The econometrics of inequality and poverty

Lecture 4: Lorenz curves, the Gini coefficient and

parametric distributions
Michel Lubrano
September 2015

Contents
1 Introduction

2 General notions
2.1 Distributions . . . . . . . . . . . .
2.2 Densities . . . . . . . . . . . . . .
2.3 Quantiles . . . . . . . . . . . . . .
2.4 Some useful math results . . . . . .
2.5 Truncated distributions and moments

.
.
.
.
.

3
3
4
5
6
6

3 Lorenz curves
3.1 A partial moment function . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 A mathematical characterisation . . . . . . . . . . . . . . . . . . . . . . . . . .

8
8
8
9

4 The Gini coefficient revisited

4.1 Gini coefficient as a surface . . . . .
4.2 Gini as a covariance . . . . . . . . .
4.3 S-Gini . . . . . . . . . . . . . . . .
4.4 Gini as mean of absolute differences

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.

11
11
11
12
13

5 Estimation of the Gini coefficient

14
5.1 Numerical evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
5.2 Inference for the Gini coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . 15

6 Lorenz curve and other inequality measures

15
6.1 Schutz or Pietra index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
6.2 Other inequality measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
7 Main parametric distributions and their properties
7.1 The Pareto distribution . . . . . . . . . . . . . .
7.2 LogNormal distribution . . . . . . . . . . . . . .
7.3 Singh-Maddala distribution . . . . . . . . . . . .
7.4 Weibull distribution . . . . . . . . . . . . . . . .
7.5 Gamma distribution . . . . . . . . . . . . . . . .
7.6 Double Pareto . . . . . . . . . . . . . . . . . . .
7.7 Which density should we select? . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

17
17
21
26
29
30
32
32

8 Pigou-Dalton transfers and Lorenz ordering

32
8.1 Lorenz Ordering for usual distributions . . . . . . . . . . . . . . . . . . . . . . 34
8.2 Generalised Lorenz Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
9 Parametric Lorenz curves

35

10 Exercises
10.1 Empirics . . . .
10.2 Gini coefficient
10.3 LogNormal . .
10.4 Uniform . . . .
10.5 Singh-Maddala
10.6 Logistic . . . .
10.7 Weibull . . . .

38
38
38
38
38
39
39
39

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

A Hypergeometric series
41
A.1 The family of the generalised Beta II . . . . . . . . . . . . . . . . . . . . . . . . 42
A.2 The Burr system as a useful particular case . . . . . . . . . . . . . . . . . . . . . 43
A.3 Playing with the Burr system . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

1 Introduction
Some authors like Sen (1976) prefer to use a discrete representation of income, which is based
on the assumption that the population is finite. Atkinson (1970) and his followers prefer to
suppose that income is a continuous variable. It implies that the population is implicitly infinite,
but the sample can be finite. Discrete variables and finite population are at first easy notions to
understand while continuous variables and infinite population are more difficult to accept. But as
far as computations and derivations are concerned, continuous variables lead to integral calculus
which is an easy topic once we know some elementary theorems. Considering a continuous
random variable opens the way for considering special parametric densities such as the Pareto
2

or the lognormal which have played an important role in studying income distribution. Discrete
mathematics are quite complicated.

2 General notions
We are interesting in the income distribution. Income is supposed to be a continuous random
variable with distribution F (.).

2.1 Distributions
Definition 1 The distribution function F (x) gives the proportion of individuals of the population
having a standard of living below or equal to x.
F is a non-decreasing function of its argument x. We suppose that F (0) = 0 while F () = 1.
F (x) gives the percentage of individual with an income below x. We usually call p that proportion.
A natural estimator is obtained for F (.) by considering
n
1X

F (x) =
1I(xi x).
n 1=1

where 1I(.) is the indicator function. This estimator is easy to implement. The resulting graph of
the density might seem discontinuous for very small sample sizes, but get rapidly smoother as
soon as n > 30. So there is in general no need for non-parametric smoothing.
Let us now order the observations by increasing order from the smallest to the largest and
call x[j] the observation which has rank j. We can write the natural estimator of F as
F (x[j] ) = j/n.
It is common to call x[j] an order statistics. They will play an important role for estimation. We
can give a short example written in R. We draw two samples from a normal distribution of size
n = 100 and then n = 1000.
n = 100
xr = rnorm(n)
x = sort(xr)
y = seq(0,1,length=n)
plot(x,y,type="l",xlim=c(-3,3),ylab="Cummulative",xlab="X")
n = 1000
xr = rnorm(n)
x = sort(xr)
y = seq(0,1,length=n)
lines(x,y,col=2)
3

1.0
0.8
0.6
0.4
0.0

0.2

Cummulative

n=1000
n=100

Figure 1: Natural estimator of the cumulative distribution for a Gaussian random variable
The abscises of the cumulative are obtained by ordering the draws, while the ordinates are simply
an ordered index between 0 and 1. The curve in black corresponds to the sample of size 100. It
is rather rough. But the curve in red, corresponding to a sample of 1000 is perfectly smooth.

2.2 Densities
We shall suppose that F is continuously differentiable so that there exist a density defined by
f (x) = F 0 (x).
So, for a given x, the value of p such that X < x can be defined alternatively as
p=

f (t) dt = F (x).

Densities are much more complicated to estimate. There exist no natural estimator as for distributions, simply because F (.) is not differentiable. Some kind of non-parametric smoothing is
needed. Non-parametric density estimation will be detailed in Lecture 6, Modelling the Income
Distribution.
If f (x) is the density, then the probability that the random variable X belongs to the interval
[xk1 , xk ] is given by
p(xk1 < x < xk ) ' f (xk )xk
4

where xk = xk xk1 . If the interval [a, b] is sliced in m smaller slices, then

m
X

p(a < x < b) '

f (xk )xk .

k=2

If we take the limit for m , we have

p(a < x < b) ' lim

m
X

f (xk )xk =

k=2

b
a

f (x)dx.

Of course this limit exists only if F (.) is sufficiently smooth, i.e. it has no jumps or kinks.

2.3 Quantiles
Once a distribution is given, it is always possible to compute its quantiles (this is not the case
for moments that exists only under specific conditions). Deciles are a convenient way of slicing
a distribution in intervals of equal probability, each interval being of probability 1/10. More
generally, a quantile is a function x = q(p) that gives the value of x such that F (x) = p.
Quantiles are implicitly defined by the relation
x = q(p) = F 1 (p).
q(p) is thus the living standard level below which we find a proportion p of the population. The
median of a population is the value of x such that half of the population is below x and half of the
population is above x = q(0.50). Using quantiles is also a way to normalize the characteristics
of a population between 0 and 1. This facilitates comparisons between two populations, ignoring
thus scale problems.
Quantiles are rather easy to estimate once we know the order statistics. Suppose that we have
an ordered sample of size n. The estimator of a quantile comes directly from the natural estimator
of the distribution. The p quantile is simply the observation that has rank [p n]. Quantiles are
directly estimated in R using the instruction
quantile(x,p),
where x is a vector containing the sample and p the level of the quantile.
Piketty (2000) in his book on the history of high incomes in France makes an extensive use
of quantiles to study the French income distribution and particularly its right tail concerning
high incomes. High incomes concern the last decile, which means q0.90 . That decile however
cover a variety of situations where wages, mixed incomes and capital incomes have a varying
importance. The interval q0.90 q0.95 concerns what he calls the middle class, formed mainly
by salaried executives. The interval q0.95 q0.99 is the upper middle class, formed mainly by
holder of intermediate incomes like layers, doctors. The really rich persons correspond to the
last centile q0.99 and over. It corresponds to holders of capital income.

2.4 Some useful math results

Three main rules are important to understand the next coming calculations:
1. Integration by parts. It comes from the rule giving the derivative of a product of two
functions of x, u(x) and v(x):
(uv)0 = u0v + uv 0 .
Let us take the integral of this expression.
uv =

u0v du +

uv 0 dv.

We deduce the integration by part formula by simply rearranging the terms:

Z

u v du = uv

uv 0 dv.

2. Compound derivatives.
f (u(x))/x = f 0 (u(x))u0(x).
3. Change of variable and densities. Let x f (x) and a transformation y = h(x) with
inverse x = g(y). Then the density of y is given by
(y) = |J(x y)|f (g(y)),
where J is the absolute value of the Jacobian of the transformation
J(x y) = |xi /yi |.
4. Change of variable and integrals. Consider the integral
Z

f (x) dx

and the change of variable x = h(u) with reciprocal u = h1 (x). Then the original integral
can be expressed as
Z

f (x) dx =

h1 (b)
h1 (a)

f [h(u)] h0(u) du.

