0% found this document useful (0 votes)

68 views39 pages

Normand (1999) - Meta-Analysis. Formulating, Evaluating, Combining, and Reporting

Meta-Analysis. Formulating, Evaluating, Combining, And Reporting

Uploaded by

Diego Ars

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

68 views39 pages

Normand (1999) - Meta-Analysis. Formulating, Evaluating, Combining, and Reporting

Meta-Analysis. Formulating, Evaluating, Combining, And Reporting

Uploaded by

Diego Ars

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 39

STATISTICS IN MEDICINE

Statist. Med. 18, 321–359 (1999)

TUTORIAL IN BIOSTATISTICS
META-ANALYSIS: FORMULATING, EVALUATING,
COMBINING, AND REPORTING

SHARON-LISE T. NORMAND ∗
Department of Health Care Policy, Harvard Medical School, 180 Longwood Avenue, Boston, MA 02115, U.S.A., and
Department of Biostatistics, Harvard School of Public Health, 677 Huntington Avenue, Boston, MA 02115, U.S.A.

SUMMARY
Meta-analysis involves combining summary information from related but independent studies. The objectives
of a meta-analysis include increasing power to detect an overall treatment e ect, estimation of the degree of
beneÿt associated with a particular study treatment, assessment of the amount of variability between studies,
or identiÿcation of study characteristics associated with particularly e ective treatments. This article presents
a tutorial on meta-analysis intended for anyone with a mathematical statistics background. Search strategies
and review methods of the literature are discussed. Emphasis is focused on analytic methods for estimation of
the parameters of interest. Three modes of inference are discussed: maximum likelihood; restricted maximum
likelihood, and Bayesian. Finally, software for performing inference using restricted maximum likelihood
and fully Bayesian methods are demonstrated. Methods are illustrated using two examples: an evaluation of
mortality from prophylactic use of lidocaine after a heart attack, and a comparison of length of hospital stay
for stroke patients under two di erent management protocols. Copyright ? 1999 John Wiley & Sons, Ltd.

1. INTRODUCTION
Meta-analysis may be broadly deÿned as the quantitative review and synthesis of the results
of related but independent studies. The objectives of a meta-analysis can be several-fold. By
combining information over di erent studies, an integrated analysis will have more statistical
power to detect a treatment e ect than an analysis based on only one study. For example, Hine
et al.1 conducted a meta-analysis of death rates in randomized controlled trials in which prophy-
lactic lidocaine was administered to patients with proved or suspected acute myocardial infarction.
Table I describes mortality at the end of the assigned treatment period for control and intravenous
lidocaine treatment groups for six studies. The unadjusted total mortality rates were 6·6 per cent
(37=557) in the lidocaine group and 3·8 per cent (21=549) in the control group. The question of
interest is whether there is a detrimental e ect of lidocaine. Because the studies were conducted
to compare rates of arrhythmias following a heart attack, the studies, taken individually, are too
small to detect important di erences in mortality rates.

∗ Correspondence to: Sharon-Lise T. Normand, Department of Health Care Policy, Harvard Medical School, 180 Longwood
Avenue, Boston MA 02115, U.S.A.
Contract=grant sponsor: National Cancer Institute
Contract=grant number: CA-61141

CCC 0277–6715/99/030321–39 $17.50 Received January 1998

Table I. Prophylactic use of lidocaine after a heart attack: evaluating mortality

from prophylactic use of lidocaine in acute myocardial infarction. Source:
reference 1
Source Number randomized Number dead
Lidocaine Control Lidocaine Control

1. Chopra et al. 39 43 2 1
2. Mogensen 44 44 4 4
3. Pitt et al. 107 110 6 4
4. Darby et al. 103 100 7 5
5. Bennett et al 110 106 7 3
6. O’Brian et al. 154 146 11 4
Total 557 549 37 21

Figure 1 displays the individual study mortality risk di erences and corresponding 95 per cent
con dence intervals. Note that the rst ve studies indicate no e ect of lidocaine on mortality; the
sixth and largest study, although not indicating a statistically signi cant e ect, does provide some
evidence of a detrimental e ect of lidocaine.
When several studies have con icting conclusions, a meta-analysis can be used to estimate an
average e ect or to identify a subset of studies associated with a bene cial e ect. For example,
the Cochran Database of Systematic Reviews2 collected information on all trials where specialist
inpatient stroke care was compared to the conventional non-specialist care. Table II lists nine
studies, and for each, the average length of stay (LOS) during the acute hospital admission and
its standard deviation. The central hypothesis of interest is whether specialist stroke unit care will
result in a shorter length of hospitalization compared to routine management. Figure 2 displays
the di erence in the average length of stay for stroke patients managed in specialty units from
the average length of stay for the group managed in the routine manner. Four of the nine studies
(studies 1, 3, 4 and 8) resulted in statistically signi cant shorter stays in the specialty units
compared to those in the general wards.
A systematic approach to synthesizing information can provide estimates of the degree of ben-
e t from a particular therapy and whether the bene t depends upon speci c characteristics of
the studies. This latter question capitalizes on the di erences across studies; it is possible to
test whether there are di erences in the size or direction of the treatment e ect associated with
study-speci c variables. For example, length of stay may be substantially shorter in publicly-
owned hospitals for specialty managed stroke patients than for routinely managed patients but
no di erent in for-pro t hospitals. Moreover, it may be of more interest to nd a particularly
e ective treatment than in determining whether all studies, on average, involve e ective
treatments.
This tutorial is intended for readers with a mathematical statistics background. Meta-analysis in
which the individual study summary statistics are Normal variables arising from two-arm studies
(such as those introduced above) are considered. For moderately large study sizes, the summary
statistics should be asymptotically normally distributed. The situation in which there are multiple
dependent variables measured on each subject requires an additional level of data synthesis. For
example, if each study involves measuring the results of several tests per subject, then in order to
make ecient use of within-subject information, multivariate methods should be utilized. Typically,

Figure 1. Prophylactic lidocaine after a heart

attack. The x-axisdisplays the risk di erence, di = p̂Ti −p̂Ci , and corresponding
p̂Ti qˆTi p̂Ci qˆCi
95 per cent con dence intervals sd2 = nTi
+ nCi
; the y-axis indicates the study and total sample size
i

however, this is not an easy task and space limitations do not permit an adequate discussion of
methods to combine multiple summary statistics across several studies.
In Section 2 issues relating to the de nition of the study objectives, the domain of the liter-
ature search, and the search strategies are reviewed. Section 3 brie y describes how to evaluate
the literature once it is retrieved. Section 4 summarizes several analytic methods for combining
information across studies including the choice of outcome and estimation in xed-e ects and
random-e ects models. Because software accommodating estimation in a random-e ects model
is not streamlined, two packages for conducting inference in a random-e ects model are de-
scribed in Section 5. Section 6 identi es features of the meta-analysis that should be reported.
Finally, methods are illustrated in Section 7 using the lidocaine data and in Section 8 using the
stroke data. Throughout this tutorial, studies comprising the meta-analysis are denoted primary
studies.

Table II. Specialist care for stroke patients from nine studies: comparing specialist
multidisciplinary team care for managing stroke inpatients with routine management
in general medical wards. Source: reference 2

Source Specialist care Routine management

Mean Mean
N LOS SD N LOS SD

1. Edinburgh 155 55·0 47·0 156 75·0 64·0

2. Orpington-Mild 31 27·0 7·0 32 29·0 4·0
3. Orpington-Moderate 75 64·0 17·0 71 119·0 29·0
4. Orpington-Severe 18 66·0 20·0 18 137·0 48·0
5. Montreal-Home 8 14·0 8·0 13 18·0 11·0
6. Montreal-Transfer 57 19·0 7·0 52 18·0 4·0
7. Newcastle 1993 34 52·0 45·0 33 41·0 34·0
8. Umea 1985 110 21·0 16·0 183 31·0 27·0
9. Uppsala 1982 60 30·0 27·0 52 23·0 20·0
Total 548 610

LOS = length of stay measured in days; SD = standard deviation.

Text references on meta-analysis include Cooper and Hedges,3 Hedges and Olkin4 and
Rosenthal.5 Cook et al.6 describe guidelines for reviewing randomized controlled trials. An anal-
ogy is worth bearing in mind throughout: conducting a meta-analysis is conceptually no di erent
than conducting primary research. It is multidisciplinary and therefore requires a research team
comprising several experts: a subject-matter specialist (for example, neurologist, cardiologist); a
biostatistician to help in the design and analytic aspects of the research; an information librarian
to provide guidance regarding printed sources and databases; data coders to translate the data from
the literature into the data base, and a group of subject-matter specialists to aid in judging the
relevance of the retrieved documents.

2. FORMULATING
2.1. Beginning a meta-analysis
As in primary research, a meta-analysis begins with a well-formulated question and design. What
are the study objectives? Is the objective of the study to validate results in a broader population?
For example, is lidocaine prophylaxis during a heart attack related to any mortality e ect? Or is
the goal to guide new studies? For example, which aspects of stroke care are bene cial?
What are the operational de nitions of the research outcome (treatment period mortality or
total mortality), the treatment or intervention (bolus ¿ 50 mg of lidocaine followed by continuous
infusion greater than 1·0 mg=min for at least 24 hours or single-dose therapy), and the population
(patients who have a con rmed heart attack or patients with a suspected heart attack).
What types of designs will be included in the search? Will only randomized trials testing the
research hypothesis be included or will the results from non-experimental studies be permitted?
Will randomized trials with poor compliance be included?
The answers to these questions impact on the methods of review, the modes of statistical
inference, and the interpretation of the results. If interest is centred on making inferences for the
very populations that have been sampled, then the treatment levels are considered xed and the

Figure 2. Stroke example. x-axis displays xTi − xCi , the di erence in average length of stay, measured in days,
1
for each of the nine studies and their corresponding 95 per cent con dence intervals si2 = spi 2 + n1 where
n 2 2
o n Ti Ci
2 = (nTi −1)sTi +(nCi −1)sCi
spi n Ti +nCi −2
. Shorter lengths of stay are assumed to re ect better care management

only source of uncertainty is that resulting from the sampling of people into studies. This type of
variation may be characterized as within-study variation that is a function of the number of patients
in the primary study and the variability in the patient responses within the primary studies. In this
case a xed-e ects (Section 4.2.1) model would be used for statistical inferences. Inferences are
similar to those made when performing an analysis of variance (ANOVA) when there is no inter-
study variation in the mean outcome. The intuition underpinning the xed-e ects model is that
other levels of treatments are suciently like those in the sample of primary studies that inferences
would be the same. The population to which generalizations are to be made consists of a set of
studies having identical characteristics and study e ects. Thus, in the stroke management example,
inferences regarding the reduction in bed days in a hospital based on a xed-e ects model are
applicable to management teams identical to those in the nine primary studies.

On the other hand, if inferences are to be generalized to a population in which the studies are
permitted to have di erent e ects and di erent characteristics, then a random-e ects (Section 4.2.2)
model would be appropriate. The intuition underpinning random-e ects models is that because there
are many di erent approaches to conducting a study by perturbing the design in a small way, then
there are many di erent potential treatment e ects that could arise. This situation corresponds to
an ANOVA model in which there is inter-study variation in the mean outcome in addition to the
within-study variation. Thus, the population in a random-e ects model is the one in which there
are in nitely many possible populations. Inferences regarding the e ect of lidocaine based on a
random-e ects model applies to the population that would be formed if additional studies were
sampled in a manner similar to that used to obtain the six primary studies.
There are conceptual diculties linked to both the xed-e ects and random-e ects points of
view. In both models, it may be dicult to characterize precisely the universe to which we are
inferring. In the random-e ects model, the universe may too big to imagine and in the xed-e ects
model, too small to be of any practical importance. There is a long debate as to the choice of
appropriate model that cannot be adequately covered in this tutorial. What should be noted is
that it is almost always reasonable to believe that there is some between-study variation and few
reasons to believe it is zero. However, if all the lidocaine studies indicate that lidocaine is related
to increased mortality then it may be of little concern that some studies favour it more strongly
than others. On the other hand, when studies con ict, such as in the stroke example, it is dicult
to ignore the between-study variation.

2.2. The domain of the literature search

Once the researcher has established the goals of the meta-analysis, an ambitious literature review
needs to be undertaken, the literature obtained, and then summarized. Sources to be searched
include the published literature, unpublished literature, uncompleted research reports, and work in
progress. The meta-analyst begins with searches of regular bibliographic reports: citation indexes
(for example, the Social Sciences Citation Index) and abstract databases (for example, Mental
Health Abstracts) provide information regarding published reports. These publications are retrieved
and the references therein are searched for more references, the new publications retrieved, and
the process is repeated again and again.
Reliance on only published reports lead to publication bias – the bias resulting from the tendency
to selectively publish results that are statistically signi cant. Study design features, such as small
sample size or failure to randomize, may be positively associated with the bias. As a rst step to-
wards eliminating publication bias, the meta-analyst needs to obtain information from unpublished
research. For example, ERIC (Educational Resources Information Center) is a database that in-
cludes references to unpublished reports and conference papers in addition to published works. The
NTIS (National Technical Information Service) is a bibliographic database summarizing completed
research sponsored by more than 600 U.S. federal agencies; UNIVRES (Directory of Federally
Supported Research in Universities) lists approximately 180,000 university-based research projects
sponsored in Canada.
Clinical research and clinical trials registers are another valued source of information. These
data sources contain information on all initiated studies and are typically maintained by their
funding institution or by groups of individuals with a particular interest in the subject area. For
example, the Neurosurgery Clinical Trials Registry lists information on all completed, active and
planned clinical trials in neurosurgery; the International Registry of Vision Trials is a univer-

sity funded database containing information of completed and active opthamology and optometry
trials.
Unpublished dissertations and master’s theses can also be searched in databases; for example,
the database produced by the University Micro lms International Dissertation Service (Ann Arbor,
MI) contains citations for theses published as early as 1961. Early dissemination of scienti c
results often occurs at conferences and so the meta-analyst must also obtain this literature (for
example, conference indexes such as British Library Lending Division Conference Index).
In summary, the literature search needs to be a well-formulated and co-ordinated e ort involving
several researchers. It is well-advised to seek the guidance of an information scientist to oversee
this aspect of the meta-analysis.

