Professional Documents
Culture Documents
doi:10.1111/ecca.12231
Standard measures of multidimensional inequality (implicitly) assume common preferences for all
individuals, and hence are not sensitive to preference heterogeneity among members of society. In this
paper, we measure the inequality of the distribution of equivalent incomes, which is a preference-sensitive
multidimensional wellbeing measure. To quantify the contribution of preference heterogeneity to
wellbeing inequality, we use a decomposition method that calculates wellbeing inequality in dierent
counterfactual distributions. We focus on four sources of wellbeing inequality: the correlation between
outcomes and preferences, the preference heterogeneity, the correlation between the outcome dimensions,
and the inequality within each of the outcome dimensions. We nd that preference heterogeneity accounts
for a considerable part of overall wellbeing inequality in Russia for the period 19952005.
INTRODUCTION
It is now widely accepted that a concern for inequality in society should go beyond an
exclusive focus on income distribution and should also consider the distribution of other
dimensions that make life go well (see Stiglitz et al. 2009). This immediately raises the
issue of how to aggregate the outcomes in the dierent life dimensions into a single
measure of wellbeing inequality. Should we aggregate rst across individuals or rst
across dimensions in the measure of wellbeing inequality? Should we take individual
preferences into account in the aggregation across dimensions? If one decides to use a
preference-sensitive measure of multidimensional wellbeing inequality, then how large is
the contribution of preference heterogeneity to overall inequality?
The rst question has been addressed in the literature before. Although it is not
always made explicit, most existing multidimensional inequality measures perform a
two-step aggregation, with one aggregation across dimensions and another across
individuals. There are two possible sequences to perform these two aggregations, and in
general both sequences lead to dierent results (see Kolm 1977; Dutta et al. 2003).
Decancq and Lugo (2012) have illustrated with Russian data that the dierence between
both sequences can be empirically relevant. In the rst sequence, one aggregates rst
across the individuals in each dimension and then across the dimensions (see, for
example, Gajdos and Weymark 2005). From a normative point of view, this procedure
has a crucial drawback. It does not capture the cumulative deprivation that occurs if the
positions of the individuals across the dierent dimensions are correlated. This problem
can be solved if one follows the second sequence, in which one rst aggregates the
outcomes across the dimensions of wellbeing into a measure of wellbeing for each
individual, and then aggregates the wellbeing measures across individuals.
This brings us to the second question. In the existing literature on multidimensional
inequality, the specication of the individual wellbeing measure that is used in the rst
step does not relate to the literature on the measurement of individual wellbeing (for a
discussion, see Decancq et al. 2015b). It is typically determined by axioms that are
formulated over the entire multidimensional aggregation process. More specically, the
PigouDalton transfer principle is generalized to more dimensions and then directly
imposed in the multidimensional space of outcomes.1 Since it has been shown that there
2017 The London School of Economics and Political Science. Published by Blackwell Publishing, 9600 Garsington Road,
Oxford OX4 2DQ, UK and 350 Main St, Malden, MA 02148, USA
2017] WELLBEING INEQUALITY 211
for the Russian Federation between 1995 and 2005.2 To compute equivalent incomes
and the inequality in their distribution, we estimate rst thepotentially
heterogeneouspreferences of the respondents over their expenditures, health,
housing quality, unemployment and wage arrears on the basis of a life satisfaction
equation. Using the proposed decomposition, we nd that along with inequality in
the expenditure and health dimension, preference heterogeneity accounts for a
considerable part of wellbeing inequality. This nding is conrmed by the subgroup
decomposition of wellbeing inequality within and between population subgroups
with the same preferences.
The paper is structured as follows. Section I discusses how wellbeing inequality can
be measured while respecting preference heterogeneity. Section II illustrates how the
proposed approach can be implemented with data from Russia. Section III presents a
decomposition of wellbeing inequality in four components. It is the core of our paper.
Section IV presents a subgroup decomposition of wellbeing inequality based on
population subgroups with the same preferences. Section V discusses how
multidimensional dominance approaches relate to our measure of wellbeing inequality,
and how they tackle (or do not tackle) preference heterogeneity. In Section VI, we
conclude and briey discuss the normative implications of our ndings.
1 IL; A I WB1 ; a1 ; . . .; WBn ; an :
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 213
with dierent specic proposals corresponding to specic choices of I() and of the
wellbeing measure WB(,).4 In the rest of this section, we rst explain why the
simplication embodied in equation (2) is unavoidable if one wants to respect the
multidimensional PigouDalton transfer principle, and then introduce the equivalent
income measure as one specic preference-sensitive proposal to measure wellbeing.
Economica
2017 The London School of Economics and Political Science
214 ECONOMICA [APRIL
Dimension 2
1
i
2
i
2
j
1
j
3
j
4
j
4
i
3
i
O Dimension 1
FIGURE 1. The impossibility of a Paretian egalitarian.
less realistic. Consider as an example the number of hours worked, where the optimal
value may dier for a typical academic and a typical low-skilled blue-collar worker
(Fleurbaey and Blanchet 2013).
