Accelerated Life Testing of One Shot Devices Data Collection and Analysis Narayanaswamy Balakrishnan Full Chapter

Accelerated Life Testing of One-shot
Devices: Data Collection and Analysis

Narayanaswamy Balakrishnan
Visit to download the full and correct content document:
https://ebookmass.com/product/accelerated-life-testing-of-one-shot-devices-data-coll
ection-and-analysis-narayanaswamy-balakrishnan/
Accelerated Life Testing of One-shot Devices
Accelerated Life Testing of One-shot Devices
Data Collection And Analysis
Narayanaswamy Balakrishnan
McMaster University
Hamilton, Canada
Man Ho Ling
The Education University of Hong Kong
Tai Po, Hong Kong SAR, China
Hon Yiu So
University of Waterloo
Waterloo, Canada
This first edition first published 2021
© 2021 by John Wiley and Sons, Inc.
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or
transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or
otherwise, except as permitted by law. Advice on how to obtain permission to reuse material from this
title is available at http://www.wiley.com/go/permissions.
The right of Narayanaswamy Balakrishnan, Man Ho Ling, and Hon Yiu So to be identified as the
author(s) of this work has been asserted in accordance with law.
Registered Office
John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, USA
Editorial Office
111 River Street, Hoboken, NJ 07030, USA
For details of our global editorial offices, customer services, and more information about Wiley
products visit us at www.wiley.com.
Wiley also publishes its books in a variety of electronic formats and by print-on-demand. Some content
that appears in standard print versions of this book may not be available in other formats.
Limit of Liability/Disclaimer of Warranty

While the publisher and authors have used their best efforts in preparing this work, they make no
representations or warranties with respect to the accuracy or completeness of the contents of this work
and specifically disclaim all warranties, including without limitation any implied warranties of
merchantability or fitness for a particular purpose. No warranty may be created or extended by sales
representatives, written sales materials or promotional statements for this work. The fact that an
organization, website, or product is referred to in this work as a citation and/or potential source of
further information does not mean that the publisher and authors endorse the information or services
the organization, website, or product may provide or recommendations it may make. This work is sold
with the understanding that the publisher is not engaged in rendering professional services. The
advice and strategies contained herein may not be suitable for your situation. You should consult with
a specialist where appropriate. Further, readers should be aware that websites listed in this work may
have changed or disappeared between when this work was written and when it is read. Neither the
publisher nor authors shall be liable for any loss of profit or any other commercial damages, including
but not limited to special, incidental, consequential, or other damages.
Library of Congress Cataloging-in-Publication Data
Names: Balakrishnan, Narayanaswamy., 1956- author. | Ling, Man Ho, author. | So, Hon
Yiu, author.
Title: Accelerated life testing of one-shot devices : data collection and
analysis / Narayanaswamy Balakrishnan, McMaster University, Hamilton,
Canada, Man Ho Ling, The Education University of Hong, Kong, New
Territories, Hong Kong, Hon Yiu So, University of Waterloo, Waterloo,
Canada.
Description: First edition. | Hoboken, NJ, USA : Wiley, 2021. | Includes
bibliographical references and index.
Identifiers: LCCN 2020035725 (print) | LCCN 2020035726 (ebook) | ISBN
9781119664000 (cloth) | ISBN 9781119664017 (adobe pdf) | ISBN
9781119663942 (epub)
Subjects: LCSH: Accelerated life testing. | Failure analysis (Engineering)
Classification: LCC TA169.3 .B35 2021 (print) | LCC TA169.3 (ebook) | DDC
620/.00452–dc23
LC record available at https://lccn.loc.gov/2020035725
LC ebook record available at https://lccn.loc.gov/2020035726
Cover Design: Wiley

Cover Image: © Piergiov/Getty Images
Set in 9.5/12.5pt STIXTwoText by SPi Global, Chennai, India
10 9 8 7 6 5 4 3 2 1
With great love and affection, we dedicate this book to
Sarah and Julia Balakrishnan, and Colleen Cutler NB

Grace Chu, Sophia Ling, and Sheldon Ling MHL
Tian Feng and Victoria So HYS
vii
Contents
Preface xi
About the Companion Website xiii
1 One-Shot Device Testing Data 1

1.1 Brief Overview 1
1.2 One-Shot Devices 1
1.3 Accelerated Life-Tests 3
1.4 Examples in Reliability and Survival Studies 4
1.4.1 Electro-Explosive Devices Data 4
1.4.2 Glass Capacitors Data 5
1.4.3 Solder Joints Data 5
1.4.4 Grease-Based Magnetorheological Fluids Data 6
1.4.5 Mice Tumor Toxicological Data 7
1.4.6 ED01 Experiment Data 7
1.4.7 Serial Sacrifice Data 7
1.5 Recent Developments in One-Shot Device Testing Analysis 10
2 Likelihood Inference 13
2.2 Under CSALTs and Different Lifetime Distributions 13
2.3 EM-Algorithm 14
2.3.1 Exponential Distribution 16
2.3.2 Gamma Distribution 18
2.3.3 Weibull Distribution 21
2.4 Interval Estimation 26
2.4.1 Asymptotic Confidence Intervals 26
2.4.2 Approximate Confidence Intervals 28
2.5 Simulation Studies 30
2.6 Case Studies with R Codes 41
viii Contents
3 Bayesian Inference 47
3.2 Bayesian Framework 47
3.3 Choice of Priors 49
3.3.1 Laplace Prior 49
3.3.2 Normal Prior 49
3.3.3 Beta Prior 50
3.5 Case Study with R Codes 59
4 Model Mis-Specification Analysis and Model Selection 65

4.2 Model Mis-Specification Analysis 65
4.3 Model Selection 66
4.3.1 Akaike Information Criterion 66
4.3.2 Bayesian Information Criterion 67
4.3.3 Distance-Based Test Statistic 68
4.3.4 Parametric Bootstrap Procedure for Testing Goodness-of-Fit 70
5 Robust Inference 79
5.2 Weighted Minimum Density Power Divergence Estimators 79
5.3 Asymptotic Distributions 81
5.4 Robust Wald-type Tests 82
5.5 Influence Function 83
6 Semi-Parametric Models and Inference 95

6.2 Proportional Hazards Models 95
6.3 Likelihood Inference 97
6.4 Test of Proportional Hazard Rates 99
7 Optimal Design of Tests 105

7.2 Optimal Design of CSALTs 105
Contents ix
7.3 Optimal Design with Budget Constraints 106

7.3.1 Subject to Specified Budget and Termination Time 107
7.3.2 Subject to Standard Deviation and Termination Time 107
7.5 Sensitivity of Optimal Designs 113
8 Design of Simple Step-Stress Accelerated Life-Tests 119

8.2 One-Shot Device Testing Data Under Simple SSALTs 119
8.3 Asymptotic Variance 121
8.3.1 Exponential Distribution 121
8.3.2 Weibull Distribution 122
8.3.3 With a Known Shape Parameter 𝑤2 124
8.3.4 With a Known Parameter About Stress Level 𝑤1 125
8.4 Optimal Design of Simple SSALT 126
8.5.1 SSALT for Exponential Distribution 128
8.5.2 SSALT for Weibull Distribution 131
9 Competing-Risks Models 141

9.2 One-Shot Device Testing Data with Competing Risks 141
9.3 Likelihood Estimation for Exponential Distribution 143
9.3.1 Without Masked Failure Modes 144
9.3.2 With Masked Failure Modes 147
9.4 Likelihood Estimation for Weibull Distribution 149
9.5 Bayesian Estimation 155
9.5.1 Without Masked Failure Modes 155
9.5.2 Laplace Prior 156
9.5.3 Normal Prior 157
9.5.4 Dirichlet Prior 157
9.5.5 With Masked Failure Modes 158
10 One-Shot Devices with Dependent Components 173

10.2 Test Data with Dependent Components 173
10.3 Copula Models 174
10.3.1 Family of Archimedean Copulas 175
10.3.2 Gumbel–Hougaard Copula 176
x Contents
10.3.3 Frank Copula 177

10.4 Estimation of Dependence 180
11 Conclusions and Future Directions 187

11.2 Concluding Remarks 187
11.2.1 Large Sample Sizes for Flexible Models 187
11.2.2 Accurate Estimation 188
11.2.3 Good Designs Before Data Analysis 188
11.3 Future Directions 189
11.3.1 Weibull Lifetime Distribution with Threshold Parameter 189
11.3.2 Frailty Models 189
11.3.3 Optimal Design of SSALTs with Multiple Stress Levels 189
11.3.4 Comparison of CSALTs and SSALTs 190
Appendix A Derivation of Hi (a, b) 191
Appendix B Observed Information Matrix 193
Appendix C Non-Identifiable Parameters for SSALTs Under Weibull

Distribution 197
Appendix D Optimal Design Under Weibull Distributions with Fixed

𝒘1 199
Appendix E Conditional Expectations for Competing Risks Model Under

Exponential Distribution 201
Appendix F Kendall’s Tau for Frank Copula 205
Bibliography 207
Author Index 217
Subject Index 221
xi
Preface
Lifetime information obtained from one-shot devices is very limited as the entire
data are either left- or right-censored. For this reason, the analysis of one-shot
device testing data poses a special challenge. This book provides several statisti-
cal inferential methods for analyzing one-shot device lifetime data obtained from
accelerated life-tests and also develops optimal designs for two mainstream accel-
erated life-tests – constant-stress and step-stress accelerated life-tests – that are
commonly used in reliability practice. The discussions provided in the book would
enable reliability practitioners to better design their experiments for data collec-
tion from efficient accelerated life-tests when there are budget constraints in place.
This is important from estimation and prediction point of view as such optimal
designs would result in as accurate an inference as possible under the constraints
imposed on the reliability experiment. Moreover, R codes are presented within
each chapter so that users can try out performing their own analysis on one-shot
device testing data.
In addition, the inferential methods and the procedures for planning accel-
erated life-tests discussed in this book are not only limited to one-shot devices
alone but also can be extended naturally to accelerated life-tests with periodic
inspections (interval-censoring) and those with continuous monitoring and
censoring (right-censoring). The book finally concludes by highlighting some
important issues and problems that are worth considering for further research.
This may be especially useful for research scholars and new researchers interested
in taking on this interesting and challenging area of research in reliability theory
and practice.
It is possible that some pertinent results or references got omitted in this book,
and we assure you that it is due to inadvertency on our part and not due to
scientific antipathy. We will appreciate greatly if the readers inform us of any
corrections/omissions, or any comments pertinent to any of the discussions in
the book!
xii Preface
Our sincere thanks go to the entire Wiley team, Ms. Mindy Okura-Marszycki,
Ms. Kathleen Santoloci, and Mr. Brett Kurzman, for taking great interest in this
project from day one, for all their help and encouragement during the whole
course, and for their fine assistance during the final production stage of the book.
Our thanks also go to our research collaborators and graduate students for their
incisive comments and queries, which always benefited us greatly and helped
clarify some of our own ideas! We express our sincere appreciation to Ms. Elena
Maria Castilla Gonzalez, a doctoral student of Professor Leandro Pardo in the
Department of Statistics and Operations Research at Complutense University
of Madrid, Spain, for her careful reading of Chapter 5 and also for sharing with
us some R codes that she had developed concerning robust inferential methods
for one-shot device test analyses. Last but not least, our special thanks go to our
families for their patience and understanding, and for providing constant support
and encouragement during our work on this book!
Finally, the first author (NB) wishes to state to his older daughter, Ms. Sarah
Balakrishnan, that though she lost out on getting his Volvo car due to a major
car accident, she should be heartened by the fact that the accident resulted in the
germination of his interest and ideas on one-shot devices (airbags), and ultimately
this book solely dedicated to the topic!
July, 2020 Narayanaswamy Balakrishnan

Man Ho Ling
Hon Yiu So
xiii
About the Companion Website
This book is accompanied by a companion website:
www.wiley.com/go/Balakrishnan/Accelerated_Life_Testing
The Student companion site will contain the codes and case studies.
1
One-Shot Device Testing Data
1.1 Brief Overview

One-shot device testing data analyses have recently received great attention in
reliability studies. The aim of this chapter is to provide an overview on one-shot
device testing data collected from accelerated life-tests (ALTs). Section 1.2 surveys
typical examples of one-shot devices and associated tests in practical situations.
Section 1.3 describes several popular ALTs, while Section 1.4 provides some
examples of one-shot device testing data that are typically encountered in relia-
bility and survival studies. Finally, Section 1.5 details some recent developments
on one-shot device testing data analyses and associated issues of interest.
1.2 One-Shot Devices

