0.11 0.05 0.02 0.01 0.00 E beneficial Cc 0.34 0.20 0.12 0.11 0.77 0.54 0.16 0.07 0.74 0.26 0.10 0.09 0.05 0.03 0.01 0.01 Ud 0.20 0.16 0.14 0.18 0.67 0.57 0.34 0.22 0.78 0.46 0.14 0.08 0.03 0.02 0.01 0.02 Mb 0.60 0.83 0.88 0.86 0.26 0.49 0.82 0.90 0.25 0.74 0.94 0.89 0.01 0.04 0.05 0.06 E harmful Cc 0.66 0.80 0.88 0.89 0.23 0.46 0.84 0.93 0.26 0.74 0.90 0.91 0.02 0.04 0.06 0.06

35

Ud 0.80 0.84 0.86 0.82 0.33 0.43 0.66 0.78 0.22 0.54 0.86 0.91 0.03 0.04 0.05 0.04

Linear

Quadratic

Model averaging

95% posterior credible interval; bM: data included in frequentist meta-analysis (Miller et al. [14]);cC: corrected data from that study; U: data published since that study.

Vitamin E Harmful 2.5

Vitamin E Beneficial 0 500 1000 Dose (IU) 1500 2000

Figure 2 Population relative risk of mortality according to vitamin E dose level. Based on data published by Miller et al. [14] The observed relative risk is plotted with the area of the circles proportional to the number of deaths in the study. The probability of no vitamin E effect (relative risk 1, dotted line) is $0.94 for all dose levels and the 95% probability interval contains relative risk 1.00 for all doses. The posterior mean is the solid curve and the 95% posterior probability interval is the dashed curve. To better show the other study results, the results from the study of patients with amyotrophic lateral sclerosis [46], which used a dose of 5000 IU, are not shown in the figure (relative risk 1.001)

uncertainty in the mean structure. Specifically, considering the four mean structures for the dose response relationship: (1) no E effect, (2) spline, (3) linear, and (4) quadratic, we give those four structures the same probability a priori and find the posterior probabilities that each is the correct http://ctj.sagepub.com

model among those considered. Consistent with the standard rules of probability, we average the results of the four models to find a final posterior probability, accounting for uncertainty in the choice of models. The major difference between the models with a single mean structure, as well as standard meta-analysis approaches, and the Bayesian model averaging approach is that each of the models with a single mean structure assumes the probability that the effect of E is equal to that of S is 0, regardless of the data. We list the posterior probabilities for each model in Table 4 for the updated data, with similar results obtained for the data published by Miller et al. and the corrected data. The most likely conclusion by far is that there is no effect of vitamin E. Looking at the posterior model probabilities and using the model that most likely allows for no effect of E gives a very different conclusion than assuming there is one! Accounting for model uncertainty is important and is left out in many meta-analyses [1416]. The model used for the remaining analyses is a weightadjusted average of the four mean structure models of the doseresponse relationship: no E effect, spline, linear, and quadratic. The posterior probabilities in Table 4 are the weights used in the Bayesian model averaging process. As shown in Table 4, the most likely models are the no E effect, with posterior probability 0.94, and the spline, with posterior probability 0.04. We first use a Bayesian meta-analysis modeling to analyze the posterior probabilities of the various Clinical Trials 2009; 6: 2841

Mortality relative risk

0.5

1.0

1.5

2.0

3.0

36

D Berry et al.
no E effect decreases from 0.948 under the assumption of homogeneity to 0.940 under the assumption of greatest heterogeneity. For our analyses, we used the prior distribution that allowed for the greatest heterogeneity, although our principal conclusion is not sensitive to this assumption. In analyses not shown we eliminated each trial in turn and found that the probability of no E effect ranged from 0.940 to 0.975 for the 22 possible combinations of 21 trials. Thus, no single trial had much of an impact on our analyses; our principal conclusion is robust in this sense.

Table 4 Prior and posterior probabilities for each model under consideration using the updated data Model Doseresponse Relationship Prior 1 No E effect Spline Linear Quadratic Prior probability Posterior probability

0.25 0.25 0.25 0.25

0.94 0.04 0.02 < 0.01

Similar results were obtained using the data published by Miller et al. [14] and the corrected data.

