Below are the first 10 and last 10 pages of uncorrected machineread text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapterrepresentative searchable text on the opening pages of each chapter.
Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.
OCR for page 101
9
An Introduction to a Bayesian Method for
MetaAnalysis: *
The Confidence Profile Method
DAVID M. EDDY, VIC HASSELBLAD, and ROSS SHACHTER
The Confidence Profile Method is a form of metaanalysis. It is a Bayesian
method for interpreting, adjusting, and combining evidence to estimate a proba
bility distribution for a parameter. Examples of parameters are health outcomes,
economic outcomes, and variables that might be used in models, such as the
sensitivity of a diagnostic test or the prevalence of a risk factor.
This paper introduces some of the mathematics, indicates the scope of the
method, and gives a few examples of formulas. Additional information can be
found in Eddy (1~; Eddy, Hasselblad, and Shachter (21; and Shachter, Eddy, and
Hasselblad (3~.
BASIC FORMULAS
Let £ be the parameter of interest. Designate as X~ the results of a piece of
evidence about £, say, the results of an experiment. Our task is to estimate the
distribution for £, conditional on the results of the experiment, Xl. Using the
conventional notation for a conditional probability, we denote this distribution
as ~~£ ~ X'). By Bayes's formula, this posterior distribution is calculated as the
product of a prior distribution for £ [which we denote as ~~£~] and the likelihood
function for the experiment.
~~£ ~ Xl) = k L(XI ~ £) ~~)
*This paper was previously published in Medical Decision Making 1990;10:1523.
101
(1)
OCR for page 101
102
DAVID M. EDDY ET AL.
The likelihood function, L(X~ ~ £), gives the likelihood of observing the actual
results of the experiment (X'), conditional on any possible value of the true
effect of the technology (en. "k" is a normalizing constant.
Equation 1 is quite general. A specific example is the formula for analyzing
the effect of a single diagnostic test on the probability that a patient has a dis
ease.
1
P(Disease ~ Test Pos) = P(Test Pos) P(Test Pos ~ Disease) P(Disease)
The "predictive value positive" [P(Disease ~ Test Positive)] corresponds to the
posterior distribution, the sensitivity of the test [P(Test Positive ~ Disease)] cor
responds to the likelihood function, the prior probability of disease [P(Disease)]
corresponds to the prior distribution, and the denominator [P(Test Positive)]
corresponds to the normalizing constant.
Now suppose a second piece of evidence gives results X2. The updated pos
terior distribution for £ that incorporates both pieces of evidence can be calcu
lated by inserting its likelihood function in the equation.
~£ ~ Xl,X2) = k L2(X2 ~ £,X`)L`(X' ~ £) ~£)
(2)
If the experiments are dependent, the likelihood function for the second experi
ment is conditional on the results of the first experiment, as shown in Equation
2. If the experiments are independent, which is very frequently the case, then:
and:
L2 (X2 ~ £,X~) = L2(X2 ~ £)
1~£ ~ Xl,X2) = k L2 (X2 ~ £)L'~X' ~ £) ~£)
Biases
(3)
An important problem in the evaluation of evidence is the presence of bias
es. An important difference between the Confidence Profile Method and other
metaanalysis techniques is the explicit modeling of biases and their incorpora
tion in the distribution for the parameter of interest. Again designate £ as the
parameter of interest. For a variety of reasons, a particular experiment might
estimate a related but slightly different parameter. Call this £' or the "study
parameter." If the study parameter is not identical to the parameter of interest
(i.e., if £ ~ £'), the evidence is biased.
A wide variety of factors can bias an experiment. For example, biases to
internal validity of a twoarm prospective controlled trial include:
· Inaccurate measurement of outcomes
· Incorrect determination of who actually received a technology
· Crossover: some patients who are offered a technology might not receive
OCR for page 101
THE CONFIDENCE PROFILE METHOD
103
it ("dilution") and some patients in the control group might receive it anyway
("contamination")
· Differences in the patients in the two groups ("patientselection bias")
· Loss of patients to followup, and
· Uncertainty about the actual number of cases or outcomes.
Biases to external validity include:
· Differences between the population involved in the experiment and the
population of interest
· Differences between the technology used in the experiment and the tech
nology of interest (e.g., type of equipment, dose of a drug, skill of practitioners)
· Differences in followup times across experiments, and
· Differences in effect measures across experiments.
