calculations - university of sheffield · partial evpi sample sizes, oakley et al. 2 9 summary 10...
TRANSCRIPT
Sample Sizes for Monte Carlo Partial EVPI1
Calculations2
Jeremy E. Oakley
Department of Probability and Statistics, University of Sheffield, UK
Alan Brennan, Paul Tappenden and Jim Chilcott
School of Health and Related Research, University of Sheffield, UK.
3
February 5, 20074
Running title: Partial EVPI sample sizes.5
Address for Correspondence: Dr Jeremy Oakley, Department of Probability and6
Statistics, University of Sheffield, The Hicks Building, Hounsfield Road, Sheffield,7
S3 7RH, UK. E-mail: [email protected]
1
Partial EVPI Sample Sizes, Oakley et al. 2
Summary9
Partial expected value of perfect information (EVPI) quantifies the10
economic value of removing uncertainty concerning parameters of interest in11
a decision model. EVPIs can be computed via Monte Carlo methods. An12
outer loop samples values of the parameters of interest, and for each of13
these, an inner loop samples the remaining parameters from the appropriate14
conditional distribution. This nested Monte Carlo approach can result in15
biased estimates if very small numbers of inner samples are used and can16
require a large number of model runs for accurate partial EVPI estimates.17
We present a simple algorithm to estimate the EVPI bias and EVPI18
confidence interval for a specified number of inner and outer samples. The19
algorithm uses a relatively small number of model runs (we suggest 630), is20
quick to compute, and can help determine how many outer and inner21
iterations to use for a desired level of accuracy. This algorithm is tested22
using two case studies, a simple illustrative cost-effectiveness model with23
easily computed partial EVPIs, and a much more complex health economic24
model of multiple sclerosis. The simple case study results demonstrate that25
our algorithm produces robust estimates of the EVPI bias and EVPI26
confidence interval for different numbers of inner and outer samples. In the27
complex case study, it is established that very large Monte Carlo sample28
sizes would be required, and Monte Carlo estimation would not be feasible29
due to the computing time involved. This algorithm to establish numbers of30
model runs required for partial EVPI estimation will be of use in any31
decision model when parameter uncertainty is important.32
Partial EVPI Sample Sizes, Oakley et al. 3
KEY WORDS: Economic Model, Expected Value of Perfect Information,33
Monte Carlo estimation, Bayesian Decision Theory.34
1 Introduction35
It is common practice to use decision models to estimate the expected net benefit36
of alternative strategy options open to a decision maker [Raiffa, 1968], [Brennan37
and Akehurst, 2000]. Invariably in health economic cost-effectiveness models, the38
input parameters’ true values are not known with certainty. A probabilistic39
sensitivity analysis will then be required to investigate the consequences of this40
input parameter uncertainty [Briggs and Gray, 1999]. Simple Monte Carlo41
propagation of input uncertainty through the model can provide an estimate of the42
distribution of net benefit, thus giving its expected value, and the probability that43
the incremental net benefit will be positive [Van Hout et al., 1994].44
More detailed analysis can compute the partial expected value of perfect45
information (partial EVPI), that is, the value to the decision maker of learning the46
true value of the uncertain parameter input before deciding whether to adopt the47
new treatment [Raiffa, 1968], [Claxton and Posnett, 1996]. Partial EVPIs are48
recommended because they quantify the importance of different parameters using49
decision-theoretic arguments [Claxton, 1999], [Meltzer, 2001], and thus can50
quantify the societal value of further data collection to help design and prioritise51
research projects [Claxton and Thompson, 2001], [Chilcott et al., 2003a].52
Unfortunately, evaluating partial EVPIs is not a trivial computational exercise.53
Partial EVPI Sample Sizes, Oakley et al. 4
Generally, a two level Monte Carlo procedure is needed, which will involve many54
runs of the economic model. A developing literature has described this more and55
more clearly [Felli and Hazen, 1998], [Brennan et al., 2002], [Felli and Hazen,56
2003], [Ades et al., 2004], [Yokota and Thompson, 2004], [Koerkamp et al., 2006],57
[Brennan et al., 2007], [Brennan and Kharroubi, 2007b]. The procedure begins58
with an outer loop sampling values of the parameters of interest, and for each of59
these, an inner loop sampling the remaining parameters from their conditional60
distribution [Brennan et al., 2002] [Ades et al., 2004].61
It is obvious that increasing the inner and outer loop sample sizes will result in62
more accurate estimates of partial EVPI. It is also known that the nested Monte63
