Académique Documents
Professionnel Documents
Culture Documents
Abstract Background: Markov models have been the standard framework for pre-
dicting long-term clinical and economic outcomes using the surrogate marker
endpoints from clinical trials. However, they are complex, have intensive data
requirements and are often difficult for decision makers to understand.
Recent developments in modelling software have made it possible to use
discrete-event simulation (DES) to model outcomes in HIV. Using published
results from 48-week trial data as model inputs, Markov model and DES
modelling approaches were compared in terms of clinical outcomes at 5 years
and lifetime cost-effectiveness estimates.
Methods: A randomly selected cohort of 100 antiretroviral-naive patients
with a mean baseline CD4+ T-cell count of 175 cells/mm3 treated with lopi-
navir/ritonavir was selected from Abbott study M97-720. Parameter esti-
mates from this cohort were used to populate both a Markov and a DES
model, and the long-term estimates for these cohorts were compared. The
models were then modified using the relative risk of undetectable viral load as
reported for atazanavir and lopinavir/ritonavir in the published BMS 008
study. This allowed us to compare the mean cost effectiveness of the models.
The clinical outcomes included mean change in CD4+ T-cell count, and
proportion of subjects with plasma HIV-1 RNA (viral load [VL]) <50
copies/mL, VL 50400 copies/mL and VL >400 copies/mL. US wholesale
acquisition costs (year 2007 values) were used in the mean cost-effectiveness
analysis, and the cost and QALY data were discounted at 3%.
Results: The results show a slight predictive advantage of the DES model
for clinical outcomes. The DES model could capture direct input of CD4+
T-cell count, and proportion of subjects with plasma HIV-1 RNA VL <50
copies/mL, VL 50400 copies/mL and VL >400 copies/mL over a 48-week
period, which the Markov model could not. The DES and Markov model
estimates were similar to the actual clinical trial estimates for 1-year clinical
results; however, the DES model predicted more detailed outcomes and had
160 Simpson et al.
slightly better long-term (5-year) predictive validity than the Markov model.
Similar cost estimates were derived from the Markov model and the DES.
Both models predict cost savings at 5 and 10 years, and over a lifetime for the
lopinavir/ritonavir treatment regimen as compared with an atazanavir
regimen.
Conclusion: The DES model predicts the course of a disease naturally, with few
restrictions. This may give the model superior face validity with decision makers.
Furthermore, this model automatically provides a probabilistic sensitivity anal-
ysis, which is cumbersome to perform with a Markov model. DES models allow
inclusion of more variables without aggregation, which may improve model
precision. The capacity of DES for additional data capture helps explain
why this model consistently predicts better survival and thus greater savings
than the Markov model. The DES model is better than the Markov model in
isolating long-term implications of small but important differences in crucial
input data.
2009 Adis Data Information BV. All rights reserved. Pharmacoeconomics 2009; 27 (2)
Markov Model vs Discrete Event Simulation for HIV 161
2009 Adis Data Information BV. All rights reserved. Pharmacoeconomics 2009; 27 (2)
162 Simpson et al.
Table I. Markov and discrete-event simulation (DES) model clinical value and the corresponding average patients
predictive validity for the lopinavir/ritonavir patient sample[16] survival estimates are plotted together.
Parameter Actual Markov DES range
clinical value prediction prediction
Probabilistic Sensitivity Analysis (PSA) and
CD4 count at 258 NA 24770 Evaluation of Parameter Relationships
1 year (mean)
VL <400 copies/mL 65 65 6166 A DES model can be programmed to perform
at 1 year (%)
100 iterations, each based on a randomly selected
VL <50 copies/mL 54 NA 5153
at 1 year (%)
combination of the model variables, as specified
CD4 count at 841 NA 82258
by the values and distributions of the parameters
5 years (mean) of the model. Each estimate can be set for a spe-
VL <400 copies/mL 56 50 5760 cific cohort size. This allows the user to perform
at 5 years (%) probabilistic sensitivity analysis (PSA), as well as
VL <50 copies/mL 54 NA 5556 examine the effects of applying model decisions
at 5 years (%) to small and large practice settings.