Valis et al. (2008) defined one-shot devices as units that are accompanied by
an irreversible chemical reaction or physical destruction and could no longer
function properly after its use. Many military weapons are examples of one-shot
devices. For instance, the mission of an automatic weapon gets completed
successfully only if it could fire all the rounds placed in a magazine or in ammu-
nition feed belt without any external intervention. Such devices will usually get
destroyed during usual operating conditions and can therefore perform their
intended function only once.
Shaked and Singpurwalla (1990) discussed the submarine pressure hull damage
problem from a Bayesian perspective and assessed the effect of various strengths
of underwater shock waves caused by either a nuclear device or a chemical device
on the probability of damage to a submarine pressure hull. A record is made of
whether a copy of a diminutive model of a submarine pressure hull is damaged
or not, and a specific strength of the shock wave on the model. Fan et al. (2009)
Accelerated Life Testing of One-shot Devices: Data Collection and Analysis, First Edition.
Narayanaswamy Balakrishnan, Man Ho Ling, and Hon Yiu So.
© 2021 John Wiley & Sons, Inc. Published 2021 by John Wiley & Sons, Inc.
Companion Website: www.wiley.com/go/Balakrishnan/Accelerated_Life_Testing
2 1 One-Shot Device Testing Data
considered electro-explosive devices in military applications, which induct a

current to excite inner powder and make them explode. Naturally, we cannot
adjudge the functioning condition of the electro-explosive device from its exterior,
but can only observe it by detonating it directly. After a successful detonation,
the device cannot be used anymore; if the detonation becomes a failure, we will
also not know when exactly it failed. Nelson (2003) described a study of crack
initiation for turbine wheels. Each of the 432 wheels was inspected once to
determine whether it had started to crack or not. Newby (2008) provided some
other examples of one-shot devices, such as fire extinguishers or munitions. A
full test would require the use of the considered devices and, therefore, their
subsequent destruction. The test carried out would show whether a device is still
in a satisfactory state, or has failed by that inspection time.
One-shot device testing data also arise in destructive inspection procedures,
wherein each device is allowed for only a single inspection because the test itself
results in its destruction. Morris (1987) presented a study of 52 Li/SO2 storage
batteries under destructive discharge. Each battery was tested at one of three
inspection times and then classified as acceptable or unacceptable according to a
critical capacity value.
Ideally, reliability data would contain actual failure times of all devices placed
on test (assuming, of course, the experimenter could wait until all devices fail),
so that the observed failure times can reveal the failure pattern over time, and we
could then estimate the reliability of the device reasonably. But, in practice, many
life-tests would get terminated before all the units fail. Such an early stoppage of
the life-test by the experimenter may be due to cost or time constraints or both.
This would result in what is called as “right-censored data” because the exact fail-
ure times of the unfailed devices are unknown, but all we know is that the failure
times of those devices are larger than the termination time. Considerable liter-
ature exists on statistical inference for reliability data under right-censoring; for
example one may refer to the books by Cohen (1991), Balakrishnan and Cohen
(1991), and Nelson (2003).
Moreover, when nondestructive and periodic inspections are carried on devices,
their exact failure times will not be observed, but the intervals wherein the failures
occurred will only be available. If a failure is observed by the first inspection,
then it is known that the failure time of the device is less than the first inspection
time, resulting in “left-censoring.” Similarly, if a failure is observed between two
consecutive inspection times, then it is known that the failure time is between
these two corresponding inspection times, resulting in “interval-censoring.”
Finally, the failure times of all surviving units at the final inspection time are
right-censored as their exact failure times will not be observed. Exact failure
times can only be observed from a life-testing experiment with continuous
1.3 Accelerated Life-Tests 3
surveillance. The periodic inspection process with nondestructive evaluation

would actually provide a reasonable approximation to failure times of devices
under test, especially when the inspection time intervals are short, even though
the precision of inference will be less in this case.
It is useful to note that in all the preceding examples of one-shot devices, we
will not observe the actual lifetimes of the devices. Instead, we would only observe
either a success or a failure at the inspection times, and so only the corresponding
binary data would be observed, consequently resulting in less precise inference.
In this manner, one-shot device testing data differ from typical data obtained by
measuring lifetimes in standard life-tests and, therefore, poses a unique challenge
in the development of reliability analysis, due to the lack of lifetime information
being collected from reliability experiments on such one-shot devices. If successful
tests occur, it implies that the lifetimes are beyond the inspection times, leading to
right-censoring. On the other hand, the lifetimes are before the inspection times,
leading to left-censoring, if tests result in failures. Consequently, all lifetimes are
either left- or right-censored. In such a setting of the lifetime data, Hwang and
Ke (1993) developed an iterative procedure to improve the precision of the max-
imum likelihood estimates for the three-parameter Weibull distribution and to
evaluate the storage life and reliability of one-shot devices. Some more examples
of one-shot devices in the literature include missiles, rockets, and vehicle airbags;
see, for example, Bain and Engelhardt (1991), Guo et al. (2010), and Yun et al.
(2014).
1.3 Accelerated Life-Tests

As one-shot devices (such as ammunition or automobile airbags) are usually kept
for a long time in storage and required to perform its function only once, the
reliability required from such devices during their normal operating conditions
would naturally be high. So, it would be highly unlikely to observe many fail-
ures on tests under normal operating conditions within a short period of time.
This renders the estimation of reliability of devices to be a challenging problem
from a statistical point of view. In this regard, ALTs could be utilized to mitigate
this problem. In ALTs, devices are subject to higher-than-normal stress levels to
induce early failures. In this process, more failures could likely be obtained within
a limited test time. As the primary goal of the analysis is to estimate the reliability
of devices under normal operating conditions, ALT models would then typically
extrapolate (from the data obtained at elevated stress levels) to estimate the relia-
bility under normal operating conditions. ALTs are known to be efficient in cap-
turing valuable lifetime information, especially when there is a need to shorten
the life-testing experiment. For this reason, ALTs have become popular and are
commonly adopted in many reliability experiments in practice. One may refer to
the detailed reviews presented by Nelson (1980), Cramer and Kamps (2001), Pham
(2006), and Meeker and Escobar (2014), and the excellent booklength account pro-
vided by Nelson (2009).
Constant-stress accelerated life-tests (CSALTs) and step-stress accelerated
life-tests (SSALTs) are two popular ALT plans that have received great attention
in the literature. Under a CSALT, each device gets tested at only one prespecified
stress level. To mention a few recent works, for example, Wang et al. (2014)
considered CSALTs with progressively Type-II right censored samples under
Weibull lifetime distribution; for pertinent details on progressive censoring,
see Balakrishnan (2007) and Balakrishnan and Cramer (2014). Wang (2017)
discussed CSALTs with progressive Type-II censoring under a lower truncated
distribution. Lin et al. (2019) studied CSALTs terminated by a hybrid Type-I
censoring scheme under general log-location-scale lifetime distributions. SSALTs
are an alternative to apply stress to devices in a way that stress levels will increase
at prespecified times step-by-step. For SSALTs, there are three fundamental
models for the effect of increased stress levels on the lifetime distribution of a
device: The tampered random variable model proposed by DeGroot and Goel
(1979), the cumulative exposure model of Sedyakin (1966) and Nelson (1980);
see also (Nikulin and Tahir, 2013), and the tampered failure rate model proposed
by Bhattacharyya and Soejoeti (1989). All these models of SSALTs have been
discussed extensively by many authors. Gouno (2001) analyzed data collected
from SSALTs and presented an optimal design for SSALTs; see also Gouno
(2007). Zhao and Elsayed (2005) analyzed data on the light intensity of light
emitting diodes collected from SSALTs with four stress levels under Weibull and
log-normal distributions. For the case of exponential lifetime distribution, by
considering a simple SSALT under Type-II censoring, Balakrishnan et al. (2007)
developed exact likelihood inferential methods for the model parameters; see
also Balakrishnan (2008) for details, while Xiong et al. (2006) considered the
situation when the stress changes from a low-level stress to a high-level stress at
a random time.
1.4 Examples in Reliability and Survival Studies

1.4.1 Electro-Explosive Devices Data
Fan et al. (2009) considered data, presented in Table 1.1, on 90 electro-explosive
devices under various levels of temperature at different inspection times. Ten
devices under test at each condition were inspected to see whether there were any
Table 1.1 Failure records on electro-explosive devices under CSALTs

with temperature (K).
Test Inspection Number of Number of

group time Temperature samples failures
1 10 308 10 3
2 10 318 10 1
3 10 328 10 6
4 20 308 10 3
5 20 318 10 5
6 20 328 10 7
7 30 308 10 7
8 30 318 10 7
9 30 328 10 9
Source: Fan et al. (2009).
failures or not at each inspection time for each temperature setting. These data
were then used to estimate the reliability of electro-explosive devices at different
mission times under the normal operating temperature.
1.4.2 Glass Capacitors Data

Zelen (1959) presented data from a life-test of glass capacitors at four
higher-than-usual levels of temperature and two levels of voltage. At each
of the eight combinations of temperature and voltage, eight items were tested.
We adopt these data to form one-shot device testing data by taking the inspection
times (hours) as 𝜏 ∈ {300,350, 400,450}, which are summarized in Table 1.2.
These data were then used to estimate the mean lifetime of glass capacitors for
250 V and 443 K temperature.
1.4.3 Solder Joints Data

Lau et al. (1988) considered data on 90 solder joints under three types of printed
circuit boards (PCBs) at different temperatures. The lifetime was measured as the
number of cycles until the solder joint failed, while the failure of a solder joint is
defined as a 10% increase in measured resistance. A simplified dataset is derived
from the original one and presented in Table 1.3, where two stress factors con-
sidered are temperature and a dichotomous variable indicating if the PCB type is
“copper-nickel-tin” or not.
Table 1.2 Failure records on glass capacitors under CSALTs with two stress
factors: temperature (K) and voltage (V).

group time Temperature Voltage samples failures
1 450 443 200 8 1

2 400 453 200 8 0
3 350 443 250 8 0
4 300 453 250 8 1
5 450 443 300 8 3
6 400 453 300 8 4
7 350 443 350 8 3
8 300 453 350 8 2
Source: Zelen (1959).
Table 1.3 Failure records on solder joints under CSALTs with temperature (K) and a
dichotomous variable indicating if the PCB type is “copper-nickel-tin (CNT)” or not.

group time Temperature CNT samples failures
1 300 293 Yes 10 4

2 300 333 Yes 10 4
3 100 373 Yes 10 6
4 1300 293 No 20 10
5 800 333 No 20 3
6 200 373 No 20 4
Source: Lau et al. (1988).
1.4.4 Grease-Based Magnetorheological Fluids Data

Zheng et al. (2018) studied grease-based magnetorheological fluids under SSALTs
with four levels of temperature and observed whether their viscosities or shear
stresses decreased by more than 10% after tests. Twenty samples of grease-based
magnetorheological fluids were subject to higher-than-normal operating tempera-
ture. Then, each sample was inspected only once and only whether it had failed or
not at the inspection time was observed, and not the actual failure time. The data
collected in this manner, presented in Table 1.4, were then used to estimate the
mean lifetime of grease-based magnetorheological fluids under the normal oper-
ating temperature.
1.4.5 Mice Tumor Toxicological Data

It is important to point out that one-shot device testing data arise from diverse
fields beyond reliability engineering, such as in mice tumor studies from
tumorigenicity experiments; see Kodell and Nelson (1980). In such a study, each
mouse received a particular dosage of benzidine dihydrochloride in its drinking
water and was later sacrificed to detect whether some tumors had developed by
then or not. Tumor presence can be detected only at the time of mouse’s sacrifice
or natural death. These data are summarized in Table 1.5. The data collected in
this form were then used to measure the impact of the chemical dosage on the
risk of tumor development.
1.4.6 ED01 Experiment Data

Lindsey and Ryan (1993) described experimental results conducted by National
Center for Toxicological Research in 1974. 3355 out of 24 000 female mice were ran-
domized to a control group or groups that were injected with a high dose (150 ppm)
of a known carcinogen, called 2-AAF, to different parts of the bodies. The inspec-
tion times on the mice were 12, 18, and 33 months and the outcomes of mice were
death without tumor (DNT) and death with tumor (DWT), and sacrificed with-
out tumor (SNT) and sacrificed with tumor (SWT). Balakrishnan et al. (2016a),
in their analysis, ignored the information about parts of mouse bodies where the
drugs were injected and combined SNT and SWT into one sacrificed group, and
denoted the cause of DNT as natural death and the cause of DWT as death due to
cancer. These data are summarized in Table 1.6. They then estimated the chance
of death without tumor.
1.4.7 Serial Sacrifice Data

Ling et al. (2020) were primarily concerned with the data (Berlin et al., 1979), pre-
sented in Table 1.7, on the presence or absence of two disease categories – (a)
Table 1.4 Failure records on grease-based magnetorheological fluids under SSALTs with
temperature (K).
Inspection Number of Number of

Stage time (h) Temperature samples failures
1 864 333 5 1
2 1512 339 5 1
3 1944 345 5 2
4 2160 351 5 2
Source: Zheng et al. (2018).

