Decision Trees
Decision Trees
Decision Trees
Hemant Kombrabail
DECISION TREES
No technique can eliminate the risks involved in taking decisions, but managers can help
themselves greatly if they adopt a logical approach to decision-making One method of
considering all the options available and the chances of them occurring is known as
decision trees A decision tree is a graphical device depicting the sequences of action-
event combinations This device is a diagram that is drawn to represent three main
features of a decision
• all of the options open to a manager;
• the different possible outcomes resulting from these options,
• the chances of these outcomes occurring
In brief, a decision tree may be defined as a diagram that sets out the options connected
with a decision and the outcomes that may result from 'chance', following these options.
It is constructed from left to right
Each branch of the tree represents an option together with a range of consequences or
outcomes and the chances of these occurring
Decision points are denoted by a square - these are decision nodes.
A circle shows that a range of outcomes may occur - a chance node.
Probabilities are shown alongside each of these possible outcomes. These
probabilities are the numerical values of an event occurring - they measure the
'chance' of an outcome occurring
The pay-offs are the expected financial gains or losses of a particular outcome
By comparing the likely financial results from each option, the manager can minimize the
risks involved
e.g. The manager of an event management company has to decide between holding a
fund-raising auction indoors or outdoors. The financial success of the event depends not
only on the weather but also on the decision to hold it indoors or outdoors.
1
TYBMS Prof. Hemant Kombrabail
Using our tree diagram above, which option would give the highest expected value -
holding the event indoors or outdoors? The answer is gained by calculating the expected
value at each of the chance nodes. The cost of each option is then subtracted from this
expected value to find the net return. This is done by working through the tree from right
to left as follows
2
TYBMS Prof. Hemant Kombrabail
Advantages
1. The decision tree approach structures the decision process and thus helps one in
making a decision in a systematic manner
2. The approach necessitates that the decision maker considers all possible outcomes
regardless of whether they are desirable or undesirable Thus, no possible outcome is
likely to be left out in decision making
3. The decision tree approach is helpful in communicating the decision-making process
to others in a very succinct manner, clearly indicating the assumptions used
4. Attention can be focused on each individual financial figure, probability, as also the
underlying assumption, one at a time
5. The decision tree can be used with a computer which means that different sets of
assumptions can be used to ascertain their influence on the final outcome
Limitations
1. Decision trees need time and money to complete As such, they are unsuitable for
minor decisions where their cost may exceed the benefit to be derived from them
3
TYBMS Prof. Hemant Kombrabail
BAYESIAN ANALYSIS
If the analysis based on the decision tree is not helpful to the management in its decision
making, a more sophisticated technique known as the Bayesian analysis can be tried out
Sometimes managers find that prior probabilities regarding certain states of nature are no
longer applicable For example, a firm manufacturing readymade garments finds that
certain garments are not selling in the market on account of their color It, therefore, has to
use a different color combination In the meanwhile, it has to revise its prior probabilities
It may be noted that prior probabilities are changed after getting some additional
information As probabilities can be revised on account of the availability of new
information, probability theory is of considerable importance in decision making
There can be three types of analysis while using the Bayesian approach - prior analysis,
posterior analysis and pro-posterior analysis.
Prior Analysis
While deciding which course of action should be chosen, the decision maker uses prior
probabilities only These probabilities are prior to the receipt of any new information
Posterior Analysis
This involves the use of posterior or revised probabilities while deciding on the course of
action Prior probabilities are revised by the decision maker on receiving new information
on the states of nature
Pre-posterior Analysis
This analysis deals with the strategic question of whether new information should be
obtained and, if so, how much before deciding the course of action
4
TYBMS Prof. Hemant Kombrabail
An example will make these steps clear. It covers all the three analyses—prior, posterior
and pre-posterior
Prior Analysis On the basis of this information, prior analysis will give the expected
monetary value This will be
EMV(A1) = (Rs 1,00,00,000 x 0.7) + Rs 10,00,000 x 0 1) + (- Rs 50,00,000 x 0.2)
= Rs 70,00,000 + Rs 1,00,000 - Rs 10,00,000 = Rs 61,00,000
EMV(A2) = Rs 0
5
TYBMS Prof. Hemant Kombrabail
In the first column of the table, three possible outcomes have been identified The second
column gives prior probability of each of the three possible outcomes Now suppose a pre-