2.5 Truncated distributions and moments

We are now going to investigate a particular given interval of the income distribution and see
how various quantities can be computed for this interval, such as for instance the mean. As a
by-product, we shall obtain useful formulae to define the Lorenz curve in the next subsection.
We thus start from an income distribution with continuous density f (x) and consider an income
interval [a, b]. We first compute the number of individuals inside that cluster:
H(a, b) = n

f (x)dx.

where n is the size of the total population. The average standard of living in the total population
is given by the total mean
Z
Z

xdF (x) =

xf (x)dx.

The total income of the population inside that cluster [a, b] is

X(a, b) = n

xf (x)dx,

while the average standard of living of a cluster is obtained as a ratio

n

X(a, b)

H(a, b)

xf (x)dx
a
=
.
Z b
Z b
n
f (x)dx
f (x)dx
a

xf (x)dx

When b tends to infinity and a to 0, we recover the average standard of living of the entire
population.
We now consider a threshold z and the population which is below that threshold, sometimes
the population over that threshold. Starting from the previous equation, X(a, b)/H(a, b), we
can compute the average standard of living of the first group, the one which is below z. This is
equivalent to the expectation of a truncated distribution
Z

1 =

z
0

xf (x)dx
.

F (z)

Using integration by parts with u = x and v 0 = f (x), we can rewrite the integral in the numerator
as:
Z z
Z z
xf (x)dx = [xF (x)]z0
F (x)dx
0

Noting that z =

Rz
0

= zF (z)

dx, it comes that

1 =

xf (x)dx
=

F (z)

F (x)dx.

F (x)
F (z)

dx.

Incidently, if we now let z tend to infinity, we arrive at an alternative expression for the the mean
=

[1 F (x)]dx.

Note also that another expression of the mean can be obtained as follows, using the quantiles.
Start from
Z
=
xf (x)dx.
0

By the change of variable x = F

=

(p) and p = F (x), we have dp = f (x)dx and thus

xf (x)dx =

1
0

F 1 (p)dp =

This expression will be used for explaining the Lorenz curve.

7

q(p)dp.

3 Lorenz curves
The Lorenz curve is a graphical representation of the cumulative income distribution. It shows
for the bottom p1 % of households, what percentage p2 % of the total income they have. The
percentage of households is plotted on the xaxis, the percentage of income on the yaxis. It
was developed by Max O. Lorenz in 1905 for representing inequality in the wealth distribution.
As a matter of fact, if p1 = p2 , the Lorenz curve is a straight line which says for instance that
50% of the households have 50% of the total income. Thus the straight line represents perfect
equality. And any departure from this 45 line represents inequality.

3.1 A partial moment function

The standard definition of the Lorenz curve is in term of two equations. First, one has to determine a particular quantile, which means solving for z the equation
p = F (z) =
and then write

f (t)dt

1 z
L(p) =
t f (t) dt.
0
So the Lorenz curve is an unscaled partial moment function. Unscaled, because it is not divided
by F (z).
A notation popularised by Gastwirth (1971) used the fact that z = F 1 (p) to write the Lorenz
curve in a direct way, using a change of variable:
Z

L(p) =
Alternatively, the relation =

R1
0

1Z

q(t) dt =

1Z

F 1 (t) dt.

q(t) dt, we can have another writing

Z

L(p) = Z01
0

q(t) dt
.
q(t) dt

The numerator sums the incomes of the bottom p proportion of the population. The denominator
sums the incomes of all the population. L(p) thus indicates the cumulative percentage of total
income held by a cumulative proportion p of the population, when individuals are ordered in
increasing income values.

3.2 Properties
The Lorenz curve has several interesting properties.
8

1. It is entirely contained into a square, because p is defined over [0,1] and L(p) is at value
also in [0,1]. Both the xaxis and the yaxis are percentages.
2. The Lorenz curve is not defined is is either 0 or .
3. If the underlying variable is positive and has a density, the Lorenz curve is a continuous
function. It is always below the 45 line or equal to it.
4. L(p) is an increasing convex function of p. Its first derivative
dL(p)
q(p)
x
=
= with x = F 1 (p)
dp

is always positive as incomes are positive. And so is its second order derivative (convexity).
The Lorenz curve is convex in p, since as p increases, the new incomes that are being
added up are greater than those that have already been counted. (Mathematically, a curve
is convex when its second derivative is positive).
5. The Lorenz curve is invariant with positive scaling. X and cX have the same Lorenz curve.
6. The mean income in the population is found at that percentile at which the slope of L(p)
equals 1, that is, where q(p) = and thus at percentile F () (as shown on Figure 2). This
can be shown easily because the first derivative of the Lorenz curve is equal to x/.
7. The median as a percentage of the mean is given by the slope of the Lorenz curve at
p = 0.5. Since many distributions of incomes are skewed to the right, the mean often
exceeds the median and q(p = 0.5)/ will typically be less than one.
The convexity of the Lorenz curve is revealing of the density of incomes at various percentiles. The larger the density of income f (q(p)) at a quantile q(p), the less convex the Lorenz
curve at L(p). On Figure 2, the density is thus visibly larger for lower values of p since this is
where the slope of the L(p) changes less rapidly as p increases.
By observing the slope of the Lorenz curve at a particular value of p, we know the pquantile
relative to the mean, or, in other words, the income of an individual at rank p as a proportion of
the mean income. An example of this can be seen on Figure 2 for p = 0.5. The slope of L(p) at
that point is q(0.5)/, the ratio of the median to the mean. The slope of L(p) thus portrays the
whole distribution of mean-normalised incomes.

3.3 A mathematical characterisation

Lorenz curves were defined by reference to a given distribution function F (.). Is it possible to
characterise a Lorenz curve directly, without making reference to a particular distribution? Let
us consider directly the expression of function that we consider to be a potential Lorenz curve.
In this case, this curve has to verify some properties in order to be a true Lorenz curve. From
Sarabia (2008), we have a first theorem:
9

Figure 2: Lorenz curve (source Duclos and Araar 2006)

Theorem 1 Suppose L(p) is defined and continuous on [0,1] with second derivative L00 (p). The
function L(p) is a Lorenz curve if and only if L(0) = 0, L(1) = 1, L0 (0+) 0, L00 (p) 0 in
(0,1).
If a curve is a Lorenz curve, it determines the distribution of X up to a scale factor which is the
mean . How could we find it? Let us take the definition of the Lorenz curve
Z
1 p 1
LX (p) =
F (t) dt
X 0 X
and express it as:
Z x
L(F (x)) =
ydF (y)).
0

Let us differentiate it using the derivative of a compound function:

L0 (F (x))f (x) = xf (x).
We simplify by f (x) and derivate it a second time so that
L00 (F (x))f (x) = 1.

We get the following theorem from Sarabia (2008):

Theorem 2 If L00 (p) exists and is positive everywhere in an interval (x1 , x2 ), then FX has a finite
0
positive density in the interval (L0 (x+
1 ), L (x2 )) which is given by
1
.
fX (x) =
00
L (FX (x))
10

4 The Gini coefficient revisited

In this section, we shall suppose that the mean of F exists. As a consequence
lim tF (t) = lim t(1 F (t)) = 0,
t

t0

which simplifies greatly the computation of some integrals when considering an infinite bound.
The next computations owe to the survey of Yitzhaki (1998) and to that of Xu (2003).

4.1 Gini coefficient as a surface

If everybody had the same income, the cumulative percentage of total income held by any bottom
proportion p of the population would also be p. The Lorenz curve would then be L(p) = p:
population shares and shares of total income would be identical. A useful informational content
of a Lorenz curve is thus its distance, p L(p), from the line of perfect equality in income.
Compared to perfect equality, inequality removes a proportion p L(p) of total income from the
bottom 100.p% of the population. The larger that deficit, the larger the inequality of income.
There is thus an interest in computing the average distance between these two curves or the
surface between the diagonal p and the Lorenz curve L(p). We know that the Lorenz curve is
contained in the unit square having a normalized surface of 1. The surface of the lower triangle is
1/2. If we want to obtain a coefficient at values between 0 and 1, we must take twice the integral
of p L(p), i.e.:
G=2

(p L(p)) dp = 1 2

L(p)dp,

which is nothing but the usual Gini coefficient. Xu (2003) gives a good account of the algebra of
the Gini index. We have given above an interpretation of the Gini index as a surface. The initial
definition we gave was in term of a mean of absolute differences in the previous chapter. There
are other formula too. All of these formula are equivalent. We have to prove this. A large survey
of the literature can also be found in the article Gini coefficient of Wikipedia.

4.2 Gini as a covariance

Let us us start from the above definition of the Gini coefficient and use integration by parts with
u0 = 1 and v = L(p). Then
G = 12
= 1

L(p)dp

2 [pL(p)]10
Z 1

= 1 + 2

+2

pL0 (p) dp

pL0 (p) dp.

