A Theil-like Class of Inequality Measures, its Asymptotic Normality Theory and Applications

07/20/2018
by   Pape Djiby Mergane, et al.
0

In this paper, we consider a coherent theory about the asymptotic representations for a family of inequality indices called Theil-Like Inequality Measures (TLIM), within a Gaussian field. The theory uses the functional empirical process approach. We provide the finite-distribution and uniform asymptotic normality of the elements of the TLIM class in a unified approach rather than in a case by case one. The results are then applied to some UEMOA countries databases.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

03/05/2019

A multinomial Asymptotic Representation of Zenga's Discrete Index, its Influence Function and Data-driven Applications

In this paper, we consider the Zenga index, one of the most recent inequ...
12/12/2018

Divergence measures estimation and its asymptotic normality theory : Discrete case

In this paper we provide the asymptotic theory of the general phi-diverg...
07/22/2018

On the influence function for the Theil-like class of inequality measures

On one hand, a large class of inequality measures, which includes the ge...
08/29/2018

Estimation of income inequality from grouped data

Grouped data in form of income shares have been conventionally used to e...
07/14/2018

About the lower bounds for the multiple testing problem

Given an observed random variable, consider the problem of recovering it...
10/09/2018

Fair Division Minimizing Inequality

Behavioural economists have shown that people are often averse to inequa...
08/08/2021

Unbottling the Gini: New Tools from an Old Concept

The Gini index signals only the dispersion of the distribution and is no...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1. Introduction

In this paper, we deal with a modern weak theory for some large class of inequality indices that, further, will allow to handle easy comparison studies with different kinds of statistics.

According to earlier economists, inequality indices are functional relations between the income and the economic welfare (see Dalton (1920)). This explains, among others, the wide variety of such indices in the literature (See, e.g., Cowell (1980a, b, 2000)).


Such statistics, of course, have been widely studied with respect to a great variety of interests, including statistical characterizations and asymptotic properties (See Davidson and Duclos (2000), Barrett and Donald (2009), for recent studies).


Recently, Greselin et al. (2009) provided a mathematical investigation of these indices in a modern setting including Vervaat processes, L-statistics and empirical processes.


Having in mind the necessity of comparing inequality measures with different kind of statistics such as growth statistics, we aim at providing a coherent asymptotic weak theory for some class of inequality measures. Indeed we propose the functional empirical process setting (see Van der Vaart and Wellner (1996)) which provide natural Gaussian field in which many statistics used in Economics may be represented in.


Our best achievement consists of the asymptotic representations for the elements of our class of inequality measures, in terms of the above mentioned Gaussian field. The results are illustrated in data driven applications, on Senegalese data for instance.


The class on which we focus here is a functional family of inequality measures which gathers various ones around the central Theil measure. This class named after the Theil-Like Inequality Measure (TLIM) will be the central point of our study. It includes the Generalized Entropy Measure, the Mean Logarithmic Deviation (Cowell (2003); Theil (1967); Cowell (1980a)), the different inequality measures of Atkinson (1970), Champernowne and Cowell (1998), Kolm (1976a), and the divergence of Renyi (1961).


This means that, here, we will not discuss other inequality statistics such as the Gini, the Generalized Gini, the S-Gini, the E-Gini (See Barrett and Donald (2009)). Those statistics and similar ones will be treated in separate papers.


Now we are going to introduce our family. For that, let

denote the income (or expense) random variable related to a given population. We assume that

and its independent observations are defined on the probability space

and take their values in on some interval

and have common cumulative distribution function

(cdf), , . In this paper, we only use Lebesgue-Stieljes integrals and for any measurable function , we have, whenever it makes sense,


where is the measure image of by , but is also Lebesgue-Stieljes probability measure characterized by: for any .

Now, consider a sample of individuals or households of that population and observe their income . We define the following family of inequality indices, indexed by as follows

(1.1)

where is the empirical mean while , , and are real and measurable functions of and . The exact form of is not important here, in opposite to the conditions on the functions , , and under which the results are valid. In a future paper on the uniform limits in , that class will be crucial.


We will see below that under specific hypotheses on and , converges to the exact inequality measure

(1.2)

where is the mathematical expectation of that we suppose finite here. We will come back later on the function classes and in which and are supposed to lie.


Each measure of this Theil-like family has its own particular properties, that are derived from the combination of different concepts. One may mention the concept of welfare criteria (Atkinson (1970), Sen (1973)), that of the analogy with analysis of risks (Harsanyi (1953), Harsanyi (1955), Rothschild and Stiglitz (1973)), that of the complaints approach (Temkin (1993)) etc. The Theil inequality itself finds all its interest in the information-theoretic idea following that of main components (Kullback 1959). It is based on the three following axioms: Zero-valuation of certainty, Diminishing-valuation of probability, Additivity of independent events. A deep review of such of individual properties for a number inequality measures can be found in Cowell (Cowell (1980a, b, 2000)) for instance.

It is worth mentioning that the TLIM presented here, is rather a mathematical form gathering a number of different measures.


The rest of our paper is organized as follows. In Section 2, we describe the TLIM family and show how the particular indices are derived from it. In Section 3, we briefly recall the functional empirical processes setting. In section 4, we deal with the asymptotic theory of the TLIM, state and describe our main results and demonstrate them. Section 5 is devoted to datadriven applications. We finish by a conclusion in Section 7.