Carlo approach can result in biased estimates if the inner loop sample size is small64
[Brennan and Kharroubi, 2007a]. Most studies recommend ‘large enough’ inner65
and outer samples e.g. 1,000 or 10,000 in order to produce ‘accurate’ partial EVPI66
estimates. If the decision model is complex, requiring substantial computation67
time for each model run, then this can become infeasible. One can use instead a68
Gaussian process meta-model, which approximates the economic model, enabling69
more efficient Monte Carlo sampling, but with the disadvantages that it is far70
more complex to program initially and is not always feasible for models with large71
numbers of uncertain input parameters (e.g., in excess of 100) [Oakley, 2003],72
[Oakley and O’Hagan, 2004], [Stevenson et al., 2004]. The fact is however, that73
none of the existing studies discuss in detail how to decide on the number of inner74
and outer samples required, nor their relationship with the scale of any bias in the75
Partial EVPI Sample Sizes, Oakley et al. 5
partial EVPI estimate.76
In this paper we present an algorithm for determining the inner and outer loop77
sample sizes needed to estimate a partial EVPI to a required level of accuracy. We78
first review the theory of estimating partial EVPIs via Monte Carlo. We then79
propose an algorithm, using a moderate number of model runs, to produce an80
estimate of the bias and confidence interval for partial EVPI estimates, and hence81
to help determine the inner and outer sample sizes needed. The results of applying82
and testing the algorithm’s performance in two case studies are given, followed by83
a discussion of how this approach can be applied more generally to support Monte84
Carlo estimation of partial EVPI.85
2 METHODS86
2.1 Evaluating Partial EVPI via Monte Carlo Sampling87
Suppose we have T treatment options, and our economic model computes the net88
benefit NB(t,x) for treatment t = 1, . . . , T , when provided with input parameters89
x. We will denote the true, uncertain values of the input parameters by90
X = {X1, . . . , Xd}, so that the true, uncertain net benefit of treatment t is given91
by NB(t,X). When considering the partial EVPI of a particular parameter Xi, we92
will use the notation X−i = {X1, . . . , Xi−1, Xi+1, . . . , Xd} to denote all the inputs93
in X except Xi. Note that all of the equations presented, and indeed the proposed94
algorithm, are the same if we are considering a group of parameters Xi rather than95
Partial EVPI Sample Sizes, Oakley et al. 6
a single scalar parameter. We use the notation NB(t,X) = NB(t,Xi,X−i) and for96
expectations EX denotes an expectation over the full joint distribution of X, that97
is EX{f(X)} =∫X f(X)dX. The partial EVPI for the ith parameter Xi is given98
by99
EV PI(Xi) = EXi
[max
tEX−i|Xi
{NB(t, Xi,X−i)}]−max
tEX{NB(t,X)}. (1)
The second term in the RHS of (1) can be estimated by Monte Carlo. We sample100
X1, . . . ,XN from the distribution of X, evaluate NB(t,Xn) for t = 1, . . . , T and101
n = 1, . . . , N , and then estimate NB∗ = maxt EX{NB(t,X)} by102
NB∗ = maxt
1
N
N∑
n=1
NB(t,Xn). (2)
We do not consider the choice of N in this paper. In practice, it is usually feasible103
to have N fairly large such that NB∗ is an accurate estimate of NB∗, and can be104
used in the partial EVPI estimate of any parameter or group of parameters.105
In this paper we concentrate on estimating the first term in the RHS of (1), which106
has the form of a maximum of inner expectations nested within an outer107
expectation. We define the maximum of the conditional expected net benefits108
given Xi as:109
m(Xi) = maxt
EX−i|Xi{NB(t,Xi,X−i)} , (3)
and hence, the first term on the RHS of equation (1) can be written as110
EXi{m(Xi)}.111
Typically, we cannot evaluate m(Xi) analytically, and so we estimate it using112
Monte Carlo. We randomly sample J values of X−i from the conditional113
Partial EVPI Sample Sizes, Oakley et al. 7
distribution of X−i|Xi to obtain X−i,1, . . . ,X−i,J . We then run the economic114
model for each of the J sampled inputs to obtain NB(t,Xi,X−i,j) for t = 1, . . . , T115
and j = 1, . . . , J , and estimate m(Xi) by116
m(Xi) = maxt
1
J
J∑
j=1
NB(t,Xi,X−i,j). (4)
We now approximate EXi{m(Xi)} by EXi
{m(Xi)}, and estimate EXi{m(Xi)} by117
randomly sampling K values from the distribution of Xi i.e. Xi,1, . . . , Xi,K , to118
compute the estimator:119
EXi{m(Xi)} =
1
K
K∑
k=1
m{Xi,k}. (5)
We refer to the process of obtaining the maximum conditional net benefit120
estimator m{Xi} for a single given Xi using J samples from the distribution of121
X−i|Xi as the inner level Monte Carlo procedure. The process of calculating (5)122
(given the values m(Xi,1), . . . , m(Xi,K)) using the K samples Xi,1, . . . , Xi,K is the123
outer level Monte Carlo procedure. The full description of our Monte Carlo124