2.3. Quantitative aspects of the search

Although the meta-analyst wants to perform as complete a search as possible, it is clearly not
feasible to obtain every piece of literature that is related to the research topic. Two concepts in
information retrieval can be used to describe the success of the search process: recall and precision.
Recall is de ned as
Number of relevant documents retrieved
Recall = × 100
Total number that should be retrieved

and measures the success of the retrieval process with a higher per cent recall corresponding to a
more successful search. Strictly speaking, however, the denominator is unknown; there is no way
of knowing whether the set of located studies is representative of the full set of existing studies.
More elaborate methods, such as estimation of population size using capture-recapture models7
can be employed.
Precision, the second concept, relates to the false positive rate and is calculated as
Number retrieved and relevant
Precision= × 100:
Number retrieved

The goal is to have high recall and high precision literature retrievals. Increasing recall and pre-
cision can be accomplished by utilizing multiple methods of search strategies. Regardless of how
successful the meta-analyst feels the search was, the meta-analyst needs to assess the presence and
impact of publication bias on subsequent inferences (see Diagnostics in Section 4.4).

2.4. Methods for searching the literature

A manual search of databases requires speci cation of a search statement and a method of search-
ing. Most research libraries add controlled vocabulary terms to the bibliographic record. For ex-
ample, classi cation codes and subject headings for books and descriptors for articles comprise
controlled vocabulary. In contrast, natural language terms are those that may appear in the title
or abstract and are not assigned by the research library. Thus, it is important to combine the two
sets of terms when de ning the search statement, for example
Select State of the Art Review AND Stroke AND Randomized-Controlled-Trial.
There are also di erent methods of searching the literature. A backward search involves iden-
tifying a publication and then moving to earlier items in the citation. A forward search identi es
a publication and then searches all items that later cite the publication. The two methods could

yield di erent publications. The selection principles of the search, such as the scope, including
the subject headings, language constraints and the domain, including sources that failed to yield
studies, should be documented throughout the data synthesis and subsequently reported.

3. EVALUATING THE RETRIEVED LITERATURE

Once the literature has been retrieved, then the study results need to be coded into a database and
the criteria used in accepting or rejecting a study to be meta-analysed need to be decided upon.
Given the vast quantities of heterogeneous literature, the rst task is more daunting than it may rst
appear. For example, what information should be collected and coded, how many coders should
be utilized, how should the coders be trained, and how should they be assessed? The type of items
that should be collected should include the characteristics of: the report (author, year, source of
publication); the study (scope of the sample, types of populations, and overall characteristics of
the study such as the sociodemographic level of the study population); the patients (demographic
and clinical features of the study participants); the research design and features (observational
or randomized, sampling mechanism, treatment assignment mechanism, compliance rates, attrition
rates, type of survey, non-response rates); the treatment (duration, dose, timing, mode of delivery);
and the e ect size (sample size, nature of outcome, estimate and standard error). The amount and
quality of information in any report will depend on several factors not limited to the author’s
professional training and experience as well as his=her writing capabilities and the journal. A well-
designed study to determine the quality of the coding process should also be implemented; see the
article by Sands and Murphy8 describing the inter- and intra-coder reliability in synthesizing the
literature.
Rules for assessing the quality of the studies and determining their relevance to the objective
of the meta-analysis are not clear. The quality and relevance of a study hinges upon the current
state-of-the-art. To this end, Chapter 2 of Cook and Campbell9 provides a framework describing
several features of a study, that if threatened, impact on the interpretation and quality of a primary
study. For example, suppose that there is bias in who received lidocaine and who did not in
study 1 (see Table I) in such a way that sicker patients were given the control treatment. If this
randomization bias occurred, then one would expect mortality di erences to be smaller between
the treatment and control groups if lidocaine was truly detrimental. Such a bias would impact the
internal validity of the study. Alternatively, suppose that the power to detect a di erence between
the treatment and control group is low in study 1; this fact may threaten the statistical validity of
the study. Threats of the extent to which one can generalize to other groups of people or settings
compromise the external validity of a study. For example, can a relationship observed in a private
hospital be generalized to all hospitals?
The approach by Chalmers and his colleagues10 is similar to the framework proposed by Cook
and Campbell but is restricted to focusing on randomized trials. Three areas are assessed in their
framework: study design; implementation, and analysis. Two readers blinded to the authors, source,
results, and discussion of the primary study make ratings regarding the study’s quality. The end
result is a percentage score that may be incorporated into a sensitivity analysis at the analytic
stage. For example, the meta-analyst may want to examine how overall inferences change if the
analysis is restricted to high quality studies.
The validity framework proposed by Cook and Campbell as well as the scoring methods pro-
posed by Chalmers et al. provide a set of criteria for making decisions regarding the inclusion=
exclusion of studies. A formal approach to deciding the ultimate inclusion status of a study may

be undertaken using a panel of judges=experts. For example, in the lidocaine meta-analysis:

The studies were accepted or rejected by two reviewers. A third review was con-
ducted with blinded ‘Methods’ and ‘Results’ sections, and there was agreement on
the acceptability of all studies included and excluded,1 p. 2694.
Note that both methods of quality assessment provide a systematic approach to describe primary
studies, to explain heterogeneity, and to assess sensitivity of results.

4. COMBINING THE STUDIES

Once the primary studies have been collected and coded, the meta-analyst needs to identify a
summary measure common to all studies and subsequently combine the measure. A review of
summary measures based on discrete data, such as risk di erences, and those based on continuous
data, such as standardized mean di erences, follows. The xed-e ects and random-e ects models
are formalized and inferential methods in each are presented. The section concludes with an
introduction to model diagnostics.

4.1. De ning the study outcome

Often the meta-analyst has little control over the choice of the summary measure because most
of the decision is dictated by what was employed in the primary studies. For example, if risk
di erences are reported in the primary studies rather than survival times, then the analyst has little
choice but to utilize the average risk di erence as the summary statistic in the meta-analysis. In
many settings, however, di erent summary measures will be reported across the primary studies.
Success may be de ned as 30-day survival in one study, as symptom-free survival in another,
and as in-hospital survival in yet another. It now becomes the job of the analyst to create a
summary statistic that is comparable across all the studies. In some situations this task will be
impossible. In this section, three classes of outcome measures are described: measures based on
discrete outcome data such as di erences in proportions; those based on continuous data that may
generally be thought of as means, and a miscellaneous set of outcome measures that may be based
on test statistics. The three classes are not exhaustive but are meant to introduce the reader to the
challenges involved in creating a summary statistic comparable across studies.

4.1.1. Risk di erences, relative risks and odds ratios

Table III demonstrates three potential study summary statistics for binary measurements using
the lidocaine data: the di erence between two probabilities (risk di erence), the ratio of two
probabilities (relative risk), and the ratio of the odds for the treated group to the odds for the
control group (odds ratio). Risk di erences are easy to interpret, are de ned for boundary values
(proportions of 0 or 1), and are approximately normally distributed for modest sample sizes.
Relative risks and odds ratios are typically analysed on the logarithmic scale, but, unlike the risk
di erence, are not de ned for boundary values.
Generally inferences in the lidocaine studies remain the same regardless of the choice of sum-
mary statistic. The direction and signi cance of the study-speci c e ects are essentially the same
regardless of the summary statistic selected. Each 95 per cent con dence interval for the risk dif-
ference covers 0 and similarly, the relative risks and odds ratios cover 1. For example, in study 4
with a total sample size of 213 patients, an excess of 1·8 per cent of patients treated with lidocaine

Table III. Study summaries for the lidocaine example. T denotes treatment group, C denotes control
group, qi = 1 − pi ; nTi and nCi denote the total number of treated and control patients, respectively; and
a, b, c, and d denote the number of observations in each of the cells de ned by the treatment (li-
docaine or control) and outcome (dead or alive) table. The con dence intervals for the relative risk
and odds ratio are computed on the logarithmic scale and transformed back to the original scale

Risk di erence Relative risk Odds ratio

PT =(1−PT )
Parameter D = P T − PC R = PT =PC = PC =(1−PC )
p̂ p̂Ti qˆCi
Estimator di = p̂Ti − p̂Ci ri = p̂Ti !i = qˆTi p̂Ci
√ pTi qTi √ qTi √ 1
Ci

pCi qCi qCi 1 1 1

Standard error s di = nTi
+ nCi
s Log(ri ) = nTi pTi
+ nCi pCi
s Log(!i ) = na
+ nb
+ nc
+ nd

Study Sample di (%) 95% CI ri 95% CI !i 95% CI

size

1 82 2·8 (−5·5, 11·1) 2·2 (0·2, 23·4) 2·3 (0·2, 26·1)

2 88 0·0 (−12·0, 12·0) 1·0 (0·3, 3·8) 1·0 (0·2, 4·3)
3 217 2·0 (−3·6, 7·6) 1·5 (0·5, 5·3) 1·6 (0·4, 5·7)
4 213 1·8 (−4·7, 8·3) 1·4 (0·4, 4·1) 1·4 (0·4, 4·5)
5 216 3·5 (−2·0, 9·1) 2·2 (0·6, 8·5) 2·3 (0·6, 9·3)
6 300 4·4 (−0·5, 9·3) 2·6 (0·8, 8·0) 2·7 (0·8, 8·8)

died compared to the control patients with 95 per cent CI for the underlying (unknown) risk dif-
ference ranging from – 4·7 per cent to 8·3 per cent. The relative risk and odds ratio estimates are
both 1·4, implying that a treated patient is 1·4 times as likely as an untreated patient to die, but
both con dence intervals cover 1·0. To compute the approximate 95 per cent con dence interval
for the log relative risk corresponding to the ith study, the lower and upper limits are calculated
1−p T i 1−pCi 1
using log(ri ) ± 1·96si where si is the standard error of the log relative-risk, as ( nTi p T i + nCi pCi ) 2 ,
with p and n denoting the mortality rate and sample size, respectively. In the case of study 4, the
lower and upper 95 per cent limits are
1=2
0·068 0·932 0·950
(lower; upper) = log ± 1·96 +
0·050 103 × 0·068 100 × 0·005
= 0·3075 ± 1·96(0·32307)1=2
= (−0·81; 1·42):
Exponentiating yields a lower limit of e−0·81 = 0·44 and an upper limit of e1·42 = 4·14 for the
relative risk.

4.1.2. Means and e ect sizes

When the primary studies report means, x , in each treatment arm, such as in the stroke data
displayed in Table II, the analyst may calculate the mean di erence and the associated measure
of precision for each study. Let i index study, T the treatment group, C the control group, and

n T i and nCi , the respective sample sizes in the two arms. A potential summary measure is the
di erence in means, Yi = x T i − x Ci with standard error, si , calculated as

1 1 (n T i − 1)s2T i + (nCi − 1)sCi
2
si2 = spi2 + 2
with spi =
nTi nCi n T i + nCi − 2
where s2T i and sCi
2
are the treatment and control group sample variances, respectively, for the
ith study. Figure 2 displays the study means, Yi , and 95 per cent intervals based on si for the
stroke study. In Figure 2, study 3 comprised 146 patients, specialist stroke unit care patients
remained in the hospital, on average, 55 days less than patients managed routinely, with 95 per
cent con dence intervals for the true di erence ranging from – 61 days to – 48 days. In the case
when there is no direct measure common to all the studies, it may be possible to transform the
study-speci c summary to a standardized (scale-free) statistic denoted an e ect size. One common
estimator of e ect size is the standardized mean di erence which is calculated as the di erence of
means divided by the variability of the measures. For example, using N(; 2 ) to denote normally
distributed with mean and variance 2 , if
YijT ∼ N( T ; 2 ); j = 1; 2; : : : ; n T i
YijC C
∼ N( ; );2
j = 1; 2; : : : ; nCi
then the standardized mean di erence is de ned as
T − C
= :

represents the gain (or loss) as the fraction of the variability of the measurements. An estimator
T C
Y −Y
of , denoted Hedges’ g, is de ned as hi = i sp i . The consequences of dividing by an estimate of
the standard deviation is to have a unitless summary measure so that in instances when ‘success’
is measured in di erent ways across the studies, the results from the primary studies can be
transformed to unitless measures and then pooled. The estimated variance of hi is
2
1 1 ˆ
+ + ;
nTi nCi 2(n T i + nCi )
2
where ˆ is the sample estimate of 2 .

4.1.3. Other measures

When the summary data from the primary studies consist of test statistics, then it is sometimes
possible to recover the estimated e ect size if the appropriate pieces of information are also
reported. For example, if the z-statistic is reported, the estimated standardized mean di erence
may be calculated as s

ˆ 1 1
=z + :
nTi nCi

If the study summaries are signi cance levels (p-values) then these may also be combined (Hedges
and Olkin,4 Chapter 3) although this method adds little insight in terms of the size of the e ect and
its direction. The reader is referred to Chapter 2 of Rosenthal5 for a summary of the relationship
between e ect sizes and tests of signi cance.

