Uncertainty and Sensitivity Analyses

Marcelo Coca Perraillon

University of Colorado
Anschutz Medical Campus

Cost-Effectiveness Analysis
HSMP 6609
2020

1 / 46
Outline

Defining terms: variability, heterogeneity, uncertainty

Sensitivity analysis: “deterministic” and “probabilistic”
Base case, one-way, two-way, three-way, scenarios
Influential variables: tornado diagrams
More advanced methods: probabilistic sensitivity analysis (PSA)
Probability distributions
How to choose probability distributions for probabilistic sensitivity
analysis
Example: HIV model

2 / 46
What is sensitivity analysis?
We have already done sensitivity analyses in homework and during
class
It is essentially the study of how changes in model inputs affect
model outputs
Inputs are quantities, prices, probabilities, and so on. In essence, all
we measure in a CEA study
Outputs are the model outcomes we care about: average or expected
costs, average or expected years of life gained, average or expected
ICER or Net Monetary Benefit. Most often than not, the focus is on
ICER
The objective of a sensitivity analysis is to understand 1) which are
the most important inputs/parameters that affect our
study/model and 2) quantify the overall uncertainty of our models
We often cannot do traditional statistical analyses (p-values,
confidence intervals, etc) because we don’t have individual data
3 / 46
Deterministic and probabilistic sensitivity analysis

We’ll cover two types of sensitivity analyses:

1 Deterministic: We choose values for one or more parameters keeping
the rest constant. For example, min or max or a case that has policy
relevance. This is what we have done so far
2 Probabilistic: We assign parameters a probability distribution and use
simulations to compute new ICERs or other outcomes of interest
One confusing part is that, once again, we are dealing with concepts
that are called different names in different fields
We need to spend some time defining terms

4 / 46
Defining terms

Variability: Variation due to chance; random variability. Even with

identical patients facing the same probability of an event, some
experience the event and some do not. Sometimes called stochastic
uncertainty or first-order uncertainty (less common)
Heterogeneity: Differences between patients that can be attributed
or explained by patient’s characteristics (think sex, age, income, and
so on)
Uncertainty: What sensitivity analysis tries to measure. We do not
know the true value of some input (parameter) or the true way a
process is generated
Two types of uncertainty: parameter uncertainty and model
uncertainty

5 / 46
Parameter uncertainty

We do not know the true value of a parameter. For example, the true
average cost for a procedure in the population
We could estimate the mean cost in a sample but the estimated
parameter (the mean) itself has some variability
In statistics, a measure of parameter uncertainty is the standard
error (do not confuse it with the standard deviation)
We can reduce some of this uncertainty if we had a larger sample; the
standard error depends on sample sizes. The larger the sample the
more precise the estimate and the smaller the standard error
(A related concept: consistency. In probability, the larger the sample
size, the closer the estimated parameter will be to its true value)
For the purpose of this class: Each model parameter has some
uncertainty that we somehow should take into account

6 / 46
Parameter uncertainty

Last week when we covered the SIR/SEIR models used to understand

the coronavirus, I mentioned parameter uncertainty
We just don’t have good information right now (as of mid April) to
build good models
What is the transmission rate, β? What is the recovery, γ? Even
more basic, what is the mortality rate?
In the HIV example: is the relative risk really 0.509? What was the
confidence interval of the study? Intuitively, if the 95% confidence
interval was, say, (0.45 - 0.55) we’d argue that the parameter
uncertainty less than if it were (0.2 - 0.70)
All that is parameter uncertainty
(I’m going to say more about confidence intervals in about 30 mins)

7 / 46
Model uncertainty

Models are simplifications of reality; some models capture the key

elements of reality and some don’t
Any model that you see in the literature could be modified to make it
more “realistic” but it adds complexity. More realistic models are not
necessarily better
The uncertainty we have about model assumptions is called model
uncertainty
For example, when we were modeling breast cancer recurrence, we
were wondering if not keeping track of time from recurrence was a
problem–we used tunnels to sidestep that problem. Was it the wrong
assumption?
Is ignoring the Exposed health state in the SIR model a terrible
mistake?