estimate for the partial EVPI for parameter Xi is therefore given by:125
EV PI(Xi) = EXi{m(Xi)} − NB∗ (6)
=1
K
K∑
k=1
max
t
1
J
J∑
j=1
NB(t,Xi,k,X−i,j|k)
− NB∗. (7)
The inner and outer level sample sizes, J and K respectively, mean that the total126
number of runs of the economic model required for the partial EVPI estimate is127
J ×K. The objective in this paper is to determine what values of J and K should128
be used in order to obtain a sufficiently accurate estimate of the partial EVPI.129
Partial EVPI Sample Sizes, Oakley et al. 8
2.2 Uncertainty and Bias in Monte Carlo Estimates of130
Partial EVPI131
Any Monte Carlo estimate of an expectation is subject to uncertainty due to132
random sampling, and the larger the number of samples, the more this uncertainty133
is reduced. We will assume that N is sufficiently large such that uncertainty in134
NB∗ is small relative to uncertainty in EXi{m(Xi)}. If K is sufficiently large then135
a normal approximation will apply, and a 95% confidence interval for EV PI(Xi) is136
given by137
1
K
K∑
k=1
m{Xi,k} −NB∗ ± 1.96
√V ar{m(Xi)}
K. (8)
The Monte Carlo process described earlier provides the first two terms of (8). For138
the final term, we can estimate V ar{m(Xi)} by simply using the sample variance139
of m(Xi,1), . . . , m(Xi,K). We note from equation (4) that m(Xi) and hence the140
variance V ar{m(Xi)} will depend on the value of J . Equation (8) appears to141
suggest that to obtain a sufficiently accurate estimate of the partial EVPI, we just142
need K to be very large, thereby making the 3rd term of (8) negligible.143
Unfortunately, this is not the case, because, as we now show, m(Xi) is an upwards144
biased estimator of m(Xi). The bias is independent of K, the number of outer145
samples, but it does depend on J and can be reduced to an acceptably small level146
if we choose J sufficiently large. We define the conditional expected net benefit147
given a particular Xi for each treatment t as:148
µt(Xi) = EX−i|Xi{NB(t,Xi,X−i)} (9)
Partial EVPI Sample Sizes, Oakley et al. 9
and, its estimator based on J random samples as:149
µt(Xi) =1
J
J∑
j=1
NB(t,Xi,X−i,j). (10)
In our estimate of EVPI in equation (7), we estimate the maximum conditional150
expected net benefit over the T treatments given a particular Xi,151
m(Xi) = max{µ1(Xi), . . . , µT (Xi)} (11)
by the maximum of the Monte Carlo estimates for each treatment152
m(Xi) = max{µ1(Xi), . . . , µT (Xi)}. (12)
The problem is that, although µt(Xi) is an unbiased estimator of µt(Xi), i.e.153
EX−i|Xi{µt(Xi)} = µt(Xi), when the maximisation is applied to the estimators154
then m(Xi) (12) is not an unbiased estimator of m(Xi) (11). In fact, it is upwards155
biased, i.e., it tends to over-estimate m(Xi). This is because, for random variables156
Z1, . . . , Zn it is straightforward to show that157
E{max(Z1, . . . , Zn)} ≥ max{E(Z1), . . . , E(Zn)}. Hence we have that for the158
expectation of the estimator m(Xi)159
EX−i|Xi{m(Xi)} = EX−i|Xi
[max{µ1(Xi), . . . , µT (Xi)}] (13)
≥ max[EX−i|Xi
{µ1(Xi)}, . . . , EX−i|Xi{µT (Xi)}
](14)
= max{µ1(Xi), . . . , µT (Xi)} (15)
= m(Xi). (16)
That is, EX−i|Xi{m(Xi)} ≥ m(Xi). Thus, we expect m(Xi) to overestimate m(Xi)160
for any value of Xi and consequently we expect EV PI(Xi) to be an over-estimate161
Partial EVPI Sample Sizes, Oakley et al. 10
of the true partial EVPI. The bias is the difference between the two terms i.e.162
Bias(Xi) = EX−i|Xi{m(Xi)} −m(Xi). (17)
2.2.1 Example163
We illustrate this with a simple example. Suppose we have two treatments T = 2,164
and165
µ(Xi) ∼ N2
0
0.5
,
1
J
1 0.1
0.1 1
. (18)
In figure 1 we plot the density functions of µ1(Xi) and µ2(Xi) for J = 1 and166
J = 100. When J = 1, there is considerable overlap between the two densities,167
with the result that E[max{µ1(Xi), µ2(Xi)}] (marked as the vertical line) is168
noticeably greater than max[E{µ1(Xi)}, E{µ2(Xi)}] = 0.5. When J = 100, the169
overlap is reduced and E[max{µ1(Xi), µ2(Xi)}] is much closer to 0.5, i.e. the bias170
is negligible.171
2.3 Estimating the bias172
To understand and quantify the bias in the estimator m(Xi) we can investigate the173
sampling distribution (the probability distribution, under repeated sampling of the174
population) of the vector of Monte Carlo estimators for conditional expected net175
benefit µ(Xi) = {µ1(Xi), . . . , µT (Xi)}′. If J is sufficiently large then, then the176
distribution of µ will be approximately multivariate normal:177
µ(Xi) ∼ NT
{µ(Xi),
1
JV (Xi)
}, (19)
Partial EVPI Sample Sizes, Oakley et al. 11
−2 −1.5 −1 −0.5 0 0.5 1 1.5 20
0.2
0.4
0.6
net benefit
J=1
−2 −1.5 −1 −0.5 0 0.5 1 1.5 20
1
2
3
4
net benefit
J=100
Figure 1: The sampling distributions of µ1(Xi) (solid line) and µ2(Xi) (dashed line)
for J = 1 and J = 100 . The vertical line shows the expectation of the maximum of
µ1(Xi) and µ2(Xi) in each case.
