Journal of Medical Economics

ISSN: 1369-6998 (Print) 1941-837X (Online) Journal homepage: http://www.tandfonline.com/loi/ijme20

Challenges in economic modeling of anticancer

therapies: an example of modeling the survival
benefit of olaparib maintenance therapy for
patients with BRCA-mutated platinum-sensitive
relapsed ovarian cancer

Robert Hettle, John Posnett & John Borrill

To cite this article: Robert Hettle, John Posnett & John Borrill (2015) Challenges in economic
modeling of anticancer therapies: an example of modeling the survival benefit of olaparib
maintenance therapy for patients with BRCA-mutated platinum-sensitive relapsed ovarian cancer,
Journal of Medical Economics, 18:7, 516-524, DOI: 10.3111/13696998.2015.1024682

To link to this article: http://dx.doi.org/10.3111/13696998.2015.1024682

Accepted author version posted online: 03 Submit your article to this journal
Mar 2015.
Published online: 07 Apr 2015.

Article views: 967 View related articles

View Crossmark data Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=ijme20

Download by: [F. Hoffmann - La Roche Inc] Date: 31 May 2017, At: 13:30
 Journal of Medical Economics Vol. 18, No. 7, 2015, 516524

1369-6998 Article 0135.R1/1024682

doi:10.3111/13696998.2015.1024682 All rights reserved: reproduction in whole or part not permitted

Original article
Challenges in economic modeling of anticancer
therapies: an example of modeling the survival
benefit of olaparib maintenance therapy for
patients with BRCA-mutated platinum-sensitive
relapsed ovarian cancer

Robert Hettle Abstract

John Posnett Objective:
HERON Commercialization, PAREXEL Consulting,
London, UK The aim of this paper is to describe a four health-state, semi-Markov model structure with health states
defined by initiation of subsequent treatment, designed to make best possible use of the data available from
John Borrill a phase 2 clinical trial.
AstraZeneca, Research and Development, Alderley
Park, Cheshire, UK
Method:
The approach is illustrated using data from a sub-group of patients enrolled in a phase 2 clinical trial of
Address for correspondence: olaparib maintenance therapy in patients with platinum-sensitive relapsed ovarian cancer and a BRCA
John Posnett, HERON Commercialization, mutation (NCT00753545). A semi-Markov model was developed with four health states: progression-free
Evergreen Building North, 160 Euston Road, survival (PFS), first subsequent treatment (FST), second subsequent treatment (SST), and death. Transition
London NW1 2DX, UK.
probabilities were estimated by fitting survival curves to trial data for time from randomization to FST, time
Tel: +44 (0)207.121.1755;
from FST to SST, and time from SST to death.
john.posnett@parexel.com

Results:
Keywords:
Survival projections generated by the model are broadly consistent with the outcomes observed in the
Ovarian cancer Economic model Olaparib
clinical trial. However, limitations of the trial data (small sample size, immaturity of the PFS and overall
Anti-cancer agents
survival [OS] end-points, and treatment switching) create uncertainty in estimates of survival.
Accepted: 26 February 2015; published online: 7 April 2015
Citation: J Med Econ 2015; 18:51624
Conclusion:
The model framework offers a promising approach to evaluating cost-effectiveness of a maintenance
therapy for patients with cancer, which may be generalizable to other chronic diseases.

Introduction
Cancer is one of the leading causes of morbidity and death globally1. In Europe,
2.7 million new cancer cases were diagnosed in 2012, and cancers accounted for
1.3 million deaths2. Four of the top 10 causes of premature death in high income
countries were cancers1. The disease burden is increasing. The annual number of
cancer cases globally is forecast to rise from 14 million in 2012 to 22 million
within two decades1.
The growing burden of cancer in high income countries has stimulated
research into the development of new therapies designed to prolong survival

516 Challenges in economic modeling of anti-cancer therapies Hettle et al. www.informahealthcare.com/jme ! 2015 Informa UK Ltd

