PRESENTATION OVERVIEW Stochastic Modeling Monte Carlo Simulation (MCS) Modeling—An
Overview Benefits of MCS
STOCHASTIC MODELING IN PHARMACOECONOMICS: COMMON MISTAKES AND HOW TO AVOID THEM COMMON MISTAKES AND HOW TO AVOID THEM
Software to Support MCS Software to Support MCS Three common mistakes in stochastic modeling: Uncertainty versus Variability
Huybert Groenendaal, PhD, MBA, Francisco Zagmutt, DVM, MPVM (EpiX Analytics) Jane Castelli‐Haley, MBA (Teva Pharmaceuticals)
Multiplication of Distributions Duplications of Uncertainties
INTRODUCTION
UNCERTAINTY
Clinical, humanistic and economic data guide important
decisions about patient care
Modeling studies in are typically based on data from a limited number of patients
In addition, typically variability between patients (groups) dd ll bl b ( ) in the outcomes of treatments
As a consequence, always uncertainty in the results of a modeling study (e.g. ICER)
Monte Carlo the most common technique to quantitatively deal with uncertainty and/or variability
Researchers have an ethical mandate to evaluate and
disclose the accuracy and reliability of their findings Research results can be compromised when appropriate R h lt b i d h i t
techniques are not used accurately to evaluate and disclose the level of uncertainty in the findings
UNCERTAINTY
4
UNCERTAINTY
Results of a study shown with and without uncertainty: 1. ICER = $23,000 2. ICER, shown as a cost‐effectiveness acceptability curve:
Sensitivity Analysis: Answers the following question: What uncertainty mostly affects the overall uncertainty of the results?
Cost‐effectiveness acceptability curve 100.00%
Cummulative probability
80.00%
60.00%
40.00%
20.00%
0.00% 0
10000
20000
30000
40000 ICER (US$)
50000
60000
70000
80000
5
6
1
MONTE CARLO SIMULATION Monte Carlo Simulation (MCS) is the most common technique to deal with uncertainty, as it allows us to approximate complex‐‐and often impossible to solve scenarios‐‐via simulation. g g Benefits of Stochastic Modeling Using MCS MCS takes into account variability and/or randomness of patient characteristics and treatment outcomes that cannot be addressed mathematically. MCS addresses statistical uncertainty about model parameters MCS is a useful (most common) tool to develop a cost‐ effectiveness acceptability curve.
COMMON MISTAKE #1: Mistake: Incorrect use of variability distribution instead of uncertainty distribution
Interpretation of uncertainty in CE studies is related to the level of knowledge associated with the parameters, and not with the underlying random variation.
Decision‐makers interested in overall ICER, not on the basis of an individual patient
MONTE CARLO SIMULATION Available Software to Support MCS Excel Goldsim Crystal Ball / @RISK Treeage
COMMON MISTAKE #1: Two ways to perform Monte Carlo simulation: 1. First‐order uncertainty (randomness/variability) – following (large number of) individuals through the Markov Model 2. Second‐order uncertainty – reflects uncertainty in parameters within the model In CE‐analysis, we are interested in mean treatment costs and mean treatment effect across patient population ‐ variability is generally not important for decision‐ makers!
Difference: What is the difference between variability and uncertainty? Uncertainty (lack of knowledge):
Variability only:
e.g. N(,)
Variability and parameter uncertainty together:
N(
,
)
Parameter Uncertainty Model Uncertainty (can be dealt with using Bayesian Model Model Uncertainty (can be dealt with using Bayesian Model Averaging, BMA)
Especially relevant when there is little ‘data’ Variability:
Patients within a clinical trial are randomly selected and will differ from patient to patient in ‘outcome’
How do we deal with variability (e.g. targeted patients/subpopulations)?
In CE studies, typically we are interested in parameter uncertainty only!
2
COMMON MISTAKE #1 – example A:
COMMON MISTAKE #1 – example B:
Example 1:
Example 2:
“The Monte Carlo simulations of the model were performed, using microsimulations trails with 1000 hypothetical patients.”
“The costs per month were modeled using a Lognormal distribution with a mean of $860 and 95‐percentile of $530 – 10,000”
COMMON MISTAKE #1 – example C:
COMMON MISTAKE #2: Multiplication of Distributions
Example 3:
Theoretically impossible from the mathematicians
“The base case simulations in Table 6 use mean costs and mean utilities for the Markov states, ignoring the variability and distribution.”