where µ(Xi) = {µ1(Xi), . . . , µT (Xi)}′ and element p, q of the matrix V (Xi) is given178
by179
Vp,q(Xi) = CovX−i|Xi{µp(Xi), µq(Xi)} . (20)
Recall that the bias is given by180
EX−i|Xi{max{µ1(Xi), . . . , µT (Xi)}} −max
t{µ1(Xi), . . . , µT (Xi)}, (21)
and so will depend on µ(Xi), V (Xi) and J through the distribution of µ(Xi).181
Note that the bias will decrease as J increases, as increasing J decreases the182
variance of µ(Xi). In the next section we describe an algorithm for estimating183
µ(Xi) and V (Xi) using a moderate number of model runs. This leads to an184
Partial EVPI Sample Sizes, Oakley et al. 12
estimate of the bias in the EVPI estimate for a specified J , and then to a 95%185
confidence interval for EV PI for a specified J and K.186
3 Algorithm to estimate bias and CI in EV PI187
for a specified J and K188
We present the algorithm in three stages. In the first stage we run the economic189
model a moderate number of times (630 is proposed) to obtain estimates of µ(Xi)190
and V (Xi), for different values of Xi. In the second stage, we use these estimates191
to sample from the distribution for the conditional expected net benefit estimators192
µ and estimate the bias in EV PI(Xi) for different specified J ’s. In the third193
stage, we estimate the variance in the maximum conditional expected net benefits194
var{m(Xi)}, enabling us to estimate the confidence interval for EV PI(Xi) using195
equation (8) for different specified K’s.196
Stage 1: Estimating µ(Xi) and V (Xi)197
1. Choose a small number of design points (say 21) Xi,1, . . . , Xi,21 to cover the198
sample space of Xi. If Xi is a scalar, these can be evenly spaced. If if Xi is a199
vector then a Latin Hypercube sample can be used.200
2. For each Xi,k generate a random sample (say 30) of the remaining inputs201
from their conditional distribution X−i|Xi,k. Denote these by202
X−i,k,1, . . . ,X−i,k,30.203
Partial EVPI Sample Sizes, Oakley et al. 13
3. Run the economic model to obtain the net benefits NB(t,Xi,k,X−i,k,j) for204
t = 1, . . . , T ; k = 1, . . . , 21; j = 1, . . . , 30.205
4. Estimate µt(Xi,k) by206
µt(Xi,k) =1
30
30∑
j=1
NB(t,Xi,k,X−i,k,j), (22)
for t = 1, . . . , T ; k = 1, . . . , 21 to obtain207
µ(Xi,k) = {µ1(Xi,k, . . . , µT (Xi,k))} , (23)
for k = 1, . . . , 21.208
5. Estimate V (Xi,k) for k = 1, . . . , 21 by V (Xi,k) where element p, q of V (Xi,k)209
is given by210
1
29
30∑
j=1
{NB(q, Xi,k,X−i,k,j)− µp(Xi,k)} {NB(q,Xi,k,X−i,k,j)− µq(Xi,k)}
(24)
We now have estimates of µ(Xi) and V (Xi) at a range of different values of Xi,211
and can now do a simulation to identify a suitable value of J .212
Stage 2: Estimating bias and determining J213
1. Choose a candidate value of J .214
2. For k = 1, . . . , 21:215
(a) Approximate the distribution of µ(Xi,k) by216
µ(Xi, k) ∼ NT
{µ(Xi,k),
1
JV (Xi,k)
}(25)
Partial EVPI Sample Sizes, Oakley et al. 14
(b) Generate µ1(Xi, k), . . . , µN(Xi, k) for large N (say 10,000) from the217
distribution of µ(Xi, k), with218
µn(Xi,k) = {µ1,n(Xi,k), . . . , µT,n(Xi,k)}′ , (26)
for n = 1, . . . , N .219
(c) Estimate the bias in m(Xi,k) for each of the 21 sampled Xi,k by220
b(Xi,k) =1
N
N∑
n=1
max {µ1,n(Xi,k), . . . , µT,n(Xi,k)}−max {µ1(Xi,k), . . . , µT (Xi,k)}
(27)
3. Estimate the expected bias in m(Xi) by221
b(Xi) =1
21
21∑
k=1
b(Xi,k). (28)
Note that if Xi,1, . . . , Xi,k are chosen deterministically, numerical integration can be222
used to estimate the expected bias, rather than Monte Carlo sampling as in (28).223
This process is repeated for different J in order to determine the relationship224
between J and the scale of bias. When put into the context of the overall EVPI225
and the partial estimate EV PI(Xi), this can help determine a J which produces226
an acceptably small bias.227
Finally, we consider estimating the width of the confidence interval for the228
estimator in (5) that will result from choosing an outer sample size K.229
Stage 3: Estimating the 95% CI for EV PI(Xi) to determine K230
1. Calculate σ2(Xi), the estimated variance of m(Xi,k), by231
σ2(Xi) =1
K − 1
K∑
k=1
{m(Xi,k)− m(Xi)}2 , (29)
Partial EVPI Sample Sizes, Oakley et al. 15
with232
m(Xi) =1
K
K∑
k=1
m(Xi,k). (30)
2. Estimate the width of the 95% confidence interval for EV PI(Xi) as233
±1.96√
σ2(Xi)/K.234
Technically, this gives us the estimated width of the confidence interval when235
J = 30; the variance of m(Xi) will decrease with J , though we would not expect it236
to change substantially relative to σ2(Xi). In practice, once J has been chosen, we237
can re-estimate σ2(Xi).238
The most important purpose of the algorithm is to estimate the bias and239
determine J . This is because the size of the bias cannot be observed when we240