NA = not available (the Markov model is not designed to produce
In addition, this probabilistic feature allows
these estimates); VL = viral load.
the user to plot the relationships between key
model outputs to get an understanding of how
the Markov and DES models are very stable, and outputs and variable values move together, as
lie very close to the values actually observed in shown in figure 1. This plot makes it very clear that
the cohort underlying the model (see table I). these two outcomes are correlated as expected.
Similar cost estimates are derived from Markov This paper presents results from a traditionally
models and DES (table II). Both models predict structured HIV model, and compares these re-
cost savings at 5 and 10 years, and over a lifetime sults with those of a more innovative healthcare
for the lopinavir/ritonavir treatment regimen. application,[17] a DES model. The estimated
The values output of the DES model depends on outcomes for the two structures are very close,
how many patients are specified for the simula- but not identical. The DES model has slightly
tion, and how many iterations the model runs. better predictive ability over 5 years, and is able
The larger the number of estimates (i.e. entities; to provide many more details about what may be
e.g. n = 2000), the less variation may be expected expected to happen to a population than the
between each model run. However, each time a more traditional Markov model. However, the
model is saved with a new name it will draw from performance of the Markov model is quite good.
a new random set of starting values, and the re- Its results fall clearly within the 3% margin of
sulting estimates will differ from the estimates in error (see table I) defined by other modellers
the previous model iteration. Figure 1 illustrates as the acceptable level of variability for a point
this variation where the estimated 48-week VL validation of an economic model.[18,19]
Table II. Comparing the cost-effectiveness estimates of the Markov model (MM) and discrete-event simulation (DES) model
Model/regimen Costs ($US, year 2007 values) Total QALYs Mean cost per QALY
5-year 10-year total
MM: lopinavir/ritonavir 105 808 189 704 310 194 10.55 29 402
DES: lopinavir/ritonavir 90 335 152 912 340 022 12.40 27 421
MM: atanazavir 119 551 207 649 318 882 10.11 31 541
DES: atanazavir 97 283 163 288 352 843 12.11 29 136
2009 Adis Data Information BV. All rights reserved. Pharmacoeconomics 2009; 27 (2)
Markov Model vs Discrete Event Simulation for HIV 163
63.5
tions to allow the analysis required to estimate
transitions. However, once well documented
63.0
transitions for a disease are available for a
62.5 specific severity group, they can form the basis
62.0 for model adaptations. The DES model may
not need a large amount of data, and some of
61.5
12.4 12.5 12.6 12.7 12.8 12.9 the models parameters can come from com-
QALYs monly reported outcomes or risks, but it requires
data that are recorded over a much longer time
Fig. 1. Correlation between model predictions of 48-week viral load
(VL) and QALY estimates. Each point is one model run. because it does not have the repetitive assump-
tion of a Markov model. Thus, as is often the
case in science, an aspect that is a strength in
Decision-analysis models provide an im- one case may become a weakness in the next
portant mechanism for translating the results of application.
clinical trials into measures of quality of life Furthermore, a Markov model uses mean
(QOL) and costs. This has especially been the values for all the variables in the model to calcu-
case in HIV since the mid-1990s, when the ac- late outcomes, while a DES model uses a random
cepted clinical trial outcomes changed from dif- draw from the distributions depicted by mean
ferences in progression to AIDS or death, to values and their measures of variation to estimate
differences in surrogate markers,[20] usually VL outcomes. Thus, DES outcomes are always esti-
suppression and CD4+ T-cell count increase. mated a number of times and the mean of these
The most common analytical structure used estimates are reported. This provides some
for HIV economic models has been the Markov, information about the effect of the statistical
or state-transition model.[21-26] One such model, uncertainty in the model input values. Stable
the Cost Effectiveness of Preventing AIDS models that are based on strong statistical para-
Complications (CEPAC) model, has been used meters will show little variation in the estimates,
extensively to estimate outcomes expected for while unstable model estimates will vary greatly.
treatments for individuals with HIV.[25,26] In only Examining the high and low values for each
rare cases are these models ever compared with parameter provides a good representation of the
the estimations derived from differently struc- stability of a model.