4.2. Modelling variation in meta-analysis

There are at least three sources of variation to consider before combining summary statistics across
studies. First, sampling error may vary among studies. For example, sample sizes range from 82 to
300 in the lidocaine example and from 21 to 311 in the stroke example resulting in study summaries
estimated with varying degrees of precision. Second, study-level characteristics may di er among
the studies. The stroke studies were conducted at both for-pro t and not-for-pro t hospitals and
there may be reason to believe the treatment e ect is di erent in these two hospital types.
Third, there may exist inter-study variation. The xed-e ects model introduced in Section 4.2.1
assumes each study is measuring the same underlying parameter and that there is no inter-study
variation. Conversely, the random-e ects model (introduced in Section 4.2.2) assumes each study
is associated with a di erent but related parameter.

4.2.1. Fixed-e ects model

A xed-e ects model assumes that each study summary statistic, Yi , is a realization from a popula-
tion of study estimates with common mean (Figure 3). Let be the central parameter of interest
and assume there are i = 1; 2; : : : ; k independent studies. Assume that Yi is such that E(Yi ) = and
let si2 = var(Yi ) be the variance of the summary statistic in the ith study. For moderately large
study sizes, each Yi should be asymptotically normally distributed (by the central limit theorem)
and approximately unbiased. Thus
indep:
Yi ∼ N(; si2 ) for i = 1; 2; : : : ; k (1)
si2
and assumed known. The central parameter of interest is which quanti es the average treatment
e ect.

4.2.2. Random-e ects model

The random-e ects framework postulates that each study summary statistic, Yi , is a draw from a
distribution with a study-speci c mean, i , and variance, si2 :
indep:
Yi | i ; si2 ∼ N(i ; si2 ): (2)
Furthermore, each study-speci c mean, i , is assumed to be a draw from some superpopulation of
e ects (see discussion in Section 2.1) with mean and variance 2 as depicted in Figure 4, with
indep:
i | ; 2 ∼ N(; 2 ): (3)
and 2 are referred to as hyperparameters and represent, respectively, the average treatment
e ect and inter-study variation.
Note that, given the hyperparameters, the distribution of each study summary measure, Yi , after
averaging over the study-speci c e ects, is Normal with mean and variance si2 + 2 . As in
the xed-e ects model, is a parameter of central interest; however, the between-study variation,
2 , plays an important role and must also be estimated. In addition to the average treatment
e ect, it is also possible to derive estimates of the study-speci c e ects, i , that are useful for
inferences regarding identifying particularly e ective studies. The distribution of i , conditional on
the observed data and the hyperparameters, denoted the posterior distribution, is
i | y; ; 2 ∼ N(Bi + (1 − Bi )Yi ; si2 (1 − Bi )) (4)

Figure 3. Fixed-e ects model. The distribution of ve hypothetical study statistics under the assumptions of the xed-e ects
model. Each study sample mean, Yi , provides an estimate of a common mean, (denoted by the dashed vertical line). The
di erence among the ve studies rests only on, si2 , how well each study sample mean estimates

where y = (Y1 ; Y2 ; : : : ; Yk ). Bi , de ned as si2 =(si2 + 2 ), is commonly referred to as the shrinkage

factor for the ith study. The larger the inter-study variation, 2 , the smaller the shrinkage Bi of
the observed study e ects. Because 06Bi 61, the mean of i in equation (4) is a compromise
between the average treatment e ect, , and the observed study summary statistic, Yi . When 2 = 0,
shrinkage is maximized with Bi = 1 so that 1 = 2 = · · · = k = and the random-e ects model
corresponds to the xed-e ects model.

4.3. Inference
In order to account for di erences in sample sizes and study-level characteristics, studies are
strati ed and then combined. That is, rather than estimating the true e ect of lidocaine as the
37 21
di erence between the total fraction dying in the treatment and control groups, 557 − 549 (Table I),

Figure 4. Random-e ects model. The distribution of ve hypothetical study statistics under the assumptions of the ran-
dom-e ects model. Each e ect, i , is drawn from a superpopulation with mean and variance 2 (upper plot). The
study-speci c summary statistics, Yi , are then generated from a distribution with mean determined by i (denoted by ×
in the upper plot) and variance si2 (lower plots). In the example, each of the ve e ects generated the ve study results
(lower plots)

a weighted average of the estimates from each study is taken. Similarly, a weighted average of
the estimates of the treatment e ects from each for-pro t hospital study and a weighted average
of the estimates from the not-for-pro t hospital studies could be taken in the stroke example.
However, the distinction of whether each study (or each set of studies) measures a common pa-
rameter remains. Therefore the convention is to rst perform a test of homogeneity of means.
If no signi cant inter-study variation is found, then a xed-e ects approach is adopted; otherwise
the meta-analyst either adopts a random-e ects approach or identi es study characteristics that
strati es the studies into subsets with homogeneous e ects. The test of homogeneity is next de-
scribed and is followed by a description of inferential modes for a xed-e ects and random-e ects
model. Maximum likelihood, restricted maximum likelihood and Bayesian methods are given for
both types of models and summarized at the conclusion of this section.

4.3.1. Test of homogeneity

The xed-e ects model (equation (1)) assumes that the k study-speci c summary statistics share
a common mean . A statistical test for the homogeneity of study means is equivalent to testing
H0 : = 1 = 2 = · · · = k against

H1 :
At least one i di erent:
Pk P P
Under H0 , for large sample sizes, QW = i Wi (Yi − ˆMLE ) 2 ∼ k−12
where ˆMLE = Wi Yi = Wi
and Wi = 1=si2 . If QW is greater than the 100(1 − ) percentile of the k−12
distribution, then the
hypothesis of equal means, H0 , would be rejected at the 100 per cent level. If H0 is rejected,
the meta-analyst may conclude that the study means arose from two or more distinct populations
and proceed by either attempting to identify covariates that stratify studies into the homogeneous
populations or estimating a random-e ects model. If H0 cannot be rejected the investigator would
conclude that the k studies share a common mean, , and estimate using ˆMLE . Tests of homo-
geneity have low power against the alternative var(i )¿0. Note that not rejecting H0 is equivalent
to asserting that the amount of between-study variation is small.

4.3.2. Fixed-e ects model

When si2 is assumed known, the log-likelihood for , log(L( | y; s 2 )) is proportional to
P (Yi −) 2
i( s2 i
) leading to the maximum likelihood estimator (MLE):

Pk
Wi Yi 1
ˆMLE = Pi=1 with Wi = (5)
k
i=1 Wi
si2

where
P s −1 = (s12 ; s22 ; : : : ; sk2 ). Standard inferences about are available using the fact that ˆMLE ∼ N(;
( i Wi ) ). A Bayesian approach may be adopted by specifying a prior distribution for , for
example, ∼ N(0; 02 ), and calculating the posterior distribution
" #−1 ! " #−1 
X X X
| y; s; 02 ∼ N  Wi + 0−2 Wi Yi ; Wi + 0−2  :
i i i

The estimator of is the posterior mean

" #−1 !
X X
˜B = Wi + 0−2 Wi Yi : (6)
i i

If 02 is large, then the posterior mean coincides with the maximum likelihood estimator.

4.3.3. Random-e ects model

If 2 is known then the MLE of is given by
P
ˆ Wi ()Yi 1
()MLE = Pi with Wi () = 2 : (7)
k Wi () si + 2

However, in the more realistic case of unknown 2 , two common methods of inference can be
employed: restricted maximum likelihood (REML) or Bayesian.

Restricted Maximum Likelihood (REML) This is a method for estimating variance components
in a general linear model.11; 12 Using the marginal distribution for y, the log-likelihood to be
maximized is
( )
X (Y − ˆR ) 2
X
i
log(L(; 2 | s 2 ; y) ∝ log(si2 + 2 ) + 2 + log (si
2
+ 2 −1
) :
i
si + 2

The REML of 2 is the solution to

P
2
i wi ()
ˆ k
k−1 (Yi − ˆR ) 2 − si2
ˆR2 = P :
i wi2 ()
ˆ
The estimator for the population mean is then calculated as
Pk
ˆ wi (ˆR )Yi 1
R = Pi k ; wi (ˆR ) = 2 (8)
i wi (ˆR )
s i + ˆ2R
P
and inferences are made using ˆR ∼ N(; ( i wi (ˆR ))−1 ). An estimator for i can be calculated by
substituting the REML estimates for the hyperparameters in equation (4). This type of approxima-
tion to the posterior distribution is known as empirical Bayes and results in ˆRi = (1− B̂Ri )Yi + B̂Ri ˆR
si2
where B̂Ri = is the shrinkage estimate. Inferences for the study-speci c e ects are made using
si2 +ˆ2
R
ˆRi ∼ N(i ; si2 (1 − B̂Ri )): Models can be estimated using the SAS procedure Proc Mixed (see Section
5). Note that the empirical Bayes approximation is de cient in that it ignores the uncertainty in
the hyperparameters, {; 2 }.

Fully Bayesian In order to re ect the uncertainty in the estimates of hyperparameters and 2
(equation (3)), a fully Bayesian approach can be adopted.13 – 16 Prior distributions on the unknown
parameters are speci ed and inferences about the population e ect (and the i s) can be made by
integrating out the unknown parameters over the joint posterior distribution of all the parameters.
Let ∼ N(0; a 2 ) and −2 ∼ gamma(c;
d) with E(−2 ) = c=d and var(−2 ) = c=d 2 . Then the joint
posterior distribution for V = ; 1 ; : : : ; k ; 2 is calculated as:
Y
p(V | y; s 2 ) ∝ p(i | yi ; si2 )p(i | ; 2 )p()p( 2 ):
i

Inferences are conducted using summaries of the posterior distribution, for example
Z Z
ˆB = E( | y; s 2 ) = {p(V ) di d 2 } d: (9)
i ; 2

The integral in equation (9) may be analytically tractable when the prior and likelihood are con-
jugate. Typically, though, the integral must be evaluated numerically. In cases such as these,
Monte Carlo approximations to the posterior, such as those employed in BUGS (see Section 5),
may be utilized. Other approximations to the posterior distributions are also available. For ex-
ample, Morris17 and Morris and Normand14 proposed an approximation to mean of the posterior

distribution for 2 , denoted the adjusted likelihood estimator derived as a result of applying a
Pearson approximation to the posterior density for 2 .

Method of Moments (MOM) A third estimator of 2 is provided by the homogeneity test.

By equating QW with its corresponding expected value, DerSimonian and Laird18 proposed a
non-iterative (method of moments) estimator of 2 de ned as
 

 
 Q − (k − 1)  
2 W
ˆDL = max 0; P P 2 :

 W 
 Wi − P i  
Wi

This leads to
P
ˆ wi (ˆDL )Yi 1
DL = Pi with wi (ˆDL ) = : (10)
i wi (ˆDL ) si2 2
+ ˆDL

ˆDL is also denoted Cochran’s semi-weighted estimator of and can be easily programmed using
most software packages. Table IV contains a summary of the estimators that we presented above.

4.4. Diagnostics
Once the data have been collected and analyzed the meta-analyst needs to assess the appro-
priateness of the assumptions that have been made. Two aspects of diagnostics are discussed.
A systematic approach to investigating how sensitive the results are to the method of analysis or
to changes in the data, denoted a sensitivity analysis, is next introduced. Methods for assessing
and adjusting the meta-analysis when there is a biased sampling mechanism are also presented.
Note that prior to the analysis of the set of primary studies, several quantitative summaries will
have already been collected. The success of the literature retrieval as measured by the recall and
precision or as estimated using capture-recapture models give an indication as to the representa-
tiveness of the collected literature (Section 2.3). If a panel of raters has been used to determine the
appropriateness of the studies, then inter- and intra-rater reliability statistics are useful measures
of quality; similarly, inter- and intra-coder reliability statistics provide guidance as to the accuracy
of the data underlying the meta-analysis.

