karnon et al, modeling assumptions impact analysis...karnon et al, modeling assumptions impact...
TRANSCRIPT
Karnon et al, Modeling assumptions impact analysis
1
A critique and impact analysis of decision modeling assumptions
(running head: Modeling assumptions impact analysis)
Jonathan Karnon, PhD, School of Health and Related Research, University of
Sheffield
Alan Brennan, MSc, School of Health and Related Research, University of Sheffield
Ron Akehurst, Hon MFPHM, School of Health and Related Research, University of
Sheffield
Conflict of Interests: The School of Health and Related Research at The University of
Sheffield, in which the authors are based, developed a model of the cost-
effectiveness of clopidogrel to be included in a submission to the National Institute for
Clinical Excellence on behalf of Bristol-Myers Squibb and Sanofi-Synthelabo. The
work described in this manuscript has not been funded by these companies, none of
the authors is in personal receipt of payment from these companies, and the
companies have had no editorial or drafting input into this manuscript.
Correspondence: Jonathan Karnon, School of Health & Related Research, University
of Sheffield, Regent Street, Sheffield, S1 4DA
Email: [email protected]
Karnon et al, Modeling assumptions impact analysis
2
A critique and impact analysis of decision modeling assumptions
Abstract
Background: Numerous guidelines have been published defining good practice for
the conduct of economic evaluations in general, and model-based evaluations in
particular. The extent to which guidelines are accepted is unknown, and the impact of
deviations from good practice is not generally recorded. We identified four specific
issues in applied studies that may impact on the accuracy and comparability of
different evaluations.
Methods: A descriptive analysis of four modeling issues (inclusion of incident cases
over a model time horizon, appropriate time horizon, parsimonious model structure,
and the handling of age-specific sub-groups) is presented. A case study model is
analysed to illustrate the quantitative impact of three of the issues.
Results: In the case study model, alternative specifications of the modeling
framework are shown to alter the estimated cost-effectiveness by large percentages.
The combined effect of including incident cases and reduced follow-up yielded the
highest divergence from the reference case results, by between 20% and 40%
depending on the age group. Reference case results of an age-weighted population
were almost 14% different from the middle single age cohort.
Discussion: The identified issues are all generalisable to a wide range of treatment
areas and are, or should be, addressed by evaluative guidelines. The authors call for
the continued development, dissemination, and application of guidelines for the
conduct of economic evaluation in general, and model-based economic evaluations
in particular.
Keywords: Guidelines; Models, Decision Support; Cost-effectiveness; Antiplatelet
Drugs
Karnon et al, Modeling assumptions impact analysis
3
A critique and impact analysis of decision modeling assumptions 1
Introduction 2
The estimation of the full economic effects of health technologies generally requires 3
the extrapolation of clinical trial evidence beyond the period of follow-up through the 4
use of decision modeling techniques to synthesise data from disparate sources. 5
Numerous guidelines have been published detailing appropriate conduct for the 6
application of modeling evaluations, which have been the subject of a methodological 7
systematic review.[1] The review recognises that methodological aspects of the 8
modeling process may be explicitly defined in the form of guidelines, though such 9
checklists cannot be expected to inform the appropriateness of issues such as the 10
specification of the model structure. 11
Through general use and review of the health economics literature we have identified 12
a number of specific issues relating to the application of decision modelling as a 13
framework for the estimation of the long-term costs and effects of alternative health 14
care interventions. The primary issue concerns the inappropriate inclusion of incident 15
cases over the time horizon of an evaluation, which was identified in an application of 16
the Coronary Heart Disease Policy (CHDP) model,[2] to analyse the cost-17
effectiveness of alternative antiplatelet therapies.[3] Secondary issues emanating 18
from our investigation include the use of an appropriate time horizon, a parsimonious 19
model structure, and the handling of (age-specific) sub-groups. This paper describes 20
the theoretical aspects of each of the above issues and quantifies the impact of 21
applying alternative approaches to three of the issues (an impact analysis of the 22
parsimony issue would require a multi-modelling centre approach that was 23
considered infeasible for this study). The aim is to highlight the need for the common 24
application of guidelines that address these specific modeling issues. 25
Karnon et al, Modeling assumptions impact analysis
4
Identifying the issues 26
The following sections describe the basis for our assessment of the four identified 27
issues: inclusion of incident cases over the time horizon of an evaluation, appropriate 28
time horizon, a parsimonious model structure, and the handling of age-specific sub-29
groups 30
Incident cases 31