tured models[23] or with actual clinical trial In this study, the older Markov approach re-
results.[27] This is because the models are pro- quired substantially fewer data parameters for
prietary, and even well described and extensively estimation and much less time for implementa-
published proprietary models (such as the tion. The extra time requirement for the DES
CEPAC) rarely have enough data in a manu- model might be because of our relatively lesser
script to allow others to replicate their complete experience in working with the Arena soft-
structure. However, the structure of a model is an ware.[29] However, the relatively large data re-
important component to assess in the evaluation quirement may also be an issue in the use of DES
of the quality of a modelling estimate.[28] models in the estimation of outcomes and costs
Each of these two model structures has for HIV treatments, because a structure of a
advantages and disadvantages that must be model may be substantially influenced by the
considered before choosing one over the other data available for modelling.[30] The issue of
(table III). In many cases, the choice must be which structure is best for HIV has not yet been
guided by the data available for the modelling. determined. It will depend on which of the two
The transition matrix in the Markov model models is most transparent to decision makers,
requires large amounts of data, especially if the since model transparency is a critical character-
2009 Adis Data Information BV. All rights reserved. Pharmacoeconomics 2009; 27 (2)
164 Simpson et al.
Table III. Summary of strengths and weaknesses of the Markov and discrete-event simulation (DES) models for use in estimation of the cost
effectiveness of antiretroviral therapies in HIV
Markov model DES model
Computationally simple, can be implemented in Microsoft Excel Requires special, often expensive, software with archaic
programming conventions not commonly used in healthcare
Cumbersome to modify when the lack of memory assumption does Can easily accommodate risk changes, such as ageing, over time
not hold
Uses the raw data frequencies from the clinical trial and may Uses statistical parameters, such as relative risk of viral load
therefore account for all differences between treatment groups breakthrough, or mean CD4 increase per time period, which may not
capture subtle variations in distributions between the measurement
times
Does not lend itself easily to probabilistic sensitivity analysis Probabilistic sensitivity analysis is inherent in the model
Can be visually inspected for programming errors, and is easy to test Parameter checks require the examination of hidden logic in nodes
for technical replication and tables, which can be difficult to do
A small increase in the number of health states requires a large Very parsimonious data requirements, but may need populations
increase in the data input to get stable transitions observed over long time periods to identify flow change points
Sequences in treatment over time require the population to transit to New treatments are easily accommodated by simple flow
a new Markov model within the model structure specifications
Requires mutually exclusive and jointly exhaustive health states, Creates many individual entities that represent patients. Entities can
which is an unnatural way for clinicians to think about a disease carry permanent traits (sex) and changing traits (age) and contact
probabilistic events (AIDS or myocardial infarction)
Is limited to reporting mean values estimated across the specific Can provide a report for any variable at any time; however, reports can
health states for the time increments in the model, but results are be difficult to program
easy to program and to check visually
All characteristics of individuals must be defined in relation to the No limit to the characteristics of individuals that can be included, as
health states in the model, and sub-models may be needed to long as both parameter estimates and the correlation among the
capture important conditions variables included are available
istic, and one that can only be determined by Because of the limitations of the Markov
exposing the model to critique from those who model, researchers end up using categorical
consider using it to inform decisions. groupings for complex interacting continuous
measures, with a potential for short-term ag-
gregation bias leading to long-term prediction
Conclusion
errors. DES models allow inclusion of individual
The DES model predicts more detailed out- variables without a need for creating compound
comes and has slightly better long-term predictive health states, thus improving the model precision
validity than the Markov model. It represents the as demonstrated above. However, they may also
course of a disease much more naturally, with require much more data analysis and program-
fewer restrictions. This may give the model ming time, and at this point it is not clear to us
superior face validity with decision makers. Most which structure has the greater transparency.
importantly, this model automatically provides
a PSA, which is cumbersome to perform with a
Acknowledgements
Markov model.
The capacity of the DES model for additional This study was funded by a grant from Abbott Labora-
data capture helps explain why it consistently tories to the Medical University of South Carolina. Drs Dietz
predicts better survival and greater savings. The and Rajagopalan are Abbott employees and as such are
DES model is better than the Markov model in eligible to receive stock options. Kit Simpson has received
consultancy fees from Abbott. The authors acknowledge the
isolating long-term implications of small but im- assistance of Menaka Bhor, PhD, Abbott Laboratories, in
portant differences in crucial input data. editing and organizing the manuscript.
2009 Adis Data Information BV. All rights reserved. Pharmacoeconomics 2009; 27 (2)
Markov Model vs Discrete Event Simulation for HIV 165
2009 Adis Data Information BV. All rights reserved. Pharmacoeconomics 2009; 27 (2)