4.4.1. Sensitivity analysis

An exploratory analysis of the primary data, for example, the study-speci c estimates, should rst
be undertaken in order to understand important features of the data. For example, a box plot of the
study e ects will indicate typical values, spread (skewness, multi-modal etc.), and tails (presence
of outliers). The box plots may be strati ed by characteristics of the studies (including quality
scores if available) in order to understand how and why studies di er. However, if the number
of studies is small, as in the two examples in this article, then the meta-analyst is limited to the
range of descriptive analyses that can be undertaken.
The meta-analyst should estimate both a xed-e ects and a random-e ects model and compare
the results of both. Sensitivity to the distributional assumptions can be assessed by assuming
di erent distributions for the study e ects and comparing subsequent inferences. For example, the
analyst may assume that the underlying study e ects, i , arise from a Student-t distribution, thereby
permitting heavier tails than those arising from a Normal distribution. Moreover, within a model,

Table IV. Summary of estimators for xed-e ects and random-e ects models. Observed Fisher information
denotes the matrix inverse of the second derivatives of the log-likelihood function evaluated at the REML
estimates. V = {; 1 ; : : : ; k ; 2 } and g(V ) denotes a function of the parameters, for example, P(i ¿0)

Method Parameter Estimator Variance

Fixed-e ects model: Yi ∼ N(; si2 ) P

Wi Yi P
MLE ˆMLE = Pi ( i Wi )−1
W
i i
Wi = 1=si2 assumed known

P P P
Bayesian ˜B = [ i
Wi + 0−2 ]−1 ( i
Wi Yi ) [ i
Wi + 0−2 ]−1

Wi = 1=si2 , 02 assumed known

Random-e ects model Yi | i ∼ N(i ; si2 ); i | ; 2 ∼ N(; 2 )

 

 

 
QW −(k−1)
DerSimonian and Laird 2 ˆ2DL = max 0; P None proposed

 P W2 
 Wi − Pi 
P Wi

w (ˆ )Y P
(Method of moments) ˆDL = i i DL i
P ( i
wi (ˆDL ))−1
wi (ˆDL )
i
Wi = 1=si2 , wi (ˆDL ) = 1
assumed known
si2 +ˆ2DL
P
w2 ()( k (Y −ˆ )2 −s2 )
ˆ k−1 i R
REML 2 ˆ2R = P i i i
Observed Fisher information
2
P i wi ()ˆ
wi (ˆR )Yi P
ˆR = P
i ( i
wi (ˆR ))−1
wi (ˆR )
i
Empirical Bayes i ˆiR = B̂iR ˆR + (1 − B̂iR )Yi si2 (1 − B̂i R )
si2
wi (ˆR ) = 1
; B̂iR = assumed known
si2 +ˆ2R si2 +ˆ2R

R
Bayesian 2 ˆ2B = 2 p̂(V | y; s)di d d2 From empirical distribution
R
ˆB = p̂(V | y; s) di d2 d From empirical distribution
R
i ˆBi = i p̂(V | y; s) dj d d2 di From empirical distribution
R
g(V) ĝ(V ) = g(V )p̂(V | y; s)dV From empirical distribution

Prior distribution for hyperparameters assumed known

the meta-analyst should determine how sensitive the combined estimate is to any one study or
group of studies. This can be accomplished by leaving one study out, calculating the combined
e ect of the remaining studies, and comparing the results with the combined e ect based on all
the studies.

Figure 5. Two funnel plots based on simulated data. 230 e ects simulated assuming i ∼ N(0·02; (0·02)2 );
yi | i ; ni ∼ N(i ; (0·01)2 =ni ) with ni ranging from 10 to 150. The rightmost plot displays those studies statistically sig-
ni cant at 0·05 level

4.4.2. Publication bias

As indicated in previous sections, a major concern of the meta-analyst is whether the collection
of analysed studies were selected using a biased mechanism. One may conjecture several study
features that may be correlated with the propensity with which a study appears in the published
literature. One such feature is the signi cance level associated with the statistical test used in the
primary study. For example, journals are typically more likely to publish results that establish a
di erence than those that do not; moreover, even if a non-signi cant result is published, there is
a tendency to publish few statistical details of such an (apparently) uninteresting result. With this
in mind, a graphic device known as a funnel plot19 can be employed to detect the presence of
publication bias. A funnel plot is a scatter plot of sample size or other measure of precision on
the y-axis versus the estimated e ect size on the x-axis for a group of studies. Because smaller
studies, such as phase I studies, will demonstrate more variability among e ects and will be more
prevalent than larger studies, then the plot should look like a funnel. Publication bias is suspected
if there is a ‘piece’ missing from the plot.
Figure 5 demonstrates, using 230 simulated study results, the impact on the funnel plot when
there is bias against publishing non-signi cant results. The leftmost plot displays the simulated
summaries for all the studies while the rightmost plot displays the simulated summaries for studies
that are statistically signi cant at the 0·05 level. Note that a section of the funnel in the lower
portion of the plot is missing. Unfortunately, detecting bias via a funnel plot is not as obvious as
it is in this contrived example. There may be several types of biasing mechanism present at any
given time; for example, there may be both a bias in publishing results from small studies (even
if signi cant) as well as against publishing non-signi cant results.
If publication bias is suspected, the meta-analyst may model the selection process into the model
in order to correct for the bias.20 – 23 One possibility is to view the selection problem as a missing
data problem and assume that the studies are missing with probabilities that are a function of their
lack of statistical signi cance.23 For example, the second plot in Figure 5 was created using
(
0 if z61·96
pi (z) =
1 if z¿1·96

where pi (z) denotes the probability of publication of the ith study which depends on its z-value.
These proposed methods can be used to provide a broad indication of whether selection bias is
present, and if it is, the impact of the bias on estimation.

5. SOFTWARE
Estimation of xed-e ects models in the context of a meta-analysis is generally straightforward
and can be coded using standard software packages such as S-plus.24 Moreover, there are several
PC-based packages available to perform such analyses (see for example the review by Normand25 ).
Software for estimating random-e ects models are generally not as accessible. This is particularly
true in the case of meta-analysis because the user will typically want to force the package to
make use of the known variances, si2 , in the estimation process. For these reasons, two packages
for performing inference in a random-e ects model are next introduced. The rst package is the
Statistical Analysis System (SAS) and provides estimates using restricted maximum likelihood
estimation. The second package is a more recently developed system called BUGS (Bayesian
Inference Using Gibbs Sampling) and performs inference within a fully Bayesian framework. This
list of software is not exhaustive. For example, DuMouchel26 has developed an S-plus function
to implement a fully Bayesian meta-analysis and is publicly available. Rather, the description that
follows is meant to acquaint the meta-analyst with the computational tactics (and tricks) utilized
when performing inference in the random-e ects model.

5.1. SAS: Proc Mixed

Version 6·11 of SAS provides a procedure denoted Proc Mixed that ts mixed linear models
including variance component models. A restatement of the random-e ects model in terms of a
variance components model is given by the following:

Y = 1k + +e; ∼ Nk (0; 2 Ik )

e ∼ Nk (0; R) where R = diag(s12 ; s22 ; : : : ; sk2 )

where Y is a k-vector of observations from the primary studies, 1k is a k-vector of ones, and Ik
is the k × k identity matrix. It follows that the study-speci c e ect for study i is i = + i . The
general model estimated in Proc Mixed is

Y = X + Z + e;

∼ N(0; G); e ∼ N(0; R)

where is the xed e ect, is the random e ect, and e is the error at the study level. The
procedure permits many di erent parameterizations for R and G including xing G and estimating
the sampling variance, R, but not vice-versa. Unfortunately, in the meta-analysis problem, the
user typically wants to x the sampling variance R and estimate G. However, users may still use
Proc Mixed by reversing the roles of the within-study and between-study speci cations and then
post-processing the SAS output.

Figure 6. SAS model statements in Proc Mixed. Call to procedure to estimate a variance components model in order to
perform a random-e ects meta-analysis. Comments are denoted by a # symbol

5.1.1. An Example
Figure 6 displays the SAS statements used in analysing the stroke data. The data is read in from
the le ‘los.data’, the study outcomes are de ned (di ) as well as the variance of the study
outcomes (vdi ) in the rst data step. Because the procedure does not permit xing G, in order
to specify information regarding the known sampling variances, the user must reverse the roles
of G and R. In other words, let G = diag(s12 ; s22 ; : : : ; sk2 ) and R = 2 Ik . The diagonal elements of G
are next speci ed in the data step using the keywords row,col and value. This is the sparse
representation of G by which the user speci es the value of the entries of the G matrix using the
row and column locations (row, col); all other unspeci ed locations are assumed to be 0. Note

that rather than tting the model

Yi = + i + ei ; i ∼ N(0; 2 ); ei ∼ N(0; si2 ) (11)

the model SAS estimates is

Yi = + ∗i + ei∗ ; ∗i ∼ N(0; si ); ei∗ ∼ N(0; 2 ): (12)

The data set is printed and next the procedure Proc Mixed is called using the data set stroke
and sorted by the order in which they appear in the data set. Class designates which variables
will be used as strati ers or classi ers in the subsequent analyses. The model statement indicates
that the dependent variable is di and is a function of an intercept term only (note the default is
to include an intercept term). The user speci es the ‘p’ option to print the predicted values, for
example, + ∗i , and to print an estimate of the overall mean, , using the ‘s’ option. The random
statement designates the random-e ects and does not include an intercept by default. Thus, ‘styid’
is speci ed as a random-e ect and G is xed by using the gdata option. This option indicates
that the G matrix is to read in from the SAS data set ‘stroke’ and using the keywords col, row,
and value. The ‘s’ option in the random statement requests that the estimated random-e ects, ∗i
be printed. The repeated statement is used to specify the R matrix. In the example, R is speci ed
using the keyword ‘diag’ to indicate that R is a diagonal matrix, R = 2 Ik . By default, REML is
the estimation method for the covariance parameters.

Post-Processing SAS Output to Get Shrinkage Estimates. Because the roles of G and R have
been reversed, in order to obtain the correct study-speci c e ect estimates, i , the user needs to
process the output further. Speci cally, the shrinkage estimate for the ith study is calculated as
i = + i = Yi − ∗i :
In order to calculate this, the user must rst create SAS data sets containing the components listed
in equation (13). The make statements request that an SAS data set containing the predicted values
( + ∗i and its standard error) and a data set comprising the random e ects (∗i and corresponding
standard error) be created. The nal data step merges the original data with the predicted values
and random e ects. The study-speci c estimates (denoted thetai) are then calculated as described
in equation (13) above and the probability that the study-speci c estimates are greater than 0 is
computed using the probnorm function in SAS. Although the SAS Predicted values need to be
manipulated, the standard errors corresponding to the estimates are correct.
Figure 7 displays partial output from the call to Proc Mixed. The data set stroke and the
full SAS output are printed in the Appendix. Three iterations were used before convergence was
reached. The estimate of inter-study variation, 2 , is 685·09 and is printed under the Covariance
Parameter Estimation along with an estimate of the standard error and a Wald test. The p-value
of 0·0713 associated with the Wald test suggests non-homogeneous study means. The estimates of
and {∗i ; i = 1; 2; : : : ; 9} are printed under the Solution for Fixed Effects and Solution
for Random Effects sections, respectively. The combined estimate of di erence in length of
stay, ,
ˆ is −15 days with a standard error of 9 days.
To examine the shrinkage estimates for any particular study, the meta-analyst refers to the post-
processed values under the column labelled THETAI. For example, the estimate corresponding to
Study 4 is de ned as + 4 and is estimated to be −54 days. The probability that this e ect is
positive is 3 per cent (0·02986); alternatively, there is strong evidence (97 per cent) that specialty

Figure 7. Output from Proc Mixed in SAS

Figure 7. Continued

stroke care is associated with shorter length of stay compared to routine management. Examination
of the shrinkage estimate for study 6 indicates that shorter lengths of stays are equally probable
when managed under routine management or specialty care (P GT 0 = 0·54).

5.2. BUGS
BUGS (version 0·501) is a software package written in Modula-2 and distributed as compiled
code.27; 28 The software conducts Bayesian inference using a Monte Carlo Markov chain technique
called Gibbs sampling. The basic sampling approach employed in BUGS is adaptive rejection
sampling using log-concave distributions. The syntax is surprisingly similar to S-plus so many
users will feel comfortable using BUGS.

Figure 8. BUGS code for the lidocaine meta-analysis. Model statements for estimating a random-e ects model in BUGS:
mu represents , the population mean; sigma is 1=2 , the precision of the study-speci c e ects. Comments are proceeded
by a # symbol

5.2.1. An example
Figure 8 contains the code for tting the model
di | i ; si2 ∼ N(i ; si2 )

i | ; 2 ∼ N(; 2 )

∼ N(0; 1e6) and = 1= 2 ∼ gamma(0·001; 0·001)

to the lidocaine data where = pT − pC , i = pTi − pCi , and di = p̂Ti − p̂Ci . The data are in a
matrix format (2 columns, 6 rows) in the le called lidnorm.data. Starting values for ; , and
{i } were set to 0, 1, and {0,0,0,0,0,0}, respectively and are in a le called lidnorm.in.
The model was run in the background using the backbugs command. A portion of the BUGS
output based on 2000 iterates after a burn-in, de ned as an initial run of iterations, of 1000
iterations is displayed in the log- le in Figure 9. The entire log- le is presented in the Appendix.
A 95 per cent credible interval for the combined risk di erence, , is given by (−0·017, 0·073)
with posterior mean estimated as 0·0274. The median of the estimated posterior distribution for
the inter-study variation, 2 , is 0·0011. The study-speci c e ect corresponding to study 2, 2 , is
estimated as 0·021 with standard deviation 0·034.

6. REPORTING
Once the analytic stages of the meta-analysis have been completed, the results must be reported.
A structured abstract summarizing the study objectives, the operational de nitions of the treatment

Figure 9. Portion of the BUGS log le for lidocaine example

and population, the meta-analytic design, the search strategy, the results, and the implications for
clinical practice as well as for clinical research should be developed. Abstracts that appear in The
Journal of the American Medical Association are an excellent guide for summarizing the important
features of a meta-analysis. Within the body of the text, the data collection measures should be
reported in detail as well as the potential biases in the retrieved literature. What was the estimated
size of all possible studies to be collected? What were the recall and precision of the retrieved
literature? A table of the key elements of each primary study should be included and the overall
treatment e ect compared to the e ects reported in the primary studies qualitatively. A graphical
display of the primary data and the estimated pooled e ects must be included in the report. The
clinical signi cance of the statistical results should be clearly stated. Do the results imply that
lidocaine may increase mortality, for example? Moreover, implications of the results for future
research need to be emphasized. The process of data synthesizing permits researchers to identify
areas where more research is needed. Are there hospitals for which there was strong evidence in

decrease length of stay even though the overall estimate covers 0? Finally, the methodological
limitations should be stated.