E effects at vitamin E doses of 200, 400, 800, and 1600 IU (Table 3). At all doses and in all analyses, the posterior probability of no E effect is large: 0.94 for the data published by Miller et al. [14] and the updated data, and 0.93 for the corrected data. For all three data sets, the probability of no E effect at a dose of 400 IU is approximately 24 times more likely than that of E being harmful, and 47 times more likely than that of E being beneficial. These analogous ratios are similar at other doses. Regarding model fit, the overall model estimates for the Bayesian meta-analysis approach agree very well with the actual observations for all 22 trials (data not shown) in the sense that discrepancies are consistent with random variability. In Figure 3, we show the posterior mean and 95% probability interval for the relative risks of mortality according to vitamin E dose level. Trials with more events are more informative and so tend to have shorter 95% probability intervals. All 95% probability intervals for the relative risks contain 1. Modeling doseresponse allows for borrowing strength from other doses and improves precision (i.e., shorter probability intervals). We include uncertainty in model selection and average over four models, thus our Bayesian meta-analysis with model averaging approach has wider intervals than does any single model. In Figure 4, we show the relative risk of mortality over a vitamin E dose range from 0 to 2000 IU, based on the updated data. The 95% probability intervals for the relative risk contain 1 (no E effect, either positive or negative) for all dose levels.

Discussion
Our goal has been to demonstrate Bayesian methods by assessing whether there is an effect or a doseeffect of vitamin E on all-cause mortality. We synthesized information from published randomized trials that examined the effects of vitamin E intake. We used Bayesian hierarchical models [54,55], in which there are two levels of experimental units: (i) subjects within trials and (ii) the trials themselves. The observed data are used to assess whether there is a doseeffect and also to estimate aspects of the population of trials [56]. Hierarchical models in the spirit of the one considered here have been developed and used by many researchers [5762]. We applied hierarchical models and used Bayesian model averaging to account for the uncertainty in the model choice. An important decision in conducting a metaanalysis is whether to treat trial as a fixed effect or a random effect. Bayesian hierarchical analysis takes a randomeffect approach and is a compromise between combining all trials into one and treating the results of each trial as a distinct conclusion. In this approach, the data themselves determine the extent of borrowing across trials. We considered a variety of models of the effect of vitamin E dose. The data dictate the weight given to these various models in our final conclusions. By including a variety of models, we accounted for uncertainty in the model development and selection process. Rather than dropping models that had less than the highest posterior probability, we included them with an appropriately discounted weight and averaged over the results of the individual models. Using a dataderived average of the various candidate models more accurately reflects the uncertainty that is present in the data. The difference between our conclusions and those of a frequentist study, such as that of Miller et al. [14] or Bjelakovic et al. [16], lies in the modeling process. Our model averaging approach is http://ctj.sagepub.com

Sensitivity analysis of the degree of trial heterogeneity In Table 5, we list the posterior probabilities for the various E effects at a dose of 400 IU for various prior distributions of the study population variance (s2). This is the critical parameter because it controls the amount of trial heterogeneity. The probability of Clinical Trials 2009; 6: 2841

Vitamin E supplementation and mortality

(a) MIN.VIT.AOX, 15 IU SU.VI.MAX, 30 IU Linxian A, 33 IU ATBC, 50 IU Linxian B, 60 IU Linqu, 200 IU GISSI, 300 IU PPP, 300 IU AREDS, 400 IU PPS, 440 IU WHS, 447 IU HOPE-TOO, 544 IU REACT, 600 IU MRC/BHF, 600 IU VECAT, 745 IU WAVE, 800 IU CHAOS, 910.4 IU SPACE, 1192 IU ADCS B, 2000 IU ADCS A, 2000 IU DATATOP, 2000 IU ALS, 5000 IU 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0

37

Vitamin E beneficial

Vitamin E harmful Relative Risk

(b)

MIN.VIT.AOX, 15 IU SU.VI.MAX, 30 IU Linxian A, 33 IU ATBC, 50 IU Linxian B, 60 IU Linqu, 200 IU GISSI, 300 IU PPP, 300 IU AREDS, 400 IU PPS, 440 IU WHS, 447 IU HOPE-TOO, 544 IU REACT, 600 IU MRC/BHF, 600 IU VECAT, 745 IU WAVE, 800 IU CHAOS, 910.4 IU SPACE, 1192 IU ADCS B, 2000 IU ADCS A, 2000 IU DATATOP, 2000 IU ALS, 5000 IU 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0

