Page 1

Investigating heterogeneity in studies of resting energy expenditure

in persons with HIV/AIDS: a meta-analysis1–3

Marijka J Batterham

ABSTRACT

Background: There is conflict in the literature about the extent of

alterationsofrestingenergyexpenditure(REE)inpersonswithHIV.

Objective: The study was conducted to ascertain the mean differ-

enceinREE(inkJ)perkilogramoffat-freemass(FFM;REE/FFM)

between HIV-positive subjects and control subjects and to investi-

gate heterogeneity in the literature.

Design:Ameta-analysiscomparingclassicalandBayesianmethods

was conducted. Heterogeneity was investigated by using subgroup

analysis, metaregression, and a mixed indirect comparison.

Results: Of 58 studies meeting the inclusion criteria, 32 included

bothHIV-positiveandcontrolgroups;24ofthese32wereincluded.

Thirty-seven studies were used in the mixed indirect comparison,

and 30 were used in the subgroup comparisons of the HIV-

symptomatic, lipodystrophy, weight-losing, and weight-stable sub-

groups and the healthy (HIV-negative) control group. Mean REE/

FFM was significantly higher in 732 HIV-positive subjects than in

340 control subjects [11.93 kJ/kg (95% CI: 8.44,15.43 kJ/kg) and

12.47 kJ/kg (95% CI: 8.19,16.57 kJ/kg), classical and Bayesian

random effects, respectively]; the test for heterogeneity was signif-

icant (P ? 0.001). Both the mixed indirect comparison and the

subgroupanalysisindicatedthatREE/FFMwashighestinthesymp-

tomatic subgroup; however, the small number of studies investigat-

ing symptomatic subjects limited statistical comparisons. The pres-

ence of lipodystrophy, use of highly active antiretroviral therapy,

subject age, and method of body-composition measurement could

not explain the heterogeneity in the data with the use of metaregres-

sion.

Conclusions: REE/FFM (kJ/kg) is significantly higher in HIV-

positivesubjectsthaninhealthycontrolsubjects.SymptomaticHIV

infection may contribute to the variations reported in the

literature.

Am J Clin Nutr 2005;81:702–13.

KEY WORDS

abolicrate,meta-analysis,metaregression,Bayesianmethods,body

composition, lipodystrophy

HIV, resting energy expenditure, resting met-

INTRODUCTION

Controversy surrounds the role of resting energy expenditure

(REE) in HIV-related metabolic abnormalities such as wasting

and lipodystrophy syndrome. Early in the HIV pandemic, it was

proposedthatREEmaybeelevatedinpersonswithHIV,because

it was thought that neither malabsorption or decreased energy

intake alone nor both together could explain the weight loss that

wascharacteristicofuntreatedinfection(1).Innoninfectivemal-

absorption, malnutrition, or underfeeding, an appropriate meta-

bolic response is a compensatory drop in REE to preserve the

fat-free mass (FFM) (2). The first 2 published reports were con-

flicting; Kotler et al (3) reported significantly lower REE/kg

FFM in HIV-positive subjects than in healthy control subjects,

whereas Hommes et al (4) reported significantly higher REE

adjustedforFFMinHIV-positivesubjectsthaninhealthycontrol

subjects. The literature published since those initial reports is

difficulttosummarizeinformally,becausedifferentauthorshave

presented the results by using different summary statistics and

different population subgroups, and because methods of mea-

suringbodycompositionvaryinthestudies.BecauseFFMisthe

primary determinant of REE, accounting for 70–80% of the

variation in REE in healthy subjects (5), comparisons between

studies must investigate whether the use of reference or field

methods of body-composition measurement explains some of

the discrepancies in the results.

Since the widespread introduction of highly active antiretro-

viral therapy (HAART) in westernized countries in 1996, a new

metabolicabnormality,lipodystrophy,hasbeendescribed(6);its

etiology remains unknown, but the syndrome is usually associ-

ated with the use of HAART, particularly the protease inhibitor

classofdrugsbutalsothenucleosideanaloguereversetranscrip-

tase inhibitors (7). More recently, there have been conflicting

reports about the differences in REE between HIV-positive sub-

jects with and without lipodystrophy syndrome (8, 9).