Now consider two individuals who both reach their own optimal outcome level in all
non-income dimensions. Our basic assumption for making interpersonal wellbeing
comparisons is that when comparing the wellbeing level of these two individuals, we can
restrict ourselves to comparing their incomes, independently of their actual preferences.
Why should their preferences matter if they reach their optimal outcome level in all non-
income dimensions? As shown in Decancq et al. (2015a), combining this assumption with
respect for preferences characterizes the wellbeing ordering that compares individuals in
terms of their equivalent incomes.7 The equivalent income is formally dened as the
solution 1
i of the equation
3 1 ^2 ^m
i ; i ; . . .; i Iai i ; i ; . . .; i :
1 2 m
In other words, it is the hypothetical level of income that, combined with the optimal
outcome level in the non-income dimensions, keeps the individual on the indierence
curve corresponding to her actual situation. If preferences are monotonic with respect to
income, then this equivalent income 1 i cannot be larger than the actual income level i
1
1 ^2 ^
because i ; i ; . . .; i Rai i ; i ; . . .; i .
m 1 2 m
The equivalent income can be interpreted as the income corrected for the loss in
wellbeing associated with a suboptimal outcome level for the non-income dimensions.
This is an intuitively attractive way of capturing the idea of multidimensional
deprivation, we believe. Moreover, it is conveniently measured in monetary units, which
provides a simple and familiar cardinal scale. In the following, we measure wellbeing
using equivalent incomes; that is, we have that WBi ; ai 1 i :
Once we have calculated an equivalent income for all individuals in society, we can
implement equation (1) using any unidimensional inequality measure. In our empirical
application we will work with the generalized entropy class of inequality measures (see
Cowell (2011) and the references therein):
" #
1 Xn
WBi ; ai a
4 GEa L; A 1 ;
aa 1n i1 l
1X n
WBi ; ai :
n i1
The lower the value of the parameter a, the more we focus on the bottom part of the
distribution of wellbeing measures. We will concentrate on the mean logarithmic
deviation (GE0) and the Theil index (GE1) in this paper, because these inequality
measures have attractive decomposability properties.
This approach obviously satises the PigouDalton transfer principle in the space of
wellbeing measures. Given the impossibility result discussed in the previous section, we
know that equation (4) does not satisfy the multidimensional PigouDalton transfer
principle. However, the measure does satisfy a restricted PigouDalton transfer principle
Economica
2017 The London School of Economics and Political Science
216 ECONOMICA [APRIL
that applies only to situations involving the optimal outcome levels in the non-income
dimensions ^2i ; . . .; ^m
i .
0 0
Restricted PigouDalton transfer principle (L , A ) is strictly better than (L, A) if for all
individuals k 6 i, j we have that k 0k , and for individuals i and j with outcomes
i 1i ; ^2i ; . . .; ^m
i and j 1j ; ^2j ; . . .; ^m
j we have that for d 2 R ,
i i d; i ; . . .; i and 0j 1j d; ^2j ; . . .; ^m
0 1 ^2 ^m
j with i d \ j d.
1 1
X
5
5 Sit ai ct bj lj0 Dit Cj jit d0 Zit uit :
j1
This life satisfaction equation includes ve life dimensions.9 The rst three life
dimensions are measured by continuous variables: 1it denotes real equivalized
household expenditures (with the square root of household size as the equivalence
scale). The dimension 2it captures health, measured as a composite index of objective
health and lifestyle indicators such as indicators of diabetes, a heart attack, anaemia,
a recent operation or hospitalization, a regular health check-up, and indicators of
regular exercises and jogging. The weights in the composite index are obtained from
an ordered logit regression with self-assessed health as the dependent variable. All
individuals with the same objective indicators have the same health score. Housing
quality is captured by 3it and is measured as the predicted value of a hedonic
regression of self-reported housing values on a number of housing characteristics
such as size, presence of metered gas, hot water and central sewerage (after
controlling for regional price dierences, a time trend and household size). The nal
two life dimensions are binary indicators of unemployment (4it and wage arrears
(5it . The latter indicator captures the phenomenon that wages were often not paid
on time in Russia during the late 1990s.
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 217
We allow for the non-linearity of the life satisfaction equation (and hence for less
than perfect substitutability between the dimensions) through a so-called BoxCox
transformation of the continuous dimensions j = 1, 2, 3 (see Box and Cox 1964):
( j
j j jit h 1 =hj when hj 6 0;
C it
lnjit when hj 0:
For the other two binary indicators, j is the identity function so that these
dimensions are not transformed. The scaling of life satisfaction in equation (5) is allowed
to be inuenced by a number of sociodemographic characteristics Zit (education, social
status and marital status) that are introduced together with time dummies ct as control
variables.10 We include individual xed eects ai to control for unobserved individual
heterogeneity in time-invariant characteristics including personality traits. To model
preference heterogeneity we include interaction eects between the outcomes and four
dummy variables contained in Dit. These dummies capture whether the respondent is
living in a rural area, is young (below the age of 33), is male, and reports to have obtained
higher education. Finally, uit is a disturbance term.
