Modelling Post Progression Utility
Data for Economic Evaluation of
Cancer Drugs
•
•
•
Iftekhar Khan
Senior Lecturer in Medical Statistics & Research Methodology, King’s College
London
Senior Research Fellow in Health Economics (Warwick Evidence / Evidence
Review Group), University of Warwick
Consultant Statistician
All views expressed here are my own
© 2017
Objectives
1. Health Economic Evaluation
2. Generic and Condition Specific HRQoL in
Cancer
3. Measuring Utility in Cancer Trials
4. Measuring EQ-5D after cancer progression
5. The behaviour of post-progression utility
6. Modelling approaches
7. Example
8. Conclusions and Future Direction
1. Health Economic Evaluation
• An estimated 14.1 million new cases of cancer were
identified across the world in 2012.
• 32.5 million people die within five years of diagnosis.
• In 2008, about 170 million years of healthy life was lost
globally because of cancer
• Resulting economic impact likely to run into hundreds
of Billions of pounds.
1. Health Economic Evaluation
• Healthcare resource is finite and scarce (Morris, 2006; Santerre &
Neun, 2009).
• Constraint on health care resource: results in governments taking a
hard look at all public expenditure, including the medicines budget.
• Patients can’t get access to treatments (weakens the reason for
licensing)
• Demonstrating value of a drug is now considered necessary for
patient access
• To demonstrate value , need to show it is cost-effective
1. Health Economic Evaluation
• To demonstrate value (cost-effectiveness), in the UK, the National
Institute of Health and Care Excellence (NICE) require the ratio of
differences in costs to differences in benefit to be above £20 - £30 K and
in exceptional cases £50k
• This relative benefit is called ICER: Dc/De < l [Dc = mean difference in
costs, De = mean difference in effectiveness and l is the CE threshold or
WTP]
• We are concerned here with improving estimates of De
• De comprises of measures of survival time and quality of life (HRQoL).
This forms the QALY
• Very few situations where cancer drugs have QALYs <£20-£30K per QALY
Treatment
Cost (£)
QALY
Cost/QALY (£)
Year
Source
Paclitaxel
28210
0.53
53227
2011
Goulart et al., Value in Health [51]
27902
0.923
30230
2010
Brown et al., Health Tech Assess [52]
21967
NR
NR
2000
Berthelot et al., JNCI [53]
24216
NR
NR
2000
Berthelot et al., JNCI [53]
26228
NR
NR
2000
Berthelot et al., JNCI [53]
33685
0.4513
74639
2009
Klein, R; J., Thoracic Oncology, Vol 4 [21]
27837
0.934
29804
2010
Brown et al., Health tech Assess [52]
27401
0.966
28365
2010
Brown et al., Health tech Assess [52]
18129
NR
NR
2000
Berthelot et al., JNCI [53]
47876
1.96
24427
2013
Wang et al., PLSone 8(3) [54]
38859
0.4676
83102
2009
Klein, R; J., Thoracic Oncology, Vol 4 [21]
23516
0.888
26482
2010
Brown et al., Health tech Assess [52]
16678
NR
NR
2000
Berthelot et al., JNCI [53]
17482
NR
NR
2000
Berthelot et al., JNCI [53]
6901
NR
NR
2010
Maniadakis: Annals of Oncology[55]
4129
0.1606
25712
2012
Thongprasert et al. [56]
13956
0.206
67748
2010
Lewis et al., [57]
27409
0.42
65260
2010
Asukai et al., [58]
24798
0.225
110215
2008
Araujo et al., Rev Port Pneumol [59]
24904
0.42
59296
2008
Carlson, Lung Cancer Sep 61(3) 405 [60]
11622
0.42
27672
2011
Vergnenge et al., J. Thoracic Oncology[61]
20903
NR
NR
2011
Cromwell et al., J. Thoracic Oncology [62]
Gemcitabine
Vinorelbine
Docetaxel
Gefitinib
Erlotinib
29387
0.52
56514
2010
Asukai et al., (2010, BMC, Cancer) [58]
27764
0.241
115205
2008
Araujo et al., Rev Port Pneumol. 2008 [59]
37119
0.41
90533