and/or enhance health-related quality-of-life; and regula-
tors have also responded by establishing fast-track proced-
ures designed to make effective treatments available
to patients more quickly. The US Food and Drug
Administration (FDA) has established four different mech-
anisms designed to accelerate patient access to drugs that
treat serious diseases and which offer benefits over existing
treatments3. Similar procedures are in place in Europe4.
One of the implications is that regulatory trials for anti-
cancer therapies are often relatively small, single-arm trials
with limited follow-up beyond the primary end-point.
Trials in oncology comprised 9% of all of the trials
registered with ClinicalTrials.gov between 200720105.
These trials were typically small (74.9% enrolled 5100
patients and median enrolment was 43); most were
single-arm, non-randomized studies (64.7% compared
with 31.2% for all trials); and most were pre-phase 3
(70.6% compared with 41.8% for all trials). Phase 2
trials comprised 38.2% of trials in oncology compared
with 20.7% of all trials. Between 19922010, the FDA
granted accelerated approval for 47 new indications for
anti-cancer drugs; 19 were based on randomized trials
and 28 (60%) were based on single-arm studies6. Of the
19 accelerated approvals based on randomized trials, none
had overall survival as a primary end-point. Approvals
based on single-arm studies were most commonly based
on response rate or response duration6.
The development of new treatments for cancer has also
placed a focus on the importance of economic evaluation
to estimate the opportunity costs of new treatments for
cancer in terms of health benefits forgone from other
uses of limited healthcare budgets. Clinical trials are the
primary source of data used to examine the incremental
health benefits and costs associated with new healthcare
technologies, but limitations in the information available
from some anti-cancer trials creates problems for economic
modeling.
This paper discusses some of the challenges in modeling
the cost-effectiveness of new therapies in oncology, and
reports on a modeling approach that has been developed
to address some of these challenges. The approach is illu-
strated using data from a sub-group of patients in a phase 2
placebo-controlled clinical trial of olaparib maintenance
therapy in patients with platinum-sensitive relapsed ovar-
ian cancer and a BRCA mutation (BRCAm). The aim of
this paper is to describe a four health-state, semi-Markov
model structure with health states defined by initiation of
subsequent treatment, designed to make best possible use of
the data available from the phase 2 olaparib clinical trial.
Olaparib maintenance therapy
Recent advances in the development of treatments for
patients with advanced ovarian cancer include the poly
! 2015 Informa UK Ltd www.informahealthcare.com/jme
Journal of Medical Economics Volume 18, Number 7 July 2015
(ADP-ribose) polymerase (PARP) inhibitor olaparib,
which is being investigated for use as a maintenance ther-
apy in patients with platinum-sensitive relapsed ovarian
cancer who carry the BRCA mutation. The role of main-
tenance therapy is to delay disease progression, postpone
the need for subsequent chemotherapy, and potentially
improve the long-term survival of patients who achieve a
response to platinum-based chemotherapy.
Evidence on olaparib in this population is available
from a phase 2, placebo-controlled, randomized trial in
patients who have experienced at least two platinum-
based lines of therapy (NCT00753545)7,8. Patient-level
data were available from a pre-defined sub-group of
patients (n 136) with the BRCA-1 or BRCA-2 mutation
enrolled in the phase 2 trial. Eligible patients had recurrent
ovarian or fallopian tube cancer, or peritoneal cancer.
Patients entering the study had received two or more pre-
vious courses of platinum-based chemotherapy with com-
plete or partial response to the last chemotherapy course
prior to enrolment. Patients were randomized at a ratio of
1:1 to olaparib 400 mg twice daily or placebo twice daily.
Study medication was discontinued on disease progres-
sion7,8. This study provides an illustration of some of the
typical challenges in economic modeling: small patient
numbers, lack of data collection after the primary end-
point, and treatment switching that occurred outside of
the study design.
The primary end-point in the phase 2 study was radio-
logical progression-free survival (PFS). In the BRCAm
sub-group, the number of events at data cut-off for the
primary PFS analysis (June 30, 2010) was 26/74 (35.1%)
for olaparib and 46/62 (74.2%) for placebo: an overall
maturity of 53% in the BRCAm sub-group8. No radio-
logical measurements of disease progression were collected
after the primary PFS analysis. To establish whether PFS
benefits were maintained with longer follow-up, the
European Medicines Agency (EMA) requested additional
post-hoc exploratory analysis. Additional analysis was car-
ried out on the time to treatment discontinuation or death
(TDT), time to first subsequent treatment or death
(TFST), and the time to second subsequent treatment or
death (TSST). The TDT and TFST end-points may be
considered to be proxies for symptomatic disease progres-
sion with the advantage that they have a longer follow-up
than PFS. The number of first subsequent treatment (FST)
events at second data cut-off (November 2012) in the
BRCAm sub-group was 46/74 (62.2%) for olaparib and
54/62 (87.1%) for placebo: an overall maturity of
74.7%8. An analysis of TSST was conducted as a proxy
for second disease progression and used to assess the dur-
ability of treatment benefit beyond initial progression. The
time to initiation of subsequent treatment was not limited
in patients receiving olaparib due to safety or other reasons