7. PROPHYLACTIC LIDOCAINE USE IN HEART ATTACKS

The objective of the meta-analysis is to determine whether there is a detrimental e ect of lidocaine
on mortality for hospitalized patients with a con rmed heart attack. The primary data include six
studies and are reported in Table I. To begin, assume that each estimated risk di erence, di , is
di ∼ N(pT − pC ; si2 ); i = 1; 2; : : : ; 6:

7.1. Homogeneity of study means

Pk
A test of equality of means yields QW = i Wi (Yi − Yw )2 ∼ k−1
2 2
= 0·86 and 0·95; 5 = 11·07: The
null hypothesis of homogeneous study means would not be rejected at the 5 per cent level so that
the analyst may conclude that the di erences between the studies are so small that any di erences
are negligible.

7.2. Fixed-e ects results

Table V displays the study-speci c risk di erences, variances, samples sizes and precisions corre-
sponding to a total of 1106 patients. The study with the largest weight, Wi , corresponds to the study
with the largest number of patients (study 6; 300 patients). Using a xed-e ects model, the com-
bined estimate of pT − pC is dW = 2·94 per cent with var(dW ) = 1=5855·5 = 0·0171 per cent. A 95
per cent CI for the risk di erence is (0·4 per cent, 5·5 per cent), indicating an increase in mortality
during the treatment period for lidocaine recipients. Code for calculating the combined estimate
and its corresponding standard error using the S-plus software is contained in the Appendix.

7.3. Random-e ects results

Two random-e ects models were estimated: a model assuming that the distribution of study ef-
fects arise from a Normal distribution and a model in which it was assumed that the study
e ects arise from a Student-t distribution with 4 degrees of freedom. This latter model permitted
the tails of the distribution for the random e ects to be heavier than the former. For example,

Table V. Lidocaine example. Calculations for the xed-e ects model. Wi = 1=si2 is the known
weight in the xed-e ects analysis
Wi Wi ×di
Study di sd2 i ni Wi P P
Wi Wi

1 0·028 0·001778 82 563·1 0·0962 0·002695

2 0·000 0·003757 88 266·2 0·0455 0·000000
3 0·020 0·000813 217 1229·7 0·2100 0·004139
4 0·018 0·001090 203 917·5 0·1567 0·002814
5 0·035 0·000801 216 1248·2 0·2132 0·007532
6 0·044 0·000613 300 1630·7 0·2785 0·01226
Total 1106 5855·5 1·0000 dW = 0·02944

E(pTi − pCi | pT − pC ; 2 ; 4) = pT − pC and var(pTi − pCi | pT − pC ; 2 ; 4) = 22 , twice that of a

Normal distribution. REML estimates using Proc Mixed were obtained for the model that assumed
the study e ects are from a Normal distribution; Bayesian estimates using BUGS were obtained
for both random-e ects models. In the Bayesian analyses, non-informative proper priors for the
population mean and variance were assumed for both random-e ects models. In particular, pT −pC
was given as N(0·0; 1·0 e6 ) and −2 ∼ Gamma(0·001; 0·001). See the article by Smith et al.13 re-
garding speci cation of clinical priors for the population hyperparameters. A burn-in period of
1000 iterations were used and inference conducted using the subsequent 2000 iterates.
Convergence was achieved after 14 iterations using REML estimation in SAS and resulted in an
estimate of 0 for 2 . Because QW is less than the corresponding degrees of freedom, the method
of moments (MOM) estimate of inter-study variation, 2 , is also 0. Consequently, the estimates for
the pooled risk di erence using a xed-e ects model, and using the MOM and REML estimates
random-e ects model are identical.
The Bayesian estimate of 2 (posterior mean) is the same whether a Student-t distribution or
a Normal distribution is assumed for the study-speci c risk di erences. As a rough guide, one
may examine the interval p̂T − p̂C ± 1·96ˆ to determine whether the magnitude of 2 is clinically
important. If the distribution of the true risk di erences is approximately Normal, then one may
expect 95 per cent of the studies to have true risk di erences in the range −3·1 per cent to −2·4
per cent (using the estimates obtain from the Bayesian-t model). Regardless of the model, the
pooled estimate of pT − pC is about 3 per cent. Moreover, even though the Bayesian intervals
for the pooled estimate include 0, the probability that pT − pC ¿0 is 92 per cent. Thus, there is
evidence to believe that there is an increase in mortality for lidocaine recipients.

7.4. Diagnostics
The assumption of normality seems reasonable given the observed sample sizes. The combined
estimate of the risk di erence pooled over the six studies is approximately 3 per cent regardless
of whether a xed-e ects model or a random-e ects model is employed and lies in the range of
observed primary study summaries. Additionally, the results under the random-e ects model do
not change if we hypothesize that the study-speci c e ects arise from a Normal distribution or
from a Student-t distribution.
The pooled estimate appears insensitive to any one study. Each row in Table VI displays
the pooled risk di erence when dropping the corresponding study under a xed-e ects model.
For example, the combined estimate of pT − pC when dropping study 1 and assuming 2 = 0 is
3 per cent. The results were similar when performing the sensitivity analysis using a random-e ects
model.

7.5. Summary
Table VII displays a comparison of the results using xed-e ects and random-e ects models; Figure
10 displays the estimated study-speci c risk di erences and corresponding 95 per cent con dence
intervals based on the primary study data and credible intervals based on the posterior means.
The pooled estimate based on the xed-e ects model appears more precise than that based on the
Bayesian model because the former assumes inter-study variation is 0. Conversely, the individual
study-speci c estimates using the Bayesian model are associated with more precision than the raw
data because some of the within-study variability is allocated to between-study variability in the

Table VI. Sensitivity analysis. Fixed-e ects estimate of

pT − pC appear to be insensitive to exclusion of studies.
The pooled estimate under the xed-e ects model using
all the studies is 2·94 per cent

Lidocaine study
√
Study di dW {var(dW )}

1 0·028 0·030 0·0137

2 0·000 0·031 0·0134
3 0·020 0·032 0·0147
4 0·012 0·032 0·0142
5 0·035 0·028 0·0147
6 0·044 0·023 0·0154

Table VII. Prophylactic lidocaine after MI. Comparison of estimates of the population e ect and
between-study variance using xed-e ects and random-e ects methods

Method Parameter Estimate Variance 95% CI

Fixed-e ects model

MLE 2 Assumed to be 0
p T − pC 2·94% (1·31%)2 (0·4%, 5·5%)
p Ti − pCi 2·94% ∀i (1·31%)2 (0·4%, 5·5%)

Random-e ects model

DerSimonian & Laird 2 0 – –
p T − pC 2·94% (1·31%)2 (0·4%, 5·5%)
REML 2 0 – –
p Ti − pCi ∼ N(p T − pC ; 2 ) p T − pC 2·94% (1·31%)2 (0·4%, 5·5%)
p Ti − pCi 2·94% ∀i (1·31%)2 (0·4%, 5·5%)
Bayesian - (Normal) 2 0·17% (0·24%)2 (0·03%, 0·70%)
p Ti − pCi ∼ N(p T − pC ; 2 ) p T − pC 2·74% (2·21%)2 (−1·72%, 7·29%)
p Ti − pCi See Figure 10

Bayesian - (t) 2 0·18% (0·27%)2 (0·03%, 0·71%)

p Ti − pCi ∼ t(p T − pC ; 2 ; 4) p T − pC 2·75% (2·20%)2 (−1·71%, 6·87%)
p Ti − pCi See Figure 10

random-e ects model. Note that the REML and MOM study-speci c estimates are estimated to be
constant and identical to the pooled xed-e ects estimate.
In conclusion, the results suggest that lidocaine administered in the hospital phase may increase
mortality among patients with a proven heart attack.

Figure 10. Estimated study-speci c and pooled estimates for lidocaine example. Note that ˆ2DL = ˆ2R = 0 implying that
(p̂T − p̂C )MLE = (p̂T − p̂C )DL = (p̂T − p̂C )R . Full Bayes model assumed (I) pTi − pCi ∼ N(p T − pC ; 2 ) (Normal) or (II)
p Ti − pCi ∼ t(pT − pC ; 2 ; 4) (Full Bayes (t)). In either case, proper but vague priors were used for the hyperparameters
(pT − pC ) ∼ N(0; 1·0e6 ); −2 ∼ Gamma(0·001; 0·001)

8. MULTIDISCIPLINARY CARE FOR STROKE INPATIENTS

The objective of this meta-analysis is to determine whether specialist stroke care results in a shorter
length of hospitalization compared to routine management on a general medical ward. The data
consist of nine randomized trials that are listed in Table II. For each of the i = 1; 2; : : : ; 9 studies,
we assume that the di erence in length of stay is

Yi ∼ N(T − C ; si2 ):

Table VIII. Stroke example. Calculations for xed-e ects model. Wi = 1=si2
Wi Wi ×Yi
Study Yi sY2i ni Wi P P
Wi Wi

1 −20 40·59 311 0·02464 0·01507 −0·30131

2 −2 2·05 63 0·48856 0·29873 −0·59745
3 −55 15·28 146 0·06544 0·04001 −2·20080
4 −71 150·22 36 0·00666 0·00407 −0·28890
5 −4 20·19 21 0·04952 0·03028 −0·12113
6 1 1·22 109 0·81731 0·49975 0·49975
7 11 95·38 67 0·01048 0·00641 0·07052
8 −10 8·03 293 0·12450 0·07613 −0·76127
9 7 20·69 112 0·04832 0·02955 0·20684
Total 1158 1·63544 1·0000 Y W =−3·494

8.1. Homogeneity of study means

2
A test of equality of study means yields QW = 241·059. Because 0·95; 8 = 15·5 the null hypothesis
of homogeneous study means would be rejected at the 5 per cent level. This is consistent with
the message displayed in a plot of the primary study summaries (Figure 2).

8.2. Fixed-e ects results

Table VIII displays the study-speci c di erences in length of hospital stay, variances, samples sizes,
and statistics for estimating the xed-e ects pooled estimate. Note that, under a xed-e ects model,
the study receiving the largest weight is study 6 having 109 patients (four studies had a larger
sample size than study 6). The combined estimate of T −C is YW = −3·5 (standard error = 0·78)
days implying that a decrease in hospital length of stay for stroke patients managed in specialty
units compared to stroke patients managed routinely. Code for calculating the combined estimate
and its corresponding standard error using the S-plus software is contained in the Appendix.

8.3. Random-e ects results

As in the lidocaine example, two random-e ects models were estimated: a model assuming that
the distribution of di erences arise from a Normal distribution and a model in which it was
assumed that the di erences arise from a Student-t distribution with 4 degrees of freedom. REML
estimates using Proc Mixed were obtained for the model assuming the study e ects are from
the Normal distribution; Bayesian estimates using BUGS were obtained for both random-e ects
models. In the Bayesian analyses, non-informative proper priors for the population mean and
variance were assumed for both random-e ects models (for example, T − C ∼ N(0·0; 1·0e6 ) and
−2 ∼ gamma(0·001; 0·001)).
Convergence was achieved after three iterations using REML estimation in SAS and resulted
in an estimate of inter-study variation, 2 , of 685 (standard error = 380). Proc Mixed does not
provide an interval estimate for 2 . The MOM estimate of 2 is 219, substantially smaller than the
REML estimate. Furthermore, the pooled estimate of 2 relies on the distributional assumptions
and on the mode of inference within the random-e ects model. The estimate of T − C does vary
with the mode of inference but not as much as the estimate of between-study variance. It appears

Table IX. Sensitivity analysis. Posterior summaries of T − C and 2 when excluding one study at a time.
Estimates based on the model that assumes the study-speci c e ects arise from a Student-t distribution
(Bayesian (t))

Stroke study
Study Yi ˆT − ˆC Posterior standard deviation ˆ2 Posterior standard deviation

1 −20 −9 11 714 1169

2 −2 −13 12 778 1187
3 −55 −5 7 285 554
4 −71 −5 7 259 462
5 −4 −12 12 786 1021
6 1 −13 11 760 1008
7 11 −13 11 693 992
8 −10 −11 12 806 1146
9 7 −14 11 714 1021

Table X. Multidisciplinary care for stroke inpatients. Comparison of estimates for the com-
bined e ect and between-study variance using xed-e ects and random-e ects models

Method Parameter Estimate Variance 95% CI

Fixed-e ects model

MLE 2 Assumed to be 0
T − C −3 days (1)2 (−5, −1)
Ti − Ci −3 days ∀i (1)2 (−5, −1)
Random-e ects model
DerSimonian and Laird 2 219 – –
T − C −14 days (5)2 (−24, −4)
REML 2 685 (380)2 NA
Ti − Ci ∼ N(T − C ; 2 ) T − C −15 days (9)2 (−32, 2)
Ti − Ci See Figure 11
Bayesian (N) 2 892 (619)2 (267, 2591)
Ti − Ci ∼ N(T − C ; 2 ) T − C −15 days (10)2 (−35, 3)
Ti − Ci See Figure 11
Bayesian (t) 2 545 (478)2 (96, 1812)
Ti − Ci ∼ t(T − C ; 2 ; 4) T − C −10 days (9)2 (−29, 7)
Ti − Ci See Figure 11

NA: Con dence intervals are not provided as part of the SAS output

that the hospital stay is between 10 and 15 (±10)days shorter for those patients managed under
specialty care compared to patients managed in the routine manner.