Clinically, it is of value to establish whether REE is indeed

elevated in HIV-positive persons (or in subsets of the HIV-

positivepopulation)sothatappropriatenutritionaladviceonthe

necessary energy intake to achieve the desirable weight can be

provided. An integration of the currently available literature is

necessary to provide this clinical information.

The primary aim of this study was, by conducting a meta-

analysis combining previous research, to ascertain an overall

mean difference in REE between HIV-positive subjects and

healthy control subjects. Two secondary aims of this study were

toinvestigatevariationsinthismeandifferencebetweenvarious

clinical subgroups (eg, persons with lipodystrophy, those who

1From the Smart Foods Centre, University of Wollongong, Australia.

2Supported by the Australian Research Council.

3Reprints not available. Address correspondence to M Batterham, Smart

FoodsCentre,UniversityofWollongong,NorthfieldsAvenue,Wollongong,

NSW 2522, Australia. E-mail: marijka@uow.edu.au.

Received June 16, 2004.

Accepted for publication November 16, 2004.

702

Am J Clin Nutr 2005;81:702–13. Printed in USA. © 2005 American Society for Clinical Nutrition

by guest on June 3, 2013

ajcn.nutrition.org

Downloaded from

C1.html

http://ajcn.nutrition.org/content/suppl/2005/03/09/81.3.702.D

Supplemental Material can be found at:

Page 2

arelosingweight,thosewhoaresymptomatic,andthosewhoare

weightstable)andtoinvestigatetheeffectofpotentialconfound-

ingcovariatesonthatmeandifference.Apriori,thesecovariates

were the presence or absence of lipodystrophy in the HIV-

positive group, the choice of method of body-composition

measurement, the use of HAART in the HIV-positive group,

and the mean age of the subjects. In addition, because of the

rapidly increasing popularity of the Bayesian framework over

classical analysis, which results from the Bayesian frame-

work’s allowing the incorporation of external evidence and

accounting for all sources of variability (10), both types of

analysis were reported.

MATERIALS AND METHODS

Search strategy and identification of reports

A database search of Medline, CINAHL, Current Contents

Connect, HIV/AIDS database, and Expanded Academic Index

(from1981,whenAIDSwasfirstdescribed,toSeptember2004)

was conducted by using the search terms “resting energy expen-

diture and HIV,” “resting metabolic rate and HIV,” “indirect

calorimetry and HIV,” and “fat oxidation and HIV.” Identified

articles were then searched to ascertain whether they met the

following inclusion criteria: 1) the studies were conducted in

humans; 2) the subjects were adults; 3) the subjects were mea-

sured after an overnight fast or under a strict postprandial pro-

tocol(becausediet-inducedthermogenesishasbeenshowntobe

prolonged past 300 min in HIV-positive subjects) (11); and 4)

data on body composition were collected.

Only studies published in English were considered. The ref-

erences of the articles identified were also searched, and authors

were contacted to identify additional publications.

Data analysis

Crude, classical, and Bayesian methods of meta-analaysis

were performed for comparison and to maximize the use of the

available data. To use all of the available data comparing HIV-

positive subjects and healthy control subjects when estimates of

variance were not provided, crude meta-analysis based on the

method of Gotzsche (12) was performed. This technique, in

which the mean REE of each group is divided by the mean FFM

of each group, and that value is then used to calculate an overall

mean difference and 95% CIs, has been used previously in a

similar meta-analysis (13).

A traditional meta-analysis, in the classical framework, com-

paring REE divided by FFM (kJ/kg) in the HIV-positive and

control groups and within the HIV subgroups, was performed.

The random-effects model based on the method-of-moments

estimatorasproposedbyDerSimonianandLaird(14)wasused,

with the inverse-variance fixed-effects model shown for com-

parison; algebraic descriptions of these models are available

elsewhere (15, 16). These analyses were performed by using the

METAN (17) command in STATA software (version 7.0; Stata

Corporation, College Station, TX).