Table 1.5 The number of mice sacrificed, with tumor from tumorigenecity experiments
data.
Number of Number of
Test Inspection Dosage mice mice
group time (mo) Sex (ppm) sacrificed with tumor
1 9.33 F 60 72 1
2 14.00 F 60 48 3
3 18.67 F 60 36 18
4 9.33 F 120 48 0
5 14.00 F 120 47 14
6 18.67 F 120 26 25
7 9.33 F 200 47 4
8 14.00 F 200 45 38
9 9.33 F 400 24 16
10 14.00 F 400 10 9
11 9.33 M 120 48 0
12 14.00 M 120 44 7
13 18.67 M 120 42 11
14 9.33 M 200 47 3
15 14.00 M 200 32 5
16 18.67 M 200 19 8
17 9.33 M 400 24 0
18 14.00 M 400 22 11
19 18.67 M 400 15 11
Source: Kodell and Nelson (1980).
thymic lymphoma and/or glomerulosclerosis and (b) all other diseases – for an
irradiated group of 343 female mice given 𝛾-radiation and a control group of 361
radiation-free female mice to study the onset time and the rate of development of
radiation-induced disease. All of the mice in both groups were sacrificed at various
times, with the presence of a disease indicating that the disease onset occurred
before sacrifice, while the absence of a disease indicating that the disease onset
would occur after sacrifice.
Table 1.6 The number of mice sacrificed, died without tumor, and died with tumor from
the ED01 experiment data.
Number of mice
Inspection High
Test time dose Died Died
group (mo) of 2-AAF Sacrificed without tumor with tumor
1 12 No 115 22 8
2 12 Yes 110 49 16
3 18 No 780 42 8
4 18 Yes 540 54 26
5 33 No 675 200 85
6 33 Yes 510 64 51
Source: Lindsey and Ryan (1993).
Table 1.7 Serial sacrifice data on the presence or absence of two disease categories: (a)
thymic lymphoma and/or glomerulosclerosis and (b) all other diseases.
Number of mice
Test Sacrifice With With With
group time (d) 𝜸-radiation Healthy (a) only (b) only (a) and (b)
1 100 No 58 13 0 1
2 200 No 40 23 1 1
3 300 No 18 41 1 3
4 400 No 8 25 1 6
5 500 No 1 21 1 16
6 600 No 1 11 0 21
7 700 No 0 9 1 39
8 100 Yes 54 12 1 0
9 200 Yes 36 24 3 5
10 300 Yes 13 35 1 17
11 400 Yes 0 13 2 28
12 500 Yes 0 3 1 35
13 600 Yes 0 0 1 30
14 700 Yes 0 0 1 28
Source: Berlin et al. (1979).

1.5 Recent Developments in One-Shot Device Testing

Analysis
We now provide a brief review of some recent developments on one-shot device
testing data analyses under ALTs. For CSALTs, Fan et al. (2009) compared three
different prior distributions in the Bayesian approach for making predictions on
the reliability at a mission time and the mean lifetime of electro-explosive devices
under normal operating conditions. In a series of papers, Balakrishnan and Ling
(2012a,b, 2013, 2014a) developed expectation-maximization (EM) algorithms
for the maximum likelihood estimation of model parameters based on one-shot
device testing data under exponential, Weibull and gamma lifetime distributions.
In addition to parameter estimation, different methods of confidence intervals for
the mean lifetime and the reliability at a mission time under normal operating
conditions have also been discussed by these authors. The maximum likelihood
estimation as well as associated tests of hypotheses, though are most efficient
when the assumed model is indeed the true model, are known to be non-robust
when the assumed model is violated, for example, by the presence of some
outlying values in the data. With this in mind, weighted minimum density power
divergence estimators have recently been developed for one-shot device testing
data under exponential, Weibull, and gamma distributions (Balakrishnan et al.,
2019a,b, 2020a,b).
Some other important inferential aspects, apart from the works on the estima-
tion of model parameters and hypothesis tests described above, have also been
addressed by a number of authors. Balakrishnan and Ling (2014b), for instance,
have developed a procedure to obtain CSALT plans when there are budget con-
straints for testing one-shot devices. Ling and Balakrishnan (2017) also studied
model mis-specification effects on one-shot device testing data analyses between
Weibull and gamma distributions, while Balakrishnan and Chimitova (2017) con-
ducted comprehensive simulation studies to compare the performance of several
goodness-of-fit tests for one-shot device testing data.
In the framework of competing risks analysis, Balakrishnan et al. (2015, 2016a,b)
discussed the analysis of one-shot device testing data when the devices contain
multiple components and hence having multiple failure modes. Another exten-
sion that has been provided for one-shot device testing is by Ling et al. (2016) who
have developed proportional hazards models for analyzing such data. Optimal
1.5 Recent Developments in One-Shot Device Testing Analysis 11
SSALT plans for one-shot device testing experiment with lifetimes following expo-
nential and Weibull distributions have been discussed by Ling (2019) and Ling and
Hu (2020).
Pan and Chu (2010) have investigated two- and three-stage inspection schemes
for assessing one-shot devices in series systems of components having Weibull life-
time distributions. Finally, Cheng and Elsayed (2016–2018) have examined several
approaches to measure the reliability of one-shot devices with mixture of units
under various scenarios and have presented reliability metrics of systems with
mixtures of nonhomogeneous one-shot units subject to thermal cyclic stresses and
further optimal operational use of such systems.
In the chapters that follow, we shall elaborate on all these developments and
also highlight their applications.
13
Likelihood Inference
2.1 Brief Overview

Likelihood inference is one of the classical methods for the estimation of model
parameters, and it relies on the maximization of likelihood function of an assumed
model based on the observed data. In this chapter, we detail the procedure of
finding the maximum likelihood estimates (MLEs) of model parameters, mean
lifetime, and reliability under normal operating conditions on the basis of one-shot
device testing data under constant-stress accelerated life-tests (CSALTs) for differ-
ent lifetime distributions. Several associated interval estimation methods are also
discussed for all lifetime parameters of interest. The discussions provided here are
primarily from the works carried out by Balakrishnan and Ling (2012a,b, 2013,
2014a). Then, the results of a simulation study evaluating the performance of the
developed point and interval estimation methods are presented. Finally, two data
sets presented in the last chapter are used to illustrate all the inferential results
discussed here.
2.2 Under CSALTs and Different Lifetime Distributions

Suppose CSALTs are performed on I groups of one-shot devices. Suppose for i =
1, 2, … , I, in the ith test group, Ki one-shot devices are subject to J types of acceler-
ating factors (such as temperature, humidity, voltage, and pressure) at stress levels
xi = (xi1 , xi2 , … , xiJ ) and get inspected at time 𝜏i for their condition. The corre-
sponding numbers of failures, ni , are then recorded. The data thus observed can
be summarized as in Table 2.1.
For notational convenience, we use z = {𝜏i , Ki , ni , xi , i = 1, 2, … , I} to denote the
observed data, and 𝜽 for the model parameter. Suppose the lifetime in the ith test
group has a distribution with cumulative distribution function (cdf) F(t; 𝚿i ), with
Accelerated Life Testing of One-shot Devices: Data Collection and Analysis, First Edition.
Narayanaswamy Balakrishnan, Man Ho Ling, and Hon Yiu So.
© 2021 John Wiley & Sons, Inc. Published 2021 by John Wiley & Sons, Inc.
Companion Website: www.wiley.com/go/Balakrishnan/Accelerated_Life_Testing
14 2 Likelihood Inference
Table 2.1 One-shot device testing data under CSALTs with multiple accelerating factors,
various stress levels, and different inspection times.
Test Inspection
group Stress levels time Number of tested devices Number of failures
1 (x11 , x12 , … , x1J ) 𝜏1 K1 n1

2 (x21 , x22 , … , x2J ) 𝜏2 K2 n2
⋮ ⋮ ⋮ ⋮ ⋮
I (xI1 , xI2 , … , xIJ ) 𝜏I KI nI
𝚿i = 𝚿(xi ; 𝜽). The observed likelihood function is then given by

∏
I
L(𝜽; z) = C [F(𝜏i ; 𝚿i )]ni [R(𝜏i ; 𝚿i )]Ki −ni , (2.1)
i=1
where R(𝜏i ; 𝚿i ) = 1 − F(𝜏i ; 𝚿i ) is the reliability at inspection time 𝜏i and C is the

normalizing constant. We then make use of the data to determine some lifetime
characteristics of devices, from the likelihood function in (2.1), under normal
operating conditions x0 = (x01 , x02 , … , x0J ), such as the model parameter, 𝜽, mean
lifetime, 𝜇, and reliability at mission time t, R(t).
2.3 EM-Algorithm
Likelihood inference on one-shot device testing data has been discussed exten-
sively for many prominent lifetime distributions; see Balakrishnan and Ling
(2012a,b, 2013,2014a). In this section, we describe the expectation-maximization
(EM) framework for determining the MLEs of all parameters of interest.
In one-shot device testing, as mentioned earlier, no actual lifetimes are
observed, and, as such, all observed data are censored. EM algorithm is known
to be a convenient and efficient method for estimating model parameters in the
presence of censoring; see, for example, McLachlan and Krishnan (2008) for all
pertinent details concerning this method, its variations, and applications. This
method has now become a standard method of model-fitting in the presence of
missing data. The EM algorithm involves two steps in each iteration of the numer-
ical method of maximizing the likelihood function such as the one presented in
(2.1): expectation-step (E-step) in which the missing data are approximated by
their expected values, and maximization-step (M-step) in which the likelihood
function, with imputed values replacing the missing data, gets maximized.
2.3 EM-Algorithm 15
Suppose the lifetimes of devices follow a distribution with probability density

function (pdf) f (t; 𝜽). We first note that the log-likelihood function, based on the
complete data, is
∑ ∑
I Ki
𝓁c (𝜽) = ln(f (tik ; 𝜽)). (2.2)
i=1 k=1
In the mth step of the iterative process, the objective then is to update the esti-
mate of parameter 𝜽 by the value that maximizes the function
Q(𝜽, 𝜽(m) ) = E𝜽(m) [𝓁c (𝜽)|z], (2.3)
based on the current estimate 𝜽(m) . The current estimate 𝜽(m) is used in obtaining
the updated estimate 𝜽(m+1) in the M-step, and then the conditional expectation
Q(𝜽, 𝜽(m+1) ) in (2.3) is obtained based on 𝜽(m+1) in the E-step. These two steps are
then repeated until convergence is achieved to a desired level of accuracy. It is
evident that we actually solve the incomplete data problem of maximizing the like-
lihood function in (2.1) by first approximating the missing data and then using the
approximated values to find the estimate of the parameter vector as the solution
for the complete data problem.
However, in the maximization step, a closed-form solution may not always
be found, and in such a case, a one-step Newton–Raphson method can be
employed for this purpose. This would require the second-order derivatives of the
log-likelihood function with respect to the model parameters. In the present case,
let us introduce
[ ]
(m) 𝜕Q(𝜽, 𝜽(m) )
I = (2.4)
𝜕𝜽 𝜽=𝜽(m)
and
[ 2 ]
𝜕 Q(𝜽, 𝜽(m) )
J(m)
=− . (2.5)
𝜕𝜽 𝜕𝜽′ 𝜽=𝜽(m)
Then, the updated estimate obtained through the one-step Newton–Raphson
method is given by
𝜽(m+1) = (J(m) )−1 I(m) + 𝜽(m) . (2.6)
In addition to estimates of model parameters, our interest may also be on
some lifetime characteristics, such as mean lifetime and reliability under normal
operating conditions. As these quantities are functions of model parameters, their
estimates can be readily found by plugging in the MLEs of model parameters
into the respective functions, and then the delta method can be employed to
determine the corresponding standard errors as well; see, for example, Casella
and Berger (2002).
2.3.1 Exponential Distribution

Let us first start with the exponential distribution, which is one of the most popular
lifetime distributions; see, for example, Johnson et al. (1994) and Balakrishnan
and Basu (1995) for developments on various methods and applications of the
exponential distribution. In this case, Fan et al. (2009) and Balakrishnan and Ling
(2012a,b) studied inferential methods for one-shot device testing data under expo-
nential distribution with single and multiple accelerating factors. Let us use E to
denote the exponential distribution in what follows. In this case, in the ith test
group, the cdf and pdf of the exponential distribution, with rate parameter 𝜆i > 0,
are given by
FE (t; 𝜆i ) = 1 − exp(−𝜆i t), t > 0, (2.7)
and
𝜕FE (t; 𝜆i )
fE (t; 𝜆i ) = = 𝜆i exp(−𝜆i t), t > 0, (2.8)
𝜕t
respectively. Its mean lifetime and reliability at mission time t under normal oper-
ating conditions x0 = (x01 , x02 , … , x0J ) are then given by
∞
1
𝜇E = tfE (t; 𝜆0 )dt =
∫0 𝜆0
and
RE (t) = 1 − FE (t; 𝜆0 ) = exp(−𝜆0 t),
respectively. As it is necessary to extrapolate the data observed from the elevated

stress levels xi = (xi1 , xi2 , … , xiJ ), to lifetime characteristics of devices under nor-
mal operating conditions x0 = (x01 , x02 , … , x0J ), the rate parameter 𝜆i in (2.7) gets
related to the stress levels in a log-linear link function of the form, with xi0 = 1,
∑
J
ln(𝜆i ) = ej xij , (2.9)
j=0
where 𝜽E = (e0 , e1 , … , eJ ) now becomes the new model parameter.

