Decision Tree analyses in HTA

Vilasini Devi Nair

 Decision Tree Modelling
A method used in health economics is
decision tree modelling, which extrapolates
the cost and effectiveness of competing
interventions over time.
Such decision tree models are the basis of
reimbursement decisions in countries using
health technology assessment for decision
making.
09/04/23 2
 Decision Tree
A decision tree is a flow chart with one
main idea and then branches out based
on the consequences of your decision.
It is called a decision tree because the
model typically looks like a tree with
branches.
These trees are used for decision tree
analysis which involve visually outline
the potential outcomes, cost and
consequences of a complex decision.
09/04/23 3
 A decision tree consists of a series of
'nodes' where branches meet: each node
may take the form of a 'choice' (a decision
about which alternative intervention to use)
or a 'probability' (an event occurring or not
occurring, governed by chance).
Probabilities at any specific node must
always add to 1.

09/04/23 4
09/04/23 5
09/04/23 6
 What Decision Tree explains?
Decision tree analysis involves visually
outlining the potential outcomes, costs, and
consequences of a complex decision. These
trees are particularly helpful for analyzing
quantitative data and making a decision
based on numbers.

09/04/23 7
 Decision analysis is now extensively
used for economic evaluation modelling
in health care.
It is used when the outcomes of an
event are uncertain, but it is possible to
assign a probability to each possible
different outcome.

09/04/23 8
et.

09/04/23 9
 For example, a drug may not cure every case
that it is administered for, but is known to
work in 80% of cases and has known side-
effects in 5% of cases; or
A surgical operation has been found to
succeed in 75% of cases but has an operative
mortality rate of 1%.

09/04/23 10
 Decision analysis is particularly valuable when
decisions are not simply about a single choice
between actions, but require strategies that
incorporate multiple choices about different
actions.
For example, if a drug does not work then a
second-line drug may be used, or if surgery
does not work then drug therapy may be
tried.  Decision analysis is a way of
structuring the different options so that the
best strategy can be determined.
09/04/23 11
 Decision analysis relies on the concept
of expected values.
The expected value of something can be
thought of as its average value when it is
repeated many times.
For example, if a treatment produces a
gain of 2 Quality Adjusted Life Years
(QALYs) per person in 80% of cases and a
gain of 1 QALY per person in the remaining
20%, then out of 100 patients the total
gain will be (2x80) + (1x20) = 180. 

09/04/23 12
 This means that the average gain is 1.8
QALYs per patient, which is the
expected value.
In reality, it is unlikely that the
probabilities of 0.8 and 0.2 would
produce exactly 80 and 20 in a
population of 100, which is why the
assumption of large numbers is made.

09/04/23 13
 Generalizing this, the expected value of an
event is calculated as the sum over all possible
outcomes of the probability of each outcome
multiplied by its value.
For example, if an activity X has two
outcomes a and b which have probabilities
P(a) and P(b) and values U(a) and U(b), the
expected value of X is
E(X)=P(a)*U(a)+P(b)*U(b)

09/04/23 14
 In economic appraisals in health
care, decision analysis calculates
the expected values of both costs
and benefits and also ICERs,
though the last of these is more
complicated.

09/04/23 15
 They usually take the following form:
Structure the problem by constructing a
mathematical model of decision-making,
usually a decision tree:
* Identify the decisions to be made between
alternative actions and describe those
alternatives
* List the possible outcomes of each
alternative for each decision
* Specify the sequence of events that follow
from each decision

09/04/23 16
1. Assign probabilities to chance events. 
2. Assign values to all possible outcomes
of chance events. 
3. Calculate the expected value of each
possible strategy.
4. Perform sensitivity analyses by
systematically changing the
assumptions.

