Content uploaded by Saamah Abdallah
Author content
All content in this area was uploaded by Saamah Abdallah on Jan 23, 2017
Content may be subject to copyright.
October 2012
1
Version: WB/V3/20121019
Task 8 – Inequalities and disparities
Brussels, October 2012
Analysis, implementation and
dissemination of well-being indicators
(contract number: 50201.2010.001-2010.580)
October 2012
2
Table of contents
TABLE OF CONTENTS 2
INTRODUCTION 5
1. MEASURES OF INEQUALITY 7
1.1 Introduction ___________________________________________________________ 7
1.2 Inequality beyond income __________________________________________________ 7
1.3 Types of inequality metric ________________________________________________ 8
1.3.1. Comprehensiveness vs. non-comprehensiveness .................................................. 8
1.3.2. Ratio vs. non-ratio ................................................................................................... 9
1.4 Possible inequality metrics ______________________________________________ 10
1.4.1. Standard deviation (comprehensive, non-ratio) ................................................... 10
1.4.2. Average deviation from the mean (comprehensive, non-ratio) ........................... 10
1.4.3. Mean pair distance (comprehensive, non-ratio) .................................................. 11
1.4.4. Coefficient of variation (comprehensive, ratio) .................................................... 11
1.4.5. GINI coefficient (comprehensive, ratio) ................................................................ 11
1.4.6. Atkinson Index (comprehensive, ratio) ................................................................. 13
1.4.7. Theil’s measure of entropy (comprehensive, ratio) .............................................. 14
1.4.8. Percentile ratios (non-comprehensive, ratio) ....................................................... 14
1.4.9 Quintile share ratios (non-comprehensive, ratio) .................................................. 14
1.4.10 Percentile ranges and quintile share differences (non-comprehensive, non-
ratio) ...................................................................................................................... 15
1.4.11 Bounded scales: a further issue ........................................................................... 15
1.5 Summary ______________________________________________________________ 16
October 2012
3
2. UNIVARIATE INEQUALITY 18
2.1 Choosing a data set - European Health Interview Survey ______________________ 18
2.1.1 Selection process .................................................................................................... 18
2.1.2 Countries and sample sizes .................................................................................... 18
2.1.3 Demographic variables ........................................................................................... 19
2.1.4 Subjective well-being variable ............................................................................... 20
2.2 Univariate inequality _____________________________________________________ 21
2.3 Relationship with income inequality ________________________________________ 30
2.4 Inequality adjustments ___________________________________________________ 31
2.4.1 Options ................................................................................................................... 32
2.4.2 Results .................................................................................................................... 32
2.5 Discussion ______________________________________________________________ 35
3. BIVARIATE AND MULTIVARIATE ANALYSES 36
3.1 Comparisons with Task 7 results __________________________________________ 36
3.2 Bivariate analyses _____________________________________________________ 37
3.2.1 Gender .................................................................................................................. 37
3.2.2 Age ........................................................................................................................ 38
3.2.3 Marital status ........................................................................................................ 39
3.2.4 Employment status ............................................................................................... 40
3.2.5 Education level ...................................................................................................... 42
3.2.6 Income .................................................................................................................. 43
3.2.7 Place of birth / migrant status .............................................................................. 46
3.2.8 Household type ..................................................................................................... 47
3.2.9 Urbanisation level ................................................................................................. 47
3.2.10 Body-mass index ................................................................................................... 48
3.2.11 Support from friends and families ........................................................................ 48
3.3 Multivariate analyses __________________________________________________ 50
October 2012
4
3.3.1 Beta coefficients ................................................................................................... 51
3.3.2 Overall fit............................................................................................................... 53
4. CONCLUSIONS AND RECOMMENDATIONS 55
4.1 Key findings __________________________________________________________ 55
4.2 Recommendations _____________________________________________________ 56
4.3 Limitations and future work _____________________________________________ 58
ANNEX 1- TABLE OF POTENTIAL DATA SETS FOR ANALYSIS 60
ANNEX 2 - SF-36 61
October 2012
5
Introduction
This report is the draft final output of Task 8 of the “Analysis, implementation and
dissemination of well-being indicators” study being conducted on behalf of Eurostat. The
overall purpose of this study is to revise and develop the list of indicators identified in the
Feasibility Study that we conducted between 2008 and 2010 (hereafter referred to as FS),
incorporating micro-level analyses and taking into consideration recent developments in the
field of well-being measurement. This Implementation Study will conclude in the
dissemination of a set of well-being indicators on the Eurostat website.
Task 8 examines distributional issues in relation to well-being. Much has been said recently
regarding the importance of moving beyond averages to look at distributions of outcomes
within populations. For example, the Sponsorship Group on Measuring Progress, Well-Being
and Sustainable Development final report included a section on “Providing information on
the distribution of income, consumption and wealth” (Section 3.1.2, page 18).1 The French
Commission on the Measurement of Economic Performance and Social Progress final report
also looked at how distributions of income, consumption and wealth can be better captured
(Section 4.2, page 32).2 In the UK, the Office for National Statistics National Well-being
Framework includes equality as a cross-cutting issue across all domains of well-being.3
Meanwhile, the United Nations’ 2010 and 2011 Human Development Reports have included
an innovative Inequality-adjusted Human Development Index (IHDI).4
In most cases (although there are exceptions), inequality measurement focuses on inequality
in income. This is of course very policy-relevant in that policies can be directly aimed at
reducing income inequality, and is very relevant to well-being given that there is a strong
relationship between individual income (absolute and relative) and subjective well-being.
High income inequality itself has also been found to be associated with lower average life
satisfaction (see the output of Task 6 and the output of Task 7).
In this study, however, we look at inequalities and disparities in subjective well-being itself.
Can we produce comparable measures of inequality of subjective well-being within
countries? What factors (including income or demographic ones) appear to most strongly
drive inequality in subjective well-being within countries?
If we take well-being to be an important outcome of policy, then measuring inequality in
well-being is necessary. Two countries with the same average reported well-being might
have quite different distributions – with implications for the lives of people within those
countries. Politically, there may well be debates about whether it is more important to
1 European Statistical System Committee (2011) Sponsorship Group on Measuring Progress, Well-being and Sustainable
Development: Final Report. Available from
http://epp.eurostat.ec.europa.eu/portal/page/portal/pgp_ess/about_ess/measuring_progress
2 Stiglitz J, Sen A & Fitoussi J-P (2009) Report by the Commission on the Measurement of Economic Performance and Social
Progress. Available from http://www.stiglitzsen-fitoussi.fr/documents/rapport_anglais.pdf
3 Beaumont, J. (2011) Measuring National Well-being – Discussion paper on domains and measures. Newport: Office for
National Statistics.
4 UNDP (2011) Human Development Report 2011. Available at http://hdr.undp.org/en/reports/global/hdr2011/
October 2012
6
maximise average well-being or minimise well-being inequality, just as there are political
debates about the importance of reducing income inequality. But at the moment, such
debates cannot take place as data on well-being inequality are not readily available.
This report presents analyses on inequality and disparities in subjective well-being based on
data from the first wave of the European Health Interview Survey (EHIS), conducted
between 2006 and 2010.5
This report is divided into 4 sections.
In Section 1 we introduce the different approaches to measuring overall inequality. We
consider the advantages and disadvantages of each approach, particularly in the context of
measuring inequality in subjective well-being (whereas most measures have typically been
applied to inequality in income).
In Section 2, after introducing the EHIS data set, we present the different measures of
inequality for the 15 countries for which data were available in EHIS. The inequality
measures presented are univariate inequality measures – i.e. inequality based on data on
only one variable (subjective well-being). Discussion focuses on whether and how the
different measures tell different stories. We also explore the potential to produce an
‘inequality-adjusted’ mean subjective well-being score.
In Section 3, we try to determine what demographic and other factors might be responsible
for the inequalities in well-being. This section is based on both bivariate analysis of the
relationships between different independent variables and subjective well-being
individually, as well as multivariate regressions bringing together independent variables. The
analysis has been conducted on the same EHIS data set, though certain variables are not
available for all 15 countries for which subjective well-being data is available. The aims are
both to identify which factors account for differences in subjective well-being in all
countries, and look at different patterns from country-to-country.
The final section draws conclusions from this Task, including recommendations for how to
apply the findings to the final well-being indicator set.
A Technical Annex has also been produced providing details of the precise methodologies
used for carrying out the analyses.
5 Section 2.1.2 presents information on the EHIS.
October 2012
7
1. Measures of inequality
1.1 Introduction
This section will consider measures of overall inequality, i.e. not inequalities between
specific population groups, but the overall level of inequality in a given variable, e.g. income
or well-being. This can be called univariate inequality, as analysis only involves a single
variable (the quantity whose inequality is being assessed) rather than requiring analysis of
any independent variables alongside it.
According to the Oxford Handbook of Economic Inequality, such metrics of inequality should
meet the following four criteria:
- Anonymity (i.e. it does not matter who has more or less – this excludes indicators that
compare specific population groups, e.g. males vs. females)
- Scale invariance (statistically independent of mean and a measure without dimension)
- Replication invariance (Statistically independent of size of sample)
- Respect the Transfer principle: that if some of the quantity whose inequality is being
measured is transferred from a ‘rich’ person to a ‘poor’ person, while still preserving the
order of income ranks, then the measured inequality should decrease.
- Statistically robust (incl. methods, sample sizes and data)
Furthermore, for their use in official statistics, we would add the following criteria against
which measures should be evaluated:
- Easy to communicate
- Comprehensive (i.e. sensitive to the most politically relevant changes in distribution)
Inequality metrics are most commonly calculated for income distributions. In this context,
the best-known inequality metrics are the GINI coefficient, Atkinson Index, and various share
or percentile ratios. However, well-being is not income, and cannot be treated in the same
way. Therefore, in this study we will need to take a step back from these better known
metrics and explore a broader range. Section 1.2 of this Chapter will discuss the reasons why
well-being may not be best treated like income and what this implies for choosing a suitable
well-being inequality metric. In Section 1.3 we will review a range of inequality metrics and
consider how they meet the requirements set out. In Section 1.4 we will conclude with
recommendations on the most suitable well-being inequality metrics, considering statistical
robustness, comprehensiveness and ease of communication.
1.2 Inequality beyond income
The analysis of inequalities beyond income or wealth is relatively uncommon. The most well-
studied area is probably health inequality, though this tends to be looked at in terms of
differences between specific population groups, rather than an overall assessment of
inequality such as that calculated in a GINI coefficient. One paper that considers how such an
assessment might be done presents a series of thought experiments to help untangle what
kinds of distribution of life expectancy might be considered more or less equal.6 A key
6 Gakidou E, Murray C & Frenk J (2000) ‘Defining and measuring health inequality: an approach based on the distribution of
health expectancy’ Bulletin of the World Health Organization 78:42-54.
October 2012
8
message from this paper, published by the World Health Organisation, is that one needs to
consider the actual thing that is distributed – life expectancy is not the same as income.
There may be normative implications which are specific to the variable in question. The
authors propose using an inequality measure like the GINI index, but calibrating it based on
survey data reflecting people’s preferences in relation to inequality.
The Atkinson Index has also been used to measure health inequality, in the UN Human
Development Reports (see 2010 and 2011 editions). As we will discuss in Section 1.3, this
may be somewhat conceptually flawed, as the Atkinson Index is based on a notion of
diminishing returns which may or may not be applicable to life expectancy. The Human
Development Report also uses the Atkinson Index methodology to calculate an inequality-
adjusted measure of educational attainment.
Early studies looking at inequality within well-being tended to use standard deviation as the
inequality metric.7 8 To economists used to using GINI coefficients, this was apparently seen
as a shortcoming. To address these concerns, the leading researchers in the field of well-
being inequality (Ruut Veenhoven and colleagues) produced a paper in 2005 arguing the
case for the use of the standard deviation for measuring well-being inequality.9 This
recommendation appears not to have been contradicted in the literature until 2011, when
Jan Delhey and Ulrich Kohler published a paper arguing for an instrument-effect-correction
to be made to the standard deviation.10
The following section (1.3) will consider the two key distinctions to bear in mind when
considering inequality metrics. Section 1.4 will then go through some options, explaining
what they are, and what their strengths and weaknesses are.
1.3 Types of inequality metric
There are two key distinctions to be made in the categorisation of inequality metrics:
comprehensive vs. non-comprehensive, and ratio vs. non-ratio.
1.3.1. Comprehensiveness vs. non-comprehensiveness
The calculation of some metrics incorporates every single response. Such metrics include the
standard deviation and the GINI coefficient, but also the Atkinson Index and the co-efficient
of variation. Other metrics sacrifice comprehensiveness for communicability and ease of
calculation. For example, the interquartile range is the range between the individual who is
25% along the distribution (the 25th percentile) and the individual who is 75% along the
distribution (the 75th percentile). In effect, changing the values for any other individual
within the sample will not change this interquartile range, provided they do not cross the
boundaries of the 25th percentile and the 75th percentile. Another example of a non-
7 Veenhoven R (1990) ‘Inequality in Happiness, Inequality in Countries Compared between Countries’, Paper in the 12th Work
Congress of Sociology, Madrid, Spain
8 Veenhoven R & Ehrhardt J (1995) ‘The Cross-national Pattern of Happiness; Test of Predictions Implied in Three Theories of
Happiness’ Social Indicators Research 34:33 – 68.
9 Kalmijn W & Veenhoven R (2005) ‘Measuring inequality of happiness in nations’ Journal of Happiness Studies 6:357-396.
10 Delhey J & Kohler U (2011) ‘Is happiness inequality immune to income inequality? New evidence through instrument-effect-
corrected standard deviations’ Social Science Research 40:742-756.
October 2012
9
comprehensive measure is the income quintile share ratio used by Eurostat (S80/S20), which
is the ratio between the mean (or total) income of the richest quintile of the income
distribution, and the mean (or total) income of the poorest quintile of the income
distribution. Any exchange of income within those quintiles, or indeed amongst anyone in
the middle three quintiles, will not change the value of this measure. On the other hand, it is
meaningful and easy to communicate.