8 / 46
Analogy with regression

Mandatory slide in cost-effectiveness... Cover your ears if you don’t

know regression analysis (but don’t get too distracted)
A standard linear regression model can be written as:
yi = β0 + β1 X1i + ... + βk Xki + i
where i is assumed to distribute normal with mean 0 and variance σ 2
After collecting data on variables y and X1 to Xk , the estimateed
model is:
E [ŷ ] = βˆ0 + βˆ1 X1 + ... + βˆk Xk
Standard errors for the coefficients SE (β̂j ) are also estimated
Each coefficient is a random variable with a (theoretically) known
distribution. Once you know the estimated coefficients and their SEs,
it is possible to test hypotheses, build confidence intervals, etc

9 / 46
Analogy with regression

The random error  corresponds to stochastic uncertainty or

variability due to chance
Parameter uncertainty is measured by the standard error SE (β̂j ) of
the estimated parameter β̂j
Heterogeneity of effects is represented by coefficients β̂j
Model uncertainty is the uncertainty around the assumptions of the
model. Is it correct to assume that the error, and consequently the
outcome Y , distribute normal? Are we including all the relevant
covariates X ? Should we model Y or log (Y )?
Anyhow, most textbooks and articles introducing uncertainty use this
analogy so there you go

10 / 46
Cheat sheet

Of course, there is an article trying to come up with an agreement on

names
From Briggs et al (2012). I used red to underline what I think is the
most common term (apparently I don’t agree with Briggs)

11 / 46
Deterministic sensitivity analysis

We’ll focus on simple deterministic sensitivity analyses

One-way: Change one parameter at a time keeping all others
constant
Standard way of presenting one-way sensitivity analyses results is to
plot the parameter you are changing in the x-axis and an output of
interest on the y-axis
In the HIV example, we could change the relative risk and analyze
what happens to average costs in the combination therapy group and
the ICER
You did that in one homework

12 / 46
HIV transition diagram, reminder

See Excel file for more details

13 / 46
One-way sensitivity analysis of relative risk with respect
combination therapy cost

Lower RR implies combo therapy more effective, so more people stay

in states that are costly. It would be unrealistic to choose very high or
very low relative risks. Combination therapy at the time of the trial
was not a miracle medication
14 / 46
One-way sensitivity analysis of relative risk with respect to
ICER

ICER shows more nuance because of more moving parts. More

effective drug increases costs but it also increases life expectancy.
Less effective combo therapy becomes more similar to mono therapy
but costs keep going up
15 / 46
For special assignment II

Most of the time we want to understand how changing a parameter

changes the ICER, not just one component of the ICER
Ccombo −Cmono
The ICER has four components: Yearscombo −Yearsmono
You need to understand how the four components are changing
Some analyses result in weird-looking graphs and non-linearity, usually
because of dividing by number close to zero
A ratio in which the denominator is close to zero tends to infinity,
positive or negative
In the HIV example, if the relative risk is close to 1, then there is no
difference in effectiveness between mono and combo therapy. That
means that Yearscombo − Yearsmono ∼ 0

16 / 46
Using a macro for sensitivity analysis in Excel

It is easy to change values in Excel but it is not the most practical

way to perform sensitivity analyses
Excel has a macro language that is very flexible and relatively easy
to use, although it can take a while to master it
It’s based on a version of a programming language, Visual Basic
(Visual Basic for Applications, VBA)
Fortunately, you don’t need to master it in order to use it for simple
analyses
We will use a simple macro for quick sensitivity analyses
But it’s worth learning macros. Add it to resume (if you learned
how to use macros, of course)