If biases exist, indiscriminate use of metaanalytic methods that fail to adjust
for them will be incorrect. In the case of the Bayesian approach, if an experi
ment contains biases to internal validity, the likelihood function will apply to £'
rather than £. That is,
~(£ ~ X) ~ L (X ~ ~ ) ~(£)
The Confidence Profile Method can correct for this by defining a function that
relates the study parameter (£') to the parameter of interest (£). Call this func
tion: f(£). This function can be substituted for c' in the likelihood function,
restoring the correctness of Bayes's formula.
~(£ ~ X) = L[X ~ i(~)] ~(£)
This last formula illustrates the three basic ingredients of the Confidence Profile
Method. The method requires prior distributions, likelihood functions, and
functions that describe biases. It also requires functions that define the measures
of effect (which will be introduced below).
Prior Distributions
The most conservative and widely used approach uses noninformative prior
distributions. The choice of a prior distribution then has a minimal effect on the
posterior distribution. Berger (4) has described methods for determining nonin
formative prior distributions, depending on the interval over which the parame
ter of interest is defined. For parameters defined on the entire real line ~ ~ (me,
so) the appropriate prior distribution is () = 1. For parameters defined on the
positive real line, ~ ~ (O. or), the appropriate prior is () = 1/~. For probabili
ties defined on the interval (0,1), the method of Jeffreys (5) gives beta distribu
tion with parameters 1/2, 1/2. For the multinomial model, the comparable prior for
the Hi ~ (0,1) is a Dirichlet distribution with parameters 1/2, 112, . . . 1/2.
OCR for page 101
104
DAVID M. EDDY ET AL.
Likelihood Functions
At the heart of the Confidence Profile Method are likelihood functions. A
different likelihood function is needed for each type of experiment, each type of
outcome, and each type of effect measure. The possible combinations are
shown in Table 9.1.
TABLE 9.1 Likelihood functions for various types of experimental designs,
outcomes, and effect measures
Outcomes
Designs Dichotomous Categorical Count Continuous
OneArm Rate Rate Mean Count Mean Score
Prospective Score Median Score
l~voArm Difference Difference Difference Difference
Prospective Ratio Ratio Ratio Ratio
Odds Ratio
% Difference
eArm
Prospective
Coefficients of Coefficients of Coefficients of Coefficients of
Logistic Linear Linear Linear
Regression Regression Regression Regression
Equation, pi Equation, pi Equation, pi Equation, pi
2x2Case Odds Ratio NA NA NA
Control
2 x n Case Coefficients of NA NA NA
Control Logistic
Regression
Equation, pi
Matched Case Odds Ratio NA NA NA
Control
Cross Coefficients of Coefficients of Coefficients of Coefficients of
Sectional Logistic Linear Linear Linear
Regression Regression Regression Regression
Equation, pi Equation, pi Equation, pi Equation, pi
NA, not applicable.
OCR for page 101
THE CONFIDENCE PROFILE METHOD
TABLE 9.2 Results of a hypothetical randomized controlled clinical trial
105
Study
No. Design No. Survive
Controls
Treated
No. Survive
RCT 100 53 104 72
There are four basic outcomes: dichotomous, categorical, counts, and con
tinuous. There is also a large number of experimental designs, including one
arm prospective trials (e.g., clinical series), twoarm prospective trials (e.g., ran
domized and nonrandomized controlled trials), multiarm prospective trials
(e.g., multidose drug trials), 2 x 2 case control studies, 2 x n case control stud
ies, matched case control studies, and crosssectional studies. Finally, there are
a variety of measures of effect. For example, in a twoarm controlled trial
involving dichotomous outcomes, the effect of the intervention can be measured
as the difference in rates of the outcomes in the two groups, the ratio of rates,
the odds ratio, and the percent difference. For case control studies, the measure
of effect usually is the odds ratio. For multiarm prospective studies, 2 x n case
control studies, and crosssectional studies, the parameters of interest might be
the coefficients of a logistic regression equation, and so forth. The Confidence
Profile Method includes likelihood functions for each type of outcome, experi
mental design, and effect measure (2~.
