Professional Documents
Culture Documents
Dutta Developmentutilityfunction 2010
Dutta Developmentutilityfunction 2010
Dutta Developmentutilityfunction 2010
REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/40608181?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
Operational Research Society and Palgrave Macmillan Journals are collaborating with JSTOR to
digitize, preserve and extend access to The Journal of the Operational Research Society
www.palgrave-journals.com/jors/
Insurance as a financial instrument has been used for a long time. The dramatic increase in competition
within the insurance sector (in terms of providers coupled with awareness for the need for insurance) has
concomitantly resulted in more policy options being available in the market. The insurance seller needs to know
the buyer's preference for an insurance product accurately. Based on such multi-criterion decision-making,
we use a logarithmic goal programming method to develop a linear utility model. The model is then used to
develop a ready reckoner for policies that will aid investors in comparing them across various attributes.
Journal of the Operational Research Society (2010) 61, 585-593. doi:10.1057/jors.2009.26
Published online 25 March 2009
1. Introduction
development of a utility function is very important in the
The liberalization of the Indian insurance sector has resulted Indian context. This paper is also possibly one of the first
in a number of insurance companies entering the market. attempts to develop a utility model of insurance products
using LGPM. We believe this work will help the buyer of an
This has led to a plethora of choices both in terms of service
providers as well as products to the consumers. With the insurance product to compare the existing products thereby
making a better-informed buying decision.
huge untapped market that still exists, the insurance market in
This paper is organized as follows. Section 2 discusses the
India is expected to increase rapidly. In this paper we attempt
overview of the insurance market in India and the increasing
to develop a ready reckoner to match the buyer's require-
importance it is garnering in that context. Section 3 outlines
ment with the products that the insurance companies are
a literature survey on the subject. In Section 4, we describe
offering. This will aid policyholders and potential investors
in comparing the various policies being offered. the proposed model. In Section 5, we describe the loga-
rithmic goal programming model. In Section 6, we outline the
In brief, we attempt to find
methodology followed to implement the model and arrive at
the results. In this section, we also show the results obtained
(a) the attributes of the product that a policyholder is looking
for, using a sample data set and then extend it to the larger case.
Section 7 talks about the findings from the models and expla-
(b) given these attributes, the relative weights of various
attributes, nation of the output. Finally, Section 8 concludes the paper
(c) a way to compare the existing insurance products based as well as outlines the possible extensions that could possibly
on the framework developed, and arise from the work discussed in this paper.
In terms of the insurance domain, Basu et al (2004) vidual attributes, given some type of compositional rule, that
discuss the use of different operational research and manage- are more consistent with the respondents' overall preference.
ment science models in various insurance applications. Das On the contrary, the LGPM is a compositional and build-
and Basu (2003) also have used an optimization techniqueup approach in which total utility for some multi-attribute
to obtain the optimal premium in the case of automobile product (or service in this case) is found to be a weighted
sum of the products (or services) perceived and associated
Table 1 Comparison of per capita insurance value ratings. Moreover, it is important to point out that a key
distinction between these approaches lies in the predominant
Country Life premium per capita US $ in 2005
purpose for which it is used. Conjoint Analysis generally
Japan 2956.3 emphasizes predictive validity and regards explanation as
UK 3287.1
the desirable (but secondary) objective, whereas the LGPM
USA 1753.2
that we are discussing in this paper is true for an expectancy
India 10
value theorist.
1 Equally important Two decision elements (eg, indicators) equally influence the parent
decision element
3 Moderately more important One decision element is moderately more influential than the other
5 Strongly more important One decision element has a stronger influence than the other
7 Very strongly more important One decision element has significantly more influence over the
other
9 Extremely more important The difference between influences of the two decision elements is
extremely significant
2, 4, 6, 8 Intermediate judgment values Judgment values between equally, moderately, strongly, very
strongly, and extremely
Reciprocals If ν is the judgment value when i is compared to j, then 1/v is
the judgment value when j is compared to i
In this paper, we use the ΑΗΡ- and LGPM-based method- d. We examine for consistency because during the pair-
ology over the Conjoint Analysis methodology. The reasons wise comparison, discrepancies might occur between
for the use of the ΑΗΡ- and LGPM-based methodology are the results of the comparison and the decision. In ΑΗΡ,
due to the fact that no such usage has been seen in the context before computing the weights based on pairwise judg-
of insurance products and that this method also allows for an ments, the degree of inconsistency is measured by the
alternate approach to be used for obtaining customer prefer- Inconsistency Index (II). Perfect consistency implies a
ences and then using them for further analysis. value of zero for II. However, perfect consistency cannot
be demanded because, as human beings, we are often
4. The model biased and inconsistent in our subjective judgments.
Therefore, it is considered acceptable if II ^0.1. For II
In order to enable a customer to choose the right insurance values greater than 0.1, the pairwise judgments may be
plan/policy, we decided to use the linear utility value concept.
revised before the weights are computed.