The scalars bj and hj , as well as the vectors lj and d, are coecients to be estimated.
Since the observed life satisfaction responses are measured on an ordinal scale, we
estimate an ordered logit model. We incorporate individual xed eects into the
estimation using the approximation proposed by Jones and Schurer (2011) of the method
discussed by Ferrer-i-Carbonell and Frijters (2004), and Frijters et al. (2004, 2006). The
three BoxCox parameters of the continuous variables are chosen on the basis of a grid
search to maximize the overall t of the model. Standard errors are corrected for
clustering at the household level.
The estimation results are summarized in Table 1. The results for the full model are
shown in the rightmost pair of columns (Model 4). The results for the life dimensions and
the sociodemographic control variables are in line with what is usually found in the
literature. Interesting for our purposes are the interaction coecients lj, as these
coecients capture the heterogeneity in preferences. These coecients will be treated as
the preference parameters ai in the wellbeing measure. Even with our restricted set of ve
life dimensions, many interactions would need to be estimated. We therefore simplify the
model by including only the interaction eects that are signicant at the 10% level. In
fact, not dropping the insignicant coecients would lead to imprecisely computed
equivalent income wellbeing measures and wellbeing distributions. The remaining
interaction terms in Model 4 can be interpreted easily. Unemployment has a stronger
negative eect on the life satisfaction of older, higher-educated males. Expenditures are
relatively more important for the higher educated in urban areas. Young females give a
relatively smaller weight to health. Housing matters more in rural areas, and less for the
highly educated respondents. Finally, the BoxCox parameters of the continuous
variables are shown at the bottom of the columns, and are equal to 0.08, 0.46 and
0.36 (for expenditures, health and housing quality, respectively).
The other columns in Table 1 contain the results for restricted versions of the full
model. Model 3 imposes that hj h for the continuous dimensions. The resulting estimate
of h equals 0.05. This restriction is close to being rejected by a standard likelihood ratio test
(v2(2; 4.32) = 0.115), and is signicant from an economic point of view. Model 2 keeps the
dierentiated BoxCox parameters but removes all preference heterogeneity. This
restriction is clearly rejected on statistical grounds (v2(9; 59.24) = 0.000). Model 1 is the
most restricted model without interactions and with the same BoxCox parameter for the
Economica
2017 The London School of Economics and Political Science
218
Economica
TABLE 1
FOUR LIFE SATISFACTION MODELS: MODELS 1 AND 2 WITHOUT INTERACTIONS, MODELS 1 AND 3 WITH COMMON BOXCOX PARAMETER
Economica
TABLE 1
CONTINUED
Notes
Clustered standard errors in parentheses.
+
, *, **, *** indicate p < 0.10, p < 0.05, p < 0.01, p < 0.001, respectively.
219
220 ECONOMICA [APRIL
three rst life dimensions. It is therefore very close to the common homothetic specication
of wellbeing without preference heterogeneity that is often (implicitly) used in the standard
multidimensional inequality measures (as in equation (2)). This restricted model is strongly
rejected with respect to the full Model 4 (v2(11; 64.23) = 0.000).
We will use the estimates of Model 4 as presented in the rightmost pair of columns of
Table 1 to compute the inequality in equivalent incomes in the next section, except for
one simplication. Since the BoxCox parameter of expenditures is close to zero, we
equalize it to zero and work with the logarithm of equivalized expenditures as the
relevant monetary life dimension. This brings us closer to the specication that is used in
the bulk of the literature on life satisfaction (for instance, see Layard et al. 2008). As will
become clear in the following subsection, this restriction also prevents the equivalent
incomes from becoming negative, which causes computational problems when
calculating the generalized entropy inequality measure.
Under this consistency assumption, the life satisfaction function is one possible utility
function that provides a representation of the preference ordering of individual i, just as
the equivalent income wellbeing measure. Equivalent incomes and the life satisfaction
function have dierent ways of attaching a label to the indierence curves, however. In
Decancq et al. (2015a), we have argued that equivalent incomes are interpersonally
comparable in a normatively attractive way, whereas the life satisfaction functions are not.
To compute equivalent incomes, we rst determine for each individual her optimal
outcome level in the non-income dimensions ^2i ; . . .; ^m
i . The estimates presented in
Table 1 lead to preferences that are monotonic with respect to all life dimensions, so that
the optimal values will be the same maximal value for all respondents. To be precise, they
are set at being in perfect health, having a high housing quality,11 not being unemployed,
and not suering from wage arrears.