17 / 46
Macros in Excel

First, make sure that you can see the Developer ribbon in Excel:
1 Click the File tab.
2 Click Options.
3 Click Customize Ribbon.
4 Under Customize the Ribbon and under Main Tabs, select the
Developer check box

18 / 46
Record a macro

The easiest way to use macros is to record a macro of a repetitive

task like changing the RR and then copying and pasting the resulting
ICER into a new column
Excel will create VBA code to do the repetitive task
It is then (relatively) easy to modify the code so the repetitive task is,
well, repeated several times
See the Excel file HIV “sensitivity m.xlsm” for an example (macros
could be used as a computer virus so you’ll get a warning before
opening the file)
See the video “Recording a macro” to see how I created the macro
(one good thing about the coronavirus: I learned how to make screen
videos)

19 / 46
Final code

20 / 46
Logic

For each copy-paste step, Excel creates a chunk of code

We used this chunk of code to create a loop so Excel does the same
in a range of cells, from rows 5 to 17
We added
1 Dim i As Integer
2 For i = 5 To 17
3 Next i
See http://www.excel-easy.com/vba.html for more Examples
Remember, this code only repeats interactions with Excel that you
could do “by hand” (pointing, clicking, or using the keyboard)
VBA is quite flexible so you can do many more things

21 / 46
Tornado diagrams

Tornado diagrams are a nice way of depicting results of one-way

sensitivity analyses
Provides a way to quickly figure out which parameters have the
most influence on outcomes
This is very helpful because one needs to pay more attention to the
most important parameters
In other words, we may not need to worry too much about one
parameter value if it turns out that the parameter is not that
“important”
We will do a tornado diagram of the the HIV model

22 / 46
Finished tornado diagram

The top variable has more uncertainty or is more sensitive to possible

range of values

23 / 46
Tornado diagram HIV model

We need to come up with a range of values for the parameters

To make things simpler, I’ll focus on only three variables
I came up with a range for them:

24 / 46
Where do we get these numbers?

The base-case scenario is our best estimate (sometimes called

baseline) scenario or baseline values
The range could be obtained from multiple sources, like drug prices
that are common for hospitals or prices after discounts
The relative risk was obtained from a meta-analysis, so a confidence
interval could be available
We can then calculate the ICER corresponding to each of the values
in the table to obtain a lower and upper bound estimate of the new
ICER corresponding to lower and upper bound range of values

25 / 46
Table for tornado diagram

Lb = lower bound; Ub = upper bound

The range is the absolute value of the difference between the upper
and the lower bound
Table is sorted from lowest to largest range
A tornado diagram is a graphical representation of this table

26 / 46
Tornado diagram

Vertical line is the ICER for base case ($11,499); the most influential
variable is the drug price for combination therapy
See the video “Tornado diagram” to learn how to do this graph in
Excel
27 / 46
Tornado diagram in Excel

A tornado diagram is a essentially a bar chart

Select the upper and lower bound ICER estimates, click insert and
select a 2-D horizontal bar chart
Select the horizontal axis (where the ICER values are) and find the
option “Vertical axis crosses.” Choose Axis value and enter the
baseline ICER 11499
Select one bar and find the option that says Series Overlap and
increase the overlap to 100
That’s it.
You can then change the colors, labels, etc

28 / 46
Why are they useful? Department of Big Pictures

Let’s go to basics. You (somehow) come up with a range of plausible

values. Then you calculate ICERs for the range. Then you graphically
see which variable has the largest range (the “top” of the tornado)
This way you identify –and communicate– the most important
(sensitive to changes) variables in your model, in the sense that
changing the value of those variables changes the ICER (or any other
outcome of interest, but mostly the ICER)
About the “somehow” part. If, for example, the estimate comes from
a study, you could use the 95% confidence interval
If it’s the price of a medication, you could use different prices:
wholesale, what the VA pays, what an average hospital pays
You could ask a bunch of experts... There are no hard rules about
how one come up with these numbers