We are then going to apply a change of variable p = F (y) and use the fact proved above that
L0 (p) = y/. We have
2
G=

2
yF (y)f (y)dy 1 =

11

Z

yF (y)f (y)dy
.
2


This formula opens the way to an interpretation of the Gini coefficient in term of covariance as
Cov(y, F (y)) = E(yF (y)) E(y)E(F (y)).
Using this definition, we have immediately that
G=

2
Cov(y, F (y)),

which means that the Gini coefficient is proportional to the covariance between a variable and its
rank. The covariance interpretation of the Gini coefficient open the way to numerical evaluation
using a regression.
R
Meanwhile, noting that Cov(y, F (y)) = y(F (y) 1/2)dF (y), using integration by parts,
we get
1Z
Cov(y, F (y)) =
F (x)[1 F (x)]dx,
2
so that we arrive at the integral form
1Z
F (x)[1 F (x)]dx.
G=

We can remark that F (x)(1 F (x)) is largest at F (x) = 0.5, which explains why the Gini index
is often said to be most sensitive to changes in incomes occurring around the median income.
The above integral form can also be written as
1
G=1

[1 F (x)]2 dx.

We shall prove this equivalence by considering the last interpretation of the Gini which is the
scaled mean of absolute differences.

4.3 S-Gini
We underlined that the Gini coefficient was very sensitive to changes in the middle of the income
distribution. A generalization of the Gini coefficient, obtained by adding a aversion for inequality
parameter as in the Atkinson index, was proposed in the literature by Donaldson and Weymark
(1980) and other papers following this contribution. Starting from
y
G = 2Cov( , 1 F (y)),

the S-Gini is found by introducing so as to modify the shape of the income distribution
y
G = Cov( , (1 F (y))1).

For = 2, of course, we recover the usual Gini index. With a value of greater than 2, a greater
weight is attached to low incomes.
We can run a small experiment, generating n = 1000 observations of a lognormal distribution
and then computing the Gini according to the above formula, with various values of . We then
compare the result to the Gini computed using the usual formula corresponding to = 2.
12

Table 1: Computing the -Gini

using the empirical cumulative distribution

-Gini

Usual Gini

1.2 0.2077537

2.0 0.5288477

0.5277905

3.0 0.6692843

4.0 0.7362263

n = 10000
x = sort(rlnorm(n))
y = seq(0,1,length=n)
for (alpha in c(1.2,2,3,4)){
g = -alpha*cov(x/mean(x),(1-y)(alpha-1))
cat("Gini = ",g," alpha = ",alpha,"\n")}
Gini(x)
For = 1, the modified Gini is equal to zero. For = 2, this method based on the empirical
covariance is only approximate. In small samples, the difference can be substantial. For n = 100,
the covariance method gives G = 0.5413686, while the correct methods gives G = 0.5305954.

4.4 Gini as mean of absolute differences

The initial definition of the Gini coefficient is the mean of the absolute differences divided by
twice the mean. If y and x are two random variables of the same distribution F , this definition
implies
Z
Z
1
IG =
|x y|dF (x)dF (y).
2 0 0
As F (x) and 1 F (x) are simply the proportions of individuals with incomes below and above
x, integrating the product of these proportions across all possible values of x gives again the Gini
R
coefficient, in its form 1 F (x)[1 F (x)]dx. If we decide to proceed step by step, we first note
that |x y| = (x + y) 2Min(x, y), so that the expectation of this absolute difference is
= E|x y| = 2 2E(Min(x, y)).
To compute the last expectation, we need the distribution of the Min of two random variables
having the same distribution. We know or we can show that it is equal to 1 (1 F (y))2, while
its derivative is d(1 F (y). So that
= 2 + 2

y d(1 F (y))2.

13

The last integral can be transformed using integration by parts with u = y and v = (1 F (y))2:
Z

y d(1 F (y)) = y(1

So that we get the integral form of the Gini

i
F (y))2
0

1
IG =
=1
2

[1 F (y)]2dy.

[1 F (x)]2 dx,

because the first right hand term is zero.

5 Estimation of the Gini coefficient

5.1 Numerical evaluation
The definition of the Gini coefficient in term of the mean of absolute differences yield several
ways of estimating it, without any assumption on the shape of F . The direct approach using a
double summation is not feasible. We have first to order the observations to compute the order
statistics x[i] . Several methods were proposed in the literature:
Deaton (1997) in his book orders the observations and proposes to use
G=

X
n+1
2

(n + 1 i)x[i] .
n 1 n(n 1)

Note that this formula points out that there are n(n 1) distinct pairs.
Sen (1973) uses a slight simplification of this with
G=

n+1
2 X
2
(n + 1 i)x[i] .
n
n

The interpretation of the Gini coefficient in term of covariance between the variable and
its rank implies that a simple routine can be used
G=

2
Cov(y[i] , i).
n

For the covariance approach, we note that the mean of the ranks is given by

So the covariance is estimated by

X
n+1
i = 1
i=
.
n
2

1X
1X
n+1
(i i)y[i] =
i y[i]
,
n
n
2
and the Gini coefficient is obtained as:
2 X
n+1
G= 2
i y[i]
.
n
n
Cov(i, y[i] ) =

14

5.2 Inference for the Gini coefficient

The main question is to find a standard deviation for the Gini coefficient. This is not an easy task
because the observations are ordered and thus are not independent. We can find essentially two
methods in the recent literature.
Giles (2004) found that the Gini can be estimated as
IG =

2 n + 1

,
n
n

(1)

where is the OLS estimate of in the weighted regression

q

i x[i] = x[i] + ui x[ i].

(2)

where x[i] is an order statistics and i its rank. An appropriate standard error for the Gini coefficient
is then
q

2 Var()
SE(IG ) =
.
(3)
n
This estimation is biased because the usual regression assumptions are not verified in the above
regression. For instance the residuals are dependent.
Davidson (2009) gives an alternative expression for the variance of the Gini which is not
based on a regression, but simply on the properties of the empirical estimate of F (x). If we note
IG the numerical evaluation of the sample Gini, we have:
V ar(IG ) =

1 X
2,
(Zi Z)
(n
)2

(4)

P
where Z = (1/n) ni=1 Zi is an estimate of E(Zi ) and

i
2i 1
2X

x[i]
x[j] .
Zi = (IG + 1)x[i] +
n
n j=1

This is however an asymptotic result which is general gives lower values than those obtained with
the regression method of Giles. Small sample results can be obtained if we adjust a parametric
density for y and use a Bayesian approach.

6 Lorenz curve and other inequality measures

Simple summary measures of inequality can readily be obtained from the graph of a Lorenz
curve. The share in total income of the bottom p proportion of the population is given by L(p);
the greater that share, the more equal is the distribution of income. Analogously, the share in
total income of the richest p proportion of the population is given by 1 L(p); the greater that
share, the more unequal is the distribution of income.
15

6.1 Schutz or Pietra index

An interesting but less well-known index of inequality is given by the Pietra index. What is the
proportion of total income that would be needed to be reallocated across the population in order
to achieve perfect equality. This proportion is given by the maximum value of p L(p), which
is attained where the slope of L(p) of the Lorenz curve is 1 (i.e., at L(p = F ())). It is therefore
equal to
F () L(F ()).

This index is called the Schutz coefficient in Duclos and Araar (2006), but is also known under
the name of the Pietra index. In a stricter mathematical framework and following Sarabia (2008),
the Pietra index is defined as the maximal deviation between the Lorenz curve and the egalitarian
line
PX = max {p LX (p)}.
0p1

If we assume that F is strictly increasing on its support, the function p LX (p) will be differentiable everywhere on (0, 1) and its maximum will be reached when its first derivative in
p
1 F 1 (x)/
is zero, that is, when x = F (). The value of p LX (p) at this point is given by
PX = F ()

F ()
0

[ F 1 (t)]dt =

1
2

|t |dF (t).

Consequently
E|X |
,
2
which is an alternative formula for the Pietra index.
PX =

6.2 Other inequality measures

It is possible also to give a formulation of the Atkinson index and of the Entropy index as transformations of the Lorenz curve. We first give the expression of these two indices when X is a
continuous random variable.
The Atkinson inequality indices are defined as
IA () = 1

Z

(x/)

dF (x)

1/(1)

, > 0,

where is the parameter that controls inequality aversion. The limiting case 1 is
1
IA (1) = 1 exp

Z

log(x)dF (x) .

The family of generalised entropy indices is

1
IG (c) =
c(c 1)

[(x/)c 1]dF (x),

16

c 6= 0, 1

The two particular cases obtained for c = 0 and c = 1 are

IG (0) =
and
IG (1) =

log(/x) dF (x),

(x/) log(x/) dF (x).