The utility function U (X) is defined as follows:
Now we try to implement the LGPM methodology developed
U{X) = Yjwixi (1) by Bryson and Joseph (1999) to ascertain the weights of the
i
criteria in the final model. We start with a set of criteria that
In this paper, we describe a method of finding the weights (w[) 5. Logarithmic goal programming model
by LGPM. The methodology followed includes the LGPM
In goal programming we minimize underachievement and
method developed by Bryson and Joseph (1999) and ΑΗΡ
overachievement from the required goal. In conventional goal
methodology developed by Saaty (1980) and Aczel and Saaty
programming we have these variables (underachievement and
(1983).
overachievement) in linear forms. In LGPM, we take the
We first follow the steps detailed in the ΑΗΡ
product of underachievement and overachievement and take
methodology - the steps are as follows:
their logarithms. The basic idea as suggested by Bryson and
Joseph ( 1 999) is that minimizing a function with two variables
1 . Establish the hierarchical structure by interviewing experts,
(linear in the objective function and summed over indices) is
conducting surveys and analyzing the elements that might
similar to minimizing logarithms of the product of the vari-
affect the survey results. The elements of similar impor-
ables (product computed over the same indices).
tance are collected on the same level in this step.
We first define the sets, indices and parameters of the model
2. The calculation of the weight of different attributes is
obtained through the following four steps: / = set of first criterion I = (1,2,3, i ... I max) indexed
by i
a. The attributes comparison is conducted in this step inJa = set of second criterion J = (1,2,3 ... j ... J max)
pairwise manner. The scale used in this comparison is indexed by j
shown in Table 2.
L = Link or pair of criteria (/, j) where / e I and j e J
b. The priority vector is calculated by dividing each ίφί
comparison value by the sum of the values in corre-Τ = set of decision-makers indexed by t, Τ =
sponding fields for the aggregation of the rows; namely, (1,2, . ..*... Γ max)
the sum of the proportion each comparison value
a'- = the ratio of the response to the /th attribute with
occupies in its corresponding row. respect to the response for the 7 th attribute for the
tth respondent, where t e Τ and (/, j) e L
2_^cell'alueilcolumnsumi (2) pt.j = the value generated by the methodology used in this
/
work for a given respondent t for the pair (/, j)
Equation (2) shows the sum of the percentage each where t € Τ and (i, j) e L. The computation of the
comparison value occupies in its corresponding row. value of p'- is explained in Equation (3) given later.
An η χ 1 matrix is acquired in this step. qjj = a value generated by the methodology used in this
c. The maximum eigenvalue is computed by multiplying work for a given respondent / for the pair (/, j)
the entire matrix with the acquired priority vector to where t e Τ and (/, j) e L
produce a n χ 1 matrix and then dividing this matrix by Vj = is the decision variable of the LGPM (not
the priority vector to acquire unit vectors. The average normalized)
of the unit vectors is sequentially calculated to acquire Wj = normalized decision variable or the weights of
the maximum eigenvalue. different attributes
LGPM involves a linear goal programming model in which The optimal solut
the objective is to generate a group mean priority point vector normalized vector
w = (w', W2, . . . , wn) such that for the comparison between normalized to giv
each pair of criteria '/' and 'y', the difference between the vector w = (w', w
ratio (wi/wj) and the decision-makers specified a'- is mini- each (i, j).
mized. Please note that data value of / max defined in this Properties of the
model is N.
(a) This logarithm
Let there be real numbers p'- > 1, q'- > 1 such that infeasible.