29 / 46
The end of deterministic sensitivity analysis
That’s all about deterministic sensitivity analysis. Same rules apply to
changing two variables at the same time, although it’s harder to
communicate what happens
I covered two-way sensitivity analyses in the decision tree class
Sometimes you want to do a scenario analysis
It could involve changing several parameters at the same time. Say, if
the cost of the combo medication is reduced by 25% and the
government manages to reduce hospital costs, would the new drug be
cost effective?
This is a preview of our policy class, but one reason drug prices
(technology in general) are cheaper in countries that that use
cost-effectiveness is because companies could do a CEA study and
figure out at which price the new drug or technology will be cost
effective, which means that it’s more likely to be approved. In the US,
pricing is more or less like this: how high can we get away with?
30 / 46
Probabilistic Sensitivity Analysis (PSA)

Instead of changing one or more variables at a time we will change all

(or many) variables simultaneously using a simulation (sometimes
called second-order Markov simulation)
Instead of coming up with a range of possible values for each variable
we will assume that a parameter has a probability distribution
Remember that a probability distribution assigns a probability to each
possible outcome of a random variable
We can describe a distribution using its parameters
For example, a normal distribution can be described by its mean and
variance (two-parameters)
Another way: assuming a probability distribution for each parameter
is equivalent to assuming many, many possible values a parameter
could take (as many as infinity in most cases)

31 / 46
Stats 101 to understand the logic of PSA

Confidence intervals: If we could repeat the experiment many, many

times, 95 percent of the time the value of the parameter would be
contained within the upper and lower bound of the confidence interval
In the frequentist framework, the confidence interval is not a
“probability” interval in the sense that the confidence interval is NOT
defined as saying that we are “95 percent confident” that the true
value of the parameter will be within the confidence interval
That’s a bit closer to the Bayesian interpretation of confidence
internals (aka credible intervals)
Language is so important. How we call things changes how we
understand them. And it can lead to confusion

32 / 46
Stats 101 to understand the logic of PSA

Currently the fights between Bayesians and Frequentists have mostly

turned into love affairs
Under flat priors, both produce similar results so no need to argue so
much about which approach to statistics is better
Flat priors mean that we approach an analysis assuming that we don’t
know anything about the true value of a parameter
But the idea that we repeat the experiment many times to come up
with a confidence interval is key to make sense of probabilistic
sensitivity analysis
So what is PSA? A way to come up with a confidence interval for a
CEA study by “repeating” the study many many times (with a
simulation) to understand the possible values the ICER could take –
and thus quantify “uncertainty” (of the parameters)

33 / 46
Digression: Hypothesis testing
There is a direct connection between confidence intervals and
hypothesis testing you learned in stats 101
Suppose we know that the 95% confidence interval for the relative
risk is (0.4 - 0.6) and we want to test the hypothesis that the relative
risk is zero. So the null is H0 = 0. Would we reject the null?
Yes, of course. Because if we repeat the experiment many times, 95%
percent of the time zero is not within the confidence interval (thus
the error is 5%)
What about testing if the relative risk is 0.5? We would not reject the
null
“Counting” and repeating experiments is fundamental to statistics.
Think about calculating the probability of obtaining a 2 rolling a dice.
We calculate that by rolling the dice many times and counting the
times 2 comes up

34 / 46
PSA and Groundhog Day

Think of PSA as a kind of Groundhog Day experiment

Bill Murray’s character, Phil, keeps repeating the same day. Many
times. Let’s say 1000 times (we are never told how many times)
He is trying to get Andie MacDowell’s character, Rita, but he gets
rejected and slapped pretty much all the time. Being a Hollywood
movie, he does succeed – once (after he becomes a nice person, of
course)
We could use this “experiment” to quantify the probability that Phil
gets Rita or the uncertainty: 1 in 1000 times
Everybody’s behavior in the movie kept changing with every new day,
within certain parameters. But only once he succeeded