Under the consistency assumption of equation (6), we can use the denition of
equivalent incomes given by equation (3) and the econometric specication of the life
satisfaction equation given by equation (5) to write
X
5
Sit ai ct b1 l10 Dit ln1it bj lj0 Dit Cj jit d0 Zit uit
j2
X
5
ai ct b1 l10 Dit ln1
it bj lj0 Dit Cj ^jit d0 Zit uit ;
j2
which yields
" #
X5
bj lj0 Dit j j j ^j
7 1
it 1it exp C it C it :
j2 b l Dit
1 10
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 221
The shape of the indierence curves as determined by the ratio of the coecients of
0 0
the transformed non-income dimension and income (bj + lj Dit)/(b1 + l1 Dit), and the
transformation functions j and 1 are of crucial importance in equation (7). On the
contrary, the conditioning variables Zit, the xed eects ai, the time trends ct and the
idiosyncratic disturbance term do not appear in equation (7). These variables only shift
the level of reported life satisfaction upwards or downwards, without aecting the
marginal rates of substitution between the life dimensions. These shifts can be interpreted
as changes in aspirations and expectations, and are considered irrelevant in making
wellbeing comparisons by means of equivalent incomes.12
from the sample. Since the resulting outcome matrix L~ is not unique, we will again
generate 2000 dierent reshued outcome matrices.
Note that in the outcomes-rst decomposition we cannot measure the eect of the
correlation between outcomes and preferences on wellbeing inequality. When all
individuals have the same outcome vector, permuting the preference vectors does not
~ by construction.
aect inequality, therefore IL; A IL; A
As is common with this type of decomposition, the results are path-dependent.
Therefore the results of the preferences-rst and outcomes-rst decompositions will be
dierent. Yet we do not consider this path-dependence to be problematic since our aim is
not to obtain a unique decomposition, but rather to understand the contribution of the
dierent components and the interactions between them. Combining the results from both
decompositions will therefore yield useful additional insights. However, the preferences-
rst decomposition opens the possibility to identify the eect of the correlation between
preferences and outcomes, which is essential in the interpretation of a preference-based
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 223
approach to wellbeing inequality. Moreover, the sum of the two rst terms in that
decomposition (I(L,A)I(L,A)) can be interpreted easily as the dierence between
inequality in wellbeing taking preferences into account and traditional multidimensional
inequality under the assumption of uniform preferences. This result is interesting in itself.
Empirical results
~ A, L~ and L for each
Based on the preference estimates in Table 1, we can construct A,
considered period. With these building blocks we then construct various counterfactual
wellbeing distributions and compute the corresponding generalized entropy inequality
measures. Table A1 in the Appendix provides the results. We rst discuss the results of
the preferences-rst decomposition, which are summarized in Figures 2 and 3 for the
mean logarithmic deviation (GE0) and the Theil index (GE1), respectively. We will then
describe the additional insights that can be obtained from the outcomes-rst
decomposition for the same inequality measures (Figures 4 and 5).
The curves at the tops of the gures show the evolution of GE0(L,A) and GE1(L,A)
over the considered period. When using the mean logarithmic deviation (GE0), wellbeing
inequality remains roughly at the same level between 1995 and 2005, but peaks in 1996
and 2000. When using the less bottom-sensitive Theil index (GE1), there is no peak in
2000 and inequality declines steadily after 1996.
1.5
GE0 (L, A)
~
GE0 (L, A)
1
GE0 (L, A)
GE0 (L1, A)
0.5
GE0 (L1, A)
GE0 (L2, A)
GE0 (L, A)
0
is the loss in wellbeing of not reaching the optimal outcome level; hence the negative
correlations for the housing dimension increase the dierence between the wellbeing of
those individuals who do better and the wellbeing of those who do worse on that
dimension.
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 225
1
GE1 (L, A)
0.8
GE1 (L, A)
GE1 (L, A)
0.6
GE1 (L, A)
0.4
GE1 (L1, A)
0.2
GE1 (L2, A)
GE1 (L4, A) GE1 (L3, A)
GE1 (L, A)
0
Economica
TABLE 2
SPEARMAN RANK CORRELATION COEFFICIENT BETWEEN OUTCOME LEVEL AND RATIO OF THE COEFFICIENTS OF THE TRANSFORMED NON-
INCOME DIMENSION AND INCOME
Unemployment 0.0351 0.0644 0.0655 0.0842 0.0674 0.0754 0.0781 0.0797 0.0812
Wage arrears 0.0556 0.0870 0.1326 0.0080 0.0213 0.0277 0.0324 0.0624 0.0377
common wellbeing measure for all individuals and, in that sense, resemble the standard
multidimensional inequality measures of equation (2). As noted before, the dierence
between this standard approach based on a common wellbeing measure and our
heterogeneous approach shows up in the dierences between I(L,A) and I(L,A) in Figures
2 and 3. After neutralizing the correlation between preferences and outcomes, the net
eect of preference heterogeneity shows up in the dierences between IL; A ~ and I(L,A).
The results are striking. Removing preference heterogeneity leads to a substantial
decrease in wellbeing inequality. While the traditional multidimensional inequality
measures would not record an increase in inequality in 2000 for GE0, our measure does.
The eect of substituting the averaged preference matrix A for the reshued matrix A~ is
in general larger than that of substituting the reshued matrix A~ for the actual matrix A.
However, the increase in wellbeing inequality in the tumultuous period between 1998 and
2000 seems to be largely driven by the contribution of the correlation between the
outcomes and preferences.