It is worth mentioning that many well-known stress-rate models, such as
Arrhenius, power law, inverse power law, and Eyring models, are all special cases
of the log-linear link function in (2.9), with suitable transformations on the stress
variables; see, for example, Wang and Kececioglu (2000).
Under the EM framework, the log-likelihood function based on the complete
data, obtained readily from (2.8), is given by
{ J ( J ) K }
∑ ∑
I Ki
∑
I
∑ ∑ ∑ i
𝓁c = ln(fE (tik ; 𝜆i )) = Ki ej xij − exp ej xij tik ,

i=1 k=1 i=1 j=0 j=0 k=1
2.3 EM-Algorithm 17
and the conditional expectation in (2.3) becomes

{ J ( J ) }
(m)
∑I
∑ ∑ (m)
Q(𝜽E , 𝜽E ) = Ki ej xij − exp ej xij Ai , (2.10)
i=1 j=0 j=0
where Ai(m) = E[Ti |z, 𝜃E(m) ] can be computed in the E-step, as described below.
The first-order derivatives with respect to the model parameters, required for
the maximization of the quantity Q(𝜽E , 𝜽E(m) ) in (2.10), are then given by
𝜕Q(𝜽E , 𝜽E(m) ) ∑
I
= Ki xij (1 − 𝜆i Ai(m) ), j = 0, 1, … , J. (2.11)
𝜕ej i=1
A closed-form solution is not available for (2.11), however. So, in the M-step of
the EM algorithm, we employ the one-step Newton–Raphson method, as men-
tioned earlier. This requires the second-order derivatives of (2.10) with respect to
the model parameters, which are given by
𝜕 2 Q(𝜽E , 𝜽E(m) ) ∑
I
= − Ki xip xiq 𝜆i Ai(m) , p, q = 0, 1, … , J.
𝜕ep 𝜕eq i=1
Next, the conditional expectation required in the E-step is easy to derive as it

only involves the failed and unfailed devices. For the failed devices, the proportion
is ni ∕Ki , and the lifetimes in this case are right-truncated at the inspection time 𝜏i
(as they cannot be more than 𝜏i ). Similarly, for the unfailed devices, the proportion
is 1 − ni ∕Ki , and the lifetimes are left-truncated at the inspection time 𝜏i (as they
cannot be less than 𝜏i ). Consequently, we obtain the conditional expectation of Ti ,
given the observed data and the current estimate 𝜽(m) E
of 𝜽E , as follows:
Ai(m) = E[Ti |z, 𝜽E(m) ]

( ) 𝜏 ( ) ∞ (m)
ni ∫0 i tfE (t; 𝜽E(m) )dt ni ∫𝜏i tfE (t; 𝜽E )dt
= + 1 −
Ki FE (𝜏i ; 𝜽E(m) ) Ki 1 − F (𝜏 ; 𝜽(m) )
E i E
( )( (m)
) ( )( )
ni 1 𝜏i exp(−𝜆i 𝜏i ) ni 1
= − + 1− 𝜏i + (m)
Ki 𝜆(m)
i
FE (𝜏i ; 𝜽(m)
E
) Ki 𝜆i
( )
1 ̂ i)
F(𝜏
= (m) + 𝜏i 1 − , (2.12)
𝜆i FE (𝜏i ; 𝜽E(m) )
∑J
̂ i ) = ni ∕Ki is the empirical probability of fail-
where 𝜆i(m) = exp( j=0 ej(m) xij ) and F(𝜏
ure by inspection time 𝜏i .
When the MLEs of the model parameters 𝜽̂ E = (̂e0 , ê 1 , … , ê J ) are determined, we
can readily have the MLEs of mean lifetime and reliability at mission time t under
normal operating conditions x0 = (x01 , x02 , … , x0J ), for example, to be

( J )
1 ∑
𝜇̂ E = = exp − ê j x0j
𝜆̂ 0 j=0
and
{ ( ) }
∑
J
R̂ E (t) = exp(−𝜆̂ 0 t) = exp − exp ê j x0j t ,
j=0
respectively.
2.3.2 Gamma Distribution

As the exponential distribution would only be suitable for modeling constant
hazard rate, it would be natural to consider the life distribution to be gamma as
it possesses increasing, decreasing, and constant hazard rate properties; see, for
example, Johnson et al. (1994). Moreover, it includes exponential distribution as
a particular case. For this reason, Balakrishnan and Ling (2014a) extended the
work to gamma lifetime distribution with multiple accelerating factors with both
scale and shape parameters varying over stress factors.
Let us use G to denote gamma distribution in what follows. Specifically, in the
ith test group, we assume the lifetimes to follow a gamma distribution with shape
parameter 𝛼i > 0 and scale parameter 𝛽i > 0, with cdf and pdf given by
( )
t
FG (t; 𝛼i , 𝛽i ) = 𝛾 𝛼i , , t > 0, (2.13)
𝛽i
and
( )
𝜕FG (t; 𝛼i , 𝛽i ) t𝛼i −1 t
fG (t; 𝛼i , 𝛽i ) = = 𝛼 exp − , t > 0, (2.14)
𝜕t Γ(𝛼i )𝛽i i 𝛽i
s
respectively, where 𝛾(z, s) = ∫0 xz−1 exp(−x)dx∕Γ(z) is the lower incomplete
∞
gamma ratio and Γ(z) = ∫0 xz−1 exp(−x)dx is the complete gamma function. The
mean lifetime and reliability at mission time t under normal operating conditions
x0 = (x01 , x02 , … , x0J ) are, respectively, given by
∞
𝜇G = tfG (t; 𝛼0 , 𝛽0 )dt = 𝛼0 𝛽0
∫0
and
( )
t
RG (t) = 1 − FG (t; 𝛼0 , 𝛽0 ) = Γ 𝛼0 , ,
𝛽0
where Γ(z, s) = 1 − 𝛾(z, s) is the upper incomplete gamma ratio. For extrap-
olating the data observed from the elevated stress levels xi = (xi1 , xi2 , … , xiJ )
2.3 EM-Algorithm 19
to lifetime characteristics of devices under normal operating conditions x0 =

(x01 , x02 , … , x0J ), in this case, both parameters 𝛼i and 𝛽i are related to the stress
levels in log-linear link functions of the form, with xi0 = 1,
∑
J
∑
J
ln(𝛼i ) = aj xij and ln(𝛽i ) = bj xij . (2.15)
j=0 j=0
Thus, the new model parameter of the considered model is
𝜽G = (a0 , a1 , … , aJ , b0 , b1 , … , bJ ).
Here, under the EM framework, the log-likelihood function, based on the com-
plete data, is readily obtained from (2.14) to be
∑ ∑
I Ki
𝓁c = ln(fG (tik ; 𝛼i , 𝛽i ))
i=1 k=1
( )
∑
I
∑Ki
1∑
Ki
= −Ki ln(Γ(𝛼i )) − Ki 𝛼i ln(𝛽i ) + (𝛼i − 1) ln(tik ) − t .
i=1 k=1
𝛽i k=1 ik
Let Ai(m) = E[ln(Ti )|z, 𝜽(m)

G
] and B(m)
i
= E[Ti |z, 𝜽G(m) ]. Then, the conditional expec-
tation in (2.3) becomes
( )
(m)
∑ I
(m)
B(m)
i
Q(𝜽G , 𝜽G ) = Ki − ln(Γ(𝛼i )) − 𝛼i ln(𝛽i ) + (𝛼i − 1)Ai − . (2.16)
i=1
𝛽i
the maximization of the quantity Q(𝜽G , 𝜽(m)
G
) in (2.16), are given by
𝜕Q(𝜽G , 𝜽G(m) ) ∑
I
= − Ki xij 𝛼i (Ψ(𝛼i ) + ln(𝛽i ) − A(m) ),
𝜕aj i=1
i
( )
𝜕Q(𝜽G , 𝜽G(m) ) ∑
I
B(m)
i
= − Ki xij 𝛼i − ,
𝜕bj i=1
𝛽i
for j = 0, 1, … , J. The corresponding second-order derivatives are

𝜕 2 Q(𝜽G , 𝜽G(m) ) ∑
I
= − Ki xip xiq 𝛼i (𝛼i Ψ′ (𝛼i ) + Ψ(𝛼i ) + ln(𝛽i ) − Ai(m) ),
𝜕ap 𝜕aq i=1
𝜕 2 Q(𝜽G , 𝜽G(m) ) ∑ I
Ki xip xiq Bi(m)
=− ,
𝜕bp 𝜕bq i=1
𝛽i
𝜕 2 Q(𝜽G , 𝜽G(m) ) ∑
I
= − Ki xip xiq 𝛼i ,
𝜕ap 𝜕bq i=1
for p, q = 0, 1, … , J, where Ψ(z) = 𝜕 ln Γ(z)∕𝜕z and Ψ′ (z) are the digamma and
trigamma functions, respectively; see Abramowitz and Stegun (1972) for a
discussion on the computation of these functions.
Here again, the conditional expectations required in the E-step, Ai(m) and B(m) i
,
only involve the failed and unfailed devices. For the failed devices, the proportion
is ni ∕Ki , and the lifetimes are right-truncated at the inspection time 𝜏i , while for the
unfailed devices, the proportion is 1 − ni ∕Ki , and the lifetimes are left-truncated
at the inspection time 𝜏i . Consequently, we obtain the conditional expectations,
given the observed data and the current estimate 𝜽G(m) of 𝜽G , as follows:
[ ( ) ]
(m) Ti | (m)
Ai = E ln |z, 𝜽G + ln(𝛽i(m) )
𝛽i(m) |
( )
𝜏 t (m)
( ) ∫0 i ln 𝛽 (m) fG (t; 𝜽G )dt
ni i
= ln(𝛽i(m) ) +
Ki FG (𝜏i ; 𝜽(m) )
( ) G
∞ t (m)
( ) ∫𝜏i ln 𝛽 (m) fG (t; 𝜽G )dt
ni i
+ 1−
Ki 1 − FG (𝜏i ; 𝜽(m)
G
)
( ) 𝜏 ∕𝛽i(m) (m)
ni ∫0 i ln(t)t𝛼i −1
exp(−t)dt
= ln(𝛽i(m) ) +
Ki Γ(𝛼i(m) )FG (𝜏i ; 𝜽(m)
G
)
( )∫∞ ln(t)t𝛼i
(m)
−1
exp(−t)dt
n 𝜏i ∕𝛽i(m)
+ 1− i
Ki Γ(𝛼 (m) )(1 − FG (𝜏i ; 𝜽G(m) ))
( i )( )
(m) (m) 𝜏i ̂ i)
F(𝜏
= ln(𝛽i ) + H1 𝛼i , (m)
𝛽 FG (𝜏i ; 𝜽(m) )
{ ( i )} ( G )
(m) (m) 𝜏i ̂ i)
1 − F(𝜏
+ Ψ(𝛼i ) − H1 𝛼i , (m)
𝛽i 1 − FG (𝜏i ; 𝜽(m)
G
)
and
Bi(m) = E[Ti |z, 𝜽(m)G
]
( ) 𝜏i ( ) ∞ (m) (m)
ni ∫𝜏i tfG (t; 𝛼i , 𝛽i )dt
(m) (m)
ni ∫0 tfG (t; 𝛼i , 𝛽i )dt
= + 1−
Ki FG (𝜏i ; 𝜽G(m) ) Ki 1 − FG (𝜏i ; 𝜽(m) )
( )( ) G
𝜏i ̂ i)
F(𝜏
= 𝛼i(m) 𝛽i(m) 𝛾 𝛼i(m) + 1, (m)
𝛽i FG (𝜏i ; 𝜽G(m) )
( )( )
(m) (m) (m) 𝜏i ̂ i)
1 − F(𝜏
+ 𝛼i 𝛽i Γ 𝛼i + 1, (m) ,
𝛽i 1 − FG (𝜏i ; 𝜽(m)G
)
2.3 EM-Algorithm 21
where
b
1
H1 (a, b) = ln(x)xa−1 exp(−x)dx
Γ(a) ∫0
ba F (a, a; a + 1, a + 1; −b)
= ln(b)𝛾(a, b) − 2 2 ,
a2Γ(a)
and n Fm (a1 , a2 , … , an ; b1 , b2 , … , bm ; z) is the generalized hypergeometric func-
tion; one may refer to Slater (2008) for properties and computational formulas for
generalized hypergeometric functions. The generalized hypergeometric function
is available in mathematical programs such as Matlab and Maple. Details on the
derivation of the above expression of H1 (a, b) are presented in Appendix A.
We then have the MLEs of mean lifetime and reliability at mission time t under
normal operating conditions x0 = (x01 , x02 , … , x0J ), using the MLEs of the model
parameters 𝜽̂ G = (â0 , â1 , … , âJ , b̂ 0 , b̂ 1 , … , b̂ J ), to be
( J )
∑
𝜇̂ = 𝛼̂ 𝛽̂ = exp
G 0 0 (â + b̂ )x
j j 0j
j=0
and
( )
t
R̂ G (t) = Γ 𝛼̂ 0 , ,
𝛽̂0
where, from (2.15),
( J ) ( )
∑ ∑
J
𝛼̂ 0 = exp âj x0j and 𝛽̂0 = exp b̂ j x0j .
j=0 j=0
2.3.3 Weibull Distribution