(for example, there was no mandated wash-out period).
Challenges in economic modeling of anti-cancer therapies Hettle et al. 517
Journal of Medical Economics Volume 18, Number 7 July 2015
At the second data cut-off (November 2012), 45.9% of
patients in the olaparib group had received at least
12 months of treatment, and 17.6% of patients had
received at least 3 years of treatment8. After discontinuing
maintenance therapy, 75% of olaparib and 90% of placebo
patients received a subsequent line of treatment. After
discontinuing maintenance therapy, 45.5% of olaparib
and 58.6% of placebo patients received at least two subse-
quent lines of treatment8.
Subsequent therapies following first disease progression
included PARP inhibitors (including olaparib), which
were not permitted within the trial, but could be accessed
following enrolment into other clinical studies.
Subsequently, 14 out of the 62 (22.6%) patients with a
BRCAm randomized to placebo were treated with
a PARP inhibitor after study drug withdrawal. No patients
randomized to olaparib subsequently received a PARP
inhibitor. The imbalance in the subsequent use of PARP
inhibitors between treatment groups is a potential source
of confounding in estimates of differences in overall
survival (OS).
Challenges for economic modeling
Limitations of the clinical trial evidence base, primarily
short follow-up, and uncertainty around estimates of
survival benefit have created challenges for economic
modeling, but also for health technology assessment
(HTA) agencies responsible for assessing the cost-
effectiveness of new drugs. In a recent review of funding
recommendations for new cancer treatments made by
HTA agencies in Canada and England between 2002
20139, two-thirds of submissions (55/83) presented no
evidence of a gain in OS, but lack of clinical evidence
was cited as the reason for rejection in only in six of 24
cases. Seventy-five per cent of negative recommendations
(18/24) were based on lack of cost-effectiveness, and 39%
of positive recommendations (23/59) were conditional on
a pricing arrangement or patient access scheme. Economic
modeling undertaken from a limited evidence base
generates a high degree of uncertainty in estimates of
cost-effectiveness, but cost-effectiveness is nonetheless
likely to be an important determinant of a positive recom-
mendation for funding. The olaparib phase 2 clinical trial
illustrates many of the challenges that are common in
oncology modeling: relatively small sample size, limited
follow-up, surrogate outcomes and potential bias caused
by treatment switching after the discontinuation of study
drug.
It is not uncommon for patients with cancer to receive
long-term treatment involving a number of subsequent
lines of therapy following disease progression10. If the
drug resistance profile of a tumor is affected by an experi-
mental therapy, current treatment could impact the
518 Challenges in economic modeling of anti-cancer therapies Hettle et al.
activity of next-in-line treatments. A maintenance ther-
apy with benefit in terms of prolonged time to first disease
progression (PFS1) may theoretically lead to shorter over-
all survival because of the effect of the drug on next-in-line
treatments. Consequently, PFS alone may not be an appro-
priate surrogate for OS. The EMA recognizes this concern
and recommends11 that clinical studies should be designed
to test for a difference in OS or, where this is not feasible,
for a difference in PFS after next-in-line therapy (PFS2),
measured by time from randomization to second disease
progression or death.
When treatment switching occurs after follow-up,
observations of OS may be biased by the confounding
effect of the experimental treatment on survival in the
control group. If treatment switching occurs after the
point of first disease progression, measures of PFS1 are
not confounded, but measures of PFS2 and extrapolations
of OS will likely be biased. The National Institute for
Health and Care Excellence (NICE) Decision Support
Unit (DSU) reviewed all 45 technology appraisals (TAs)
in advanced or metastatic cancer assessed by NICE up to
December 200912. Treatment switching occurred in piv-
otal clinical trials in 25 of these TAs (56%). There is no
agreement on the most appropriate means of adjusting for
switching, and there was no consistency in the methods
used in the TAs reviewed by the DSU: 20 of 25 TAs either
made no adjustment (unadjusted intention to treat [ITT]
analysis), simply excluded patients who switched treat-
ments, or modeled outcomes on PFS only. None of these
approaches is likely to have been appropriate. Other
approaches included the assumption of no difference in
OS between treatments, or applying the same risk of
death in both trial arms following disease progression.
Few TAs (3/25) attempted to adjust for treatment switch-
ing using one of the more complex methods, such as
inverse probability of censoring weights, rank preserving
structural failure time model, or iterative parameter esti-
mation algorithm. Different methods of adjusting for the
effects of switching can have significantly different effects
on estimated outcomes of an economic model12,13, but as
yet there is no single method that is likely to be suitable in
all cases.
Methods
The health state structure of the economic model
developed for olaparib is an extended version of the
conventional three-state structure typically used in the
evaluation of anti-cancer treatments: PFS, progressive dis-
ease (PD), and death. This structure was used in the two
most recent NICE appraisals of maintenance treatments in
oncology (pemetrexed maintenance treatment following
induction therapy with pemetrexed and cisplatin for
non-squamous non-small-cell lung cancer (TA309)14;
www.informahealthcare.com/jme ! 2015 Informa UK Ltd