Therefore, if p'- = q'- = 1 for each pair of criteria Τ and 'y ', Dominance Dominance occurs when one option pe
then the set of point estimates provided by the decision-maker least as well as another on all criteria and strictly be
7' is consistent; otherwise the data are inconsistent and our the other on at least one criterion. In principle, o
problem then is to minimize the product Πϊε/Π/ε//7!/^//· might dominate all others, but in practice this i
Aczel and Saaty (1983) suggested that the group Once the decision-maker has ranked the different insurance
'consensus' pairwise comparison values should be the products on the basis of the criteria, an initial step can be to
geometric mean of the individual pairwise comparisonsee if any of the options dominate others.
values. Rather than focusing on each pairwise comparison,
we will focus on the entire set of pairwise comparison values. Linear additive model According to a study by Cochrane
Therefore, if Τ is the index set of the decision-makers and and Zeleny (1973), if dominance is not observed, and it can
M = | Γ |, then it follows that for the group, our problem iseither be proved, or reasonably assumed, that the criteria are
to minimize the product for all respondents t e Τ and each preferentially independent of each other and if uncertainty is
value of (/, j) e L. Thus the problem is to minimize the termnot formally built into the MCDA model the simple linear
additive evaluation model is applicable. The linear model
Y[teTi'ielY'je(iJ)eLPijaij·
This translates to solving the following linear goal shows how an option's values on the many criteria can be
programming problem where the decision variables are the combined into one overall value. This is done by multi-
un-normalized vector (v', i>2, · · · , vn) · In this case, insteadplying the value score on each criterion by the weight of that
of minimizing the product, we minimize the logarithm of thecriterion (determined by LGPM), and then adding all those
product and use it in a linear programming problem. weighted scores together. In case there are two parameters,
which are not preference-independent, then one approach can
be to club together the two variables and treat them as one
variable. Otherwise a more complicated MCDA-based model
Minimize Ζ = ln(0) = ( - j J^ 1η(Θ' ) (4)that calculates correlation among the parameters can be used,
but we will not go into that.
subject to Into) - 1η(υ7) + ln(^) - Ιη(^) = ln(^)
VteT;(iJ)eL (5)
6. Methodology
In this section, we describe the data collection method
^ ' iel jeJ
employed in this study and the modelling details. Further,
in Section 6.2, we describe the exact workings of the model
V t e T (6) using a smaller data set and a reduced set of factors. This, we
hope, will significantly enhance the readability of the paper.
where Κ = Ν χ (Ν
Finally, -
Section 6.3 talks 1); / =implemen-
about the mathematical {1, 2
are tation of the full scale model described in this paper.
non-negative.
6.1. Data collection is less than 1 min. The other advantages of using AMPL are
as follows:
We collected information on the life insurance policies and
conducted a survey to determine the relative importance given
(a) The model in the computer code can be written in the
to these factors. We first demonstrate the computational part
same way as the formulation is written.
by having a sample data set which we solve in the Excel
(b) Model data independence.
solver (this is explained in Appendix A). This sample data has
(c) Model solver independence.
10 respondents and eight attributes. In this sample study, we
(d) There is no restriction on the number of variables.
found that there is no added advantage to having a scale of
1-100 and hence, we reduced the choice set to 1-9. The ques-
We grouped the data suitably to identify if respondents in
tionnaire is provided in Appendix B. For the life insurance
different salaried classes had allotted different weights.
study, we had 102 respondents. However, some responses
were inaccurate or had missing data points; excluding these,
we were left with 87 respondents. It should be noted7.that
Model findings
the proposed method of generating priority point vectors in
We evaluated the existing life insurance policies using the
the literature suggests that each respondent should specify
weights. Specifically, we looked at money-back life insurance
values for each pairwise comparison. Since we had a large
number of factors for each category of policy, we decided to (where maturity benefits are paid in installments at
policies
fixed intervals).
reduce the number of values input to 1 for each factor. We
then derived the pairwise comparison values. The surveyKey
wasparameters for life insurance policies
conducted among predominantly salaried or self-employed
persons.
1. Low premium
In the second part of the survey, we requested the respon-
2. Flexibility in payment structure
dents to rate the existing products available in the market. Tax benefits in insurance plan
3.
4. Benefits on death
Each respondent was told to rate each product on each
5. Benefits on survival
attribute.
6. Good customer service
Table 3 Weights assigned from the model Table 5 Description and abbreviation of the parameter
Table 4 Respondent's profile and consistency indicators Since different attributes have different weights, an insur-
ance provider will concentrate on those factors where the
Salary per month Number of Consistency
responses indicator weights are higher. So an insurance provider will be interested
in improving on BOD than SS.
Less than Rs. 10,000 8 0.5578
Assuming that there are only these six insurance products
Rs.l0,000-Rs.20,000 30 0.7690
Rs.20,000-Rs.30,000 30 0.7460 in the market, in Table 7, we calculate the utility score of
Rs.30,000-Rs.40,000 8 0.7795 each insurance product as assigned by the buyer. We then use
Rs.40,000-Rs.50,000 7 0.4962 a multinomial logit model (Luce, 1959; Luce and Suppes,
More than Rs. 10,000 4 0.7350 1965; McFadden, 1980). The multinomial logit model then
Aggregate 87 calculates the probability that a customer will buy one of the
products.