1.3.2. Ratio vs. non-ratio
This distinction is particularly important to understand in the context of moving from
inequality measures of income to inequality measures of well-being. The easiest way to
compare a ratio and a non-ratio metric of inequality is to consider the coefficient of variation
and the standard deviation. The former is calculated by dividing the latter by the mean for a
sample. Standard deviation is therefore a non-ratio metric – which has its own units, whilst
the coefficient of variation is a dimensionless ratio.
The drawback of using a metric like the standard deviation or indeed the interquartile range
for measuring inequality in something like income, is that they are not scale invariant and
are not dimensionless. For example, as Kalmijn and Veenhoven put it, when the Netherlands
changed its currency to the Euro in 2002, the standard deviation of income would have
decreased by a factor of 2.2, but of course income inequality had not fallen.11
A second problem is the following: in a country where the mean income is $3000 a year, a
standard deviation of $1000 is a lot – but in a country where the mean income is $30,000 a
year, it’s not very much. There are two reasons for this – one to do with the nature of
measurement and the other related to what is being measured. The measurement issue is
that income is measured at the ratio level. $10,000 is twice $5,000. Compare this to for
example temperature as measured in Celsius 30° is not double 15°, or indeed life satisfaction
there is no reason to believe that a score of 8 out of 10 is double 4 out of 10. The other
reason that one prefers a measure of income inequality to be scale invariant is that we have
the notion that income provides diminishing returns – an increase in income of $1,000 is
substantial for someone on $3,000 a year, but it would make little difference to someone
earning $300,000 a year.
For these reasons, income inequality measures must always be scale invariant and
dimensionless – the GINI is such a measure, as is the coefficient of variation, and the
S80/S20 – where the income of the top quintile is divided by the income of the bottom
quintile. Were the difference calculated through subtraction (i.e. the top quintile minus the
bottom quintile), the metric would not be scale invariant, and would not be suitable for
measuring inequality. In the case of the coefficient of variation, division by the mean in
effect ‘standardises’ the standard deviation so that it is expressed as a proportion of the
mean. The assumption is that, as a result, it is structurally independent of the mean.12
11 Kalmijn & Veenhoven (2005) op cit.
12 Of course, there may be intrinsic dependencies, for example because of socio-economic processes, but these we hope to
emerge separately.
October 2012
10
Subjective well-being on the other hand, is dimensionless already. More importantly, it is not
measured at the ratio level, but at the ordinal level (or interval level at best). It is not clear
that doubling one’s life satisfaction from 2 to 4 has anything in common with doubling from
4 to 8. Nor is there any reason to believe that a standard deviation of 1 when a sample has
mean well-being of 4 is akin to a standard deviation of 2 when the sample has a mean well-
being of 8.
Now, if dividing by the mean serves to produce an inequality metric that is independent of
the mean in the case of a ratio variable like income, it is likely to do the opposite for a
dimensionless ordinal variable like life satisfaction. So, dividing a standard deviation of 1 by a
sample mean to produce a coefficient of variation would lead to penalising countries with
lower means.
Having said that, whilst we cannot be sure that subjective well-being is not measured on a
ratio scale, nor can we be sure that it isn’t, nor indeed that it is acceptable to treat is an
interval scale (though other papers have demonstrated that treating life satisfaction as
interval data is acceptable13). In other words, there are no strong logical grounds for arguing
that it is better to think of changes in well-being as absolute values (i.e. an increase of 1 or 2
points), rather than as proportional values (i.e. an increase of 5% or 10%). As such, for the
time being, we believe it valuable to consider both ratio indicators and non-ratio indicators.
1.4 Possible inequality metrics
In this section we will consider some of the possible inequality metrics we could use in
relation to well-being, both comprehensive and non-comprehensive, ratio and non-ratio.
1.4.1. Standard deviation (comprehensive, non-ratio)
The standard deviation is perhaps the most widely used measure of dispersion, through
rarely reported as a measure of inequality. Conceptually, it can be understood as an
assessment of the average distance between any given respondent and the mean. However,
calculation of the standard deviation involves taking the square of each distance, and then
square rooting the mean of these distances (amongst other reasons, this is simply the most
computationally efficient way to remove the valency of negative differences when
comparing respondents that are below the mean). The result is that standard deviations are
affected more by extreme values, though this may be less of an issue for well-being where
response scales are bounded.
1.4.2. Average deviation from the mean (comprehensive, non-ratio)
Theoretically, one can also calculate the average absolute distance from the mean of each
response without taking squares or square rooting. In practice this is less used for reasons of
computational ease, but also because this average deviation does not have a simple
relationship to normal distribution parameters.
13 Ferrer-i-Carbonell A & Frijters P (2004) ‘How Important is Methodology for the estimates of the determinants of Happiness?’
Economic Journal 114:641-659.
October 2012
11
1.4.3. Mean pair distance (comprehensive, non-ratio)
The mean pair distance is the average distance between any two randomly selected
respondents. This intuitively appealing metric is rarely used, perhaps because of the
difficulties in its computation. However, it is actually a close relative of the GINI coefficient,
effectively being its non-ratio equivalent. For large samples, to derive the GINI from the
mean pair distance, one must simply divide by 2 times the mean for the sample.14
1.4.4. Coefficient of variation (comprehensive, ratio)
The coefficient of variation is the standard deviation of a variable, divided by its mean. Its
simplicity is a key advantage, as well as the fact it utilises all the data in a distribution. But
the coefficient of variation has rarely been used for income inequality. Several reasons for
this have been given.
Firstly, unlike most other ratio variables and owing to the square function in the standard
deviation, the figures have no clear maximum (though in the case of well-being data, given
the bounded scale, this is likely to be an issue).
Secondly, some have highlighted the issue that the coefficient of variation is affected by
changes at the top and bottom of a distribution equally, whereas there is a sense that
changes at the bottom should have a larger weighting.15 It is not clear whether this should
necessarily be seen as a problem for a measure of well-being inequality.
Lastly, there has been no clear criterion for determining whether it ranks distributions
correctly.
1.4.5. GINI coefficient (comprehensive, ratio)
Perhaps the best known measure of inequality is the GINI coefficient. The GINI coefficient is
best understood in relation to a Lorenz curve (Figure 1).
Figure 1: Lorenz Curve
14 One should also divide by N/(N-1) – but, with large Ns, this approximates to 1.
15 Foster J (1985) ‘Inequality measurement’ Fair Allocation 33:31-69.
October 2012
12
The Lorenz Curve plots the cumulative income of a population, starting with the lowest
income households first. At first the curve rises slowly, but as higher income households are
added to the cumulative income, the gradient increases.
The curve can be compared against the ‘line of equality’, which is the diagonal. This would
be the cumulative income function were every household to have the same income.
Similarly, one can hypothesise that, if there was perfect inequality within a society (i.e. 1
household received all the income), then the ‘curve’ would lie on the x-axis up until the final
household at which point it would shoot up.
In reality, of course, all distributions lie somewhere in between these two extremes. The
GINI coefficient assesses the income distribution by calculating what proportion of the area
under the line of equality, lies between that line and the Lorenz Curve. If the Lorenz Curve of
a population lies on the line of equality it would be 0. If it were just below the line, it would
be a low number. If the Lorenz Curve approximates the most unequal distribution, the
number approaches 1.
The GINI coefficient related to income is the most widely used inequality measure, included
in the World Bank World Databank,16 the CIA World Factbook,17 the United Nations
International Human Development Indicators,18 and Eurostat’s data sets.
Whilst it is mathematically very difficult to calculate, it is able to visualise what it means
quite easily, and the numbers have an intuitive meaning. On the other hamd whilst many
people may have heard of GINI coefficients, it is likely that most do not actually understand
what they represent. Furthermore, it is calculated by dividing a dispersion metric (in this
case half the mean pair distance) by the mean, which automatically incorporates the
assumption that it is more meaningful to talk about percentage changes to well-being rather
16 http://data.worldbank.org/data-catalog/world-development-indicators
17 https://www.cia.gov/library/publications/the-world-factbook/
18 http://hdrstats.undp.org/en/indicators/default.html
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0% 20% 40% 60% 80% 100%
% of income
% of households
Line of equality
Lorenz curve
October 2012
13
than absolute changes. Nevertheless, at least one study has calculated the GINI coefficient of
well-being.19
1.4.6. Atkinson Index (comprehensive, ratio)
The Atkinson Index is an example of an inequality index based on the social welfare model.20
In essence the aim of the social welfare model is to derive total utility associated with a
particular income distribution, with the recognition that the relationship between income
and utility is non-linear and results in diminishing returns.
The Atkinson Index methodology assumes that the relationship can be described by an nth
root equation, where n is specified using the parameter ε (n=1- ε). If ε is set as 0.5, then the
relationship between income and utility is treated as a square root. As such differences in
income at the higher end of the distribution have less positive impact on utility and
therefore the total utility is calculated as falling short of the total income. The difference
between the two allows one to calculate the Atkinson Index – which is the deficit in utility
resulting from inequality.
If ε is set as 0, then the relationship between income and utility is treated as linear, and
inequality has no effect on the total utility – Atkinson Indices are not calculated. If it is set as
1, then the utility function becomes the geometric mean of incomes, and the Atkinson Index
can be calculated from the difference between the arithmetic mean and the geometric
mean. One study identified an ε of 0.8 to be the closest to people’s preferences in the UK –
we will be using this figure in Section 2.21
Like the GINI coefficient, the Atkinson Index ranges from 0 to 1, where 0 indicates perfect
equality, and 1 total inequality.
The Atkinson Index is a conceptually solid approach to calculating the utility lost from an
unequal distribution of income.22 However, this conceptual framework loses its meaning
somewhat when looking at the inequality of well-being. What are the diminishing returns
from increases in well-being higher up the spectrum? Also, its appeal has tended to be
restricted to the research community – beyond that the GINI coefficient has dominated the
field. Furthermore, as with the GINI coefficient, it is a ratio metric, and inappropriate if we
think that changes in well-being should be understood as percentage changes rather than
absolute values.
19 Easterlin R (2009) ‘Lost in Transition: life satisfaction on the road to capitalism’ Journal of Economic Behavior & Organization
71:130-145.
20 Jackson T, Marks N, Ralls J / Stymne S (1997) Sustainable economic welfare in the UK 1950-1996 (London: nef)
21 ibid.
22 Although most recent research on the relationship between income and well-being suggests that a logarithmic relationship is
the most appropriate approach (see Task 6 Annex, Section 1.1).
October 2012
14
1.4.7. Theil’s measure of entropy (comprehensive, ratio)
Theil’s measure of entropy applies the concepts of entropy to information theory.23 It is
rarely used, and given our concerns regarding ratio measures, we shall not discuss it further
in this study.
1.4.8. Percentile ratios (non-comprehensive, ratio)
Percentile ratios are calculated by taking the incomes (or levels of well-being) of individuals
who stands at two specific points in the distribution and dividing one by the other.
With respect to income, often cited percentile ratios include:
80/20 (i.e. the income of the person at the 80th percentile versus the income of the
person at the 20th percentile).
90/10
90/median (i.e. the income of the person at the 90th percentile versus the median
income)
The advantage of such ratios is that they are very easy to communicate – saying that the
richest 10% have an income of (for example) 8 times the poorest 10% makes intuitive sense
to people. They also avoid being influenced by real extremes – the 0.1% richest, for example,
will not influence the 80/20, 90/10 or even 99/1 ratio. The disadvantage is that they are
insensitive to the rest of the distribution – large changes in equality levels might go
unnoticed if they are taking place anyway other than the 2 data points that are chosen.
1.4.9 Quintile share ratios (non-comprehensive, ratio)
Slightly different from percentile ratios are quintile share ratios. Rather than use two
percentile boundaries to calculate a ratio (e.g. the 80th vs. the 20th percentile), in this
methodology, the means for two given groups (e.g. the bottom quintile vs. the top quintile)
are compared.
There are three reasons why a quintile share ratio may be preferable to a percentile ratio.
Firstly, it is arguably easier to communicate. Secondly, it does not rely on just two data
points, making use of (in the case of comparing two quintiles) 40% of the data. Thirdly,
Eurostat already publishes the S80/S20 for income, which is the quintile share ratio. The one
disadvantage of such ratios – that they are more sensitive to extreme values, this is not an
issue for well-being given its bounded scale.
23 Hoeller P et al (2012) ‘Less Income Inequality and More Growth – Are They Compatible? Part 1. Mapping Income Inequality
Across the OECD’, OECD Economics Department Working Papers, No. 924, OECD Publishing.
http://dx.doi.org/10.1787/5k9h297wxbnr-en , pg.32
October 2012
15
1.4.10 Percentile ranges and quintile share differences (non-comprehensive, non-ratio)
As with the comprehensive measures, non-ratio versions of the non-comprehensive metrics
can also be calculated. The simplest approach is to simply subtract one value from the other,
rather than dividing. So, if the mean life satisfaction for the top quintile of a sample is 8 and
for the bottom quintile is 4, then the percentile ratio would be 2 (8/4), but the percentile
range would be 4 (8-4). Contrast this to a distribution where the bottom quintile had a mean
of 2 and a top one had a mean of 6. In this case the percentile ratio would be higher – 3 –
but the percentile range would be the same.
1.4.11 Bounded scales: a further issue
Delhey and Kohler throw up another issue that needs to be considered in calculating
inequality metrics for well-being – that of bounded scales.24 Consider two groups, one with a
mean life satisfaction (on a scale of 0-10) of 8 and one with a mean of 5. For the former
group, the largest possible standard deviation would be 4, for the latter it would be 5.25 As
such, if these two groups both had standard deviations of 4, one would be at its theoretical
maximum in terms of inequality (the one with the mean of 8), the other would not be (the
one with the mean of 5). Intuitively, one feels that the latter group is less unequal than the
former (after all, inequality could yet increase in it), but the standard deviation does not
capture that.