Vitamin E beneficial

Vitamin E harmful Relative risk

Figure 3 Relative risk of mortality according to vitamin E dose level. (a) All-cause relative risk of mortality, and 95% confidence interval based on each study individually. Each circle locates the empirical relative risk estimate; its area is proportional to the number of deaths in the study, (b) Relative risk of mortality, and 95% posterior probability intervals for each study based on borrowing across studies and doses. Each circle locates the posterior mean relative risk of the indicated study; its area is proportional to the number of deaths in the study

not dependent on any particular model and so gives rise to substantially less precision. Our probability intervals are wider because we include model uncertainty and do not assume a specific relationship between dose and mortality rate. In addition, our approach is less sensitive to local fluctuations in the observed doseresponse relationship. No model http://ctj.sagepub.com

can be known either a priori or a posteriori to be right. But our approach of considering a wide class of models has a greater chance of considering models that are close to being right. And models that are close to the process that gave rise to the actual data will tend to have high posterior probabilities. Clinical Trials 2009; 6: 2841

38
3.0

D Berry et al.
Vitamin E harmful 2.5

Vitamin E beneficial 0 500 1000 Dose (IU) 1500 2000

Figure 4 Population relative risk of mortality according to vitamin E dose level, based on updated data. The observed relative risk is plotted with the area of the circles proportional to the number of deaths in the study. The probability of no vitamin E effect (relative risk 1, dotted line) is again approximately 0.94 for all dose levels and the 95% probability interval contains relative risk 1.00 for all doses

Table 5 Sensitivity analyses for the prior distribution of the study population variance (s2) Heterogeneity No vitamin E effect 0.948 0.945 0.944 0.943 0.942 0.940 Vitamin E beneficial 0.015 0.032 0.034 0.019 0.021 0.020 Vitamin E harmful 0.037 0.037 0.038 0.038 0.039 0.040

Least (most homogeneous) Very little Little Moderate High Highesta

a

each compound in terms of the biological activity of one form, dl-alpha-tocopheryl acetate, and allows for comparison of the different compositions of vitamin E used across the studies. If doses were left as reported clinically, within-dose comparisons would be made among biologically different amounts of alpha-tocopherol. We used what we believe to be the appropriate conversion, putting all trials on the same footing. However, this issue had little effect on our analyses and conclusions. Although recent studies support the hypothesis that vitamin E has no effect on mortality, others have suggested benefits for certain populations. Three trials that we excluded from our metaanalysis because their mean intervention and follow-up times were <1 year illustrate such claims. One trial demonstrated that vitamin E supplementation may have a neuroprotective effect in cancer patients experiencing chemotherapy-induced peripheral nerve damage [6]. Another trial showed that vitamin E supplementation has a beneficial effect on clinical outcome for patients with acute myocardial infarction [64]. Still another trial demonstrated a protective effect of vitamin E supplementation on upper respiratory tract infections for elderly residents in nursing homes [65]. In summary, our Bayesian meta-analysis of data from 22 clinical trials concludes that vitamin E is unlikely to affect all-cause mortality, and that this is true regardless of dose.

Mortality relative risk

0.5

2.0

1.5

2.0

References
1. Goodman SN. Toward evidence-based medical statistics 2: The Bayes factor. Ann Intern Med 1999; 130: 100513. 2. Goodman SN. Toward evidence-based medical statistics 1: The p-value fallacy. Ann Intern Med 1999; 130: 9951004. 3. Berry DA. Statistics: A Bayesian Perspective Duxbury Press, 1995. 4. Berry DA. Bayesian clinical trials. Nat Rev Drug Discov 2006; 5: 2736. 5. Radimer K, Bindewald B, Hughes J et al. Dietary supplement use by US adults: Data from the National Health and Nutrition Examination Survey, 1999-2000. Am J Epidemiol 2004; 160: 33949. 6. Argyriou AA, Chroni E, Koutras A et al. Vitamin E for prophylaxis against chemotherapy-induced neuropathy: A randomized controlled trial. Neurology 2005; 64: 2631. ` 7. Mantovani G, Maccio A, Madeddu C et al. A phase II study with antioxidants, both in the diet and supplemented, pharmaconutritional support, progestagen, and anticyclooxygenase-2 showing efficacy and safety in patients with cancer-related anorexia/cachexia and oxidative stress. Cancer Epidemiol Biomarkers Prev 2006; 15: 10304. 8. Huang HY, Caballero B, Chang S et al. The efficacy and safety of multivitamin and mineral supplement use to prevent cancer and chronic disease in adults: A systematic review for a National Institutes of Health state-of-thescience conference. Ann Intern Med 2006; 145: 37285.