Weights^ 0.1141 0.01059 0.1362 0.1319 0.1211 0.1014 0.0978 0.0856 0.1059
PI 453445345 4.07
P6 343444454 3.83
P3 343334353 3.37
P5 343433432 3.22
P4 534232343 3.21
P2 333224332 2.74
Table 7 Results
Cochrane JL and Zeleny M (1973). Multiple Criteria Decision Making.
University of South Carolina Press: USA.
Product SCORE(Ui) EXP(Ui) Prob(Pi)
Das S (2003). Joint life insurance policies with differential benefits
name
and premiums to the policyholders. Working Paper No. 208, Indian
PI 4.07 58.556962 0.294268 Institute of Management, Bangalore.
P6 3.83 46.062538 0.231479 Das S and Basu S (2003). A Markov chain based no-claims discount
P3 3.37 29.078527 0.146129 system. / Ins Risk Mngt 2/3: 61-72.
P5 3.22 25.028120 0.125774 Fourer R, Gay DM and Kernighan BW (1993). AMPL (A Modeling
P4 3.21 24.779086 0.124523 Language for Mathematical Programming), The Scientific Press
P2 2.74 15.486900 0.077827 Series. Boyd & Fraser Publishing Company: California, USA.
Green PE and Srinivasan V (1978). Conjoint analysis in consumer
Total 198.992133 1 research: Issues and outlook. J Consum Res 5: 103-123.
Green PE and Srinivasan V (1990). Conjoint analysis in marketing:
New developments with implications for research and practice. /
The linear additive model can also be extended to the Mark 54(4): 3-19.
India Core (2007). Overview of Insurance Sector, www.indiacore.com/
context of revenue management. As revenue management is
insurance.html, accessed 13 March 2008.
an emerging research and application area, one of the impor-
Luce R (1959). Individual Choice Behaviour: A Theoretical Analysis.
Wiley:isNew York.
tant considerations from the point of the service provider
Luce has
the utility of a customer. We believe no research work R and Suppes Ρ (1965). Preference utility and subjective
probability. In: Luce R, Bush RR and Galanter EH (eds).
been done in the Indian context and the logarithmic goal-
Handbook of Mathematical Psychology, Vol. 3. Wiley: New York,
programming model may be applicable in this context asppwell.
249-441.
The second extension is based on the assumption of HM
Markowitz (1952). Portfolio selection. J Financ 7: 77-91.
linearity. We have assumed the utility function to be linear. D (1980). Econometric models for probabilistic choice
McFadden
This may be a good approximation for the development amongof the products. J Bus 53(3 Part 2): S13-S29.
Saaty TL (1980). The Analytic Hierarchical Process. McGraw Hill
utility function. However, in reality, the utility function may
Book Company: New York.
be non-linear. In such cases, it may be possible to find the
Srinivasan V (1980). Comments on 'On conjoint analysis and quantal
second and higher order terms for each attribute in the choice
model. models'. J Bus 53(3 Part 2): S45-S46.
The real challenge of this problem will be to collect data V (1998). A strict paired comparison linear programming
Srinivasan
approach to nonmetric conjoint analysis. In: Aronson JE and Zionts
through market surveys for computing the higher order terms.
S (eds). Operations Research: Methods, Models and Applications.
Although we have shown that LGPM is a method to develop
Quorum Books: Westport, CO, USA, pp 97-112.
a linear utility function for the insurance method, it does
Srinivasan V, Jain AK and Malhotra NK (1983). Improving predictive
not show how this method compares with that of Conjoint power of conjoint analysis by constrained parameter estimation.
Analysis. A comparative study will be an interesting extension
J Mark Res 20: 433-438.
Table A.5 Optimal solution that has been used in the computation
12 3 4 5 6 7 8
We now show the optimal solution of the model with the normalized values in Table A.5.
Appendix B.
Dear Sir/Madam
We are doing a study on Utility Function of Life Insurance Buyers in the Indian Institute of Management, Ahmedabad (IIMA
We would like to know the factors you would consider when choosing a life insurance policy. Filling this questionnaire shou
not take you more than 3 minutes.
Thank you
Your Name:
Following are some factors people consider while buying an insurance policy. Indicate the importance of each facto
influencing your decision to buy a particular life insurance policy. Indicate the importance on a scale of 1 (not at all imp
to 9 (very highly important).
• Online payment
• Renegotiation of term/insured amount
7 Bonus