As such the authors propose calculating an instrument-effect-corrected standard deviation.
They propose two methods. As the results of the two methods are almost identical, we only
consider the conceptually simple one here. It involves dividing the standard deviation of a
sample by the maximum possible standard deviation given its mean (see Equation 1). The
result is a ratio, but not as in the sense we have already critiqued – one is not dividing by the
mean.
Equation 1:
Where U is the upper bound of the response scale (i.e. 10 for life satisfaction), L is the lower bound,
and N is the sample size.
The authors argue that the previously found correlation between well-being equality and
well-being mean (i.e. countries with higher well-being means have more equal
distributions)26 27 28 can thus be seen as an artefact of the instruments used. When the above
24 Delhey & Kohler (2011) op cit.
25 For small N’s, if one uses the sample standard deviation (i.e. divides by N-1) as opposed to the standard deviation of the
sample (dividing by N), numbers can be higher than those presented here.
26 Cummins R (2003) ‘Normative life satisfaction: measurement issues and a homeostatic model’ Social Indicators Research
64:225-256.
27 Fahey T & Smith E (2004) ‘Do subjective indicators measure welfare? Evidence from 33 European societies’ European
Societies 6:5-27.
28 Ott J (2005) ‘Level and inequality of happiness in nations: does a greater happiness of a greater number imply greater
inequality of happiness?’ Journal of Happiness Studies 6:397-420.
October 2012
16
correction is made, this correlation drops out. Furthermore, the previously elusive
correlation between income inequality and well-being inequality emerges, with an R of 0.41
(significant at the p < 0.01 level). Countries with high income inequality but high well-being
such as Colombia and Mexico see the biggest effect of this adjustment.
Such an instrument-effect adjustment can also be made for other metrics, such as the mean
pair distance and the GINI coefficient. However, it is not without controversy. Both Kalmijn
and Veenhoven have critiqued the method saying that it is ‘correcting’ for a genuine
phenomenon.29 30 They argue that the fact that countries that have high average well-being
tend to have little well-being inequality is not something to be ‘corrected’ for, but a genuine
artefact of the bounded nature of well-being (not just the well-being scale). One of Delhey &
Kohler’s approaches to adjusting for instrument effects (not the one reviewed above), for
example, assumes that life satisfaction has a normal distribution and that the upper bound
of 10 on a life satisfaction scale is somehow artificial – that there are people who are above
this, but respond with a 10 as that is the highest score available. But what if this is not the
case? What if satisfaction itself is conceptually bounded (in the same sense that
temperature is conceptually bounded by a lower bound of -273°C, i.e. 0 Kelvin)? What if 10
does genuinely represent the most satisfied an individual can be? If that is so, it makes no
sense to ‘correct’ for instrument effects.
1.5 Summary
It should be clear from sections 1.3 and 1.4 that the measurement of well-being inequality is
still a nascent field, with no consensus as to the best approach as yet. It does seem that the
two sets of authors whose opinions were discussed in section 1.4.11 are in consensus as to
the preference for a non-ratio metric, as opposed to a metric measure such as the GINI
coefficient, but others (e.g. Richard Easterlin) have used the GINI. It is not certain that the
assumptions one has to make in comparing inequality values based on non-ratio metrics are
any better than those one has to make in comparing values based on ratio metrics.
Two things are clear. Firstly, well-being is not income. Perhaps the clearest distinction
highlighted by Kalmijn and Veenhoven is that income is a ‘capacity’ variable whereas well-
being is an ‘intensity’ variable. The total income of a group of people is the sum of all their
incomes. It does not make sense to say that their total well-being is the sum of all their well-
beings. Given this, and the fact that it is definitely not measured on a ratio scale, and
technically not measured on an interval scale, one must be aware of the assumptions one
makes when calculating anything based on the data – not just inequality metrics, but also
means, regression coefficients and other parametric variables.
Secondly, the key to resolving some of the debates being had is to understand what people
are thinking when they respond to subjective well-being questions. Is Delhey right that there
are people who report a 10 out of 10 on life satisfaction but actually feel more satisfied than
29 Veenhoven R (2012) ‘The medicine is worse than the disease: Comment on Delhey and Kohler’s proposal to measure
inequality in happiness using ‘instrument-effect-corrected’ standard deviations’ Social Science Research 41:203-205.
30 Kalmijn W (2012) ‘Happiness is not normally distributed: A comment on Delhey and Kohler’ Social Science Research 41:119-
202.
October 2012
17
that? Does a 1-point difference in life satisfaction represent the same, less or more
difference depending on where in the scale it occurs?
Related to this is the point of understanding what people perceive to be well-being
inequality. For example, do we consider that the inequality in a population group with a
standard deviation of 1 and a mean of 3 is greater than that in a population group with the
same standard deviation and a mean of 7? Unfortunately, this is not simply a matter of
identifying which distribution is normatively preferable – of course the second one (with a
higher mean) is preferable. But should that be reflected in our inequality metrics? Kalmijn
and Veenhoven highlight the distinction between inequality and inequity, suggesting that
people often conflate the two.
In section 2, we will calculate a range of inequality metrics, including both comprehensive
and non-comprehensive, ratio and non-ratio, instrument-effect-corrected and non-
instrument-effect-corrected. Ultimately, we are looking to identify one best comprehensive
measure and one best non-comprehensive measure. Beyond that, the results will help us
draw conclusions as to what to recommend. If all the different metrics intercorrelate highly,
then some of the debates had in this section will be less important.
October 2012
18
2. Univariate inequality
2.1 Choosing a data set - European Health Interview Survey
2.1.1 Selection process
We considered five possible data sets, based on four criteria:
Countries – The data set used should compare several countries, preferably all of
Europe
Sample size – inequality analysis requires large samples.
Demographic data – data set should include basic economic and demographic variables
including gender, age, income, education and ethnicity. Rural/urban split would also be
desirable
Subjective well-being data – data set should include life satisfaction or some other
measure of subjective well-being
Annex 1 lists the five data sets considered and summarises our evaluations of how well they
meet these four criteria.
The first wave of the European Health Interview Survey (EHIS) was chosen by Eurostat and
the contractors as it was the only large cross-national survey that included any measure of
subjective well-being.
Sections 2.1.2 to 2.1.4 outline the key features of the Survey.
2.1.2 Countries and sample sizes
The European Health Interview Survey is a comprehensive and coordinated set of surveys
performed within the European Statistical System and under the responsibility of Eurostat.
The responsibility for data collection rests at the EU-member level, but for the first wave
(which we are using) there was no legal requirement for member states to implement the
survey. 19 countries committed to administering the survey between 2006 and 2010, with
absent countries tending to be from Western or Northern Europe (notable absences include
the UK, Sweden and Italy).
The micro-data available to us did not include Germany or Switzerland. Germany had
submitted to Eurostat aggregated data, whilst Switzerland did not carry out the survey. The
complete set of subjective well-being questions was not included in the surveys
administered by Belgium and Estonia. As such, we had relevant data for 15 countries, listed
alphabetically in Table 1.
We have been advised by the EHIS team in Eurostat to treat the data from France with
caution as it may not be a fully representative sample.
October 2012
19
The total sample size for respondents with most of the data on subjective well-being was
over 150,000. However sample sizes varied considerably between countries ranging from
1,955 in the Czech Republic to 35,100 in Poland. Table 1 lists the countries for which we
have subjective well-being data and the number of respondents in each country for which
we were able to calculate a measure of subjective well-being.
Table 1: Countries and numbers of respondents
Country
N
Austria
15250
Bulgaria
5135
Cyprus
6196
Czech Republic
1928
Spain
21451
France
14520
Greece
6113
Hungary
5047
Latvia
6275
Malta
3520
Poland
30329
Romania
15213
Slovenia
2116
Slovakia
4940
Turkey
14634
2.1.3 Demographic and economic variables
The following variables were available to us to be treated as independent variables:
Gender
Age (continuous variable)
Income (10 bands in each country)
Education (based on ISCED)
Urbanisation level (rural, semi-urban, or urban)
Place of birth (native, other EU country, other non-EU country)
Marital status (married, divorced, single, widowed)
October 2012
20
Labour status (employed, unemployed, education, retired, permanently disabled,
military or community service, domestic tasks, other)
Household type (living alone, lone parent, couples with or without children)
Household size
Number of people whom one can count on when faced with serious problems
As there was some flexibility as to whether the items in the survey were brought together in
a single national survey, or included in a range of surveys within that country, some variables
are missing for some countries (e.g. household type and urbanization level).
2.1.4 Subjective well-being variable
Our primary subjective well-being variable during these analyses was an aggregate score
measuring mental health and vitality from a subset of nine questions from the SF-36 Health
Survey (‘SF’ stands for ‘short form’, see Annex 2 for more information on the origins of the
SF-36).31 32 The nine questions are asked as follows:
The next questions are about how you feel and how things have been with you during the
past 4 weeks. For each question, please give the answer that come closest to the way you
have been feeling:
How much of the time, during the past 4 weeks…
- Did you feel full of life?
- Have you been very nervous?
- Have you felt so down in the dumps that nothing could cheer you up?
- Have you felt calm and peaceful?
- Did you have a lot of energy?
- Have you felt down-hearted and depressed?
- Did you feel worn out?
- Have you been happy?
- Did you feel tired?
The 5-point response scale ranges from ‘all of the time’ to ‘none of the time’. Scale scores
(which we call SF scores) are easy to calculate, by simply summing the scores for each item,
taking care to reverse the scoring for positively worded questions. The total SF scores range
from 1 (lowest well-being) to 37 (highest well-being).
31 Ware JE, Snow KK, Kosinski M & Gandek B (1993) SF-36® Health Survey Manual and Interpretation Guide (Boston, MA: New
England Medical Center, The Health Institute).
32 Ware JE, Kosinski M &Dewey JE (2000) How to Score Version Two of the SF-36 Health Survey (Lincoln, RI: QualityMetric,
Incorporated).
October 2012
21
We followed the protocol recommended by the EHIS team in Eurostat which was to
calculate an overall score for all respondents for who answered more than half of the 9
questions (i.e. 5 or more), filling in missing data by taking an average of their responses to
other questions (syntax available in Technical Annex).
The SF score can be interpreted as a measure of hedonic well-being, incorporating both
positive and negative affect (see e-Frame report, as well as OECD Handbook for advantages
and disadvantages of using hedonic well-being) 33 34.
Rather than being understood as a proxy for life satisfaction, it should be seen as more
closely related to the measure of vitality in the set of indicators recommended by the
consortium (feeling active and vigorous, taken from the European Quality of Life Survey -
EQLS)35. Country means for that question, from the EQLS, correlate with the country means
for SF at R=0.61.36 SF scores correlate with the complete WHO Mental Health Index (also
from the EQLS) at R=0.59, but life satisfaction (EQLS) at only R=0.39.
The correlation with the WHO Mental Health Index is respectable given that the data
predominantly comes from different years, and that the WHO Mental Health Index only
includes positive affect items whereas the SF includes both positive and negative affect
items.
2.2 Univariate inequality
Table 2 & 3 present mean SF scores and several measures of inequality in SF scores for the
15 countries. Countries are ranked according to SF means in descending order. Table 2
presents the comprehensive inequality metrics, whilst Table 3 presents the non-
comprehensive metrics. Table 4 presents country ranks on a selection of these.
Table 2: Mean and selected comprehensive inequality metrics
Mean
SD
Mean Pair
Distance
GINI
SD adj
MPD adj
GINI adj
Atkinson
index
Cyprus
28.0
6.1
6.4
0.11
0.39
0.47
0.61
0.027
Austria
27.6
5.8
6.2
0.11
0.37
0.45
0.58
0.023
Spain
27.5
6.7
7.3
0.13
0.42
0.52
0.68
0.031
Greece
27.3
7.3
8.0
0.15
0.46
0.56
0.73
0.041
Slovakia
27.0
5.7
6.3
0.12
0.36
0.43
0.56
0.023
Slovenia
26.6
5.7
6.2
0.12
0.35
0.42
0.55
0.023
Bulgaria
26.4
6.6
7.2
0.14
0.40
0.48
0.63
0.034
Malta
25.6
5.6
6.2
0.12
0.34
0.40
0.53
0.024
Poland
25.5
5.8
6.4
0.12
0.34
0.40
0.53
0.026
33 Abdallah S & Mahony S (2012) ‘Stocktaking report on subjective wellbeing’ Part of e-Frame (European Framework for
Measuring Progress. Forthcoming at http://www.eframeproject.eu/index.php?id=28
34 OECD (forthcoming) Guidelines on the Measurement of Subjective Well-Being.
35 We used the data from the 2007 European Quality of Life Survey (wave 2). The survey was carried out in 31 countries
including all the EU member states and Turkey, with sample sizes of over 1000 people for all countries and larger sample sizes
for the bigger countries such as Germany (see Task 7 for more information on this data set).
36 France is excluded from these analyses given the concerns raised about the sampling in the country. If it is included, then the
correlations weaken to R=0.47 for ‘feeling active and vigorous’, R=0.48 for the WHO index, and R=0.27 for life satisfaction.
October 2012
22
Latvia
25.4
5.8
6.4
0.13
0.35
0.41
0.54
0.026
Czech Rep.
