You are on page 1of 20

1

STATISTICAL INFERENCE UNDER


INEQUALITY CONSTRAINTS FOR
NON STANDARD MODELS

A thesis submitted for the degree of Doctor of Philosophy

Suman Rakshit
M.Sc. Statistics I.I.T. Kanpur, India

Department of Econometrics and Business Statistics


Faculty of Business and Economics
Monash University
Australia
March 2011
2

Declaration

I hereby declare that this thesis contains no material which has been accepted for the
award of any other degree or diploma in any university or equivalent institution, and that,
to the best of my knowledge and belief, this thesis contains no material previously by
another person, except where due reference is made in the text of the thesis.

SUMAN RAKSHIT
3

Abstract
4

Acknowledgements
5

Contents
6

List of Tables
7

List of Figures
8

Chapter 1
Simultaneous test of Non-inferiority and
Superiority : Lorenz curve perspective

Introduction

1.1 Motivation

Comparison of populations using welfare measures:


Comparing populations with respect to their income distribution has been a well-known
research topic. Especially, the last two decades have witnessed considerable use and
development of statistical theory for inferring the dominance of one distribution (income,
wealth, wages, etc.) over another. Different welfare measures have been applied, such
as first and second-order stochastic dominance, Lorenz dominance, transfer-sensitivity
dominance and comparisons of poverty deficit curves (Beach and Davidson (1983), Bishop,
Formby and Thistle (1992), Howes (1993), Beach, Davidson and Slotsve (1994)). These
criteria comprise quite common principles of “symmetry”, “scale independence” and
“principle of transfer” (see, for instance, Shorrocks (1983) and Davies and Hoy (1994)).
The comparisons usually aims to establish inequality and social welfare rankings for two
comparative populations.

Lorenz dominance:
Recent interest in poverty issues is centered on unambiguous ranking of income distribu-
tions. Lorenz dominance provides an unanimous ranking of populations based on income
9

inequality (Atkinson (1970)).1 An extensive research is bestowed to test for Lorenz domi-
nance and stochastic dominance (see, for instance, Kaur, Rao and Singh (1994), Mcfadden
(1989), Davidson and Duclos (2006), Anderson (1996), Barrett and Donald (2003)). While
Lorenz dominance continues to play the central role in ranking populations, experience
shows that in many cases of practical interest, the Lorenz curves being compared do
intersect and hence there is no Lorenz dominance. Therefore, there is a definite need to
identify the limitations of this criterion and develop new statistical methodology that
should be able to rank the populations under much broader circumstances.

Limitations of Lorenz dominance:


A disadvantage of the Lorenz dominance tests is that they are very conservative. These
tests are able to rank the economies only if the Lorenz curve of one population entirely
lies above another. However, there are many practical scenarios where there is no Lorenz
dominance albeit the two populations could be ranked on the basis of stronger normative
judgments (see for instance, Shorrocks and Foster (1987), Dardanoni and Lambert (1988),
Davies and Hoy (1994,1995), Gajdos (2004))2 . For example in fig 1.1.1(b) it can be seen
that the major part of population−1 has less income inequality than population−2. In a
practical sense, population-1 should have been considered superior as the fall of its Lorenz
curve is practically negligible. Borrowing terminology from biostatistics, we shall refer to
this as Lorenz improvement of practical significance.

New methodology:
A related statistical inference problem has been studied in recent literatures of medical
statistics for comparing two different treatments when the response variable is multivariate
(see for instance, Bloch, Lai, Tubert-Bitter (2001), Hung, Wang, Tsong, Lawrence, O’Neil
1 For ease of understanding, the Lorenz dominance is considered in accordance with the Lorenz curve

dominance, i.e. population-1 dominates population-2 by Lorenz order if Lorenz curve of population-1
completely lies above the Lorenz curve of the population-2.
2 Davies and Hoy (1995) made the aversion to downside inequality (ADI) criterion more fully operational

by providing a simple procedure for establishing whether any two distributions whose Lorenz curves cross a
finite number of times can be ranked under ADI
10

(2003), O’Brein (1984), Laska, Tang, Meisner (2002), Tamhane, Logan (2004) Perlman, Wu
(2004)). It is very rare that one treatment will dominate another in terms of every outcome
variable of interest. In such cases, one may conclude that there is a difference of practical
significance when treatment-1 performs better than treatment-2 with respect to at least
one variable but not substantially worse with respect to every variable. In this chapter,
we adopt this approach to suggest a new way of formulating an income inequality testing
procedure, what we may call near Lorenz dominance. This would not be as strong as Lorenz
dominance.