ILLUSTRATION
Imagine a randomized controlled trial with 100 patients in the control group
and 104 patients in the group offered treatment (see Table 9.21. Imagine that 53
of the patients in the control group survive five years, compared with 72
patients in the treatment group. Suppose we are interested in the probability
that the difference in survival resulted from the treatment. That is, let £ be the
difference in survival rates in the two groups.
To derive the appropriate likelihood function for the difference in survival,
we begin by looking at the outcomes in each group. Let Be be the true survival
rate in the control group, let at be the true survival rate in the treated group, and
let £ be the difference in rates caused by treatment, £ = at  ~C
A joint likelihood function for Be and at based on observing 53 survivors of
100 patients in the control group and 72 survivors of 104 patients in the treated
group can be derived from the binomial distribution.
L(53 of 100, 72 of 104 1 ~c, at) ~ ~C53 (1  )47 ~t72 (1  ~t)32
The probability of success in the control group (8c) is raised to the power of
the observed number of successes in the control group (53), and so forth. Using
OCR for page 101
106
DAVID M. EDDY ETAL.
the definition of £ = fit  ~C' we can solve for fit in tees of £ and Bc, and sub
stitute to obtain a joint likelihood for as and £.
L(53 of 100, 72 of 104 ~ ~c' £) a' ~C53 (1~c)47~£ + ~C)72(l£~C)32
The likelihood function for £ can be obtained by integrating over Bc (6,7),
using a beta distribution with parameters a = 1/2, ~ = 1/2 as a noninformative
prior for tic
L(53 of 100, 72 of 104 1 £)
of; ~C53 (1~c)47~£ + ~C)72(l£~C)32 p1/2,l/2~ec' d`eCy
(4)
This likelihood function can be used in Bayes's formula to calculate a poste
rior distribution for £. The result is illustrated in Figure 9.1.
The horizontal axis shows the range of possible values for £. Because as and
at can each range from O to 1, the range of £, which is at  Bc, is from 1 to +1.
In this case the distribution for £ is centered approximately over 0.16, indicating
that treatment increases the probability of survival by approximately 16 percent.
The uncertainty about that estimate is indicated by the shape of the distribution.
From this distribution it is easy to calculate the probability that the true
effect, £, lies between any set of limits the assessor cares to specify. The distri
bution itself can be used directly in any additional calculations the assessor
cares to perform (e.g., decision trees, mathematical models).
Experiment 1
Face Value
0.5 0 0.5
1
FIGURE 9.1 Probability distribution A for an increase in fiveyear survival as a result of treat
ment. Based on a randomized controlled Dial of 204 patients.
OCR for page 101
THE CONFIDENCE PROFILE METHOD
107
TABLE 9.3 Results of a hypothetical randomized controlled clinical trial win
dilution
Study Controls Treated
No. Design No.
Survive No. Survive Biases
RCT 100
53 104 72 Dilution 20%
Now, suppose there is a bias in this trial. Suppose the best available informa
tion indicates that 20 percent of the patients offered the treatment did not get it.
That is, there is a dilution bias of approximately 20 percent (see Table 9.3~.
If that is true, the likelihood function just derived (Equation 4) no longer esti
mates the parameter of interest, i.e., the effect of treatment in people who actu
ally receive treatment. Rather, the trial estimates a different parameter, £',
which is the effectiveness of offering treatment in the setting of the trial.
L(53 of 100, 72 of 104 ~ £ ~ = ~ ecs3 (1oC)47~£ + oC)72~1  £ ~C)32
pl12,l128C) d(8c)
(5)
This likelihood function cannot be used for £ in Equation 1 without furler work.
To adjust for this dilution, we need a model for how dilution affects the
results of the trial. As before, let 0' be the true probability of survival in people
who actually receive treatment. Let 8'' be the true probability of survival in the
people who are offered treatment in the trial. Finally, let a be the fraction of
people who are offered treatment but do not receive it. In that case, the
probability of survival in patients offered treatment, 8'', is the probability of sur
vival in people who actually receive treatment, 8', multiplied by the proportion
who do receive treatment, (1  a), plus the probability of survival in people who
do not receive treatment, Bc, multiplied by the proportion who do not receive it, a.
8'' = (1  abet + Alec
If the dilution is thought to be 20 percent, set a to 0.2 to obtain a formula for
0~' in terms of 8~ and Bc
8~ = 0.80~ + 0.23
Substituting for 8~' in the formula for the effect measured by the experiment,
£' = 8''  Oc implies that the dilution causes £' to be equal to 0.8£.
