CHAPTER - I

INTRODUCTION

Applied statisticians frequently face problems dealing with the

analysis of time-to-event data. Diverse fields such as medicine, biology,

public health, epidemiology, engineering, economics, and demography

present such data. Further the analysis of survival experiments is

complicated by issues of censoring, where an individual's life length is

known to occur only in a certain period of time, and by truncation, where

individuals enter the study only if they survive a sufficient length of time

or individuals are included in the study only if the event has occurred by a

given date. Some events which cause the individual to be randomly

censored, with respect to the event of interest, are accidental deaths,

migration of human population, death due to some cause other than the

one of interest, patient withdrawal from a clinical trial and so forth.

Outliers make statistical analysis difficult. When a sample contains

outliers, they may give rise to two superficially distinct problems. Usually,

one wishes to use the outliers in the sample to make model inferences

that are minimally affected by the number, nature or the values of any

observations. Such accommodation methods are covered under the

headings of robust and nonparametric inference.

The exponential life distribution provides a good description of the

life length of a unit which does not age with time. A property of the

exponential distribution which makes it especially important in reliability

theory and applications is that the remaining life of a used component is

independent of initial age. This property is referred to as "memory less"

 2

property. Other important life distributions are Weibull distribution, Linear

Failure Rate distribution and Makeham distribution.

The Weibull distribution has been used to describe fatigue failure

(Weibull, 1939), vaccum tube failure (Kao, 1958) and ball bearing failure

(Lieblein and Zelen, 1956). It is perhaps the most popular parametric

family of failure distributions at the present time. The Weibull distribution

belongs to the class of increasing failure rate (IFR) distribution for values

of shape parameter a> I.

1.1 Organisation of the Thesis:

This dissertation contains 6 chapters. Chapter I is an introductory

chapter containing the basic ideas needed for further development.

Chapter II deals with outlier robust estimation of the mean of

an exponential distribution under the presence of multiple outliers.

Suppose that observations Xx,.-.Xn_p arise at random from an

exponential distribution with density function

f(x:0) = -e'x/0 x >0. 0>0

0

and observations Xn_p+]....,Xn from

f(x;014) = 10 with £ < 1.

This is, of course, a scale contamination model for an exponential sample

of size n with p contaminants, which might well be expected to reveal

themselves as upper outliers. For situation of this type, Kale and Sinha

(1971) obtained the estimator for one outlier model. Joshi (1972)

extended some of the results of Kale and Sinha (1971) and obtained the

values of m for which Tm has the smallest mean squared error (MSE).
 J

For further references on this, one can refer Sinha (1973), Kale (1975),

Patel, Mudholkar and Fernando (1988) 3nd the review discussion in

Barnett and Lewis (1993).

A particular case of some interest is the full sample L-estimator of

Chikkagoudar and Kunchur (1980) in the class

d = t<*iX{i) (i.i)
;=l

where

n(n +1)
«, =-------------
n

where X(i) < X(2) < . . . <X(n) are the order statistics of the sample. This

can be interpreted as a linear combination

/=i

of separate estimators

where X = with weight

/=i

Pi = —------------- jT •

n(n +1)

This estimator in (1.1) has been shown to present varying degrees of

relative efficiency when compared with alternative estimators proposed by

Kale and Sinha (1971, a winsorized mean) and Joshi (1972), and the (0,1)

trimmed mean. It is sometimes more efficient, sometimes less so.

The principal characteristic of the estimator (1.1) is that the more

extreme upper sample values are down weighted-but only moderately so.
 4

The weights do not decrease rapidly enough to provide adequate

compensation against contaminants, particularly if there are several of

these and they are manifest as upper outliers.

Balakrishnan and Barnett (2002) proposed a modification of the

Chikkagoudar and Kunchur estimator with weights that decrease more

rapidly at the upper end of the sample. The generalised estimator of

Balakrishnan and Barnett is given below

n (k)
(k + l)(i + k-\)
A’ </)
(A' + l)
(/? + k)

where (i+k-l)(l<) = (i+k-l)(i+k-2)...i .

This class includes Chikkagoudar and Kunchur estimator as a special case

corresponding to k=l. For comparative purpose, the estimator

nX
T-=^hXw = n +1
is also included. In this chapter we propose estimators based on the

probabilities of different order statistics being outliers in the p outlier

model. Based on these probabilities bias and mean squared error of the

proposed estimators are obtained for different values of n, p and l. A

paper containing the results of this chapter has been presented at

National Conference on Statistical Inference and its applications, held at

Shivaji University, Kolhapur and sent for publication.

Increasing failure rate average (IFRA) distributions naturally arise

when coherent systems are formed from components with independent

increasing failure rate distributions. Also it naturally arises when one

considers cumulative damage shock models. The class of increasing

failure rate average distributions contains the class of increasing failure

 5

rate distributions as subclass. IFRA class of distributions is the smallest

class of distributions which contains the exponential distribution and is

closed under formation of coherent systems.

