W P N - 18 Setting Weights in Multidimensional Indices of Well-Being and Deprivation

Oxford Poverty & Human Development Initiative (OPHI)
Oxford Department of International Development

Queen Elizabeth House (QEH), University of Oxford
OPHI WORKING PAPER NO. 18

Setting Weights in Multidimensional Indices of Well-being
and Deprivation
Koen Decancq1 and María Ana Lugo2
August 2008 [rev. March 2009]
Abstract
Multidimensional indices of well-being and deprivation have become increasingly popular, both in the
theoretical and in the policy-oriented literature. By now, there is a wide range of methods to construct
multidimensional well-being indices, differing in the way they transform, aggregate and weight the
relevant dimensions. We present a unifying framework that allows us to compare the different
approaches and to analyze the specific role of the dimension weights in each of them. Through
interaction with choices about the transformation and aggregation of the different attributes, the weights
play a crucial role in determining the trade-offs between the dimensions. Setting weights thus reflects
important value judgements about the exact notion of well-being. We survey six methods to set weights
usually employed in the literature. Three principles guides our assessments: first, weights should be made
explicit and clear so that they can be subject to public scrutiny; second, weights should be set taking into
consideration their role determining the trade-offs between dimensions; finally, weights should respect
people’s preferences about these dimensions.
Keywords: weights, multidimensional well-being index, multidimensional deprivation index
JEL classification: I32, D63
1Katholieke Universiteit Leuven, Belgium; 2 Department of Economics, University of Oxford, Manor Road Building, Manor
Road, OX1 3UQ, Oxford, UK, e-mail: [email protected]
This study has been prepared within the OPHI theme on multidimensional poverty.
OPHI gratefully acknowledges support for its research and activities from the Government of Canada through the
International Development Research Centre (IDRC) and the Canadian International Development Agency (CIDA), the
Australian Agency for International Development (AusAID), and the United Kingdom Department for International
Development (DFID).
ISSN 2040-8188 ISBN 978-1-907194-07-8
Decancq & Lugo Weights in Multidimensional Indices of Well-Being
Acknowledgements
Comments and suggestions are most welcome. This paper was originally drafted background for the
OPHI workshop on Weighting Dimensions (Oxford, 26–27 May 2008). We thank Tony Atkinson,
Sabina Alkire, James Foster as well as participants at the OPHI workshop for valuable comments and
remarks. The usual disclaimer applies.
The Oxford Poverty and Human Development Initiative (OPHI) is a research centre within the
Oxford Department of International Development, Queen Elizabeth House, at the University of
Oxford. Led by Sabina Alkire, OPHI aspires to build and advance a more systematic methodological
and economic framework for reducing multidimensional poverty, grounded in people’s experiences and
values.
This publication is copyright, however it may be reproduced without fee for teaching or non-profit purposes, but not for
resale. Formal permission is required for all such uses, and will normally be granted immediately. For copying in any other
circumstances, or for re-use in other publications, or for translation or adaptation, prior written permission must be obtained
from OPHI and may be subject to a fee.
Oxford Poverty & Human Development Initiative (OPHI)

Oxford Department of International Development
Queen Elizabeth House (QEH), University of Oxford
3 Mansfield Road, Oxford OX1 3TB, UK
Tel. +44 (0)1865 271915 Fax +44 (0)1865 281801
[email protected] http://ophi.qeh.ox.ac.uk/
The views expressed in this publication are those of the author(s). Publication does not imply endorsement by OPHI or the
University of Oxford, nor by the sponsors, of any of the views expressed.
Decancq & Lugo Working Paper No.18
1 Introduction
The notion that well-being is inherently multidimensioned has by now become well-established
in the theoretical and policy-oriented literature. Rooted in a tradition going back to Aris-
totle, philosophers such as Rawls (1971), Dworkin (2000), Sen (1985), Nussbaum (2000),
or Townsend (1979) have advocated a multidimensional perspective on the good life and
well-being, exposing the deficiencies of a sole focus on income as indicator of well-being.
In line with these views, the World Bank lead survey collecting the voices of more than
60,000 poor women and men from 60 countries, conclude -among other things– that for the
poor, well-being and deprivation are multidimensional with both material and psychological
dimensions (Narayan 2000).1
When it comes to operationalizing the multidimensional approaches, one quickly runs

into the crucial problem of how to describe individuals’ multidimensional well-being by one
single index.2 This is the so-called indexing problem (Rawls 1971, p. 80). In the literature
this problem has been addressed from two slightly different but complementary perspectives.
First, from an operational perspective rooted in measurement theory, the approach has
been to provide clear and relatively simple guidelines on how to construct composite or
social indicators in many fields, including well-being. The most popular example of such a
well-being composite indicator is probably the Human Development Index (HDI), used to
compare the performance of countries in terms of their combined achievements in income
–as command over resources–, health and education.3 The second approach uses a social
choice perspective to define measures of multidimensional welfare, inequality or poverty
and puts the emphasis on the measurability and comparability of the different dimensions,
and on desirable properties of the obtained indices.4 The results of this second strand are
1
A further support of the view of well-being as being multidimensional comes from the rapidly emerging
literature on the determinants of happiness and life-satisfaction in psychology and economics. According
to various studies, people’s happiness is affected by many aspects of life such as their health, employment,
marital status, and material resources (see Kahneman and Krueger (2006)).
2
One could reasonable argue that there is no need to aggregate dimensions of well-being into a single
index, and thus full cardinalization is unwarranted. Two alternatives are generally proposed. The first
involves considering each dimension at a time, the item-by-item approach. This might a valid approach in
the case of, for instance, a social planner interested in how the policies in each separate area are performing
and there is no need to arrive to one ranking (see, for instance, the proposal by Atkinson et al. (2004) on
indicators for social inclusion in the European Union to assess the performance of Member States in the Joint
Report on Social Inclusion). However, if she were interested instead in how her ‘citizens’ are performing in
each of the dimension of well-being, the item-by-item approach is rather inappropriate as it is oblivious to
the possible dependence between these dimensions. In other words, she would equally value a society where
all the wealth and all the health is concentrated in a group of people and one where the rich in income are
poor in health and vice-versa. The second approach uses dominance criteria based on the multidimensional
distribution function (see for instance, Atkinson and Bourguignon (1982) and Muller and Trannoy (2004) for
a theoretical formulation and Duclos et al. (2006), Crawford (1999), and Gravel et al. (2005) for statistical
formulation and applications) which requires minimal agreement of the characteristics of the underlying
welfare function. In many situations, especially when one needs to rank more than two states, dominance
conditions are not able to provide a complete ranking of distributions. Though incomplete rankings can be
to some extent be informative, an analyst or policy maker is often more satisfied with a complete ranking.
3
Many international institutions are active in using and promoting the use of composite indicators - see
Nardo et al. (2005, 2005) for a survey of some alternative composite indicators. To give one example,
the reader is referred to the detailed information server on composite indicators hosted by the European
Commission on http://composite-indicators.jrc.ec.europa.eu/.
4
See Weymark (2006) or Maasoumi (1999) for an overview of the literature on multidimensional inequal-
www.ophi.org.uk 1
mainly theoretical, although they are increasingly empirically applied.5 In this paper, we
survey the different approaches to deal with the indexing problem, particularly focussing
on the question how to weight the different dimensions of well-being. Thereby we combine
the insights from both the theoretical and operational branches of the literature.
To order and compare the plethora of existing well-being indices from the literature,
the paper starts by proposing a unifying framework in section 2. This framework reduces the
differences between the well-being indices to differences in the chosen transformations of the
original variables, the aggregation function and the weighting scheme for the dimensions.
The focus of this paper is on the weights. In section 3 we analyze the meaning of the weights
within the proposed unifying framework. Together with the choices about transformation
and aggregation, the weights will be shown to play a crucial role in the imposed trade-offs
between the dimensions. Inescapably, the weights reflect important value judgements about
the (vague) notion of well-being. Researchers should therefore be as clear as possible about
how the weights are set. As Anand and Sen argued:
“Since any choice of weights should be open to questioning and debating in

public discussions, it is crucial that the judgments that are implicit in such
weighting be made as clear and comprehensible as possible and thus be open to
public scrutiny” (Anand & Sen 1997, p. 6)
Section 4 critically surveys six proposed methods to set the weights in multidimensional
measures of well-being: equal weighting, frequency based weighting, most favorable weight-
ing, multivariate statistical weighting, regression based weighting and normative weighting.
We argue that whether the weights are set reasonably should be judged upon the accept-
ability of the implicitly imposed trade-offs by them. Section 5 concludes.
2 A unifying framework
Let us assume that agreement has been reached on the domains of well-being that are
relevant for the assessment of persons’ standard of life and, moreover, that the achievement
in all of these dimensions can be measured in a interpersonally comparable way. Let xj
denote the achievement of an individual on dimension j = 1, ..., q and let the well-being
vector x = (x1 , ..., xq ) ∈ Rq+ summarize these achievements across all dimensions.
The indexing problem can be described as the search for an appropriate well-being
index W , that maps the well-being bundle x on the real line, so that it can be naturally
ordered and used to assess the position of any two persons and the distance between them.6
ity and Bourguignon and Chakravarty (2003) for a survey of the multidimensional poverty or deprivation
literature.
5
See Justino (2005) for an overview.
6
The unit of analysis considered in this section is the individual, though one could use the present
framework for other relevant units, such as countries, regions or districts. If the unit of analysis is the
www.ophi.org.uk 2
Inescapably, the choice of a specific well-being index entails important value judge-
ments about the meaning of well-being and about the respective contribution of its com-
ponents. In the present paper we confine ourselves to the following wide class of well-being
indices:
1/β
w1 I1 (x1 )β + ... + wq Iq (xq )β