Gaspoz et al used the CHDP model to estimate the incremental cost-effectiveness of 32
aspirin, clopidogrel, or combined aspirin plus clopidogrel for the secondary 33
prevention of coronary heart disease.[3] It is stated that the evaluation ‘modelled U.S. 34
patients, 35 to 84 years of age, in whom coronary disease developed during or 35
before 2003 to 2027 and who survived their first month with it’. 36
The evaluation configured the CHDP model to predict the costs and QALYs 37
associated with CHD in the US population over a 25 year period. This means that 38
incident patients are added to the model every year over the 25 year period, but that 39
the model follows all patients only to the end of the 25 year period, i.e. patients 40
diagnosed in year 20 are followed up for 5 years, irrespective of their age at the end 41
of the follow-up period. This approach provides an incomplete analysis of the full 42
impact of the interventions being evaluated, for which the downstream costs and 43
benefits of avoiding secondary cardiovascular events in the short-term are excluded. 44
If incident cases are to be included in a model-based evaluation, then the time 45
horizon of the evaluation should be extended such that the costs and effects are 46
measured over the same timeframe (i.e. to the same upper age limit) for every 47
individual that enters the model. If this rule is implemented, in most treatment 48
evaluations the modelling of incident cases will affect only the absolute cost and 49
effect estimates as the relative cost-effectiveness of alternative interventions will 50
remain the same. Exceptions will occur in cases in which the characteristics of the 51
Karnon et al, Modeling assumptions impact analysis
5
patient population vary over time, such as changes in the age distribution or disease 52
severity. Though, if identifiable characteristics are shown to significantly alter cost-53
effectiveness results then sub-group analyses may be relevant. 54
Time horizon 55
A commonly identified area of variation between modelling studies is the defined time 56
horizon of models, which describes the maximum age to which individuals are 57
followed. In the applied CHDP model,[3] the choice of a 25 year time horizon for the 58
evaluation of a population aged between 35 and 84 years may also be questioned. 59
The youngest individuals entering the model are followed up to the age of only 60 60
years, and even individuals aged 50 years (an age at which CHD incidence is 61
considerable) may gain significant QALYs after the age of 75 years. 62
There are numerous examples of the specification of non-lifetime horizons in models 63
evaluating interventions in a wide range of disease areas.[4,5] The impact on the 64
accuracy of the cost-effectiveness results will vary, though it will increase in line with 65
decreasing mortality rates, time horizons, and discount rates. 66
Parsimony 67
Weinstein et al advise that “[t]he structure of the model should be as simple as 68
possible, while capturing underlying essentials of the disease process and 69
interventions”,[pg12] in their report on Principles of Good Practice for Decision 70
Analytic Modeling.[15] 71
The CHDP model aimed to estimate the health effects and costs of coronary heart 72
disease in the US population.[2] This is a complex state transition model comprising 73
three linked models describing the incidence of coronary disease in the unaffected 74
population, the first 30 days of an initial coronary event, and the disease history of 75
patients who survive the first 30 days following a coronary event. The final section of 76
the CHDP model (the Disease History model) includes four health states in months 77
Karnon et al, Modeling assumptions impact analysis
6
2-12 following diagnosis (angina, acute myocardial infarction (AMI), cardiac arrest, or 78
AMI plus cardiac arrest). Eight states describe the subsequent natural history, 79
representing patients’ coronary disease history as alternative combinations of the 80
experience of three events: cardiac arrests, AMIs, and revascularisation in the form 81
of a coronary artery bypass graft (CABG). Patients not experiencing any of these 82
events have, by default, angina. From these eight states patients are assigned 83
annual transition probabilities of experiencing an AMI, cardiac arrest, CABG, or non-84
coronary death. 85
The evaluation of antiplatelet therapies using the CHDP model illustrates the link 86
between unnecessary model complexity and potential modelling errors. Firstly, as a 87
result of identified errors the authors’ were required to revise the originally published 88
results.[6] In particular, one error occurred in the estimation of the benefits in 89
reducing the risk of stroke, an area that was not covered by the original CHDP model 90
and was incorporated into the model in the form of a derived reduction in the overall 91
rate of death from non-coronary causes. 92
Secondly, a lifetime duration of clopidogrel therapy is assumed in the original Gaspoz 93
analysis, though in response to de Lemos and McGuire,[7] Gaspoz et al specify a 94
therapy option of clopidogrel plus aspirin for all patients in the 1st year post-CHD 95
diagnosis. The results of the short-term treatment option are counter-intuitive, 96
showing that Clopidogrel is less cost-effective when treatment is confined to the first 97
year post-diagnosis – the time period over which patients are at increased risk of 98