1.1.2 Near Lorenz dominance and outline of the problem

The need for statistical methods to establish that a particular treatment is better than an-
other, is encountered frequently in many areas of studies. In most of the existing empirical
studies, the treatment effect is usually captured by a scalar parameter and a studentized t
test statistic is used. Often the treatment effect is reduced to a scalar parameter for sim-
plicity. There are many practical settings where the parameter representing the treatment
effect is either a finite dimensional vector parameter or a smooth function. In such cases,
the methodological issues can be challenging.

The basic formulation of the problem:


As an example, suppose that we wish to establish that population-2 dominates population-
1 in terms of individual income levels. Let Xj and Fj denote the income variable and its
cumulative distribution function for population-j (j = 1, 2). To establish that population-2
dominates population-1 in terms of income levels we need to establish that θ(t) ≥ 0 for
every t with strict inequality holding for at least one value of t, where θ(t) = F1 (t)−F2 (t). In
this case, we say population-2 dominates population-1 with respect to first order stochastic
dominance.
11

Suppose that income data are available only in grouped form. For example data are
available only for the income groups (t0 , t1 ], · · · , (tk , tk+1 ]. Let θi = F1 (ti ) − F2 (ti ), i = 1, · · · , k.
In this case, to establish that population-2 dominates population-1 in terms of income
level, we need to establish that θi ≥ 0 for every i with strict inequality holding for at least
one i (i = 1, · · · , k).

Few instances where the same formulation could be implemented:


The structure of the foregoing inference problem arises naturally in many other areas.
Let us mention some examples briefly: (a) θ(t) = ROC2 (t) − ROC1 (t) where ROC1 (t) and
ROC2 (t) are the receiver operating characteristic curves for a test and standard (reference).
(b) θ(t) = f2 (t) − f1 (t) where fj (t) is the dose-response functions for treatment j = 1, 2. (c)
Rt
θ(t) = L2 (t)−L1 (t) where Li is the Lorenz curve for population-i. (d) θ(t) = 0 {F1 (t)−F2 (t)} dt
; if θ(t) 0 then population-2 is said to dominate population-1 with respect to sec-
RtRs
ond order stochastic dominance. (e) θ(t) = 0 0 {F1 (x) − F2 (x)} dx ds ; if θ(t) 0 then
population-2 is said to dominate population-1 with respect to third order stochastic domi-
nance. Lorenz/stochastic dominance are fundamental concepts used in welfare economics.
Stochastic dominance plays a fundamental role in reliability engineering (see Shaked and
Shanthikumar (2006), Muller and Stoyan(2002)).

Combined superiority and non-inferiority test as the basis of test formulation:


The foregoing examples show clearly that the inference problem studied in this project is of
fundamental importance. However, the inferential procedure we propose is fundamentally
motivated by a widely used testing procedure in clinical trials, known as “combined
superiority and non-inferiority testing”. An important ingredient of our approach is to
demonstrate that θ(t) is greater than −ε(t) for all t (to demonstrate the non-inferiority),
and is greater than δ(t) for some t ( to demonstrate superiority at some point t). The
nonnegative functions ε(·) and δ(·) are specified by the researchers respectively as the
allowable level of deterioration and the desired level of improvement for one entity (in
12

most cases, stochastic functions such as Lorenz curve, response functions, ROC curve e.t.c)
over another under comparison.

Comparison of populations based on the income inequality:


Our main focus in the present context is Lorenz curve. For simplicity, let us explain the
ideas for the case of two points. Let θi denote θ(ti ), the vertical distance of L2 from L1 at
ti ∈ (0, 1). Now, we are proposing that a fall of εi (ε(ti )) of the Lorenz curve at ti for i = 1, 2
is considered practically insignificant; similarly, an increase of δi (δ(ti )) of the Lorenz curve
at either of the ti ’s is considered practically significant. Therefore, the choice of ε and δ
would depend on what the researcher considers to be practically (in)significant in a given
study. In this present framework, population-2 is considered less unequal(in terms of
income distribution) than population-1 if θ1 > −ε1 and θ2 > −ε2 , and moreover, if either of
the following two scenarios θ1 > δ1 , θ2 > δ2 occurs.