W (x|β) = . (1)
w1 + ... + wq
The individual well-being index W (x|β) is defined as a weighted mean of order β of the
transformed achievements Ij (xj ).7 The dimension-weights w1 , . . . wq are all non-negative,
and are often assumed to sum up to one so that the denominator of expression (1) drops
out. The interpretation of these weights and how to set them is the topic of this paper.
Before turning to the weights, though, we discuss briefly the other two components of the
well-being index, that is, the transformation functions Ij (.) and the parameter β.
Appropriate transformation functions for well-being indices should satisfy at least two
criteria. First, since the achievements xj are often measured in different measurement
units –such as income in pounds and health in years –, they need to be transformed or
standardized to a common basis before they can be sensibly aggregated. Transformation
functions typically make the achievements scale independent. Common examples include
the z-score and the ratio to the mean standardisations. Second, if the original distribution
is skewed the transformation functions should avoid that excessive relative importance is
given to outliers or extreme values. One example of such transformation is the logarithmic
transform. 8
Expression (1) can also be used to construct an index of multidimensional poverty or

deprivation. In this setting, the function Ij (.) is defined as a negative function of xj and
transforms the achievement in dimension j into the shortfall in that dimension. These trans-
formation functions typically request also a dimension-specific poverty-line zj to be defined.
Let us call these transformation function Dj (xj |zj ). The most widely used transformation
individual but the assessment is done at a higher level, say the country of these individuals, then this
section can be seen as the first step into the full aggregation from the space of individuals/attributes to a
real number. The second step would be the aggregation across individuals. See Decancq and Lugo (2008)
and Dutta et al. (2008) on the order of aggregation of multidimensional distributions of well-being and
deprivation.
7
Blackorby and Donaldson (1982) provide an axiomatic characterization of the weighted mean of order
β. In the literature on multidimensional inequality, Maasoumi (1986) provides an information-theoretic
justification of this class of well-being indices. Further, it belongs to the wider class of well-being indices
proposed by Bourguignon (1999) while it is similar to Foster et al (2005)’s proposal for a distribution-sensitive
measure of human development. Decancq and Lugo (2008) axiomatize it as part of a multidimensional Gini
measure and Decancq et al. (2007) have used it to analyze the trend in multidimensional global inequality.
Furthermore, in the related literature on the measurement of multidimensional poverty and deprivation, this
class of indices has been suggested by Anand and Sen (1997) and is a special case of the class proposed by
Bourguignon and Chakravarty (2003).
8
In the context of building an index of deprivation for small areas, Noble et al. (2006, 2008) include
two additional criteria for standardization: first, it should imply an appropriate degree of substitutability or
cancelation between domains, second, it should facilitate the identification of the most deprived.
www.ophi.org.uk 3
for poverty measure is the relative shortfall defined as
zj − x j
Dj (xj |zj ) = . (2)
zj
We can then defined a generalized class of deprivation indices D(x|β) as follows:

1/β
w1 D1 (x1 |z1 )β + ... + wq Dq (xq |zq )β

D(x|β) = , (3)
w1 + ... + wq
Table 1 in the Appendix surveys some widely used transformation functions in the
literature. In this paper, we do not prioritize one transformation method over another, but
limit ourselves to presenting them while highlighting the crucial role they play on the inter-
pretation of relative weights -as shown in the next section. We refer the interested reader
to Jacobs et al. (2004) and Nardo et al. (2005) for an extensive survey of the alternative
transformation methods and their properties. In general, we see that transformation func-
tions used to construct a well-being index are increasing in xj , whereas the transformation
functions used to construct an index of deprivation are decreasing in the achievements.
The second component in expression (1) is the parameter β. One useful interpretation
of β is related to the elasticity of substitution between transformed achievements of well-
1
being, σ where σ = 1−β . The smaller the β, the smaller the allowed substitutability
between dimensions, that is, the more one has to give up of one attribute to get an extra
unit (of transformed achievement) of a second attribute while keeping the level of well-
being constant. Generally, for β ≤ 1 the well-being index is a weakly concave function,
which reflects a preference for well-being bundles that are more equally distributed. When
expression (1) is used to describe multidimensional deprivation, a restriction of β ≥ 1 seems
more appropriate.
For β = 1, the weighted mean of order β is reduced to the standard weighted arithmetic
mean of the following form:
w1 I1 (x1 ) + · · · + wq Iq (xq )
W (x|1) = . (4)
w1 + · · · + wq
Due to its simplicity and clarity of procedure, expression (4) is used frequently to construct
composite indices.9 However, the consequence of setting β = 1 might not always be de-
sirable. The elasticity of substitution between (transformed) achievements is infinite and
dimensions are perfect substitutes, meaning that there is a fixed rate at which transformed
attributes can be exchanged which is constant for all possible levels of all attributes. One
9
See Jacobs et al. (2004) and Nardo et al. (2005).
www.ophi.org.uk 4
could argue that the amount of money needed to compensate for a less year of life should
be quite different whether the person is in her youth or at the end of a normal life. In other
words, it might be desirable to allow for the rate of substitution between dimensions to vary
depending on their levels. An even stronger argument is made by the Human Development
Report (2005).
“Losses in human welfare linked to life expectancy, for example, cannot be com-
pensated for by gains in other areas such as income or education” Human De-
velopment Report (2005)
At the same time, however, the leading index of the Human Development Report –the Hu-
man Development Index– makes use of a linear aggregation assuming perfect substitutability
between the transformed achievements.10
Two other, equally-simple choices –though not as widely used– are worth exploring as
they might better capture the feeling of the previous arguments. The first option is to set
β = 0 which makes the well-being index a weighted geometric mean,
W (x|0) = I1 (x1 )w1 /(w1 +···+wq ) ∗ · · · ∗ Iq (xq )wq /(w1 +···+wq ) . (5)
The well-being index in (5) has unit elasticity of substitution between all pairs of dimensions.
This means that a one percent decrease in one of the dimensions can be compensated by a
one percent increase in another dimension. Note that this formulation allows for the rate of
substitution between transformed attributes to change as the levels of achievements vary.
As argued above, this characteristic of index (5) might be considered sensible and desirable
both in theory and in practice. Figure 2 shows in a two dimension space (health and income)
the lines where the level of well-being is maintained constant, for different options of β (for
clarity of exposition of the graph we assume Ij (xj ) = xj ).
We can express (5) by its ordinarily equivalent logarithmic version so that it becomes
10
Note that the rate at which dimensions are traded off, measured in its original units -and not transformed,
is constant (though not perfect) between the pair health-education, and non constant for health-income and
education income pairs. Due to the log transformation employed for per capita GDP, the tradeoff between,
say, per capita GDP and life expectancy depends also on the level of income the country achieves. In
particular, the amount of money required to compensate for a less year of life expectancy is increasing in
income; for a rich country such as Belgium an extra year of life expectancy is valued at nearly 7,000 US$
(in PPP terms) which for a relative poor country, such as Cote d’Ivoire this is merely 300 US$. Therefore,
contrary to the claim above the Human Development Index does indeed allow for compensation between
dimensions, even when this compensation might vary across levels. In the area of poverty and deprivation,
the UNDP suggest two Human Poverty Indices (HPI) setting β = 3. “This gives an elasticity of substitution
of 1/2 and places greater weight on those dimensions in which the deprivation is larger” (Anand and Sen
1997, p. 16). For very high levels of any one attribute the compensation necessary for decreases in that
attribute tend to very a very large number -depending on the extent of deprivation- but still it allows for
some substitutability.
www.ophi.org.uk 5
140
120
100
80
Health
60 β = -∞
40
β =0
20
β =1
0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Income
Figure 1: Iso- well-being curves for β = 0, 1, −∞
a weighted arithmetic mean of the logs of the transformed achievements.
w1 wq
W (x|0) = ln (I1 (x1 )) + · · · + ln (Iq (xq )) . (6)
w1 + · · · + wq w1 + · · · + wq
Note if in addition we set the the identity function as transformation, Ij (xj ) = xj , this
expression becomes equivalent to choosing β = 0 and Ij (xj ) = ln(xj ) for all j. While one
could interpret it either way, we prefer to refer to it as W (x|0) emphasizing the curvature
of the iso-well-being lines as shown in the previous figure.
A potentially crucial problem of setting β = 0 is that when a person has no achievement

in one of the (transformed) dimensions the overall well-being index will be insensitive to
the achievements in the other dimensions.
A second alternative is to let β go to −∞(+∞) where the elasticity of substitution