Prior used in our other analyses. Results are based on the data in Table 1. The prior distributions are indexed by percent increase in relative risk so that the probability of the indicated percent increase in relative risk for any one trial over another, at the same dose level of vitamin E, is 97.5%. The tabulated results are for a dose of 400 IU

The analyses by Miller et al. [14] and Bjelakovic et al. [16] have been criticized for excluding trials with fewer than 10 deaths [63]. We used the same criterion. The information content in any trial depends on the number of events in the trial. Very small trials contain little information about the relevant scientific question. Especially in the context of a hierarchical analysis (which considers trialspecific effects), data from very small trials add little strength to the overall conclusions. Further, we performed a sensitivity analysis that included the data from trials with <10 deaths and obtained nearly identical conclusions. Concern has also been raised regarding the conversion of various forms of alpha-tocopherol to international units of dl-alphatocopheryl acetate [63]. This conversion expresses Clinical Trials 2009; 6: 2841

http://ctj.sagepub.com

Vitamin E supplementation and mortality

9. Schroder FH, Roobol MJ, Boeve ER et al. Randomized, double-blind, placebo-controlled crossover study in men with prostate cancer and rising PSA: Effectiveness of a dietary supplement. Eur Urol 2005; 48: 92230discussion 9301. 10. Touvier M, Boutron-Ruault MC, Volatier JL, Martin A. Efficacy and safety of regular vitamin and mineral supplement use in France: Results from the ECCA study. Int J Vitam Nutr Res 2005; 75: 2019. 11. Hayden KM, Welsh-Bohmer KA, Wengreen HJ et al. Risk of mortality with vitamin E supplements: The Cache County study. Am J Med 2007; 120: 1804. 12. Bairati I, Meyer F, Jobin E et al. Antioxidant vitamins supplementation and mortality: A randomized trial in head and neck cancer patients. Int J Cancer 2006; 119: 22214. 13. Dabbagh OC, Aldawood AS, Arabi YM et al. Magnesium supplementation and the potential association with mortality rates among critically ill non-cardiac patients. Saudi Med J 2006; 27: 8215. 14. Miller ER, Pastor-Barriuso R, Dalal D et al. Metaanalysis: High-dosage vitamin E supplementation may increase all-cause mortality. Ann Intern Med 2005; 142: 3746. 15. Bjelakovic G, Nikolova D, Gluud LL et al. Mortality in randomized trials of antioxidant supplements for primary and secondary prevention: Systematic review and meta-analysis. JAMA 2007; 297: 84257. 16. Bjelakovic G, Nikolova D, Gluud LL et al. Antioxidant supplements for prevention of mortality in healthy participants and patients with various diseases. (Review). Cochrane Database Syst Rev 2008; 2: CD007176. 17. Berry DA. Statistical innovations in cancer research. In Kufe D, Bast R Jr, Hait W et al. (eds). Cancer Medicine. (7th edn). BC Decker, 2005, pp. 41125. 18. Madigan D, Raftery AE. Model selection and accounting for model uncertainty in graphical models using Occams window. J Am Stat Assoc 1994; 89: 153546. 19. Volinsky CT, Madigan D, Raftery AE, Kronmal RA. Bayesian model averaging in proportional hazard models: Assessing the risk of a stroke. Appl Stat 1997; 46: 43348. 20. Viallefont V, Raftery AE, Richardson S. Variable selection and Bayesian model averaging in case-control studies. Stat Med 2001; 20: 321530. 21. Draper D. Assessment and propagation of model uncertainty (with discussion). In Marsden PV. (ed.). Sociological Methodology. Blackwell Publishers, Oxford, UK, 1995, pp. 11163. 22. Raftery AE. Approximate Bayes factors and accounting for model uncertainty in generalized linear models. Biometrika 1996; 83: 25166. 23. Raftery AE, Madigan D, Hoeting JA. Model selection and accounting for model uncertainty in linear regression models. J Am Stat Assoc 1997; 92: 17991. 24. Hoeting JA, Madigan D, Raftery AE, Volinsky CT. Bayesian model averaging: A tutorial. Stat Sci 1999; 14: 382401. 25. Ferrigno L, Aldigeri R, Rosmini F et al. Associations between plasma levels of vitamins and cataract in the Italian-American Clinical Trial of Nutritional Supplements and Age-related Cataract (CTNS): CTNS Report #2. Opthalmic Epidemiol 2005; 12: 7180. 26. Bairati I, Meyer F, Gelinas M et al. Randomized trial of antioxidant vitamins to prevent acute adverse effects of radiation therapy in head and neck cancer patients. J Clin Oncol 2005; 23: 580513. 27. Arad Y, Spadaro LA, Roth M et al. Treatment of asymptotic adults with elevated coronary calcium scores with atorvastatin, vitamin C, and vitamin E: The