The new approach is broader in scope:


The foregoing discussion proposes a new testing procedure that offers a more practical
and flexible framework of comparing two populations. Furthermore, the new approach
would be able to compare income inequality for any particular segment of the population.
For example, a question of particular interest is the following: Has the proportion of total
income increased by a practically significant amount for some lower income groups after
tax, while for all other income groups it did not get significantly lower? By choosing
appropriate ti ’s and εi , δi ’s it would be possible to address similar income inequality
related queries.

An overview of the testing procedure and its challenges:


In the present formulation, we propose to apply a combination of intersection-union and
union-intersection test. Furthermore, to overcome the problem of unknown least favorable
null configuration a bootstrap procedure is proposed. The method developed here is
mainly based on previous work of Tamhane and Logan (2004) and Bloch et al. (2001,2007).
13

The underlying assumption of all of these earlier works was that the test statistic is pivotal
i.e. the corresponding asymptotic covariance matrix say V , does not depend on the original
parameter value θ. Consequently, that made their testing problem quite simple and the
bootstrap algorithm straightforward. However, the testing problem we are focused on
is not pivotal in nature (see Holly and Monfront (1980) Kodde and Palm (1987), Wolak
(1980)). Therefore, the bootstrap procedure described in Tamhane and Logan (2004) and
Bloch et al. (2001,2007) does not directly solve the problem of unknown least favorable
null configuration. Accordingly, we adhere to the local nature of the hypothesis test for
solving the intractable least favorable null configuration problem. It makes the problem
nonstandard and more challenging. Thus, there is a definite need to develop new improved
inferential methodologies for problems like ours. This chapter focuses on these issues.

Outline of the chapter:


In section 2, a brief review of the recent developments on combined superiority and
non-inferiority testing method is presented. There are different test statistics proposed by
various authors. Basic properties of these test statistics are discussed in this section.
In section 3, a general study of the union-intersection (UI) and intersection-union (IU)
principle is considered. Our testing procedure is based on a combination of these two
testing principles. Thus, the important results are referred as well.
In section 4, the methodology is described as the combination of UI and IU principles. The
difference between the proposed test (UI-IU) and a standard pivotal test are identified and
addressed. To overcome the problems we refer to a local test.
In section 5, the local test is constructed. The nature of the local test is discussed. All
the advantages of this construction are specified as well. This allows us to perceive an
appropriate bootstrap procedure for the test.
Section 6 illustrates the bootstrap algorithm.
In section 7, a thorough simulation study is conducted to evaluate the performance of the
proposed test. The study is aimed at investigating the type-I error rate of the UI-IU test, as
14

well as analyze its power properties. The study compares the performance of the UI-IU
test with that of the Perlman and Wu test.
Section 6 provides an empirical study to illustrate the methodological ideas. The com-
parison of income inequality between year 2001 and year 2002 is derived for households
throughout the Australia. The annual disposable income estimates are used to conduct
the study.
In section 7, the main issues are summarized. Based on the findings and recent develop-
ment in the area, a conclusion is made.
Section 8, 9 and 10 form the appendix for the proofs, tables and figures respectively.

1.7 Simulation study

A simulation study is conducted to investigate the type-I error rate of the UI-IU test,
as well as analyze its power properties. The important aspect of this study is to assess
the accuracy of the theories those are developed in earlier section. Accordingly, the null
configuration θ = δ n is investigated for various populations. In line with proposition 1.5.2,
the type-I error rate for configurations on the boundaries are investigated and reported.
To evaluate the performance of the UI-IU test, a comparison is made with the Perlman and
Wu test statistic.