becomes 0, and the well-being index becomes the minimum (maximum) of the transformed
achievements across the dimensions,
W (x| − ∞) = min[I1 (x1 ), ..., Iq (xq )]. (7)
In this extreme case, there is no substitution between dimensions possible, which seems to
reflect better the philosophy of the above UNDP Report quote.
We assumed a common degree of substitution for all pairs of dimension. For example,
if income, health and education are taken as the components of well-being, all the above
indices assume that the rate at which income and health are substituted is the same as that
www.ophi.org.uk 6
between income and education or health and education. We chose to do so for simplicity
of exposition. Nonetheless, this might not be always a sensible assumption to make. One
alternative is to use a nested approach where, first, several subsets of dimensions are aggre-
gated using expression (1) with each subset having a different β and, second, these subsets
are combined using again the same expression.11
The choice of the substitutability parameter β is intimately linked to the choice of the
transformation function Ij (.). In an interesting paper, Ebert and Welsch (2004) investigate
the extent to which the ordering of the well-being bundles is invariant to the choice of
the specific transformation function. Building on results from social choice theory, they
conclude that the multiplicative aggregation (β = 0) is the only aggregation form that
makes the ordering of the well-being bundles robust to the choice of a dimension-specific
ratio scale transformation (rows 5-7 in table 1). Since well-being indices typically aggregate
very different dimensions, a dimension-specific transformation is most often needed. Instead,
the more general form defined in (1) is robust only to transformations involving the same
rescaling across all dimensions, that is, attributes are being divided or multiplied by the
same factor.12 In other words, apart from some very restricted choices for β and Ij (.), the
decision on the transformation function to use typically affects the ordering of the bundles.
The lesson is not necessarily that we should restrict the transformation functions and β to
those case, but rather that this decision is to be handled with care and, preferably, guided by
either a theoretical framework or empirical observations (or both) about the true meaning
of well-being.
In sum, the framework proposed reduces the decisions to be made to three: the value
of the parameter β, the transformation functions I1 (.), . . . , Iq (.) and the weights w1 , . . . , wq .
Table 2 gives an overview of the common choices made in the literature with respect to these
decisions.13 These choices reflect alternative viewpoints on the meaning of the notion well-
being and will potentially have a non-trivial impact on the resulting ordering of bundles.14
In the next section we go deeper into the meaning of the weights. Before we will do so, we
introduce an example to illustrate the effect that the choice of the parameters in expression
(1) has on the ordering of persons in terms of their well-being.
11
Another alternative is to allow the substitutability parameter β to be a function of the achievements,
as in Bourguignon and Chakravarty (2003).
12
See Ebert and Welsch (2004) and the reference therein for more details. Note that Ebert and Welsch’s
argument implies that indices such as the Human Development Index -which uses dimension-specific linear
transformations- or the Index of Multiple Deprivation -which relies on an exponential transformation- are
not meaningful.
13
The table includes some indices used in practice, such as the HDI, and studies that provide empirical
applications, while it leaves out studies that are solely theoretical.
14
A striking example can be found in the work by Becker, et al. (1987). The authors studied the quality
of life in 329 metropolitan areas of the U.S. by ordering them according to standard variables such as quality
of climate, health, security, economical performance. The authors find that, depending on the weighting
scheme chosen, there were 134 cities that could be ranked first, and 150 cities that could be rank last.
Moreover, there were 59 cities that could be rated either first or last, using the same data, but by selecting
alternative weighting schemes. Based on this example, Diener and Suh (1997) conclude that a procedure for
resolving how to weight the dimensions is lacking.
www.ophi.org.uk 7
We compare the well-being of two persons –Ann and Bob– in two dimensions –income
and health– denoted y and h respectively, with the former being measured in dollars and
the latter in expected years of life. Ann is healthier than Bob, her life expectancy is 90 years
whereas his is only 50 years. But Bob is richer; he has an income of $2,000, whereas she
earns only $1,000. We use expression (1) to evaluate who is better off of the two. Figures
1 to 3 depict the position of Ann and Bob in the income-health space, and the iso-well-
being curves connecting all the points leading to the same level of the well-being index for
alternative definitions of the index.
We define a benchmark case as follows: the dimension-weights are set equal (wy =
wh = 21 ), transformed attributes are perfectly substitutable (β = 1), and the achievements
x
are rescaled by the median achievement (Iy = M yey for y and Ih = Mxheh for h), where
M ey = $2, 500 and M eh = 80 years. The benchmark iso-well-being curves are represented
in the graphs by the dotted lines. In figure 2 Ann’s bundle lies on a higher iso-well-being
line than Bob’s, which means that, according to this particular index, she is better off than
Bob. Let us now look at three alternative parameter choices. First, we increase the relative
weight assigned to income so that wy = 34 and wh = 41 . The corresponding iso-well-being
curves are represented in figure 1 by the solid lines. The iso-well-being curves are steeper
than the benchmark case and Bob is now considered to be better off than Ann, hence
reversing the ordering between them.
Figure 2: Iso- well-being curves. Alternative dimension-weights.
www.ophi.org.uk 8
In the second case –figure 2– we use a different transformation function. Let us assume
that the achievements of other individuals in the society have deteriorated, leading to a drop
of the median achievement to an income of $1,000 (instead of $2,500) and a life expectancy
of 60 years (instead of 80 years). In the new situation is represented by the solid line where
Bob turns out to be better off than Ann.
Figure 3: Iso- well-being curves. Alternative transformation (median).
Finally, in figure 2 we use a lower degree of substitutability between dimensions β =

0.1, in place of the previous from β = 1. The previously linear iso-well-being curves become
now convex. This means that the rate at which one can substitute income for health is
no longer constant across the well-being space. Compared to Bob, Ann -relatively rich in
health- would be willing to give up many more years of life in the future for an extra dollar
today. On this new assessment of their well-being, averages are preferred to extreme cases,
which positions Bob at a higher well-being curve than Ann.
These stylized examples illustrate that the ordering of the well-being bundles can be
very sensitive to the choice of the parameters. The lesson is that one should be careful when
deciding about the parameters. In the following section we go deeper into the meaning of
the parameters, in general, and of the weights, in particular.
3 What do weights mean?
A straightforward way to look at the meaning of the weights within the framework of the
previous section, is to study some of the properties of the well-being index in terms how
www.ophi.org.uk 9
Figure 4: Iso- well-being curves. Alternative β.
it reacts to changes in the parameters and the achievements in the different dimensions.
We do this by analyzing the partial derivatives and the corresponding marginal rate of
substitution, and by looking specifically at the role played therein by the weights. (See also
Anand and Sen (1997) for a similar approach).
The derivative of the well-being index W (.) with respect the weight of dimension j
will tell us how the well-being index reacts to small changes in wj , while keeping all other
parameters and achievements constant.
Proposition 1. For all well-being indices defined by expression (1), it holds that:
Ij (xj )β − W (x|β)β

∂W (x|β)
= . (8)
∂wj β [w1 + ... + wq ] W (x|β)β−1
Proof. The proof uses some straightforward algebraic manipulations and is extended to the
appendix, together with all other proofs of this section.
Irrespective of the selected β, an increase in the weight of dimension j leads to an

increase the well-being index if the transformed achievement in dimension j is larger than the
total well-being, in other words if the individual is performing relatively well in dimension
j. Indeed, it is intuitive that increasing the weight of a dimension on which the individual
performs well leads to an increase in overall well-being, whereas increasing the weight of
a dimension on which the individual performs relatively weak leads to a decrease in her
well-being.
www.ophi.org.uk 10
Expression (8) also provides interesting insights into the cases when the obtained well-
being index W (.) is insensitive to the exact choice of the weight wj . This happens when
Ij (xj )β = W (X)β for all j, in other words when the transformed achievements are very
alike across dimensions, or the correlation between them is sufficiently high.
Intuitively, one expects changes in dimensions with a higher weight to have more im-
pact on total well-being, than dimensions with a lower weight. In a recent paper, Chowdhury
and Squire (2006) write:
“The ideal approach would presumably involve using as weights the impact of
each component on the ultimate objective ...”. Chowdhury and Squire (2006, p.
762)
We, therefore, investigate the first derivative of the well-being index with respect the
achievement in dimension j itself. This derivative captures how the well-being index reacts
to small changes of the achievement in a given dimension, keeping all the other variables
and parameters constant.
W (x|β) 1−β

∂W (x|β) wj 0
= I (xj ) , (9)
∂xj w1 + ... + wq j Ij (xj )
∂Ij (xj )
where Ij0 = ∂xj .
The impact of a small change of the achievements of dimension j on total well-being

depends on three terms. The first term is the relative dimension-specific weight. As one
might expect, the larger the relative weight of a certain dimension, the larger the impact
of a small change in the achievement of that dimension. Secondly, the impact depends on
the derivative of the transformation curve. The larger this derivative, in other words, the
steeper the transformation curve, the larger the effect of a small increase in the achievement
on the transformed achievement and hence on total well-being. For multidimensional well-
being indices, the derivative tends to be positive, whereas it is negative for multidimensional
poverty indices. Finally, the effect depends on the ratio W (X)
Ij (xj ) to the power 1−β. For values
of β ≤ 1, if the person performs worse in that dimension than in the overall well-being, an
increase in achievement of such dimension will have a positive effect on overall well-being.
Note the the parameter β offers an instrument to increase the relative impact of a dimension
on total well-being. The lower the β the more sensitive the index is to weak performing
dimensions. A policy maker seeking to maximize the well-being W (X) will spend more
effort on the relatively weak performing dimensions if β ≤ 1, leading to a more equalized
development across dimensions.
For the simple additive well-being indices (β = 1), the term between square brackets
in expression (9) drops out, and the effect of a small change of one of the achievements only
www.ophi.org.uk 11
depends on its relative weight and the steepness of the transformation function, that is,
∂W (x|1) wj
= I 0 (xj ). (10)
∂xj w1 + ... + wq j
If, moreover, the transformation function is the identity function (Ij (xj ) = xj ) the
impact of a small change in achievement j is only determined by the weight of dimension
wj
j, i.e. w1 +...+wq
. Therefore, for this specific choice of parameters, the relative weight of a
dimension captures the impact of small change in the achievement of that dimension. The
total well-being is indeed more sensitive to changes in a dimension with larger weight.
If instead β = 0 the effect of a marginal increase in dimension j on the well-being

index is defined as
∂W (x|1) wj W (x|0)
= I 0 (xj ) , (11)
∂xj w1 + ... + wq j Ij (xj )
where the effect depends not only on the weight but, crucially, on the level at which the
change has taken place. This seems to be a reasonable feature. Using the same example as
in the previous section, an extra year of life for a young person might be valued differently
to an additional year given to the elder.
An alternative but related meaning of the weights is as substitution rates between two
dimensions –y and h– denoted M RSyh .15 Let us reconsider the previous example of Anne
and Bob where dimension h represents health and dimension y is income. The marginal
rate of substitution between these dimensions is the amount of health an individual would
like to gain if she were to sacrifice one unit of income, while maintaining the same level of
well-being. In graphical terms, the M RSyh reflects the slope of the iso-well-being curves
and is formally defined as:
dxh ∂W (x|β) . ∂W (x|β)