and erlotinib monotherapy for maintenance treatment of
non-small-cell lung cancer (TA227)15). The cost-effect-
iveness model for rituximab for first-line maintenance
treatment of non-Hodgkins lymphoma (TA226)16 had
four health states: PFS in the first-line maintenance
phase (PFS1), PFS after receiving second-line induction
treatment with rituximab in combination with chemo-
therapy (PFS2), PD, and death.
None of these recent cost-effectiveness evaluations
exactly matches the decision problem in the olaparib
case, although the erlotinib case comes closest. In order
to capture the expected benefits of olaparib in delaying
progression and postponing the transition to a subsequent
line of chemotherapy, and to provide a framework that is
flexible enough to address some of the limitations of the
data available from the olaparib clinical trial, we devel-
oped a four health-state model structure. The health states
in the model are: PFS, FST, second subsequent treatment
(SST), and death.
The health states were defined in terms of FST and
SST, using the exploratory end-points of the olaparib
phase 2 clinical trial8. These end-points were preferred
to the primary end-point of PFS for two main reasons.
First, FST and SST are more meaningful health states
than radiological defined PFS for the purpose of an eco-
nomic model designed to estimate differences in expected
costs and patient utilities. The initiation of a new line of
treatment is likely to trigger a change in resource use and
costs and, where therapy is changed as a result of worsening
symptoms, FST and SST are also likely to be correlated
with a reduction in patient utility. Second, because extra-
polating on the basis of FST may reduce uncertainty in
estimates of OS because of the larger number of observa-
tions on FST at data cut-off (100/136) than on PFS (72/
136). Where it is not feasible to determine the PFS of the
next-line therapy within a clinical study, the EMA
acknowledges that the time to next-in-line therapy (FST
to SST) may be a reasonable proxy for second disease
progression11.
The effects of treatment were simulated using semi-
Markov state transitions estimated by fitting survival
curves to trial data for time from randomization to FST
or death, time from FST to SST or death, and time from
SST to death. This technique is more flexible than the
traditional partitioned survival approach typically used
to simulate the outcomes of fixed regimen chemotherapy.
The incremental effects of treatment in the semi-Markov
model are determined by the FST end-point. Extrapolating
the benefits observed at the FST end-point through to
survival via the Markov state transition structure is
expected to lead to a more robust assessment of the differ-
ence in OS between treatment arms because of the longer
follow-up period (compared with PFS). Partitioned sur-
vival modeling relies on the assumption that the observed
trend in OS is representative of the future trend in OS.
! 2015 Informa UK Ltd www.informahealthcare.com/jme
Journal of Medical Economics Volume 18, Number 7 July 2015
This technique can be well suited to cases where survival
follow-up is mostly complete and the time period of the
extrapolation is limited. The robustness of this approach in
cases of limited maturity (560%) is unclear, and has not
been explored in previous published studies. The Markov
modeling technique has been used in other chronic relap-
sing-remitting diseases (e.g., relapsing-remitting multiple
sclerosis), and was recently applied in a NICE technology
appraisal of bevacizumab in combination with gemcita-
bine and carboplatin for treating first recurrence of
platinum sensitive advanced ovarian cancer (TA 285)17.
The olaparib model allows for different adjustments for
the potentially confounding effects of treatment switching
on estimates of OS. In the phase 2 study, 14/62 (22.6%) of
patients randomized to placebo subsequently received a
PARP inhibitor8. In the base case, adjusted analysis, the
risks of transitioning from the FST state or SST state in
the placebo group were set equal to the risks observed in
the olaparib group (where PARP inhibitors were not avail-
able post-discontinuation). This approach assumes that
the effect of the intervention lasts only until first disease
progression. Beyond that point there is no difference in the
risk of further progression or death between the treatment
arms and a gain in time to FST will be translated into a
gain in OS12. The model is flexible enough that this
assumption can be varied from the most conservative
approach (higher risk of further progression or death
such that OS is the same in both arms) to the more gen-
erous assumption (any benefit in FST is maintained
throughout the lifetime of the patient).
The process for fitting parametric survival curves to
patient level data was based on the guidance document
prepared by the NICE DSU18. Parametric survival
models were fitted to time to event data for the following
end-points: time from randomization to FST or death
(event one); time from FST to SST or death (event
two): time from SST to death (event three). Transition
probabilities were derived using the following equations:
TPState  i, state  j, time 1, time 2
 