2008
Carlson Lung Cancer Sep 61(3) 405 [60]
14239
0.41
34729
2011
Vergnenge et al., J. Thoracic Oncology [61]
17455
0.97
17995
2010
Greenhalgh et al. [63]
41731
0.5016
83195
2009
Klein, R; J. Thoracic Oncology, Vol 4 [21]
8905
0.41
21720
2012
Fragoulakis, Lung Cancer. [64]
3973
0.1745
22766
2012
Thongprasert et al. (2012, J clin Oncol) [56]
NR
1.111
NR
2010
Brown et al., Health Tech Assess [52]
19787
0.79
25047
2013
Zhu, MC Cancer [65]
7704
0.79
9752
2013
Zhu, MC Cancer [65]
28471
0.91
31287
2012
Gilberto de lima lopes, 2012 [66]
8980
0.2881
31170
2010
EGFR Testing, Ontario Health [67]
10536
0.3188
33048
2010
EGFR Testing, Ontario Health [67]
13730
0.238
57689
2010
Lewis et al., [57]
22439
0.25
89756
2008
Araujo et al., Rev Port Pneumol. 2008 [59]
23567
0.42
56112
2008
Carlson, Lung Cancer Sep 61(3) [60]
5286
0.1745
30292
2012
Thongprasert et al. (2012,J Clin Oncol) [56]
25546
1.4
18247
2013
Wang et al., PLSone 8(3) [54]
Health Economic Evaluation Methods for Cancer Trials using Clinical Trial and Real
World Data (Khan, Crott & Bashir, Chapman & Hall , 255 pages; due 2017)
2. HRQoL in Cancer
• Consider HRQoL
• HRQoL consists of generic and condition specific measures
Examples:
QLQC-30
HAQ
KHQ
Examples:
LCSS
FACT-O
(Very
condition
specific)
HRQoL Instruments
Condition
Specific
HRQoL
Very
Condition
Specific
HRQoL
Generic
HRQoL
Preference
Based HRQoL:
EQ-5D
HUI
SF-6D
Non
Preference
Based
SF-36
• EQ-5D-3L and EQ-5D-5L , 3 point or 5 point scale, most widely used generic measure
• EQ-5D is a measure of preference for a given ‘health state’
2. HRQoL in Cancer
• Health State: Anxiety =1, Mobility=1, Self-Care=1, Usual Activities =1
and Pain=1 forms the health state 11111
• Combinations of health states are associated with utilities
• 11111 is associated with a utility of 1 (or Full health), 33333 is a state
worse than death (-0.549)
• These utilities are used to adjust clinical outcomes to construct a
quality adjusted survival time (quality adjusted life year, QALY)
• EQ-5D-5L more, or as sensitive to detect treatment benefit compared
to the QLQ-C30’s 15 domains (Khan et al 2017)
2. HRQoL in Cancer
• To construct the QALY in cancer
• QALYAUC = S [(Qi +Qi+1)/2] * [(Si +Si+1) /2 ]*[ti+1- ti]
• where, Qi are the utility measures at time point i, and Si are the
corresponding survival rates.
2. HRQoL in Cancer
• EQ-5D not intended to be a measure of treatment benefit but as a way
showing public (tax payer) preference for a health state.
• However the interpretation given is as though it is a measure of clinical
benefit (Which is why sensitivity issues were raised against it)
• Some have suggested EQ-5D is also a measure of clinical benefit and
may be just as valid as other condition specific measures (like the QLQC30)
• Can contradict (EQ-5D says one thing and condition specific measure
concludes the opposite – or vice versa)
3. Measuring Utility in Cancer Trials
• Cancer trials (assume phase III confirmatory) HRQoL data collected
• However, disease progression (PD) might occur
• Patient leaves trial
• Patient dies
• Toxicity
• EQ-5D is the vehicle used to collect utility
• After PD, this data is often not collected
• Patient too ill
• Leaves the protocol follow up
• Collection until death not possible (too long)
• From a clinical perspective we may be interested in HRQoL benefits over a
short-time horizon
• For cost-effectiveness, interested in the life time costs and benefits
attributable to the experimental drug
3. Measuring Utility in Cancer Trials
• Why do we need utility data for cancer trials ?