M RSyh = − =− . (12)
dxy ∂xy ∂xh
By substituting expression (9) into (12) we obtain the following expression:

wy Iy0 (xy ) Ih (xh ) 1−β

M RSyh =− . (13)
wh Ih0 (xh ) Iy (xy )
15
See, for instance, Munda and Nardo, (2005).
www.ophi.org.uk 12
The marginal rate of substitution between dimension income and health also consists
of three parts. We will relate each of these components to the cases illustrated in the
previous section in figures 1 to 3. The first component is the ratio of the dimension-specific
weights wy /wh . The larger the weight assigned to income the more the years of life (health)
that the person needs to gain to compensate for the loss of one dollar of income. Going back
to figure 2, the new income-weight wy is increased, leading to a larger ratio wy /wh , a larger
M RSyh and a steeper iso-well-being curve. The second part of expression (13) is the ratio
of the derivatives of the transformation functions of dimension h and y. The steeper the
transformation function of income, xy –or equally, the flatter the transformation function of
health, xh – the larger the amount of dimension h necessary to compensate for the loss in xy .
In figure 2, the deteriorated medians of the society lead to a larger ratio Iy0 (xy )/Ih0 (xh ),and
hence to a steeper iso-well-being curve.16 Finally, the marginal rate of substitution depends
on the ratio of the transformed achievements to the power 1 − β. For β < 1, the amount
of dimension h needed to compensate for the loss in dimension y is greater, the smaller the
original achievement in dimension y. This makes sense; achievements are more valuable as
they become more scarce. In figure 2 using the alternative iso-well-being curves, the poorer
the person the steeper the iso-well-being curve becomes. Anne should be given more health
than Brian to compensate for a unit decrease in income.
In the linear case (β = 1), the trade-off is assumed constant at all levels of achieve-
ments.
wy Iy0 (xy )
M RSyh = − . (14)
wh Ih0 (xh )
If, in addition, the ratio of the derivatives of the transformation functions is unity, the
marginal rate of substitution between two dimensions is uniquely defined by the ratio of
their weights.
wy
M RSyh = . (15)
wh
In the geometric case, instead, the trade-off between dimensions depends on the levels
where the exchange is taking place. Specifically,
wy Iy0 (xy ) Ih (xh )

M RSyh = − . (16)
wh Ih0 (xh ) Iy (xy )
When the transformation is the identity, the marginal rate of substitution between
health and income results
wy x h
M RSyh = − . (17)
wh x y
0
16 Iy (xy ) 1/M ey M eh
To be precise, the ratio raise from 0 (x )
Ih
= 1/M eh
= M ey
= 80/2500 = 0.032 to 60/1000 = 0.06.
h
www.ophi.org.uk 13
In short, the analysis of some properties of the general class of the well-being index
proposed in expression (1), gives us three lessons. First, that the higher the weight the
better the person will perform if she fares better in that dimension than in the others.
Second, in close interplay with the other parameters, the dimension-weights determines the
contribution of that dimension to total well-being. Finally, dimension-weights form part of
the trade-off between attributes and can be interpreted directly as the trade-offs only under
certain assumptions –perfect substitutability between dimensions and no transformation of
the original variables. In the next section we survey some procedures to set the weights
from this perspective.
4 How can we set weights in a reasonable way?
In the previous section we concluded that whether the weights are set reasonably or not,
can and should be evaluated based on the trade-offs they imply between the dimensions
of well-being. In this section we survey from this perspective some of the most commonly
used methods to set the weights in practice.17
4.1 Equal weights
The most commonly used approach to weighting in multidimensional indices of well-being

has been equal weighting.18 Despite its popularity, equal weighting is far from uncontro-
versial. Chowdhury and Squire refer to equal weighting as “obviously convenient but also
universally considered to be wrong.” (Chowdhury & Squire 2006, p. 762).
Equal weighting has often been defended from an agnostic viewpoint, by its simplicity
or indeed from the recognition that all indicators are equally important.19 As an example
of the agnostic viewpoint, Mayer and Jencks defend equal weighting by remarking that:
“ideally we would have liked to weight ten hardships according to their relative importance
in the eyes of legislators and the general public, but we have no reliable basis for doing this”
(Mayer & Jencks 1989, p. 96).
17
An alternative method, not reviewed here, would be to use market or personalized prices as weights,
so that the well-being index (with identity transformations and β = 1) coincides with the individual’s
expenditures. Srinivasan (1994) advocates such an approach. Smeeding et al. (1993) present a related
approach, where non-monetary dimensions are given valued (for instance, education and health services are
imputed global values based on the amount the government spends on them) and added to the current
expenditure. On the other hand, as stated by Foster and Sen (1997), prices do not exist for many relevant
dimensions of well-being and are in general inappropriate for well-being comparisons, a task for which they
are not constructed.
18
Examples are the Human Development Index, the Human Poverty Indices, the Commitment to Devel-
opment Index (Roodman 2007), the English Index of Local Conditions (Department of Environment, 1994),
and the Townsend Material Deprivation Score (Townsend, Phillimore & Beattie 1988), among others.
19
Strictly speaking, equal weighting assigns zero weight to all dimensions of well-being ignored by the
index.
www.ophi.org.uk 14
However, there is a fallacy in setting the weights equally motivated from an agnostic
viewpoint. As has been shown in the previous section, there is no escape from the fact that
the weights reflect an important aspect of the trade-offs between the dimensions. As any
other weighting scheme, the equal weighting scheme implies in interplay with choices about
the transformation and substitutability specific trade-offs between the dimensions, that can
and should be made explicit, and might be considered reasonable or not. In a paper on
the HDI, Ravallion (1997) looks at the implied marginal rates of substitution in the HDI
and finds that: “The HDIs implicit monetary valuation of an extra year of life rises from a
remarkably low level in poor countries to a very high level in rich ones. In terms of both
absolute dollar values and the rate of GDP growth needed to make up for lower longevity,
the construction of the HDI assumes that life is far less valuable in poor countries than in
rich ones; indeed, it would be nearly impossible for a rich country to make up for even one
year less of life on average through economic growth, but relatively easy for a poor country”
He concludes: “The value judgements underlying these trade-offs built into the HDI are not
made explicit, and they are questionable.” (Ravallion 1997, p. 633).20
In sum, researchers that would like to avoid the hazardous question of how to set
the weights, and therefore choose for equal weighting, should be aware that the equal
weighting scheme is actually a weighting scheme as any other without specific normative
attractiveness, and just as any other weighting scheme it implies trade-offs that might be
reasonable or not.21
4.2 Data-driven weighting schemes
Many methods to obtain dimension weights rely in some way or another on the data at
hand. We compare four approaches and will criticize them on similar grounds.
4.2.1 Frequency-based weights
One method to determine the weights is to set them in terms of the proportion of the pop-
ulation suffering deprivation in that dimension. Two different approaches can be found in
the literature, taking quite opposite perspectives. First, in the context of multidimensional
deprivation measurement, Desai and Shah (1988) and Cerioli and Zani (1990) argue that the
smaller the proportion of individuals with a certain deprivation, the higher should be the
weight, on the grounds that a hardship shared by few has more impact than one shared by
20
Decanq et al. (2007) make a similar point, based on the most recent calculation method of the Human
Development Index.
21
One could use the transformation functions to obtain reasonable trade-offs while keeping the weights
equal across dimensions. But this seems rather to turn the problem on its head and might even defeat the
purpose of simplifying the index if the resulting transformation functions were less than simple.
www.ophi.org.uk 15
many.22 On the other hand, in their work on well-being indices, Osberg and Sharpe (2002)
make use of frequency-based weights to aggregate subcomponents of the economic security
component whereby smaller weight is given to dimensions with a smaller proportion of the
population at risk.23
A related way of setting the weights is based on the quality of the data. Jacobs et al.
(2004) suggest to give less weight to those variables where data problems exist or with large
amounts of missing values. The advantage is that the reliability of the well-being index can
be improved by giving more weight to good quality data.
Apart from the apparent disagreement how the weights should depend on the relative
proportions, the fundamental question seems to be why the weights and the implied trade-
offs should depend on the relative proportion achieved by the population or on the data
quality.
4.2.2 Most favorable weights
When applying the same weighting scheme to all individuals, some of them might feel that
the evaluation of their well-being is submitted to someone else’s perspective on what well-
being exactly is.24 Therefore, a researcher might want to give all individuals the “benefit
of the doubt” and select for each individual the most favorable weighting scheme. This
method has originally be proposed for evaluating macro-economic performance (Melyn
& Moesen 1991) and has recently been used in the construction of composite indicators
(Despotis 2005, Mahlberg & Obersteiner 2001). The weights are individual-specific and
endogenously determined such that they maximize the obtained well-being of the individ-
ual.25 The highest relative weights are given to those dimensions on which the individual
performs best. To avoid that all weight is given to one dimension (the best dimension of
the individual), extra constraints can be imposed upon the weights assuring that minimal
weight is given to each dimension of well-being.
Drawbacks of this approach can be the following: First, since every individual has
22
In a recent paper, Brandolini (2007) points out that when applying Desai and Shah’s weighting formula
to Italian data, he comes to a rather questionable and unbalanced weighting scheme.
23
The index of Economic Well-being proposed by Osberg and Sharpe aggregate four components of well-
being (consumption, accumulation, income distribution, and economic security) using equal weights but
frequency weights to aggregate indicators within the last component.
24
In recent social choice literature, the question of whose preferences or ideas about trade-offs should
matter has been taken up under the label of the “indexing dilemma” –see Fleurbaey (2007). In his elegant
paper, Fleurbaey investigates the apparent impossibility of finding a weighting scheme that is individual
specific and, at the same time, is not susceptible to the critiques formulated against welfarism, as expressed
by Sen (1985). He argues in favor of a way out the impossibility based on an approach that takes information
on the individual iso-well-being curves into account.
25
In case β = 1, this problem reduces to linear programming problem, see Cherchye et al. (2006) for
technical details. The authors provide an overview composite indicators that set the weights based on this
most favorable weighting scheme, which is an application of so-called Data Envelopment Analysis. Mahlberg
and Obersteiner (2001) proposes a specific application to the Human Development Index, and Ramos and
Silber (2005) compare the approach to alternative ones. See also Despotis (2005).
www.ophi.org.uk 16
her own weighting scheme, the comparison of well-being levels across individuals is not
straightforward.26 Second, the obtained results depend highly on the exact formulation
of the technical constraints chosen by the analyst, making it a less transparent procedure.
Finally, and most importantly, there is no guarantee that the most favorable weights lead
to reasonable trade-offs between the dimensions. There seems to be no a priori reason, why
a certain dimension on which the individual performs relatively well should have a larger
impact on total well-being, because the individual performs well on that dimension.
4.2.3 Statistical weights
There are two sets of techniques that are employed to choose weights for multidimensional
indices: descriptive and explanatory models.27
The first approach relies on multivariate statistical methods to summarize the data.
The most commonly used techniques are based on principal components (Klasen 2000,
Noorbaksh 1998) and cluster analysis (Hirschberg, Maasoumi & Slottje 1991). The use of
these statistical techniques is motivated by a concern for the so-called problem of double
counting. In many empirical applications the dimensions of well-being are found to be
strongly correlated.28 Loosely speaking, most multivariate statistical techniques adjust for
the correlation between indicators by either choosing the dimensions that are not correlated
or by adjusting the weights so that correlated dimensions get less weight (Nardo, Saisana,
Saltelli & Taranto 2005). For instance, in principal component analysis, a given set of di-
mensions is transformed into an equal number of mutually uncorrelated linear combinations
of dimensions. One can compute the proportion of the variance explained by each linear
combination. If a small group of those linear combinations can explain a large proportion of
the variance, then the information contained by the initial dimensions is largely contained
in the small group of combinations that are, by definition, uncorrelated and which solves the
double counting problem. The two most commonly used methods to obtain weights from
the linear combinations, is to use either the principal component that explains the largest
proportion of the variance, or to use a weighted average of all the linear combinations.
The second approach, sometimes known as latent variable models is an explanatory ap-
proach that assumes that some observed variables (dimensions) are dependent on a certain
26
Despotis (2005) propose a way of using the individual most favourable weights to construct a common
weighting structure. This is done minimizing the distance between the individual (country)-specific weight
and the global weights.
27
For a detailed overview of the statistical properties of some methods to set the weights based on multi-
variate statistics, we refer the reader to Kirshnakumar and Nadar (2008).
28
For instance, Srinivasan (1994) reports a correlation coefficient of about 0.8 between the dimensions of
the Human Development Index. Whether double counting is really a problem, is open for discussion. One
could argue that the correlation between the dimensions in a society reflects an important aspect of the real
situation and as such it should be included, not eliminated from the analysis. The pluralistic egalitarian
notion of Walzer (1981), for instance, considers that the correlation between the dimensions is one of the
most essential characteristics of the society. From that perspective, correcting for correlation between the
dimension might be completely inappropriate.
www.ophi.org.uk 17
number of unobserved latent variables (Krishnakumar & Nadar 2008). Factor analysis is
possibly the simplest case of a latent variable model, imposing that the observed dimensions
are in fact different manifestations of the latent component, or ‘factor’. In the context of
well-being and deprivation indices, factor analysis has been widely employed (Maasoumi &
Nickelsburg 1988, Schokkaert & Van Ootegem 1990, Nolan & Whelan 1996, Noble, McLen-
nan, Wilkinson, Whitworth, Barnes & Dibben 2008). More advanced latent models include
other exogenous variables that also might influence the latent variable but are not part of
the selected set of dimensions used to construct the index. In this line, Multiple Indicator
and Multiple Causes Model (MIMIC) and structural equation models (SEM) have been
proposed to construct multidimensional indices, particularly among those supporting the
capability approach (Di Tommaso 2006, Kuklys 2005, Krishnakumar 2007, Krishnakumar
& Ballon 2007).
There are, however, some drawbacks to these multivariate statistical approaches. First,
the obtained linear combinations of dimensions might be hard to interpret as a facet of hu-
man well-being (Srinivasan 1994). Additionally, the derivation of weights through principal
component or latent variables models is, by no means, straightforward and hence it lacks
transparency, which makes these technique less attractive as a method to informing lo-
cal and international policy makers(de Kruijk & Rutten 2007). Most crucially, statistical
approaches can lead to normatively inappropriate results. For instance, in the construc-
tion of the Environmental Sustainability Index, the principal component method was found
to assign negative weights to some sub-indicators (World Economic Forum, 2002). We
obtain the same result when we apply this method to the example of Ann and Bob.29
Moreover, there is a priori no reason to believe that statistical weights are in accord with
people’s perceptions about priorities and relative importance of each dimension (de Kruijk
& Rutten 2007). Along that line, Brandolini (2007) warns the reader that “we should be
cautious in entrusting a mathematical algorithm with a fundamentally normative task”.
Multivariate statistical techniques, especially principal component analysis, are developed
to summarize the data in a statistically reasonable and parsimonious way. As such, they
can be useful to aggregate indicators within dimensions. But this is quite a different task
to setting weights that are normatively reasonable.
4.2.4 Regression based weights
The final approach to set the weights based on data is to estimate the coefficients αj =
α1 , . . . , αq of the following equation:
Yi = α1 I1 (x1i ) + ... + αq Iq (xqi ) + εi , (18)