The formula for £' can then be substituted in the right side of Equation 5 to
obtain a likelihood function in terms of £, the parameter of interest.
L(53 of 100, 72 of 104 ~ £) = ~ ecs3 (1~C)47~0.8£ + ~C)72(l0.8£~C/2
p~l2,ll2~8c) d(8c) (6)
OCR for page 101
08
Adjusted for Fixed Dilution
DAVID M. EDDY ET AL.
Experiment 1
Face Value
.
0.5 0 0.5
FIGURE 9.2 Probability distribution B for an increase in fiveyear survival as a result of treatment.
Based on a randomized controlled trial of 204 patients in which 20 percent of the patients offered
treatment did not actually receive Reagent (dilution bias of 20 percent).
Use of this "adjusted" likelihood function in Bayes's formula results in a poste
rior distribution that corrects for the bias (see Figure 9.2~. The result is shown
as the solid line in the figure, which includes for comparison the original distri
bution that took the experiment at face value, without adjusting for dilution.
The presence of dilution caused the experiment to underestimate the true effect
of the treatment in patients who actually receive treatment; the best estimate is
now a 20 percent increase in survival for people who receive treatment.
Now suppose we are uncertain about the magnitude of dilution. Suppose all
we can say is that we are 95 percent confident that the proportion of patients
offered treatment who did not receive it (a) is between 6 percent and 42 percent
(see Table 9.41.
TABLE 9.4 Results of a hypothetical randomized controlled clinical trial with
dilution and uncertainty
Study Controls Treated
.
No. Design No. Survive No. Survive Biases
1 RCT 100 53 104 72 Dihedron 20~o
(6  42%)
OCR for page 101
THE CONFIDENCE PROFILE METHOD
109
This uncertainty can be incorporated in the likelihood function by using a
distribution for a [say, ·~3a beady and integrating over that distribution.
L(53 of 100, 72 of 104 1 £) =
11 ~cS3 (1)47(0.8£ + ~C)72(1  0.8£)32 p1l2'll2(8c) pa b(a) do dca
The result is shown in Figure 9.3. The doped line represents the posterior
distribution if the study is taken at face value; the dashed line takes into account
a dilution factor of 0.2; the solid line incorporates uncertainty about the magni
tude of that dilution.
Additional biases and nested biases can be incorporated in the analysis. For
example, in addition to dilution, there might be errors in measurement of out
comes (e.g., there might be a 5 percent probability that a patient in the control
group labeled as dead from the disease actually died of other causes). Or we
might suspect that patients who dilute from the group offered treatment have an
inherently lower risk of the outcome. Or some who dilute might have gotten a
modified treatment that was, say, halfway between the treatment offered the
"treated" and control groups. As in the illustration, it is possible to incorporate
uncertainty about any parameter used to define a bias.
Now consider a second experiment that has 50 patients in the control group
with 23 survivors, and 50 patients in the group offered treatment win 38 sur
vivors (see Table 9.5).
Experiment 1
Face Value
Adjusted for Fixed Dilution
Adjusted for Uncertain Dilution
· ~
· ~
· if.
t
, ~
0.5 0 0.5
FIGURE 93 Probability distribution C for an increase in fiveyear survival as a result of treat
ment. Based on a randomized controlled trial of 204 patients assuming (1) no biases (dotted line),
(2) dilution bias of 20 percent (dashed line), and (3) dilution bias of uncertain magnitude (solid line).
OCR for page 101
110
DAVID M. EDDY ET AL.
TABLE 9.5 Results of two hypothetical randomized controlled clinical trials
Study Controls Treated
No. Design No. Survive No. Survive Biases
RCT 100 53 104 72 Dilution 20%
(6  42%)
RCT 50 23 50 38 None
Suppose there are no biases in this expenment. The likelihood function for
this experiment [L2 (X2 ~ c)] is also based on the binomial distribution and is
derived in the same fashion as for the first experiment (Equation 4~. The results
are indicated in Figure 9.4, which includes for comparison the first shady, after
adjustment for dilution (the dotted line).
Bayes's formula can be used to combine Me information in the two expen
ments to derive a new posterior distribution (Equation 3~. This distribution is
shown as the solid line in Figure 9.5, with the distributions for We two individu
al studies shown as the dashed lines.