8.4. Diagnostics
Because of lack of homogeneity of means, a xed-e ects model is not appropriate for these data.
Each row in Table IX displays the posterior mean and standard deviation of the pooled e ect
and between-study variance when dropping the corresponding study from the overall analysis. The

Figure 11. Estimated study-speci c and pooled estimates for stroke example. REML estimates converged after
three iterations. Full Bayes model assumed (I) Ti − Ci ∼ N(T − C ; 2 ) (Normal) or (II) Ti − Ci ∼
t(T − C ; 2 ; 4) (Full Bayes (t)). In either case, proper but vague priors were used for the hyperparameters
(T − C ) ∼ N(0; 1·0e6 ); −2 ∼ gamma(0·001; 0·001))

estimated model employed a Student-t distribution for the underlying study-speci c e ects and
used priors as described earlier. Studies 3 and 4 have an impact on both the estimate of T − C
and on the estimate of 2 ; when each of these individual studies are excluded from the meta-
analysis, the magnitude of the pooled di erence is decreased from approximately −11 days to −5
days and the between-study variance is estimated to be about 40 per cent of what it is when all
studies are included. This should come as no surprise given the data displayed in Figure 2.

8.5. Summary
Table X displays a comparison of the results using both xed-e ects and random-e ects models.
Unlike the lidocaine example, the estimates vary widely depending on whether a xed-e ects or a

random-e ects model is assumed. However, it is clear that a xed-e ects model is not supported
by the data.
Figure 11 displays the estimated study-speci c di erences in length of stay and corresponding
95 per cent con dence intervals based on the primary study data and credible intervals based on
the posterior means. Because the estimate of 2 is large compared to the within-study variances,
there is little shrinkage of the posterior means to the prior mean (for example, Bi = 0) so that the
random-e ects estimates of Ti − Ci and corresponding credible intervals are almost identical to
those inferred by using the primary study data alone.
In conclusion, specialty stroke care generally resulted in shorter lengths of stay compared to
routine care. Further investigation is needed to determine which aspects of the specialty care are
associated with shorter length of stay and which are generalizable.

9. SUMMARY
In this tutorial, methods were described for combining information across related but independent
studies when the primary studies report a single summary statistic such as a mean di erence.
Analytic methods were con ned to those cases in which normality of the summary statistic holds.
There are many situations for which combining multivariate statistical summaries will be the
primary focus. The methods outlined in this article, however, present an introduction to meta-
analysis and a general strategy for inference.
Although three potential sources of variation (within-study, systematic between-study variation,
and unexplained between-study variation) were identi ed as important when synthesizing results,
the examples in the tutorial focused primarily on within-study and unexplained between-study
variation. It is important to note that systematic between-study variation, as measured through
xed characteristics of the studies, zi , may exist and can be accounted for in both the xed-e ects
and random-e ects frameworks. In the case of a xed-e ects model, this would be interpreted
to mean that treatment ecacy varies in a systematic fashion between studies characterized by
zi and studies not characterizied by zi . In the random-e ects framework, between-study variation
would be modelled by both study characteristics and random variation. This would be interpreted
to mean that the ecacy within studies characterized by zi is more similar than studies without
zi , and moreover, the ecacies within studies characterized by zi di er because of unexplained
between-study variation.

APPENDIX
The S-plus code for performing a xed-e ects analysis is included below for both examples utilized
in the tutorial. The complete outputs from two estimation procedures for a random-e ects model
are also included in this Appendix. The rst corresponds to the SAS Procedure Proc Mixed for
the stroke study and the second to the BUGS simulation for the lidocaine study.

S-Plus code for the xed-e ects lidocaine meta-analysis

gg_matrix(scan("../examples/lidocaine.data"),ncol=4,byrow=T)
deadtreat_gg[,3]; deadcontrol_gg[,4]
ptreat_gg[,3]/gg[,1]; pcontrol_gg[,4]/gg[,2]
ntreat_gg[,1]; ncontrol_gg[,2]

###COMPUTE RISK DIFFERENCES############

diff_ptreat - pcontrol
vardiff_ptreat*(1-ptreat)/gg[,1] + pcontrol*(1-pcontrol)/gg[,2]
cilower_diff-1.96*sqrt(vardiff); ciupper_diff+1.96*sqrt(vardiff)
## WEIGHTED; FIXED EFFECTS COMBINED ESTIMATE###########
dweights_1/vardiff
dbarweighted_sum(dweights*diff)/sum(dweights) #combined estimate
vardbarweighted_1/sum(dweights) #variance of estimate
print(c(dbarweighted,sqrt(vardbarweighted)))

S-Plus code for the xed-e ects stroke meta-analysis

los_matrix(scan("../examples/los.data"),ncol=6,byrow=T)
ntreated_los[,1]; meantreated_los[,2]; sdtreated_los[,3]
ncontrol_los[,4]; meancontrol_los[,5]; sdcontrol_los[,6]
difflos_meantreated - meancontrol
pooleds_((ntreated-1)*sdtreated^2 + (ncontrol-1)*sdcontrol^2)/(ntreated +
ncontrol -2)
vardifflos_pooleds*(1/ntreated + 1/ncontrol)
cilower_difflos - 1$\cdot$96*sqrt(vardifflos); ciupper_difflos+1$\cdot$96*sqrt
(vardifflos)
## - WEIGHTED; FIXED EFFECTS COMBINED ESTIMATE###########
losweights_1/vardifflos
losbarweighted_sum(losweights*difflos)/sum(losweights) #combined estimate
varlosbarweighted_1/sum(losweights) #variance of estimate
print(c(losbarweighted,sqrt(varlosbarweighted)))

SAS output for the random-e ects stroke meta-analysis

Stroke Study 14:31 Tuesday, April 16, 1996 1
L N
O S C L
N S D O O S
T T T N S D V S V
R R R T C C D D T A
O E E E R O O I I Y R C L
B A A A O N N F S F I O O U
S T T T L T T F P F D W L E
1 155 55 47 156 75 64 -20 3155.55 40.586 1 1 1 40.586
2 31 27 7 32 29 4 -2 32.23 2.047 2 2 2 2.047
3 75 64 17 71 119 29 -55 557.33 15.281 3 3 3 15.281
4 18 66 20 18 137 48 -71 1352.00 150.222 4 4 4 150.222
5 8 14 8 13 18 11 -4 100.00 20.192 5 5 5 20.192
6 57 19 7 52 18 4 1 33.27 1.224 6 6 6 1.224
7 34 52 45 33 41 34 11 1597.18 95.376 7 7 7 95.376
8 110 21 16 183 31 27 -10 551.83 8.032 8 8 8 8.032
9 60 30 27 52 23 20 7 576.46 20.694 9 9 9 20.694
Stroke Study 14:31 Tuesday, April 16, 1996 2

The MIXED Procedure

Class Level Information
Class Levels Values
STYID 9 1 2 3 4 5 6 7 8 9

REML Estimation Iteration History

Iteration Evaluations Objective Criterion
0 1 63.77953195
1 3 63.33581077 0.00054886
2 1 63.31711993 0.00000851
3 1 63.31684770 0.00000000
Convergence criteria met.

Covariance Parameter Estimates (REML)

Cov Parm Estimate Std Error Z Pr > |Z|
DIAG 685.09397633 379.91352734 1.80 0.0713
Residual 1.05643873 . . .

Model Fitting Information for DIFF

Description Value
Observations 9.0000
Variance Estimate 1.0564
Standard Deviation Estimate 1.0278
REML Log Likelihood -39.0099
Akaike’s Information Criterion -40.0099
Schwarz’s Bayesian Criterion -40.0497
-2 REML Log Likelihood 78.0199

Solution for Fixed Effects

Parameter Estimate Std Error DDF T Pr > |T|
INTERCEPT -15.12158889 8.95356233 8 -1.69 0.1297

Solution for Random Effects

Parameter Estimate SE Pred DDF T Pr > |T|
STYID 1 -0.27284292 6.21024468 0 -0.04 .
STYID 2 0.03908618 1.42879248 0 0.03 .
STYID 3 -0.87007274 3.87112874 0 -0.22 .
STYID 4 -10.04910369 11.21602341 0 -0.90 .
STYID 5 0.31841048 4.43620482 0 0.07 .
STYID 6 0.02874070 1.10526162 0 0.03 .
STYID 7 3.19213098 9.21507235 0 0.35 .
STYID 8 0.05934993 2.81953230 0 0.02 .
STYID 9 0.64860109 4.48951554 0 0.14 .
Predicted Values
Observed Predicted Var Pred SE Pred L95M U95M Residual
-20.0000 -15.3944 109.7662 10.4769 . . -4.6056
-2.0000 -15.0825 81.7301 9.0405 . . 13.0825
-55.0000 -15.9917 91.6538 9.5736 . . -39.0083
-71.0000 -25.1707 177.1315 13.3091 . . -45.8293
-4.0000 -14.8032 95.2559 9.7599 . . 10.8032
1.0000 -15.0928 81.1020 9.0057 . . 16.0928
11.0000 -11.9295 145.4908 12.0620 . . 22.9295
-10.0000 -15.0622 86.2581 9.2875 . . 5.0622
7.0000 -14.4730 95.6211 9.7786 . . 21.4730

OBS NTREAT LOSTREAT SDTREAT NCONTROL LOSCONT SDCONT DIFF SP

1 155 55 47 156 75 64 -20 3155.55
2 31 27 7 32 29 4 -2 32.23
3 75 64 17 71 119 29 -55 557.33
4 18 66 20 18 137 48 -71 1352.00
5 8 14 8 13 18 11 -4 100.00
6 57 19 7 52 18 4 1 33.27
7 34 52 45 33 41 34 11 1597.18
8 110 21 16 183 31 27 -10 551.83
9 60 30 27 52 23 20 7 576.46
OBS VDIFF STYID ROW COL VALUE THETAI SE_PRED P_GT_0
1 40.586 1 1 1 40.586 -19.7272 10.4769 0.02986
2 2.047 2 2 2 2.047 -2.0391 9.0405 0.41078
3 15.281 3 3 3 15.281 -54.1299 9.5736 0.00000
4 150.222 4 4 4 150.222 -60.9509 13.3091 0.00000
5 20.192 5 5 5 20.192 -4.3184 9.7599 0.32908
6 1.224 6 6 6 1.224 0.9713 9.0057 0.54294
7 95.376 7 7 7 95.376 7.8079 12.0620 0.74129
8 8.032 8 8 8 8.032 -10.0593 9.2875 0.13938
9 20.694 9 9 9 20.694 6.3514 9.7786 0.74200

BUGS output for the random-e ects lidocaine meta-analysis

Welcome to BUGS on 24 th Sep 1996 at 16:23:8
BUGS : Copyright (c) 1992 .. 1995 MRC Biostatistics Unit.
All rights reserved.
Version 0.501 for SPARC.
For general release : please see documentation for disclaimer.
The support of the Economic and Social Research Council (UK)
is gratefully acknowledged.
Bugs>compile("lidnorm.bug")
model lidnorm;
const
N = 6; #NUMBER OF STUDIES
var
diff[N], vdiff[N], sinv[N], #DECLARING VARIABLES
theta[N], mu, sigma, tau;
data diff, vdiff in "lidnorm.data"; #READ DATA IN FROM FILE
inits in "lidnorm.in"; #STARTING VALUES IN FILE
LIDNORM.IN
{
for (i in 1:N) {
sinv[i] <- 1/vdiff[i]; #TRANSFORM TO GET PRECISIONS
diff[i] ~ dnorm(theta[i],sinv[i]); #DIFF ~ N(THETA)I,S_I^2)
theta[i] ~ dnorm(mu,sigma); #POPULATION MEAN IS MU AND
#PRECISION SIGMA
}
mu ~ dnorm(0.0,1.0e-6); #SPECIFY DISTRIBUTION FOR
sigma ~ dgamma(0.001,0.001); #HYPERPARAMETERS
tau <- 1/(sigma);
}
Parsing model declarations.
Loading data value file(s).
Loading initial value file(s).
Parsing model specification.
Checking model graph for directed cycles.
Generating code.

Generating sampling distributions.

Checking model specification.
Choosing update methods.
compilation took 00:00:00
Bugs>update(1000) time for 1000 updates was 00:00:00
Bugs>monitor(mu)
Bugs>monitor(tau)
Bugs>monitor(sigma)
Bugs>monitor(theta)
Bugs>update(2000) time for 2000 updates was 00:00:01
Bugs>
Bugs>stats(mu)
mean sd 2.5% : 97.5% CI median sample
2.742E-2 2.209E-2 -1.716E-2 7.294E-2 2.691E-2 2000
Bugs>stats(tau)
mean sd 2.5% : 97.5% CI median sample
1.734E-3 2.443E-3 2.745E-4 7.012E-3 1.101E-3 2000
Bugs>stats(theta)
mean sd 2.5% : 97.5% CI median sample
[1] 2.710E-2 2.881E-2 -3.069E-2 8.544E-2 2.656E-2 2000
[2] 2.088E-2 3.434E-2 -4.806E-2 8.696E-2 2.143E-2 2000
[3] 2.263E-2 2.314E-2 -2.132E-2 6.747E-2 2.220E-2 2000
[4] 2.246E-2 2.518E-2 -2.727E-2 7.116E-2 2.284E-2 2000
[5] 3.179E-2 2.264E-2 -1.247E-2 7.577E-2 3.165E-2 2000
[6] 3.725E-2 2.072E-2 -4.153E-3 7.889E-2 3.727E-2 2000
Bugs>stats(sigma)
mean sd 2.5% : 97.5% CI median sample
1.172E+3 9.322E+2 1.417E+2 3.626E+3 9.068E+2 2000
Bugs>q()

ACKNOWLEDGEMENTS

This work was supported by research grant CA-61141 from the National Cancer Institute. I would
like to thank William DuMouchel, AT&T Labs Research, Florham Park, NJ, two anonymous
referees and the Associate Editor for helpful comments on earlier versions of the manuscript.