further vascular events. Gaspoz et al do not discuss the foundations of this counter-99
intuitive result, and one cannot discount the possibility that a more parsimonious 100
model, describing only the relevant events and providing more flexibility in the 101
representation of the transition probabilities, would have produced alternative results. 102
Karnon et al, Modeling assumptions impact analysis
7
Sub-group analysis 103
Dewilde & Anderson present a very interesting paper that shows the importance of 104
incorporating the age distribution of the eligible population when modelling the cost-105
effectiveness of screening, rather than simply modelling the pathways of patients 106
from the age at which they first become eligible for screening.[8] Using cervical 107
cancer screening as a case study, they show that the standard approach of 108
modelling a single screening cohort (from age 15 years) and an age-weighted 109
approach in which costs and effects are estimated for 11 separate age-based cohorts 110
result in mean incremental cost-effectiveness ratios of Aus$47,361 and Aus$61,031, 111
respectively (discounted at 5%). 112
The single cohort approach estimates the cost-effectiveness of screening once the 113
screening programme reaches equilibrium, i.e. all eligible women were first offered 114
screening at age 15 years. In the case of cervical cancer screening, equilibrium 115
would not be reached for 45 years based on five year intervals between age cohorts 116
and a final screening age of 65 years. An alternative approach would be to treat the 117
different age cohorts as sub-groups of the eligible population, with the possibility of 118
defining age-based screening policies. This latter approach has been adopted in 119
various model-based screening evaluations.[9,10] 120
The findings of Dewilde & Anderson[8] are likely to be applicable to treatment 121
evaluations, particularly in areas such as CHD where event rates have been shown 122
to vary significantly with respect to age.[11] Where age-based sub-groups are 123
feasible from a policy implementation perspective, separate analyses by age should 124
be undertaken. If it is not possible to implement age-specific policy 125
recommendations, the age distribution of the eligible population should be 126
represented, and the approach presented by Dewilde & Anderson provides a suitable 127
method for undertaking such analyses. 128
Karnon et al, Modeling assumptions impact analysis
8
Modeling the impact of alternative approaches 129
A comparison of the use of a parsimonious versus a non-parsimonious model 130
structure is difficult to undertake and require the application and independent 131
analysis of multiple models as the hypothesised advantages of parsimony relate to a 132
reduced probability of error. However, the three other issues raised above may be 133
investigated further by the application of alternative methods of analysis to a single 134
case study model, which is used to demonstrate the impact of including incident 135
cases over the course of the specified time horizon, and the use of a short time 136
horizon relative to expected survival. The case study model is also analysed to 137
estimate the impact of modelling alternative age-based patient sub-groups: 138
- a single age cohort based on the mean age of the patients informing the input data, 139
- multiple age cohorts with the results presented separately for each cohort, 140
- multiple age cohorts with the results aggregated based on the age distribution of the 141
eligible population. 142
Methods 143
The case study model is one of our own previously published models incorporating a 144
wide range of age-specific data. The chosen model evaluates the cost-effectiveness 145
of one year of clopidogrel in addition to aspirin compared to aspirin alone in patients 146
with newly diagnosed acute coronary syndromes (ACS).[11] 147
A decision tree covering the one-year treatment period describes patient pathways 148
with respect to alternative forms of revascularisation over the one-year treatment 149
period (Figure 1). Major bleeding event rates are linked to the incidence of coronary 150
artery bypass graft (CABG), and percutaneous coronary intervention (PCI) rates. The 151
summed proportions of patients in each of the end states of the decision tree (event 152
free, non-fatal MI, non-fatal stroke or vascular death) inform the distribution of 153
patients across the start states in a lifetime Markov model. The Markov model 154
Karnon et al, Modeling assumptions impact analysis
9
describes the annual probability of patients experiencing a MI, stroke, or death from 155
vascular or non-vascular causes (Figure 2). Following the experience of either a MI 156
or stroke, patients may either remain in that state or progress to the death state. 157
Alternative states are used to describe mortality rates for patients in the first year of a 158
MI or stroke event, and for patients who have survived their first year following a MI 159
or stroke. 160
Data sources used to populate the model are described in more detail elsewhere.[11] 161
In summary, revascularisation and bleeding event rates were based on aggregate 162
(non-age-specific) analysis of prospective and retrospective studies of ACS patients 163
in the UK.[12,13] Patient-level data from the Prospective Registry of Acute Ischaemic 164