We make two further observations. First, the contribution of preference heterogeneity
to wellbeing inequality remains quite stable over time. This is not surprising since we have
assumed that preferences are constant over time for each person. Second, the contribution
of preferences is relatively larger for GE0 than for GE1.15 One possible interpretation is
that individuals at the top of the wellbeing distribution score well on their non-income
dimensions, so there is only limited room for preference heterogeneity to aect their
wellbeing. For individuals at the bottom of the wellbeing distribution, on the contrary, the
relative weighting of their dierent deprivations, or the dierences in their willingness to
pay, become potentially more important for wellbeing inequality.
~
Inequality in outcomes (IL;A IL; A) Let us nally look at the fourth term of the
decomposition, which captures the contribution of the inequality in each of the outcome
dimensions to overall wellbeing inequality. Again, the results for the preferences-rst
Economica
2017 The London School of Economics and Political Science
228 ECONOMICA [APRIL
TABLE 3
SPEARMAN RANK CORRELATION COEFFICIENT BETWEEN OUTCOME DIMENSIONS
decomposition (Figures 2 and 3) are similar to the results that would be obtained by the
standard approach to multidimensional inequality.17
We neutralize the inequality in the dierent dimensions in a specic order: rst we
average expenditures, then we average consecutively health, housing quality,
unemployment and wage arrears. In principle, this specic sequence may aect the
results, and other sequences may lead to dierent results. Yet since the equivalent income
wellbeing measure as dened by equation (7) is close to being additively separable, this
eect is quite small, and reversing the sequence hardly changes the results.18
Overall, the most important contributors to overall wellbeing inequality are the
inequality in the expenditure and health dimensions. Moreover, there is a remarkable
increase in wellbeing inequality due to the presence of wage arrears around 1996. The
eect of wage arrears tapers o over time, however.
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 229
1.5
GE0 (L, A)
GE0 (L, A)
1
GE0 (L1, A)
GE0 (L2, A)
GE0 (L4, A)
0.5
GE0 (L3, A)
GE0 (L, A)
GE0 (L, A)
0
Economica
2017 The London School of Economics and Political Science
230 ECONOMICA [APRIL
GE1 (L, A)
0.8
GE1 (L, A)
0.6
GE1 (L1, A)
0.4
GE1 (L4, A)
GE1 (L2, A)
GE1 (L3, A)
GE1 (L, A)
0.2
GE1 (L, A)
0
captures inequality in the situation where all individuals have the same averaged
outcomes, but their own preferences. According to the standard approach, which uses
only information about L, this inequality is necessarily equal to 0. As we discussed in the
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 231
GEa L; A GEW
a L; A GEa L; A;
B
where the between component is computed by setting all equivalent incomes in each
preference group equal to their group average, and the within component is given by
K h
X i
a L; A
GEW vk 1a sk a GEa Lk ; Ak ;
k1
with Lk and Ak being the outcome matrix and preference matrix for preference subgroup
k (k = 1, . . . ,K), vk = nk/n being the population share, and sk being the equivalent
income share. When a = 0, the inequality within the preference groups is weighted by the
population shares, whereas for a = 1, the equivalent income shares are used.19
Tables 4 and 5 present the results. For I(L,A), the within component dominates: since
preferences are the same within each preference group, this component captures the
dierences in outcomes between the individuals within each preference group. The
between component captures both outcome and preference dierences between the
preference groups.
Additional insights can be obtained by looking at the decomposition of I(L,A). Since
all preferences are equalized in this counterfactual distribution, the between component
reects only dierences in outcomes between the preference groups. The between
component is now much smaller, suggesting that preference heterogeneity is an
important contributor to the dierences between the preference groups. As can be seen
from comparing Tables 4 and 5, the within-group inequality is larger for the mean
logarithmic deviation with, rather than without, preference heterogeneity; that is,
GEW 0 L; A [ GE0 L; A, while the opposite is true with the Theil index, that is,
W
GEW 1 L; A \ GE 1 L; A. This is in line with our earlier nding that taking preference
W
Economica
2017 The London School of Economics and Political Science
232 ECONOMICA [APRIL
TABLE 4
SUBGROUP DECOMPOSITION ANALYSIS FOR GE0
TABLE 5
SUBGROUP DECOMPOSITION ANALYSIS FOR GE1
heterogeneity into account has a larger eect for inequality measures that focus more on
the bottom of the wellbeing distribution.
Finally, the subgroup decomposition of I(L,A) conrms the ndings of the results
subsections of Section III. The within component now becomes zero, while the between
component captures the eect of preference heterogeneity in the counterfactual situation
when all outcomes are averaged.
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 233
however. The resulting wellbeing inequality ranking may turn out to be incomplete,
meaning that some comparisons will be indecisive.
To illustrate, we implement the wellbeing inequality ranking20
where A^ is the set of all preference matrices in which all individuals share a common
preference vector a^ that is observed in the sample. As we have seen in the previous
section, there are 16 dierent preference groups in our empirical analysis, and hence 16
dierent preference matrices A. ^ Clearly, this dominance idea could be further generalized
by checking the inequality in equation (10) for various members of some class of
inequality measuresby testing Lorenz dominance, for instance. However, testing
dominance is not the purpose of this paper and we illustrate the approach only for a
single measure, which is the mean logarithmic deviation GE0.
