You are on page 1of 25

Inequality and

Human
Development

Maria Ana Lugo


University of Oxford

20th Sept 2004


maria.lugo@economics.ox.ac.uk
Purpose
 To present key issues (debates) around
inequality

 To explore methodologies to compare


distributions

 To analyse how much HDI incorporates


(or can incorporate) inequality

2
Outline

1. Initial Remarks – definitions

2. Issues on Inequality

3. Why does inequality matter?

4. Methods

5. Inequality and HDI

6. Example (Arg)
3
1. Initial Remarks
“Inequality is like an elephant: You can’t
define it, but you know it when you see it”
(Fields 2001, 14)

 The meaning of inequality: encapsulating ethical concepts vs.


statistical dispersion (distribution rather than inequality)

 Inequality, poverty and welfare. Related but distinct


 broader than poverty – defined over the whole distribution, not only
below a certain poverty line
 narrower than welfare – if indep of mean, only concerned with the
second moment (relative vs absolute measures)
 But 3 closely related, sometimes used in composite measures (e.g.
Sen’s measure contains the Gini among the poor)

4
2. Issues on Inequality
 WHY should we care about inequality?
 HOW do we compare distributions? (we always
compare!)
 Empirical issues (dimensions, variables, unit of
analysis, methods -indices and orderings-,
between/within inequality)
 WHY is there inequality? [ineq as ‘output]
 Theories to explain determinants: factor share and
market imperfections theories – especially LM i.e. info
asymmetries, efficiency wages, selection models
 Tools: decompositions, vertical vs. horizontal inequality
 WHAT are the effects of inequality? [ineq as
‘input’]
 On growth, HD, poverty, conflict, democracy, etc
5
3. Why does inequality International unw International w World

matter?
Source National accounts National accounts Household surveys
Unit of analysis Country Country Individuals
“.. every normative theory of social
Welfare concept GDPpc GDPpc Mean pc income
arrangement
Inequality
National currency
 that has
good Marketat
or all
ER or PPP stood
bad ER
for the test of
Within -country
(income)
Ignoredtime growth?
seems to demand
Ignored equality of
Included
 “Because it isMeasure:
high something
and
relative – something-that
increasing!”
or absolute is regarded
Magnitude ofas
INEQ-Levels: Kuznets inverted
particularly important on that theory” ‘U’ hyp
the problem INEQ to Growth: HK Loss (-), Access to K
(Sen 1992,(+),
12)
mkt (+/-), Efficiency wages (-), Savings
 Maps ‘equality political of instability (-)
what?’ [space of equality:
 World inequalityUtilitarians,
… increasing?
 Human Rawlsians,
Development: Nozicksean, and
effects on levels
Senians] (of edu, health, others) achieved, political
 International: between countries
egalitarian participation,in some space, anti-egalitarian in
etc.
some
 National: Within other space.
country-inequality
 Poverty: sensitivity of poverty to growth,
 World: between +effect within of specific policies
 Social cohesion (Stewart):
 “Because it is fair” Horizontal
- TheoriesIneq
of justice
(vs vertical) culturally
 defined groups => social stability,
Sense of justice and self-worth
instrumental and direct welfare reasons
 “Because it affects dev/growth/HD/ democracy /
social cohesion”
 Mechanisms from inequality to variables 6
4. How do we compare two
distributions?
“It is generally agreed that, other things being equal, a considerable
reduction in inequality of incomes found in most modern communities
would be desirable.
But it is not generally agreed how this inequality should be measured”
Hugh Dalton (1920)
“The measurement of the inequality of incomes”
Economic Journal, 1920, Vol. 30, p. 348

 Ineq of what and among whom?


 Variable/Indicator: one or many? Which one(s)? If income,
pre/post taxes; includes profits of capital and land or not.
 Unit of analysis: individuals, households, regions, countries
(≠ recipient unit)
 Adjustments/corrections: e.g. adult equivalent (eq scales),
economies of scales, cost-of-living differentials (rural-urban,
diff regions) 7
4. How do we compare two
distributions?
One ‘x’
or
many ‘xs’?

UNI- MULTI-
DIMENSIONAL DIMENSIONAL

1. Indices 2. Orderings
Item-by-Item Aggregative Non-aggreg
(real #) (CDFs, LC)

2 step MD Indices

8
4. How do we compare two
distributions?
UNIDIMENSIONAL
1. Indices
E.g. Gini coefficient, Theil index, Atkinson Index, 90-10 Ratio

Which one shall we use?


How do we choose between indices?

 Axiomatic approach: desirable properties


 Anonymity, Scale Independence, Population Independence,
Transfer Principle (Pigou-Dalton) => M of Relative Inequality
 Others: Decomposability, Transfer Sensitivity
 ‘Ad hoc’:
 mathematical appeal (GE)
 neat 9
statistical or graphical interpretation (Gini)
4. How do we compare two
distributions?