Syndromes in the UK[12] and the Nottingham Heart Attack Register (personal 165
communication D Gray) were available for analysis to inform age-specific event rates 166
for ACS patients from diagnosis, as well as following the experience of an MI. The 167
estimated event rates accounted for higher event rates in the period immediately 168
following an event by undertaking logit regression analysis to estimate initial event 169
rates, and exponential survival analysis to estimate constant longer term event rates 170
(time varying distributions were tested and were not found to be statistically 171
significantly different). General population mortality rates, minus age-specific 172
proportion of deaths due to vascular causes, were also applied in the model. 173
No data were identified describing subsequent event rates in ACS patients 174
experiencing stroke. Data are available describing event rates in all stroke patients 175
(i.e. not limited to stroke patients previously diagnosed with ACS),[13] but these rates 176
are unrealistically low in comparison to the observed event rates in patients 177
remaining event-free following ACS. A conservative assumption is made that ACS 178
patients experiencing stroke during the one-year treatment period have the same 179
subsequent event rates as ACS patients remaining event-free. 180
Karnon et al, Modeling assumptions impact analysis
10
Treatment effectiveness was represented as the relative risk (RR) of four clinical 181
outcome events - vascular death, non-fatal MI, non-fatal stroke, and major bleeding. 182
The RRs were obtained from the CURE trial, which randomised 12,562 patients 183
within 24 hours of the onset of non-ST-segment-elevation ACS to standard therapy 184
including aspirin (75-325 mg), or to clopidogrel (300 mg loading dose and then 75 mg 185
daily) on top of standard therapy including aspirin.[14] Cost and utility data were 186
obtained from published sources and are described in Table 1, as are the RRs. 187
Model analyses 188
The reference case analysis of the model is defined to exclude incident cases, 189
following a patient cohort of 60 year olds newly diagnosed with ACS to a maximum 190
age of 100 years. Costs and QALYs are both discounted at an annual rate of 3.5% 191
(currently recommended in the UK). Subsequent analyses cover all combinations of 192
including and excluding incident cases, modelling patient cohorts of 50, 60, and 70 193
year old patients, and maintaining the maximum age 100 years timeframe or 194
reducing the timeframe to a 20 year follow-up period. The results of these analyses 195
are also combined to estimate age-weighted aggregate estimates of costs and 196
QALYs. 197
Results 198
The results of the analyses based on alternative sets of framework assumptions are 199
presented in Table 2 for a single age patient cohort. The reference case analysis 200
(number 1: 60 year old patients, excluding incident cases, followed to a maximum 201
age of 100 years) shows average costs per patients of almost £16,000 (difference 202
between treatment groups of £393), QALY gains of around 9 QALYs (difference 203
between treatment groups of 0.036), leading to an incremental cost-effectiveness 204
ratio (ICER) of £11,054. 205
Karnon et al, Modeling assumptions impact analysis
11
Compared to the reference case, analysis number 2 (in which a shorter time horizon 206
of 20 years is specified) shows a very small increase in the ICER (0.1%), which is 207
due to the high mortality rates associated with ACS – after 20 years over 85% of a 60 208
year old cohort receiving ASA+clopidogrel are expected to have died. In analysis 209
number 3, the lone effect of modelling incident cases over a 40 year time horizon is 210
not large (7.3%) as the effects on patients presenting beyond the 1st 20 years of the 211
time horizon are reduced due to discounting, whilst patients presenting up to 20 212
years are mostly followed up to death. 213
Analysis number 4 shows that the combined impact of reducing the time horizon and 214
including incident cases across the time horizon has a more substantial impact on 215
the estimated ICER than the sum of the parts, increasing it by over £3,000 (over 216
27%). 217
Comparisons of the same set of analyses within the 50 and 70 year old patient 218
cohorts show a similar pattern. The difference from the reference case results is 219
highest for the combined case (incident cases and short time horizon) for the cohort 220
of 50 year old patients as the shorter time horizon assumption has a greater 221
proportional impact in the case study model. The absolute effects in the 60 and 70 222
year old cohorts are similar for the shorter time horizon assumption, and the 223
percentage change differs only due to the magnitude of the reference case estimate. 224
Table 3 presents a comparison of the results within the separate age groups with the 225
results of a age-weighted aggregate patient cohort that reflects the age distribution of 226
patient cohort within the general population. Analyses numbers 1, 5, 9, and 13 show 227
that the reference case ICER (no incident cases, and a time horizon that follows 228
patients to a maximum age of 100 years) for the youngest of the three single age 229
patient cohorts was almost double the value of the ICER for the oldest age cohort, 230
and that there was a 14% difference between the age-weighted ICER and the ICER 231