Experiment 1
Adjusted for Uncertain Dilution
Experiment 2
Face Value
: I
l
l
61~\
\
· ~
\
\
\
..\
J
1
0.5 0 0.5
FIGURE 9.4 Probability distribution D for an increase in fiveyear survival as a result of treat
ment. Based on a randomized controlled trial of 1,000 patients (solid line), compared with a ran
domized controlled trial of 204 patients adjusted for dilution bias (dotted line).
1
OCR for page 101
THE CONFIDENCE PROFILE METHOD
Experiment 1 n
Adjusted for Uncertain Dilution_  
Experiment 2 l l
Face Value  
Combined Evidence I ~ \~
1' !\ 1
L ~
0.5 0 0.5
111
FIGURE 95 Probability distribution E for an increase in fiveyear survival as a result of treatment.
Based on He combined results of two randomized controlled trials (solid line). The probability dis~i
budons based on die results of Be individual randomized controlled trials are shown as dashed lines.
BASIC FORMULAS IN THE CONFIDENCE PROFILE METHOD
The Confidence Profile Method contains likelihood functions for all the
experimental designs, outcome measures, and effect measures shown in Table
9.1 (21. There is no requirement that all the studies to be combined have the
same design. In general, likelihood functions for studies with dichotomous out
comes are based on the binomial distribution; those with categorical outcomes
are based on the multinomial distribution; those with counts are based on the
Poisson distribution; and those with continuous outcomes are based on the nor
mal distribution. This paper illustrated one likelihood function: a twoarm
prospective study with dichotomous outcomes, whose effect is measured as the
difference in rates of outcomes. The Confidence Profile Method also contains
models for all the biases listed previously (1, 2), one of which (dilution) was illus
trated in this paper. It also incorporates models for compound or nested biases (2~.
ADDITIONAL FORMULAS
The Confidence Profile Method contains a number of formulas for handling
problems that are more complex than the ones just described. These include a
hierarchical Bayes method, formulas for analyzing indirect evidence, and for
mulas for analyzing technology families.
OCR for page 101
112
DAVID M. EDDY ET AL.
Hierarchical Bayes
The hierarchical Bayes method addresses the following problem. Again, let
~ be the true effect in which we are interested. However, it is possible that
Mother Nature does not have a single particular value for this effect. For exam
ple, the success rate of a surgical procedure might be slightly different in New
York than in Chicago, due to factors that we cannot identify or adjust for explic
itly. In such cases, it is reasonable to act as though Mother Nature has a distri
bution for the true effect; our task is to estimate the distribution. The hierarchi
cal Bayes method accomplishes that (8~. An analogous approach using classical
statistical techniques (called the "random effects model") has been described by
DerSimonian and Laird (9~.
Indirect Evidence
The problem posed by indirect evidence is that experiments frequently relate
a technology (e.g., exercise), not to the health outcomes in which we are really
interested (e.g., a heart attack), but to an intermediate outcome (e.g., blood pres
sure, obesity, or serum cholesterol). Another body of evidence must then be
used to relate the intermediate outcomes to health outcomes.
Diagram of indirect evidence:
Technology ~ Intermediate Outcomes ~ Health Outcomes
The Confidence Profile Method includes formulas for combining the two
bodies of evidence, including the possibility that the intermediate outcome is
not a perfect indicator of the health outcome (1~. For example, exercise might
have an independent effect on the chance of a heart attack not mediated through
a change in serum cholesterol.
Technology Families
The formulas for analyzing technology families address another common
problem of technology assessment. Frequently, there are a variety of technolo
gies for the same health problem. For example, breast cancer can be treated
with many different combinations of surgery, radiation, chemotherapy, and hor
monal therapy. A review of the literature might uncover studies that relate
many pairs of technologies, represented as the solid lines in Figure 9.6, but not
all. For example, suppose we are interested in comparing technology B with
technology E, as indicated by the dashed line in Figure 9.6. Even though there
is no direct evidence for this comparison, it is possible to compare these two
technologies using information about other technologies that have been compared.
The Confidence Profile Method contains formulas for accomplishing that (1~.
OCR for page 101
THE CONFIDENCE PROFILE METHOD
technology B )
(Technology Dy
/
~3
113
X /
Am/
/
\
/, \K

\~Technology C)

FIGURE 9.6 Diagram of technology families. Solid lines indicate the existence of trials relating
two technologies; dashed line indicates the two technologies to be compared.