Stime 2
Pnon  fatal event  1 
Stime 1
TPState  i, death, time 1, time 2
 
Stime 2
Pfatal event  1 
Stime 1
where TP(state i, state j, time 1, time 2) is the transition
probability from state-i to state-j (or death) over the period
between time 1 and time 2. P(fatal event) and P(non-fatal
event) 1  P(fatal event) are the proportions of fatal and
non-fatal events. S(time) is the cumulative survival prob-
ability derived from the parametric survival model. In
these equations, time 1 and time 2 are expressed as time
since entering state-i.
Challenges in economic modeling of anti-cancer therapies Hettle et al. 519
Journal of Medical Economics Volume 18, Number 7 July 2015
Progression-free survival
First progression
event
First subsequent treatment
Second
progression event
Second subsequent
treatment
Death
Figure 1. Health state structure of the economic model.
Diamond boxes correspond to events and rectangular boxes correspond to health
states.
Patients who experience a non-fatal event immedi-
ately transition to their next subsequent treatment state
in the model (Figure 1). For these patients, the time
since entering the state (time 1) is set to zero, and
the risk of subsequent health state transition is deter-
mined by the next parametric survival model in the
series. Patients who do not transition in a given cycle
are assumed to remain in their current health state. For
these patients, the time since entering their current
health state (time 1 or time 2) advances by 1 month
(cycle length). The probability of transition in the next
cycle is calculated using the updated time 1 and time 2
variables. The parametric survival modeling was con-
ducted using the flexsurv package in the statistical soft-
ware, R19. The semi-Markov transition model was
developed in MS Excel 2007.
Transition probabilities were calculated using the
log-normal distribution for event one, and the Weibull
distribution for event two and event three. The choice of
statistical distribution was based on goodness of fit statis-
tics, and visual fit to the Kaplan-Meier curves (Figure 2).
The transition probabilities for event two and event three
were modeled using a common survival distribution
(Weibull). Other statistical distributions (generalized
gamma for placebo and log-normal for olaparib) produced
an equally plausible fit to these event rates. However, dif-
ferent distributions impose different assumptions on the
behavior of transition rates when extrapolated, which
can ultimately influence the difference in survival between
arms in the model. To reduce the risk of survival being
driven by these assumptions, a common survival distribu-
tion (Weibull) was used to model events two and three for
both olaparib and placebo.
520 Challenges in economic modeling of anti-cancer therapies Hettle et al.
Results
In terms of the proportions of patients on olaparib treat-
ment and alive at 1, 2, and 3 years, the predictions of
the model are close to the values observed in the trial
(Table 1). The estimated median time to olaparib treat-
ment discontinuation (11 months) is similar in both cases
(1111.5 months).
At data cut-off for the primary end-point in the trial
(PFS), patients receiving olaparib had a significantly
longer median PFS than patients in the control group
(11.2 months vs 4.3 months; hazard ratio [HR] 0.18;
p50.0001)8, a difference of 6.9 months. The median
follow-up was 5.6 months and there were a total of 72/
136 events (53% maturity). In the BRCAm sub-group,
the median time to FST observed in the clinical trial was
significantly longer in the olaparib arm (15.6 months vs
6.2 months; HR 0.33; p50.0001)8, a difference of 9.4
months. The economic model predicts a difference in the
median time to FST, unadjusted for treatment switching,
of 12 months (19.5 months vs 7.5 months) (Table 1).
Median times to event and the difference between treat-
ment arms are longer than was observed in the trial.
Adjusting for treatment switching does not change the
time to FST, because the adjustment applies only after
the initiation of FST. A common feature of the trial data
and modeled output is that the difference between treat-
ments in time to FST is longer than the observed differ-
ence in PFS.
In the clinical trial, the median time to SST was also
significantly longer in the olaparib arm (23.8 months vs
15.2 months; HR 0.44; p50.001)8, a difference of 8.6
months (Table 1). The economic model predicts a
difference of 9 months (27.5 months vs 18.5 months)
unadjusted for treatment switching. In the trial and in
the economic model, the difference in time to SST (and
FST) is longer than the difference in PFS. In addition, the
difference in time to SST and time to FST is approximately
the same, and this would be consistent with the expect-
ation that the benefit of olaparib does not increase after
treatment discontinuation, or conversely that there is no
significant negative impact on the efficacy of next-in-line
therapy.
In the trial the difference in median OS was not sig-
nificantly different between treatments (34.9 months
[olaparib] vs 31.9 months; HR 0.73; p 0.19)6
(Table 1). The number of events in the BRCAm sub-
group was 71/136 (52% maturity) at a median follow-up
of 37.3 months. The economic model predicts a difference
of 7 months in median survival (38.5 months vs 31.5
months) unadjusted for treatment switching. In the ana-
lysis adjusted for switching the difference is 12.5 months
(38.5 months vs 26.0 months), which is the same as the
estimated difference in time to FST. The difference in
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
! 2015 Informa UK Ltd www.informahealthcare.com/jme
Journal of Medical Economics

Figure 2. KaplanMeier survival plot, fitted parametric survival curves and 95% confidence intervals. (a) Time to treatment discontinuation; (b) time to first subsequent therapy; (c) time from first to second
subsequent therapies; (d) time from second subsequent therapy to death.
Volume 18, Number 7

Challenges in economic modeling of anti-cancer therapies Hettle et al.

July 2015

521
Journal of Medical Economics Volume 18, Number 7 July 2015

Table 1. Observed versus modelled clinical outcomes (undiscounted) in the BRCAm subgroup of the phase 2 study.

Arm End-point Trial reported Modelled outcome

(log-normal)

Olaparib Proportion of patients on treatment and alive at 1 year 45.9% 49.6%

Proportion of patients on treatment and alive at 2 years 28.4% 23.6%
Proportion of patients on treatment and alive at 3 years 17.6% 13.5%
Median time to discontinuation of treatment 11.0 months 11.5 months
Median time to first subsequent treatment or death 15.6 months 19.5 months
Median time to second subsequent treatment or death 23.8 months 27.5 months
Median time to death (overall survival) 34.9 months 38.5 months
Placebo Median time to first subsequent treatment or death 6.2 months 7.5 months
Median time to second subsequent treatment or death: no adjustment 15.2 months 18.5 months
for cross-over
Median time to death (overall survival): no adjustment for cross-over 31.9 months 31.5 months
Median time to death (overall survival): adjusted for cross-over NR 26.0 months
Incremental Difference in median time to first subsequent treatment or death 9.4 months 12 months
(olaparib vs placebo)
Difference in median time to second subsequent treatment or death: 8.6 months 9.0 months
no adjustment for cross-over
Difference in median time to death: no adjustment for cross-over 3.0 months 7.0 months
Difference in median time to death: adjusted for cross-over NR 12.5 months

NR not relevant to trial data.