Table 6 presents the results for all pairwise year-by-year tests of the dominance test
given by equation (10). A cell with a 1 denotes that the row year has a lower wellbeing
inequality according to GE0 for each of the 16 common preference matrices A. ^ We see
that all years are less unequal than 1996, and that most of the years after 2000 are less
unequal than 2000.
It is important to stress that the dominance approach does not take into account the
diversity in preferences in a given society at a given point in time. There is an important
dierence between on the one hand looking for a unanimous inequality ranking for
dierent wellbeing measures, each of them common to all individuals in society, and on
the other hand measuring wellbeing inequality while respecting preference heterogeneity.
Although one has to interpret our ndings cautiously, this dierence can be illustrated
by comparing the results in Table 6 with those in Figures 2 and 4. According to both
approaches, we nd that 1996 is more unequal than 1995 (and more unequal than 1998).
Yet while 2000 was more unequal than 1998 in Figures 2 and 4, this is not found in the
dominance results. This dierence may have to do with the underlying causes of the
increases in wellbeing inequality. The increase in wellbeing inequality in 1996 is due to the
sharp wellbeing loss as a result of the presence of wage arrears, which is an objective
phenomenon that aects the inequality for all preference matrices A.^ The increase in 2000,
TABLE 6
DOMINANCE TESTA 1 MEANS THAT THE ROW YEAR HAS A LOWER GE0 THAN THE
COLUMN YEAR FOR ALL OBSERVED PREFERENCE PARAMETERS
Economica
2017 The London School of Economics and Political Science
234 ECONOMICA [APRIL
however, is mainly due to an increase in the eect of the correlation between outcomes, and
the correlation between outcomes and preferences. Preference dierences are important for
the evaluation of the former correlation, and essential for the latter correlations. This may
explain why the sharp inequality increase in 2000 is not reected in the dominance results.
VI. CONCLUSION
We have shown that preference heterogeneity constituted an important part of wellbeing
inequality in Russia between 1995 and 2005. All in all, we have found that the main
drivers of wellbeing inequality in the considered period were preference heterogeneity,
expenditure inequality, health inequality and wage arrears inequality (during the late
1990s).
Some caveats apply, however. First, our empirical ndings are based on one dataset
only. In the period between 1995 and 2005, Russia was a specic setting, characterized by
large social and economic changes in a heterogeneous society. It is not clear whether
preference heterogeneity would be equally important in other settings. Second, our method
to estimate preferences on the basis of a satisfaction equation is arguably rather primitive.
In particular, the consistency assumption in equation (6) is debatable and is hard to test
empirically. Yet one could argue that the fact that we cannot identify individual
preferences with this method, but have to limit ourselves to only 16 dierent preference
groups, strengthens our conclusion on the empirical relevance of preference heterogeneity.
More important than our specic ndings for Russia, however, are the normative and
methodological questions that are raised by these ndings. Preference heterogeneity is
completely neglected by the standard approach to multidimensional inequality
measurement. Leaving pragmatic considerations of the availability of preference
information aside, this position has been justied on normative grounds. There seems to be
a certain distrust of individual preferences in the capability approach, for instance (Sen
1985). The capability approach has been very inuential in shaping the multidimensional
approach towards the measurement of wellbeing, inequality and poverty. Already before
Sen (1985), Kolm (1977) suggested in his seminal article that a common wellbeing measure
could be seen as the observers evaluation of the individual welfare, and Scanlon (1975)
wrote that the common objective opinion on what a good life is and what constitutes
wellbeing is rooted in some reasoned social agreement on basic components of wellbeing
and on the relative urgency of claims to dierent goods. As we have seen, the dominance
approach does not depart from the basic idea that there is one underlying common
wellbeing measure, but introduces the additional twist that there may be uncertainty about
this reasoned social agreement or a lack of consensus between dierent ethical observers.
As emphasized by preferentialists, however, the argumentation in favour of
neglecting individual preferences has a strong perfectionist avour. They claim that in a
pluralist society with widely divergent opinions about what constitutes a good life, public
policy in general, and inequality measurement in particular, cannot neglect these
divergences and should therefore take up preference heterogeneity.
The dierence between the two approaches is perhaps illustrated most strikingly by
their evaluation of the hypothetical situation in which all individuals in society have the
same objective outcomes. According to the standard multidimensional measurement
literature, no ethically relevant inequality remains in that situation. If one takes
preferences into account, however, then the match between the outcomes and the
preferences is brought into the picture, and it is seen as ethically relevant that dierent
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 235
individuals can attach dierent weights to the dierent dimensions and may therefore
have dierent wellbeings, even when their objective outcomes are the same.