for the middle age cohort (the most likely single age cohort to be presented in a singe 232
Karnon et al, Modeling assumptions impact analysis
12
age cohort analysis). A 60% difference in the ICER is noted when comparing the 233
reference case results for the 50 year old cohort and the age-weighted cohort. 234
Analysis number 8, involving a 20 year time horizon, incident cases and a 50 year old 235
cohort increases the ICER by over 120% relative to the reference case age-weighted 236
analysis. Comparing the age-weighted cohort with the 60 year old cohort, analysis 237
number 12 shows that if incident cases were combined with a shorter time horizon, 238
the estimated ICER would be 45% higher than the age-weighted reference case. 239
Discussion 240
This paper explores the impact of implementing alternative modeling assumptions 241
relating to the inclusion of incident cases over the course of the specified time 242
horizon; the use of a non-lifetime time horizon; and the handling of age-specific 243
patient sub-groups. The analysis of a published model that incorporated significant 244
age-specific parameter values shows that the effect of assumptions alternative to 245
defined reference case scenarios can have a significant impact on the resulting cost-246
effectiveness ratios. 247
The case study evaluation involves relatively high levels of disease-specific mortality. 248
This would be expected to reduce the estimated impact of including incident cases 249
and short time horizons as fewer long-term costs and effects are omitted. Other 250
models in disease areas with lower mortality effects would be expected to identify 251
even more significant differences in the incremental cost per QALY gained estimated 252
by alternative sets of assumptions. 253
Philips et al,[1] in their review of guidelines for good practice in decision-analytic 254
modeling in health technology assessment, state that existing guidelines (including 255
their own amalgamated guideline) require further development, but that guidelines 256
are likely to be ‘too general to pick up on the specific nuances of each model’ (pg iii). 257
These guidelines identify the appropriate time horizon as one that extends ‘far 258
Karnon et al, Modeling assumptions impact analysis
13
enough into the future to reflect important differences between options’ (pg21), and 259
recognise that a lifetime horizon is the most relevant option for most models.[1] 260
Weinstein et al also state that “Lifetime horizons are appropriate for many models 261
and are almost always required for models in which options have different 262
timevarying survival rates”[pg12][15] We have identified various model-based 263
evaluations that have not adopted these recommendations,[3-5] a factor that reduces 264
the comparability of published evaluations and may contribute to flawed policy 265
decisions. 266
The issue of modeling incident cases over the duration of a model is not explicitly 267
addressed by current guidelines. It is linked to the time horizon of the model as all 268
important effects of an intervention for each patient included in a model should be 269
described, i.e. all patients entering the model should be followed for their lifetime. 270
Few cost-effectiveness models have included incident cases to the end of the 271
specified time horizon, as done by Gaspoz et al,[3] though this study serves to clarify 272
the erroneous nature of this approach and to demonstrate the potential effects on the 273
cost-effectiveness results. 274
The definition of the appropriate patient cohort, specifically with regard to the age 275
distribution of an eligible patient population is important. Many model-based 276
treatment evaluations do not include any age-specific data other than general 277
population (non-disease specific) mortality rates,[16,17] which may reduce the impact 278
of modelling separate age cohorts. However, in cases in which more widespread 279
age-specific data are available the impact on aggregate cost-effectiveness may be 280
significant. 281
The fourth issue raised in this paper concerns the parsimony of decision analytic 282
models for the cost-effectiveness analysis of health care interventions, as 283
recommended by Weinstein et al.[15] The model that highlighted this issue was an 284
application of the CHDP model for the cost-effectiveness analysis of antiplatelet 285
Karnon et al, Modeling assumptions impact analysis
14
therapies.[3] This initial publication of this model was amended due to the post-286
publication identification of modelling errors, and the amended version retained some 287
seemingly counter-intuitive results that were not addressed by the authors. Though it 288
is not possible to prove a causal link between a non-parsimonious model and model 289
errors, we believe this example does confirm the value of model parsimony. 290
The potential advantage of using a broad model such as the CHDP model, to 291
evaluate a range of alternative interventions in the same broad disease area is that it 292
may provide a more suitable basis for the comparison of their relative cost-293
effectiveness. There is a balance to be struck regarding the benefits derived from 294
increased comparability of different evaluations and the extent to which reduced 295
parsimony complicates the analysis and increases the probability of error. 296
The increased awareness of the economic aspects of health care, and the need for a 297