Research Planning
The posterior distribution for the parameter of interest, estimated from exist
ing information, can be used as a prior distribution for calculating the probabili
ty that future experiments of various types (e.g., different designs, different
sample sizes) will yield certain results. The simplest example arises when cal
culating the power of an experiment. Power calculations require postulation of
a particular magnitude of effect; the formulas calculate the probability of a sta
tistically significant result at a specified level of significance, conditional on the
assumed magnitude of the effect. The distribution for the effect calculated by
the Confidence Profile Method can be used in these calculations to obtain a
power conditional on the existing evidence for the effect, rather than a hypothe
sized effect. Because the Confidence Profile Method delivers a distribution, it
can also calculate the probability an experiment will yield results within a speci
fied range (rather than simply a statistically significant result, as in a power cal
culation). For example, the Confidence Profile Method can be used to estimate
the probability that a third randomized controlled trial with 100 patients in each
group will show that treatment increases survival between 15 percent and 25
percent, taking into account the evidence from the first two trials.
Additional techniques in the Confidence Profile Method enable calculation
of the covariance matrix for all parameters incorporated in the analysis. For
OCR for page 101
114
DAVID M. EDDY ET AL.
example, the covariance matrix indicates how a change in the variance of the
distribution for a, the dilution in the first experiment, affects the posterior distri
bution for the parameter of interest. This feature enables calculation of the sen
sitivity of the result to the magnitude and range of uncertainty about any param
eters used in the calculations.
IMPLEMENTATION
To apply the method a problem must be formulated in a way that uses these
ingredients accurately and efficiently, and a solution must be calculated. There
are two basic approaches, which we call the stepwise approach and the integrat
ed approach. The stepwise approach, described in this paper, basically consists
of evaluating one experiment at a time, adjusting each to ensure that it estimates
the parameter of interest, and combining them according to Bayes's formula.
This approach works well for problems that are relatively straightforward. For
more complex assessment problems, the Confidence Profile Method uses an
integrated approach that takes into account the multivariate nature of many
assessment problems, with dependencies between parameters, biases, and pieces
of evidence. The integrated approach is extremely powerful, although more dif
ficult to conceptualize (5~. Both approaches involve considerable mathematics.
We are producing a number of aids to help make the Confidence Profile
Method available. These include a book that pulls all the information together,
with examples; software that implements the stepwise approach; and a comput
erbased, interactive tutorial that will lead a novice through a complete exposi
tion of the method.
RELATIONSHIP TO OTHER METAANALYSIS TECHNIQUES
The Confidence Profile Method differs from metaanalysis techniques based
on classical statistics in several important ways. First, because it is based on
Bayesian statistics, the Confidence Profile Method gives marginal probability
distributions for the parameters of interest and, if the integrated approach is
used, a joint probability distribution for all the parameters. Other metaanalysis
techniques calculate a point estimate for a single effect measure and confidence
intervals for the estimate under an assumption of large sample sizes. The value
of probability distributions is that they can be used to calculate the probability
that the "true value" of a parameter lies within any specified range. Probability
distributions also can be used in models of varying complexity, including sim
ple transformations (e.g., logs, powers), simple operations (e.g., addition, sub
traction by convolution), decision trees, and stochastic models (e.g., Markov
chains).
A second distinguishing feature is that the Confidence Profile Method allows
the assessor to derive probability distributions for parameters that are functions
OCR for page 101
THE CONFIDENCE PROFILE METHOD
115
of other parameters. Classical metaanalysis, as currently formulated, enables
one to combine evidence about a single parameter. For example, the production
of probability distributions enables the Confidence Profile Method to analyze
indirect evidence and technology families, neither of which can be analyzed by
other metaanalysis techniques.