These two indices can be written in terms of the Lorenz Curve. We have for the Atkinson
index
Z 1
1/(1
1
0
IA () = 1
[LX (p)] dp
, > 0.
0

For the generalised entropy index:

1
IG (c) =
c(c 1)

{[L0X (p)]c 1}dp, c 6= 0, 1.

These formulas allow these indices to be obtained directly from the Lorenz curve without the
necessity of knowing the underlying cumulative distribution function.

7 Main parametric distributions and their properties

Several densities have been proposed in the literature to model the income distribution. Of course
all these densities are defined for a positive support. The most simple distributions, and consequently the widely used ones are the Pareto and the log-normal. These distributions have two
parameters. The gamma and the Weibull are also two parameter distributions. In order to fit
better the tails, three parameters distributions were proposed. We shall examine the mainly the
Singh-Maddala distribution. We must note that all these densities are uni-modal. Four parameter
densities were proposed in the literature, without solving the question of multi-modality. At this
stage, mixture of simple distributions offer more flexibility without having an overwhelming cost
in term of parsimony.

7.1 The Pareto distribution

Pareto (1897) observed that in many populations the income distribution was one in which the
number of individuals whose income exceeded a given level x could be approximated by Cx for
some choice of C and . More specifically, he observed that such an approximation seemed to be
appropriate for large incomes, i.e. for x above a certain threshold. If one, for various values of x,
plots the logarithm of the income level against the number of individuals whose income exceeds
that level, Paretos intuition suggests that an approximately linear plot will be encountered.
The important role of the Pareto laws in the study of income and other size distributions is
somewhat comparable to the central role played by the normal distribution in many experimental
sciences. In both settings, plausible stochastic arguments can be advanced in favour of the models, but probably the deciding factor is that the models are analytically tractable and do seem to
adequately fit observed data in many cases.
17

A random variable X follows a Pareto distribution if its survival function is

F (x) = P (X > x) =

x
xm

x > xm .

The use of the survival function comes from the intuitive characterisation of the Pareto. The
cumulative function is simply 1 F which implies
x
F (x) = P (X < x) = 1
xm


The density is obtained by differentiation

f (x) = xm x1 ,

x > xm .

Moments are given in Table 2. We can already see that this density has a special shape. It is
Table 2: Moments of the Pareto distribution
parameters
value
domain
scale

xm

xm > 0

shape

>0

support
median

x [xm ; +)

xm 2
xm

mode
xm

mean
variance

x2m

( 1)2 ( 2)

>1
>2

always decreasing. So it is valuable only to model high or medium incomes. Its moments are
restricted to exist only for certain values of . This is the price to pay for its long tails. In Figure
3, we give the graph of the density for xm = 1 and various plausible values of . The Gini index
(see Table 4 for its expression) is very sensitive to the value of . Table 3 shows that the most
Table 3: Gini and Pietra indices for the Pareto

1.2 1.5 2.0 2.5 3.0 3.5

Gini

0.71 0.50 0.33 0.25 0.20 0.17

Pietra

0.58 0.39 0.25 0.19 0.15 0.12

18

3.0
2.5
2.0
1.5

alpha=2

alpha=1

0.0

0.5

1.0

alpha=3

1.0

1.5

2.0

2.5

3.0

Figure 3: Pareto density

plausible values of the Gini correspond to the very small range [2, 2.5].
The tails of the Pareto distribution have an interesting property which is nice for an empirical
test. On a log-log graph, the tail of the Pareto distribution is a straight line as
log(Pr(X x)) = log(xm ) log(x).
Because the distribution is available analytically, many interesting characteristics for inequality analysis are directly available and given in Table 4. These expressions are particularly simple.
In particular the Lorenz curve of two Pareto distributions can never intersect as soon as the are
different. This is a strong restriction. In Figure 4, we have displayed Lorenz curves associated
to the Pareto densities for various values of . The Pareto density is very unequal for low values
of . It is particularly able to give a good place to rich people in the income distribution. These
Lorenz curve are totally different from those that will be obtained for the log-normal density.
Many variants of the Pareto distribution were proposed in the literature, see for instance
Arnold (2008). Usual generalisations are Pareto II-IV which introduce more parameters.
Pareto and more generally power function distributions can appear in a variety of context that
are nicely summarised in Mitzenmacher (2004). For instance Champernowne (1953) considers
a minimum income xm and then breaks income and small intervals with bounds defined as xm j
with > 1. Over each time step, An individual can move from class i to class j with a probability
pij that depends only on the value of j i. Champernowne (1953) shows that the equilibrium
distribution is a Pareto.
19

Table 4: Various coefficients for the Pareto distribution

Coefficient
expression
domain
(2 2)1/2

Coefficient of variation

>2

L(p) = 1 (1 p)(1)/

Lorenz curve

>1

Pietra index

( 1)

Gini index

(2 1)1

>1

Atkinson

> 1/2

+1

1/(1)

>1

0.6

0.8

1.0

Generalised entropy

alpha=3.2

0.4

alpha=2.2

0.2

alpha=1.6

0.0

alpha=1.2

0.0

0.2

0.4

0.6

0.8

1.0

Figure 4: Lorenz curves for the Pareto density

20

>1

Lp = function(p,alpha) {1-(1-p)((alpha-1)/alpha)}
p = seq(0,1,0.01)
plot(p,p,type="l")
lines(p,Lp(p,1.2),col=2)
lines(p,Lp(p,1.6),col=3)
lines(p,Lp(p,2.2),col=4)
lines(p,Lp(p,3.2),col=5)
text(0.8,0.15,"alpha=1.2",col=2)
text(0.8,0.35,"alpha=1.6",col=3)
text(0.8,0.48,"alpha=2.2",col=4)
text(0.8,0.58,"alpha=3.2",col=5)

7.2 LogNormal distribution

The log-normal density is convenient for modeling small to medium range incomes. A random
variable X has a log normal distribution if its logarithm log X has a normal distribution. If Y is
a random variable with a normal distribution, then X = exp(Y ) has a log-normal distribution;
likewise, if X is log-normally distributed, then Y = log X is normally distributed.
Let us suppose that y is N(, 2) and let us consider the change of variable x = exp y. The
Jacobian of the transformation from y to x is given by
J(y x) =

y
log x
1
=
=
x
x
x

So, the probability density function of a log-normal distribution is:

fX (x; , ) =

x 2

exp

(ln x )2
, x > 0.
2 2

The cumulative distribution function has no analytical form and requires an integral evaluation:
"

1
ln x
ln x
,
FX (x; , ) = erfc
=
2

2
where erfc is the complementary error function, and is the standard normal cdf. However,
these integrals are easy to evaluate on a computer and built-in functions are standard.
The moment are easily obtained as functions of and . If X is a lognormally distributed
variable, its expected value, variance, and standard deviation are
1

E[X] = e+ 2 ,
2

Var[X] = (e 1)e2+ ,
s.d[X] =

Var[X] = e+ 2

21

e2 1.

Equivalently, the parameters and can be obtained if the values of the mean and the variance
are known:

Var[X]
,
= ln(E[X]) 12 ln 1 +
2
E[X]

Var[X]
.
2 = ln 1 +
E[X]2
The mode is

Mode[X] = e .
The median is
Med[X] = e .

1.0

1.5

The above graph was made for = 0. The two densities have the same median, but of course

0.5

y1

sigma = 0.25

0.0

sigma = 1

Figure 5: Log-normal density

not the same mean.

library(ineq)
x = seq(0,4,0.01)
y1 = dlnorm(x,meanlog=0,sdlog=0.25)
y2 = dlnorm(x,meanlog=0,sdlog=1.0)
plot(x,y1,type="l")
lines(x,y2,type="l",col="red")
22

text(1.8,1,"sigma = 0.25")
text(3,0.20,"sigma = 1")
The log-normal has some nice properties.
1. Suppose that all incomes are changed proportionally by a random multiplicative factor,
which is different for everybody and that follows a gaussian process. Then the distribution
of the population income will converge to a log-normal, if the process is active for a long
enough period.
2. The log normal fits well to many data sets
3. Lorenz curves associated to the log-normal are symmetric around a line which is given by
the points corresponding to the mean of x. This is a good visual test to see if the log-normal
fits well to a data set.
4. Inequality depends on a single parameter which uniquely determines the shape of the
Lorenz curves. The latter do not intersect. The Gini coefficient also depends uniquely on
this parameter.
5. Close form under certain transformations
We know that if X N(, 2 ), then Y = a + bX is also normal with Y N(a + b, b2 2 ).
Let us now consider a log-normal random variable Y (, 2) and the transformation Y =
aX b . Then Y (log(a) + b, b2 2 ). There is a nice application for this property. It has been
observed in many countries that the tax scheduled can be approximated by
t = x axb
The disposable income is given by
y = axb
So if the pre-tax income follows a log-normal, the disposable income will also follow a lognormal.
The right tail of the lognormal density behaves very differently from the Pareto tail, just
because the log normal has got all its moment when the Pareto in general has no finite moment
when is too small. However, for large values of , the two distributions might have quite
similar tails. This can be seen on a log-log graph. Let us take the log of the density
log f (x) = log x log
=
' (

log2 x
2 2

+(

(log x )2
2 2

1) log x log

1) log x log

2
2 2

2
2 2

for large

The left tail of the log density behaves like a straight line for a large range of x when is large
enough.
23