guided rationale in allocating scarce health care resources that provides explicit 298
reasons for the choices made, led to a rapid increase in the number of published 299
economic evaluations of health care technologies over the last decade. Though 300
there is a growing acceptance of the role for economic evaluation in informing clinical 301
and policy decisions over the provision of alternative technologies, there remains 302
some suspicion around the methodological validity of economic evaluations. 303
Critics believe that the results of model-based evaluations can be manipulated more 304
easily and with more subtlety than other forms of evaluation,[18] and that the ‘black 305
box’ feel about the analysis of such models does not enhance confidence in the 306
derived results.[19] Such criticisms have less foundation when a transparent 307
description of model assumptions and data sources is provided. Halpern et al accept 308
that there are problems associated with modeling studies, but argue that many of the 309
cited problems could be countered through 'rigorous peer review and methodological 310
development'.[20] 311
Karnon et al, Modeling assumptions impact analysis
15
The main limitation of this paper is that the effects of three of the identified issues 312
were only examined in a single case study model, and so the study is not able to 313
generalise the expected magnitude of the observed effects across a range of case 314
studies in different disease areas. It was also not feasible to undertake a quantitative 315
analysis of the effects of model parsimony, which in our view, would require a multi-316
centre study, in which separately developed models are subject to blinded review by 317
independent expert reviewers. 318
This paper has demonstrated the effects of three specific issues to the extent that 319
they are important issues that have the potential to significantly affect cost-320
effectiveness results and that they should now be included in modelling guidelines. 321
The continued development, dissemination, and application of guidelines for model-322
based economic evaluations is necessary to further enable evidence based decision 323
making. 324
Karnon et al, Modeling assumptions impact analysis
16
References
1 Philips Z, Ginnelly L, Sculpher M, Claxton K, Golder S, Riemsma R, Woolacott
N, Glanville J, Review of guidelines for good practice in decision-analytic
modelling in health technology assessment, Health Technology Assessment
2004; Vol 8: number 36
2 Weinstein, M.C., Coxson, P.G., Williams, L.W., Pass, T.M., Stason, W.B.,
Goldman, L. (1987) Forecasting coronary heart disease incidence, mortality,
and cost: the Coronary Heart Disease Policy Model. American Journal of
Public Health 77: 1417-26.
3 Gaspoz, J-M., (2002) Cost effectiveness of aspirin, clopidogrel, or both for
secondary prevention of coronary heart disease. New England Journal of
Medicine 346(23): 1800-6. (corrected version published on NEJM website on
February 6, 2003).
4 Hillner BE, Benefit and Projected Cost-Effectiveness of Anastrozole versus
Tamoxifen as Initial Adjuvant Therapy for Patients with Early-Stage Estrogen
Receptor–Positive Breast Cancer, Cancer 2004;101:1311–22,
5 St Leger, A. S., Spencely, M., McCollum, C. N., and Mossa, M. Screening for
abdominal aortic aneurysm: a computer assisted cost-utility analysis. European
Journal of Vascular & Endovascular Surgery 1996; 11(2) 183-190
6 Gaspoz J-M, Coxson P, Goldman L, Aspirin, Clopidogrel, or Both for
Secondary Prevention of Coronary Disease, New England Journal of Medicine
2003, 348(6): 560-563
7 De Lemos, J.A and McGuire, D.K. (2003) Aspirin, clopidogrel, or both for
secondary prevention of coronary disease (letter to the editor). New England
Journal of Medicine 348(6): 561.
Karnon et al, Modeling assumptions impact analysis
17
8 Dewilde S, Anderson R, The Cost-Effectiveness of Screening Programs Using
Single and Multiple Birth Cohort Simulations: A Comparison Using a Model of
Cervical Cancer, Medical Decision Making 2004, Vol. 24, No. 5, 486-492
9 Koopmanschap MA, Lubbe KT, van Oortmarssen GJ, van Agt HM, van
Ballegooijen M, Habbema JK, Economic aspects of cervical cancer screening,
Soc Sci Med. 1990;30(10):1081-7,
10 Boer R, de Koning H, Threlfall A, Warmerdam P, Street A, Friedman E,
Woodman C, Cost effectiveness of shortening screening interval or extending
age range of NHS breast screening programme: computer simulation study,
BMJ. 1998; 317(7155): 376–379
11 Karnon J, Brennan A, Bakhai A, Pandor A, Flather M, Warren E, et al, A cost-
utility analysis of Clopidogrel in patients with non-ST elevation acute coronary
syndromes in the UK, Int J Cardiology 2006, 190:307-16
12 Collinson, J, Flather, MD, Fox, KAA et al. Clinical outcomes, risk stratification
and practice patterns of unstable angina and myocardial infarction without ST
elevation: Prospective Registry of Acute Ischaemic Syndromes in the UK
(PRAIS-UK). European Heat Journal 2001;21:1450-7
13 Wolfe CDA, Rudd AG, Howard R, et al. Incidence and case fatality rates of
stroke subtypes in a multiethnic population: the South London Stroke Register.
J Neurol Neurosurg Psychiatry 2002;72:211-6.
14 The Clopidogrel in Unstable Angina to Prevent Recurrent Events Trial
Investigators (2001) Effects of clopidogrel in addition to aspirin in patients with
acute coronary syndromes without ST-segment elevation. New England
Journal of Medicine 345: 494-502.