29
The weights wh , wy obtained by principal component of the transformed data (z-scores) are respectively
-0.7071 and 0.7071.
www.ophi.org.uk 18
where Yi is some output variable capturing the well-being of individual i. This expression
shows great similarity with the linear well-being index as defined in expression (2), with
the role played by the coefficients αj corresponding to the weights wj . The main problem
to operationalize this approach to find a reasonable Yi for every individual, approximating
her well-being.
In a recent paper, Schokkaert (2007) proposes to rely on the emerging measures of

life satisfaction as proxy for individual well-being. He writes “On the one hand, the robust
statistical relationship between functionings and life satisfaction may provide useful infor-
mation on the relative weights to be given to the various dimensions in the calculation of
individual living standards. On the other hand, from a non-welfarist point of view we do
not want idiosyncratic individual factors to wipe out the effects of conditions of material
deprivation, linked, for example, to unemployment or job satisfaction” (Schokkaert 2007,
p.423). Schokkaert proposes an approach in which individual life satisfaction is used as
left-hand-side variable in expression (18), and where all idiosyncratic individual factors are
set at their mean value for the population to take account of the second point. His ap-
proach can be made more flexible by allowing for non-linearities in expression (18) or for
the coefficients αj to vary across different groups in the population.
In general the regression based weights have the drawback that if the well-being could
be measured in an appropriate way by the single variable Yi , there would be not need
to construct a well-being index in the first place. A less critical concern is that the the
coefficients of αj might suffer from the problem of multicollinearity were the dimensions of
well-being be strongly correlated.
Data-driven approaches offer an interesting way to obtain weights for multidimensional