25.3
5.9
6.6
0.13
0.35
0.42
0.55
0.028
Romania
24.7
5.7
6.3
0.13
0.33
0.39
0.51
0.027
Hungary
24.1
7.1
7.9
0.16
0.41
0.48
0.64
0.045
France
24.0
6.5
7.2
0.15
0.37
0.43
0.58
0.037
Turkey
22.1
5.2
5.9
0.13
0.29
0.34
0.46
0.025
Table 3: Mean and selected non-comprehensive inequality metrics
Mean
S80/S20
ratio
Differences
90/10
80/20
90/median
S80/S20
Cyprus
28.0
1.89
14.0
9.0
5.0
16.2
Austria
27.6
1.86
14.0
8.0
5.0
15.8
Spain
27.5
2.07
16.0
11.0
6.0
18.2
Greece
27.3
2.27
18.0
11.0
6.0
19.9
Slovakia
27.0
1.88
14.0
8.0
5.0
15.8
Slovenia
26.6
1.86
14.0
9.0
6.0
15.5
Bulgaria
26.4
2.12
16.0
11.0
6.7
18.1
Malta
25.6
1.90
13.0
9.0
6.0
15.5
Poland
25.5
1.97
14.3
9.0
6.0
16.0
Latvia
25.4
1.99
15.0
9.0
6.0
16.1
Czech Rep.
25.3
2.03
15.0
9.0
6.0
16.5
Romania
24.7
1.98
14.0
9.0
6.0
15.7
Hungary
24.1
2.50
19.0
12.0
8.0
19.8
France
24.0
2.28
16.0
10.0
6.0
17.9
Turkey
22.1
1.99
14.0
9.0
7.0
14.5
Table 4: Country ranks on selected inequality metrics
SD
SD
adj
Mean Pair
Distance
MPD
adj
GINI
GINI
adj
Atkinson
index
S80/S20
ratio
S80/S20
difference
Cyprus
6
5
8
5
14
5
7
12
7
Austria
9
7
12
6
15
6
14
15
10
Spain
3
2
3
2
5
2
5
5
3
Greece
1
1
1
1
3
1
2
3
1
Slovakia
11
8
11
8
13
8
13
13
11
Slovenia
13
10
14
9
12
9
15
14
14
Bulgaria
4
4
4
3
4
4
4
4
4
Malta
14
13
13
13
11
13
12
11
13
Poland
10
12
9
12
10
12
10
10
9
Latvia
8
11
7
11
9
11
9
8
8
Czech Rep.
7
9
6
10
7
10
6
6
6
Romania
12
14
10
14
8
14
8
9
12
Hungary
2
3
2
4
1
3
1
1
2
France
5
6
5
7
2
7
3
2
5
Turkey
15
15
15
15
6
15
11
7
15
October 2012
23
Figures 2 to 11 plot country means against these inequality metrics (for reasons of space we
exclude a couple of the percentile difference metrics.
Figure 2: SF mean vs. SF Standard Deviation
Figure 3: SF mean vs. SF Mean Pair Distance
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
5,0
5,5
6,0
6,5
7,0
7,5
8,0
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
Standard Deviation
SF mean
CY
AT
GR
ES
BG
SK
SI
LV
PL
MT
CZ
RO
HU
FR
TR
5,5
6,0
6,5
7,0
7,5
8,0
8,5
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
Mean Pair Distance
SF mean
October 2012
24
Figure 4: SF mean vs. SF GINI coefficient
Figure 5: SF mean vs. adjusted SF Standard Deviation
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
0,10
0,11
0,12
0,13
0,14
0,15
0,16
0,17
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
GINI coefficient
SF mean
October 2012
25
Figure 6: SF mean vs. adjusted SF Mean Pair Distance
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
0,20
0,25
0,30
0,35
0,40
0,45
0,50
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
Standard Deviation (adjusted)
SF mean
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
0,30
0,35
0,40
0,45
0,50
0,55
0,60
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
Mean Pair Distance (adjusted)
SF mean
October 2012
26
Figure 7: SF mean vs. adjusted SF GINI Coefficient
Figure 8: SF mean vs. SF Atkinson Index (ε=0.8)
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
GINI (adjusted)
SF mean
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
0,020
0,025
0,030
0,035
0,040
0,045
0,050
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
Atkinson Index (e=0.8)
SF mean
October 2012
27
Figure 9: SF mean vs. SF S80/S20 quintile share ratio
Figure 10: SF mean vs. SF S80/S20 quintile difference
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
1,60
1,70
1,80
1,90
2,00
2,10
2,20
2,30
2,40
2,50
2,60
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
S80/S20 ratio
SF mean
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
13,0
14,0
15,0
16,0
17,0
18,0
19,0
20,0
21,0
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
S80/S20 difference
SF mean
October 2012
28
Figure 11: SF mean vs. SF 90th percentile to median difference
A few observations can be made based on these data:
On all metrics, either Hungary or Greece are found to be the most unequal country in
terms of well-being, with France, Bulgaria and Spain also regularly amongst the top 5
most unequal.
All four permutations of mean and inequality can be exemplified with one or more
countries:
o High average, low inequality (e.g. Austria or Cyprus)
o High average, high inequality (e.g. Greece)
o Low average, low inequality (e.g. Turkey)
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
4,00
5,00
6,00
7,00
8,00
9,00
21,0 22,0 23,0 24,0 25,0 26,0 27,0 28,0 29,0
90/median difference
SF mean
October 2012
29
o Low average, high inequality (e.g. Hungary or France)
Comparing the ratio measures (GINI, Atkinson and S80/S20 ratio) with the non-ratio
measures (e.g. standard deviation, mean pair difference, S80/S20 difference), one can
see that the former, as discussed in the literature, tend to downplay inequality in high
well-being countries in comparison to the former. So the most unequal country according
to the GINI, the Atkinson Index and the S80/S20 ratio is Hungary, but according to the
non-ratio measures, it is Greece.
Choosing to adjust indices for instrument effects makes a bigger difference to country
rankings. In all three cases, adjustment increases the assessed inequality in those
countries with higher means. For example, Greece has the third highest GINI coefficient,
but the highest adjusted GINI coefficient. This adjustment also takes Spain from fifth to
second. The result is that, whilst there appears to be little correlation between mean and
most of the unadjusted dispersion metrics, significant positive correlations are found
when adjustments are made (up to R=0.66 for the adjusted mean pair distance). Smaller
but similar patterns happen when standard deviation and mean pair distance are
adjusted.
As a result there are also big differences in terms of which countries are ranked most
equal. According to all the non-ratio measures except for some of the percentile
difference metrics, Turkey (the country with the lowest mean well-being) is actually the
country with the lowest inequality in well-being. According to the ratio measures, it is
either Austria or Slovenia – two countries with relatively high well-being.
The percentile difference measures appear to be less useful, given the fact that they are
limited to a small range of integer values, For example, 9 countries have the same score
on the 90/median difference.
The 90/median difference is the only measure that correlates negatively with mean well-
being (-0.66), with the highest 90/median differences found in Hungary and Turkey. The
fact that countries like Greece and Spain have relatively low 90/median differences, but
high inequality according to other metrics, highlights that inequality in those countries is
driven more by large differences in the bottom half of the well-being distribution rather
than the top half.
Most measures of inequality correlate very highly (see table 5). If one categorises the
measures into a) non-ratio metrics (excluding the percentile differences), b) adjusted
metrics, and c) ratio metrics, all metrics within each category correlate with all other
indicators within that category at R=0.9 or higher (indeed R=0.99 or 1.00 in most cases).
Between categories, the picture is more complicated. All the non-ratio metrics correlate
at 0.7 or higher with all other metrics except the 90/median difference. Correlations
breakdown a little when comparing the instrument-adjusted non-ratio metrics and the
ratio metrics. For example, the unadjusted GINI does not correlate with any of the
adjusted metrics (including the adjusted GINI). The S80/S20 quintile ratio does not
correlate with the adjusted Mean Pair Distance.
October 2012
30
Table 5: Correlations between inequality metrics
GINI
Atkin
son
Index
S80/
S20
ratio
Stan
dard
devia
tion
Mea
n
pair
dista
nce
S80/
S20
differ
ence
SD
(adju
sted)
MPD
(adju
sted)
GINI
(adju
sted)
90/1
0
differ
ence
80/2
0
differ
ence
90/
medi
an
diffe
renc
e
GINI
0.93
0.99
0.73
0.82
0.79
0.88
0.85
0.78
Atkinson Index
0.93
0.97
0.91
0.95
0.94
0.73
0.62
0.69
0.96
0.90
0.61
S80/S20 ratio
0.99
0.97
0.81
0.88
0.87
0.58
0.54
0.93
0.86
0.71
Standard deviation
0.73
0.91
0.81
0.99
0.99
0.94
0.89
0.92
0.92
0.87
Mean pair distance
0.82
0.95
0.88
0.99
1.00
0.89
0.81
0.86
0.95
0.91
S80/S20 difference
0.79
0.94
0.87
0.99
1.00
0.90
0.83
0.88
0.94
0.89
SD (adjusted)
0.73
0.58
0.94
0.89
0.90
0.99
1.00
0.77
0.72
MPD (adjusted)
0.62
0.89
0.81
0.83
0.99
1.00
0.68
0.64
GINI (adjusted)
0.69
0.54
0.92
0.86
0.88
1.00
1.00
0.74
0.70
90/10 difference
0.88
0.96
0.93
0.92
0.95
0.94
0.77
0.68
0.74
0.88
0.59
80/20 difference
0.85
0.90
0.86
0.87
0.91
0.89
0.72
0.64
0.70
0.88
0.70
90/median
difference
0.78
0.61
0.71
0.59
0.70
2.3 Relationship with income inequality
How does income inequality relate to well-being inequality? None of the metrics correlated
significantly with the income GINI (2008 data from Eurostat).37 The strongest correlation was
at R=0.25 for the 80/20 percentile difference. Of the main potential metrics, the strongest
correlations with the income GINI were R=0.18 for the standard deviation, mean pair
distance and Atkinson Index (and indeed the coefficient of variation, though we do not
report on that metric). We plot the income GINI against Mean Pair Distance in Figure 12.
Figure 12: Income GINI coefficient vs. SF Mean Pair Distance
37 These analyses exclude Turkey, which is an outlier in terms of income inequality.
CY
AT
GR
ES
BG
SK
SI
LV
MT
PL
CZ
RO
HU
FR
TR
5,0
5,5
6,0
6,5
7,0
7,5
8,0
8,5
0,20 0,25 0,30 0,35 0,40 0,45 0,50
Mean pair distance
Income GINI
October 2012
31
As one can see, the country with the highest income GINI is also the country with the
smallest Mean Pair Distance (Turkey). Beyond that, one can see other countries with high
income inequality and low well-being inequality (e.g. Latvia and Romania), and countries
with low income inequality and high well-being inequality (e.g. Hungary). This mirrors the
findings of academics looking at a larger range of countries, who have not found much of a
relationship between income inequality and well-being inequality.38 Unlike Delhey & Kohler,
we do not find that the instrument-effect adjustment increases the correlation with income
inequality (but of course, our sample is a lot smaller).
These are important findings. It means that some countries that have not been concerned
about inequality so much, given their low income GINI coefficients, may need to reconsider
their policies.
It is worth noting that such differences between different types of inequality can be seen in
other domains. For example, looking at the 2011 Human Development Report, which
calculates inequality in health, education and income, there are stark differences. Greece,
which has very low health inequality (4.8%), has very high education inequality (14.3%).39
Latvia, whilst suffering very high income inequality (21.0%), actually has relatively low health
inequality (7.1%) and very low educational inequality (3.8%), a finding that corroborates the
low well-being inequality we report here.
Putting aside the concerns about the interval and ratio scales, might any interpretation be
available for the fact that SF GINI coefficients are much lower than income GINI coefficients
(ranging between 0.11 and 0.16)? This is considerably lower than the lowest income GINI
coefficient (0.23 for Slovenia).
Two substantive reasons can be given for the low inequality values. Firstly, the SF score is on
a bounded scale, whereas income spectrums can stretch much higher. Secondly, as we know
from studies of the relationship between income and subjective well-being, there is a
pattern of diminishing returns, whereby the positive effect of increasing income diminishes
as one goes up the income spectrum.40 The large incomes of the richest in any country do
not translate into substantially higher levels of well-being. As such it is likely that, caveats
regarding the scale aside, well-being inequality is indeed lower than income inequality.
2.4 Inequality adjustments
Economists interested in inequality have often attempted to produce what is often called a
social welfare function. Such a function combines the mean and a measure of distribution to
produce some sort of utility function which one might desire to maximise.
38 Berg M & Veenhoven R (2010) ‘Income inequality and happiness in 119 nations’ In B Greve (ed) Social policy and happiness in
Europe (Edward Elger: Cheltenham, UK)
39 Percentages represent losses to the mean resulting from inequality, as calculated using an Atkinson Index, with ε=1. See
2011 Human Development Report for details.
40 See Section 1.1. of Task 6 Annex.
October 2012
32
For example, the UN’s Inequality Adjusted Human Development Index, uses the Atkinson
Index to adjust each of the three components of the Index. Making these adjustments
changes the rankings of countries. For example, in the 2011 rankings whilst Norway and
Australia maintain first and second place respectively after adjusting for inequality, the USA
falls from 4th to 23rd, whilst Sweden rises from 8th (excluding countries for which the IHDI
could not be calculated) to 3rd.
2.4.1 Options
Obviously, how the mean and the measure of inequality are combined is a political issue as it
implies somehow weighting the relative importance of maximising average welfare and
increasing equality in welfare.
We calculated the following adjustments or values:
Median – The well-being of the ‘middle’ person. Often used instead of the mean when
there is a large skew (for example in the case of income), so as to avoid extreme values
influencing the measure.
Geometric mean – Calculated by multiplying all the well-being values together and then
rooting (rather than adding them all together and dividing by the number of values,
which is how an arithmetic mean is calculated). More sensitive to inequality than the
arithmetic mean, producing smaller values.
Atkinson adjusted mean – Based on an Atkinson Index calculated with ε=0.8.
Mean of bottom half, bottom 20% and bottom 10% - The last three approaches involve
zooming in on the worst off in each country.
2.4.2 Results
The above functions are presented in Table 6, with countries in order of decreasing ‘raw’
mean. Figure 13 shows how the Atkinson Index adjustment affects the mean for each
country. With an ε=0.8 there is very little difference to the means, and only one change in
rank order of countries (Greece and Slovakia swapping places). Figure 14 plots the overall
mean compared with the means for the bottom half, bottom 20% and bottom 10%.