come to actually estimate a partial EVPI using Monte Carlo sampling. Although241
it is useful at stage 3 to estimate the width of the confidence interval for242
EV PI(Xi), it is possible to observe directly the variability in EVPI estimates as243
more outer samples are used in the two-level procedure.244
4 Case studies245
Two case studies have been used to explore the feasibility, accuracy and usefulness246
of the proposed algorithm in predicting the bias and confidence interval in EVPI247
estimates using a specified J and K.248
Partial EVPI Sample Sizes, Oakley et al. 16
4.1 Case study 1: A simple decision tree model249
The first case study concerns a simple hypothetical cost-effectiveness model, which250
compares two strategies: treatment with drug T0 versus treatment with drug T1.251
Table 1 shows the nineteen uncertain model parameters, with prior mean values252
shown for T0 (column a), T1 (column b) and hence the incremental analysis253
(column c). Costs include cost of drug and cost of hospitalisations i.e the product254
of the percentage of patients admitted to hospital, days in hospital, and cost per255
day. Thus, mean cost of strategy T0 = $1000 + 10% x 5.20 x $400 = $1208. Health256
benefits are measured as QALYs gained and come from two sources: responders257
receive a utility improvement for a specified duration, and some patients have side258
effects with a utility decrement for a specified duration i.e. QALY for strategy T0259
= 70% responders x 0.3 x 3 years + 25% side effects x -0.1 x 0.5 years = 0.6175.260
The illustrative willingness to pay i.e. threshold cost per QALY is set at261
λw = $10000. (Note that this is a purely illustrative model, the value of λw is262
arbitrarily chosen and is lower than the threshold of many western countries).263
Thus, the net benefit of T0 is = $10000 x 0.6175 - $1208 = $4967. Effectively this264
is a simple decision tree model with a net benefit function of sum-product form i.e.265
NB(T0) = λw(θ5θ6θ7 + θ8θ9θ10)− (θ1 + θ2θ3θ4)266
NB(T1) = λw(θ14θ15θ16 + θ17θ18θ19)− (θ11 + θ12θ13θ4)267
Uncertainty about the parameters is described with independent normal268
distributions. Standard deviations for the model parameters are shown in columns269
(d) and (e). Each parameter can be informed by collection of further data on270
Partial EVPI Sample Sizes, Oakley et al. 17
individual patients. Given current knowledge, the basic model results show $5405271
expected net benefit for T1 compared with $4967 for T0 (difference = $437.80),272
which means that our baseline decision given current information should be to273
adopt strategy T1. Probabilistic sensitivity analysis [Briggs and Gray, 1999] shows274
that T1 provides greater net benefits than T0 on only 54.5% of 1000 Monte Carlo275
samples. This suggests that knowing more about some of the uncertain parameters276
could affect our decision. The results of the partial EVPI analyses we have277
undertaken are described on an indexed scale, where the overall expected value of278
perfect information, $1319 per patient in our model, is indexed to 100.279
Our analysis of the accuracy of the algorithm in predicting the expected bias and280
confidence interval of EVPI estimates focusses on parameter θ5, the % responding281
to treatment T0. For inner sample sizes of J=10, 100 and 500 and outer samples282
also of K= 10, 100 and 500 we undertake comparisons between the predictions of283
our algorithm and the results of re-running the Monte-Carlo estimation of EVPI284
1000 different times. By generating 1000 real estimates of EV PI(θ5), using for285
example J = 10 inner and K = 10 outer samples, we can compute the mean bias286
exhibited and the 95% confidence interval exhibited in these EVPI estimates. We287
then compare the mean bias exhibited with that predicted by the algorithm and288
similarly the confidence interval width exhibited with that predicted by the289
algorithm.290
Partial EVPI Sample Sizes, Oakley et al. 18
ab
cd
ePara
mete
rsPara
mete
rM
ean
Valu
es
giv
en
Unce
rtain
tyin
Means
Exis
ting
Evid
ence
Sta
ndard
Devia
tions
T0
T1
(T1-
T0)
T0
T1
Cos
tof
Dru
g(θ
1,θ
11)
$10
00$1
500
$500
$1$1
%A
dm
issi
ons
(θ2,θ
12)
10%
8%-2
%2%
2%D
ays
inH
ospit
al(θ
3,θ
13)
5.20
6.10
0.90
1.00
1.00
Cos
tper
day
(θ4)
$400
$400
...
$200
$200
%R
espon
din
g(θ
5,θ
14)
70%
80%
10%
10%
10%
Uti
lity
Chan
geif
resp
ond
(θ6,θ
15)
0.30
0.30
...
0.10
0.05
Dura
tion
ofre
spos
e(y
ears
)(θ
7,θ
16)
3.0
3.0
...
0.5
1.0
%Sid
eeff
ects
(θ8,θ
17)
25%
20%
-5%
10%
5%C
han
gein
uti
lity
ifsi
de
effec
t(θ
9,θ
18)
-0.1
0-0
.10
-0.0
00.
020.
02D
ura
tion
ofsi
de
effec
t(y
ears
)(θ
10,θ
19)
0.50
0.50
...
0.20
0.20
Centr
alEst
imate
Resu
lts
Tot
alC
ost
$120
8$1
695
487
Tot
alQ
ALY
0.61
750.