09/04/23 17
 This will be explained using a decision tree first
presented by Morris, Devlin, Parkin and
Spencer (2012).
A decision tree uses a diagrammatic illustration
of a decision problem. Nodes, representing key
elements of the decision problem are
connected by lines, called paths, representing
links between them.
The paths run from left to right and imply a
sequence in that order. Decisions that are
controlled by a decision maker are
called decision nodes, conventionally drawn as
squares:
 

09/04/23 18
 This shows that the decision is
between surgery and drug therapy.
The paths running from this are the
consequences of choosing a
particular alternative. It is possible
to have more than two alternatives,
but they must all be mutually
exclusive

09/04/23 19
 Any chance occurrences that are
not controlled by the decision-
maker are called chance
nodes or probability
nodes, conventionally drawn as
circles

09/04/23 20
 In this case, disease may be present or
absent. The important aspect of this is
that each of the alternatives must have a
probability attached to it - in this case
the probability that a disease is present
or absent.
There can be more than one outcome,
but they must also be mutually exclusive
and moreover they must be collectively
exhaustive; that is, the sum of their
probabilities must exactly equal one.

09/04/23 21
 The final outcome associated with a
path is a terminal node,
conventionally drawn as a triangle:

09/04/23 22
 Every path that does not lead to a chance
event or decision must have a terminal node.
In this case, the final outcome of a set of
decisions and chance events is that the patient
is cured.
A full decision tree would contain all of the
relevant nodes and the paths between them.
To this would be added the probabilities
assigned to chance events and the values
attached to each potential final outcome.

09/04/23 23
09/04/23 24
 In this case, the initial choice of surgery leads to a
chance node associated with surgery, where the
possibility of operative mortality is included. Assuming
that the patient survives, a further chance node
represents the probability that they will be cured, leading
to the final outcomes of being cured by surgery and not
cured despite surgery.
The initial choice of drug therapy leads to a node
depicting the chance that the patient is cured or not.
Being cured is a final outcome and leads to a terminal
node. If the patient is not cured, a second decision is
taken, to offer counselling, which again leads to a final
outcome, or a second line drug, which again has a
cured/not cured chance node, although this time both
lead to final outcomes .
09/04/23 25
 The expected value of the different
options can then be calculated in stages,
working from right to left. The expected
value to a patient who survives surgery
is (0.8*15) + (0.2*0.5) = 12.1 QALYs.
This value is used to calculate the
expected value of surgery for all patients,
whether they survive or not, which is
(0.05*0) + (0.95*12.1) = 11.495 QALYs.

09/04/23 26
 The expected value of using Drug 2 is
(0.5*10) + (0.5*1) = 5.5 QALYs. Since
this is greater than the value of
counselling, the decision following a
failure of Drug 1 should be to try Drug
2, so the expected value of Drug 2
therapy is used in the calculation of the
expected value of Drug 1, which is
(0.7*13) + (0.3*5.5) = 10.75.

09/04/23 27
 Surgery therefore produces, on
average, a greater QALY gain, and the
incremental benefit of surgery is
(11.495 - 10.75) = 0.745 QALYs.
Of course, in an economic evaluation, it
would also be necessary to measure the
costs of each terminal node and to
calculate the expected value of costs for
each decision option.

09/04/23 28
 Markov Model
The Markov model is an analytical framework that is
frequently used in decision analysis, and is probably
the most common type of model used in economic
evaluation of healthcare interventions.
Markov models use disease states to represent all
possible consequences of an intervention of interest.
These are mutually exclusive and exhaustive and so
each individual represented in the model can be in
one and only one of these disease states at any given
time. 

09/04/23 29
 Examples of health states that might be
included in a simple Markov model for a
cancer intervention are: 
progression-free,
post-progression and dead.
Individuals move ('transition') between disease
states as their condition changes over time.

09/04/23 30
 Time itself is considered as discrete time periods
called ‘cycles’ (typically a certain number of
weeks or months), and movements from one
disease state to another (in the subsequent time
period) are represented as ‘transition
probabilities’.
Time spent in each disease state for a single model
cycle (and transitions between states) is associated
with a cost and a health outcome.

09/04/23 31
  Costs and health outcomes are aggregated for a
modelled cohort of patients over successive cycles
to provide a summary of the cohort experience,
which can be compared with the aggregate
experience of a similar cohort, for example one
receiving a different (comparator) intervention for
the same condition. Markov models are limited in
their limited ability to ‘remember’ what occurred
in previous model cycles

09/04/23 32
 For example the probability of what occurs after
disease progression may be related to the time to
progression.
Although to some extent health states can be
defined ingeniously to address this complexity,
other modelling approaches may be required for
more complex diseases.

09/04/23 33
09/04/23 34