Similar to gamma distribution, Weibull distribution can also model increasing,
decreasing, and constant hazard rates, and it also includes exponential distribu-
tion as a special case; see, for example, Johnson et al. (1994) and Murthy et al.
(2003) on various methods and applications of Weibull distribution. In this case,
Balakrishnan and Ling (2013) studied one-shot device testing data for the Weibull
lifetime distribution with both scale and shape parameters varying over multiple
stress factors.
Let us use W to denote the Weibull distribution in what follows. Specifically,
in the ith test group, we assume that the lifetimes of devices follow the Weibull
distribution, with shape parameter 𝜂i > 0 and scale parameter 𝛽i > 0, with cdf and
pdf as
( ( )𝜂 i )
t
FW (t; 𝜂i , 𝛽i ) = 1 − exp − , t > 0, (2.17)
𝛽i
and
( )𝜂i −1 ( ( )𝜂 i )
𝜕FW (t; 𝜂i , 𝛽i ) 𝜂 t t
fW (t; 𝜂i , 𝛽i ) = = i exp − , t > 0, (2.18)
𝜕t 𝛽i 𝛽i 𝛽i
respectively. The mean lifetime and reliability at mission time t under normal oper-
ating conditions x0 = (x01 , x02 , … , x0J ) are
∞ ( )
1
𝜇W = tfW (t; 𝜂0 , 𝛽0 )dt = 𝛽0 Γ 1 + (2.19)
∫0 𝜂0
and
( ( )𝜂 0 )
t
RW (t) = 1 − FW (t; 𝜂0 , 𝛽0 ) = exp − , (2.20)
𝛽0
respectively. For extrapolating the data observed from the elevated stress levels
xi = (xi1 , xi2 , … , xiJ ) to the lifetime characteristics of devices under normal oper-
ating conditions x0 = (x01 , x02 , … , x0J ), in this case, both parameters 𝜂i and 𝛽i are
related to the stress levels in log-linear link functions of the form, with xi0 = 1,
∑
J
∑
J
ln(𝜂i ) = rj xij and ln(𝛽i ) = sj xij . (2.21)
j=0 j=0
Thus, the new model parameter of the considered model is

𝜽W = (r0 , r1 , … , rJ , s0 , s1 , … , sJ ).
Instead of working with Weibull lifetimes for lifetime data, it is often more con-
venient to work with extreme value distribution for log-lifetimes. This is because
the Weibull distribution belongs to the scale-shape family, while the extreme value
distribution belongs to a location-scale family, which makes it convenient to work
with; see Johnson et al. (1995). For this reason, we consider here the extreme value
distribution with corresponding pdf and cdf as
𝜉i exp(−𝜉i )
fV (𝑣; 𝜇i , 𝜎i ) = , −∞ < 𝑣 < ∞, (2.22)
𝜎i
and
FV (𝑣; 𝜇i , 𝜎i ) = 1 − exp(−𝜉i ), −∞ < 𝑣 < ∞, (2.23)
∑J
where 𝑣 = ln(t), 𝜉i = exp((𝑣 − 𝜇i )∕𝜎i ), with link functions 𝜇i = ln(𝛽i ) = j=0 sj xij
∑J
and 𝜎i = 1∕𝜂i = exp(− j=0 rj xij ).
We now define 𝑣ik = ln(tik ) and 𝜉ik = exp((𝑣ik − 𝜇i )∕𝜎i ). Then, under the EM
framework, the log-likelihood function, based on the complete data, is obtained
readily from (2.22) as
∑ ∑
I Ki
𝓁c = ln(fV (𝑣ik ; 𝜇i , 𝜎i ))
i=1 k=1
2.3 EM-Algorithm 23
( )
∑
I
∑
J
∑
Ki
∑
Ki
= Ki rj xij + ln(𝜉ik ) − 𝜉ik .
i=1 j=0 k=1 k=1
Let Ai(m) = E[ln(𝜉i )|z, 𝜽(m)

W
] and B(m)
i
= E[𝜉i |z, 𝜽(m)
W
]. Then, the conditional expecta-
tion, from (2.3), is
( J )
(m)
∑ I
∑ (m) (m)
Q(𝜽W , 𝜽W ) = Ki rj xij + Ai − Bi . (2.24)
i=1 j=0
the maximization of the quantity Q(𝜽W , 𝜽(m)
W
) in (2.24), are given by
(m)
𝜕Q(𝜽W , 𝜽W ) ∑
I
= Ki xij (1 + A(m) − Ci(m) ),
𝜕rj i=1
i
(m)
𝜕Q(𝜽W , 𝜽W ) ∑ I
Ki xij (1 − B(m)
i
)
=− ,
𝜕sj i=1
𝜎i
for j = 0, 1, … , J, where Ci(m) = E[𝜉i ln(𝜉i )|z, 𝜽W (m)

]. The corresponding second-order
derivatives are
𝜕 2 Q(𝜽W , 𝜽(m)
W
) ∑ I
= Ki xip xiq (Ai(m) − Ci(m) − D(m) ),
𝜕rp 𝜕rq i=1
i
𝜕 2 Q(𝜽W , 𝜽(m)
W
) ∑ I
Ki xip xiq B(m)
i
=− ,
𝜕sp 𝜕sq i=1 𝜎i2
𝜕 2 Q(𝜽W , 𝜽(m)
W
) ∑ I
Ki xip xiq (1 − B(m)
i
− Ci(m) )
=− ,
𝜕rp 𝜕sq i=1
𝜎i
for p, q = 0, 1, … , J, where D(m)

i
= E[𝜉i (ln(𝜉i ))2 |z, 𝜽(m)
W
].
The required conditional expectations in the E-step, namely, Ai(m) , Bi(m) , Ci(m)
and Di(m) , again involve only the failed and unfailed devices. For the failed
devices, the proportion is ni ∕Ki and the lifetimes are right-truncated at the
inspection time 𝜏i , while for the unfailed devices, the proportion is 1 − ni ∕Ki
and the lifetimes are left-truncated at the inspection time 𝜏i . Upon defin-
ing 𝜈i(m) = exp((ln(𝜏i ) − 𝜇i(m) )∕𝜎i(m) ), we can derive the required conditional
expectations to be as follows:
Ai(m) = E[ln(𝜉i )|z, 𝜽(m)
W
]
( ) ln(𝜏i ) ( ) ∞ (m)
ni ∫0 ln(𝜉)fV (𝑣; 𝜽(m)
W
)d𝑣 ni ∫ln(𝜏i ) ln(𝜉)fV (𝑣; 𝜽W )d𝑣
= (m)
+ 1 −
Ki FV (ln(𝜏i ); 𝜽W ) Ki 1 − FV (ln(𝜏i ); 𝜽(m)
W
)
( ) 𝜈i(m) ( )∫∞ ln(x) exp(−x)dx
ni ∫0 ln(x) exp(−x)dx n 𝜈i(m)
= (m)
+ 1− i (m)
Ki FW (𝜏i ; 𝜽W ) Ki 1 − FW (𝜏i ; 𝜽W )
{ }( )
ln(𝜈i(m) ) ̂ i)
F(𝜏
=− 𝛾+ + EI(𝜈i(m) )
exp(𝜈i(m) ) (m)
FW (𝜏i ; 𝜽W )
{ }( )
ln(𝜈i(m) ) ̂ i)
1 − F(𝜏
+ + EI(𝜈i(m) ) ,
exp(𝜈i(m) ) 1 − FW (𝜏i ; 𝜽(m)
W
)
Bi(m) = E[𝜉i |z, 𝜽W(m)

]
( ) ln(𝜏i ) ( ) ∞ (m)
ni ∫ln(𝜏i ) 𝜉fV (𝑣; 𝜽W )d𝑣
(m)
ni ∫0 𝜉fV (𝑣; 𝜽W )d𝑣
= + 1−
Ki FV (ln(𝜏i ); 𝜽(m)
W
) Ki 1 − F (ln(𝜏 ); 𝜽(m) )
V i W
( ) 𝜈i(m) ( )∫∞ (m) x exp(−x)dx
ni ∫0 x exp(−x)dx n 𝜈i
= (m)
+ 1− i
Ki FW (𝜏i ; 𝜽W ) Ki 1 − F (𝜏 ; 𝜽(m) )
W i W
{ } ( )
1 + 𝜈i(m) ̂ i)
F(𝜏
= 1−
exp(𝜈i(m) ) (m)
FW (𝜏i ; 𝜽W )
{ }( )
1 + 𝜈i (m)
̂ i)
1 − F(𝜏
+ ,
exp(𝜈i(m) ) (m)
1 − FW (𝜏i ; 𝜽W )
Ci(m) = E[𝜉i ln(𝜉i )|z, 𝜽W(m)

]
( ) ln(𝜏i )
ni ∫0 𝜉 ln(𝜉)fV (𝑣; 𝜽(m)
W
)d𝑣
= (m)
Ki FV (ln(𝜏i ); 𝜽W )
( ) ∫ ∞ 𝜉 ln(𝜉)f (𝑣; 𝜽(m) )d𝑣
n ln(𝜏i ) V W
+ 1− i (m)
Ki 1 − FV (ln(𝜏i ); 𝜽W )
( ) 𝜈i(m) ( ) ∞
ni ∫0 x ln(x) exp(−x)dx ni ∫𝜈i(m) x ln(x) exp(−x)dx
= (m)
+ 1−
Ki FW (𝜏i ; 𝜽W ) Ki 1 − FW (𝜏i ; 𝜽(m)
W
)
{ }( )
1 + ln(𝜈i(m) ) + 𝜈i(m) ln(𝜈i(m) ) ̂ i)
F(𝜏
= 1−𝛾 − EI(𝜈i(m) ) −
exp(𝜈i(m) ) (m)
FW (𝜏i ; 𝜽W )
{ }( )
1 + ln(𝜈i(m) ) + 𝜈i(m) ln(𝜈i(m) ) ̂ i)
1 − F(𝜏
+ EI(𝜈i(m) ) + ,
exp(𝜈i(m) ) (m)
1 − FW (𝜏i ; 𝜽W )
Di(m) = E[𝜉i (ln(𝜉i ))2 |z, 𝜽W

(m)
]
( ) ln(𝜏i ) (m)
ni ∫0 𝜉(ln(𝜉))2 fV (𝑣; 𝜽W )d𝑣
= (m)
Ki FV (ln(𝜏i ); 𝜽W )
2.3 EM-Algorithm 25
( ) ∞ (m)
ni ∫ln(𝜏i ) 𝜉(ln(𝜉)) fV (𝑣; 𝜽W )d𝑣
2
+ 1−
Ki 1 − F (ln(𝜏 ); 𝜽(m) ) V i W
( ) 𝜈 (m)
ni ∫0 i x(ln(x))2 exp(−x)dx
= (m)
Ki FW (𝜏i ; 𝜽W )
( ) ∞
ni ∫𝜈i(m) x(ln(x)) exp(−x)dx
2
+ 1−
Ki 1 − FW (𝜏i ; 𝜽(m) )
( ) W
̂
F(𝜏i )
= H2 (2, 𝜈i(m) ) (m)
FW (𝜏i ; 𝜽W )
{ }( ̂ i)
)
𝜋2 (m) 1 − F(𝜏
+ 𝛾 + 2
− 2𝛾 − H2 (2, 𝜈i ) ,
6 1 − FW (𝜏i ; 𝜽(m)
W
)
where
b
1
H2 (a, b) = (ln(x))2 x exp(−x)dx
Γ(a) ∫0
2ba ln (b)2 F2 (a, a; a + 1, a + 1; −b)
= (ln(b))2 𝛾(a, b) −
a2Γ(a)
2ba 3 F3 (a, a, a; a + 1, a + 1, a + 1; −b)
+ ,
a3Γ(a)
∞
𝛾 ≈ 0.577215665 is the Euler’s constant, EI(a) = ∫a x−1 exp(−x)dx is the exponen-
tial integral, and n Fm (a1 , a2 , … , an ; b1 , b2 , … , bm ; z) is the generalized hypergeo-
metric function. The exponential integral is available in mathematical programs
such as R, Matlab and Maple. Details on the derivation of the above expression of
H2 (a, b) are presented in Appendix A.
We then have the MLEs of mean lifetime and reliability at mission time t under
normal operating conditions x0 = (x01 , x02 , … , x0J ), when the MLEs of the model
parameters 𝜽̂ W = (̂r 0 , r̂ 1 , … , r̂ J , ŝ0 , ŝ1 , … , ŝJ ) have been determined, from (2.19)
and (2.20), to be
( )
1
𝜇̂ W = 𝛽̂0 Γ 1 +
𝜂̂ 0
and
( )𝜂̂ 0
⎛ ⎞
t
R̂ W (t) = exp ⎜− ⎟,
⎜ 𝛽̂0 ⎟
⎝ ⎠
where
( ) ( J )
∑
J
∑
𝜂̂ 0 = exp r̂ j x0j and 𝛽̂0 = exp ŝj x0j .
j=0 j=0
2.4 Interval Estimation