Table 2. Comparison of the observed overall survival with modelled survival (BRCAm ITT population).

Outcome Trial data Parametric OS model Four health state model

Olaparib Placebo
OS Olaparib Placebo
OS Olaparib Placebo
OS

6-month survival (%) 98.6 96.7 1.9 99.3 98.4 0.9 98.5 97.7 0.8
12-month survival (%) 92.9 88.2 4.7 93.8 89.0 4.8 93.3 90.1 3.2
18-month survival (%) 82.6 68.8 13.8 83.9 75.2 8.7 85.3 78.2 7.1
24-month survival (%) 73.8 59.9 13.9 72.8 61.6 11.2 75.5 65.3 10.2
Median survival (months) 34.9 31.9 3.0 38.0 30.0 8.0 38.5 31.5 7.0
Mean survival NA NA NA 48.7 39.1 9.6 46.1 34.2 11.9

NA not applicable.

mean survival predicted by the model is 11.9 months (46.1
months vs 34.2 months) (Table 2).
Table 2 shows the results of a comparison between the
OS predictions of the economic model and OS predicted
by the best fitting parametric survival curve (log normal)
fitted directly to time to death data from the olaparib phase
2 study (Table 2).
Both of the modeling approaches produce a similar esti-
mate of the difference in median survival between olaparib
and placebo (78 months), but both approaches over-
estimate the difference in median survival compared with
the trial (3 months). The main driver of this difference is
the lower olaparib survival observed in the trial (34.9
months) compared with the predictions of the models
(3838.5 months). Estimated survival for the placebo
group is similar (31.9 [trial] vs 3031.5 months [models]).
The median OS times in the placebo arm predicted by
models and observed in the trial are both longer than was
observed in a recent study of untreated patients who met
the eligibility criteria for the olaparib phase 2 study in
which the median time to death was 21.9 months20.
The parametric OS model and the four health-state
economic model produce similar estimates of the incre-
mental survival benefit of olaparib vs placebo at 6, 12,
18, and 24 months, although neither of the modeling
approaches produces a particularly close estimate of the
incremental survival benefit of olaparib observed in the
trial. At each of the time-points both modeling approaches
tend to over-estimate survival for both olaparib and pla-
cebo, but the overall effect is not consistent, with the result
that the incremental benefit is similar or lower in the
modeled approach than in the trial at 6 months and 12
months, and higher in months 18 and 24.
The economic model predicts a lower mean survival in
both the olaparib and placebo arms than the parametric
OS model (46.1 months vs 48.7 months [olaparib]; 34.2
months vs 39.1 months [placebo]), although the incremen-
tal survival benefit of olaparib is higher with the four-state
model (11.9 months vs 9.6 months). The mean survival
predicted by the four-state model is consistent with values
reported for the control arm in the economic evaluation
of trabectedin for the treatment of relapsed ovarian

522 Challenges in economic modeling of anti-cancer therapies Hettle et al. www.informahealthcare.com/jme ! 2015 Informa UK Ltd