Table 6: Means ‘adjusted’ for inequality
Country
Mean
Median
Geometric
Mean
Atkinson adjusted
mean (e=0.8)
Mean of
Bottom half
Mean of
Bottom 20%
Mean of
Bottom 10%
Cyprus
28.0
29.0
27.0
27.2
23.6
18.3
14.2
Austria
27.6
29.0
26.8
26.9
23.3
18.3
14.9
Spain
27.5
29.0
26.4
26.6
22.3
17.0
13.5
Greece
27.3
29.0
25.8
26.2
21.7
15.6
11.7
Slovakia
27.0
28.0
26.1
26.3
22.6
17.9
14.8
Slovenia
26.6
27.0
25.8
26.0
22.2
17.9
15.1
Bulgaria
26.4
27.3
25.3
25.6
21.4
16.2
12.7
Malta
25.6
26.0
24.8
25.0
21.2
17.2
14.5
October 2012
33
Poland
25.5
26.0
24.6
24.8
21.1
16.5
13.4
Latvia
25.4
26.0
24.5
24.7
20.8
16.3
13.2
Czech Rep.
25.3
26.0
24.4
24.6
20.8
16.1
13.1
Romania
24.7
25.0
23.8
24.0
20.3
16.0
13.2
Hungary
24.1
25.0
22.7
23.0
18.5
13.1
9.9
France
24.0
25.0
22.8
23.1
18.9
14.0
11.1
Turkey
22.1
22.0
21.4
21.6
18.0
14.7
12.4
Figure 13: Raw mean compared with Atkinson adjusted mean
Figure 14: Overall mean compared with means for different groups
October 2012
34
Comparing the means for different groups provides more interesting contrasts. Whilst
Cyprus has the highest mean overall, the 20% and 10% least ‘happy’ (i.e. with lowest well-
being) in that country have lower well-being than the 20% and 10% least ‘happy’ in Austria.
Whilst Greece has the fourth highest mean overall, the bottom decile (i.e. bottom 10%) in
that country ranks third from bottom comparing with the bottom decile in other countries.
Slovakia and Slovenia, two countries that are in the middle of the distribution for overall
mean, do very well in terms of well-being for the bottom deciles. For example, the bottom
10% in Slovenia have the highest well-being of any bottom 10% group. Malta is another
country that moves up the rankings when one considers only lower deciles.
Meanwhile, at the bottom end, the bottom decile and the bottom quintile in Hungary have
the lowest well-being of all 15 countries.
Another approach is to look at the ‘absolute poverty’ rate in terms of well-being – i.e. what
proportion in each country has well-being below a given level. We take a score of 10 or
below to be ‘languishing’. This means scoring on average between 1 and 2 out of 5 for each
question. For the positive questions, this means saying that you only felt a particular
emotion ‘a little of the time’ or ‘none of the time’. For the negative questions, this means
saying that you felt a particular emotion ‘most of the time’ or ‘all of the time’.
By far the highest rate was to be found in Hungary – 4.9% languishing. The next highest
country is France at 3.5%, followed by Greece (3.2%). The lowest levels of languishing are in
Hungary’s neighbours Slovenia (1.1%) and Austria (1.3%).
,0
5,0
10,0
15,0
20,0
25,0
30,0
CY
AT
ES
GR
SK
SI
BG
MT
PL
LV
CZ
RO
HU
FR
TR
Mean
Mean of Bottom half
Mean of Bottom 20%
Mean of Bottom 10%
October 2012
35
2.5 Discussion
Which, if any of the discussed metrics of well-being inequality or adjusted well-being means
should Eurostat adopt?
Considering the comprehensive inequality metrics, there was some substantial differences in
results, suggesting that choosing the best one is not a straightforward matter. The
theoretical arguments for instrument-effect-adjustment are appealing, but, for the SF, they
appeared to lead to quite strong correlations between the mean and the dispersion
measure, suggesting that they were introducing some undesirable scale-dependence.
Bearing this in mind, we would tentatively suggest one of the non-ratio non-adjusted
measures be the most appropriate. The Mean Pair Distance has the advantage of being
mathematically related to the GINI coefficient, making it our preferred candidate. However,
we appreciate that it is probably prudent for Eurostat to wait for academics to reach a
consensus before committing itself to any single metric.
With regards to a non-comprehensive inequality metric, with communication value, we
would suggest the S80/S20 difference, which may be easier communicated as the
‘top/bottom well-being quintile difference’. Theoretically, this avoids the pitfalls of a
S80/S20 ratio discussed in Section 1.3. The percentile difference could potentially be
appealing, but there is little room for variance between them, as, in most cases, they will be
integer values (with life satisfaction on a scale of 0-10 this is even more problematic).
Furthermore, the measure appeared to differentiate clearly between high inequality
countries (Greece and Hungary), medium inequality countries (France, Bulgaria and Spain)
and the rest. It correlated well with most other measures of well-being inequality
(particularly the non-ratio comprehensive measures) and did not correlate at with mean
well-being. Lastly, it nicely mirrors the S80/S20 quintile ratio that Eurostat report for income
inequality.
The Atkinson-adjusted mean does not appear to add much value, given its close correlation
with the normal mean. Furthermore, as we discussed in Section 1.4.6, it might be unsound
conceptually.
However, the graphs showing the means for the bottom groups in each country do appear
to be very informative, as is the statistic on the proportion of people below a certain level of
well-being.
October 2012
36
3. Bivariate and multivariate analyses
This section considers the different variables which explain the variation in well-being within
countries. We carried out two sets of analyses.
Firstly, we carried out bivariate analyses with the SF score as dependent variable and a range
of independent variables separately. Depending on the independent variable, we carried out
t-tests, one-way ANOVAs and correlations. Such analyses allowed us to compare the
outcome variable for people in different situations or with different demographics.
However, bivariate analyses risk omitting important causal factors. For example, it may be
that two ethnic groups have different levels of average well-being but that this difference is
due to differences in income. Were one to control for income, one might find that two
people from different ethnic groups, but with the same income actually have the same well-
being on average. Multivariate regressions allow us to explore the predictive power of
multiple independent variables at the same time, and understand their overlapping effects.
They also allow us to get an overall assessment of how much of the variation in well-being
within each country can be explained by economic and demographic factors.
We present the bivariate analyses first, looking at each variable separately, before
presenting the regressions.
Because these analyses were done prior to the univariate analyses presented in this report,
SF scores are presented here on a scale of 9 to 45 as opposed to 1 to 37. However, this
makes no difference to the analyses.
3.1 Comparisons with Task 7 results
In this section we make comparisons with the results from Task 7 of this Study. Task 7
involved the analysis of the European Quality of Life Survey (wave 2 - 2007) and the
European Social Survey (wave 3 – 2006) to explore drivers of well-being using micro-data.41
A first task in Task 7 was to look at the effects of six ‘control’ variables, all of which are also
analysed here in Task 8 (age, gender, marital status, employment status, income and
education level). As such we are able to compare the effect sizes seen in Task 7 with those
we find here. Task 7 also looked at two other variables that we consider here (urbanisation
level and support from friends), as these are potential variables to be included in the well-
being set.
A few differences between the analyses and the data sets used should be noted:
The dependent variable here is the SF-36, a measure of hedonic well-being. Task 7 used
life satisfaction as its dependent variable, a measure of evaluative well-being.
41 See Task 7 for more information on these two surveys.
October 2012
37
We present most analyses broken down by country. In Task 7, analyses are presented
for data from individuals from all countries combined.
Sample sizes here vary by country, between around 1,900 for the Czech Republic to
30,000 for Poland. The total sample size for Task7 for all countries combined was
around 20,000.
The countries covered here are biased towards Central, Eastern and Southern Europe,
with important gaps including Germany, the UK, Italy and Scandinavia.
Some minor differences in the operationalisation of some of the independent variable.
Minor differences in how different independent variables were combined.
3.2 Bivariate analyses
Where asterisks (*) are used, they indicate the significance of a difference. Two asterisks
indicate significance at the 0.01 level, and one asterisk indicates significance at 0.05 level. In
all tables, countries are ranked according to the overall mean (with highest SF scores at the
top).
Figures in grey represent means with N’s of less than 100. Where there is an N of less than
50 for a particular cell, the number is not reported. Some sub-samples are small compared
to the total sample size simply because there are very few people with that particular
demographic or economic characteristic e.g. there are few disabled people in the sample.
3.2.1 Gender
Table 7: Well-being by gender
Country
Means
Difference
Rank
(Female)
Male
Female
Cyprus
37.0
35.0
1.9**
1
Austria
36.3
34.9
1.3**
2
Spain
36.8
34.2
2.6**
5
Greece
36.4
34.3
2.1**
4
Slovakia
35.4
34.5
0.9**
3
Slovenia
35.4
33.8
1.6**
6
Bulgaria
35.2
33.8
1.4**
7
Malta
34.5
32.8
1.7**
9
Poland
34.1
32.9
1.1**
8
Latvia
34.0
32.8
1.2**
10
Czech Rep.
34.3
32.3
2.0**
11
Romania
33.4
32.1
1.3**
12
Hungary
33.2
31.2
2.0**
14
France
32.9
31.2
1.7**
13
Turkey
30.8
29.5
1.3**
15
October 2012
38
Males have significantly higher SF scores in all countries, the smallest difference being in
Slovakia (0.9) and the largest in Spain (2.6). Ranking countries according to SF scores of
females leads to Slovakia and Spain swapping 5th and 3rd places. Poland and France also both
move up one place, at the expense of Malta and Hungary respectively.
Note that such a large difference between males and females is not typically visible with life
satisfaction. In Task 7, the difference between males and females was only significant at the
0.05 level in the ESS data set.
3.2.2 Age
Figure 15: Well-being with age, all countries
Combining data from all countries, one can see a steady decline in SF scores with age, with
the biggest fall between the 65-74 and the 75-84 age bands. This pattern is consistent across
all countries. Note that this is different for patterns for life satisfaction. Amongst Western
European countries a U-shaped function is seen - with life satisfaction falling up until around
28,0
29,0
30,0
31,0
32,0
33,0
34,0
35,0
36,0
37,0
38,0
16-19
20-24
25-34
35-44
45-54
55-64
65-74
75-84
85+
SF score
Age group
October 2012
39
middle age (40s or 50s) and then rising again with age.42 The difference is likely a result of
the inclusion of vitality questions in the SF scale, which invariably declines in older people.
Whilst this pattern is seen in all countries, it is stronger in some than others, as is shown in
Table 8, which compares two age groups: 20-64 and 65+.
Table 8: Well-being by age
Country
Means
Difference
Rank (65+)
20-64
65+
Cyprus
36.2
33.3
3.0**
3
Austria
36.0
33.4
2.6**
2
Spain
35.8
33.4
2.4**
1
Greece
35.9
32.2
3.7**
5
Slovakia
35.3
31.7
3.6**
7
Slovenia
35.1
32.1
3.0**
6
Bulgaria
35.3
29.8
5.5**
12
Malta
33.7
32.4
1.3**
4
Poland
33.8
30.9
2.8**
9
Latvia
33.7
30.5
3.1**
10
Czech Rep.
33.4
31.4
2.0**
8
Romania
33.3
28.2
5.1**
15
Hungary
32.5
29.6
2.9**
13
France
32.4
30.1
2.3**
11
Turkey
30.0
28.3
1.6**
14
The biggest impact of age is seen in Bulgaria (5.5) and Romania (5.1), whilst the smallest is
seen in Malta (1.3) and Turkey (1.6). Slovakia, which did well in terms of gender equality,
does less well in terms of age equality, with a gap of 3.6. Re-ranking countries according to
the SF scores of 65+ respondents, puts Spain in top place, and sees Bulgaria dropping from
7th to 12th. Malta moves up from 8th to 4th. Romania falls from 12th place to bottom, below
Turkey.
3.2.3 Marital status
Table 9: Well-being by marital status
Country
Means
Difference
Married
Divorced
Single
Widowed
Cyprus
35.8
34.0
37.8
31.7
-1.7**
Austria
35.7
34.5
36.3
32.1
-1.3**
Spain
35.4
33.6
36.7
31.7
-1.8**
Greece
35.1
32.3
37.5
30.5
-2.9**
Slovakia
35.0
33.7
36.5
31.1
-1.3**
Slovenia
34.6
35.4
35.3
31.6
0.8
42 For example in Task 7, the EQLS data revealed a pattern whereby age groups between 25 and 64 years old all had significantly
lower life satisfaction then those in the youngest age group 18-24, but those aged 65+ did not have significantly lower life
satisfaction
October 2012
40
Bulgaria
34.1
33.8
37.6
29.6
-0.3
Malta
33.6
31.7
34.4
30.9
-1.9**
Poland
33.4
31.5
35.2
30.5
-1.9**
Latvia
33.5
31.3
35.1
29.8
-2.2**
Czech Rep.
33.5
32.3
34.4
29.8
-1.2*
Romania
32.6
31.9
35.6
27.9
-0.7**
Hungary
32.1
30.6
34.0
28.7
-1.5**
France
32.1
31.1
32.6
29.0
-1.1**
Turkey
29.8
27.4
31.8
27.2
-2.4**
According to one-way ANOVAs, people with different marital status have significantly
different well-being in all countries. Typically single people have the highest SF scores,
followed by married, then divorced then widowed. However, one should be careful in
interpreting these results, as they are probably related to age – i.e. that younger people,
who have higher SF scores, are more likely to be single, and older people, who have lower SF
scores, are more likely to be divorced or widowed.
We carried out a t-test between married and divorced groups (where we felt age was least
likely to be driving differences), and found significant differences in most countries, the
exceptions being Bulgaria and Slovenia – in the latter case possibly due to small sample sizes.