39

28.

29.

30.

31.

32.

33.

34.

35.

36.

37.

38.

39.

40.

41.

42.

St. Francis Heart Study randomized clinical trial. J Am Coll Cardiol 2005; 46: 16672. Hercberg S, Galan P, Preziosi P et al. The SU.VI.MAX study: A randomized, placebo-controlled trial of the health effects of antioxidant vitamins and minerals. Arch Intern Med 2004; 164: 233542. Lonn E, Bosch J, Yusuf S et al. HOPE and HOPE-TOO Trial investigators. Effects of long-term vitamin E supplementation on cardiovascular events and cancer: A randomized controlled trial. JAMA 2005; 293: 133847. Girodon F, Galan P, Monget AL et al. Impact of trace elements and vitamin supplementation on immunity and infections in institutionalized elderly patients: A randomized controlled trial. MIN. @VIT. @AOX geriatric network. Arch Intern Med 1999; 159: 74854. Blot WJ, Li JY, Taylor PR et al. Nutrition intervention trials in Linxian, China: Supplementation with specific vitamin/mineral combinations, cancer incidence, and disease-specific mortality in the general population. J Natl Cancer Inst 1993; 85: 148392. The Alpha-Tocopherol Beta Carotene Cancer Prevention Study Group. The effect of vitamin E and beta carotene on the incidence of lung cancer and other cancers in male smokers. N Engl J Med 1994; 330: 102935. Virtamo J, Pietinen P, Huttunen JK et al. Incidence of cancer and mortality following alpha-tocopherol and beta-carotene supplementation: A postintervention follow-up. JAMA 2003; 290: 47685. Li JY, Taylor PR, Li B et al. Nutrition intervention trials in Linxian, China: Multiple vitamin/mineral supplementation, cancer incidence, and disease-specific mortality among adults with esophageal dysplasia. J Natl Cancer Inst 1993; 85: 14928. You WC, Chang YS, Heinrich J et al. An intervention trial to inhibit the progression of precancerous gastric lesions: compliance, serum micronutrients and S-allyl cysteine levels, and toxicity. Eur J Cancer Prev 2001; 10: 25763. Gruppo Italiano per lo Studio della Sopravvivenza nellInfarto Miocardico. Dietary supplementation with n-3 polyunsaturated fatty acids and vitamin E after myocardial infarction: results of the GISSI-Prevenzione trial. Lancet 1999; 354: 44755. de Gaetano G Collaborative Group of the Primary Prevention Project. Low-dose aspirin and vitamin E in people at cardiovascular risk: A randomised trial in general practice. Lancet 2001; 357: 8995. Age-Related Eye Disease Study Research Group. A randomized, placebo-controlled, clinical trial of high dose supplementation with vitamins C and E and beta carotene for age-related cataract and vision loss: AREDS report no. 9. Arch Ophthalmol 2001; 119: 143952. Lee IM, Cook NR, Gaziano JM et al. Vitamin E in the primary prevention of cardiovascular disease and cancer: The Womens Health Study: A randomized controlled trial. JAMA 2005; 294: 5665. Greenberg ER, Baron JA, Tosteson TD et al. A clinical trial of antioxidant vitamins to prevent colorectal adenoma. Polyp Prevention Study Group. N Engl J Med 1994; 331: 1417. Chylack Jr LT, Brown NP, Bron A et al. The Roche European American Cataract Trial (REACT): a randomized clinical trial to investigate the efficacy of an oral antioxidant micronutrient mixture to slow progression of age-related cataract. Ophthalmic Epidemiol 2002; 9: 4980. Heart Protection Study Collaborative Group. MRC/ BHF heart protection study of antioxidant vitamin supplementation in 20,536 high-risk