7100
0.09
25C
ost
per
QA
LY
5267
Net
Ben
efit
give
nth
resh
old
(λ)=
$100
00)
$496
7$5
405
$438
Table 1: Summary of Model and Parameters for Case Study 1
Partial EVPI Sample Sizes, Oakley et al. 19
4.2 Case study 2: A complex cost-effectiveness model291
The second case study uses a published health economic model developed on292
behalf of the National Institute for Clinical Excellence (NICE) to evaluate the293
cost-effectiveness of beta interferon and glatiramer acetate in the management of294
multiple sclerosis in the UK. Full details are given elsewhere [Chilcott et al.,295
2003b]. Briefly, the model uses a state transition methodology to describe the296
clinical course of patients with relapsing/remitting MS and secondary progressive297
MS through the Expanded Disability Status Scale (EDSS). Owing to the298
time-dependence of the probabilities of transiting between health states, a single299
run of the model required around 7 seconds, meaning that probabilistic sensitivity300
analysis using the model was highly computationally expensive. In fact, if 10,000301
samples were used for both the inner and outer levels, then calculating the partial302
EVPI for just 1 of the models 128 uncertain parameters using the 2-level sampling303
algorithm would require around 22.2 years computation time. Using 10,000 random304
samples we estimated the overall EVPI for the model to be £8,855 per patient.305
One-way sensitivity analyses identified several model parameters as important and306
we chose to focus our analysis of the bias and confidence intervals for EVPI307
estimates on the parameter: mean cost associated with the EDSS 9.5 health state.308
Partial EVPI Sample Sizes, Oakley et al. 20
5 Results309
5.1 Results using Case Study 1310
The true EVPI for the parameter of interest θ5 is estimated at around $199, which311
is 15.1 when indexed against overall EVPI ($1319=100). The results of running312
the algorithm to predict bias and confidence intervals are shown in Table 2.313
J = 10 J = 100 J = 500 J = 1000 J = 5000 J = 10000
Bias (independent of K) 15.61 1.58 0.25 0.17 0.05 0.02
95% CI
K = 10 ±43.9 27.7 25.5 25.1 24.9 24.8
K = 100 ±13.9 8.8 8.1 7.9 7.9 7.9
K = 500 ±6.2 3.9 3.6 3.6 3.5 3.5
K = 1000 ±4.4 2.8 2.5 2.5 2.5 2.5
K = 5000 ±2.0 1.2 1.1 1.1 1.1 1.1
K = 10000 ±1.4 0.9 0.8 0.8 0.8 0.8
Table 2: Predicted bias and 95% CI for Monte Carlo partial EVPI estimate in case
study 1 using our proposed algorithm
The predicted bias in the Monte Carlo estimate for a given inner sample size is314
compared with the actual bias exhibited in Figure 1. The order of magnitude of315
the predicted expected bias is very similar to that mean bias actually exhibited316
using 1000 repeated runs of the two level partial EVPI. The absolute scale of the317
Partial EVPI Sample Sizes, Oakley et al. 21
bias is quite high for very small numbers of inner samples. For example, with just318
10 inner runs used, the bias in this case study is approximately 15.5 when indexed319
to the overall EVPI = 100, which would more than double our estimate for the320
indexed partial EVPI for parameter θ5. When larger numbers of inner runs are321
used, the bias fall substantially with an expected indexed bias of just 0.25 when322
using J=500 inner iterations. This suggests that on average, if we used J=500323
iterations we would estimate the indexed partial EVPI for parameter θ5 at 15.35324
rather than 15.1. The near equivalence of the predicted versus actuals for the325
different inner numbers of iterations tested suggests that our algorithm provides a326
robust estimate of the expected bias due to small numbers of iterations, at least in327
this case study.328
Comparison of Bias Predicted by Proposed Algorithm versus Exhibited bias in Case Study 1
-4
0
4
8
12
16
20
10 100 500
J - number of Inner Samples
Exp
ecte
d bi
as
(Ind
exed
to
over
all
EV
PI
= 1
00)
Predicted bias
Actual BiasExhibited
Figure 2: Comparison of Bias Predicted by Proposed Algorithm versus Exhibited
bias in Case Study 1.
Figure 3 shows that the predicted confidence intervals for the partial EVPI329
estimates are similar but slightly wider than the actual confidence intervals330
exhibited when using 1000 repeated computations of EV PI(θ5). The estimated331
Partial EVPI Sample Sizes, Oakley et al. 22
widths are of sufficient accuracy to be informative, and as discussed earlier, K can332
be refined once we have actually chosen J and obtained a partial EVPI estimate.333
With the smallest numbers tested, (K=J=10), the predicted 95% confidence334
interval for the indexed partial EVPI is ±43.8 i.e. an estimate could be expected335
to be biased by 15.5 indexed points as discussed earlier and vary anywhere336
between (-13 to +74) when indexed to overall EVPI = 100. Clearly using such337
small numbers to produce one estimate of EV PI(θ5) would give almost338
meaningless results. As larger numbers of iterations are used, the predicted339
confidence intervals for EV PI(θ5) reduce. When K=J=500 is used, the predicted340
95% confidence interval is 15.35± 3.6. It may be that even higher numbers of341
iterations are required if this level of accuracy is not enough for analysts or342
decision makers’ needs. It can be seen that increasing the outer sample size K343
reduces the confidence interval in proportion to√
K. Again, the near equivalence344
of the predicted versus actuals for the different numbers of outer and inner345
iterations tested suggests that our algorithm provides a robust estimate of the346
order of magnitude of 95% confidence intervals for partial EVPI estimates, at least347
in this case study.348
5.2 Results using Case Study 2349
For case study 2, the predicted bias in EVPI estimates produced by our algorithm350
for different numbers of inner samples J are shown in Table 3. The analysis351
suggests that an inner sample of 1,000 would still produce an appreciable bias352
Partial EVPI Sample Sizes, Oakley et al. 23
Comparison of Confidence Intervals Predicted by ProposedAlgorithm versus Exhibited Confidence Intervals in Case Study 1
-
5
10
15
20
25
30
35
40
45
50
10*1
0
10*1
00
10*5
00
100*
10
100*
100
100*
500
500*
10
500*
100
500*
500
Outer K * Inner J Samples Used
Con
fiden
ce In
terv
al (+
/- X
) (
Ind
exed
to o
vera
ll E
VP
I = 1
00)
Precticted CIActual CI
Figure 3: comparison of Confidence Intervals Predicted by Proposed Algorithm
versus Exhibited Confidence Intervals in Case Study 1.
(2.42 on the indexed scale).353
In this case study, we tested different versions of our algorithm, extending the354
number of inner iterations used within the algorithm from our originally specified355
J = 30 up to 40, 50 and 60. We found some small variations but no appreciable356
differences in assessing the order of magnitude of the predicted bias when using357
greater than 30 inner iterations in our algorithm.358
Table 4 shows the estimated the width of a 95% confidence interval for the partial359
EVPI for various outer sample sizes, given an inner sample size of J = 1000. We360
can see that the confidence intervals are fairly wide here, indicating a relatively361
large outer sample size will also be needed, certainly in excess of K = 1000.362
The results of our case study 2 analyses, on the basis of 630 model evaluations, led363
us to establish that a likely minimum sample size for an accurate 2-level Monte364
Partial EVPI Sample Sizes, Oakley et al. 24
Inner Predicted Bias Indexed
Samples (overall EVPI=100)
J = 500 £ 456 5.25
J = 1000 £ 214 2.42
J = 5000 £ 45 0.51
J = 10000 £ 20 0.22
Table 3: The predicted size of a bias in a Monte Carlo partial EVPI estimate for
parameter cost associated with health state EDSS 9.5, in case study 2.