A Bayesian random-effects method was also used to estimate

the common effect and the between-study variance (?2) and to

compare this estimate with the estimates obtained by using the

method of DerSimonian and Laird (18). Here, too, the Bayesian

fixed model is shown for comparison. Classical (or frequentist)

methods of statistics assume that each study (in the current case,

thismeta-analysis)isoneinalong-runningseriesofexperiments

inwhichthecurrentstudyestimate(theoverallmeandifference)

is likely to lie within the stated CIs 95% of the time. This differs

from the Bayesian approach, in which the study estimate (the

overall mean difference) has a probability distribution that ex-

presses our prior belief (prior distribution) about the mean com-

bined with the available data (likelihood) (19). Two prior distri-

butions were used: a noninformative prior distribution and an

informative prior distribution calculated by using the data pro-

videdfromthesingle-armHIVstudies.Single-armstudiescanbe

incorporated into a Bayesian analysis in the form of a prior

distribution, which allows for the use of these data (15). The

ability to incorporate external information, such as previous re-

search, by using an informed prior distribution (10) is an advan-

tage of the Bayesian technique. To ascertain the effect size and

variance for these single-arm estimates, a bootstrap resampling

procedure was used to generate control samples with replace-

ment from the pooled data of the available individual studies. In

all Bayesian models, the likelihood was initially assumed to be

from a normal distribution, the prior distribution for the mean

difference was also assumed to be normally distributed, and

thepriordistributionfortheheterogeneitytermintherandom-

effects models was assumed to be a conjugate gamma distri-

bution, because this is computationally convenient and be-

cause it provides the effect estimate (posterior distribution) in

the same form as the likelihood (normal distribution) (18).

The choice of gamma prior distributions for the heterogeneity

term was tested by comparing the results with those obtained

by using a uniform distribution on ? (20). Random-effects

models with the likelihood that the mean difference between

the studies would be coming from a Student’s t distribution

with the noninformative prior distributions were also evalu-

ated for comparison and to test the assumption of normality of

the mean difference estimate. A model with 4 df and a model

with the df modeled as an additional unknown parameter are

presented (21, 22).

A mixed indirect comparison was performed in the Bayesian

frameworkonthebasisofthemethodproposedbySpiegelhalter

et al (20). This model allows the incorporation of all available

information regardless of whether there are data for all compar-

isons. The model is set up to compare the HIV subgroups (ie,

lipodystrophy, weight-losing, weight-stable, and symptomatic)

withhealthycontrolsubjects,andwithin-groupcomparisonsare

also modeled.

All Bayesian models were performed by using WINBUGS

software (version 1.4; Imperial College & Medical Research

Council, Cambridge UK, 2003). A burn-in period of 5000 iter-

ationswasusedforallmodels,andthesubsequent5000iterations

were used to estimate the variables of interest. Diagnostic pro-

cedures including inspection of autocorrelation plots and trace

histories were performed to verify the results. WINBUGS em-

ploys a simulation-based Gibbs sampling technique, which is a

type of Markov chain Monte Carlo method (10).

An ancillary analysis was performed on the available individ-

ualdatatoobtainanestimateofREEadjustedforFFM(23).This

analysis was conducted by using the general linear model (anal-

ysis of covariance) in SPSS for WINDOWS software (version

11.5.0;SPSSInc,Chicago,IL).Twomodelswereemployed:an

analysis with HIV-positive or healthy control subjects as the

factor and an analysis considering the study site in addition to

HIV status of the subject.

HIV REE META-ANALYSIS

703

by guest on June 3, 2013

ajcn.nutrition.org

Downloaded from

Page 3

Assessment and exploration of heterogeneity

Heterogeneity was detected by visual inspection and by using

the chi-square test and was quantified by calculating the I2sta-

tistic by the equation

I2? ?Q ? df?/Q

(1)

where Q is the heterogeneity statistic, and where df (24)

reflects the percentage of the variability between studies that

can be attributed to heterogeneity rather than to sampling

error (24). Heterogeneity was initially investigated by con-

sidering the factors identified a priori as potential confound-

ers in establishing a common effect. Subgroup analyses were

used to visually assess differences between groups, and a

more rigorous investigation was provided by the mixed

indirect comparison model. Formal investigation of hetero-

geneity was performed by univariate metaregression with

lipodystrophy, technique of body-composition measure-

ment, use of HAART in the HIV-positive group, and mean

age of subjects as covariates. A random-effects model based

on a restricted maximum likelihood estimate was used, and

analyses were performed by using the METAREG (25)

command in STATA software. Bayesian metaregression was

also conducted with both normal and t distributions for the

likelihood of the mean differences between the studies.