cancer: 34.8 months in the control arm receiving pegy-
lated liposomal doxorubicin (TA222)21.
Discussion
The four health-state, semi-Markov model was developed
in part to provide a flexible framework to address some of
the challenges common to economic models in oncology:
relatively small sample size, limited follow-up, surrogate
outcomes, and potential bias caused by treatment switch-
ing after study drug discontinuation. It was also developed
to make the best use of the information available from the
olaparib phase 2 clinical trial7,8. Following analysis of the
primary end-point in the trial, radiographic follow-up was
limited. However, information was collected on the time
to initiation of FST and SST. The main differences
between this approach and the conventional approach
are the use of a semi-Markov structure, the use of the
FST state instead of PFS as an indicator of disease progres-
sion, and the introduction of the SST state. Using time to
FST rather than a RECIST-based indicator of disease pro-
gression may be expected to correlate better with worsen-
ing health status/patient utility and with changes in
treatment costs. Comparing time to SST between inter-
vention and placebo arms makes it possible to examine
some of the possible impact of an experimental treatment
on the effectiveness of subsequent lines of therapy. A fur-
ther advantage is that observational data on time to FST
and SST are relatively easy to collect outside a trial proto-
col, and will typically have longer follow-up than PFS or
OS end-points at the time of a payer submission.
The predictions of the model are plausible in the sense
that they are in line with the expected effects of a main-
tenance treatment in this population. The aim of a
maintenance treatment is to delay disease progression
and postpone the need for a further line of treatment,
and this is consistent with the hypothesis that patients
treated with olaparib will have a longer period of PFS
and longer time to FST. The clinical trial shows a differ-
ence in median PFS in favor of olaparib of 6.9 months and a
difference in time to FST of 9.4 months. Radiological PFS
is not a health state in the economic model, but the model
also predicts a difference in favor of olaparib in time to FST
(12 months).
The benefits of a maintenance therapy should be sus-
tained beyond the period of treatment. The clinical trial
shows a difference in median time to second subsequent
treatment in favor of olaparib of 8.6 months, and this is
similar to the difference in time to FST (9.4 months). The
economic model predicts a median difference in time to
SST of 9 months (unadjusted for switching), which is close
to the difference observed in the trial. The differences
between treatments in the time to FST and in time to
SST observed in the trial are similar (8.6 months vs 9.4
! 2015 Informa UK Ltd www.informahealthcare.com/jme
Journal of Medical Economics Volume 18, Number 7 July 2015
months), and this is in line with the expectation that the
benefit of olaparib does not increase beyond the period of
treatment, and, conversely, that there is no major negative
impact on the efficacy of next-in-line treatments.
The difference between the economic model and the
clinical trial are most marked in estimates of time to FST
(9.4 months [trial] vs 12 months [model]) and in OS. In the
clinical trial, the difference between treatments in median
survival was 3 months, and this difference was not statis-
tically significant. The economic model predicts a median
difference in OS of 7 months (unadjusted) and
12.5 months (adjusted for switching) in favor of olaparib.
The unadjusted estimate is in line with the difference in
PFS observed in the trial (6.9 months), but is less than
the difference in time to FST estimated by the model
(12 months).
A comparison of OS predicted by the economic model
and the best fitting parametric survival curve fitted directly
to time to death data from the olaparib phase 2 study shows
that the two modeling approaches produced broadly simi-
lar outcomes. Both produced a similar estimate of the
unadjusted difference in median survival between olaparib
and placebo (78 months), and both over-estimated the
difference in median survival compared with the trial
(3 months). The main driver of this difference is the
lower olaparib survival observed in the trial (34.9
months) compared with the predictions of the models.
The parametric OS model and the economic model
produce similar estimates of the incremental survival
benefit of olaparib vs placebo at 6, 12, 18, and 24
months, although neither of the modeling approaches pro-
duces a particularly close estimate of the incremental sur-
vival benefit of olaparib observed in the trial. At each of
the time-points, both modeling approaches tend to
over-estimate survival for both olaparib and placebo.
The economic model predicts a lower mean survival in
both the olaparib and placebo arms than the parametric
OS model (46.1 months vs 48.7 months [olaparib]; 34.2
months vs 39.1 months [placebo]), but the incremental
survival benefit of olaparib is higher with the economic
model (11.9 months vs 9.6 months). An alternative
approach would be to fit Kaplan-Meier survival curves dir-
ectly to the trial data. This would produce a better fit to the
trial data, but would not necessarily produce a better basis
for extrapolating differences in survival beyond the trial
period.
The modeling approach discussed here has a number of
limitations. Extrapolating survival from the limited obser-
vations in a clinical trial is inevitable, but the choice of
parametric functions will very likely have an important
effect on outcomes. There is no unique approach to
extrapolation that is appropriate in every case. Although
we have followed good practice guidance18, model
estimates will be sensitive to the choice of parametric func-
tions. Similarly, results are likely to be sensitive to the
Challenges in economic modeling of anti-cancer therapies Hettle et al. 523
Journal of Medical Economics Volume 18, Number 7 July 2015
method of adjusting for treatment switching. There is no
single accepted way to adjust for treatment switching, and
the method we have used is pragmatic and flexible, but
more complex statistical approaches are available. A con-