Some measures
 Gini Coefficient: most widely use inequality measure.
1 n n
Gini = 2 ∑∑ yi − yj
2n µ ( y) i =1 j =1

- Function of differences between every pair of inds’ income


- Ranges from 0 to 1

- It is a relative measure

- Has a close relation to Lorenz Curves => clear representation

- Can be easily computed from normally available data (%pop/

%income)
But
- Not decomposable (between- and within-groups)

10
4. How do we compare two
distributions?
Other two relative measures
 GE measures: from Theory of Information

α
1 1   xi  
GEα = ∑ 1−   
α (1− α ) n   µ (x )  

α weight to changes at different parts of the distribution. The lower the α,


the more sensitive to the lower part.
- When α=0 Theil’s mean log; When α=1 Theil index (equal weight);

 Atkinson index: includes subjective elements


1/ (1−ε )
 1  x 1−ε 
I ε = 1−  ∑  i  
 n  µ (x )  

ε: parameter of inequality aversion, explicit choice of weights to changes at


different parts of the distribution. The higher the ε, the more weight to
the lower part 11
4. How do we compare two
distributions?
Formula Mean Popindep Pigou- Decomposability Transfer sensitivity. Others
Indep Dalton
Range Max Yi − Min Yi • • Ignores inside the extremes
R= X
µ
Variance
V=
∑ (Y i − µ) 2 X • •

n
RelativeMean Deviation • • Not sensitive to transfers among same
M=
1
∑ µ − Yi X side of the mean income.

Gini Coefficient • • • Sensitive to centre of distrib –depends
G=
1
2n 2 µ
∑ j ∑ i Yi − Y j X on # people between. Not incomes.

Standard deviation of logs • • • More sensitive for transfer in the lower


X (might
1/ 2
  log µ − log Y 
2

H = ∑    end. Not defined for zero income
  n 
  violatefor very earnings

high income)

GE measures • • • • Depending on α
1 1  Y 
α

GE = ∑ 1 − 
α (1 − α ) n   Yi
?
 

 
Theil second α = 0 1 Y • • • • Very sensitive to lower end. Not
L=
n
∑ log
Yi
defined for zero income values

Theil index α = 1 • • • • Sensitive to lower end


1 Yi Y
T=
n
∑ Y
log i
Y
Coefficient of Variation α = 2 1 V • • • • Transfer neutrality, equally sensitive to
= ( CV ) ; CV =
2
GE2 all income transfers
2 µ

Atkinson (ε) 1/ ε • • • Depending on ε

A?
 1  Yi  ε 
=1 −  ∑   
X
 n  Y  
12
4. How do we compare two
distributions?
1
2.Orderings (C»D) Φ(y)

Lorenz Curves (1905) Percentage


of Income
Line of Perfect
Equality (LPE) A

A
Gini =
A +B B
Lorenz Curve
L(y) - LC

0 Percentage of Population 1
F (y)

Atkinson (1970): LC if do not cross, then any measure


satisfying above axioms would rank distributions equally
=> we do not have to worry about which one to use
13
4. How do we compare two
distributions?
 But if they cross? => We cannot conclude anything
unambiguously (incomplete ordering)

- Either we choose some measure of inequality


=> different measures might give us different answers, in fact,
we are always able to find 2 measure that disagree!!

- Or we leave the analysis at this point


Is this bad? Sen argues instead of this being a drawback of LD
criterion, it is in fact its greatest advantage, as it captures the
essential ambiguity of the concept of inequality

“Ambiguity of inequality should be preserved rather than


trying to remove it through some arbitrary completed
ordering” (Sen 1997, p. 121)

NB: LC for continuous (‘distributable’) unbounded variables 14


4. How do we compare two
distributions?
MULTIDIMENSIONAL
1. Item-by-item: each vle examined separately
UD methods (indices/orderings) + covariance/correlation analyses

2. Aggregative strategies: Collapsing dimensions


 Two-step procedure: (1) well-being index (2) UD methods
 Multidimensional measures of inequality
=> Complete ordering of distributions.
Decisions over: weighting (w), degree of substitution bt dimensions (β),
and degree of inequality aversion.
HDI: chooses w: [1/3, 1/3, 1/3] and β: 1 for all pair of attributes

3. Non-aggregative strategies:
Tests comparing MD distributions, as in UD ordering: Kolmogorov-
Smirnov and Pearson’s tests. Not easy to implement, few studies 15
4. How do we compare two
distributions?
Advantages Disadvantages
ITEM-BY-ITEM  No problem of aggregation Not easy to compare if
(commensurability, choice of #dimensions increase
w, β) No account of MD aspect
at unit level (only corr)

 Summarises info, easy to Have to impose


AGGREGATIVE 
compare bt units assumptions (values) on
 Gives complete ordering w, β, α
 Commensurability
(units), types of vles