1.0
0.8
0.6

0.50

0.4

1.00

0.2

45 line

1.50

0.0

Lc.lognorm(p, parameter = 0.25)

0.25

0.0

0.2

0.4

0.6

0.8

Figure 6: Log-normal Lorenz curves

24

1.0

library(ineq)
p = seq(0,1,0.01)
plot(p,Lc.lognorm(p, parameter=0.25),type="l",col="brown")
lines(p,Lc.lognorm(p, parameter=0.5),col="red")
lines(p,Lc.lognorm(p, parameter=1.0),col="blue")
lines(p,Lc.lognorm(p, parameter=1.5),col="green")
lines(p,p)
text(0.42,0.5,"45 line")
text(0.8,0.68,"0.25")
text(0.8,0.58,"0.50")
text(0.8,0.40,"1.00")
text(0.8,0.20,"1.50")
We can give some more details on this distribution, concerning Gini coefficient and the Lorenz
curve. Let us call (x) the standard normal distribution with (x) = P rob(X < x). From
Cowell (1995), we have Table 5. The Pietra index was found in Moothathua (1989).
Table 5: Various coefficients for the Log-Normal distribution
q

exp( 2 ) 1

Coefficient of variation

(1 (p) )

Lorenz curve

2( 2 /2) 1

2(/ 2) 1

Pietra index
Gini index

1 exp(1/2 2)

Atkinson

exp((2 ) 2 /2) 1

Generalised entropy

Lognormal distributions are usually generated by multiplicative models. The first explanation
of this type was proposed by Gibrat (1930). We start with an initial value for income X0 . In the
next period, this income can grow or diminish according to a multiplicative and positive random
variable Ft
Xt = Ft Xt1 .
Taking the logs and using a recurrence, we have
log Xt = log(X0 ) +

log(Fk ).

By the central limit theorem, we get a log normal distribution. Note that the mechanism designed by Champernowne (1953) was very similar. We got a Pareto distribution only because a
minimum value was imposed.

25

1.0
0.8
0.6
0.4
0.2
0.0
0

Figure 7: Singh-Maddala income distribution

The line in black corresponds to a2 = a3 = 1. Then, for the curve in
red, a2 = 2, while for the curve in green a3 = 3.

7.3 Singh-Maddala distribution

Singh and Maddala (1976) propose a justification of the old Burr XII distribution by considering
the log survival function as a richer function of x than what the Pareto does. With the Pareto we
had log(1 F ) = log(xm ) log(x). Here, the relation is no longer linear with:
log(1 F ) = a3 log(1 + a1 xa2 ),
following the notations of Singh and Maddala (1976). Consequently, the cumulative distribution
is
1
FSM (x) = 1
.
(1 + a1 xa2 )a3
The corresponding density is obtained by differentiation
fSM (x|a, b, q) = a1 a2 a3

xa2 1
.
[1 + a1 xa2 ]a3 +1

Let us plot this density for various values of the parameters. First of all, a1 is just a scale
parameter and we set it equal to 1. Then we use the following code in R:
x = seq(0,5,0.1)
f_SM = function(x,a_2,a3){
26

f = a_2*a_3*(x(a_2-1))/(1+x(a_2))(a_3+1)
}
a_2 = 1
a_3 = 1
plot(x,f_SM(x,a_2,a3),type="l",ylab="",xlab="")
a_2=2
lines(x,f_SM(x,a_2,a3),col=2)
a_3 = 2
lines(x,f_SM(x,a_2,a3),col=3)
The parameter of the Pareto distribution could easily be estimated using a linear regression
of log(1 F ) over log(x) where F is the natural estimator of the cumulative distribution. Here
a non linear regression can be applied which minimised
X

[log(1 F ) + a3 log(1 + a1 xa2 )]2 .

The uncentered moments of order h and the Gini coefficient are expressed in term of the Gamma
function and can be found in McDonald and Ranson (1979) and McDonald (1984):
E(X h ) = bh

(1 + h/a2 )(a3 h/a2 )

(a3 )

with b = (1/a1 )1/a2 as well as the Gini index

G=1

(a3 )(2a3 1/a2 )

.
(a3 1/a2 )(2a3 )

All the moment do not exist in this distribution. For a moment of order h, we must have
a3 >

h
.
a2

If a3 > 1/a2 , we can derive the Lorenz curve as

LC(p) =
=

1Z

0
ba3 Z

b[(1 y)1/a3 1]1/a2 dy

z

t1/a2 (1 t)a3 1/a2 1 dt

= Iz (1 + 1/a2 , a3 1/a2 )
where z = 1 (1 a3 )1/a3 and Iz (a, b) denotes the incomplete beta function ratio defined by
Z

IBz (a, b) = Z01

0

ta1 (1 t)b1 dt
ta1 (1 t)b1 dt

27

1.0
0.6
0.4

LCsingh(p, a, 0.7)

0.8

1.0
0.8
0.6
0.4

a2=3.5

a2=2.5

a3=0.7

a2=2.0

0.0

0.2

a3=0.9

0.0

0.2

LCsingh(p, a, 0.7)

a3=2

0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

0.4

0.6

0.8

1.0

Figure 8: Weibull and Singh-Maddala Lorenz curves

The Singh-Maddala distribution admit two limiting distributions, depending on the value of a3 .
For a3 = 1, we have the Fisk (1961) distribution. For a3 , we have the Weibull distribution,
to be detailed later on. So, depending on the value of a3 , the associated Lorenz curves are
supposed to cover a wide range of shapes. In the left panel, we kept a2 = 2 and let a3 vary
between 0.7 and 2. In the right panel, we kept a3 = 0.7 and let a2 vary between 2 and 3.5. The
two black curves are identical. In one case the modification is more in the right part and in the
other case more in the left part. However, we note that the flexibility is not very strong.
The corresponding code using R is:
LCsingh <- function(p,a,q){
pbeta((1 - (1 - p)(1/q)), (1 + 1/a), (q-1/a))}
p = seq(0,1,0.01)
a = 2
plot(p,LCsingh(p, a,0.7),type="l")
lines(p,LCsingh(p, a,0.9),type="l",col="red")
lines(p,LCsingh(p, a,2),type="l",col="blue")
lines(p,p)
text(0.8,0.24,"a3=0.7")
text(0.8,0.34,"a3=0.9",col="red")
text(0.8,0.48,"a3=2",col="blue")

28

1.2

7.4 Weibull distribution

0.6

0.8

1.0

alpha=0.5

0.2

0.4

alpha=2.0

0.0

alpha=1.2

Figure 9: Weibull income distribution

The Weibull distribution is a nice two parameter distribution where all moments exists. It is
obtained as a special case of the three parameter Singh Maddala distribution, for a3 . This
relation explains that the cumulative distribution has an analytical form:
F (x) = 1 exp((kx) ).
By differentiation, we get the density
f (x) = k (k x)1 exp (k x) .
We have a plot of this density in Figure 9. For < 1, the density has the shape of the Pareto
density, which means that it has no finite maximum. For = 1, it cuts the y axis. As grows,
there is less and less inequality and the function concentrates around its mean. Plausible values
for corresponding to usual income distributions are [1.5 2.5].
The h th moments around zero are given by
h =

(1 + h/)
kh

where (a) is the gamma function defined by

(a) =

ua exp(u) du
29

The coefficient of variation (the ratio between the standard deviation and the mean) is equal to:
cv =

(( + 2)/) ( + 1)/)2
(( + 1)/)

As we have the direct expression of the distribution, the Gini coefficient and the Lorenz curves
are directly available. We find the expression of the Lorenz curve and the Gini index for instance
in Krause (2014):
( log(1 p), 1 + 1/)
LC = 1
,
(1 + 1/)
where (x, ) is the incomplete Gamma function.
We regroup in Table 6 some of these results.
Table 6: Several indices for
the Weibull distribution
((+2)/)(+1)/)2
Coefficient of variation
((+1)/)
1

Lorenz curve

( log(1p),1+1/)
(1+1/)

Pietra index
1 21/

Gini index
Atkinson
Generalised entropy

Note that there are various ways of writing the density of the Weibull, concerning the scale
parameter k. Either (kx) or (x/k) . For inference, it might even be convenient to consider kx .
So be careful. In R, the density is available as dweibull(x, shape, scale = 1) using
the parameterization (x/k) .
The Weibull distribution shares with the Pareto, the Sing-Maddala distribution a common
feature which is to have an analytical cumulative distribution. If we rearrange its expression and
take logs, we get
log( log(1 F )) = log(kx).
So that it is easy to check if a sample has a Weibull distribution. And by the way gives a method
to estimate the parameter .