REFERENCES
1. Hine, L. K., Laird, N., Hewitt, P. and Chalmers, T. C. ‘Meta-analytic evidence against prophylactic use
of lidocaine in Myocardial Infarction’, Archives of Internal Medicine, 149, 2694 –2698 (1989).
2. Cochran Database of Systematic Reviews, 1995.
3. Cooper, H. M. and Hedges, L. V. (eds). The Handbook of Research Synthesis, Russell Sage Foundation,
New York, 1994.
4. Hedges, L. V. and Olkin, I. Statistical Methods for Meta-Analysis, Academic Press, New York, 1985.
5. Rosenthal, R. Meta-Analytic Procedures for Social Research, Sage Publications, Beverly Hills, 1984.
6. Cook, D. J., Sackett, D. L. and Spitzer, W. O. ‘Methodologic guidelines for systematic reviews of
randomized control trials in health care from the Potsdam Consultation on meta-analysis’, Journal of
Clinical Epidemiology, 48, 167–171 (1995).
7. Bishop, Y. V., Fienberg, S. E. and Holland, P. W. Discrete Multivariate Analysis, MIT Press,
Cambridge, MA, 1975, Chapter 6.
8. Sands, M. L. and Murphy, J. R. ‘Use of kappa statistic in determining validity of quality ltering
for meta-analysis: a case study of the health e ects of electromagnetic radiation’, Journal of Clinical
Epidemiology, 49, 1045–1051 (1996).
9. Cook, T. D. and Campbell, D. T. Quasi-Experimentation: Design & Analysis Issues for Field Settings,
Houghton Miin Company, Boston, MA, 1979.

10. Chalmers, T. C, Smith, H. Jr., Blackburn, B., Silverman, B., Schroeder, B., Reitman, D. and Ambroz, A.
‘A method for assessing the quality of a randomized control trial’, Controlled Clinical Trials, 2, 31– 49
(1981).
11. Patterson, H. D. and Thompson, R. ‘Recovery of interblock information when block sizes are unequal’,
Biometrika, 58, 545 –554 (1971).
12. Laird, N. M. and Ware, J. H. ‘Random-e ects models for longitudinal data’, Biometrics, 38, 963– 974
(1982).
13. Smith, T. C., Spiegelhalter, D. J. and Thomas, A. ‘Bayesian approaches to random-e ects meta-analysis:
a comparative study’, Statistics in Medicine, 14, 2685 –2699 (1995).
14. Morris, C. N. and Normand, S. L. ‘Hierarchical models for combining information and for meta-analyses’,
Bayesian Statistics 4, Oxford University Press, 1992, pp. 321–344 (with Discussion).
15. DuMouchel, W.H. ‘Bayesian meta-analysis’ in Statistical Methodology in the Pharmaceutical Sciences,
Marcel Dekker, New York, NY, 1990, pp. 509 –529.
16. DuMouchel, W.H. and Harris, J.E. ‘Bayes methods for combining the results of cancer studies in humans
and other species’, Journal of the American Statistical Association, 78, 293–315 (1983).
17. Morris, C.N. ‘Approximating posterior distributions and posterior moments’, Bayesian Statistics 3,
Oxford University Press, 1988, pp. 327–344 (with Discussion).
18. DerSimonian R. and Laird, N. ‘Meta-Analysis in Clinical Trials’, Controlled Clinical Trials, 7, 177–188
(1986).
19. Light, R. J. and Pillemer, D. B. Summing Up: The Science of Reviewing Research, Harvard University
Press, Boston, MA, 1984.
20. Iyengar, S. and Greenhouse, J. ‘Selection models and the le-drawer problem’ with Discussion,
Statistical Science, 3, 109 –135 (1988).
21. Dear, K. B. and Begg, C. B. ‘An approach for assessing publication bias prior to performing a meta-
analysis’, Statistical Science, 7, 237–245 (1992).
22. Hedges, L. V. ‘Modeling publication selection e ects in meta-analysis’, Statistical Science, 7, 246 –255
(1992).
23. Givens, G. F., Smith, D. D. and Tweedie, R. L. ‘Estimating and adjusting for publication bias using
data augmentation in Bayesian meta-analysis’, Technical Report, Department of Statistics, Colorado State
University, 1995.
24. Statistical Sciences, S-PLUS User’s Manual, Version 3.1 Supplement, Statistical Sciences, Inc., Seattle,
WA, 1992.
25. Normand, S. L. ‘Meta-analysis software: a comparative review’, American Statistician, 49, 298–309
(1995).
26. DuMouchel, W. H. ‘Predictive cross-validation of Bayesian meta-analysis’, in Bayesian Statistics 5,
Oxford University Press, 1996, pp. 107–128, (with Discussion).
27. Spiegelhalter, D. J., Thomas, A. and Best, N. G. ‘Computation on bayesian graphical Models’, in
Bayesian Statistics 5, Oxford University Press, 1996, pp. 407– 425 (with Discussion).
28. Normand, S. L. Discussion of ‘Computation on Bayesian Graphical Models’ by Spiegelhalter, D. J.,
Thomas, A. and Best, N. G. in Bayesian Statistics 5, Oxford University Press, 1996, pp. 420 – 421.

Systematic Review - and - Meta - Analysis - Vetter (2019)
No ratings yet
Systematic Review - and - Meta - Analysis - Vetter (2019)
9 pages
Ethical Considerations in Civic Engagement
80% (5)
Ethical Considerations in Civic Engagement
2 pages
Meta Analysis
100% (1)
Meta Analysis
26 pages
Analysis of Safety Data of Drug Trials An Update Free Ebook Download
100% (19)
Analysis of Safety Data of Drug Trials An Update Free Ebook Download
14 pages
Carton Packaging Knowledge
88% (8)
Carton Packaging Knowledge
93 pages
Martin Bland - An Introduction To Medical Statistics-OUP Oxford (2015) PDF
100% (4)
Martin Bland - An Introduction To Medical Statistics-OUP Oxford (2015) PDF
447 pages
Study Designs: Hani Tamim, MPH, PHD
100% (1)
Study Designs: Hani Tamim, MPH, PHD
38 pages
EPIDEMIOLOGICAL STUDIES Final (1) 1
100% (1)
EPIDEMIOLOGICAL STUDIES Final (1) 1
41 pages
Meta-Analysis in Medicine
100% (2)
Meta-Analysis in Medicine
319 pages
Materi Webinar Sysrev Dan Meta Kedokteran Klinis 2024
No ratings yet
Materi Webinar Sysrev Dan Meta Kedokteran Klinis 2024
72 pages
Meta-Analysis in Stata: History, Progress and Prospects
No ratings yet
Meta-Analysis in Stata: History, Progress and Prospects
54 pages
Evidence Based Medicine Wrap-Up: DR Meirion Evans DR Bethan Stephens
No ratings yet
Evidence Based Medicine Wrap-Up: DR Meirion Evans DR Bethan Stephens
41 pages
Screenshot 2025-01-22 at 5.56.58 PM
No ratings yet
Screenshot 2025-01-22 at 5.56.58 PM
62 pages
Università Di Genova 20230102
No ratings yet
Università Di Genova 20230102
81 pages
Meta-Analysis in Medicine and Health Policy (Chapman & Hall CRC Biostatistics Series) (PDFDrive)
No ratings yet
Meta-Analysis in Medicine and Health Policy (Chapman & Hall CRC Biostatistics Series) (PDFDrive)
418 pages
Meta Analysis PDF
No ratings yet
Meta Analysis PDF
66 pages
The Rough Guide To Systematic Reviews and Meta-Analyses: Review Article
No ratings yet
The Rough Guide To Systematic Reviews and Meta-Analyses: Review Article
13 pages
Warko Karnadihardja, Reno Rudiman: (Critical Appraisal of The Topics)
No ratings yet
Warko Karnadihardja, Reno Rudiman: (Critical Appraisal of The Topics)
53 pages
(Sici) 1097-0258 (19990915 30) 18 17 18 2331 Aid-Sim259 3.0.co 2-l
No ratings yet
(Sici) 1097-0258 (19990915 30) 18 17 18 2331 Aid-Sim259 3.0.co 2-l
11 pages
Ination of Type I Error Rate in Two Statistical Tests For The Detection of Publication Bias in Meta-Analyses With Binary Outcomes
No ratings yet
Ination of Type I Error Rate in Two Statistical Tests For The Detection of Publication Bias in Meta-Analyses With Binary Outcomes
13 pages
L10. Systematic Review
No ratings yet
L10. Systematic Review
122 pages
Meta Analysis
No ratings yet
Meta Analysis
32 pages
Critical Appraisal of The Pharmacology Literature - 2023
No ratings yet
Critical Appraisal of The Pharmacology Literature - 2023
52 pages
Systematic Reviews & Meta-Analysis: Hesham Al-Inany, MD
No ratings yet
Systematic Reviews & Meta-Analysis: Hesham Al-Inany, MD
24 pages
Revision
No ratings yet
Revision
52 pages
Evidence A Based T Medicine
No ratings yet
Evidence A Based T Medicine
28 pages
Systematic Review and Meta-Analysis: How To Conduct
No ratings yet
Systematic Review and Meta-Analysis: How To Conduct
66 pages
Critical Appraisal: Ns. Indriati Kusumaningsih, Mkep., Spkepkom
No ratings yet
Critical Appraisal: Ns. Indriati Kusumaningsih, Mkep., Spkepkom
29 pages
Systematic Review and Meta Analysis1731024542
No ratings yet
Systematic Review and Meta Analysis1731024542
44 pages
Guías Jama Daño (Etiología)
No ratings yet
Guías Jama Daño (Etiología)
10 pages
Data Synthesis and Meta Analysis
No ratings yet
Data Synthesis and Meta Analysis
29 pages
Meta-Analysis: Method For Quantitative Data Synthesis: Martin Bland
No ratings yet
Meta-Analysis: Method For Quantitative Data Synthesis: Martin Bland
14 pages
Interpreting and Understanding Meta-Analysis Graphs
No ratings yet
Interpreting and Understanding Meta-Analysis Graphs
4 pages
Test Questions in Professional Education
No ratings yet
Test Questions in Professional Education
16 pages
Meta-Analysis: Montarat Thavorncharoensap, Ph.D. Faculty of Pharmacy, Mahidol University
No ratings yet
Meta-Analysis: Montarat Thavorncharoensap, Ph.D. Faculty of Pharmacy, Mahidol University
44 pages
Normal Pulse Voltammetry
100% (2)
Normal Pulse Voltammetry
10 pages
Analysis of Safety Data of Drug Trials An Update ISBN 3030058034, 9783030058036 Enhanced Ebook Download
No ratings yet
Analysis of Safety Data of Drug Trials An Update ISBN 3030058034, 9783030058036 Enhanced Ebook Download
16 pages
Research Lecture 3
No ratings yet
Research Lecture 3
12 pages
Part 6
No ratings yet
Part 6
4 pages
Types of Clinical Studies: Noel J. Atienza, MD MS Fch2 SBCM
No ratings yet
Types of Clinical Studies: Noel J. Atienza, MD MS Fch2 SBCM
14 pages
Mbe - Interpreting and Understanding Meta-Analysis Graphs
No ratings yet
Mbe - Interpreting and Understanding Meta-Analysis Graphs
4 pages
Thalluru Tarun Meta-Analysis Summary
No ratings yet
Thalluru Tarun Meta-Analysis Summary
3 pages
MEDTHERA-1.01 Critical Appraisal of Journals
No ratings yet
MEDTHERA-1.01 Critical Appraisal of Journals
5 pages
How To Do Meta Analysis: Arindam Basu
No ratings yet
How To Do Meta Analysis: Arindam Basu
26 pages
Assessments in Occupational Therapy Mental Health An Integrative Approach, 4th Edition Full Digital Edition
100% (15)
Assessments in Occupational Therapy Mental Health An Integrative Approach, 4th Edition Full Digital Edition
16 pages
Journal Club Format
100% (1)
Journal Club Format
3 pages
1-s2 0-S0012369220304487-Main
No ratings yet
1-s2 0-S0012369220304487-Main
6 pages
Handbook of European Homicide Research - Liem & Pridemore (2012)
No ratings yet
Handbook of European Homicide Research - Liem & Pridemore (2012)
508 pages
How To Conduct Meta Analysis
No ratings yet
How To Conduct Meta Analysis
15 pages
A Medical Writer's Guide To Meta-Analysis: Adam Jacobs
No ratings yet
A Medical Writer's Guide To Meta-Analysis: Adam Jacobs
4 pages
An Introduction To Meta-Analysis: Statistics For Researchers
No ratings yet
An Introduction To Meta-Analysis: Statistics For Researchers
8 pages
Decision Making in Medicine: Clinical Decision Using An Article About Harmful Effect
No ratings yet
Decision Making in Medicine: Clinical Decision Using An Article About Harmful Effect
13 pages
CH 18
No ratings yet
CH 18
4 pages
Bluephase G4 - De, En, FR, It, Es, PT
No ratings yet
Bluephase G4 - De, En, FR, It, Es, PT
7 pages
Violence and Mental Health - Lindert Et Al. (2015)
100% (1)
Violence and Mental Health - Lindert Et Al. (2015)
423 pages
Se 221FJ01071
No ratings yet
Se 221FJ01071
3 pages
The Origins and Development of Problem Behavior Theory - Jessor (2016)
No ratings yet
The Origins and Development of Problem Behavior Theory - Jessor (2016)
229 pages
Understanding Research Studies 2010
No ratings yet
Understanding Research Studies 2010
4 pages
Multilevel Modelling For Public Health and Health Services Research - Leyland & Groenewegen (2020)
No ratings yet
Multilevel Modelling For Public Health and Health Services Research - Leyland & Groenewegen (2020)
293 pages
Anthology Poems For Nostalgia
100% (1)
Anthology Poems For Nostalgia
7 pages
Lang Aquisition - Emergent Rubric Original All Criteria
No ratings yet
Lang Aquisition - Emergent Rubric Original All Criteria
4 pages
The Police Mystique - An Insider's Look at Cops, Crime, and The Criminal Justice System - Bouza (1990)
No ratings yet
The Police Mystique - An Insider's Look at Cops, Crime, and The Criminal Justice System - Bouza (1990)
302 pages
Course Module: Course Module Code Topic Coverage Reference/s Duration Learning Outcomes
No ratings yet
Course Module: Course Module Code Topic Coverage Reference/s Duration Learning Outcomes
3 pages
Chapter 8-Performance Management
No ratings yet
Chapter 8-Performance Management
14 pages
Scherfi Gsvej 8, DK-2100 Copenhagen Ø, Denmark Tel.: +45 39 17 17 17. Fax: +45 39 17 18 18. E-Mail: Postmaster@euro - Who.int Web Site: WWW - Euro.who - Int
No ratings yet
Scherfi Gsvej 8, DK-2100 Copenhagen Ø, Denmark Tel.: +45 39 17 17 17. Fax: +45 39 17 18 18. E-Mail: Postmaster@euro - Who.int Web Site: WWW - Euro.who - Int
205 pages
Package Hmisc' - Harrell (2022)
No ratings yet
Package Hmisc' - Harrell (2022)
455 pages
Environmental Law and Jurisprudence
No ratings yet
Environmental Law and Jurisprudence
76 pages
Mitochondrial Disorders Biochemical and Molecular Analysis Methods in Molecular Biology Vol 837 2012th Edition Lee-Jun C. Wong (Editor) Download PDF
100% (2)
Mitochondrial Disorders Biochemical and Molecular Analysis Methods in Molecular Biology Vol 837 2012th Edition Lee-Jun C. Wong (Editor) Download PDF
84 pages
Module Body Fluids For Board Exam
No ratings yet
Module Body Fluids For Board Exam
8 pages
Biological Causes of Delinquency - Rosner & Schwartz (1989)
No ratings yet
Biological Causes of Delinquency - Rosner & Schwartz (1989)
434 pages
Julia de Burgos Biography - Bilingual
No ratings yet
Julia de Burgos Biography - Bilingual
2 pages
Package poLCA - Linzer & Lewis (2022)
No ratings yet
Package poLCA - Linzer & Lewis (2022)
21 pages
Package Lcda' - (Buecker, 2022)
No ratings yet
Package Lcda' - (Buecker, 2022)
19 pages
300 GPD Water Maker
No ratings yet
300 GPD Water Maker
7 pages
Peace and Conflict Studies
No ratings yet
Peace and Conflict Studies
18 pages
Death by Thomas Nagel Com
100% (1)
Death by Thomas Nagel Com
10 pages
Van Bork Et Al. (2022) - A Causal Theory of Error Scores
No ratings yet
Van Bork Et Al. (2022) - A Causal Theory of Error Scores
21 pages
Subject-Verb Agreement: by Resmy R S
No ratings yet
Subject-Verb Agreement: by Resmy R S
19 pages
Cambridge International AS & A Level: Biology 9700/51
No ratings yet
Cambridge International AS & A Level: Biology 9700/51
16 pages
SOW Ransomware Protection Solution v1
No ratings yet
SOW Ransomware Protection Solution v1
11 pages
Mandelbrot Zoom Report
No ratings yet
Mandelbrot Zoom Report
9 pages
Lectura 2 Intimate Violence Avaluacio 3 PDF
No ratings yet
Lectura 2 Intimate Violence Avaluacio 3 PDF
18 pages
Cico Plast-N: Normal Water Reducing Admixture / Plasticiser For Concrete
No ratings yet
Cico Plast-N: Normal Water Reducing Admixture / Plasticiser For Concrete
2 pages
Criminal Careers
No ratings yet
Criminal Careers
21 pages
CSC10004: Data Structures and Algorithms
No ratings yet
CSC10004: Data Structures and Algorithms
20 pages
1 Maxwell's Equations in Matter (Integrate With Next Section)
100% (1)
1 Maxwell's Equations in Matter (Integrate With Next Section)
2 pages
The New 'Peculiar Institution'. On The Prison As Surrogate Ghetto - Wacquant (2000)
No ratings yet
The New 'Peculiar Institution'. On The Prison As Surrogate Ghetto - Wacquant (2000)
13 pages
Clayton (2006) - Diasporic Otherness' Racism, Sectarianism and National Exteriority' in Modern Scotland
No ratings yet
Clayton (2006) - Diasporic Otherness' Racism, Sectarianism and National Exteriority' in Modern Scotland
19 pages
Amaravathi Bye Laws
No ratings yet
Amaravathi Bye Laws
5 pages
T150mm - Beam and Blocks PDF
No ratings yet
T150mm - Beam and Blocks PDF
2 pages
RDMC - Cairo Metro Line-3 Checklist 03-02: Rail - Greasy Status Check Preventive
No ratings yet
RDMC - Cairo Metro Line-3 Checklist 03-02: Rail - Greasy Status Check Preventive
1 page
Uniu S2466 Sti Ii Ul
No ratings yet
Uniu S2466 Sti Ii Ul
1 page
In Vivo and in Vitro Evaluation of Four Different Aqueous Polymeric Dispersions For Producing An Enteric Coated Tablet
No ratings yet
In Vivo and in Vitro Evaluation of Four Different Aqueous Polymeric Dispersions For Producing An Enteric Coated Tablet
6 pages
Arterial hypertension in clinical practice: study and analysis of biotechnological and telemedicine models
From Everand
Arterial hypertension in clinical practice: study and analysis of biotechnological and telemedicine models
Michele Karaboue
No ratings yet
Clinical Examination Skills for Healthcare Professionals
From Everand
Clinical Examination Skills for Healthcare Professionals
Mark Ranson
4.5/5 (16)
Painless Evidence-Based Medicine
From Everand
Painless Evidence-Based Medicine
Antonio L. Dans
2/5 (1)
Computed-Tomography a Powerful Tool for Diagnosis of Pediatric and Adult Congenital Heart Disease: Methodology and Interpretation Guide
From Everand
Computed-Tomography a Powerful Tool for Diagnosis of Pediatric and Adult Congenital Heart Disease: Methodology and Interpretation Guide
Jami G. Shakibi
No ratings yet
Epidemiology for Canadian Students: Principles, Methods and Critical Appraisal
From Everand
Epidemiology for Canadian Students: Principles, Methods and Critical Appraisal
Scott Patten
1/5 (1)
Red Light Revolution: Unlocking the Science, Benefits, and Real-World Applications of Red Light Therapy
From Everand
Red Light Revolution: Unlocking the Science, Benefits, and Real-World Applications of Red Light Therapy
Global Council for Health
No ratings yet
Dosimetry for Radiopharmaceutical Therapy
From Everand
Dosimetry for Radiopharmaceutical Therapy
IAEA
No ratings yet
Neuroscientific based therapy of dysfunctional cognitive overgeneralizations caused by stimulus overload with an "emotionSync" method
From Everand
Neuroscientific based therapy of dysfunctional cognitive overgeneralizations caused by stimulus overload with an "emotionSync" method
Christian Hanisch
No ratings yet
Clinical Trial Management – an Overview
From Everand
Clinical Trial Management – an Overview
Editor IJSMI
No ratings yet
Diagnostic Problems in Tumors of Central Nervous System: Selected Topics
From Everand
Diagnostic Problems in Tumors of Central Nervous System: Selected Topics
Arun Chitale
No ratings yet
The Logical Structure of Clinical Medicine
From Everand
The Logical Structure of Clinical Medicine
Ulrich Müller-Kolck
No ratings yet