• NICE (and some others) want them
• To demonstrate value – deriving QALYs.
• QALYs used across disease areas so comparative value to the tax
payer assessed (e.g. Dementia, Cancer, Heart Disease etc)
• Utility data collected from baseline until disease progression (PD)
typically
• Measurements after PD not common and restricted to essential
follow up (longer term toxicity – e.g. RT)
4. Measuring Utility after PD
• Why do we need Post progression utility data for cancer trials ?
• In cancer trial economic modelling is often carried out using
partitioned survival models: OS = PFS + PPS
• Would like to have a better estimate of the AUC presenting PPS
• Impacts the QALY and consequently the ICER / Cost per QALY
• Sometime s novel treatments are considered to be delivered
beyond PD and longer term value also needs to be assessed
(e.g. Nivolumab / pseudo progression)
4. Measuring Utility after PD
• Options for using utility after PD appear to be:
a) Use historical/published data (e.g. Nafees et al, 2008)
b) Proceed to collect data after PD in all patients
c) Proceed to collect data after PD in some patients
(sparse sampling)
d) Mapping methods
e) Modelling and Extrapolation
4. Measuring Utility after PD
a) Use historical/published data (e.g. Nafees et al, 2008)
• Aggregate estimates (not patient level)
• Different populations
• Different treatments
• No AUC (mean estimate)
• Varying follow up times of patients
• …..but have been referred to repeatedly as a source in several recent
HTA ERG’s (e.g. Nivolumab)
• Just a form of poor extrapolation
5. Behaviour of Post-Progression Utility
Utility
Time
-0.549
5. Behaviour of Post-Progression Utility
• Conducted an observational study in newly
diagnosed lung cancer patients (n=98) [Khan et al,
2015]
• Assessments made monthly (EQ-5D-3L, EQ-5D-5L,
QLQ-C30, SF-36, OS, PFS)
• Patients taking routine cancer treatment
• HRQoL assessed until 2 years or death (whichever
occurred first)
5. Behaviour of Post-Progression Utility
6. Modelling Approaches
• Given the behaviour of the post progression (PP)
utility , several models (and approaches) considered
to fit extrapolate utility after PD
• Extrapolation using parametric models acceptable
for survival data
• Consider feasibility of fitting similar models to utility
data for estimating PPS AUC:
AUCOS = AUCPFS + AUCPPS
6. Modelling Approaches
- Assume the Utility declines linearly (at a constant rate) over time
- Fit a simple Linear (GLM) model
- Predict utilities
- Use all data (pre and post progression) or use whatever post progression data is available
Progression
Last observed utility
Utility
Extrapolated Utility
Time
-0.549
6. Modelling Approaches
6. Modelling Approaches
• Given the behaviour of the post progression (PP)
utility , several models (and approaches) considered
to fit extrapolate utility after PD
• Extrapolation using parametric models acceptable
for survival data
• Consider feasibility of fitting similar models to utility
data for estimating PP AUC:
AUCOS = AUCPFS + AUCPPS
6. Modelling Approaches
Method
Survival Data
Utility
Comment
(i)
No Extrapolation
No Extrapolation
Use utilities from published data
All Patients died (no need to extrapolate survival)
Utilities may be extrapolated between PD and Death
(ii)
Extrapolated
No Extrapolation
(iii)
Extrapolated
Extrapolated
Need estimates of utility between PD and Death
Use published data or extrapolate
Both survival and utility extrapolated
AUCOS = AUCPFS + AUCPPS
extrapolated
observed
Modelling can focus on AUCos or AUCPPS , treating AUCPFS as observed
6. Modelling Approaches
Utility
Time Post Progression
6. Modelling Approaches
EQ-5D = a + b*exp{-g(X-d)2}.
EQ-5D = 1 – 1/Xa
Model
Functional form
Linear decline
Exponential decline
Bragg-Packer (1962)
Pareto
Beta
Lorentz (1979)
Rational Function
5-Parameter
Y = a+ b *X
Y = exp(-lt)
 + *exp{-X-)2}.