The biggest differences between the two groups were in Greece (2.9) and Turkey (2.4). It is
worth noting that these countries are not countries with particularly large differences in SF
scores between age groups, so it is likely that these differences are to do with marital status
and not just an artefact of age (although this will be tested further in the regressions).
3.2.4 Employment status
Table 10: Well-being by employment status
Country
Means
Differences from Employed
Ranks
Empl
Unem
Dis
Ret
Unem
Dis
Ret
Unem
Dis
Ret
Cyprus
36.6
35.7
26.1
33.4
-0.9*
-10.5**
-3.2**
1
8
3
Austria
36.6
32.8
33.6
-3.8**
-3.1**
9
2
Spain
36.4
35.7
28.5
34.0
-0.7**
-7.9**
-2.4**
2
2
1
Greece
36.5
35.0
27.6
33.1
-1.4**
-8.9**
-3.4**
3
4
5
Slovakia
35.6
34.6
29.2
32.5
-1.1**
-6.4**
-3.2**
4
1
7
Slovenia
35.5
34.0
33.2
-1.5**
-2.3**
6
4
Bulgaria
36.1
34.2
26.2
30.5
-1.8**
-9.8**
-5.6**
5
7
12
Malta
34.6
32.0
33.1
-2.6**
-1.5**
10
6
Poland
34.3
33.2
28.1
31.7
-1.1**
-6.2**
-2.6**
8
3
8
Latvia
34.3
30.8
26.4
30.5
-3.5**
-7.9**
-3.8**
13
6
11
October 2012
41
Czech Rep.
34.0
31.5
31.7
-2.5**
-2.3**
12
9
Romania
34.1
33.6
24.8
29.1
-0.4
-9.3**
-5.0**
7
11
15
Hungary
33.6
31.8
25.0
30.4
-1.8**
-8.6**
-3.2**
11
10
13
France
33.0
30.2
25.2
30.9
-2.8**
-7.8**
-2.1**
14
9
10
Turkey
30.6
29.6
26.6
30.4
-1.0**
-4.0**
-0.2
15
5
14
key: unempl – unemployed; dis – disabled; ret- retired
The employment status variable (which included options for employed, unemployed,
permanently disabled, retired, in education, community or military service, housework and
other), was significantly related to SF scores for all countries. We focus here on the SF scores
for four groups: employed, unemployed, permanently disabled and retired. As one can see
in Table 7, the unemployed, disabled and retired have consistently lower SF scores in all
countries for which enough data was available.
Looking first at unemployed people, there are significant well-being penalties in all countries
with the exception of Romania. The biggest differences are seen in Austria (3.8), and Latvia
(3.5) two countries which generally do not have high inequality in well-being. The difference
is small (but still significant) in Spain (0.7) and Cyprus (0.9). Previous research has shown that
the impact on well-being of unemployment tends to be smaller in countries with high
unemployment.43
Retirement, like being widowed, is likely to be related to age. Indeed Bulgaria and Romania
(the two countries with the biggest age differences) also have the biggest differences
between the employed and the retired (5.6 and 5.0 respectively). The difference is not
significant in Turkey. Whether the difference between retired and non-retired is a factor of
age, or whether the age effect is a factor of retirement will be explored in Section 3.3, when
we include both variables in multivariate regressions.
Sample sizes are a little smaller for disabled people (under 50 in Austria, Slovenia, Malta and
the Czech Republic, and under 100 in Cyprus and Greece). Nevertheless the difference
between being disabled and employed is significant in all countries (even when sample sizes
are small), and is the largest of all employment statuses. In Cyprus, the effect is 10.5. The
smallest effect is seen in Turkey (4.0).
How do countries rank in terms of well-being for the unemployed, retired and disabled?
Unemployed people have the highest well-being in Cyprus, followed by Spain (up from 3rd)
and then Greece. Austria drops to 9th place. Again, this may be due to unemployment having
a milder effect on well-being in countries with high unemployment; 44 or it may be because
of strong social networks mitigating the negative social and financial impacts of not having a
job. Slovakian disabled people have the highest well-being of all disabled people, followed
by Spanish, then Polish. The lowest well-being amongst disabled people is found amongst
Romanians and Hungarians. Spanish retired people have the highest well-being of retired
people, followed by Austrians, then Cypriots. Romanians retirees ranked lowest.
43 See Task 6 Annex, Section 2.1.
44 Note that Latvia’s unemployment rate is now 18% - counting it amongst the highest in Europe. However, at the time when
the EHIS was conducted in that country (2008), the rate was a lot lower (around 6%).
October 2012
42
3.2.5 Education level
The EHIS data includes a variable for education level, categorised according to ISCED
(International Standard Classification of Education) level. The levels are as follows:
1 – Primary education
2 – Lower secondary education
3 – Upper secondary eduction
4 – Post-secondary non-tertiary education
5 – First stage of tertiary education
6 – Second stage of tertiary education
0 indicates no formal education.
Figure 16: Well-being with education, all countries
As one can see in Figure 16, combining all countries, there is a steady increase in mean SF
scores from one level to the next. The only exception is between level 4 and level 5, which is
not that significant, given that few people have as their highest level of education level 4
(post-secondary non-tertiary education), and this level is quite varied in what it covers.
Overall the biggest difference is from level 1 (primary education) to level 2 (lower
secondary): 31.6 to 33.9). The increase is seen in all countries, and the planned linear
ANOVA is consistently significant.
Table 11: Well-being by education level
Country
Means – ISCED level
Unstandardised
Regression
Coefficient
1
2
3
4
5
6
7
Cyprus
30.7
33.3
36.4
36.8
37.2
37.5
1.1
28,00
29,00
30,00
31,00
32,00
33,00
34,00
35,00
36,00
37,00
38,00
0
1
2
3
4
5
6
SF score
ISCED
October 2012
43
Austria
33.2
34.1
35.9
36.7
36.3
35.9
0.7
Spain
32.3
34.8
36.2
36.3
36.5
38.4
0.7
Greece
30.4
33.5
35.9
36.4
37.0
36.5
36.2
0.7
Slovakia
31.9
33.9
35.0
35.5
36.0
0.7
Slovenia
30.0
33.2
34.5
35.8
36.1
35.4
1.0
Bulgaria
32.1
31.1
33.4
35.1
33.7
35.7
0.9
Malta
30.5
32.0
32.6
33.9
34.7
34.7
35.4
0.8
Poland
27.9
31.3
36.1
33.6
33.8
34.6
34.9
0.8
Latvia
31.4
32.9
33.3
33.0
34.6
0.6
Czech Rep.
32.2
33.4
34.2
0.6
Romania
27.7
28.6
32.4
33.3
33.4
34.7
35.7
1.3
Hungary
27.8
30.6
32.5
33.6
33.9
1.3
France
30.0
29.8
32.6
32.5
33.2
0.7
Turkey
28.1
29.5
31.6
31.4
31.8
0.8
However, the gradient of that variation45 is different from country to country. The largest
gradients are in Romania and Hungary where, on average, an extra level of education (based
on ISCED) equates to an increase of SF score of 1.3 on a scale from 9 to 45. The smallest
differences are in the Czech Republic and Latvia, where an extra level of education only
increases SF scores by 0.6.
By comparison, Task 7 found that each extra level of education increased life satisfaction (in
the EQLS) over all countries pooled together by 0.17 on a scale of 1-10 which is much lower
in absolute and in relative terms.
3.2.6 Income
Figure 17: SF scores by income decile, selected 5 countries
45 The gradient reported here is the unstandardised B coefficient for the education term in a bivariate regression with education
as the only independent variable and SF score as the dependent variable. Education was entered into the regression as a linear
term.
October 2012
44
Figure 17 shows how SF score varies by income decile within 5 countries selected for
presenting interesting patterns. Unsurprisingly, there is a general trend for higher well-being
amongst wealthier respondents. In some countries, the pattern is sharper than in others. For
example, in Turkey, moving up one income decile increases SF scores by 0.34 on average.46
In Latvia, moving up one income decile increases SF scores by 1.08 on average, i.e. the effect
is three times as large. As such, whilst high-income Latvians have much higher well-being
than high-income Turks (a difference of 4 points between the tenth deciles of the two
countries), there is little difference between the well-being of low income Turks and Latvians
– indeed Turks in the second decile have higher well-being than Latvians in the second
decile. From income decile 7 upwards (i.e. the richest 40%), Latvians have on average
marginally higher well-being than Slovenians, whilst for the first 6 deciles (i.e. the poorest
60%), Slovenians have higher well-being.
Further data is available in Table 12. The other country with a particularly steep
income/well-being gradient is Romania (1.00). The R values reported provide a picture of the
tightness of the relationship.47 One can see that, for example, whilst there is a large R value
for Bulgaria (0.33), indicating a strong relationship, the R value for Greece is only 0.08. These
R values are based on a regression which assumed a linear relationship between income
decile and well-being. Whilst the evidence is strong that the relationship between income
and well-being is logarithmic (see Tasks 6 and 7), we expect that, given the nature of the
income distribution, using income deciles will tend to capture that logarithmic pattern, as
the difference between the higher deciles in terms of absolute income is likely to be higher
than the differences between the lower deciles.
46 As with education, this gradient is estimated as the unstandardised B coeffiecient in a bivariate regression.
47 R values come from a Spearman’s rank correlation test, given that the income band data is best treated as non-parametric.
26,0
28,0
30,0
32,0
34,0
36,0
38,0
40,0
1
2
3
4
5
6
7
8
9
10
SF scores
Income Band
CY
AT
SI
LV
TR
October 2012
45
Table 12 also re-ranks countries according to the well-being of their bottom decile and their
bottom two deciles (i.e. their bottom quintile). Ranking according to bottom quintile, Austria
is the country with the highest well-being, ahead of Spain and then Cyprus. Turkey and
Hungary remain bottom of the ranking (income decile data is not available for France), but
just ahead of them is now Latvia, which ranked 10th for the overall population. Ranking
according to bottom decile, Cyprus now falls to 5th place, and Latvia falls to 2nd from bottom.
Lastly, Table 12 includes a column which attempts to identify where on the income spectrum
the income/well-being relationship tails off, if at all. This is based on comparing each
consecutive income decile with all income deciles above it. The number recorded in the
column is the decile beyond which further increases in income appear to confer no
significant well-being advantage (a ‘plenty line’).48
Whilst one needs to treat this analysis with caution as it is affected by sample size (i.e.
countries with larger samples are likely to see significant differences even when their
magnitude is smaller), it does reveal interesting information. Whilst, the relationship
between income and well-being in Austria is not particularly strong, it is consistent,
appearing to influence well-being right up to the highest income band. By contrast, the very
strong relationship between income and well-being in Romania all appears to take place
between the first three income bands. From the third band upwards there are no significant
differences (however, there may be some sampling issues in Romania, as the upper deciles
are severely underrepresented in the sample). Another country with an interesting pattern is
Greece. Beyond the 5th income band there is no increase in SF scores until the very final
income decile.
Table 12: Well-being by income decile
Country
Means – for income deciles
Unstandardised
Regression Coefficient
R
Ranks
Plenty Line
1
1&2
5
1
1&2
Cyprus
31.8
33.1
36.1
0.58
0.25
5
3
8
Austria
33.7
33.8
35.6
0.36
0.17
1
1
10
Spain
32.7
33.2
35.6
0.45
0.15
2
2
8
Greece
32.4
32.5
36.5
0.45
0.08
3
4
5
Slovakia
32.1
32.0
35.1
0.50
0.21
4
6
6
Slovenia
31.7
32.1
35.7
0.45
0.21
6
5
5
Bulgaria
30.4
31.0
34.6
0.67
0.33
8
10
10
Malta
31.2
31.5
33.9
0.35
0.18
7
8
4
Poland
29.6
30.7
33.4
0.50
0.20
11
11
9
Latvia
28.8
29.2
33.7
1.08
0.29
13
12
7
Czech Rep.
30.1
31.1
32.7
0.44
0.17
9
9
7
Romania
29.7
31.5
34.1
1.00
0.24
10
7
3
Hungary
29.0
28.9
31.7
0.56
0.23
12
13
10
France
n/a
Turkey
28.0
28.8
30.4
0.34
0.19
14
14
9
48 Woodward D & Abdallah S (2012) ‘Why we need a plenty line’. Available at http://globaltransition2012.org/challenge-
papers/global-inequality/
October 2012
46
3.2.7 Place of birth / migrant status
Table 13: Well-being by place of birth / migrant status49
Country
Means
Biggest
difference
Native
Other EU
Non-EU
All ‘migrants’
Cyprus
35.7
37.3**
37.3**
37.3**
1.5
Austria
35.8
34.8**
34.1**
34.3**
-1.7
Spain
35.4
36.1**
36.0**
36.1**
0.6
Greece
35.3
36.1
35.4
35.5
0.1
Slovakia
35.0
33.5
-1.5
Slovenia
34.7
34.0
33.8*
-0.9
Bulgaria
34.5
Malta
33.6
35.2**
34.7
34.9**
1.1
Poland
33.5
32.1**
31.0**
31.4**
-2.5
Latvia
33.7
30.8**
31.7**
31.6**
-2.8
Czech Rep.
33.3
32.5
34.5
33.1
-0.9
Romania
32.7
Hungary
32.1
33.7**
33.6**
1.5
France
32.1
30.9**
31.2**
31.1**
-1.2
Turkey
30.1
31.1
31.4**
31.2**
1.0
The ‘place of birth’ variable reflects migrant status, though some people who were born
abroad and are defined officially as migrants may not consider themselves as such (for
example they may be born to native parents who were temporarily abroad at their birth).
Being a migrant (defined like this) appears to have a particularly strong negative effect on
well-being in Latvia and Poland. But it does not always negatively predict well-being. In some
countries (e.g. Greece or the Czech Republic) there is no significant difference from natives,
and in others (particularly southern countries such as Cyprus, Malta and Spain, but also
Hungary) there is even a significant advantage. This might reflect differences in the different
migration patterns to the countries in the sample. Latvia, which has the highest effect, is an
unusual case, as many of the people identified as migrants would be Russians or other
people from the former USSR who moved to Latvia at a time when such a move would not
have involved crossing a national border.