15 Weinstein MC, O’Brien B, Hornberger J, Jackson J, Johannesson M, McCabe
C, Luce BR, Principles of Good Practice for Decision Analytic Modeling in
Karnon et al, Modeling assumptions impact analysis
18
Health-Care Evaluation: Report of the ISPOR Task Force on Good Research
Practices—Modeling Studies, Value in Health 2003, 6(1):9-17
16 Palmer S, Sculpher M, Philips Z, Robinson M, Ginnelly L, Bakhai A, et al.
Management of non-ST-elevation acute coronary syndromes: how cost-
effective are glycoprotein IIb/IIIa antagonists in the UK National Health
Service? International Journal of Cardiology 2005;100(2):229-240,
17 Karnon J, Delea T, Johnston SRD, Smith R, Brandman J, Sung J, Goss PE,
Cost Effectiveness of Extended Adjuvant Letrozole in Postmenopausal Women
after Adjuvant Tamoxifen Therapy. The UK Perspective, Pharmacoeconomics
2006; 24 (3): 237-250
18 Sheldon, T. A. (1996 Jan-Feb) Problems of using modelling in the economic
evaluation of health care. Health Economics 5[1], 1-11.
19 Drummond, M. (1992) Cost-effectiveness guidelines for reimbursement of
pharmaceuticals: Is economic evaluation ready for its enhanced status? Health
Economics 1, 85-92.
20 Halpern, M. T., McKenna, M., and Hutton, J. (1998a) Modeling in economic
evaluation: an unavoidable fact of life. Health Economics 7[8], 741-2.
21 Joint Formulary Committee. British National Formulary no.46. London: British
Medical Association and Royal Pharmaceutical Society of Great Britain, 2003.
22 Department of Health. NHS reference costs, elective surgery costs (TELIP)
HRG E04 and HRG E15 2002 http://www.dh.gov.uk
23 Lightowlers S, McGuire A, Primary Prevention of Ischemic Stroke Cost-
Effectiveness of Anticoagulation in Nonrheumatic Atrial Fibrillation in the
Primary Prevention of Ischemic Stroke, Stroke 1998;29;1827-1832
24 Chambers M, Hutton J, Gladman J. Cost-effectiveness analysis of antiplatelet
Karnon et al, Modeling assumptions impact analysis
19
therapy in the prevention of recurrent stroke in the UK. Aspirin, dipyridamole
and aspirin-dipyridamole. Pharmacoeconomics 1999;16:577-93.
25 Derdeyn CP, Powers WJ. Cost effectiveness of screening for asymptomatic
carotid atherosclerotic disease. Stroke 1996;27:1944-50.
26 Salkeld G, Phongsavan P, Oldenburg B, et al. The cost-effectiveness of a
cardiovascular risk reduction program in general practice. Health Policy
1997;41:105-19.
27 Zeckhauser R, Shepard D. Where now for saving lives? Law Contemporary
Problems 1976;40:4-45.
28 Haigh R, Castleden M, Woods K, et al. Management of myocardial infarction in
the elderly: admission and outcome on a coronary care unit. Health Trends
1991;23:154-7.
29 Kuntz KM, Tsevat J, Goldman L, et al. Cost effectiveness of routine coronary
angiography after acute myocardial infarction. Circulation 1996;94:957-65.
30 Lee TT, Solomon NA, Heidenreich PA, et al. Cost-effectiveness of screening
for carotid stenosis in asymptomatic patients. Ann Intern Med 1997;126:337-
46.
31 Tengs TO, Lin TH. A meta-analysis of quality of life estimates for stroke.
Pharmacoeconomics 2003;21:191-200.
Karnon et al, Modeling assumptions impact analysis
20
Figure legends Figure 1 Decision tree model structure describing possible patient pathways
over the 1-year treatment period
ACS=acute coronary syndromes; CABG=coronary artery bypass graft;
MI=myocardial infarction; PCI=percutaneous coronary intervention.
Figure 2 Markov model structure describing possible patient pathways over the
remainder of the patient cohort’s lifetime
Table 1 Relative risks, costs and utility values used in the model
Table 2 Comparison of incident cases and time horizon assumptions in a
single age patient cohort (60 years olds)
Table 3 Comparison of single age and age-weighted patient cohort analyses
Karnon et al, Modeling assumptions impact analysis
21
Figure 1
12 months revasc. PCI
12 months revasc. CABG
Repeat revasc.
(no repeat rev asc.)
Acute PCI
Acute CABG
12 months revasc. PCI
12 months revasc. CABG
12 month revasc.
Death
MI
Stroke
Event-f ree
1
(no revasc.)