Lorenz Curve:
The main objective of this study is to demonstrate the applicability of the testing procedure
in case of income inequality comparison. Lorenz curves have been used to compare the
income inequality between two populations. Formally, the Lorenz curve corresponding to
an income distribution F is defined as

R F−1 (p)
0
u dF(u)
L(p) = R∞ for p ∈ [0, 1] (1)
0
u dF(u)
15

L(p) represents the total income received by the bottom 100p% of the population expressed
as the proportion of the total income of the population. The empirical estimate of the
Lorenz ordinate is obtained from the sample income data. Let {X(1) ≤ · · · , X(n) } be the
ordered observations of a random sample. Then, the sample estimate of Lorenz ordinate is
given by
{X(1) + · · · + X(r) }
L̂(p) = (2)
{X(1) + · · · + X(n) }

where r = [np], the integer part of np.


The simulation study is carried out for both discrete and continuous income data.

1.7.1 Simulation study of level error

Data generation from a discrete income distribution:

Let us first consider the scenario when the income data are generated from a discrete
income distribution. The income groups for discrete income distributions are obtained
using Singh-Maddala distribution. A 3-parameter Singh-Maddala distribution is well
known for providing good fit to income data ( Branchman et al. (1996)). The functional
form of the Lorenz curve in the case of Singh-Maddala distribution is used to obtain the
income groups. The Lorenz ordinate of a Singh-Maddala distribution is expressed as

R 1−(1−p)1/c
0
t1/b (1 − t)c−1/b−1 dt
L(p) = R1 , 0≤p≤1 (3)
0
t1/b (1 − t)c−1/b−1 dt

The design that is implemented to determine the discrete income distributions is the
following. For m comparison points on x-axis the Lorenz ordinates are obtained. The
discrete income distribution is assumed to be multinomial distribution with m + 1 income
levels. Under the assumption that it produces the same Lorenz ordinates at those m grid
points, the income levels are calculated.
16

Proposition 1.7.1:
Let {L1 , · · · , Lm } are Lorenz ordinates obtained using Singh-Maddala distribution at {0 < x1 <
· · · < xm < 1}. The income levels corresponding to the discrete distribution are given as

β L1 β(L2 − L1 ) β(Lm − Lm−1 ) β(1 − Lm )


0< < < ··· < <
x1 (x2 − x1 ) (xm − xm−1 ) (1 − xm )

where β > 0 is any positive number signifies the scale of the incomes.

Design plan for the configuration θ = δn :


The scale parameter of the Singh-Maddala distribution was fixed at 100 for all simu-
lations. Different combinations of shape parameters (b, c) are used to obtain income
groups. These parameter values cover a wide range of Lorenz curves. All the simulations
are done for m = 5 and 9. The grid points are considered as (0.1, 0.3, 0.5, 0.7, 0.9) and
(0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9) respectively.

Lorenz ordinates for population 1 are obtained using the afore mentioned technique.
However, to select the Lorenz ordinates for population 2, the following result is considered
Proposition 1.7.2:
Let g : [0, 1] → [0, 1] is an increasing convex function satisfying, g(0) = 0 and g(1) = 1. Let
L(x), x ∈ [0, 1] is a Lorenz curve. Then L∗ (x) = L(g(x)) is also a Lorenz curve and L∗ (x) ≤ L(x) for
all x ∈ [0, 1].
Consequently, the Lorenz ordinate for population 2 is defined as L2 (p) = L1 (g(p)), where
g(x) = x1+η has been chosen. For this part of the simulation, ε n was chosen as half of the δ n .

The simulations are carried out for different combinations of sample size and η to evaluate
the accuracy of the asymptotic least favorable null configuration given in proposition
1.5.3. For each Singh-Maddala distribution, four combinations of (n, η) are considered:
(100, 1.4), (500, 1.2) (3000, 1.1) and (10000, 1.05). Furthermore, for each simulation, 1000
independent sets of data are generated. Five hundred bootstrap samples are generated
17

from each of the 1000 sets of data, as described in Section 1.6. The nominal level of the
test is 0.05 in every case.

Results:
The simulation results for null configuration θ = δ n are summarized in Table 1.11.1 -
Table 1.11.4. The results here are fairly consistent with the theory developed in Section
1.5. Even when the sample size is only 100, the empirical level of the test is consistently
close to the nominal level. The level error is not affected substantially by the choice of
η. However, it is worth noting that the rejection rate is almost equal to the nominal level
for small η and large n. On the other hand, rejection rates for smaller sample sizes are
relatively lower than the significance level of the test, mostly lying between 0.030-0.040.
Moreover, the difference between the results due to the difference in m is small, with m = 9
performing slightly better. Overall, the rejection rates corresponding to m = 9 are closer to
the nominal level than m = 5. In addition, the difference in results between the UI-IU and
the Perlman and Wu tests are not considerable. However, the rejection rate corresponding
to the Perlman-Wu statistic is perceived to be slightly higher. Overall, both the tests control
the type I error rates for all the cases that are considered.