7.5 Gamma distribution

The probability density function using the shape-scale parameterization is
x

xk1 e
f (x; k, ) = k
(k)

for x > 0 and k, > 0.

30

1.0

DF = 1
0.8

DF = 2
DF = 3
DF = 4
0.6
0.4
0.0

0.2

Density

DF = 5

10

Figure 10: Gamma density

Here (k) is the gamma function evaluated at k. k represent the degrees of freedom. It is also
the shape parameter. corresponds to the scale parameter in this parameterization. Using this
parameterization, we can plot this density for = 1 and various values of k.
n = 1000
x = seq(0,10,length=n)
df = 1.0
s = 1
y = dgamma(x,shape = df, scale = s)
plot(x,y,type="l",ylab="Density")
text(8,1.0-df/15,paste("DF = ",toString(df)),col=df)
for (df in c(2,3,4,5)){
y = dgamma(x,shape = df, scale = s)
lines(x,y,col = df)
text(8,1.0-df/15,paste("DF = ",toString(df)),col=df)}
The cumulative distribution function is the regularized gamma function:
F (x; k, ) =

f (u; k, ) du =

where (k, x/) is the lower incomplete gamma function.

31

k, x
(k)


The skewness is equal to 2/ k, it depends only on the shape parameter k and approaches
a normal distribution when k is large (approximately when k > 10). The mean is k and the
variance k2 .
Rather easy to estimate. Bayesian inference. In R, dgamma, pgamma, qgamma, rgamma
using the same parameterisation.

7.6 Double Pareto-Lognormal distribution

A large class of four parameter densities was proposed in McDonald (1984) and the most famous
one is the Generalized beta II. The main goal was to provide flexibility for both the left and
right tails. A more recent distribution was developed in Reed and Jorgensen (2004), applied
for income distributions in Reed (2003) and is known also as the double Pareto. It is closely
related to the lognormal and Pareto distributions. A good review of this distribution and its
comparison with the Pareto and the lognormal distributions is given in Mitzenmacher (2004).
Both the Generalized Beta II and the Double Pareto have four parameters, but are uni-modal.

7.7 Which density should we select?

In his book, Cowell (1995) is not very optimistic about the more complicated four parameter
densities. Their parameters are hard to interpret and they are difficult to estimate. He is more in
favour of the Pareto density, which is fact has a single important parameter (xm defines only the
support of the density), the two parameter lognormal and eventually the gamma density. He does
not like the more complicated densities like the Singh-Maddala and even more the generalized
Beta II. In Lubrano and Protopopescu (2004), we make use of the two parameter Weibull density
to estimate generalized Lorenz curves and rank bibliometric distributions. The three parameters
Singh-Maddala distribution is quite simple to estimate as the authors propose a method based on
a regression. The three parameter generalized gamma density has a very awkward parameterization so that it has the reputation of being not estimable by maximum likelihood on individual
data.
The Pareto density is nice for modelling high incomes. The gamma density is nice for modelling mid range incomes as well as the log-normal density. Cowell (1995) thus prefers two
parameter densities for modelling particular portions of the income distribution. We can conclude that using mixture of two parameter densities might be the best alternative for modelling
the complete income distribution.

8 Pigou-Dalton transfers and Lorenz ordering

Pigou-Dalton transfers are mean-preserving equalizing transfers of income. They involve a
marginal transfer of 1 from a richer person belonging to percentile pr to a poorer person belonging to percentile pp < pr ) that keeps total income constant. These equalizing transfers have
the consequence of moving the Lorenz curve unambiguously closer to the line of perfect equal-

32

ity. This is because such transfers do not affect the value of L(p) for all p up to pp and for all p
greater than pr , but they increase L(p) for all p between pp and pr .
Let us consider two income distributions A and B, where distribution B is obtained by applying Pigou-Dalton transfers to A. Hence, the Lorenz curve LB (p) of distribution B will be
everywhere above the Lorenz curve LA (p) of distribution A. Inequality indices which obey the
principle of transfers will unambiguously indicate more inequality in A than in B. We will also
say that if
LB (p) LA (p) 0
p
then B Lorenz dominates A.

Lorenz curve
1.0

x_A
x_B
x_C

0.8

L(p)

0.6

0.4

0.2

0.0
0.0

0.2

0.4

0.6

0.8

1.0

Figure 11: Lorenz dominance for Pigou-Dalton transfers

We are going to illustrate Pigou-Dalton transfers on a simulated example. We first generate
an income distribution xA , using a lognormal distribution with parameters 0 and 1 and n = 500
observations. We then define a flat rate of taxation equal to 0.25. A Pigou-Dalton transfer takes
money from the rich to redistribute to the poor without changing the mean income and without
changing the order of the incomes. We can thus define the transfers as
T r = sort(xA , decreasing = T )
where sort(xA , decreasing = T ) is the reverse order of xA , provided xA is sorter by increasing
values. The new income distribution xB is
xB = (1 ) xA + T r
33

Table 7: The effects of

Pigou-Dalton transfers
Distribution Mean

Gini

xA

1.76

0.55

xB

1.76

0.36

xC

1.70

0.39

We finally draw n values of xC from a lognormal with = 1 and having the same theoretical
mean as xA or xB . In Table 7, we compute the mean and the Gini coefficient of each distribution.
We illustrate these numbers in Figure 11 where we have drawn the Lorenz curve of xA in black.
It is the farthest away from the diagonal. Inequality is rather large in this income distribution.
Pigou-Dalton transfers do not change the mean, order the ordering, but reduce greatly the Gini
coefficient. The Lorenz curve corresponding to xB is in red. It does not intersect LA even if the
distribution of xB cannot be a lognormal. However, it intersects the Lorenz curve of xC which
corresponds to a similar Gini coefficient and results from totaly different transfers.

8.1 Lorenz Ordering for usual distributions

This material is adapted from Sarabia (2008). We have derived Lorenz curves for the most
important parametric densities, leaving aside those which were too complex. Lorenz curves can
be used to define an ordering in the space of the L distributions. If two distribution functions
have associated Lorenz curves which do not intersect, they can be ordered without ambiguity in
terms of welfare functions which are symmetric, increasing and quasi-concave (see Atkinson,
1970). We express this formally with the definition
Definition 2 Let A and B be two income distributions. Distribution B is preferred to distribution
A in the Lorenz sense iff
B L A LB (p) LA (p),

p [0, 1].

If B L A, then B exhibits less inequality than A in the Lorenz sense. Note that the Lorenz
order is a partial order and is invariant with respect to scale transformation.
It is fairly possible now to characterize Lorenz dominance by restrictions over the parameter
space if the two random variables have the same class of distributions. For some parametric
families the restrictions will be very simple, and by the way will imply rather simple parametric
statistical tests. We present first results for the Pareto and the log-normal
Pareto: Let Xi P (i , xmi ). Then
FX1 L FX2 1 2
34

 Log-Normal: Let Xi LN(i , i2 ). Then

FX1 L FX2 1 2
The proof of these results is straightforward because in the two cases, the Lorenz curves do not
intersect and depend on a single parameter.
The case of the Singh-Maddala distribution is more difficult to establish. Its Lorenz curve
depends on two parameters and may intersect. Let us note the normalized distribution as F =
1 1/(1 + xa )q . Then from Sarabia (2008) we get:
Theorem 3 Let Xi SM(ai , qi ), i = 1, 2 be two Singh-Maddala distributions. Then
X1 L X2 a1 q1 a2 q2 , and a1 a2 .
The proof of this result is more delicate to establish and the statistical test of these restrictions is
slightly more difficult to implement.