No way to define the outcome when two (or more)
perspective
“Acceptability curves are generated by Monte Carlo simulation, using individual draws (500 in this case) from the entire distribution of costs and utilities in different states.” …….
distributions are multiplied Still frequently used in medical decision making (i.e.
incorrectly multiplying a variability distributions (e.g. the costs of getting a certain disease) by a frequency (e.g. the number of patients getting the disease) to estimate total expenditures from a public health perspective.
45
40
35
30
25
20
5
0
15
762.9
656.3
443.2
549.7
35,000
30,000
25,000
20,000
15,000
How would we calculate the total disease cost Example model with uncertainty for next year?
10,000
Cost/person $Lognormal(10000, 7000)
230.0
Chulera outbreaks, thus next year cases could be modeled as NegBin(15, 1/(1+1/2)) # of people affected per city N(500,80)
336.6
In the last 2 years we observed 50 cities with
5,000
Possible correct approaches: Central Limit Theory Simulate individual patients Other aggregate distribution methods (e.g. Fast Fourier Transfer method, FFT)
Assume the following:
Values x 10^-4
multiplication, because it works using expected values but not once you simulate variability.
0
Issue is that people are trying to do an addition via
Estimate the cost (with uncertainty) of outbreaks of infectious disease (Chulera) in a large urban population next year. As the disease is highly infectious, the outbreaks cluster within cities and exhibits a similar pattern of spread on each disease regardless of the population size of city.
-5,000
COMMON MISTAKE #2: Multiplication of Distributions
10
Handling Multiplication of Distributions
3
Comparing the correct and wrong approach Multiplying distributions 76.73 1.0
5.0% 34.3%
178.26 90.0% 46.5%
COMMON MISTAKE #3: Duplication of Uncertainties
5.0% 19.2% Total money (CLT) Minimum €46,947,600.37 Maximum €244,399,161.41 Mean €125,003,283.41 Std Dev €30,722,030.01 2.5% €69,559,733.86 97.5% €189,396,336.10 Values 1000
0.8
0.6
Representing an uncertain variable more than once in a model or (mistake related to this) having two uncertain model or (mistake related to this) having two uncertain variables that are highly correlated
Wrong answer
0.4
Minimum €7,520,049.85 Maximum €975,690,351.98 Mean €124,243,361.31 Std Dev €95,593,736.45 2.5% €23,779,278.16 97.5% €369,987,136.43 Values 1000
0.2
0.0 Values in Millions (€)
COMMON MISTAKE #3: Duplication of Uncertainties Examples: Two separate distributions for one input in a model, obviously gives incorrect scenarios and can result in b i l i i i d l i smaller variance of the outputs (less uncertainty in results), as the variability among iterations will be less than when using only one distribution.
COMMON MISTAKE #3: Duplication of Uncertainties Good health Treatment A P= Bad health
Treatment
Costs/QALY’s
Good health
Two distributions that have common components,
Treatment B
for example the costs with and without a certain treatment, which are not independent.
P=
Bad health
Costs/QALY’s
Empirical data available suggests that ‘p’ is different between A and B
Percentiles 1% 5% 10% 50% 91% 95% 99%
ICER (incorrect) $ (47,791.95) $ (9,360.42) $ (3,994.10) $ 328.21 $ 5,326.18 $ 9,110.34 $ 45,729.60
ICER (correct) $ (49,308.53) $ (10,246.56) $ (4,401.72) $ 529.54 $ 5,526.70 $ 9,737.50 $ 50,080.15
Lesson: Parameters of which no empirical data for differences should not be simulated for each treatment arm separately!
COMMON MISTAKE #3: Duplication of Uncertainties
A related common mistake is that an assumption is made that all uncertainty distributions are independent (uncorrelated) While this is a very independent (uncorrelated). While this is a very common (and often unmentioned) assumption, it is also a very strong assumption!
If you have no dependencies in your model, be aware!
The use of the WinBUGS software can help estimate uncertainty distribution as their dependencies
Useful paper: O’Hagan et al., 2005. PharmacoEconomics, 2005; 23 (6): 529-536
4
IN SUMMARY: 1.
Only 2nd order uncertainty is important in CE studies (e.g. ICE‐ratio)
2.
Do not multiply distribution of variability;
3.
Be aware of (hidden) dependencies / correlations in your model!
Questions? [Course instructor name] [Course Instructor Email and phone] Vose Consulting
Consequence: Common mistakes can have large effects on the results that can really influence decision‐makers
Dr. H.Groenendaal Managing partner EpiX Analytics LLC
[email protected] P: 303 440 8524
5