test result on this promotion was favorable In view of this, the manager must assess the
conditional probability of getting a favorable pre-test given the various possible outcomes
Column 3 of the table gives such probabilities The joint probability of R and Si i e P(R
and S,) is obtained by multiplying the probabilities in columns 2 and 3 The joint
probabilities for all the three possible outcomes are shown in column 4 The last column
gives the posterior probabilities These have been calculated by applying Bayes' rule
Symbolically,
These are the probabilities of various outcomes given the test results Thus, the posterior
probability of very favorable outcome (S1) given favorable test result (R) = 0.42/0.47 = 0
894 Similarly, the posterior probabilities of S2 and S3 can be calculated The posterior
EMV(A1) is calculated as follows
Posterior EMV(A1) = (Rs 1,00,00,000 x 0.894) + (Rs 10,00,000 x 0.064) + (Rs -
50,00,000 x (0.042) = Rs 89,40,000 + Rs 64,000 - Rs 2,10,000 = Rs 90,04,000
6
TYBMS Prof. Hemant Kombrabail
It will be seen that prior EMVPI was higher (Rs 10,00,000) than posterior EMVPI (Rs
2.10,000) This makes sense as it shows that on account of the new information, the
degree of uncertainty has declined Consequently, the value of additional information has
reduced
• Pre-posterior Analysis
Having discussed posterior analysis, we now turn to pre-posterior analysis, which is
helpful in evaluating the worth of research before it is undertaken
Suppose that in our earlier example the marketing manager was conducting a test market
for a special soft drink promotion The test would cost Rs 80,000 The first problem is
whether to have a test market or not This is followed by the market test (if it is
undertaken), then the decision alternatives about the promotion and finally the possible
outcomes of these decision alternatives What follows is a step-by-step procedure with pre
posterior analysis
Step 1 The marketing manager thinks that there are likely to be three test market
outcomes (i) a 15 per cent increase in sales (T1), (ii) a 5 per cent increase in sales (T2),
and (iii) no increase in sales (T3). He has now to obtain conditional probabilities of test
market results as shown in Table below
The conditional probability indicates that if the test market is run as proposed and
receives a favorable reaction from consumers, there will be certain probability of
occurrence of T1, T2 and T3. Thus the probability of having a +15 per cent test result given
a very favourable test market P (T1/S1) is 0.6.
These conditional probabilities P (Ti/Sis) are multiplied by the prior probabilities P (Si), to
give joint probabilities. For example, P(S1 and T1) = P(S1) x P(T1/S1) = 0.6 x 0.7 = 0.42.
All these joint probabilities are shown in Table below. The marginal probability of each Ti
can be obtained by adding the joint probabilities where Ti occurs. Thus the marginal
probability of T1 - P(T1) is 0.47; of T2 - P (T2) is 0.29; and of T3 P(T3) is 0.24. These
probabilities are also shown in Table below.
7
TYBMS Prof. Hemant Kombrabail
Step 2 The prior probabilities of the possible outcomes can now be revised. These
probabilities are revised by using Bayes' rule. For example,
Similar calculations can be easily made for Si’s given T2 and T3. These results are shown
on the decision tree. The expected pay-off of each alternative is calculated again using the
posterior or revised probabilities. On the basis of these calculations, the best courses of
action and pay-offs are T1 = Rs 87,94,000; T2 = Rs 71,02,000; and T3 = Rs 0. These three
best pay-offs are then multiplied by their associated P(Ti).
Step 3 The sum of the best outcomes, each multiplied by its respective P(T i), comes to T1
= Rs 41,33,180; T2 = Rs 20,59,580 and T3 = Rs 0.
This is the maximum amount that can be paid for this research.
8
TYBMS Prof. Hemant Kombrabail
Step 6 Since the expected monetary gain is positive, it is advisable to undertake the test
market
Figure below presents the structure of the problem in the form of a decision tree It may
be noted that the tree diagram shows the decision problems in proper sequence
It may be pointed out that one may consider alternative designs for test market, each
having different costs and conditional probabilities In view of these factors, the EMGII
9
TYBMS Prof. Hemant Kombrabail
will also be different for each alternative design In the same manner (as in the preceding
example), we have to calculate EMGII for each research design or test market Obviously
our choice will be in favor of that design which gives the highest EMGII
It is felt in some quarters that the Bayesian analysis is an abstruse method and is more an
academic exercise than a realistic decision method It is true that such a decision method
is seldom used on account of the considerable effort involved At the same time, the
Bayesian analysis covers more information than traditional methods and, as such, it
should enable one to arrive at correct decisions more frequently than would otherwise be
possible There is another benefit of this method Many a time the marketing researcher is
given the direction by the management and is not associated with the structuring of a
problem However, as the requirement of data for the application of the Bayesian analysis
can be best fulfilled through the interaction of manager and researcher, it facilitates
communication between them
10