2.4.1 Asymptotic Confidence Intervals
To construct confidence intervals for parameters of interest, asymptotic confidence
intervals are commonly used, in conjunction with MLEs, when sample sizes are
sufficiently large. This would require the asymptotic variance-covariance matrix
of the MLEs of model parameters, which could be obtained as the inverse of the
observed information matrix.
Under the EM framework, the missing information principle (Louis, 1982) is
often employed for obtaining the observed information matrix of the MLEs of
model parameters. In the case of one-shot device testing data, if 𝜽̂ is the maximum
likelihood estimate of 𝜽, the observed information matrix obtained from the miss-
ing information principle is equivalent to the observed Fisher information matrix
computed directly as
( )( )( )
∑I
1 1 𝜕F(𝜏i ; 𝜽) 𝜕F(𝜏i ; 𝜽)
Iobs (𝜽) = Ki + . (2.25)
i=1
F(𝜏i ; 𝜽) 1 − F(𝜏i ; 𝜽) 𝜕𝜽 𝜕𝜽′
Details of derivation of the observed information matrix in (2.25) are presented in
Appendix B.
The first-order derivatives of the cdf, F(𝜏i ; 𝜽), under the exponential, gamma,
and Weibull lifetime distributions, with respect to the corresponding model
parameters (exponential: 𝜽E = (ej ), gamma: 𝜽G = (aj , bj ), Weibull: 𝜽W = (rj , sj ) for
j = 0, 1, … , J), can be found to be as follows:
𝜕FE (𝜏i ; 𝜽E )
= −𝜏i 𝜆i xij exp(−𝜏i 𝜆i ), (2.26)
𝜕ej
( ( ) ( ))
𝜕FG (𝜏i ; 𝜽G ) 𝜏 𝜏
= −𝛼i xij Ψ(𝛼i )𝛾 𝛼i , i − H1 𝛼i , i , (2.27)
𝜕aj 𝛽i 𝛽i
( )𝛼i ( )
𝜕FG (𝜏i ; 𝜽G ) xij 𝜏i 𝜏
=− exp − i , (2.28)
𝜕bj Γ(𝛼i ) 𝛽i 𝛽i
𝜕FW (𝜏i ; 𝜽W )
= xij ln(𝜈i )𝜈i exp(−𝜈i ), (2.29)
𝜕rj
𝜕FW (𝜏i ; 𝜽W ) xij
= − 𝜈i exp(−𝜈i ). (2.30)
𝜕sj 𝜎i
Then, the asymptotic variance-covariance matrix of the MLEs of model param-
eters is the inverse of the observed information matrix, i.e.
−1
V𝜽 = Iobs (𝜽).
Furthermore, as mentioned earlier in the beginning of Section 2.3, the asymp-
totic variance of the MLE of a function of model parameter 𝜽, such as mean
Table 2.2 The first-order derivatives of mean lifetime and reliability function under
normal operating conditions x0 a) with respect to model parameters under the
exponential, gamma and Weibull distributions.
Mean lifetime 𝝁
Model Model parameter The first-order derivative
x
Exponential ej − 𝜆0j
0
Gamma aj x0j 𝛼0 𝛽0
bj x0j 𝛼0 𝛽0
Weibull rj −x0j 𝜎0 exp(𝜇0 )Ψ(1 + 𝜎0 )Γ(1 + 𝜎0 )
sj x0j exp(𝜇0 )Γ(1 + 𝜎0 )
Reliability R(t)
Model Model parameter The first-order derivative
Exponential ej t𝜆0 x0j exp(−t𝜆0 )

( ( ) ( ))
Gamma aj 𝛼0 x0j Ψ(𝛼0 )𝛾 𝛼0 , 𝛽t − H1 𝛼0 , t
𝛽0
x0j
( )𝛼0 ( 0)
t
bj Γ(𝛼 ) 𝛽
exp − 𝛽t
0 0 0
Weibullb) rj −x0j ln(𝜉0 )𝜉0 exp(−𝜉0 )

x0j
sj 𝜉
𝜎0 0
exp(−𝜉0 )
a) x0 = (x01 , x02 , … , x0J , )

b) ξ0 = exp((ln(t) − 𝜇0 )∕𝜎0 ).
lifetime and reliability at mission time under normal operating conditions, can be
readily obtained by delta method; this would require the first-order derivatives of
the corresponding function with respect to the model parameters. The first-order
derivatives of mean lifetime and reliability function for the exponential, gamma,
and Weibull lifetime distributions, with respect to the corresponding model
parameters, are all presented in Table 2.2. Specifically, the asymptotic variance of
the MLE of a parameter of interest, 𝜙, is then
V̂ 𝜙 = P′ V̂ 𝜽 P,
where P = 𝜕𝜙∕𝜕𝜽 is a column vector of the first-order derivatives of 𝜙 with respect

−1 ̂
to the components of 𝜽, and V̂ 𝜽 = Iobs (𝜽) is the inverse of the observed Fisher infor-
mation matrix evaluated at the MLE of model parameter, 𝜽. ̂
√
Subsequently, the standard error se(𝝓)‚ = V̂ can be readily obtained, with
𝜙
which a 100(1 − 𝛿)% asymptotic confidence interval for the parameter 𝜙 can be
obtained as
( )
𝜙̂ − z 𝛿 se(𝝓),
‚ 𝜙̂ + z 𝛿 se(𝝓)
‚ ,
2 2
where z𝛿∕2 is the upper 𝛿∕2 percentage point of the standard normal distribution.
Truncation on the bounds of confidence intervals for reliability and mean lifetime
of devices may be needed, as the reliability has to be between 0 and 1, while the
mean lifetime has to be necessarily positive. But, such a truncation of confidence
intervals may not be accurate.
2.4.2 Approximate Confidence Intervals

Hoyle (1973) discussed various transformations for developing suitable confidence
intervals in case of estimators with skewed distributions and/or when parameters
have bounded ranges. When the sample size is small, the MLEs of mean lifetime
and reliability at mission time may not possess normal distributions, and conse-
quently, the confidence intervals constructed directly by the asymptotic method
may not maintain the nominal level of confidence. For a detailed and compre-
hensive review of all such transformations and their use in developing statistical
intervals, we refer the readers to Meeker et al. (2017).
For variance stabilization, when the asymptotic distribution is over-dispersed,
the hyperbolic transformation (Anscombe, 1948; Kim and Taylor, 1994) could be
used effectively. As the reliability lies between 0 and 1, a hyperbolic arcsecant
(arsech) transformation could be suitable for developing confidence intervals for
reliability. In this arsech-approach, we assume that
( √ )
1 + 1 − ( ̂ 2
R(t))
f̂ = ln (2.31)
̂
R(t)
is normally distributed. Then, by delta method, its standard error can be readily
found to be
̂
se(R(t))
se(f̂ ) = √ ,
̂
R(t) ̂ 2
1 − (R(t))
̂
where se(R(t)) ̂
is the standard error of R(t). Thence, an approximate 100(1 − 𝛿)%
confidence interval for f is given by
( )
f̂ − z 𝛿 se(f̂ ), f̂ + z 𝛿 se(f̂ ) .
2 2
As f̂ ≥ 0, the lower bound may have to be truncated at 0. Since f̂ in (2.31) is a

̂
monotone decreasing function of R(t), we can immediately obtain an approximate
100(1 − 𝛿)% confidence interval for the reliability R(t) to be
( )
2 2
, ,
exp(Uf ) + exp(−Uf ) exp(Lf ) + exp(−Lf )
where Uf = f̂ + z 𝛿 se(f̂ ) and Lf = max (0, f̂ − z 𝛿 se(f̂ )).

2 2
The logit transformation (Viveros and Balakrishnan, 1993) is another useful
transformation for this purpose. In this case, it is assumed that
( )
̂
R(t)
f̂ = ln (2.32)
̂
1 − R(t)
is normally distributed with the corresponding standard error, determined by delta
method, as
̂
se(R(t))
se(f̂ ) = .
̂
R(t)(1 ̂
− R(t))
Then, an approximate 100(1 − 𝛿)% confidence interval for f is
( )
f̂ − z 𝛿 se(f̂ ), f̂ + z 𝛿 se(f̂ ) ,
2 2
using which we can obtain an approximate 100(1 − 𝛿)% confidence interval for the
reliability R(t), based on the inverse of transformation in (2.32), to be
( )
̂
R(t) ̂
R(t)
, ,
̂ + (1 − R(t))S
R(t) ̂ ̂ + (1 − R(t))S
R(t) ̂ −1
where S = exp(z 𝛿 se(f̂ )).

2
In a similar manner, the log-approach can be used to construct confidence inter-
vals for mean lifetime, which would avoid having a negative lower bound for the
mean lifetime. In this approach, we will assume that ln(𝜇)
̂ is asymptotically nor-
mally distributed with the corresponding standard error as
se(𝜇)̂
se(ln(𝜇))
̂ = ,
𝜇̂
where se(𝜇)̂ is the standard error of 𝜇.
̂ This approach yields readily an approximate
100(1 − 𝛿)% confidence interval for mean lifetime 𝜇 as
( { } { })
z 𝛿 se(𝜇)
̂ z 𝛿 se(𝜇)
̂
𝜇̂ exp − 2
, 𝜇̂ exp 2
.
𝜇̂ 𝜇̂
The computation of the required standard errors of the MLEs of mean lifetime and
reliability at mission time have been discussed in the preceding subsection.
2.5 Simulation Studies

In this section, extensive simulation studies are carried out to evaluate the perfor-
mance of the described inferential methods for various sample sizes and different
lifetime distributions, in terms of bias, root mean square errors (RMSEs), cover-
age probabilities, and average widths of 95% confidence intervals, based on 1000
Monte Carlo simulations. Table 2.3 presents the CSALTs, with I = 12 test groups
and Ki = K, for all i, used in these simulation studies.
The EM algorithm was terminated when the convergence was achieved by the
criterion ||𝜽(m) − 𝜽(m+1) || ≤ 1 × 10−5 . It was observed that the number of iterations,
m, was at most 4015. The average numbers of iterations for different sample sizes
and for different distributions were found, and these are presented in Table 2.4.
It was also observed that the MLEs were obtained with less number of iterations
through the EM algorithm when sample size increased. Finally, we should men-
tion that the computational efforts for estimating the model parameters of Weibull
distribution was more than those for exponential and gamma distributions.
Tables 2.5 and 2.6 present the bias and RMSEs of the MLEs of the model param-
eters, mean lifetime, and reliability under normal operating conditions for the
exponential, gamma, and Weibull lifetime distributions. It can be observed that the
Table 2.3 CSALTs with two accelerating factors and I = 12 test

groups.
Test group (i) Stress levels (xi1 , xi2 ) Inspection time (𝝉i )
1 (30, 40) 10
2 (40, 40) 10
3 (30, 50) 10
4 (40, 50) 10
5 (30, 40) 20
6 (40, 40) 20
7 (30, 50) 20
8 (40, 50) 20
9 (30, 40) 30
10 (40, 40) 30
11 (30, 50) 30
12 (40, 50) 30
Table 2.4 Average numbers of iterations for different sample sizes under the
exponential, gamma, and Weibull distributions.
Model K = 10 K = 30 K = 50 K = 100 K = 200
Exponential 27.827 2325.08 2324.216 23.347 —

Gamma — 159.871 142.493 129.187 120.696
Weibull — 409.398 360.299 305.062 278.402
Table 2.5 Bias and RMSEs of MLEs of the model parameter, 𝜽E , mean lifetime, 𝜇E , and
reliability, RE (t), under normal operating conditions x0 = (25, 35), for various sample
sizes, K, under exponential lifetime distribution.
e0 e1 e2 𝝁E RE (10) RE (20) RE (30)

K −6.000 0.030 0.030 66.686 0.799 0.638 0.509
10 Bias 0.003 0.000 −0.001 8.800 −0.013 −0.013 −0.007