http://ctj.sagepub.com

Clinical Trials 2009; 6: 2841

40

D Berry et al.
individuals: A randomised placebo-controlled trial. Lancet 2002; 360: 2333. McNeil JJ, Robman L, Tikellis G et al. Vitamin E supplementation and cataract: Randomized controlled trial. Ophthalmology 2004; 111: 7584. Waters DD, Alderman EL, Hsia J et al. Effects of hormone replacement therapy and antioxidant vitamin supplements on coronary atherosclerosis in postmenopausal women: A randomized controlled trial. JAMA 2002; 288: 243240. Stephens NG, Parsons A, Schofield PM et al. Randomised controlled trial of vitamin E in patients with coronary disease: Cambridge Heart Antioxidant Study (CHAOS). Lancet 1996; 347: 7816. Mitchinson MJ, Stephens NG, Parsons A et al. Mortality in the CHAOS trial. Lancet 1999; 353: 3812. Graf M, Ecker D, Horowski R et al. High dose vitamin E therapy in amyotrophic lateral sclerosis as add-on therapy to riluzole: Results of a placebo-controlled double-blind study. J Neural Transm 2005; 112: 64960. Boaz M, Smetana S, Weinstein T et al. Secondary prevention with antioxidants of cardiovascular disease in endstage renal disease (SPACE): Randomised placebocontrolled trial. Lancet 2000; 356: 12138. Petersen RC, Thomas RG, Grundman M et al. Vitamin E and donepezil for the treatment of mild cognitive impairment. N Engl J Med 2005; 352: 237988. Sano M, Ernesto C, Thomas RG et al. A controlled trial of selegiline, alpha-tocopherol, or both as treatment for Alzheimers disease. The Alzheimers Disease Cooperative Study. N Engl J Med 1997; 336: 121622. Parkinson Study Group. Mortality in DATATOP: A multicenter trial in early Parkinsons disease. Ann Neurol 1998; 43: 31825. Dietary reference intakes for vitamin C, vitamin E, selenium, and carotenoids. A report of the Panel on Dietary Antioxidants and Related Compounds, Subcommittees on Upper Reference Levels of Nutrients and Interpretation and Uses of Dietary Reference Intakes, and the Standing Committee on the Scientific Evaluation of Dietary Reference Intakes Food Nutrition Board. Institute of Medicine. National Academies Press, 2000. Berry SM, Berry DA. Bayesian survival analysis with nonproportional hazards: Meta-analysis of combination pravastatin-aspirin. J Am Stat Assoc 2004; 99: 3644. Berger JO. Statistical decision theory and Bayesian analysis. (2nd edn). Springer-Verlag, New York, 1985. Lindley DV, Smith AFM. Bayes estimates for the linear model. J Royal Stat Assoc 1972; 34: 141. Berger JO, Delampady M. Testing precise hypotheses. Stat Sci 1987; 2: 31752. Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian Data Analysis. Chapman and Hall, London, 1995. Carlin BP, Louis TA. Bayes and Empirical Bayes Methods for Data Analysis. Chapman and Hall, New York, 1996. Berry SM. Understanding and testing for heterogeneity across 2x2 tables: application to meta-analysis. Stat Med 1998; 17: 235369. Berry SM. Meta-analysis vs large trials: Resolving the controversies. In Stangl DK, Berry DA. (eds). Metaanalysis in Medicine and Health Policy. Marcel Dekker, New York, 2000, pp. 6581. Stangl DK. Prediction and decision making using Bayesian hierarchical models. Stat Med 1995; 14: 217390. Stangl DK. Hierarchical analysis of continuous time survival models. In Berry DA, Stangl DK. (eds). Bayesian Biostatistics. Marcel Dekker, New York, 1996, pp. 42950. 63. High-dosage vitamin E supplementation and all-cause mortality. Letters to the editor. |Ann Intern Med| 2005; 143: 1508. 64. Jaxa-Chamiec T, Bednarz B, Drozdowska D et al. Antioxidant effects of combined vitamins C and E in acute myocardial infarction. The randomized, doubleblind, placebo controlled, multicenter pilot Myocardial Infarction and VITamins (MIVIT) trial. Kardiol Pol 2005; 62: 34450. 65. Meydani SN, Leka LS, Fine BC et al. Vitamin E and respiratory tract infections in elderly nursing home residents: a randomized controlled trial. JAMA 2004; 292: 82836. Erratum in: JAMA 2004; 292: 1305.