NON-AGGREG  Dealswith MD of unit  Difficultto implement


 Not many assumptions of  Problem of high # obs
parameters required required
 Most probable, leads to
ambiguous conclusions 16
5. HDI and Inequality
 Between-country inequality: HDI as a 1st step,
then apply UD indices

 Within-country inequality: HDI components are


national averages => not included

“An intrinsic concern for inequality in human development requires


adjustment of average achievements in each of the three dimensions
(by the extent of measured inequality in each) – as well as an
accounting of the covariances in achievements along the different
dimensions” (Anand 2000: 99)

Only GDI – including gender differentials


1991-94 HDR include Yadj=Y(1 – G)
17
5. HDI and Inequality
 Should we include w-c inequality?
 Efficiency arguments: variables as means to other
ends, in a diminishing way. That’s why income in logs

 Equity arguments: intrinsically valuable


 Basic problem: type of indicators in HDI
- LE is probability-based group measure, not individual
- Literacy: group measure (%) or individual (0-1) then no
efficiency argument for a 0-1 variable
⇒ Change indicators (for a continuous unbounded)
⇒ Sub-group inequality (gender, location, race) – municipalities
[though same for 3 attributes!] 18
5. HDI and Inequality
How shall we include inequality?
 Hicks (1997): Extending HDR91-93 income (Sen Index)
Each x dimension
IA x = µ ( x)* λx[ 1 − Gx ]

Gx: Gini coefficient for component x


Edu: years of schooling / Health: life-span attainment (age-at-death)

BUT:
- It is not group-consistent/decomposable, bc. Gini is not
- It depends on the ordering of aggregation across people and
dimensions
- Also, Anand-Sen’s critique: interpretation of reductions 19
5. HDI and Inequality
 Foster (2003): ‘General Mean of Means’ (Atkinson Index)
1
∑x
1−ε
i  1−ε

Each x dimension µ1−ε (x )= µ(x )[ 1− I ε(x )] = i 


 n 
 

Aggregate components H ε (D )= µ1− ε [ µ1− ε (x ), µ1− ε ( y), µ1− ε (z )]

- Choose ε according to concern for inequality –higher ε more weight


to lower part of distribution (more aversion to inequality)
- Group-decomposable and sequencing-free
- Sensitive to inequality across dimensions (penalising uneven
development, there is some substitutability but it’s not infinite) 20
5. HDI and Inequality
Related questions
- Should ε be the same for all dimensions? (see
Anand 2003) – bc of nature of the dimension and
of the indicator used
- How should be choose ε?
- Substitutability between inequality across
dimensions, but not between levels of
components (valid ? also for HDI).

21
6. Exercise – Argentine
provinces
 Objective: Do multivariate distribution analyses contribute to
the study of Argentine welfare inequality?
 Exercise
 Period: 1991 vs. 2001
 Dimensions:

Dimension

 Income
Units of analysis: provinces
Weights: one province one observation (as in HDI!)

22
6. Results
Item- by-item Income µ1991 f µ2001
Stoch dom and measures 1991 f 2001
Health(LE) µ2001 f µ1991
Stoch dom and measures 2001 f 1991
Education µ2001 f µ1991
Stoch dom and measures 2001 f 1991

 Income distr worsen


 Health and educ distr improved
NB: Low values of ineq (not sign different). Between-ineq is relevant

Aggregative Maasoumi Stoch dom and measures 2001 f 1991


(1 w, 4 β, 2 α)
Strategies  Different result from income distribution
Bourguignon Depends on values of α and β
(1 w, 4 β, range0-1 α) if β ? ½ => 2001 f 1991
if β < ½ => 1991 f 2001
 Two opposite results 23
Conclusions
 There are enough reasons to include distribution (inequality)
issues in the analysis. But not consensus on how, even within
UD.
 Be specific, demand specificity! “Definitional issues are too
serious to be left to footnotes or ignore altogether” (AB 2000:
14) Inequality OF WHAT (vle) AMONG WHOM (unit)?
 Atkinson’s 1st result: All measures imply assumptions
(embedded SWF) – important to make them explicit and
understand them
 Different measures might give different orderings of
distributions =>
 Atkinson’s 2nd result: Use LC criterion when possible
 Test robustness of results (diff measures/parameters)
 Or use one and give reason for choosing it over the others
24
Conclusions
 Multidimensional inequality: three strategies with adv and
disadv.
 Also they imply taking many decisions
“There is an inescapably arbitrariness in the choice of [β]. The right
way to deal with the issue is to explain clearly what is being assumed
… so that public criticism of the assumption is possible” (Anand 2003:
218)
Better than neglecting the issue altogether, as generally done,
and focus solely on income distribution
⇒ Again, use sensitivity analysis, results under different
assumptions
 Statistical Inference: for data derived from surveys,
meaningful comparisons between estimates need include
standard errors of the measures (not usually done) 25

You might also like