Let F be a probability distribution such that F(0)=0. Then F is an

increasing failure rate average distribution (IFRA) if [^(r)]';/ is

decreasing in t > 0, or equivalently, for x > 0, 0 < b < 1,

F(bx)>{F(x)}h (1.2)

where F = \-F. The equality in (1.2) holds if and only if F is an

exponential distribution.

Tests for exponentiality designed to detect the alternative

hypotheses relevant in reliability theory include those of Proschan and

Pyke(1967), Bickel and Doksum(1969), Ahmed(1975), Hollander and

Proschan(1972, 1975) and Koul(1977, 78). However, the test that have

been developed specifically for testing for increasing failure rate average

alternative only is due to Deshpande(1983), besides tests due to El-

Bassiouny (2003) and Hendi and Abouammoh (2001).

The genralization of the class of tests for exponentiality against

IFRA alternatives is considered in Chapter III. In this chapter, we propose

a new class of test statistic which are U-statistics whose kernel depends

on sub sample minima. An alternative expression based on ranks for the

proposed class of statistics is also given. The asymptotic normality and

consistency of the proposed class of tests is studied. Pitman ARE

comparisons, unbiasdeness of the tests and empirical powers of the tests

are studied. A paper based on this chapter has been sent for publication.

Testing exponentiality against nonparametric aging life classes has

been active area of research over the past three decades. One of the
 6

important classes tested is the class of 'increasing failure rate' (IFR). Let

X be a nonnegative random variable with distribution function F and

survival function F = \-F. We say that X is IFR if F(x + y)/F(x) is

decreasing in x > 0 for all \ >0. This definition is equivalent to saying

that the failure rate r(x)= f(x)/F(x) is nondecreasing provided that the

probability density function f(x) of F(x) exists. Many authors have

considered the problem of testing the hypothesis that F is exponential

against the alternative that it is IFR. The early contributors are Proschan

and Pyke (1967), Barlow and Proschan (1969), Bickel and Docksum

(1969) and Bickel (1969) to mention a few.

A major problem concerning the procedures suggested for testing

against the IFR alternative is that these procedures are not easy enough

to introduce them into undergraduate or beginning graduate

nonparametric statistics courses or texts. This is mainly because

calculating these test statistics is often too complicated to be done by

hand and may require extensive programming on computer. Even when

the statistics are easily calculable, their limiting distributions are not easy

to derive. To remedy this situation, Ahmad (2004) provided a new simple

approach and proposed tests for testing exponentiality against IFR and

NBU alternatives. In Chapter IV, we propose a new class of simple test

statistics based on U-statistics for testing exponentiality against IFR

alternatives. Its properties and performance in terms of Pitman ARE are

investigated. A paper based on the results of this paper has been

submitted for publication.

Mathisen(1943) proposed a test for the two-sample location

problem based on the number of observations in the X-sample not

 7

exceeding the median of the Y-sample. Besides Mathisen's test, a number

of distribution free tests are available in the literature. Wilcoxon rank sum

test is a popular nonparametric procedure for this problem. Mood's

median test is particularly effective in detecting shift in location in

distributions which are symmetric and heavy tailed. The normal scores

test is detecting shift in a normal distribution. Gastwirth's L and H tests

are effective in detecting shifts in moderately heavy tailed distributions.

The RS test due to Hogg, Fisher and Randles (1975) is effective in

detecting shifts in distributions that are skewed. The test proposed by

Shetty and Govindarajulu(1988), Shetty and Bhat(1993) based on

subsample medians takes care of specified number of outliers at the

extremes of both the samples. The tests proposed by Shetty and

Bhat(1994) are resistant to outliers in one of the samples.

Chapter V deals with the effect of random censoring on the

performance of a class of tests proposed by Shetty and Bhat (1994). ARE

comparisons are carried out for various distributions.

Stochastic dependence between random measurements is one of

the important aspects of many statistical investigations. The

quantification of this concept for bivariate distributions has been

attempted by several authors. One of the oldest measures based on

ranks is Spearman's rank correlation coefficient. Some other well-known

measures of concordance are Kendall's tau, Blomquist's Q and Hoeffding's

A to mention a few.

Suppose (X.Y) is an absolutely continuous random variable with

joint distribution function F(x.y) and survival function F(x,y). We

 8

consider the problem of testing the null hypothesis of independence

against positive quadrant dependence. That is

H0 : F(x. y) = F, (x)- F: (y) V x, y

against

H[ : F(x,y) > F,(.v)-F:( v) with strict inequality holding on a set of

nonzero probability

where F,(x) and F,(v) are marginal distribution functions of X and Y

respectively.

In Chapter VI, we consider Kendall's sample tau coefficient for

testing H0 against Hi when the variable X is subjected to random

censoring. The mean, variance and asymptotic distribution of Kendall's

sample tau coefficient in this setup are obtained in Section 3. In Section

4, Pitman ARE of Kendall's tau in the censored setup relative to

uncensored setup is obtained. Various censoring distributions are

investigated for Morgenstern and Woodworth family of distributions.