well-being indices. There is an expanding literature proposing their use and perfecting the
methods so that they are more than just data summarizing techniques. Two points of cau-
tion are in order. First, these methods generally assume a linear form β = 1 which implies
perfect substitutability between all pair of dimensions at all levels. Some of the techniques
could overcome this problem straightforwardly. Second, an inconvenient property of the
weights obtained by statistical techniques is that are sensitive to adding new observations
to the data-set (Nardo, Saisana, Saltelli, Tarantola, Hoffman & Giovannini 2005). In other
words, given that weights are data-specific they can change from one point in time to the
next, and from country to country, which makes any meaningful comparison of situations
problematic.
4.3 Normative weights
From the previous section on the meaning of the weights we recall that the weights are crucial
in determining the trade-offs between the dimensions of well-being. A third approach is to
www.ophi.org.uk 19
obtain more normatively inspired weights. Unfortunately, there are very few guidelines
in the ethical or philosophical literature on how the obtain reasonable trade-offs between
dimensions of well-being. Fleurbaey (2008) states: “One can of course invoke the ethical
preferences of the observer and ask her, for instance, how she trades the suicide rate off
against the literacy rate, but there is little philosophical or economic theory that gives us
clues about how to form such preferences.” (Fleurbaey 2008, p. 21).
One approach would be to ask all individuals in the society how they personally would
trade-off the different dimensions, and then aggregate these opinions somehow. In practice,
however, asking all individuals in a society might not be feasible, therefore one often relies
on the preferences of a limited group of people that are thought to represent, to some extent,
the rest of the society.30 Generally, four sets of groups are considered: a random sample
of the population, ‘experts’ from the academic and international organization communities,
and policy makers -usually deciding where and how to spend resources- and groups of
representatives of different sections of society.
In the literature, there exist some methods to elicit the preferred trade-offs between
the dimensions of the (representative group of) individuals. A first method is to survey
directly how the individuals would trade off different dimensions of well-being. Similar
approaches have been used in health economics to obtain an estimation how much health
gain one is prepared to sacrifice for a reduction in health inequality (see for instance Shah
et al. 2001 and Jacobs et al. 2004). An example within the well-being index literature
is give by de Kruijk and Rutten (2007). The authors uses the Maldivean household sur-
vey –Poverty and Vulnerability Assessment 1997/98 and 2005– where (randomly sampled)
respondents are asked to rank living standard dimensions according to their relative impor-
tance in determining the overall standard of living and deprivation. The dimension weight
for each individual is then computed as a function of the number of dimensions considered
and the specific ranking of that dimension.31 Finally, the paper uses the average weight
to compute the individual specific human vulnerability index, so that every respondent is
assigned the mean priority weight of the country. In an interesting paper, Chowdury and
Squire (2006) use electronic surveys to elicit weighting schemes to assess whether the equal
weighting scheme of the Human Development Index had support from the ‘expert com-
munity’, understood as development researchers throughout the world placed in academic
institutions. Each person was asked to weight each component of the HDI from 0 to 10 in
30
However, public opinion polls have been used in problems of eliciting the public concerns about en-
vironmental issues. In that way the concern the public opinion attaches to the different environmental
subindicators is determined. Parker (1991 p.95-98) advocates such an approach: “public opinion polls have
been extensively employed for many years for many purposes, including the setting of weights”.
31
Specifically, the weight for dimension j is determined by
1 + nd − rj
wj = Pnd ,
j=1 1 + nd − rj
where nd is the number of dimensions and rj is the ranking of dimension j with value 1 if it is the most
important, 2 if it is second most important dimension, and so on.
www.ophi.org.uk 20
order of importance, and the average of these weighting schemes was considered. Interest-
ingly enough, they find that the average weighting scheme does not statistically differ from
the present equal weighting scheme. Survey methods are in practice, a voting mechanism.
When the question is simple enough and the options are limited, standard voting procedures
can also be applied. Survey methods have been used to collect preferences of representative
individuals of the society and so called-experts, and less frequently, policy-makers.32
A second and related method is to use budget allocation. The members of the represen-
tative group are asked to distribute a budget of points to a number of dimensions, paying
more for those dimensions whose importance they want to stress. Moldan and Billharz
(1997) report a case study in which 400 German experts were asked to allocate a budget to
a set of environmental indicators related to air pollution, leading to very consistent results,
where experts came form very different social backgrounds. Budget allocation mechanism
belongs to a more general class of participatory planning.
A third method is the analytic hierarchy process. This has been proposed by Saaty
(1987) originating from multi-attribute decision making. In this procedure, all members
of the representative group are asked to compare pairwise the dimensions by asking the
question: “Which of the two is the more important? - and by how much?”. The strength
of the preference is expressed on a semantic scale of 1-9. These comparisons result in a
comparison matrix from which the relative weights can be calculated using an eigenvector
technique (see Nardo et al (2005) and the references therein for a detailed treatment).
The main source of concern with participatory methods relates to the selection of par-
ticipants, a concern that holds true for any of the above mentioned sets of groups (experts,
representative individuals, and policy-makers), and any technique employed. Selection of
participants can be biased -some groups being under-represented- or simply uninformed,
and hence the resulting weighting scheme will be skewed. A second problem is that, even
when the selection is bias-free, the participatory technique may lead to unrepresentative
preferences were the process be subject of pressures from power groups and vested inter-
ests. A final critique is that participatory approaches in general can lead to some sort of
paternalism. Although to a certain degree, the extent of paternalism can be handled with
appropriate selection of participants and elicitation mechanism, normative participatory
approaches rely on imposing some people’s ideas of how important each dimension is to
other individuals. Hence, the critique of paternalism prevails.
Although normative approaches have the disadvantages mentioned above, they are in
32
Another example of elicitation of expert preferences is given by Carlucci and Pissani (1995). The
authors make use of the Delphi methodology, where participants (‘experts’) are asked to give values to
complete predefined lotteries. In this way, they are able to elicit not only dimension weights but also some
measures of interdependence between the dimensions, in the eyes of the participants. This method seems
particularly applicable when the participants are literate enough to understand the questions possed and
give a meaningful cardinal answer.
www.ophi.org.uk 21
nature closer related to the meaning of the weights as trade-offs, and as such they can be
expected to lead to more reasonable results.
After having surveyed these six methods, a final remark is in place. Researchers might
find it difficult to pinpoint a unique weighting scheme, whereas they might find it easier
to obtain ”ranges” in which reasonable values of weights can be found. Foster and Sen
(1997, p. 206) state that while ”the possibility of arriving at a unique set of weights is
rather unlikely, that uniqueness is not really necessary to make agreed judgements in many
situations”. Such an approach of working with ranges of weights, rather than exact values
has the advantage of allowing for some degree of agnosticism. However, that agnosticism
comes at a price: an approach based on ranges of weights, is likely to lead to a partial
ordering of the well-being bundles. How incomplete the ordering becomes, or how many
bundles will become incomparable, depends on the allowed width of the ranges and the
correlation between the achievements of the individuals across the dimensions. The stronger
the correlation between the dimensions, the less important the exact specification of the
weights. A sensitivity analysis for alternative weighting schemes can be very helpful in
determining how robust the well-being index and the implied ordering of the bundles is for
alternative weighting schemes.33 Although it is clear that a sensitivity analysis can never
answer the question on how to set weights in a reasonable way, it might give an idea how
important the answer is for the obtained results and how much room there is for agnosticism,
concerning the weights.
5 Conclusion
In this paper, we surveyed different approaches for setting weights in multidimensional

indices. We provided a general framework where most methods fit in. This framework
allowed us to understand the meaning of weights as crucial factors determining the trade-off
between dimensions. Dimension-weights are, however, not the only component determining
this trade-off. The form of the transformation of the original variables into commensurable
units and the parameter of substitution between dimensions also play an important role.
However, these components are more often than not ignored in the literature.
We reviewed six approaches used to set dimension weights, highlighting their advan-
tages and drawbacks. Ultimately, the definite test for any weighting scheme should be in
terms of its reasonability in terms of implied trade-offs between the dimensions. As long as
33
For example, in the context of the measurement of multidimensional global welfare, Decancq and Ooghe
(2008) propose a normative framework in which they carry out a sensitivity analysis for all possible weighting
schemes. They find that the obtained trend in increasing welfare is robust for almost all weighting schemes,
except for the one giving almost all weight to life-expectancy. Foster et al (2008) propose a way to easily
test the robustness of weights. They apply a rank-robustness technique to assess Human Development Index
weights.
www.ophi.org.uk 22
there is no widely accepted theoretical framework how to set these trade-offs, the researcher
has no choice than to rely on her common sense and to be very cautious in interpreting
the obtained orderings of the well-being bundles. In all cases, robustness tests to determine
whether results are driven solely by the specific value of weights selected, should be called
upon.
In terms of the specific approaches surveyed we conclude the following. First, equal
weights are in no sense neutral and can be questioned on ethical grounds and do not make
explicit the underlying assumptions. Second, data-driven methods are useful when aggre-
gating indicators within a given dimension as by definition, they help reducing the dimen-
sionality of the data-set at hand. Finally, normative-based methods seem more appropriate
to aggregate across dimensions. Within those, participatory approaches are a promising
route but still too many things left subject to the democracy and efficiency of the process
-more work on this is needed.
www.ophi.org.uk 23
References
Anand, S. & Sen, A. K. (1997). Concepts of human development and poverty: A mul-
tidimensional perspective. Human development papers, Human Development Report
Office (UNDP), New York.
Atkinson, A. B. & Bourguignon, F. (1982). The comparison of multi-dimensioned distribu-

tions of economic status, Review of Economic Studies 49(2): 183–201.
Atkinson, A. B., Marlier, E. & Nolan, B. (2004). Indicators and targets for social inclusion
in teh european union, Journal of Common Markets Studies 42(1): 47–75.
Becker, R. A., Denby, L., McGill, R. & Wilks, A. R. (1987). Analysis of data from the
places rated almanac, The American Statistician 41(3): 169–186.
Blackorby, C. & Donaldson, D. (1982). Ratio-scale and translation-scale full interpersonal

comparability without domain restrictions: Admissible social-evaluation functions, In-
ternational Economic Review 23(2): 249–68.
Bourguignon, F. (1999). Comment on ‘Multidimensioned approaches to welfare analysis’ by

E. Maasoumi, in J. Silber (ed.), Handbook of income inequality measurement, Kluwer
Academic, London, pp. 477–84.
Bourguignon, F. & Chakravarty, S. R. (2003). The measurement of multidimensional

poverty, Journal of Economic Inequality 1(1): 25–49.
Brandolini, A. (2007). On synthetic indices of multidimensional well-being: health and

income inequalities in France, Germany, Italy, and the United Kingdom. Bank of Italy,
Economic Research Department.
Carlucci, F. & Pissani, S. (1995). A multiattribute measure of human development, Social

Indicators Research 36: 145–176.
Cerioli, A. & Zani, S. (1990). A fuzzy approach to the measurement of poverty, in C. Dagum
& M. Zenga (eds), Income and Wealth Distribution, Inequality and Poverty, Springer-
Verlag, Berlin.
Cherchye, L., Moesen, W., Rogge, N., Van Puyenbroeck, T., Saisana, M., Saltelli, Liska,
R. & Tarantola, S. (2006). Creating composite indicators with DEA and robustness
analysis: the case of the technology achievement index. Katholieke Universiteit Leuven,
Public Economics Working Paper 0613.
Chowdhury, S. & Squire, L. (2006). Setting weights for aggregate indices: An application
to the commitment to development index and human development index, Journal of
Development Studies 42(5): 761–771.
www.ophi.org.uk 24
Crawford, I. (1999). Nonparametric tests of stochastic dominance in bivariate distributions,

with an application to UK data. IFS Working Papers W99/28, Institute for Fiscal
Studies.
de Kruijk, H. & Rutten, M. (2007). Weighting dimensions of poverty based on peoples pri-
orities: Constructing a composite poverty index for the Maldives. Q-Squared Working
Paper No. 35, Centre For International Studies, University Of Toronto.
Decancq, K., Decoster, A. & Schokkaert, E. (2007). The evolution in world inequality in
well-being. Discussion Paper Series 07/04, Center for Economic Studies, Katholieke
Universiteit Leuven.
Decancq, K. & Lugo, M. A. (2008). Measuring inequality of well-being: A proposal based

on a multidimensional Gini index. Mimeo.
Decancq, K. & Ooghe, E. (2008). Did the world move forward? Mimeo.
Desai, M. & Shah, A. (1988). An econometric approach to the measurement of poverty,