2. Description of the TLIM


This inequality measures mentioned above are derived from (1.1) with the particular values of the mesurable functions and as described below for all


2.1. Generalized Entropy

2.2. Theil’s measure

2.3. Mean Logarithmic Deviation

2.4. Atkinson’s measure

2.5. Champernowne’s measure

2.6. Kolm’s measure

2.7. Divergence of Renyi


3. The functional empirical process

Let be a sequence of independent and identically distributed random elements defined on the probability space with values in some metric space Given a collection of measurable functions satisfying

where is the mathematical expectation of , the functional empirical process (FEP) based on the and indexed by is defined by:

This process is widely studied in Van der Vaart and Wellner (1996)

for instance. It is readily derived from the real Law of Larges Numbers (

LLN

) and the real Central Limit Theorem (

CLT) that and that , where

(3.1)

whenever .

When using the FEP, we are often interested in uniform LLN’s and weak limits of the FEP considered as stochastic processes. This gives the so important results on Glivenko-Cantelli classes and Donsker ones. Let us define them here (for more details see Van der Vaart and Wellner (1996)).


Since we may deal with non measurable sequences of random elements, we generally use the outer almost sure convergence defined as follows:


a sequence converges outer almost surely to zero, denoted by whenever there is a measurable sequence of measurable random variables such that

The weak convergence generally holds in the space of all bounded real functions defined on equipped with the supremum norm


Definition 1.

is called a Glivenko-Cantelli class for , if


Definition 2.

is called a Donsker class for , or -Donsker class if converges in to a centered Gaussian process with covariance function

Remark 1.

When and , is called real empirical process and is denoted by

In this paper, we only use finite-dimensional forms of the FEP, that is And then, any family of measurable functions satisfying (3.1), is a Glivenko-Cantelli and a Donsker class, and hence


where is the Gaussian process, defined in Definition 2. We will make use of the linearity property of both and . Let be measurable functions satisfying (3.1) and , then


The materials defined here, when used in a smart way, lead to a simple handling of the problem which is addressed here.

4. Our results

Let us introduce some notation.

for all , we define the following function

with and is the derivative of the function


The following general condition will be assumed in all the paper:

  • is not null in a neighborhood of

Here are our main results.

4.1. Pointwise asymptotic laws

Consider the following hypotheses based on the functions . The A1.x series concern the almost-sure limits and the A2.x the asymptotic normality.


  • is a continuous function on

  • for and is continuous on


  • is continuously differentiable such that

  • is continuously differentiable at


We have :

Theorem 1.

Suppose that the conditions and are satisfied, then converges almost surely to


Theorem 2.

Suppose that the conditions and are satisfied, and is finite. Then

(a) we have the following asymptotic representation in the empirical functional process

where


(b) and we have the convergence in distribution, as tends to infinity, of to centered normal Gaussian law:

where


with


Remark. The main result is the one given in Point (a). From it, Point (b) is deduced in a straightforward way.

The results above cover all the TLIM class. They should be particularized for the practitioner who would pick one of the elements of that class for analyzing data. Here are then the details for each case.

4.2. Particular cases for pointwise results

a. The Theil’s measure

The empirical form of Theil measure is defined as follows

Denote by


the continuous form of the Theil measure.


All these functions are continuous on then the assumptions defined above become for the a.s. requires that is finite and . As for the asymptotic normality, we need that

And we have . We conclude that

with


b. The Mean Logarithmic Deviation


Let

be the empirical form of the Mean Logarithmic Deviation. Its theoretical form is given as folllows

These specific functions are given by:


The consistency requires that and that while the normality is got when


In that case, we find easily that and

where


c. The Champernowne’s measure


In this case, the specific functions are given by:

And, the various forms are:


We find that and where is the Mean Logarithmic Deviation. As is continuous on we consider the same hypotheses as in the case of Mean Logarithmic Deviation.
The function is continuously differentiable, we put and then we have

with


d. Cases of the Generalized Entropy the Atkinson’s measure the Divergence of Renyi


We may gather these indices into one subclass by giving different values to the function and to the parameter , with this common expression


and then give a general description of the results. For that, Let and


We require for consistency that and that and, for asymptotic normality that

Further, let and . Then we get

which tends towards a centered Gaussian process with variance

(4.1)

Now, we may return to the individual cases.

d.1. Generalized Entropy

We find , from there, we get the variance


d.2. Atkinson’s measure

Put . We similarly get that


d.3. Divergence of Renyi

By taking we obtain by the same way, that

where is given in (4.1).


e. Case of the Kolm’s measure


This index is defined for and its specific functions are:


Its empirical form is given by


and its theoretical form is defined as follows


We need for consistency that and that and, for asymptotic normality that


and are finite.

Then we have and


Put

Then


Since we deduce that

Finally, we summarize the used abbreviations in Table 1, and, for each index, the expression of the function and in Table 2 where we can find the expressions of and


Notations Indices
Generalized Entropy with parameter
THEIL Theil
MLD Mean Logarithmic Deviation
ATK and Atkinson with parameter
CHAMP Champernowne
KOLM Kolm with parameter
DR Divergence of Renyi with parameter
Table 1. Notations of the indices

Indices
THEIL
MLD
ATK
CHAMP
KOLM
DR
Table 2. Summary of the functions for each index

5. Proof of Theorems 1 and 2

Proof of Theorem 1.

On one hand, denote by

(5.1)

by decomposing the difference of and we get the next equality


As for all the function is continuous on and using the fact that converges almost surely to then we have when tends to infinity

(5.2)

We have also

Or the sequence of the random variables is independent and identically distributed, and as by the hypothesis then the Law of Large Numbers implies that

(5.3)

Finally, using (5.2) and (5.3), we get


On the other hand, as satisfies the hypothesis then we deduce that


Proof of Theorem 2.

Using the equation (5.1), we have


Since is continuously differentiable at for we get



Then

But


then we get the next expression


Then