Publication bias

Publication bias was assessed visually by using a funnel plot

(26) that plots sample size against the treatment effect and that

has a funnel shape when publication bias is not present. Formal

testing for publication bias was conducted by using the rank

correlationtestofBeggandMazumdar(27),theregressiontestof

Egger et al (28), and the “trim and fill” method of Duval

and Tweedie (29), a nonparametric method that estimates the

numberofmissingstudiesalongwiththeirmean?SEdifference

and then adds in the estimated values, recalculating the overall

mean ? SE. Because of the limited number of studies in the

subanalyses, these tests are performed only on the primary sam-

ple estimates.

Sensitivity analysis

Several sensitivity analyses were conducted on the primary

analysis. First, the presentation and comparison of the fixed

effects as well as the random-effects analysis are regarded as a

sensitivity analysis (15). The fixed-effect model makes the as-

sumptionsthatthereisnobetween-studyvariationandthatallthe

studies are estimating a single underlying effect (10). In the

random-effects model, the incorporation of the heterogeneity

parameter (?2) allows the assumption that the studies are esti-

mating different underlying effect sizes (15). Statistically, the

decision to use the random-effects model is based on the signif-

icance of the test for heterogeneity, but both models should

produce a similar overall estimate.

Second, it was important to ascertain whether the studies not

reporting REE/FFM were in some way affecting the estimate of

the effect size. To investigate this, approximations of the effect

sizewerecalculatedbydividingthemeanREEbythemeanFFM

as discussed above (12). These values were used in the analysis

as substitutes for the 2 studies that did not directly provide the

REE/FFM estimate. The SDs were then calculated by using the

method recommended by Whitehead (18), in which a pooled

varianceiscalculatedforallthestudieswithavailablevalues,and

this value is then used to calculate the individual variance of the

studies with missing values.

Third, influence analyses were used to investigate the sensi-

tivity of the estimates to individual studies. For the classical and

Bayesian random-effects models, sensitivity to individual stud-

ies was assessed by excluding and then replacing one study at a

time and recalculating the mean difference and 95% CI (in the

classical analysis, CI ? confidence interval; in the Bayesian

analysis,CI?credibleinterval)andtheestimateofthebetween-

studyvarianceorheterogeneity(?2)and95%CIfortheBayesian

analysis (16).

RESULTS

Seventy-three studies were identified in the initial search.

Seven studies were conducted in children and were excluded on

thatbasis(30–36).Twostudieswereexcludedbecausethemea-

sureswerenotfastingorstandardized(37,38),and6studieswere

excluded because body composition was either not assessed or

not reported (39–44). Thus, 58 studies remained for consider-

ation in the analysis (3, 4, 8, 9, 11, 45–97). Of these 58 studies,

32 included an HIV-negative control group, and they were con-

sidered for the primary analysis. Twenty-six of these 32 studies

provided enough information for calculation of the crude effect

estimate, and the inclusion of both an estimate and SD or SE

allowed formal statistical analysis of 24 studies. When data for

theHIV-positivegroupwerepresentedonlyinsubsets,eitherthe

asymptomatic, the weight-stable, or the HIV as opposed to the

AIDSgroupwasusedfortheprimaryanalysis,andtheadditional

subsets were used in the mixed indirect comparison estimate for

which37studieswereincludedandinthesubgroupcomparisons

for which 30 studies were available. Authors of the 58 studies

initially considered were contacted to ascertain whether they

would provide either their original data for calculations of sum-

mary statistics or the summary statistics if the raw data were not

provided in their reports. Individual data were available for 21

studies, 9 of which included a control group; it is important that

individualdataorsummaryestimateswereprovidedforallstud-

iesthatincludedsubjectswithlipodystrophy.Relevantdetailsof

allthestudiesconductedinadultsaresummarizedelsewhere(see

Appendix A under “Supplemental data” in the current issue at

www.ajcn.org).