RMSE 1.760 0.031 0.030 38.985 0.090 0.134 0.154
30 Bias 0.000 0.000 0.000 2.512 −0.005 −0.006 −0.004
RMSE 0.998 0.016 0.017 19.444 0.050 0.078 0.091
50 Bias −0.012 0.001 0.000 1.806 −0.002 −0.002 −0.001
RMSE 0.769 0.013 0.013 14.503 0.038 0.059 0.070
100 Bias 0.002 0.000 0.000 0.828 −0.001 −0.001 −0.001
RMSE 0.537 0.009 0.009 10.043 0.026 0.042 0.050
bias of the parameters of interest are all small, except for the mean lifetime under
the Weibull distribution in case of small sample sizes. In addition, the RMSEs are
seen to decrease with increasing sample size. Further, the results show that the
estimation of mean lifetime with sample size K less than 100 is not satisfactory for
the Weibull distribution.
Tables 2.7–2.10 present the coverage probabilities and average widths of 95%
confidence intervals for the model parameters, mean lifetime, and reliability under
normal operating conditions for the exponential, gamma, and Weibull lifetime dis-
tributions.
From these tables, we observe that the asymptotic confidence intervals for
the model parameters maintain the nominal level of confidence in all cases, but
the coverage probabilities for mean lifetime and reliability are deflated (below
the nominal level of confidence) for small sample sizes. This is so because the
asymptotic distributions of the MLEs of mean lifetime and reliability become
quite skewed, especially when the sample sizes are small. Figures 2.1 and 2.2
show that the normality assumption for mean lifetime and reliability is violated
Table 2.6 Bias and RMSEs of MLEs of the model parameters, {𝜽G , 𝜽W }, mean lifetime, {𝜇G , 𝜇W }, and reliability, {RG (t), RW (t)}, under normal
operating conditions x0 = (25, 35), for various sample sizes, K, under gamma, and Weibull lifetime distributions.
Gamma distribution
a0 a1 a2 b0 b1 b2 𝝁G RG (10) RG (20) RG (30)

K −0.500 0.020 0.020 7.000 −0.060 −0.060 60.340 0.912 0.739 0.562
30 Bias −0.101 0.002 0.001 0.187 −0.003 −0.002 4.049 −0.014 −0.010 −0.015
RMSE 2.004 0.040 0.039 2.598 0.050 0.049 18.296 0.064 0.075 0.093
50 Bias 0.033 0.000 0.000 −0.002 0.000 0.000 1.875 −0.006 −0.003 −0.009
RMSE 1.660 0.031 0.030 2.104 0.038 0.037 12.634 0.052 0.061 0.071
100 Bias −0.021 0.000 0.000 0.061 0.000 −0.001 1.301 −0.003 −0.001 −0.002
RMSE 1.141 0.021 0.022 1.446 0.026 0.026 8.563 0.036 0.043 0.048
200 Bias 0.000 0.000 0.001 0.010 0.001 −0.001 0.392 −0.001 −0.001 −0.002
RMSE 0.851 0.015 0.016 1.068 0.018 0.019 5.757 0.027 0.031 0.034
Weibull distribution
r0 r1 r2 s0 s1 s2 𝝁W RW (10) RW (20) RW (30)
K −1.000 0.020 0.020 6.000 −0.030 −0.030 62.451 0.851 0.686 0.539
30 Bias −0.094 0.001 0.001 0.196 −0.002 −0.002 179.699 −0.009 −0.009 −0.022
RMSE 1.746 0.031 0.032 1.386 0.020 0.024 2174.127 0.075 0.078 0.114
50 Bias −0.016 0.001 0.000 0.125 −0.002 −0.001 18.066 −0.003 0.000 −0.009
RMSE 1.323 0.024 0.024 0.947 0.016 0.016 84.059 0.061 0.058 0.086
100 Bias −0.038 0.001 0.000 0.052 −0.001 0.000 5.611 −0.003 −0.002 −0.005
RMSE 0.883 0.016 0.016 0.615 0.010 0.010 21.300 0.042 0.039 0.052
200 Bias 0.003 0.000 0.000 0.014 0.000 0.000 2.172 −0.001 −0.001 −0.003
RMSE 0.608 0.012 0.011 0.428 0.007 0.007 12.136 0.030 0.028 0.036
Table 2.7 Coverage probabilities and average widths of 95% confidence intervals for the
model parameter, 𝜽E , mean lifetime, 𝜇E , and reliability, RE (t), under normal operating
conditions x0 = (25, 35), for various sample sizes, K, under exponential lifetime
distribution.
e0 e1 e2 𝝁E RE (10) RE (20) RE (30)

K −6.000 0.030 0.030 66.686 0.799 0.638 0.509
Coverage probability
10 asya) 0.951 0.947 0.951 0.899 0.917 0.905 0.890

arsechb) — — — — 0.945 0.941 0.937
log/logitc) — — — 0.943 0.958 0.966 0.969
30 asya) 0.943 0.955 0.941 0.942 0.937 0.931 0.921
arsechb) — — — — 0.941 0.946 0.951
log/logitc) — — — 0.936 0.939 0.940 0.945
50 asya) 0.944 0.964 0.943 0.945 0.934 0.933 0.937
arsechb) — — — — 0.939 0.938 0.937
log/logitc) — — — 0.951 0.953 0.951 0.948
100 asya) 0.945 0.947 0.953 0.947 0.937 0.935 0.935
arsechb) — — — — 0.943 0.941 0.943
log/logitc) — — — 0.940 0.939 0.938 0.940
Average width
10 asya) 6.743 0.116 0.116 140.976 0.330 0.501 0.574

arsechb) — — — — 0.312 0.474 0.563
log/logitc) — — — 163.664 0.326 0.466 0.527
30 asya) 3.858 0.066 0.066 72.804 0.188 0.296 0.350
arsechb) — — — — 0.185 0.289 0.345
log/logitc) — — — 76.227 0.188 0.288 0.336
50 asya) 2.988 0.051 0.051 55.579 0.145 0.229 0.273
arsechb) — — — — 0.143 0.226 0.271
log/logitc) — — — 57.121 0.145 0.226 0.267
100 asya) 2.112 0.036 0.036 38.605 0.102 0.163 0.194
arsechb) — — — — 0.102 0.162 0.194
log/logitc) — — — 39.134 0.103 0.162 0.192
a) asy: asymptotic confidence interval.

b) arsech: approximate confidence interval based on arsech transformation.
c) log/logit: approximate confidence interval based on log transformation for mean lifetime
and logit transformation for reliability.
Table 2.8 Coverage probabilities and average widths of 95% asymptotic confidence
intervals for the model parameters, {𝜽G , 𝜽W }, for various sample sizes, K, under gamma
and Weibull lifetime distributions.
Gamma distribution
a0 a1 a2 b0 b1 b2
−0.500 0.020 0.020 7.000 −0.060 −0.060
K = 30 0.966 0.953 0.958 0.969 0.956 0.961

K = 50 0.945 0.954 0.958 0.946 0.957 0.956
K = 100 0.958 0.956 0.941 0.959 0.957 0.941
K = 200 0.940 0.944 0.946 0.938 0.947 0.950
Average width
K = 30 8.483 0.155 0.154 10.827 0.190 0.190

K = 50 6.534 0.119 0.119 8.254 0.146 0.146
K = 100 4.585 0.084 0.084 5.783 0.102 0.103
K = 200 3.235 0.059 0.059 4.068 0.072 0.072
Weibull distribution
r0 r1 r2 s0 s1 s2
−1.000 0.020 0.020 6.000 −0.030 −0.030
K = 30 0.947 0.954 0.958 0.946 0.954 0.938

K = 50 0.938 0.947 0.938 0.944 0.952 0.942
K = 100 0.952 0.949 0.951 0.948 0.956 0.947
K = 200 0.949 0.948 0.954 0.945 0.957 0.955
Average width
K = 30 6.414 0.116 0.116 4.884 0.076 0.078

K = 50 4.878 0.089 0.089 3.522 0.057 0.056
K = 100 3.390 0.063 0.063 2.358 0.039 0.039
K = 200 2.383 0.044 0.044 1.632 0.027 0.027
Table 2.9 Coverage probabilities and average widths of 95% confidence intervals for
mean lifetime, 𝜇G , and reliability, RG (t), under normal operating conditions x0 = (25, 35),
for various sample sizes, K, under gamma lifetime distribution.
𝝁G RG (10) RG (20) RG (30)

60.340 0.912 0.739 0.562
K = 30 asya) 0.935 0.923 0.947 0.958

arsechb) — 0.959 0.960 0.974
log/logitc) 0.947 0.960 0.959 0.975
K = 50 asya) 0.942 0.906 0.943 0.950
arsechb) — 0.933 0.954 0.958
log/logitc) 0.946 0.958 0.951 0.964
K = 100 asya) 0.946 0.935 0.953 0.955
arsechb) — 0.955 0.956 0.956
log/logitc) 0.948 0.962 0.954 0.964
K = 200 asya) 0.945 0.922 0.942 0.953
arsechb) — 0.938 0.946 0.954
log/logitc) 0.946 0.947 0.948 0.952
Average width
K = 30 asya) 68.361 0.232 0.313 0.363

arsechb) — 0.245 0.302 0.357
log/logitc) 73.036 0.324 0.310 0.347
K = 50 asya) 47.479 0.187 0.241 0.279
arsechb) — 0.191 0.236 0.276
log/logitc) 48.775 0.230 0.239 0.272
K = 100 asya) 32.311 0.143 0.170 0.190
arsechb) — 0.141 0.168 0.189
log/logitc) 32.700 0.154 0.169 0.188
K = 200 asya) 22.015 0.102 0.120 0.133
arsechb) — 0.101 0.120 0.133
log/logitc) 22.138 0.106 0.120 0.132
a) asy: asymptotic confidence interval.