8.2 Generalised Lorenz Curve

We borrow this material to Sarabia (2008). The generalized Lorenz curve (GLC) introduced
by Shorrocks (1983) is the most important variation of the Lorenz curve (LC). The LC is scale
invariant and is thus only an indicator of relative inequality. However, it does not provide a
complete basis for making social welfare comparisons. The Shorrocks proposal is the generalized
Lorenz curve defined as
Z p
GLC(p) = LC(p) =
F 1 (y)dy
0

Note that GLC(0) = 0 and GLC(1) = . A distribution with a dominating GLC provides
greater welfare according to all concave increasing social welfare functions defined on individual
incomes (Kakwani 1984 and Davies et al. 1998). On the other hand, the GLC is no longer scalefree and in consequence it determines any distribution with finite mean.
The order induced by GLC is the second-order stochastic dominance that we shall study in a
next chapter. This order is a new partial ordering, and sometimes it allows a bigger percentage
of curves to be ordered than in the Lorenz ordering case.
The usual Lorenz curve when one focusses his attention on inequality only. The Generalized Lorenz curves mixes concerns for inequality and for the mean, so it is related to welfare
comparisons.

9 Parametric Lorenz curves

Detail some of the solutions exposed in Chotikapanich (2008).
We first recall in a table the expression of the Lorenz curve for some standard income distribution. We gave a theorem characterizing a Lorenz curve. This means that any function following
these properties is a Lorenz curve. So we can try to investigate this class of functions. We follow
35

Table 8: Lorenz and Gini indices for classical income distributions

Distribution
Lorenz curve
Gini index
Pareto I

L(p) = 1 (1 p)11/

Lognormal

L(p) = (1 (p) )

Singh-Maddla L(p) = Iz (q + 1/a, q 1/a)

1
2 1

2(/ 2) 1
1

(q)(2q 1/a)

(q 1/a)(2q)

Sarabia (2008), but not all the details. The first parametric form which was given in the literature
is
L(p) = p exp((1 p))

with 1 and > 0.

A family of Lorenz curves which is interesting and easy to understand is build around the
Pareto family. We can generalize the Lorenz curve of the Pareto by adding one more parameter,
so as to get
L(p) = [1 (1 p)11/ ] .

If = 1, we have the asymmetric Lorenz curve of the Pareto. If = 1/(1 1/), we obtain a
symmetric Lorenz curve, thus having a similar property to that of the Lognormal. The underlying
density to this Lorenz curve combines properties of the Pareto and of the Lognormal. More
general expressions are given in Sarabia (2008).
Let us explore these functional forms using R.
LCgen <- function(p,alpha,beta){
smlc <- (1-(1-p)(1-1/alpha))beta
smlc}
p = seq(0,1,0.01)
plot(p,LCgen(p, 1.5,1),type="l")
text(0.93,0.45,"1.5, 1.0")
lines(p,LCgen(p,3,1.5),type="l",col="red")
text(0.7,0.50,"3.0, 1.5")
lines(p,LCgen(p,4,2),type="l",col="blue")
text(0.5,0.10,"4.0, 2.0")
lines(p,p)
text(0.42,0.5,"45 line")
It is remarkable that play playing with two parameters, we can obtain very different shapes and
in particular many points of intersection in a much simpler way than with the Singh Maddala
distribution. The Gini coefficient has a simple expression and is equal to
G=1

2
B(1/(1 1/), + 1)
1 1/
36

1.0
0.8
0.6
0.4

1.5, 1.0

0.2

LCgen(p, 1.5, 1)

3.0, 1.5

45 line

0.0

4.0, 2.0

0.0

0.2

0.4

0.6

0.8

1.0

Figure 12: The flexibility of a two parameter Lorenz curve

where B(.,.) is the incomplete Beta function.
It would be nice to compute the Atkinson and GE indices using the formula given above
using the Lorenz curve. Derive the corresponding densities.

37

10 Exercises
10.1 Empirics
Using the previous FES data set, the software R and the package ineq, compare the empirical
Lorenz curve to those obtained for the Pareto and Log-normal. Say which distribution would fit
the best. Redo the same exercise limiting the data to high incomes.

10.2 Gini coefficient

We have seen that the Gini coefficient could be seen as the covariance between a variable and its
rank, namely:
2
G = Cov(y, F (y)).

As Cov(y, F (y)) =

y(F (y) 1/2)dF (y), use integration by parts to show that

Cov(y, F (y)) =

1
2

F (x)[1 F (x)]dx,

and give the corresponding form of the Gini. Give the value of F for which the Gini is maximum.
What can you deduce of this result as a property of the Gini index?

10.3 LogNormal
Compute the value of the Generalised Entropy index for = 0 and = 1. Comment your result.
Does it hold in the general case of a general distribution. Do the same calculation for the Pareto
density.

10.4 Uniform
The uniform density between 0 and xm is sometimes used in theoretical economic paper to
describe the income distribution. It writes:
f (x) =

1
1I(x xm )
xm

This density has strange properties that we shall now explore.

1. Compute the mean and the variance
2. Calculate the expression of the cumulative distribution
3. Using the inverse of this cumulative distribution compute the expression of the Lorenz
curve
Z
1 p 1
L(p) =
F (t)dt
0
38

4. Compare L(p) with that of the Pareto distribution

5. Compute the Gini index corresponding to the uniform distribution using
G = 12

L(p)dp

6. Verify that you obtain the same result using

1
G=1

xm

[1 F (t)]2 dt

10.5 Singh-Maddala
Find an example where two Lorenz curves associated to the Singh-Maddala distribution intersect.
Use the graphs produced by R for this. Mind that the parametrisation adopted in R for the
function Lc.singh is awkward. Use the function provided in the text.

10.6 Logistic
The logistic density is very close to the normal density, but it has nicer properties, such as in
particular an analytical cumulative distribution. We have
f (x) =

e(x)/s
s(1 + e(x)/s )2
1

F (x) =

e(x)/s

1+
with mean and variance s /3. Find the log logistic distribution using the adequate transformation. Find the Gini coefficient. This is the Fisk distribution.
2 2

10.7 Weibull
Show that when a3 in the Singh-Maddala distribution, we get the Weibull.

References
Arnold, B. C. (2008). Pareto and generalized pareto distributions. In Chotikapanich, D., editor,
Modeling Income Distribuions and Lorenz Curves, volume 5 of Economic Studies in Equality,
Social Exclusion and Well-Being, chapter 7, pages 119145. Springer, New-York.
Atkinson, A. (1970). The measurement of inequality. Journal of Economic Theory, 2:244263.
Champernowne, D. (1953). A model of income distribution. Economic Journal, 63:318351.
39

Chotikapanich, D. (2008). Modeling Income Distribuions and Lorenz Curves, volume 5 of Economic Studies in Equality, Social Exclusion and Well-Being. Springer, New-York.
Cowell, F. (1995). Measuring Inequality. LSE Handbooks on Economics Series. Prentice Hall,
London.
Davidson, R. (2009). Reliable inference for the gini index. Journal of Econometrics, 150:3040.
Davies, J. B., Green, D. A., and Paarsch, H. J. (1998). Economic statistics and social welfare
comparisons: A review. In Ullah, A. and Giles, D. E. A., editors, Handbook of Applied Economic Statistics, volume 155 of Statistics: Textbooks and Monographs, pages 138. Dekker,
New York, Basel and Hong-Kong.
Deaton, A. (1997). The Analysis of Household Surveys. The John Hopkins University Press,
Baltimore and London.
Donaldson, D. and Weymark, J. (1980). A single-parameter generalization of the gini indices of
inequality. Journal of Economic Theory, 22(1):6786.
Duclos, J.-Y. and Araar, A. (2006). Poverty and Equity: Measurement, Policy and Estimation
with DAD. Springer, Newy-York.
Fisk, P. (1961). The graduation of income distributions. Econometrica, 29:171185.
Gibrat, R. (1930). Une loi des reparations e conomiques: leffet proportionnel. Bulletin de
Statistique General, France, 19:469.
Giles, D. E. A. (2004). Calculating a standard error for the gini coefficient: Some further results.
Oxford Bulletin of Economics and Statistics, 66(3):425433.
Kakwani, N. (1984). Welfare ranking of income distributions. In Basmann, R. and Rhodes, G.,
editors, Advances in Econometrics, volume 3, pages 191213. JAI Press.
Krause, M. (2014). Parametric Lorenz curves and the modality of the income density function.
Review of Income and Wealth, 60(4):905929.
Lubrano, M. and Protopopescu, C. (2004). Density inference for ranking european research
systems in the field of economics. Journal of Econometrics, 123(2):345369.
McDonald, J. (1984). Some generalised functions for the size distribution of income. Econometrica, 52(3):647663.
McDonald, J. B. and Ranson, M. R. (1979). Functional forms, estimation techniques and the
distribution of income. Econometrica, 47(6):15131525.
Mitzenmacher, M. (2004). A brief history of generative models for power law and lognormal
distributions. Internet Mathematics, 1(2):226251.
40