1 – 1/X
X*(1-X)
 +  Time -)2]
  *Time)/(1+Time+*Time2)
  *Time +*Time2/ (1+Time+*Time2)
EQ-5D = aXb*(1-X)g
7. Example
N=98
No PD, Alive (n=14)
PD, Alive (n=57)
Died, No PD (n=7)
PD and Died (n=20)
PD: progressive Disease
For modelling, need those alive with PD (n=57) + those with PD and
Died (n=20) + = 77 patients (possibly add No PD and alive)
Figure 9.11: Categories of Events in Study 3 (Vertical Line Separates Pre and Post-Progression Period)
7. Example
Month
Baseline*
Mean Observed
Pre-progression
Utility [n]
0.6399
Mean Observed
Post-progression
Utility [n]
0.639954
1
0.64961
0.552166
2
0.64024
0.544205
3
0.60999
0.518493
4
0.64784
0.550661
5
0.64299
0.546538
6
0.60438
0.513719
7
0.55711
0.473531
8
0.71307
0.574423
9
0.41315
0.351181
10
0.54014
0.435588
11
0.41743
0.337865
12
0.35723
0.303646
Overall
Mean
0.566097
0.474095
Figure 9.9: Survival Rates for OS, PFS and PPS for Study 3 Data
7. Example
Model PPS and
Extrapolate until
24 months
RP 3 Knots
Using SAS Macro published for flexible parametric models (Dewar & Khan 2016)
7. Example
Model
Linear
parameter


Estimate
0.654
-0.0219
p-value
<0.001
Equation
0.654 -0.0219*Time
AIC
115.7
Exponential
Bragg-P
l








0.1415
-0.08323
0.693710
0.005245
1.697112
0.4073
0.9595
0.1270
1.3985
<0.001
0.983
0.8617
0.8990
0.6575
<0.001
0.0002
0.1440
0.0059
Exp(-0.1415* Time)
 + *exp{-*( Time -)2}
198.0
106.3
1-[1 – 1/Time0.4073]
 Time *(1-Time)
253.4
101.2
Lorentz




-0.4832
1.0939
0.00532
1.77192
0.968
0.927
0.948
0.784
 + /[1+ *(Time -)2]
106.2
Rational
Function




-0.04831
0.5997
-0.01726
0.005538
Five Parameter





-0.01618
0.5985
0.1000
0.003861
-0.00148
Pareto
Beta
0.5562 ( + *Time)/(1+*Time+*Time2)
<.0001
0.8670
0.7464
0.7419
<.0001
<.0001
0.8397
0.7431
( + *Time +*Time2/ (1+*Time+*Time2)
106.4
92.4
7. Example
Predicted from PP to 12 months
Extrapolated until 24 months PP
7. Example
Month PP
PPS1
Linear
Exponential
Bragg-P
Pareto
Beta
Lorentz
Rational
5-parameter
1
2
3
4
5
6
7
8
9
10
11
12
100%
89%
80%
71%
69%
63%
55%
55%
55%
55%
49%
43%
0.6323866
0.6104600
0.5885334
0.5666067
0.5446801
0.5227535
0.5008269
0.4789003
0.4569736
0.4350470
0.4131204
0.3911938
0.8680552
0.7535198
0.6540968
0.5677921
0.4928748
0.4278426
0.3713909
0.3223878
0.2798504
0.2429256
0.2108728
0.1830493
0.6087141
0.6101463
0.6043310
0.5914498
0.5719015
0.5462812
0.5153507
0.4800006
0.4412078
0.3999912
0.3573677
0.3143117
1.0000000
0.6699382
0.5299948
0.4488172
0.3945166
0.3550638
0.3248014
0.3006798
0.2808945
0.2643017
0.2501377
0.2378708
0.6035076
0.6193203
0.6109764
0.5919137
0.5667773
0.5378116
0.5063101
0.4731083
0.4387921
0.4038001
0.3684792
0.3331174
0.6077124
0.6097999
0.6042279
0.5912281
0.5713266
0.5452932
0.5140697
0.4786924
0.4402168
0.3996548
0.3579274
0.3158357
0.6082953
0.6104777
0.6053164
0.5924955
0.5722454
0.5453433
0.5130163
0.4767720
0.4382009
0.3987983
0.3598385
0.3223101
0.5783413
0.5606051
0.5550958
0.5720147
0.5519026
0.5255827
0.5140768
0.5421457
0.4000281
0.3997085
0.3285215
0.3172918
13*
14*
15*