Standard multidimensional inequality measures and dominance approaches are
arguably the best way to proceed if one believes that individuals do not have well-dened
conceptions of the good life, or that, even when they exist, it is impossible to know them,
or that, even when they exist and one can approximate them, one should not do so, but
rather implement an objective conception of the good life. If, on the other hand, one does
believe that an individual can form a well-considered opinion about what is important in
his or her own life, that these preferences can be reasonably (although imperfectly)
approximated, and that they should be respected in a pluralist society, then one should
introduce preference heterogeneity into the measurement of wellbeing inequality. This is
essentially a normative debate, to which we did not contribute in this paper. What we
have shown, however, are the stakes of the debate. The normative choices determining
the role of preference heterogeneity have a crucial eect on the resulting wellbeing
inequality. They do matter.
APPENDIX
TABLE A1
DECOMPOSITION RESULTS
Economica
2017 The London School of Economics and Political Science
236 ECONOMICA [APRIL
TABLE A1
CONTINUED
ACKNOWLEDGMENTS
We thank the Editor and two Referees for helpful comments and suggestions. We are grateful to
Florent Bresson, Claude dAspremont, Annemie Nys, and the participants of presentations in
Leuven, Louvain-la-Neuve, Luxembourg, the World Bank, and Clermont-Ferrand for their
comments.
Koen Decancq thanks the Center for Health and Wellbeing of Princeton University for its
hospitality during the academic year 201516.
NOTES
1. In the literature on the measurement of multidimensional inequality, the multidimensional PigouDalton
transfer principle is implemented by relying on results from majorization theory (see Arnold et al. 2011).
Weymark (2006), Aaberge and Brandolini (2015), and Chakravarty and Lugo (2016) provide overviews of
the literature on multidimensional inequality measurement.
2. To be precise, we use panel data from the Russia Longitudinal Monitoring Survey (RLMS-HSE) between
1995 and 2005. The RLMS-HSE has also been used to compute equivalent incomes by Decancq et al.
(2015a). Compared to that paper, we include two additional periods in the analysis, implement a slightly
dierent sample selection procedure, and use a more exible specication of the life satisfaction equation.
The same dataset has also been used by Decancq and Lugo (2012) to highlight the empirical dierences
between both aggregation sequences, focusing on a subset of the wellbeing dimensions considered in this
paper.
3. As we know from the booming literature on behavioural economics, these well-considered judgments are
not necessarily revealed in choice behaviour; see Fleurbaey and Schokkaert (2013).
4. Maasoumi (1986) has proposed a two-step multidimensional generalized entropy inequality measure, for
instance. Bosmans et al. (2015) interpret a normative two-step inequality measure as a measure of the social
welfare loss due to the suboptimal distribution of outcomes after removing the social welfare loss due to its
ineciency.
5. Let <, and denote the standard vector inequalities.
6. De la Vega et al. (2010) derive a class of multidimensional inequality measures consistent with this version
of the multidimensional PigouDalton transfer principle. However, in the axiomatic literature on
multidimensional inequality it is more common to consider uniform majorizationwhich captures transfers
where the transferred bundle is a fraction of the dierence between the outcome vectors of the donor and
recipient of the transferand to drop the restriction that the outcomes of the donor should be at least as
good as the outcomes of the recipient in all dimensions (see, for instance, Weymark 2006). These
modications of the multidimensional PigouDalton transfer principle do not change the impossibility
result discussed in this section, however.
7. We do not discuss the normative strengths and weaknesses of this proposal here, but we refer the interested
reader to Decancq et al. (2015a,b) for a discussion.
Economica
2017 The London School of Economics and Political Science
2017] WELLBEING INEQUALITY 237
REFERENCES
AABERGE, R. and BRANDOLINI, A. (2015). Multidimensional poverty and inequality. In A. B. Atkinson and F.
Bourguignon (eds), Handbook of Income Distribution, Vol. 2. Oxford: Elsevier, pp. 141216.
ARNOLD, B., MARSHALL, A. and OLKIN, I. (2011). Inequalities: Theory of Majorization and Its Applications.
New York: Springer.
ATKINSON, A. B. and BOURGUIGNON, F. (1982). The comparison of multidimensioned distributions of
economic status. Review of Economic Studies, 49(2), 183201.
BEEGLE, K., HIMELEIN, K. and RAVALLION, M. (2012). Frame-of-reference bias in subjective welfare. Journal of
Economic Behavior & Organization, 81(2), 55670.
BOSMANS, K., DECANCQ, K. and OOGHE, E. (2015). What do normative indices of multidimensional inequality
really measure? Journal of Public Economics, 130, 94104.
BOURGUIGNON, F. (1999). Comment on Multidimensioned approaches to welfare analysis by E. Maasoumi. In
J. Silber (ed.), Handbook of Income Inequality Measurement. London: Kluwer Academic, pp. 47784.
BOX, G. E. P. and COX, D. R. (1964). An analysis of transformations. Journal of the Royal Statistical Society,
Series B (Methodological), 26(2), 21152.
CHAKRAVARTY, S. and LUGO, M. (2016). Multidimensional indicators of inequality and poverty. In M. Adler
and M. Fleurbaey (eds), Oxford Handbook on Well-being and Public Policy. Oxford: Oxford University
Press, pp. 24685.