Oxford Economic Papers 40: 505–522.
Despotis, D. K. (2005). A reassessment of the Human Development Index via data envel-
opment analysis, Journal of the Operational Research Society 56(8): 969–980.
Deutsch, J. & Silber, J. (2005). Measuring multidimensional poverty: An empirical com-

parison of various approaches, Review of Income and Wealth 51(1): 145–74.
Di Tommaso, M. L. (2006). Measuring the well being of children using a capability approach:
An application to Indian data. CHILD - Centre for Household, Income, Labour and
Demographic economics, CHILD Working Paper 05.
Diener, E. & Suh, E. (1997). Measuring quality of life: Economic, social and subjective
indicators, Social Indicators Research 40: 189–216.
Duclos, J.-Y., Sahn, D. E. & Younger, S. D. (2006). Robust multidimensional poverty

comparisons, The Economic Journal 116(514): 943–968.
Ebert, U. & Welsch, H. (2004). Meaningful environmental indices: A social choice approach,
Journal of Environmental Economics and Management 47(2): 270–283.
Fleurbaey, M. (2007). Social choice and the indexing dilemma, Social Choice and Welfare
29: 633–648.
Fleurbaey, M. (2008). Individual well-being and social welfare: Notes on the theory. Mimeo.
Foster, J., Lopez-Calva, L. F. & Szekely, M. (2005). Measuring the distribution of human
development: methodology and an application to Mexico, Journal of Human Develop-
ment 6(1): 5–25.
www.ophi.org.uk 25
Foster, J., McGillivray, M. & Seth, S. (2008). Rank robustness of multidimensional well-
being measures. Mimeo.
Foster, J. & Sen, A. K. (1997). On Economic Inequality, Oxford University Press, Oxford.
Gravel, N., Moyes, P. & Tarroux, B. (2005). International multidimensional comparisons

of inequality in disposable income and access to public goods. IDEP Working Paper
05-12, Institut d’economie publique, Marseille.
Hirschberg, J. G., Maasoumi, E. & Slottje, D. (1991). Cluster analysis for measuring welfare
and quality of life across countries, Journal of Econometrics 50(1-2): 131–150.
Jacobs, R., Smith, P. & Goddard, M. (2004). Measuring performance: An examination of

composite performance indicators. CHE Technical Paper Series 29, Centre for Health
Economics, University of York.
Justino, P. (2005). Empirical applications of multidimensional inequality analysis. Poverty

Research Unit at Sussex, Department of Economics, University of Sussex.
Kahneman, D. & Krueger, A. B. (2006). Developments in the measurements of subjective

well-being, Journal of Economic Perspectives 20(1): 3–24.
Klasen, S. (2000). Measuring poverty and deprivation in south africa, Review of Income
and Wealth 46: 33–58.
Krishnakumar, J. (2007). Going beyond functionings to capabilities : an econometric model

to explain and estimate capabilities, Journal of Human Development 7(1): 39–63.
Krishnakumar, J. & Ballon, P. (2007). Estimating basic capabilities: A structural equation

model approach applied to Bolivian data, World Development forthcoming.
Krishnakumar, J. & Nadar, A. (2008). On exact statistical properties of multidimensional

indices based on principal components, factor analysis, MIMIC and structural equation
models, Social Indicators Research 86(3): 481–496.
Kuklys, W. (2005). Amartya Sens capability approach: Theoretical insights and empirical
applications, Springer, Berlin.
Lugo, M. A. (2007). Comparing multidimensional indices of inequality: Methods and ap-

plication, Research on Economic Inequality 14.
Maasoumi, E. (1986). The measurement and decomposition of multi-dimensional inequality,

Econometrica 54(4): 991–97.
Maasoumi, E. (1999). Multidimensioned approaches to welfare analysis, in J. Silber (ed.),

Handbook of income inequality measurement, Kluwer Academic, Boston, pp. 437–77.
www.ophi.org.uk 26
Maasoumi, E. & Lugo, M. A. (2008). The information basis of multivariate poverty assess-
ments, in N. Kakwani & J. Silber (eds), Quantitative Approaches to Multidimensional
Poverty Measurement, Palgrave-MacMillan.
Maasoumi, E. & Nickelsburg, G. (1988). Multidimensional measures of well-being and

an analysis of inequality in the Michigan data, Journal of Business and Economic
Statistics 6(3): 327–334.
Mahlberg, B. & Obersteiner, M. (2001). Remeasuring the HDI by data envelopment anal-
ysis. Interim report IR-01-069, International Institute for Applied System Analysis.
Mayer, S. E. & Jencks, C. (1989). Poverty and the distribution of material hardship, Journal
of Human Resources 24(1): 88–114.
Melyn, W. & Moesen, W. (1991). Towards a synthetic indicator of macroeconomic perfor-

mance: unequal weighting when limited information is available. Katholieke Univer-
siteit Leuven, Public Economic Working Paper 17.
Moldan, B. & Billharz, S. (1997). Indicators of Sustainable Development, John Wiley,

Chichester.
Muller, C. & Trannoy, A. (2004). A dominance approach to well-being inequality across

countries. Working Paper WP-AD 2004-27, University of Alicante.
Munda, G. & Nardo, M. (2005). Constructing consistent composite indicators: the issue of
weights. EUR 21834EN, Institute for the protection and security of the citizen.
Narayan, D. (2000). Voices of the Poor: Volume 1: Can Anyone Hear Us?, World Bank
Publication, DC.
Nardo, M., Saisana, M., Saltelli, A. & Taranto, S. (2005). Tools for composite indicators
building. EUR 21682 EN, European Communities.
Nardo, M., Saisana, M., Saltelli, A., Tarantola, S., Hoffman, A. & Giovannini, E. (2005).
Handbook on constructing composite indicators: Methodology and user guide. OECD
Statistics Directorate, OECD Statistics Working Papers 2005/3.
Nilsson, T. (2007). Measuring changes in multidimensional inequality. an empirical appli-

cation. Mimeo.
Noble, M., McLennan, D., Wilkinson, K., Whitworth, A., Barnes, H. & Dibben, C. (2008).
The English Indices of Deprivation 2007. Communities and Local Government, London.
Noble, M., Wright, G., Smith, G. & Dibben, C. (2006). Measuring multiple deprivation at
the small-area level, Environment and Planning 38(1): 169–185.
Nolan, B. & Whelan, C. T. (1996). The relationship between income and deprivation: A
dynamic perspective, Revue conomique 3: 709–717.
www.ophi.org.uk 27
Noorbaksh, F. (1998). The Human Development Index: some technical issues and alterna-
tive indices, Journal of International Development 10: 589–605.
Nussbaum, M. (2000). Women and human development, Cambridge University Press, Cam-
bridge.
Osberg, L. & Sharpe, A. (2002). An index of economic well-being for selected oecd countries,
Review of Income and Wealth 48(3): 291–316.
Ramos, X. & Silber, J. (2005). On the application of efficiency analysis to the study of the
dimensions of human development, Review of Income and Wealth 51(2): 285–309.
Ravallion, M. (1997). Good and bad growth: The human development reports, World
Development 25(5): 631–638.
Rawls, J. (1971). A Theory of justice, Belknap Press of Harvard University Press, Cam-
bridge, Mass.
Roodman, D. (2007). The Commitment to Development Index: 2007 Edition. Center for
Global Development.
Saaty, R. W. (1987). The analytic hierarchy process- what it is and how it is used, Mathe-
matical Modelling 9: 161–176.
Schokkaert, E. (2007). Capabilities and satisfaction with life, Journal of Human Develop-
ment 8(5): 415–430.
Schokkaert, E., Fleurbaey, M. & Decancq, K. (2008). What good is happiness? OPHI
Working Paper, 20.
Schokkaert, E. & Van Ootegem, L. (1990). Sen’s concept of the living standard applied to
the Belgian unemployed, Recherches Economiques de Louvain 56: 429–450.
Sen, A. K. (1985). Commodities and capabilities, North-Holland, Amsterdam and Oxford.
Smeeding, T., Saunders, P., Coder, J., Jenkins, S., Fritzell, J., Hagenaars, A., Hauser, R. &
Wolfson, M. (1993). Poverty, inequality and family living standards impact across seven
nations: the effect of non-cash subsidies for health, education and housing, Review of
Income and Wealth 39(3): 229–256.
Srinivasan, T. N. (1994). Human development: A new paradigm or reinvention of the

wheel?, American Economic Review 84(2): 238–43.
Townsend, P. (1979). Poverty in the United Kingdom, Allen Lane and Penguin, London.
Townsend, P., Phillimore, P. & Beattie, A. (1988). Health and Deprivation: Inequality and
the North, Routledge, London.
www.ophi.org.uk 28
United Nations Development Programme, (2005). Human development report 2005. Oxford
University Press, New York.
Weymark, J. A. (2006). The normative approach to the measurement of multidimensional

inequality, in F. Farina & E. Savaglio (eds), Inequality and economic integration, Rout-
ledge, London.
www.ophi.org.uk 29
Appendix
Starting from expression (1):
" #1
w1 [I1 (x1 )]β + ... + wq [Iq (xq )]β β
W (X|β) = for β 6= 0 (19)

w1 + ... + wq
W (X|0) = I1 (x1 )w1 /(w1 +···+wq ) ∗ · · · ∗ Iq (xq )wq /(w1 +···+wq ) for β = 0 (20)
For proposition 1:
∂W (X|β) 1h i 1 −1
−1
w1 [I1 (x1 )]β + ... + wq [Iq (xq )]β [Ij (xj )]β [w1 + ... + wq ] β
β
=
∂wj β
1 i1
− 1 −1
h
w1 [I1 (x1 )]β + ... + wq [Iq (xq )]β
β
− [w1 + ... + wq ] β
β
∂W (X|β) 1h i 1−β
β β β −1
= w1 [I1 (x1 )] + ... + wq [Iq (xq )] [w1 + ... + wq ] β ∗
∂wj β
h i
[Ij (xj )]β − W (X)β
Ij (xj )β − W (X)β

∂W (X|β)
= (21)
∂wj β [w1 + ... + wq ] W (X)β−1
When β = 0
!
∂W (X|0) w1 /(w1 +···+wq ) w1 /(w1 +···+wq ) 1 wj
= I1 (x1 ) ∗ ... ∗ Iq (xq ) −
∂wj w1 + ... + wq (w1 + ... + wq )2

∂W (X|0) W (X/0) wj
= 1− (22)
∂wj w1 + ... + wq w1 + ... + wq
Pq
If also j w)j = 1 then
∂W (X|0)
= W (X/0) (1 − wj )
∂wj
For proposition 2:
www.ophi.org.uk 30
" # 1−β
∂W (X|β) w1 [I1 (x1 )]β + ... + wq [Iq (xq )]β β wj
= [Ij (xj )]β−1 Ij0 (xj )
∂xj w1 + ... + wq w1 + ... + wq
Ij (xj ) β−1