16*
17*
18*
19*
20*
21*
22*
23*
24*
36%
33%
32%
29%
29%
25%
23%
20%
18%
15%
13%
11%
0.3810000
0.3600000
0.3390000
0.3180000
0.2970000
0.2760000
0.2550000
0.2340000
0.2130000
0.1920000
0.1710000
0.1500000
0.1588969
0.1379312
0.1197319
0.1039339
0.0902204
0.0783163
0.0679828
0.0590129
0.0512264
0.0444674
0.0386001
0.0335070
0.2717202
0.2303847
0.1909714
0.1540096
0.1198883
0.0888597
0.0610497
0.0364715
0.0150440
-0.0033898
-0.0190427
-0.0321654
0.2271183
0.2175969
0.2090917
0.2014369
0.1945018
0.1881820
0.1823931
0.1770658
0.1721430
0.1675768
0.1633268
0.1593588
0.2979837
0.2632678
0.2292018
0.1959969
0.1638710
0.1330590
0.1038273
0.0764966
0.0514822
0.0293806
0.0112157
~0
0.2740473
0.2330954
0.1933870
0.1552166
0.1187826
0.0842040
0.0515365
0.0207869
-0.0080745
-0.0351042
-0.0603764
-0.0839771
0.2869069
0.2540558
0.2239648
0.1966766
0.1721177
0.1501403
0.1305537
0.1131479
0.0977094
0.0840311
0.0719187
0.0611935
0.2763705
0.2368746
0.1988992
0.1627003
0.1284338
0.0961758
0.0659408
0.0376981
0.0113846
-0.0130844
-0.0358072
-0.0568892
4.648132
28.25819
3.747115
21.88337
4.160589 3.732875
18.89631 32.44121
4.304773
21.69392
4.028435
19.02141
PP QALY#
%Extrapolated@
Extrapolated
4.272484 4.135935
21.13157 18.50418
7. Example
• The extrapolated PP AUC for the linear model gives a higher QALY, but as a
proportion of the total QALY is more uncertain (28% of the overall QALY)
• Complex 5 parameter models seems to predict observed PP Utility (based on AIC)
• Suggest use model with the least extrapolation as a percent of total QALY
• Other models could be a sensitivity check on the QALY estimate
• In health economic evaluations of cancer treatments, the extrapolated QALY (AUC)
due to survival can be extensive (>50%) as in TA370 (2012) for Bortezamib in
Lymphoma patients.
• Several HTA submissions (e.g. Vinflunine for urothelial tract cancer, 2013), TA272 and
TA366 , TA272 and Nivolumab (2016) use estimates of PP utilities from a single study
(Nafees et al., 2008); for TA272 for melanoma (Nov, 2015), modelling postprogression utilities were not considered at all.
• Nafees (2008) is not based on cancer patients , but preferences of health states
common to cancer, from a general population
8. Conclusions, Limitations and Future Direction
• Feasible to extrapolate utilities
• Use patient level data
• Can make adjustment for patient level covariates
• Better than using historical data ? : comparison with those of Nafees
(2008) and others (Lloyd, 2006) ongoing – [Khan, 2017]
• Estimates are not unreasonable (simulations : Khan et al, 2017)
• Limitations/ Future :
• Need more RCT data to test these models (ongoing) to test impact on
ICER and decision impact
• Need to consider treatment switching/crossover: impact on Utility
(currently only considered for survival)
• Modelling the entire profile, from time zero and not just from
progression (criticism: why model when you have the observed utilities
prior to PD, same is true for survival models - more uncertainty?)