Normand (1999) - Meta-Analysis. Formulating, Evaluating, Combining, and Reporting

Uploaded by

Normand (1999) - Meta-Analysis. Formulating, Evaluating, Combining, and Reporting

Uploaded by

STATISTICS IN MEDICINE

Statist. Med. 18, 321–359 (1999)

CCC 0277–6715/99/030321–39 $17.50 Received January 1998

Table I. Prophylactic use of lidocaine after a heart attack: evaluating mortality

Figure 1. Prophylactic lidocaine after a heart

Source Specialist care Routine management

1. Edinburgh 155 55·0 47·0 156 75·0 64·0

LOS = length of stay measured in days; SD = standard deviation.

2.2. The domain of the literature search

2.3. Quantitative aspects of the search

2.4. Methods for searching the literature

3. EVALUATING THE RETRIEVED LITERATURE

be undertaken using a panel of judges=experts. For example, in the lidocaine meta-analysis:

4. COMBINING THE STUDIES

4.1. De ning the study outcome

4.1.1. Risk di erences, relative risks and odds ratios

Risk di erence Relative risk Odds ratio

pCi qCi qCi 1 1 1

Study Sample di (%) 95% CI ri 95% CI !i 95% CI

1 82 2·8 (−5·5, 11·1) 2·2 (0·2, 23·4) 2·3 (0·2, 26·1)

4.1.2. Means and e ect sizes

4.1.3. Other measures

4.2. Modelling variation in meta-analysis

4.2.1. Fixed-e ects model

4.2.2. Random-e ects model

where y = (Y1 ; Y2 ; : : : ; Yk ). Bi , de ned as si2 =(si2 +  2 ), is commonly referred to as the shrinkage

4.3.1. Test of homogeneity

4.3.2. Fixed-e ects model

The estimator of  is the posterior mean

4.3.3. Random-e ects model

The REML of  2 is the solution to

Method of Moments (MOM) A third estimator of  2 is provided by the homogeneity test.

4.4.1. Sensitivity analysis

Method Parameter Estimator Variance

Fixed-e ects model: Yi ∼ N(; si2 ) P

Wi = 1=si2 , 02 assumed known

Random-e ects model Yi | i ∼ N(i ; si2 ); i | ; 2 ∼ N(; 2 )

Prior distribution for hyperparameters assumed known

4.4.2. Publication bias

5.1. SAS: Proc Mixed

Y = 1k +  +e;  ∼ Nk (0;  2 Ik )

e ∼ Nk (0; R) where R = diag(s12 ; s22 ; : : : ; sk2 )

 ∼ N(0; G); e ∼ N(0; R)

that rather than tting the model

Yi =  + i + ei ; i ∼ N(0;  2 ); ei ∼ N(0; si2 ) (11)

the model SAS estimates is

Yi =  + ∗i + ei∗ ; ∗i ∼ N(0; si ); ei∗ ∼ N(0;  2 ): (12)

Figure 7. Output from Proc Mixed in SAS

 ∼ N(0; 1e6) and  = 1= 2 ∼ gamma(0·001; 0·001)

Figure 9. Portion of the BUGS log le for lidocaine example

7. PROPHYLACTIC LIDOCAINE USE IN HEART ATTACKS

7.1. Homogeneity of study means

7.2. Fixed-e ects results

7.3. Random-e ects results

1 0·028 0·001778 82 563·1 0·0962 0·002695

E(pTi − pCi | pT − pC ; 2 ; 4) = pT − pC and var(pTi − pCi | pT − pC ; 2 ; 4) = 22 , twice that of a

Table VI. Sensitivity analysis. Fixed-e ects estimate of

1 0·028 0·030 0·0137

Method Parameter Estimate Variance 95% CI

Fixed-e ects model

Random-e ects model

Bayesian - (t) 2 0·18% (0·27%)2 (0·03%, 0·71%)

8. MULTIDISCIPLINARY CARE FOR STROKE INPATIENTS

1 −20 40·59 311 0·02464 0·01507 −0·30131

8.1. Homogeneity of study means

8.2. Fixed-e ects results

8.3. Random-e ects results

1 −20 −9 11 714 1169

Method Parameter Estimate Variance 95% CI

Fixed-e ects model

S-Plus code for the xed-e ects lidocaine meta-analysis

###COMPUTE RISK DIFFERENCES############

S-Plus code for the xed-e ects stroke meta-analysis

SAS output for the random-e ects stroke meta-analysis

The MIXED Procedure

REML Estimation Iteration History

where y = (Y1 ; Y2 ; : : : ; Yk ). Bi , de ned as si2 =(si2 + 2 ), is commonly referred to as the shrinkage

The estimator of is the posterior mean

The REML of 2 is the solution to

Method of Moments (MOM) A third estimator of 2 is provided by the homogeneity test.

Fixed-e ects model: Yi ∼ N(; si2 ) P

Wi = 1=si2 , 02 assumed known

Random-e ects model Yi | i ∼ N(i ; si2 ); i | ; 2 ∼ N(; 2 )

Y = 1k + +e; ∼ Nk (0; 2 Ik )

∼ N(0; G); e ∼ N(0; R)

Yi = + i + ei ; i ∼ N(0; 2 ); ei ∼ N(0; si2 ) (11)

Yi = + ∗i + ei∗ ; ∗i ∼ N(0; si ); ei∗ ∼ N(0; 2 ): (12)

∼ N(0; 1e6) and = 1= 2 ∼ gamma(0·001; 0·001)

E(pTi − pCi | pT − pC ; 2 ; 4) = pT − pC and var(pTi − pCi | pT − pC ; 2 ; 4) = 22 , twice that of a

Bayesian - (t) 2 0·18% (0·27%)2 (0·03%, 0·71%)