The mean effect size and 95% CIs for the difference between

785 HIV-positive subjects and 403 healthy control subjects—

11.02kJ/kgFFM(95%CI:8.03,14.01)—wasobtainedbyusing

the method of Gotzsche (12), and those findings indicated that

overall REE was significantly higher in HIV-positive subjects

than in healthy control subjects. The means and sample sizes for

the studies included in this analysis are shown in Table 1.

TheforestplotinFigure1showseachstudy’seffectestimate

and 95% CI, as well as the overall estimate for the DerSimonian

and Laird random-effects primary analysis comparing the HIV-

positivesubjectsandthecontrolsubjects.Means,SDs,andsam-

plesizesareshowninTable1.Themeandifferenceestimatesand

95% CIs (or credible intervals for the Bayesian analyses) for the

fixed- and random-effects models for the traditional and Bayes-

ian analyses are shown in Table 2. All analyses indicated that

REE/FFMwassignificantlyhigherinHIV-positivesubjectsthan

in healthy control subjects.

704

BATTERHAM

by guest on June 3, 2013

ajcn.nutrition.org

Downloaded from

Page 4

For the random-effects models, the between-study variance

(?2) is shown along with the overall mean difference estimates.

The ?2is inflated in the Bayesian models on the basis of the

normal distribution and, to a lesser extent, by using the t distri-

butionwiththedfmodeledasanunknownparameter.Therefore,

the estimate based on the t distribution with 4 df is used for

comparison.

In this analysis, the informed prior distribution was similar to

the data set incorporated in the likelihood and did not alter the

posterior distribution (Table 2). In addition, the use of uniform

prior distributions on ?, instead of the gamma prior distribution,

resultedinsimilarestimatesfortherandom-effectsmodel,which

suggests that the estimates were not sensitive to the distribution

of the prior distribution (20) (Table 2).

Theuseoftherandom-effectsclassicalestimateoftheoverall

meandifference(11.93kJ/kg)andmultiplicationofthatnumber

bythemeanFFMintheHIV-positivegroups(53kg)givesadaily

differenceinenergyexpenditurebetweenHIV-positiveandcon-

trol subjects of ?630 kJ (classical estimate) or ?661 kJ [Bayes-

ian random-effects estimate based on the t distribution (12.47

kJ/kg)]. This difference reflects an elevation in REE of ?9%

whenthemeanREE/FFMforHIV-positivesubjectsisdividedby

the REE/FFM for control subjects.

By using the pooled individual-subject dataset, REE in HIV-

positive subjects (n ? 587) was significantly higher than that in

control subjects (n ? 189) after adjustment for FFM (7229 and

6729 kJ/kg, respectively; P ? 0.001) by analysis of covariance

(using the general linear model). Similarly, after adjustment for

the study site, the elevation remained significantly (P ? 0.001)

different between HIV-positive subjects (n ? 420; 7036 kJ/kg)

and control subjects (6452 kJ/kg). These differences represent

elevationsof?7%and9%,respectively,andareconsistentwith

the meta-analysis values.

Assessment and exploration of heterogeneity

The test for heterogeneity was significant (?2? 88.73, df ?

23; P ? 0.001), and the proportion of variability in the mean

differenceduetoheterogeneityratherthantosamplingerror(I2)

was 74%. Forest plots of the various subgroupings are shown in

Figure 2. Within the HIV-positive subgroups, there was a trend

towardhigherREE/FFMinthesymptomaticsubjectsthaninthe

weight-stable subjects (P ? 0.079). Differences in REE/FFM

between the weight-losing (P ? 0.590) and lipodystrophy (P ?

0.976) groups and the healthy control subjects were not signifi-

cant. All HIV subgroups showed significantly higher REE/FFM

thanwasseeninthecontrolsubjects:symptomaticsubjects,P?

0.012; weight-losing subjects, P ? 0.020l; subjects with lipo-

dystrophy, P ? 0.001; and weight-stable subjects, P ? 0.001).

Heterogeneity is evident in all subgroup comparisons (P ?