49 In this table, significant differences from natives are noted in the “means” columns for migrants so that we can highlight the
biggest differences in the last column.
October 2012
47
3.2.8 Household type
Table 14: Well-being by household type
(no significance levels reported)
Country
Means – household types
Differences
One-
person
Lone
parent
Multiple
Adult
Multiple adult – One-
person
Family with children
– Lone parent
Cyprus
33.8
34.2
36.2
-2.4**
-2.5**
Austria
n/a
Spain
n/a
Greece
33.2
34.7
35.5
-2.4**
-1.7
Slovakia
33.4
35.1
35.5
-2.1**
-0.8*
Slovenia
32.7
35.6
34.8
-2.1*
0.1
Bulgaria
31.6
36.3
34.8
-3.2**
0.0
Malta
32.5
31.7
33.8
-1.3**
-2.4**
Poland
31.6
33.4
33.7
-2.1**
-1.0**
Latvia
31.3
33.4
33.7
-2.4**
-1.7**
Czech Rep.
32.3
32.8
33.5
-1.2
-1.3*
Romania
30.0
33.1
33.1
-3.1**
-1.2**
Hungary
29.1
32.0
32.5
-3.4**
-1.2*
France
30.4
31.5
32.4
-2.0**
-1.4**
Turkey
n/a
The EHIS micro-data codes households into various categories including one-person
households, lone parents, couples, couples with children, and other households with or
without children. In Table 14, we have grouped couples, couples with children and other
households into one column multiple adult households. ANOVAs reveal that household type
was significantly related to well-being for all countries for which data was available. Planned
comparison tests50 show that people living alone (i.e. one-person households) had
significantly lower SF scores than other household categories in all countries except the
Czech Republic.
Lone-parent households also suffered a well-being deficit, when compared with other
households with children (see the last column of Table 14), with a significant difference
(based on t-tests) seen in all countries except Greece, Slovenia and Bulgaria. The difference
was biggest in Cyprus (2.5) and Malta (2.4).
3.2.9 Urbanisation level
Table 15: Well-being by urbanisation level
50 Planned comparison tests can be conducted in SPSS as part of an ANOVA analysis. Whilst an ANOVA tells you whether a
particular independent variable has a significant relationship with the dependent variable, planned comparisons allow one to
compare particular groups created on the basis of the independent variable. Best practice is to decide which comparisons are
likely to be of interest before carrying out the ANOVA, and submit these for planned comparison. Submitting all possible
comparisons is likely to lead to Type I errors.
October 2012
48
Country
Means – urbanisation level
Urban-Rural
(significant?)
Urban
Semi
Rural
Cyprus
36.3
35.6
35.5
0.8**
Austria
34.7
36.0
36.1
-1.4**
Spain
35.5
35.6
35.5
-0.1
Greece
35.3
34.6
35.5
-0.3
Slovakia
35.0
35.0
34.9
0.2
Slovenia
35.2
34.6
34.4
0.7*
Bulgaria
34.9
34.7
33.9
1.1**
Malta
33.9
33.3
Poland
33.4
33.6
33.5
-0.1
Latvia
33.4
32.8
33.4
0.0
Czech Rep.
32.9
33.8
33.3
-0.4
Romania
33.3
33.0
32.3
1.0**
Hungary
32.5
32.0
31.9
0.7**
France
n/a
Turkey
n/a
The urbanisation variable used is based on official statistics, rather than the subjective
assessment of the respondent.
Confirming findings reported in Task 6 of this project, rural populations had significantly
higher well-being than urban populations in Austria. However, in no other country is a
significant rural advantage found. Indeed, the reverse (an advantage for urban populations)
was found in several countries: Bulgaria, Romania, Cyprus, Hungary, and Slovenia. The
different directions of this effect in different countries make sense whilst living in a rural
context in Western European countries is often a lifestyle choice, often associated with
people on higher incomes, those living in Central and Eastern European countries may not
be in rural contexts out of choice and might represent low income, low education sectors of
society.
The different patterns might explain why the effect of urbanisation found in Task 7 was so
small (though significant). Pooling data from all countries (both Western and Eastern)
probably obscured potentially strong relationships within countries.
3.2.10 Body-mass index
Body-mass index was negatively correlated with SF scores in all countries, with the largest R
seen in Bulgaria (R=-0.16), and the smallest in France (R=-0.04).51
3.2.11 Support from friends and families
Table 16: Well-being by number of people respondent can rely on for support
Country
Means – number of people
Biggest
51 Perason’s parametric correlation tests were carried out.
October 2012
49
None
1-2
3-5
>5
difference
Cyprus
31.5
34.5
35.9
37.0
5.5**
Austria
Spain
31.4
33.8
35.3
36.6
5.1**
Greece
34.5
34.8
35.7
36.8
2.3**
Slovakia
29.4
33.4
35.3
36.1
6.6**
Slovenia
28.7
33.1
35.2
35.9
7.2**
Bulgaria
30.9
33.4
35.6
36.8
5.9**
Malta
31.8
33.1
34.0
34.6
2.8**
Poland
29.8
32.1
33.8
34.8
5.0**
Latvia
29.3
32.5
34.3
35.2
5.9**
Czech Rep.
25.6
31.8
33.5
34.7
9.1**
Romania
30.9
31.6
33.0
34.3
3.4**
Hungary
28.7
30.4
32.2
34.1
5.4**
France
Turkey
27.5
29.0
30.2
30.7
3.2**
The last item we looked at is a little different, but potentially might explain a lot of variation
in SF scores. Respondents were asked the number of people they could rely on in case of
serious problems – their social support network.
In all countries, there is a significant relationship between social support network size and SF
scores, with larger networks consistently resulting in higher well-being. This is consistent
with results from Task 7 (see section on 2.5.3) on a related item – there not having anyone
to rely on in difficult situations was associated with a life satisfaction of 0.7 lower than for
people who could rely on others in difficult situations.
The sharpest differences can be seen in Czech Republic and Slovenia, two of the more equal
countries on other measures. It is also worth noting which countries have the largest
proportions reporting having nobody to rely on in case of serious problems – 4% in both
Greece and Turkey. If one includes people who report only having 1 or 2 people they can
count on, the highest percentages are in Bulgaria (53%) followed by Greece and Latvia (51%
each).
October 2012
50
3.3 Multivariate analyses
Table 17: Standardised beta coefficients and other statistics from multivariate regressions
AT
BG
CY
CZ
ES
FR
GR
HU
LV
MT
PL
RO
SI
SK
TR
Female
-.08
-.06
-.19
-.14
-.18
-.12
-.14
-.10
-.06
-.11
-.07
-.07
-.10
-.02
-.09
Age
-.13
-.40
-.03
-.10
-.13
-.14
-.23
-.19
-.26
-.10
-.29
-.38
-.14
-.25
-.13
Widowed
-.06
-.07
-.04
-.11
-.05
-.02
-.04
-.01
-.06
-.05
-.02
-.06
-.04
-.15
-.02
Divorced
-.02
-.03
-.04
-.02
-.04
.01
-.09
-.01
-.07
-.02
-.04
-.04
.05
-.04
-.04
Unemployed
-.09
-.04
-.04
-.02
.00
-.07
-.05
-.04
-.13
-.04
-.02
-.01
.00
-.04
-.05
Retired
-.06
-.01
-.02
.02
.04
.01
-.01
.01
.01
-.01
.07
-.07
.06
.06
.03
Disabled
-.10
-.16
-.14
-.22
-.13
-.15
-.11
-.23
-.16
-.05
-.14
-.13
-.07
-.14
-.07
Education
.07
.06
.16
.04
.05
.11
.04
.10
.06
.05
.07
.09
.11
.05
.07
Income decile
.08
.14
.11
.15
.09
.04
.09
.15
.08
.13
.05
.12
.08
.11
EU migrant
-.02
-.02
.05
-.01
.00
-.02
-.01
.02
-.04
.05
.01
.01
-.03
.00
.02
Other migrant
-.07
-.01
.05
-.02
.01
-.02
-.04
.01
.00
-.01
-.01
.00
-.03
-.01
.02
Lives alone
.05
.00
.07
-.05
-.04
-.06
.05
.00
.03
-.01
.01
.04
Lone-parent
.01
.02
.03
-.03
.02
-.02
-.01
-.04
.00
-.01
.00
-.01
Household size
.01
-.04
-.05
-.06
-.02
-.01
-.01
-.04
.02
-.01
-.02
-.03
.00
-.03
-.01
Rural
.09
.05
.00
.08
.05
.08
.01
.05
-.01
.04
.00
-.04
.01
BMI
-.04
-.02
-.02
-.03
-.06
-.01
-.05
-.01
-.01
-.06
-.01
-.02
.00
-.01
-.05
R2
.12
.29
.16
.13
.12
.14
.17
.21
.09
.16
.30
.10
.15
.10
October 2012
51
3.3.1 Beta coefficients
Table 17 presents the main results from the set of multivariate regressions we did for each
country. The top part of the table presents the standardised beta coefficients for each
variable for each country. Beta coefficients in black were significant at the 0.05 level, those
in grey were not significant. Those in bold were significant at the 0.01 level. Cells in grey
indicate that this variable was not available for the country in question. Note that many of
the variables included here are dummy variables derived from raw variables, as many of the
raw variables from the EHIS survey are categorical (e.g. marital status or employment
status). Dummy variables are coded as 1 for yes and 0 for no.
The final row of the table provides the R2 for regressions for each country, which will be
explored in section 3.3.2.
The first step in exploring this table is to look at which variables were significant predictors in
all countries; only one was: being permanently disabled. Other variables that were almost
always significant were being were gender (females had lower SF scores in all countries
except Slovakia, all other factors held constant), age (older people had significantly lower SF
scores in all countries except Cyprus) income decile (all countries except Greece), and
education level (all countries except Malta and Czech Republic). Of these, gender is worth
highlighting – in Task 7, when other control variables were included gender was not found to
significantly predict life satisfaction in the EQLS, though it did in the ESS.
Some variables for which mixed patterns were seen in the bivariate analyses continued to
have mixed patterns. For example, being divorced continued to have a positive impact on
well-being in Slovenia, all other things being equal, whilst it had a negative impact in several
other countries. Being retired had a positive impact in Spain, Turkey and Poland, but a
negative impact in Austria and Romania. Migrants still have higher well-being in Cyprus,
Turkey, and Malta, but lower in Austria, Bulgaria, France, Greece, and Latvia.
However, for other variables, the regression analyses changed the outcomes somewhat.
Whilst urban populations have significantly higher SF scores than rural ones in several
countries (according to planned comparisons in ANOVA), controlling for other variables
eliminates this difference. This result is consistent with the findings in Task 7, where holding
control variables constant, more (self-assessed) urbanity did only have a very marginal
negative effect on satisfaction. Indeed, the only significant differences seen resulting from
the urbanisation variable are now in favour of rural populations, which have higher well-
being scores in seven countries. It is likely that what is happening is that rural populations in
many countries may have lower incomes and education levels. Once one controls for this,
comparing rural populations with urban populations with similar incomes and education
levels, one can see that the rural populations actually have higher well-being.
Similarly the advantage seen for migrants in Spain and Hungary now disappears, suggesting
that their higher well-being was due to other confounding factors – perhaps economic status
or education level.
October 2012
52
On the other hand, some variables which had consistent results in bivariate analyses, now
presented mixed results. Living alone had a negative impact on well-being in all countries in
bivariate post-hoc analyses. In multivariate regression, the effect was negative in France,
Hungary and Greece, but actually positive in Bulgaria, Latvia, Poland and the Czech Republic.
It is possible that living alone in these countries is to some extent a proxy for wealth – a
luxury only possible for a few people. For example, household size was negatively correlated
with well-being in six countries, mostly former Communist.
Table 18 presents the top three significant effects in each country (based on beta
coefficient). As one can see the same effects appear repeatedly: age, gender, being disabled,
income and education. Being widowed appears once as a top three effect, as does living in a
rural location.
Table 18: Top three independent effects, by country
Country
Top 3 effects
1
2
3
Cyprus
Gender
Education
Disabled
Austria
Age
Disabled
Rural
Spain
Gender
Age
Disabled
Greece
Age
Gender
Disabled
Slovakia
Age
Widowed
Disabled
Slovenia
Age
Income
Education
Bulgaria
Age
Disabled
Income
Malta
Gender
Age
Income
Poland
Age
Disabled
Income
Latvia
Age
Disabled
Income
Czech Rep.
Disabled
Income
Gender
Romania
Age
Disabled
Education
Hungary
Disabled
Age
Gender
France
Disabled
Age
Gender
Turkey
Age
Income
Gender
By comparison, the top three beta coefficients in the EQLS in Task 7 were for variables for
income, marital status and being aged 35-49 (as opposed to being aged 18-24 years). In the
ESS they were income, being unemployed and being permanently sick/disabled. Gender was,
as noted before, insignificant as a predictor of life satisfaction in the EQLS, and only
significant at the 0.05 level in the ESS.
Table 19 presents the three countries where each effect is strongest (based on beta
coefficient). Where effects sometimes go in different directions, we’ve only presented the
effect where it can be seen as ’negative’ (e.g. countries where retired people have higher SF
scores are not listed). The countries appearing in these rankings most commonly are Greece,
Bulgaria, France and Latvia.
Malta, Slovenia, Slovakia, Poland and Turkey all appear only once.
Table 19: Countries where independent effects are strongest
Country
3 countries with biggest effect size
October 2012
53
1
2
3
Gender
Cyprus
Spain
Greece
Age
Bulgaria
Romania
Poland
Widowed
Slovakia
Czech Rep.