35 / 46
Simulation or how we get plausible numbers
If we assume that a variable (in this case, a model parameter) has a
distribution and we have the parameters of the distribution, we can
use software to generate a sample of possible values
Think of this as asking the computer to pick random numbers for
you. The only condition is that these numbers must come from a
probability distribution with certain characteristics
For example, I used Stata to simulate 10 draws of a normal
distribution with mean of 100 and standard deviation of 50. The first
ten numbers were:
52.3, 125.6, 187.7, 113.5, 113.8, 138.1, 74.3, 77.4, -21.8, 37.0
Pay attention to the numbers. All are reasonably close to 100, the
mean. How close depends on the standard deviation, in this case, 50
Digression: technically, computers generate pseudo-random
numbers; need a “seed”
36 / 46
Normal distribution

Simulated values for a normal variable with mean 100 and standard
deviation of 50; N(100, 502 )

37 / 46
PSA

The procedure to conduct PSA is fairly simple:

1 Assign a probability distribution to each model parameter
2 Ask the computer to draw a number from these distributions
3 Calculate ICER (or any other outcome of interest but mostly the ICER)
4 Save the new ICER
5 Repeat procedure many times (1000 or 10,000 times)
6 Plot all ICERs in a cost-effectiveness plane
7 Calculate statistics (like the percent of time the ICER is below
threshold)
The resulting “cloud” of ICERs will give you a sense of the overall
uncertainty of the study

38 / 46
PSA for HIV example
Campbell et al (2014) performed a PSA of the HIV example
Assumed distributions

39 / 46
PSA for HIV example
Performed 10,000 simulation draws

40 / 46
Why is PSA important?

It gives you an idea of the overall uncertainty of the study

If the points are scattered all over the four quadrants, there is a lot of
uncertainty
In general results are presented as the percent of times the ICER is
below the threshold line
Obviously, less uncertainty means that our model is less sensitive to
the value of parameters, which is good
PSA is essentially trying to mimic frequentist statistical inference

41 / 46
Choosing probability distributions for probabilistic
sensitivity analysis
The normal distribution is the first one you learn but in CEA few
parameters distribute normal
Costs, for example, cannot distribute normal because they have to be
> 0 and costs are usually skewed (but prices are often close to
normally distributed)
Utilities and probabilities must be bounded between 0 and 1 so they
can’t be normally distributed
Common distributions in PSA are Gamma, Beta, and Log-normal,
Poisson depending on the parameter
Obviously, the domain or support of the distribution must match the
characteristics of the parameter. It would be a bad idea to use a
normal distribution to model the number of test needed, which are
integer, positive numbers like 0, 2, 3, etc

42 / 46
Example distributions

See Stata code for simulations

43 / 46
Characteristics

Gamma: γ(k, θ) has two parameters (shape and scale), support from
> 0 to infinity. Used to simulate costs because costs are positive and
skewed
Beta: β(a, b) has two shape parameters, support from 0 to 1. Good
for simulating utilities or probabilities because they are bounded
between 0 and 1
Log-normal: the exponent of a normal distribution exp(N(µ, σ 2 )).
Support from 0 to infinity
Note: support is also called the domain of a distribution; the range
of values for which the distribution is defined

44 / 46
How to simulate numbers

Most statistical packages (SAS, Stata, R, Excel) have random

numbers generators for common distributions
Inverse Transform Method: It is possible to simulate numbers from
any (continuous) distribution by using a uniform random generator
We only need 1) a random generator for the uniform distribution and
2) the formula of the inverse cumulative distribution (may not exist in
closed form)
Now most software has random number generators so no need for the
inverse transform method, but you still see this because that’s how
some people learned how to simulate random numbers
See Excel file for examples

45 / 46
Summary

We have used Markov models as examples, but everything we have

learned apply to decision trees
Sensitivity analyses are vital parts of CEA
PSA may be complicated to implement but it is very easy to
understand

46 / 46