Moothathua, T. S. K. (1989). On unbiased estimation of Gini index and Yntema-Pietra index

of lognormal distribution and their variances. Communications in Statistics - Theory and
Methods, 18: 2,(2):661672.
Piketty, T. (2000). Les hauts revenus en France au 20`eme si`ecle: Inegalites et redistributions,
1901-1998. Grasset, Paris.
Reed, W. J. (2003). The Pareto law of incomes: an explanation and an extension. Physica, A
319:469486.
Reed, W. J. and Jorgensen, M. (2004). The double pareto-lognormal distribution: A new parametric model for size distributions. Communications in Statistics - Theory and Methods,
33(8):17331753.
Sarabia, J. M. (2008). Parametric lorenz curves: Models and applications. In Chotikapanich, D.,
editor, Modeling Income Distribuions and Lorenz Curves, volume 5 of Economic Studies in
Equality, Social Exclusion and Well-Being, chapter 9, pages 167190. Springer, New-York.
Sen, A. (1976). Poverty: an ordinal approach to measurement. Econometrica, 44(2):219231.
Sen, A. K. (1973). On Economic Inequality. Clarendon Press, Oxford.
Singh, S. and Maddala, G. (1976). A function for the size distribution of incomes. Econometrica,
44:963970.
Xu, K. (2003). How has the literature on ginis index evolved in the past 80 years? Economics
working paper, Dalhousie University. Available at SSRN: http://ssrn.com/abstract=423200 or
doi:10.2139/ssrn.423200.
Yitzhaki, S. (1998). More than a dozen alternative ways of spelling gini. Research in Economic
Inequality, 8:1330.

A Hypergeometric series
On a vu que les moments partiels ou certaines distributions cumulees sexprimaient en fonction
de series hypergeometriques dont levaluation et la manipulation sont difficiles. Voir le papier
de McDonald (1984). Or il se trouve quil existe une relation entre certaines de ces series hypergeometriques et les fonctions Gamma et Beta incompl`etes.
La fonction Gamma incompl`ete est donnee par
G(x; p) =

1
(p)

tp1 et dt

On trouve cette formule dans les textes habituels comme Johnson et Kotz. Mais dans
http : //mathworld.wolf ram.com/IncompleteGammaF unction.html
41

le denominateur de normalisation a saute. Par contre ce site donne lequivalence

Rx
0

t1 et dt = 1 x ex 1 F1 (1; 1 + ; x)
= 1 x 1 F1 (1; 1 + ; x)

On remarque donc quil suffit de diviser des deux cotes par () pour retrouver les formulations
usuelles.
Passons maintenant a` la Beta incompl`ete
IB(x, p, q) =

1
Beta(p, q)

up1(1 u)q1 du

On a que
IB(x, p, q) =

x
xp (1 x)q1
)
2 F1 (1 q, 1; p + 1;
p Beta(p, q)
1x

On en deduit donc que la cumulee de la Beta2 generalisee et ses moments partiels peuvent
sexprimer au moyen de la Beta incompl`ete qui comme la Gamma incompl`ete sont programmees
dans R.

A.1

The family of the generalised Beta II

The most general distribution which has been proposed in the literature for fitting income data is
the generalised Beta II (GB2) introduced in McDonald (1984) and developed in McDonald and
Xu (1995). Its density is:
fGB2 (x|a, b, p, q) =

axap1
bap B(p, q)[1 + (x/b)a ]p+q

(5)

with x > 0 and where B(p, q) is the Beta function given by

(p) (q)
B(p, q) =
=
(p + q)

tp1

(1 + t)p+q

dt,

p, q > 0

The generalised Beta II has the advantage that many densities can be obtained as a particular
case and thus it constitutes a nice framework for discussion. In particular, we have the analytical
expression of its uncentred moments
E(X h ) = 0h = bh

B(p + h/a, q h/a)

B(p, q)
(p + h/a)(q h/a)
= bh
(p)(q)
42

(6)

The h moment exists provided aq > h.

The cumulative distribution is written as a function of Z =

FGB2 (x) =

2 F1
pB(p, q)

p, 1 q
p+1

(x/b)a
1+(x/b)a

where 2 F1 is an hypergeometric series.

A.2

The Burr system as a useful particular case

Many of the interesting particular cases we can consider when starting from the GB2 are related
to the system of distributions introduced by Burr (1942). Burr built a system of distributions F (x)
by reference to a differential equation dF/dx = A(F )g(x) which generalises Pearsons system.
The object was to fit a distribution function to the data (rather than a density as in Pearsons
system) and to obtain the density by differentiation. Burr reviews twelve different distributions.
Two are of a particular interest to us. They are now named Burr III and Burr XII by reference to
the number they occupy in the main Table of Burrs original paper. Kleiber (1996) pointed out
that Burr III was nothing but Dagum (1977) distribution. It corresponds to the GB2 with q = 1.
Burr XII is nothing but the distribution promoted some thirty years after by Singh and Maddala
(1976) for fitting income data. It received a considerable interest both in the statistical literature
and in the economic literature. It corresponds to imposing p = 1 in the GB2. Many distributions
can be found as restrictions of the Burr XII.

A.3

Playing with the Burr system

The Burr XII, also named the Singh-Maddala density function is obtained as a first particular
case, imposing p = 1 in the GB2.
fSM (x|a, b, q) = fGB2 (x|a, b, p = 1, q)
= ba qaxa1 [1 + (x/b)a ](q+1)
Its cumulative distribution has an analytical form:
FSM (x) = 1
and its moments are
E(X h ) = bh

1
(1 + (x/b)a )q

(1 + h/a)(q h/a)
(q)

The Gini coefficient is

G=1

(q)(2q 1/a)
(q 1/a)(2q)
43

This distribution has been widely studied in the statistical literature, see e.g. Rodriguez (1977)
for an interesting discussion.
The Pareto density function, is obtained as a restriction of the GB2 with a = b = p = 1 and
q = (or more simply starting directly from the Burr XII)
fP 2 (y|) = fGB2 (y|a = 1, b = 1, p = 1, q = )
= [1 + y](+1)
The obtained distribution is the Pareto of the second kind (or Lomax distribution). The usual
Pareto is obtained after a change in variable x/ = 1 + y (with Jacobian 1/) so that
f (x) = x(+1) 1I(x > )

(7)

This is the notation of Arnold and Press (1983) (see also Johnson, Kotz and Balakrishnan 1994).
The cumulative distribution function is
F (x) = 1 (/x) .

(8)

The uncentered moments are

E(X h ) = h

if h <
h

They are obtained as a particular case of the moments of GB2 using the change of variable
x/ = 1 + y and a recursion over h. Or they can be computed directly from the expression of
the density (which is a simpler solution).
The Weibull distribution is obtained from the Burr XII for q , see Rodriguez (1977) for
a derivation. Its density is
a
fW (x|a, b) = (x/b)a1 exp((x/b)a )
b
and the cdf is
FW (x) = 1 exp((x/b)a )
The uncentered moments are
E(X h ) = bh (1 + h/a)
Rodriguez (1977) notes that Burr type XII can be obtained as a smooth mixture of a two parameter Weibull distribution FW (x) = 1 exp(xa ) compounded with respect to , where has a
standard Gamma distribution with parameter q. Note finally that the Weibull is connected to the
exponential distribution p(y) = ey because it can be obtained from it by the change of variable
Y = (X/b)a .

44

The Burr III is the Dagum (1977) distribution and can be derived from the GB2 imposing
the restriction q = 1 and rearanging the terms
fD (x|a, b, p) = fGB2 (x|a, b, p, q = 1) = apba x(a+1) [1 + (b/x)a ](p+1)
with cumulative distribution
FD (x) = [1 + (b/x)a ]p
The uncentered moments are
E(X h ) = bh

(p + h/a)(1 h/a)
(p)

Kleiber (1996) showed that if X SM(a, b, q), then 1/X D(a, 1/b, p).

The generalised gamma density1 is obtained as a limit of the GB2 when q

fGG (x|a, , p) = limq fGB2 (x|a, b = q 1/a , p, q)
=

axap1
ap (p)

exp[(x/)a ]

Recalling that the Weibull distribution was obtained considering the same limit for the Burr XII,
the generalised gamma is thus a generalisation of the Weibull which is obtained for p = 1. The
cumulative distribution function is
F (x) = G[(x/)a ; p]
where G is the incomplete Gamma function
G(x; p) =

1
(p)

tp1 et dt

The uncentered moments are best derived using the properties of the gamma distribution,
E[X h ] = h

(p + h/a)
(p)

One common and alternative way to define the generalised gamma density is to start from a random variable Z
with a standard gamma distribution and make the change of variable X = bZ 1/a with Jacobian (a/b)(x/b)a1 .

45