Outer Samples K K=100 K=1,000 K=10,000 K=100,000
Absolute Value £ 11, 729 £ 3, 709 £ 1173 £ 371
Indexed to Overall EVPI 132.45 41.88 13.25 4.19
Table 4: Estimated Width of 95% confidence interval for a Monte Carlo partial
EVPI estimate for parameter cost associated with health state EDSS 9.5, for an
inner sample size J = 1000 and outer sample size K.
Carlo estimate will be of the order of 1 million model runs. Partial EVPI365
estimation using the Monte Carlo approach was clearly infeasible because this366
would require 81 days computation time. At this point we moved on to utilise the367
more efficient Gaussian Process model to approximate partial EVPI, the details of368
which are given elsewhere [Tappenden et al., 2004].369
Partial EVPI Sample Sizes, Oakley et al. 25
6 Discussion370
We have presented an algorithm for predicting the expected bias and confidence371
interval for a Monte Carlo estimate of partial EVPI. Testing of the algorithm372
against exhibited bias and confidence intervals in our first case study shows that it373
provides predictions with the correct order of magnitude. Our second case study374
provides an example where infeasibly long computation times would be required375
for accurate EVPI estimates and a Gaussian process meta-model was required to376
emulate the original model and produce much quicker model runs [Tappenden377
et al., 2004].378
The algorithm can be applied generally to any decision model, and is not limited379
to economic evaluation of health technologies. It can be applied patient simulation380
models [Brennan et al., 2006], where there is the additional complication of how381
many patient simulations per parameter set to use [O’Hagan et al., 2007]. It is also382
applicable to models with any characterisation of uncertainty and is not limited to383
continuous or parametric distributions. Although both of our case studies384
examined the partial EVPI of a single parameter, the algorithm is identically385
applicable to computing EVPI for a group of parameters Xi.386
The algorithm is very quick to compute. The time taken to compute predicted387
biases and confidence intervals for the 36 options of inner J and outer K samples in388
table 2 for case study 1 was just over 8 seconds. This compares with several days389
computation to produce the 36,000 runs of actual two level partial EVPIs with390
which to compare the predictions in Figures 1 and 2.391
Partial EVPI Sample Sizes, Oakley et al. 26
The main limitation of the approach is that it requires the computation of the392
conditional distribution for the parameters not of interest X−i|Xi. In case study 1393
these were multi-variate normally distributed and so the conditional mean and394
variance matrix were easily available with analytic formulae. This may not be the395
case for other characterisations of uncertainty. Of course this problem applies just396
as equally to the computation of EVPI via Monte carlo itself as it does to the397
prediction algorithm for the bias and confidence intervals of these estimates. A398
second limitation is that this procedure will be most useful for models that can run399
fairly quickly (i.e. a few seconds per model run). For very computationally400
expensive models, such as patient simulation models that can take an hour per401
run, we believe a Gaussian meta-model approach is currently the only viable402
option for computing EVPI.403
One implication of considering the uncertainty in EVPI estimates explicitly is that404
decision makers and analysts may need to define some limit on how accurately405
they wish to know the true EVPI for parameters of interest. Both case studies406
suggest that relatively small numbers of inner and outer loops, such as K=J=500407
and a total of 250,000 model runs, provide a reasonable estimate of the order of408
magnitude of the partial EVPI. The diminishing returns of higher numbers of409
inner and outer iterations for EVPI accuracy are an unavoidable fact of the Monte410
Carlo estimation process. In case study 1 moving from K=J=500 to K=J=1000411
(an additional 750,000 models runs taking 4 times longer) reduces the 95% CI412
from ±3.6 to ±2.5 on the indexed scale. At a further extreme, moving from413
Partial EVPI Sample Sizes, Oakley et al. 27
K=J=5,000 to K=J=10,000, an additional 75 million model runs reduces the 95%414
CI by just 0.3, from ±1.1 to ±0.8.415
As partial EVPI calculations become more common in practice, we have a need to416
be as efficient as possible in producing estimates [Claxton et al., 2004]. We have417
tested the algorithm on only two case studies here and further research would be418
useful to test the performance of the algorithm in other models. The field of value419
of information analysis is widening to consider expected value of sample420
information (EVSI) [Ades et al., 2004], [Brennan and Kharroubi, 2007a] and the421
expected value of perfect implementation (EVPImp) [Fenwick and Claxton, 2005].422
In many of these analyses, the same issue of inner and outer sampling exists and it423
would be useful to adapt our algorithm to these contexts and investigate its424
accuracy and usefulness.425
In conclusion, this novel algorithm is easily and generally applied to compute426
predicted bias and confidence intervals for Monte Carlo based estimates of partial427
EVPI in decision models. The algorithm is a relatively simple tool, using standard428
results concerning uncertainty in Monte Carlo estimates, but extending them into429
the context of nested expectations with intervening maxima. Our judgement is430
that it should provide robust estimates in all kinds of decision model but further431