The other boundary (null parameter space) configurations:


In this part of the study, type I error rates for different boundary null configurations barring
θ = δ n are assessed. Proposition 1.5.1 emphasizes upon an important set of configurations
on the boundaries of the null parameter space. To integrate that in the simulation study,
the following class of configurations are investigated.

{θt = −εnt , for exactly one t, and θj > δnj for all j , t, t = 1, · · · , m}

These configurations are asymptotic least favorable if the distance between θj and δnj
increases to infinity ( in limit) for all j , t. However, in case of Lorenz difference, the
18

configurations of this nature are not feasible. Consequently, simulations are carried out
for configurations with large differences between θj and δnj for j , t.

In addition, the null configurations corresponding to θt = −εnt for two or more t are also
included in the simulation study. The construction of these configurations amounts to a
tedious design plan.

Design plan:
The income groups for both the populations are obtained using the Singh-Maddala distri-
bution. All the configurations in the present scenario are associated to intersecting Lorenz
curves. Therefore, the parameter values for both the populations are chosen in such a way
that the corresponding population Lorenz curves intersect each other.

Let (bi , ci ) be the parameters corresponding to the ith population, i = 1, 2. Then, the two
population’s Lorenz curves intersect each other if any one of the following occurs

(i) {b1 ≥ b2 and b1 c1 < b2 c2 } (ii) {b2 ≥ b1 and b2 c2 < b1 c1 }

However, the δn is obtained as it was done in the previous section, using the proposition
1.7.1. It is calculated with respect to population 1 by choosing small η values between 1.01
and 1.05. The εn is considered in accordance with the configurations under consideration.

A sample size of 5000 is considered for this part of the simulation study. The large sample
size and the small value of η are chosen in line with the theory. The number of bootstrap
repetitions and the sets of data that are generated are kept same as before.

Results:
The results of the study are reported in Table 1.11.5-1.11.8. Parameter values for the
populations are reported under the column heading (bi , ci ) in the table. Clearly, the
estimated levels of the bootstrap test are much lower than the nominal level. Specifically,
for the configurations possessing less number of θt ’s larger than δt s, are substantially
19

lower than the significance level. Rejection rates for configurations featuring two or less
such points are between 0.01 and 0.02. Furthermore, the rate increases up to 0.03 for
configurations with more than two points.

The empirical level of the test is also influenced by the magnitude of the difference between
θ and δ n . The rejection rates are between 0.01 and 0.03 when the difference between θj ’s
and δnj ’s are not substantial. However, it is worth noting that, the rejection rate is close to
the nominal level for large difference between θ and δ n . Rejection rate increases close to
the nominal level for configurations bearing θj greater than five times of δnj . Overall, the
error rates do not surpass the nominal level for all the configurations that are studied.

Continuous income distribution:

Alongside the grouped income data, continuous income data are also encountered quite
frequently in different economic study (Bhattacharya et al. (2007)). Continuous income
data are generated from the Singh-Maddala distribution for both the populations. Simu-
lations are conducted using the same Singh-Maddala distributions those are used in the
study of discrete data. The comparison margins of the test respectively εn and δn are also
considered to be the same as before. All the simulation results are summarized in the table
1.11.9 and 1.11.10.

Simulation results corresponding to θ = δn are reported in the table 1.11.9. In general, the
results are fairly consistent with the theory. The empirical level of the tests are close to
the nominal level. Apart from few isolated cases, the rejection rates are between 0.045
and 0.05. Overall, the results are considerably similar to the case of discrete income
distribution.

Simulations corresponding to other null configurations are produced in table 1.11.10. All
the results here are close to their discrete counter parts. Overall, the simulation results
justify the theory for both kinds of data.
20

1.7.3 Simulation study of power

1.8 An example using Australian household disposable in-


come data for different time periods

1.9 Conclusion

You might also like