43. 44.

45.

46. 47.

Technical appendix
Covariate analysis The prior mean of di is important because it determines the possible doseresponse relationships for E. For each covariate, we considered 13 forms for md(Di, Zi), where Di is the dose of trial i and Zi is the covariate for trial i, and took them to be equally plausible. These 13 forms correspond to the 13 models that we considered (Supplementary Table 1). The doseresponse relationship of the 13 models is as follows: (1) no E effect; (2) a linear dose effect; (3) a quadratic dose effect; (4)(6) covariate interactions with slope and/ or intercept of linear dose effect; and (7)(13) covariate interactions with terms in quadratic dose effect models. We provide the mean structure and description of each model in Supplementary Table 1 and the posterior probability for each model in Supplementary Table 2. To determine if the covariates under consideration (vitamin E used in combination with other vitamins/minerals, subject age, sex, and follow-up duration) are related to any possible doseeffect of vitamin E, we separately included each covariate in the 13 models described in Supplementary Table 1 and compared each of the models 2 through 13 with model 1, no E effect. In Supplementary Table 2, we show the posterior probabilities of each model, which we used to assess the importance of each of the four covariates. We included each covariate individually in models 413 and compared that to models 13, which do not include the covariate. If the posterior probability of any model of models 413 is large, then we considered the covariate used in that model to be important. Bayesian meta-analysis incorporates uncertainty in model choice by averaging the posterior distributions over the various models. The weights in this averaging process are the posterior probabilities of the various models. http://ctj.sagepub.com

48.

49. 50.

51. 52.

53.

54. 55. 56. 57. 58. 59. 60.

61. 62.

Clinical Trials 2009; 6: 2841

Vitamin E supplementation and mortality

Supplementary Table 1 Model 1 2 3 4 5 6 7 8 9 10 11 12 13
a

41

Mean structure for models under consideration md(Di, Zi)a Description No E effect, no covariate effect Linear dose effect, no covariate effect Quadratic dose effect, no covariate effect Linear dose effect with different slope for each covariate group Linear dose effect with different intercept for each covariate group Linear dose effect with different slope and intercept for each covariate group Quadratic dose effect, treatment covariate interaction Quadratic dose effect, treatment covariate interaction Quadratic dose effect, treatment covariate interaction Quadratic dose effect, treatment covariate interaction Quadratic dose effect, treatment covariate interaction Quadratic dose effect, treatment covariate interaction Quadratic dose effect, treatment covariate interaction

0 0 0Di 0 0Di 0Di2 0 0Di 1DiZi 0 0Di 1Zi 0 0Di 1Zi 1DiZi 0 0Di 0Di2 1Zi 0 0Di 0Di2 1DiZi 0 0Di 0Di2 1Di2Zi 0 0Di 0Di2 1Zi 1DiZi 0 0Di 0Di2 1Zi 1Di2Zi 0 0Di 0Di2 1DiZi 1Di2Zi 0 0Di 0Di2 1Zi 1DiZi 1Di2Zi

Di is the dose of trial i; Zi is the covariate for trial i.

Supplementary Table 2 Posterior probabilities for models used to assess the importance of each covariate Model Model posterior probability assuming the indicated covariate is included in models 413 E Comba 1 2 3 4 5 6 7 8 9 10 11 12 13 0.956 0.024 0.005 0.006 0.007 0 0 0 0 0 0 0 0 Age 0.952 0.023 0.004 0.007 0.004 0.001 0.001 0.004 0.002 0 0 0 0 Sexb 0.969 0.013 0.005 0.004 0.002 0 0 0 0 0 0 0.005 0 FUc 0.959 0.023 0.006 0.006 0.003 0.001 0 0 0.003 0 0 0 0

Based on data in Supplementary Table 1. aE Comb: whether vitamin E used in combination with other vitamins/minerals; b Sex: percent men in the study population; cFU: follow-up duration.

http://ctj.sagepub.com

Clinical Trials 2009; 6: 2841

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.