0.001)exceptthatbetweenweight-losingandweight-stablesub-

jects(P?0.212).Themixedindirectcomparisonmodel(Table

3) shows that REE/FFM is significantly higher in all HIV sub-

groups than in the healthy control subjects and, in addition, that

symptomatic subjects have significantly higher REE/FFM than

doboththehealthycontrolsubjectsandtheotherHIVsubgroups.

The slope coefficients and intercepts of the metaregression

analysesareshowninTable4.Lipodystrophy,methodofbody-

compositionmeasurement,subjectage,andHAARTusageinthe

HIV-positive group were unable to explain the heterogeneity by

using metaregression in either a classical or Bayesian frame-

work, as indicated by the 95% CIs (as confidence intervals and

credible intervals, respectively) and P values of the slope coef-

ficients.

Data were available from only 2 studies that included groups

of HIV-positive subjects without lipodystrophy who were re-

ceiving and not receiving HAART. Our previous research (87)

found REE/FFM in 32 subjects taking HAART to be 9.5 kJ/kg

(95% CI: ?21.51, 2.51) lower than that in 8 subjects not taking

HAART, whereas Kosmiski et al (8) found that REE/FFM in 13

subjectstakingHAARTwas8.79kJ/kg(95%CI:?2.21,19.78)

higherthanthatin5subjectsnottakingHAART,whichledtoan

overall mean difference, combining these 2 studies, of ?0.19

kJ/kg(95%CI:?18.11,17.73).Asapreliminaryanalysis,the24

studies for the main analysis were split into 2 separate groups (5

studies in which all subjects were using HAART and 19 studies

in which HIV-positive subjects were not using HAART); the

studies in which subjects were using HAART had an overall

mean difference in REE of 15.10 kJ/kg FFM compared with

healthy controls, whereas the studies in which the HIV-positive

subjects were not using HAART showed a mean difference of

11.02 kJ/kg FFM, which equates to an estimated difference of

216 kJ/d.

TABLE 1

Resting energy expenditure (REE) and sample sizes for the study

comparing HIV-positive subjects with healthy control subjects1

Reference

REE2

HIV-positive subjects Control subjects

Kotler et al (3)

Hommes et al (4)

Hommes et al (49)

Melchior et al (47)

Melchior et al (50)

Mulligan et al (52)

Salehian et al (92)

Godfried et al (54)

Macallan et al (53)

Sharpstone et al (45)

Schwenk et al (58)

Sharpstone et al (46)

McNurlan et al (88)

Heijligenberg et al (60)

Jimenez-Exposito et al (66);

Garcia-Lorda et al (94)3

Grinspoon et al (65)4

Pernerstorfer-Schoen et al (74)4

Lane and Provost-Craig (75)

Coors et al (79)

Sekhar et al (83)

Hadigan et al (81)

Korach et al (82)

Luzi et al (86)

Kosmiski et al (8, 85)5

Batterham et al (87)

Crenn et al (93)

121 ? 22 [5]

134 ? 18 [18]

128 ? 8 [11]

156 ? 19 [50]

153 ? 14 [129]

136 ? 5 [6]

139 ? 5 [4]

136 ? 17 [12]

129 ? 13 [51]

137 ? 14 [104]

118 ? 27 [65]

148 ? 17 [10]

139 ? 11 [9]

131 ? 12 [9]

142 ? 19 [85]

155 ? 14 [6]

113 ? 10 [11]

113 ? 10 [11]

125 ? 17 [14]

132 ? 14 [31]

123 ? 5 [6]

123 ? 10 [4]

123 ? 9 [15]

114 ? 8 [12]

129 ? 12 [57]

117 ? 9 [29]

147 ? 14 [10]

129 ? 12 [9]

125 ? 7 [9]

128 ? 11 [19]

173 ? — [33]

142 ? — [29]

131 ? 7 [10]

145 ? 7 [7]

128 ? 26 [6]

126 ? 12 [19]

87 ? 8 [9]

139 ? 25 [12]

145 ? 18 [23]

135 ? 15 [70]

137 ? 13 [8]

151 ? — [26]

122 ? — [37]

125 ? 10 [16]

141 ? 13 [7]

106 ? 11 [6]

113 ? 15 [8]

93 ? 14 [9]

128 ? 19 [12]

121 ? 15 [14]

122 ? 12 [16]

111 ? 11 [9]

1All values are x ? ? SD; n in brackets.