(no acute CABG)
(no acute PCI)
ACS patients
Go to 1
Go to 1
Go to 1
Karnon et al, Modeling assumptions impact analysis
22
Figure 2
ACSEvent-free
MIYear 1
StrokeYear 1
StrokePost-Year 1
MIPost-Year 1
Dead
Karnon et al, Modeling assumptions impact analysis
23
Table 1
Relative risk Mean Source
Non-fatal MI (clopidogrel+standard therapy vs standard therapy)
0.71 14
Non-fatal stroke (clopidogrel+standard therapy vs standard therapy)
0.73 14
Vascular death (clopidogrel+standard therapy vs standard therapy)
0.93 14
Revascularisation (clopidogrel+standard therapy vs standard therapy)
0.976 14
Major bleed (clopidogrel+standard therapy vs standard therapy)
1.38 14
Cost (£s)
ASA 300 mg/day (pa) 3.47 21 Clopidogrel 75 mg/day (pa) 460.32 21 PCI revascularisation 2,455 22 (HRG E15) CABG revascularisation 6,275 22 (HRG E05) Major bleeding event 2,064 23 ACS event-free (year 1) 1,421 16 ACS event-free (post-year 1) 1,421 16 MI (year 1) 3,966 16 MI (post-year 1) 1,587 16 Stroke (year 1)* 7,466 24 Utility values
ACS event-free (year 1) 0.80 25-27 ACS event-free (post-year 1) 0.93 28-30 MI (year 1) 0.80 25-27 MI (post-year 1) 0.93 28-30 Stroke (year 1)* 0.69 31 *Based on 35% of new stroke patients being dependent.[24] No utility values are described for the post-year 1 stroke state because the model assumes patients experiencing a stroke return to the ACS event-free state due to a lack of relevant prognostic data for these stroke patients. ACS=acute coronary syndromes; ASA=acetylsalicyclic acid; CABG=coronary artery bypass graft; CI=confidence interval; MI=myocardial infarction; PCI=percutaneous coronary intervention.
Karnon et al, Modeling assumptions impact analysis
24
Table 2
Costs QALYs
Framework parameters ASA ASA+Clopidogrel ASA ASA+Clopidogrel ICER % difference from
reference case
1 40 year time horizon, no incident cases £15,604 £15,997 8.843 8.879 £11,054 - 2 20 year time horizon, no incident cases £14,558 £14,952 8.212 8.248 £11,065 0.10% 3 40 year time horizon, incident cases £7,685 £7,901 4.320 4.339 £11,860 7.3% 4 20 year time horizon, incident cases £8,007 £8,291 4.407 4.427 £14,085 27.4%
ASA – aspirin; ICER – incremental cost-effectiveness ratio
Karnon et al, Modeling assumptions impact analysis
25
Table 3
Costs QALYs Framework parameters
ASA ASA+Clopidogrel ASA ASA+Clopidogrel ICER % difference from
reference case
Weighted aggregate cohort
1 30-50 year time horizon, no incident cases £13,543 £13,936 7.574 7.614 £9,716 - 2 20 year time horizon, no incident cases £12,700 £13,093 7.063 7.104 £9,748 0.3% 3 30-50 year time horizon, incident cases £6,953 £7,182 3.840 3.862 £10,380 6.8% 4 20 year time horizon, incident cases £7,219 £7,501 3.914 3.937 £11,946 23.0%
50 years
5 50 year time horizon, no incident cases £21,296 £21,687 12.375 12.400 £15,630 60.9% 6 20 year time horizon, no incident cases £17,944 £18,337 10.342 10.367 £15,919 63.8% 7 50 year time horizon, incident cases £9,264 £9,454 5.355 5.367 £16,387 68.7% 8 20 year time horizon, incident cases £9,240 £9,525 5.199 5.212 £21,533 121.6%
60 years
9 40 year time horizon, no incident cases £15,604 £15,997 8.843 8.879 £11,054 13.8% 10 20 year time horizon, no incident cases £14,558 £14,952 8.212 8.248 £11,065 13.9% 11 40 year time horizon, incident cases £7,685 £7,901 4.320 4.339 £11,860 22.1% 12 20 year time horizon, incident cases £8,007 £8,291 4.407 4.427 £14,085 45.0%
70 years
13 20 year time horizon, incident cases £6,300 £6,581 3.334 3.362 £10,011 3.0% 14 30 year time horizon, incident cases £5,995 £6,240 3.212 3.238 £9,197 5.3% 15 20 year time horizon, no incident cases £10,417 £10,809 5.643 5.690 £8,408 13.5% 16 30 year time horizon, no incident cases £10,544 £10,936 5.719 5.766 £8,393 13.6%
ASA – aspirin; ICER – incremental cost-effectiveness ratio