∂W (X|β) wj 0
= I (xj ) (23)
∂xj w1 + ... + wq j W (X|β)
For β = 0
∂W (X|0) wj 0 W (X|β)
= I (xj )
∂xj w1 + ... + wq j Ij (xj )
For proposition 3:
i 1−β
w1 [I1 (x1 )]β +...+wq [Iq (xq )]β
h β wy
∂W (X|β)
∂xy w1 +...+wq w1 +...+wq [Iy (xy )]β−1 Iy0 (xy )
M RSyh = ∂W (X|β)
=h i 1−β
w1 [I1 (x1 )]β +...+wq [Iq (xq )]β β
∂xh
w1 +...+wq
wh
w1 +...+wq [Ih (xh )]β−1 Ih0 (xh )
∂W (X|β)
wy Iy0 (xy ) Iy (xy ) β−1

∂xy
M RSyh = = (24)
∂W (X|β) wh Ih0 (xh ) Ih (xh )
∂xh
For β = 0
∂W (X|β)
∂xy wy Iy0 (xy ) Ih (xh )
M RSyh = =
∂W (X|β) wh Ih0 (xh ) Iy (xy )
∂xh
www.ophi.org.uk 31
Tables
www.ophi.org.uk 32
Table 1: Transformation functions
Transformation method Definition Characteristics Examples

Identity I(.) = xj
Linear scale transformations
xj −x̄j
- z-scores Ij = sd(x̄j ) Imposes a standard normal distribution. Environment Sustainability Index
Decancq & Lugo
www.ophi.org.uk
It produces negative values.
Extreme values are given a large weight.
xj −min(xj )
- by range Ij = max(xj )−min(xj ) Robust to outliers Human Development Index
Index of Economic Well-Being
zj −xj
- shortfall Ij = zj where zj is the dimension-specific threshold Human Poverty Index
Ratio scale transformations
xtj −xt−1j
- over time Ij = xtj
Only feasible with longitudinal data Economic Sentiment Indicator
xj
- by mean Ij = x̄j Not robust to outliers
(or relative distance from)
xj
- by best performer Ij = max(xj ) Not robust to outliers Environment Policy Index
Increasing transformations
Logarithmic Ij = ln(xj ) Coefficients are interpreted as elasticities.
Higher weight to changes at the bottom
Exponential and rank R = 23 ∗ ln{1Rj ∗ (1e100/23 )} where Rj is the rank in each xj Multiple Deprivation Index
(or relative distance from)
Ordinal transformations
Ranking Ij = rank(xj ) Uses ordinal information only, Medicare Health Care Index
hence discarding all level information
Other transformations
xj
Number of indicators Ij = x̄j − (1 + p) p arbitrary threshold above/below the mean
above/below mean
33
Working Paper No.18
Scoring Ij = score(xj ) Discards information when original

variable is continuous
Decancq & Lugo
www.ophi.org.uk
Table 2: Selection of composite well-being indices
Composite Index Dimensions Transformation Ij (xj ) β Dimensions weights wj
Human Development Index Income, education, health linear scale - range 1 equal weight
increasing - log(pcGDP)
Human Poverty Index - 1 rates of survival, education, economic perf identity 3 equal weight
Human Poverty Index - 2 (rates for) health, knowledge, identity 3 equal weight
economic, social exclusion
i
4
hP
Gender-related development GDI = j=1 w j E(x j ) equal weight
1/β
index E(xj ) = f I(xjf )β + f I(xjm )β linear scale - range -1 gender frequency weight
xjf : achievement of female (m: male)
f : female population share (m: male)
Index of Multiple Deprivations 7 domains increasing - exponential 1 participatory for domains
38 indicators factor analysis and equal
weight for indicators
Index of Economic Well-Being 4 domains: consumption, accumulation linear scale - range 1 equal weights for domains
Osberg and Sharp (2002) income distribution, economic security frequency weights w/in security
16 indicators arbitrary/equal weights for others
34
Working Paper No.18
Table 3: Selection of underlying well-being indices in multidimensional inequality and poverty measures with empirical application
Composite Index Dimensions Transformation Ij (xj ) β Dimensions weights wj
Brandolini (2007) [1,2,5, range [0,1]
10,100,500]
De Kruijk (2007) 12 dimensions linear scale - shortfall 1 normative priority weights
indicators
Decancq & Lugo
www.ophi.org.uk
Decancq, Decoster, income, health, education linear scale - range [-5,1] equal and principal
and Schokkaert (2007) ratio scale - by mean component weights
linear scale - z-score
increasing - logarithm
Desai and Shah (1988) family composition, education, linear scale - shortfall 1 frequency weights
income, wealth, region, health,
place of birth
Despotis (2005) income, education, health linear scale - range 1 most favourable weights
Deutsch and Silber (2005) 1 equal and
frequency weights
Justino (2005) income/expenditure, education, linear scale - range [-1/3, 1] equal and
health range weights
Kirshnakumar (2007) education, health, social identity 1 latent variable model
participation
(9 indicators)
Klasen (2000) 14 indicators scoring (1,5) 1 principal component analysis
and equal weights
Lugo (2007) income, health, education linear scale - range [-20,1] equal weights
Nilsson (2007) expenditure, land holding, linear scale - range [-20,1] equal weights
education, health
Maasoumi and Lugo (2008) expenditure, health, education identity [-3, 1] equal weights
linear scale - shortfall
35
Working Paper No.18
Mahlberg and Obersteiner (2001) income, education, health linear scale - range 1 most favourable weights
Schokkaert, Fleurbaey, Decancq (2008) income, health, housing identity 1 regression-based weights
employment

W P N - 18 Setting Weights in Multidimensional Indices of Well-Being and Deprivation

Uploaded by

Copyright:

Available Formats

W P N - 18 Setting Weights in Multidimensional Indices of Well-Being and Deprivation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

W P N - 18 Setting Weights in Multidimensional Indices of Well-Being and Deprivation

Uploaded by

Copyright:

Available Formats

Oxford Poverty & Human Development Initiative (OPHI)

Oxford Department of International Development

OPHI WORKING PAPER NO. 18

Keywords: weights, multidimensional well-being index, multidimensional deprivation index

JEL classification: I32, D63

Oxford Poverty & Human Development Initiative (OPHI)

When it comes to operationalizing the multidimensional approaches, one quickly runs

“Since any choice of weights should be open to questioning and debating in

Expression (1) can also be used to construct an index of multidimensional poverty or

for poverty measure is the relative shortfall defined as

We can then defined a generalized class of deprivation indices D(x|β) as follows:

Figure 1: Iso- well-being curves for β = 0, 1, −∞

a weighted arithmetic mean of the logs of the transformed achievements.

A potentially crucial problem of setting β = 0 is that when a person has no achievement

A second alternative is to let β go to −∞(+∞) where the elasticity of substitution

W (x| − ∞) = min[I1 (x1 ), ..., Iq (xq )]. (7)

Figure 2: Iso- well-being curves. Alternative dimension-weights.

Figure 3: Iso- well-being curves. Alternative transformation (median).

Finally, in figure 2 we use a lower degree of substitutability between dimensions β =

3 What do weights mean?

Figure 4: Iso- well-being curves. Alternative β.

Irrespective of the selected β, an increase in the weight of dimension j leads to an

The impact of a small change of the achievements of dimension j on total well-being

If instead β = 0 the effect of a marginal increase in dimension j on the well-being

dxh ∂W (x|β) . ∂W (x|β)

By substituting expression (9) into (12) we obtain the following expression:

wy Iy0 (xy ) Ih (xh ) 1−β

wy Iy0 (xy ) Ih (xh )

4 How can we set weights in a reasonable way?

4.1 Equal weights

The most commonly used approach to weighting in multidimensional indices of well-being

4.2 Data-driven weighting schemes

4.2.1 Frequency-based weights

4.2.2 Most favorable weights

4.2.3 Statistical weights

4.2.4 Regression based weights

Yi = α1 I1 (x1i ) + ... + αq Iq (xqi ) + εi , (18)

In a recent paper, Schokkaert (2007) proposes to rely on the emerging measures of

Data-driven approaches offer an interesting way to obtain weights for multidimensional

4.3 Normative weights

In this paper, we surveyed different approaches for setting weights in multidimensional

Atkinson, A. B. & Bourguignon, F. (1982). The comparison of multi-dimensioned distribu-

Blackorby, C. & Donaldson, D. (1982). Ratio-scale and translation-scale full interpersonal

Bourguignon, F. (1999). Comment on ‘Multidimensioned approaches to welfare analysis’ by

Bourguignon, F. & Chakravarty, S. R. (2003). The measurement of multidimensional

Brandolini, A. (2007). On synthetic indices of multidimensional well-being: health and

Carlucci, F. & Pissani, S. (1995). A multiattribute measure of human development, Social

Crawford, I. (1999). Nonparametric tests of stochastic dominance in bivariate distributions,

Decancq, K. & Lugo, M. A. (2008). Measuring inequality of well-being: A proposal based

Desai, M. & Shah, A. (1988). An econometric approach to the measurement of poverty,

Deutsch, J. & Silber, J. (2005). Measuring multidimensional poverty: An empirical com-

Duclos, J.-Y., Sahn, D. E. & Younger, S. D. (2006). Robust multidimensional poverty

Gravel, N., Moyes, P. & Tarroux, B. (2005). International multidimensional comparisons

Jacobs, R., Smith, P. & Goddard, M. (2004). Measuring performance: An examination of

Justino, P. (2005). Empirical applications of multidimensional inequality analysis. Poverty

Kahneman, D. & Krueger, A. B. (2006). Developments in the measurements of subjective

Krishnakumar, J. (2007). Going beyond functionings to capabilities : an econometric model

Krishnakumar, J. & Ballon, P. (2007). Estimating basic capabilities: A structural equation

Krishnakumar, J. & Nadar, A. (2008). On exact statistical properties of multidimensional