2Measured in kJ per kg fat-free mass.

3Data were pooled from 2 studies with the same control group.

4A variance estimate was not available for these studies, which are

included in the crude analysis and sensitivity analysis only.

5Estimates were based on raw data provided by the authors of these 2

publications; the SD for the control group is the pooled SD, because data for

control subjects were not available.

HIV REE META-ANALYSIS

705

by guest on June 3, 2013

ajcn.nutrition.org

Downloaded from

Page 5

Publication bias

The funnel plot (Figure 3) shows both an overestimation and

underestimation of the mean effect size, with the results in 10

studies clearly greater than the mean and those in another 10

studies below the mean; 4 studies fall virtually on the mean.

There is some evidence of a “tunnel” or gap separating the out-

lyingstudybyKotleretal(3).The2studiesaroundthelineofno

difference between the groups both have a relatively small sam-

ple size, which indicates that nonsignificant studies with small

sample sizes are published. The formal tests for publication bi-

as—the tests of Egger et al (bias: ? ? ?0.45, t ? ?0.41; P ?

0.686)andBeggandMazumdar(z?0.32,P?0.747)—werenot

significant.Visually,thisfindingissupportedbypresentationof

the funnel plot of Begg and Mazumdar (Figure 3), which is

symmetricalexceptfortheoutlyingstudybyKotleretal(3).The

trimandfillmethodsuggestedtherewerenomissingstudiesand

produced an unadjusted estimate, and that, again, suggests that

there was no publication bias in this analysis.

Sensitivity analysis

Thefixed-effectestimatesprovidedinTable2weresimilarto

therandom-effectsestimatesobtainedbyusingboththeclassical

and Bayesian methods. The mean difference between studies

was similar with the use of all models. The Bayesian random-

effects analysis using the t distribution on 4 df produced the

lowest value for the heterogeneity parameter ?2and is therefore

the preferred estimate.

The calculated pooled SD was 15.15 kJ/kg FFM. This value

wassubstitutedfortheSDforthe2studiesincludedinthecrude

analysis for which a variance estimate was not available so that

thestudycouldbeincludedintheprimaryanalysis(Table1).The

random-effects meta-analysis was redone with these additional

studiesincluded.Thecombinedmeandifferencewas12.77kJ/kg

FFM(95%CI:9.45,16.10kJ/kgFFM;z?7.53,P?0.001).This

difference would represent a daily difference in energy expen-

ditureofonly16kJ(basedonanaverageFFMintheHIVstudies

of 53 kg) when compared with the random-effect Bayesian es-

timate (Student’s t distribution on 4 df).

TheBayesiananalysisoftheinfluenceofindividualstudiesis

shown in Table 5. The removal of the data from the 2002 study

by Korach et al (82) produced the highest mean difference, and

the removal of the data from the 2004 study by Crenn et al (93)

produced the lowest estimate; this difference translated into a

small change in daily energy expenditure (61 kJ/d based on the

average FFM of 53 kg). The most substantial effect on the

between-study variation (?2), which decreased by 74% of the

combinedestimatevalue,wasseenwhenthedatafromthestudy

of Kotler et al (3) were removed. The classical framework pro-

duced similar results, although the omission of the studies of

Kotleretal(mean12.72kJ/kg;95%CI:9.49,15.95)andKorach

et al (mean 12.72 kJ/kg; 95% CI: 9.33, 16.10) tied to give the

highest mean difference estimate, and the omission of the study

by Melchior et al (47; mean 11.17 kJ/kg; 95% CI: 7.82, 14.52)

gave a slightly lower mean estimate than did that of the study of

FIGURE1.Forestplotshowingtheoverallcomparisonofrestingenergyexpenditure(REE;inkJ)perkgfat-freemass(FFM)inHIV-positivesubjectsand

healthy control subjects.

706

BATTERHAM

by guest on June 3, 2013

ajcn.nutrition.org

Downloaded from