testing of this may be useful. The algorithm is particularly useful when relatively432
small numbers of inner and outer iterations are planned and is recommended for433
use when reporting all Monte Carlo based partial EVPI calculations.434
Partial EVPI Sample Sizes, Oakley et al. 28
References435
A. E. Ades, G. Lu, and K. Claxton. Expected value of sample information436
calculations in medical decision modeling. Medical Decision Making, 24:207–227,437
2004.438
A Brennan and R Akehurst. Modelling in health economic evaluation. what is its439
place? what is its value? Pharmacoeconomics, 17, 2000.440
A. Brennan and S.A. Kharroubi. Efficient computation of partial expected value of441
sample information using bayesian approximation. available at442
http://dx.doi.org/10.1016/j.jhealeco.2006.06.002. Journal of Health Economics,443
26, 2007a.444
A Brennan and SA Kharroubi. Expected value of sample information for weibull445
survival data (accepted). Health Economics, 2007b.446
A. Brennan, S. A. Kharroubi, J. Chilcott, and A O’Hagan. A two level monte447
carlo approach to calculating expected value of perfect information:- resolution448
of the uncertainty in methods. available at449
http://www.shef.ac.uk/content/1/c6/02/96/05/brennan.doc. Discussion paper,450
University of Sheffield, 2002.451
A. Brennan, SE Chick, and R Davies. A taxonomy of model structures for452
economic evaluation of health technologies. available at453
http://dx.doi.org/10.1002/hec.1148. Health Economics, 15:12951310, 2006.454
Partial EVPI Sample Sizes, Oakley et al. 29
A Brennan, SA Kharroubi, A O’Hagan, and J Chilcott. Calculating partial455
expected value of perfect information via monte-carlo sampling algorithms.456
(provisionally accepted). Med Decis Making, 2007.457
AH Briggs and A Gray. Handling uncertainty when performing economic458
evaluation of healthcare interventions. Health Technology Assessment, 1999.459
J. Chilcott, A. Brennan, A. Booth, J. Karnon, and P. Tappenden. The role of460
modelling in prioritising and planning clinical trials. Health Technology461
Assessment, 7, 2003a.462
J. Chilcott, C. McCabe, P. Tappenden, and A. et al O’Hagan. Modelling the cost463
effectiveness of interferon beta and glatiramer acetate in the management of464
multiple sclerosis. BMJ, 326, 2003b.465
K. Claxton. Bayesian approaches to the value of information: implications for the466
regulation of new health care technologies. Health Economics, 8:269–274, 1999.467
K. Claxton and J. Posnett. An economic approach to clinical trial design and468
research priority-setting. Health Economics, 5:513–524, 1996.469
K. Claxton and K. Thompson. A dynamic programming approach to efficient470
clinical trial design. Journal of Health Economics, 20:797–822, 2001.471
K Claxton, L Ginnelly, M Sculpher, Z Philips, and S Palmer. A pilot study on the472
use of decision theory and value of information analysis as part of the nhs health473
technology assessment programme. Health Technology Assessment, 8, 2004.474
Partial EVPI Sample Sizes, Oakley et al. 30
J. C. Felli and G. B. Hazen. Sensitivity analysis and the expected value of perfect475
information. Medical Decision Making, 18:95–109, 1998.476
J. C. Felli and G. B. Hazen. Correction: Sensitivity analysis and the expected477
value of perfect information. Medical Decision Making, 23:97, 2003.478
E Fenwick and Sculpher M Claxton, K and. The value of implementation and the479
value of information: Combined and uneven development. CHE Discussion480
Paper 5, 2005.481
BG Koerkamp, M Hunink, T Stijnen, and MC Weinstein. Identifying key482
parameters in cost-effectiveness analysis using value of information: a483
comparison of methods. Health Economics, 15, 2006.484
D. Meltzer. Addressing uncertainty in medical cost effectiveness analysis.485
implications of expected utility maximization for methods to perform sensitivity486
analysis and the use of cost-effectiveness analysis to set priorities for medical487
research. Journal of Health Economics, 20:109–129, 2001.488
J. E. Oakley. Sensitivity analysis for computationally expensive economic models.489
Technical report, Department of Probability and Statistics, University of490
Sheffield, 2003.491
JE Oakley and A O’Hagan. Probabilistic sensitivity of complex models: a492
Bayesian approach. J. Roy. Statist. Soc. Ser. B, 66:751–769, 2004.493
A. O’Hagan, M Stevenson, and J Madan. Monte carlo probabilistic sensitivity494
Partial EVPI Sample Sizes, Oakley et al. 31
analysis for patient level simulation models: efficient estimation of mean and495
variance using anova. Health Economics, page ??, 2007.496
H. Raiffa. Decision Analysis: Introductory Lectures on Choice Under Uncertainty.497
Addison-Wesley, Reading, Mass., 1968.498
MD Stevenson, JE Oakley, and JB Chilcott. Gaussian process modelling in499
conjunction with individual patient simulation modelling: A case study500
describing the calculation of cost-effectiveness ratios for the treatment of501
osteoporosis. Medical Decision Making, 24, 2004.502
P. Tappenden, J. Chilcott, S. Eggington, J. Oakley, and C. McCabe. Methods for503
expected value of information analysis in complex health economic models:504
developments on the health economics of interferon- and glatiramer acetate for505
multiple sclerosis. Health Technology Assessment, 8, 2004.506
BA Van Hout, Al MJ, G.S. Gordon, and F.F. Rutten. Costs, effects and c/e-ratios507
alongside a clinical trial. Health Economics, 1994.508
F. Yokota and K.M. Thompson. Value of information literature analysis509
(VOILA): A review of applications in health risk management. Medical510
Decision Making, 24, 2004.511