COWELL, F. A. (2011). Measuring Inequality. Oxford: Oxford University Press.
Economica
2017 The London School of Economics and Political Science
238 ECONOMICA [APRIL
and JENKINS, S. P. (1995). How much inequality can we explain? A methodology and an application to
the United States. Economic Journal, 105(429), 42130.
DARDANONI, V. (1996). On multidimensional inequality measurement. In C. Dagum and A. Lemmi (eds),
Research on Economic Inequality, Vol. 6. London: JAI Press, pp. 2015.
DE LA VEGA, C. L., URRUTIA, A. and DE SARACHU, A. (2010). Characterizing multidimensional inequality
measures which full the PigouDalton bundle principle. Social Choice and Welfare, 35(2), 31929.
DECANCQ, K. (2014). Copula-based measurement of dependence between dimensions of well-being. Oxford
Economic Papers, 66(3), 681701.
and LUGO, M. A. (2012). Inequality of wellbeing: a multidimensional approach. Economica, 79(316),
72146.
, FLEURBAEY, M. and SCHOKKAERT, E. (2015a). Happiness, equivalent incomes, and respect for
individual preferences. Economica, 82, 1082106.
, and (2015b). Inequality, income, and well-being. In A. B. Atkinson and F. Bourguignon
(eds), Handbook of Income Distribution, Vol. 2. Oxford: Elsevier, pp. 67140.
DUTTA, I. and FOSTER, J. (2013). Inequality of happiness in the US: 19722010. Review of Income and Wealth,
59(3), 393415.
, PATTANAIK, P. K. and XU, Y. (2003). On measuring deprivation and the standard of living in a
multidimensional framework on the basis of aggregate data. Economica, 70(278), 197221.
FERRER-I-CARBONELL, A. and FRIJTERS, P. (2004). How important is methodology for the estimates of the
determinants of happiness? Economic Journal, 114(497), 64159.
FLEURBAEY, M. and BLANCHET, D. (2013). Beyond GDP: Measuring Welfare and Assessing Sustainability.
Oxford: Oxford University Press.
and MANIQUET, F. (2011). A Theory of Fairness and Social Welfare. Cambridge: Cambridge University
Press.
and SCHOKKAERT, E. (2013). Behavioral welfare economics and redistribution. American Economic
Journal: Microeconomics, 5(3), 180205.
and TRANNOY, A. (2003). The impossibility of a Paretian egalitarian. Social Choice and Welfare, 21(2),
24363.
FRIJTERS, P., GEISHECKER, I., HAISKEN-DENEW, J. P. and SHIELDS, M. A. (2006). Can the large swings in
Russian life satisfaction be explained by ups and downs in real incomes? Scandinavian Journal of Economics,
108(3), 43358.
, HAISKEN-DENEW, J. P. and SHIELDS, M. A. (2004). Investigating the patterns and determinants of life
satisfaction in Germany following reunication. Journal of Human Resources, 39(3), 64974.
GAJDOS, T. and WEYMARK, J. (2005). Multidimensional generalized Gini indices. Economic Theory, 26(3), 471
96.
JONES, A. M. and SCHURER, S. (2011). How does heterogeneity shape the socioeconomic gradient in health
satisfaction? Journal of Applied Econometrics, 26(4), 54979.
KOLM, S.-C. (1977). Multidimensional egalitarianisms. Quarterly Journal of Economics, 91(1), 113.
LAYARD, R., MAYRAZ, G. and NICKELL, S. (2008). The marginal utility of income. Journal of Public Economics,
92(89), 184657.
MAASOUMI, E. (1986). The measurement and decomposition of multidimensional inequality. Econometrica, 54
(4), 9917.
and JEONG, J. H. (1985). The trend and the measurement of world inequality over extended periods of
accounting. Economics Letters, 19(3), 295301.
RAVALLION, M., HIMELEIN, K. and BEEGLE, K. (2016). Can subjective questions on economic welfare be
trusted? Economic Development and Cultural Change, 64(4), 697726.
SCANLON, T. (1975). Preference and urgency. Journal of Philosophy, 72(19), 65569.
SEN, A. K. (1985). Commodities and Capabilities. Amsterdam: North-Holland.
STIGLITZ, J.E.SEN, A. and FITOUSSI, J.-P. (2009). Report by the commission on the measurement of economic
performance and social progress. Mimeo.
TRANNOY, A. (2006). Multidimensional egalitarianism and the dominance approach: a lost paradise. In F.
Farina and E. Savaglio (eds), Inequality and Economic Integration. London: Routledge, pp. 284302.
TSUI, K. Y. (1999). Multidimensional inequality and multidimensional generalized entropy measures: an
axiomatic derivation. Social Choice and Welfare, 16(1), 14557.
WEYMARK, J. A. (2006). The normative approach to the measurement of multidimensional inequality. In F.
Farina and E. Savaglio (eds), Inequality and Economic Integration. London: Routledge.
Economica
2017 The London School of Economics and Political Science