Bulgaria
Divorced
Greece
Latvia
Turkey
Unemployed
Latvia
Austria
France
Retired
Romania
Austria
-
Disabled
Hungary
Czech Rep.
Latvia
Education
Cyprus
Slovenia
France
Income decile
Czech Rep.
Latvia
Bulgaria
EU migrant
Latvia
Bulgaria
Austria
Other migrant
Austria
Greece
France
Lives alone
Hungary
France
Greece
Lone-parent
France
-
-
Household size
Cyprus
Bulgaria
Hungary
Rural/Urban
Austria
Greece
Czech Rep.
BMI
Spain
Malta
Greece
Note that these countries may be different from those identified as having the strongest
differences in the bivariate analyses in Section 2. This is for two reasons. Firstly, in the
regressions, other variables are controlled for at the same time, so one is able to look at the
independent effect of each variable. Secondly, and perhaps more subtly, Table 19 represents
information derived from beta coefficients which capture the tightness of a relationship,
whereas the data presented in Section 2 often focused on differences between groups. A
difference between two groups may be at the same time large but not represent a very tight
relationship.
3.3.2 Overall fit
Returning to Table 17, the final row (R2 values) provides an assessment of the proportion of
the variance in SF scores within each country that can be explained by the economic and
demographic variables. If R2 reaches 1.00 it means that all variance in SF scores can be
attributed to the variables analysed. The R2 is calculated excluding variables for which we did
not have data for all countries (i.e. urbanisation, household type, and size of social network)
with the exception of income decile. As we judged this variable to be particularly important
we preferred to include it and therefore exclude France (for which there were doubts about
representativeness) from this analysis.
Figure 18: Summary statistics of variance produced by economic and demographic
variables in different countries
October 2012
54
Figure 18 plots the R2 against a measure of the total amount of variance in each country
(coefficient of variation). One can see that whilst Hungary and Greece had the largest
coefficients of variation in total, it was in Bulgaria and Romania that the largest proportion
of variation is explained by demographics. In other words, whilst there is more well-being
inequality in Hungary, in Romania the well-being inequality is more associated with factors
such as age and income.
In 7 out of 14 countries, the R2 values found here are higher than those found in Task 7 (0.13
for the EQLS data and 0.12 for the ESS, in section 1.7.2 of the report of Task 7). This is
probably for two main reasons. Firstly, we have separated data out by country. In Task 7
respondents from all countries were pooled. It is likely that country of residence is an
important determinant of subjective well-being. Secondly, we have included several
variables not included in the analyses in 1.7.2, including urbanisation level, migrant status
and household size.
Having said that, in 7 other countries, the different R2s are equivalent to or lower than those
found in Task 7. It is hard to draw any firm conclusions on why this is the case, but it may be
due to the different measure of subjective well-being. Some studies have found evaluative
measures such as life satisfaction are more shaped by material conditions than hedonic
measures such as the SF-36.52 Indeed many of the most important determinants hedonic
well-being (health, social relationships, sense of community) were not included in these
regressions, explaining the low R2 values.
52 Diener E, Kahneman D, Tov W & Arora R (2010) ‘Income’s association with judgements of life versus feelings’ In E Diener, J
Helliwell, D Kahneman (eds) International Differences in Well-being (Oxford: Oxford University Press)
AT
BG
CY
CZ
ES
GR
HU
LV
MT
PL
RO
SI
SK
TR
,00
,05
,10
,15
,20
,25
,30
,35
,15 ,17 ,19 ,21 ,23
R2
Coefficient of Variation (total variation)
October 2012
55
4. Conclusions and recommendations
This report has presented data exploring inequalities in well-being in 15 European countries
based on data from the European Health Interview Survey. We have explored different ways
to operationalise that inequality, and the different economic and demographic factors which
appear to explain it.
Section 4.1 below highlights the key results from these analyses.
Section 4.2 presents recommendations to Eurostat on how well-being inequality should be
incorporated into the well-being indicator set. Note that these recommendations would
supplement Eurostat reporting income inequality, or other material inequalities (wealth
inequality being an area that deserves further exploration). As discussed in the Annex report
to Task 6, there is evidence that income inequality is related to average levels of well-being
within a country and that it could be seen as a “driver” of well-being.
Lastly, section 4.3 acknowledges some of the limitations of this work, with suggestions for
further work in the future.
4.1 Key findings
The countries with the highest inequality in well-being were Hungary and Greece, with
France, Bulgaria and Spain also regularly amongst the top in terms of inequality. Whilst
this well-being inequality was found alongside low mean well-being in Hungary and
France, mean well-being in Greece was actually high (4th highest out of 15 countries).
Whilst measures of well-being inequality intercorrelated highly, there were differences
depending on whether one used ratio measures, non-ratio measures or instrument-
effect-adjusted measures. Furthermore the 90/median percentile ratio behaved slightly
differently with Hungary and then Turkey presenting the greatest inequality. The only
well-being inequality metric to correlate negatively with mean was the 90/median ratio.
The instrument-effect-adjusted measures actually correlated positively with the mean.
It is important to note that well-being inequality and income inequality did not show the
same patterns. The countries identified to have high well-being inequality (Hungary,
France and Greece) did not have particularly high income inequality.
The Atkinson Index adjustment to the mean only made a slight difference to the rank
ordering of countries.
However, ranking countries according to the mean well-being of the people with the
lowest well-being does make a difference. Looking at these figures, Hungary, Greece and
Bulgaria do considerably worse than when looking at means for the overall population.
On the other hand, Slovakia, Slovenia and Malta do considerably better.
In general, the most important independent economic and demographic determinants of
well-being are age, gender, disability, income and education level, though these differ
October 2012
56
from country to country. Some of these match the findings from Task 7, with the addition
of age and gender likely to be a result of the different well-being measures used.
Looking at bivariate analyses, the countries with the biggest differences for these five
variables were:
o Gender – Spain
o Age – Bulgaria
o Disability – Cyprus
o Income – Latvia, Romania
o Education level – Romania, Hungary
Looking at regressions (controlling for other variables), the countries where these
variables had the strongest effects were:
o Gender – Cyprus
o Age – Bulgaria, Romania
o Disability – Hungary
o Income – Czech Republic, Latvia
o Education level – Cyprus
The countries with the highest well-being inequality in general were not necessarily those
with the biggest differences between economic or demographic groups. For example,
well-being inequality in Greece was the second highest amongst fifteen countries,
according to coefficient of variation, but the amount of variance attributable to these
variables was only fifth highest. Meanwhile, Romania had only the eighth highest
coefficient of variation, but differences associated with economics and demographics
were the second highest across the countries. As such, the regression to model well-
being based on economic and demographic variables had an R2 of 0.30 for Romania, but
only 0.14 for Greece (and only 0.09 for Malta). This means that 30% of variation in well-
being in Romania can be attributed to the variables we used for analysis, whereas only
14% of variation in well-being in Greece is associated with them – the remaining 86%
being related to other factors.
4.2 Recommendations
Well-being inequality is not the same as income inequality. Whilst rankings of well-being
inequality can be compared to rankings of income inequality, the best measures are not
the same, and the actual values of inequality for the two metrics should not be
compared.
October 2012
57
As a comprehensive measure of inequality, given the differences between the results for
the different measures, we only tentatively recommend the Mean Pair Distance. We
would suggest that Eurostat remain abreast of the debates within the academic
community before committing to any single measure. More research is also needed to
explore how people understand inequality in the context of well-being.
We do argue, however, that a non-comprehensive measure of inequality could be
deployed immediately: the S80/S20 difference, though it may need to be slightly
renamed.
Breakdowns for the well-being of the bottom half, 20% and 10% of the population are an
effective way to communicate dispersion of well-being in the bottom half of the
distribution.
The following variables are strongly related to well-being and, as such, breakdowns of
well-being should be reported using them:
o Gender
o Age (with the SF scores a downward trend was seen, though we know that with life
satisfaction there is a U-shaped curve in many countries)
o Employment status (particularly highlighting the unemployed, permanently disabled
and retired)
o Income
Other independent variables that were included in our analyses and would also be useful
for understanding well-being inequality include:
o Education level
o Urbanisation level (noting different patterns in different countries)
o Migrant status (although ethnicity may be more useful in many countries)
o Household type and size
Further independent variables that might be of value (not included in the current survey)
include:
o Ethnicity
o Working hours
o Occupation
As well as reporting means for different economic and demographic groups, Eurostat
should ensure that regression analyses are carried out to understand the way different
independent variables interact and play their role. The example of urbanisation is
particularly poignant – illustrating that, even in countries where rural populations have
October 2012
58
lower well-being, this is due to confounding variables, rather than the fact of living in a
rural area per se, which in fact appears to positively contribute to well-being.
Visualising the contrast between total variance/inequality within a country, and the
amount of variance that can be attributed to economic and demographic variables, may
be a useful way to conceptualise and understand inequality within a country.
4.3 Limitations and future work
We were only able to do this study using the SF36 mental health and vitality measures,
which have not been recommended in the well-being indicator set proposed by the
consortium. Whilst data from the SF is likely to correlate with other measures of positive
and negative affect, it is likely that a similar analysis using life satisfaction would produce
different results (for example the age pattern is different). We would advise replicating
this research with the life satisfaction item from the 2013 SILC Well-Being module.
Unfortunately, the first wave of the EHIS did not achieve much take up in Western and
Northern Europe and, where it did, there were problems meaning that we were unable
to analyse data for those countries. In the end, the only EU15 countries that we analysed
were Austria, France, Greece, and Spain – and for France there were limitations to the
data. It is advisable to replicate the research when data from more Western European
countries becomes available, either through the next wave of EHIS, or the 2013 SILC Well-
Being module.
The scope of this project has not allowed us to investigate the harmonisation of the EHIS.
Given that countries had the flexibility of deploying items from the EHIS in different
surveys, it is possible that different survey effects influenced the results of the SF
questions in different countries. Further study would be necessary to confirm that this
did not have an undue effect on results.
There is an extensive literature in health inequality which we touched upon in Section 1,
but were unable to do full justice to. We would recommend commissioning experts on
health inequality to produce a review of that field so as to provide lessons for well-being
inequality.
Also, as mentioned in 4.2, Eurostat should keep abreast of the nascent literature on well-
being inequality with a view to identifying the best univariate measures. It may be
necessary for research to explore how people understand inequality in well-being – for
example do people see a difference of 2 on a 0-10 scale to be the same if that difference
occurs at the top of the scale (e.g. between 7 and 9) or the bottom of the scale (e.g.
between 1 and 3)?
Task 1 of this Implementation Study explores how cultural biases might influence
responses to subjective well-being questions. One of the biases identified – Extreme
Response Style (ERS) – could be being conflated with well-being inequality. Indeed,
looking at the map of ERS in Europe based on the European Social Survey (page 53 of the
Interim Report), it is worth noting that some of the countries with the strongest ERS are
also countries which we have identified here as showing high well-being inequality –
October 2012
59
particularly Hungary and France. Other studies cited in that report found high ERS in
Mediterranean countries such as Greece and Spain. However, one of the countries with
the lowest well-being inequality in the EHIS was Austria, which was also found to have
high ERS in Task 1. Further work should test the impact of controlling for ERS, by taking
responses to unrelated questions. This would not have been possible using the EHIS
questionnaire as there were no unrelated questions with a similar response scale to the
SF.
October 2012
60
Annex 1- Table of potential data sets for analysis
The table considers the following five data sets:
SILC (EU Survey of Individual Living Conditions) – 2009, with module on material
deprivation, most relevant
EHIS (European Health Interview Survey) – wave 1 (2008/09)
EQLS (European Quality of Life Survey) – wave 2 (2007)
ESS (European Social Survey) – wave 5 (2010) is latest, with extra well-being data in
wave 3 (2006)
IHS (UK Integrated Household Survey) – 2011/12
Data set
Countries
Sample size
Demographics
Subjective Well-Being
SILC (2009 best)
All EU
Large – > 270,000
All (maybe not
urban)
None
EHIS (wave 1)
15* EU countries
Large: 2,000 - 35,000
per country
All (inc urban)
SF-10
EQLS (wave 2)
All EU + 4 others
Small ~ 35,000
All (inc urban)
Life satisfaction
ESS (wave 3 or 5)
20 EU countries +
6 others
Small ~ 50,000
All (inc urban)
Life satisfaction
IHS
UK only
Large – 200,000
All (but not urban)
Life satisfaction
For each criterion, we have used red, yellow and green to summarise our evaluation.
* Wave 1 of the EHIS was actually administered in 18 countries. However, for 2
countries (Belgium and Estonia), subjective well-being questions were not asked; micro-data
is not available for Germany.
October 2012
61
Annex 2 - SF-36
The SF-36 was developed by John E. Ware and colleagues at the RAND Corporation in the
early 1990s, and is now owned by Quality Metric Inc. It ‘was designed as a generic indicator
of health status for use in population surveys and evaluative studies of health policy’,53 and
is itself based on a range of pre-existing scales. As well as an overall score, users can
calculate a set of scales for different aspects of health using date from the tool. In 1996,
Version 2.0 of the SF-36 was introduced, which included several changes in questionnaire-
presentation, question-wording and, in response categories.
The SF-36 is widely used. By 1992, data from 1 million respondents had been collected with
the tool, even before it had been validated. The producers report around 4,000 publications
citing the scale. The reliability and validity of the tool has been demonstrated in several
studies.54 For example, test-retest correlations for the separate scales of the tool range from
0.6 to 0.9. Meanwhile scores on the SF-36 predict later visits to doctors, hospital admissions
and mortality over the following four years.55
53 McDowell I, Newell C. Measuring Health: A Guide to Rating Scales and Questionnaires (second edition). New York: Oxford
University Press, 1996.
54 Ibid.
55 McHorney CA. Measuring and monitoring general health status in elderly persons: practical and methodological issues using
the SF-36 Health Survey. Gerontologist 1996; 36:571–583.