A third distinguishing feature of the Confidence Profile Method, again
enabled by the use of Bayesian statistics, is the explicit modeling of biases to
internal and external validity. Other metaanalysis techniques take biases into
account either by a "take it or leave it" approach, or by assigning weights. In
the latter approach, the assessor assigns each study a weight designed to
decrease its influence compared with the other studies being synthesized. The
main problem with this approach is that weights do not accurately correct for
the effects of biases. Biases cause a piece of evidence to misestimate the mag
nitude and range of uncertainty of a parameter. The use of weights assumes the
study is correctly estimating the magnitude of the parameter; the effect of the
weight is only to modify the variance of the estimate. A second problem with
weights is largely due to the first; there is no theoretical basis for estimating the
appropriate weights to adjust for a specific bias or collection of biases. In the
"take it or leave it" approach, the assessor decides whether to accept a study for
inclusion in a synthesis, which is tantamount to assuming it has no biases, or
decides to reject it, which is tantamount to assuming its biases invalidate its
results. This is equivalent to assigning a weight of either 1 or 0.
In contrast, the Confidence Profile Method models biases explicitly and
incorporates the models in the formulas that synthesize the evidence. These
models allow the assessor to think about each bias individually, in natural units.
For example, an assessor who wants to adjust a randomized controlled trial for
dilution describes the proportion of people who "dilute" who are offered treat
ment but do not receive it. To estimate the effect of possible errors in measure
ment of outcomes (e.g., errors in claims data, chart notes, or patient recall), the
assessor can describe the applicable error rates. The estimates of the magni
tudes of biases can be based on records, separate experiments, or if necessary,
subjective judgments. The Confidence Profile Method also allows for the nest
ing of biases and dependencies between biases. Finally, the method enables the
assessor to describe uncertainty about the magnitude of any bias. Uncertainty
can be present if a bias is estimated empirically, due to the inherent imprecision
of the experiment (e.g., sample size), or if a bias must be estimated subjectively.
The ability of the Confidence Profile Method to incorporate subjective judg
ments about biases is one example of its third important feature, which is to pro
vide a formal, axiomatically based method for incorporating subjective judg
ments in a metaanalysis.
The fourth main difference that distinguishes the Confidence Profile Method
from other metaanalysis methods is that it is a unified set of techniques. The
assessor can describe a system of equations that incorporates simultaneously all
OCR for page 101
116
DAVID M. EDDY ET AL.
the basic parameters (e.g., population parameters), functional parameters
(parameters that are functions of other parameters), experimental evidence, and
subjective judgments. This enables the assessor to represent the multivariate
nature of the assessment problem, taking into account dependencies between
variables and pieces of evidence, and functional relationships as complicated as
the assessor cares to define. The solution of the system of equations yields a
joint probability distribution for all the parameters.
SUMMARY
To summarize, the Confidence Profile Method can be used to assess tech
nologies when the available evidence involves a variety of experimental
designs, types of outcomes, and effect measures; a variety of biases; combina
tions of biases and nested biases; uncertainty about biases; an underlying vari
ability in the parameter of interest; indirect evidence; and technology families.
The result of an analysis with the Confidence Profile Method is a posterior dis
tribution for the parameter of interest, posterior distributions for other parame
ters, and a covariance matrix for all the parameters in the model. The posterior
distributions incorporate all the uncertainty He assessor chooses to describe
about any parameter used in the analysis.
REFERENCES
1. Eddy DM. The Confidence Profile Method: A Bayesian method for assessing
health technologies. Operations Research 1989;37:21~228.
2. Eddy DM, Hasselblad V, Shachter R. A Bayesian method for synthesizing evi
dence: The Confidence Profile Method. International Journal of Technology
Assessment in Health Care, in press.
3. Shachter R. Eddy DM, Hasselblad V. An influence diagram approach to the
Confidence Profile Method for health technology assessment. Technical Report,
Center for Health Policy Research and Education, Duke University, Durham, N.C.,
1988.
4. Berger JO. Statistical Decision Theory and Bayesian Analysis. New York:
SpringerVerlag, 1985.
5. Jeffreys H. Theory of Probability. London: Oxford University Press, 1961.
6. Basu D. On the elimination of nuisance parameters. Journal of the American
Statistical Association 1977;72:355366.
7. Berger J. Wolpert R. The Likelihood Principle (2nd edition). Hayward, Calif.:
Institute of Mathematical Statistics, 1988.
8. Wolpert RL, Hasselblad V, Eddy DM. Hierarchical Bayes methods for confidence
profiles. Technical Report, Center for Health Policy Research and Education, Duke
University, Durham, N.C., 1987.
9. DerSimonian R. Laird NM. Metaanalysis in clinical trials. Controlled Clinical
Trials 1986;7:177188.