firmatory phase 3 trial of olaparib maintenance therapy in
BRCAm PSR ovarian cancer is ongoing (NCT01874353),
with an estimated primary completion date of Q4, 2015.
The outcomes from this trial will be used to refine the
existing economic model.
Conclusion
The economic model will ultimately be used to evaluate
the cost-effectiveness of olaparib compared with standard
of care, and the estimated survival benefits of olaparib will
be a key driver of the incremental cost-effectiveness ratio
(ICER). The four-state economic model provides a good
representation of the triggers of treatment change and
symptomatic progression (and hence costs and patient util-
ity), and for this reason the proposed approach could be
expected to reduce uncertainty in estimates of cost and
utility. It remains to be seen if there are also benefits in
terms of reducing uncertainty around estimates of the
ICER. The cost-effectiveness analysis will be completed
as a separate exercise. The four-state model ultimately
depends on the availability of information on time to
FST and SST, but these are observational data that can
be collected with the clinical trial.
Transparency
Declaration of funding
PAREXEL was commissioned by AstraZeneca PLC to conduct
this research and develop the manuscript.
Declaration of financial/other relationships
JP and RH are employees of PAREXEL Consulting. JB is an
employee of AstraZeneca PLC.
Acknowledgments
The authors would like to thank Gaurav Suri (PAREXEL) and
Stuart Spencer (AstraZeneca) for their work in helping to
develop the olaparib economic model, and David Parry
(AstraZeneca) for his helpful comments on earlier drafts of this
paper.
References
1. World Health Organization (WHO). 2014. Cancer, Fact Sheet no. 297.
(published on-line) www.who.int/mediacentre/factsheets/fs297/en/.
Accessed September 2014
2. International Agency for Research on Cancer. GLOBOCAN 2012: Estimated
Cancer Incidence, Mortality and Prevalence Worldwide in 2012. Paris,
France: WHO, International Agency for Research on Cancer. Published
2014. http://Globocan.iarc.fr/Pages/fact_sheets_cancer.aspx. Accessed
September 2014
524 Challenges in economic modeling of anti-cancer therapies Hettle et al.
3. United States Food and Drug Administration (FDA). Fast track, breakthrough
therapy, accelerated approval and priority review. Philadelphia, United
States. http://www.fda.gov/ForConsumers/ByAudience/ForPatientAdvocates/.
Accessed September 2014
4. European Medicines Agency. Guideline on the procedure for accelerated
assessment pursuant to Article 14(9) of Regulation (EC) no 726/2004.
London: EMA; 2006. www.ema.europe.eu. Accessed September 2014
5. Califf RM, Zarin DA, Kramer JM, et al. Characteristics of clinical trials regis-
tered in ClinicalTrials.gov, 2007-2010. JAMA 2012;307:1838-47
6. Johnson JR, Ning Y-M, Farrell A, et al. Accelerated approval of oncology
products: The Food and Drug Administration Experience. J Natl Cancer Inst
2011;103:1-9
7. Ledermann J, Harter P, Gourley C, et al. Olaparib maintenance therapy
in platinum-sensitive relapsed ovarian cancer. N Eng J Med 2012;366:
1382-92
8. Ledermann J, Harter P, Gourley C, et al. Olaparib maintenance therapy in
patients with platinum-sensitive relapsed ovarian cancer: a preplanned retro-
spective analysis of outcomes by BRCA status in a randomized phase 2 trial.
Lancet Oncol 2014;15:852-61
9. Chabot I, Rocchi A. Oncology drug health technology assessment recommen-
dations: Canadian versus UK experiences. Clin Econ Outcomes Res
2014;6:357-67
10. Armstrong D. Relapsed ovarian cancer: challenges and management strate-
gies for a chronic disease. Oncologist 2002;7(5 Suppl):20-8
11. European Medicines Agency. Guideline on the evaluation of anticancer
medicinal products in man (report and Appendix 1). London: EMA, 2011.
www.ema.europe.eu. Accessed September 2014
12. Latimer NR, Abrams KR. NICE DSU Technical Support Document 16:
adjusting survival time estimates in the presence of treatment switching.
School of Health and Related Research, University of Sheffield, Sheffield,
UK. 2014
13. Tappenden P, Chilcott J, Ward S, et al. Methodological issues in the economic
analysis of cancer treatments. Eur J Cancer 2006;42:2867-75
14. Pemetrexed maintenance treatment following induction therapy with peme-
trexed and cisplatin for non-squamous non-small-cell lung cancer. National
Institute for Health and Care Excellence (NICE) technology appraisal guidance
309. London, UK: National Institute for Health and Care Excellence, 2014.
guidance.nice.org.uk/ta309. Accessed September 2014
15. Erlotinib monotherapy for maintenance treatment of non-small-cell lung
cancer. NICE technology appraisal guidance 227. London, UK: National
Institute for Health and Care Excellence, 2011. guidance.nice.org.uk/ta227.
Accessed September 2014
16. Rituximab for the first-line maintenance treatment of follicular non-Hodgkins
lymphoma. NICE technology appraisal guidance 226. London, UK: National
Institute for Health and Care Excellence, 2011. guidance.nice.org.uk/ta226.
Accessed September 2014
17. R Core team. R: a language and environment for statistical computing.
Vienna, Austria: R Foundation for Statistical Computing, 2013. http://
www.R-project.org/. Accessed September 2014
18. Bevacizumab in combination with gemcitabine and carboplatin for treating the
first recurrence of platinum-sensitive advanced ovarian cancer. NICE tech-
nology appraisal guidance 285. London, UK: National Institute for Health and
Care Excellence, 2013. guidance.nice.org.uk/ta285. Accessed September
2014
19. Latimer N. NICE DSU technical Support Document 14: Undertaking survival
analysis for economic evaluations alongside clinical trials extrapolation with
patient-level data. 2011
20. Hirst C, Parry D, Alsop K, et al. Survival in patients with BRCA mutation-
positive platinum-sensitive recurrent ovarian cancer. J Clin Oncol
2014;32:(Suppl abstr e16519). http://meetinglibrary.asco.org/content/
131746-144. Accessed September 2014
21. Trabectedin for the treatment of relapsed ovarian cancer. NICE technology
appraisal guidance 222. London, UK: National Institute for Health and Care
Excellence, 2011. guidance.nice.org.uk/ta222. Accessed September 2014
www.informahealthcare.com/jme ! 2015 Informa UK Ltd