b) arsech: approximate confidence interval based on arsech transformation.
c) log/logit: approximate confidence interval based on log transformation for mean lifetime
and logit transformation for reliability.
Another random document with
no related content on Scribd:
women; and I was glad in my heart, when, taking ten sheep and a fat
bullock, we left these poor creatures to their fate, as, had more
Arabs arrived, they would most certainly have stripped them of every
thing. We halted, after dark, at a place called Mull.
Feb. 2.—Our road, as yesterday, was an extensive valley,
bounded to the right and left by low hills; about noon we descended
slightly, and found ourselves in a productive plain of great extent,
thickly planted with trees and underwood, not unlike a preserve in
England. About an hour before sun-set, we came to what had the
appearance of the bed of a lake, and here was the wished-for well of
water. The horses had not drank since noon on the 31st, and
although ready to drop on the road from faintness, were, on reaching
the well, quite unmanageable. The name of the well was Kofei.
On the 31st, Boo-Khaloom had thought it right to send on a
Tibboo with the news of our approach to the sheikh El-Kanemy, who,
we understood, resided at Kouka, and one was despatched with a
camel and a man of Mina Tahr: the Gundowy accompanied him on
the arrival at Kofei of the Arabs, who preceded us for the purpose of
clearing the well. The Tibboo who had been despatched was found
alone and naked; some Tibboo Arabs, of a tribe called Wandela, had
met them near the well on the preceding evening and robbing him
even to his cap, and taking from him the letters, saying, they cared
not for the sheikh or Boo-Khaloom, tied him to a tree, and then left
him. In this state was he found by our people; and Mr. Clapperton
coming up soon after, gave him, from his biscuit-bag, wherewithal to
break his fast, after being twenty-four hours without eating. Eighteen
men had stripped him, he said, and taken off the camel and Mina
Tahr’s man, who, they also said, should be ransomed, or have his
throat cut. Mina Tahr represented these people as the worst on the
road in every sense of the word: “They have no flocks,” said he, “and
have not more than three hundred camels, although their numbers
are one thousand or more; they live by plunder, and have no
connexion with any other people. No considerable body of men can
follow them; their tents are in the heart of the desert, and there are
no wells for four days in the line of their retreat. Giddy-ben-Agah is
their chief, and I alone would give fifty camels for his head: these are
the people who often attack and murder travellers, and small kafilas,
and the Gundowy, who respect strangers, have the credit of it.”
The men of Traita, with their chief, Eskou-ben-Coglu, came in the
evening to welcome us: the well Kofei belongs to them; and greatly
enraged they appeared to be at the conduct of the Wandelas. This
chief returned to Boo-Khaloom his letters, which, he said, “the chief
of the Wandelas had sent him that morning, begging that he would
meet the kafila at the well, and deliver them to Boo-Khaloom: had he
known then what had taken place, the slave,” he said, “should have
been stabbed at his father’s grave before he would have delivered
them.” Boo-Khaloom was greatly enraged; and I was almost
apprehensive that he would have revenged himself on the Traita
chiefs. However, the Tibboo courier was again clothed and mounted,
and once more started for Bornou. The Traita Tibboos are more
important-looking fellows than the Gunda, but they want their
quickness and activity: they are said not to be more than eight
hundred strong in males.
Feb. 3.—Our course, during the early part of the day, was due
south, and through a country more thickly planted by the all-tasteful
hand of bounteous Nature. We disturbed a flock of what we at first
thought were deer, but they were only a large species of antelope;
they are of a deeper fawn colour, and have black and white stripes
under the belly. The Guinea fowl were in great numbers, but
extremely shy. The whole day our route lay through most pleasing
forest scenery. It was near sunset when we arrived at Mittimee,
which, in the Bornou language, means warm, tepid: the wells exceed
fifty in number, and lie in a woody hollow, where there are clumps of
the tulloh and other species of the mimosa tribe, encircled by
kossom and various parasitical and twining shrubs, which,
embracing their stems, wind to the extremities of their branches, and
climb to the very tops, when, falling over, they form weeping bowers
of a most beautiful kind: it was indeed a lovely and a fair retreat.
Boo-Khaloom, myself, and about six Arabs, had ridden on in front:
it was said we had lost the track, and should miss the well: the day
had been oppressively hot, my companions were sick and fatigued,
and we dreaded the want of water. A fine dust, arising from a light
clayey and sandy soil, had also increased our sufferings: the
exclamations of the Arab who first discovered the wells were indeed
music to our ears; and after satisfying my own thirst, with that of my
weary animals, I laid me down by one of the distant wells, far from
my companions; and these moments of tranquillity, the freshness of
the air, with the melody of the hundred songsters that were perched
amongst the creeping plants, whose flowers threw an aromatic odour
all around, were a relief scarcely to be described. Ere long, however,
the noisy kafila, and the clouds of dust which accompanied it,
disturbed me from the delightful reverie into which I had fallen.
Feb. 4.—Previously to arriving at Lari, we came upon two
encampments of the Traita Tibboos, calling themselves the sheikh’s
people: their huts were not numerous, but very regularly built in a
square, with a space left in the north and south faces of the
quadrangle, for the use of the cattle. The huts were entirely of mats,
which, excluding the sun, yet admitted both the light and the air:
these habitations, for fine weather, are preferable to the bete shars,
or tents, of the Arabs of the north. The interior was singularly neat:
clean wooden bowls, with each a cover of basket-work, for holding
their milk, were hung against the wall. In the centre of the inclosure
were about one hundred and fifty head of cattle feeding from cradles:
these were chiefly milch cows, with calves and sheep. The Tibboos
received us kindly at first, but presumed rather too much on sheikh
Kanemy’s protection, which they claim or throw off, it is said, as it
suits their purpose. The modest request of a man, with two hundred
armed Arabs, for a little milk, was refused; and ready as the Arabs
are to throw down the gauntlet, a slight expression of displeasure
from their leader was followed by such a rapid attack on the Tibboos,
that before I could mount, half the stock was driven off, and the
sheikh well bastinadoed. Boo-Khaloom was, however, too kind to
injure them; and after driving their cattle for about a mile, he allowed
them to return, with a caution to be more accommodating for the
future. Accustomed as these people are to plunder one another, they
expect no better usage from any one who visits them, provided they
are strong enough, and vice versa; they are perfect Spartans in the
art of thieving, both male and female.
An old woman, who was sitting at the door of one of the huts, sent
a very pretty girl to me, as I was standing by my horse, whose massy
amber necklace, greased head, and coral nose studs and ear-rings,
announced a person of no common order, to see what she could
pick up; and after gaining possession of my handkerchief and some
needles, while I turned my head, in an instant thrust her hand into
the pocket of my saddle-cloth, as she said, “to find some beads, for
she knew I had plenty.”
Another and much larger nest of the Traitas lay to the east of our
route, a little further on, with numerous flocks and herds. About two
in the afternoon we arrived at Lari, ten miles distant from Mittimee.
On ascending the rising ground on which the town stands, the
distressing sight presented itself of all the female, and most of the
male inhabitants, with their families, flying across the plain in all
directions, alarmed at the strength of our kafila. Beyond, however,
was an object full of interest to us, and the sight of which conveyed
to my mind a sensation so gratifying and inspiring, that it would be
difficult in language to convey an idea of its force or pleasure. The
great lake Tchad, glowing with the golden rays of the sun in its
strength, appeared to be within a mile of the spot on which we stood.
My heart bounded within me at this prospect, for I believed this lake
to be the key to the great object of our search, and I could not refrain
from silently imploring Heaven’s continued protection, which had
enabled us to proceed so far in health and strength, even to the
accomplishment of our task.
It was long before Boo-Khaloom’s best endeavours could restore
confidence: the inhabitants had been plundered by the Tuaricks only
the year before, and four hundred of their people butchered; and but
a few days before, a party of the same nation had again pillaged
them, though partially. When, at length, these people were satisfied
that no harm was intended them, the women came in numbers with
baskets of gussub, gafooly, fowls, and honey, which were purchased
by small pieces of coral and amber of the coarsest kind, and
coloured beads. One merchant bought a fine lamb for two bits of
amber, worth, I should think, about twopence each in Europe; two
needles purchased a fowl; and a handful of salt four or five good
sized fish from the lake.
Lari is inhabited by the people of Kanem, who are known by the
name of Kanemboo: the women are good-looking, laughing
negresses, and all but naked; but this we were now used to, and it
excited no emotions of surprise. Most of them had a square or
triangular piece of silver or tin hanging at the back of the head,
suspended from the hair, which was brought down, in narrow plaits,
quite round the neck.
Feb. 5.—By sun-rise I was on the borders of the lake, armed for
the destruction of the multitude of birds, who, all unconscious of my
purpose, seemed as it were to welcome our arrival. Flocks of geese
and wild ducks, of a most beautiful plumage, were quietly feeding at
within half pistol shot of where I stood; and not being a very keen or
inhuman sportsman, for the terms appear to me to be synonymous,
my purpose of deadly warfare was almost shaken. As I moved
towards them they only changed their places a little to the right or
left, and appeared to have no idea of the hostility of my intentions. All
this was really so new, that I hesitated to abuse the confidence with
which they regarded me, and very quietly sat down to contemplate
the scene before me. Pelicans, cranes, four and five feet in height,
grey, variegated, and white, were scarcely so many yards from my
side, and a bird, between a snipe and a woodcock, resembling both,
and larger than either; immense spoonbills of a snowy whiteness,
widgeon, teal, yellow-legged plover, and a hundred species of (to me
at least) unknown water fowl, were sporting before me; and it was
long before I could disturb the tranquillity of the dwellers on these
waters by firing a gun.
From a Sketch by Major
Engraved by E. Finden.
Denham’s.
KANEMBOO MARKET UNMARRIED WOMAN
WOMAN. OF SOUDAN.
Published Feb. 1826, by John Murray, London.
The soil near the edges of the lake was a firm dark mud; and, in
proof of the great overflowings and recedings of the waters, even in
this advanced dry season, the stalks of the gussub, of the preceding
year, were standing in the lake, more than forty yards from the shore.
The water is sweet and pleasant, and abounds with fish; which the
natives have a curious way of catching. Some thirty or forty women
go into the lake, with their wrappers brought up between their legs,
and tied round their middles, as I should say, by single files, and
forming a line at some distance in the water, fronting the land, for it is
very shallow near the edges, and absolutely charge the fish before
them so close, that they are caught by the hand, or leap upon the
shore. We purchased some, and the best flavoured was a sort of
bream.
A circumstance happened whilst I was on the margin of the lake,
which was a further proof that the little kindnesses I had shown the
Arabs were not lost upon them; and which supported my favourite
position, that no people on earth are so savage, but that gentle kind
treatment, with a frank and liberal manner, will gain their confidence
and regard. A lamb, the most harmless thing that breathes, alarms a
child who for the first time sees such an animal. I had suffered my
horse to go loose, in order to approach close to the flights of birds
around me, and he probably thinking the tents might afford him
better fare than where I left him, first rubbed off his bridle, and then
quietly returned to the encampment. About the same time one of the
freed women found my bornouse, which had fallen from the saddle,
and brought it to Boo Khaloom. All this created an alarm, and it was
then found out, that two boats or canoes had been seen coming from
the south-east, in which direction are islands inhabited by the
Biddoomah, a people who live by plundering on the main land, and
carry off any thing they can pick up. This was quite enough to make
Boo Khaloom think I was already gone, or in great danger; and not
only several Arab chiefs armed themselves, and mounted, to seek
me, but some of the merchants also. They found me, after a long
search, on the lake among the gussub stalks, loaded with more birds
than I could carry, and would scarcely believe that I had seen neither
enemies nor boats. The dread which the natives appear to have of
these koorie, or islanders, is almost equal to their fear of the
Tuaricks; but the former are less rapacious and bloody in their visits.
Their habitations are three or four days distant to the southward of
east, towards the centre of the lake.
In the evening I visited the town of Lari: it stands on an eminence,
and may probably contain two thousand inhabitants. The huts are
built of the rush which grows by the sides of the lake, have conical
tops, and look very like well thatched stacks of corn in England. They
have neat inclosures round them, made with fences of the same
reed, and passages leading to them like labyrinths. In the inclosure
is a goat or two, poultry, and sometimes a cow. The women were
almost all spinning cotton, which grows well, though not abundantly,
near the town and lake. The interior of the huts is neat: they are
completely circular, with no admission for air or light, except at the
door, which has a mat, by way of safeguard. I entered one of the
best appearance, although the owner gave me no smiles of
encouragement, and followed close at my heels, with his spear and
dagger in his hand. In one corner stood the bed, a sofa of rushes
lashed together, and supported by six poles, fixed strongly in the
ground. This was covered by the skins of the tiger-cat and wild bull;
round the sides were hung the wooden bowls, used for water and
milk: his tall shield rested against the wall. The hut had a division of
mat-work, one half being allotted to the female part of the family. My
host, however, continued to look at me with so much suspicion, and
seemed so little pleased with my visit, notwithstanding all my
endeavours to assure him I was a friend, that I hurried from the
inhospitable door, and resumed my walk through the town.
Feb. 6.—A gratifying scene took place this morning, in the
departure of nearly thirty freed slaves, natives of Kanem, who here
left us for their homes, three days’ journey to the eastward. I had
been applied to, the night before, to intercede with Boo Khaloom for
this indulgence; for as he had heard that the sheikh was at war with
some of the chiefs of Kanem, he had determined on first taking them
to Bornou, for fear of their being plundered on the road of the little
they had saved in slavery. These poor creatures had, however,
found one or two of their countrymen at the market of Lari, who
assured them of their safety on the road between that place and their
homes. The good man complied with evident reluctance on their own
account, and they took leave, kissing his hand, with tears and
blessings. They had most of them been in the service of the bashaw,
some for a term of years, and were returning to die at home at last.
One poor deaf and dumb woman, whom the rapacity of Mukni, the
former sultan of Fezzan, who spared neither age, sex, nor infirmity,
had induced him to march to Tripoli, had shed torrents of tears ever
since she had been made acquainted, by signs, that she was to go
to Bornou. She had left two children behind her; and the third, which
was in her arms when she was taken by the Arabs, had been torn
from her breast after the first ten days of her journey across the
desert, in order that she might keep up with the camels. Her
expressive motions in describing the manner in which the child was
forced from her, and thrown on the sand, where it was left to perish,
while whips were applied to her, lame and worn out as she was, to
quicken her tottering steps, were highly eloquent and interesting.
They had all been my friends for more than five months, and to some
I had rendered little services by carrying their bag of zumeeta, or
salt. They were not ungrateful, and our parting had something in it
affecting, which, considering negroes in the degraded light they do,
seemed greatly to astonish the Arabs.
On quitting Lari, we immediately plunged into a thickly-planted
forest of acacias, with high underwood; and at the distance of only a
few hundred yards from the town, we came upon large heaps of the
elephants’ dung, forming hillocks three and four feet in height, and
marks of their footsteps: the tracks of these animals increased as we
proceeded. Part of the day our road lay along the banks of the
Tchad, and the elephants’ footmarks, of an immense size, and only a
few hours old, were in abundance. Whole trees were broken down,
where they had fed: and where they had reposed their ponderous
bodies, young trees, shrubs, and underwood, had been crushed
beneath their weight. We also killed this day an enormous snake, a
species of coluber; it was a most disgusting, horrible animal, but not
however venomous. It measured eighteen feet from the mouth to the
tail; it was shot by five balls, and was still moving off, when two
Arabs, with each a sword, nearly severed the head from the body.
On opening the belly, several pounds of fat were found, and carefully
taken off by the two native guides who accompanied us. This they
pronounced a sovereign remedy for sick and diseased cattle, and
much prized amongst them. Scarcely a mile further, a drove of wild
red cattle, which I at first took for deer, were seen bounding to the
west. I had no gun, but got extremely close to them, and found they
were what the Arabs call “bugra-hammar-wahash” (red cow wild).
They appeared to partake of the bullock and buffalo, with a tuft or
lump on the shoulder.
We bivouacked near a small parcel of huts, called Nyagami, in a
beautiful spot, so thick of wood, that we could scarcely find a clear
place for our encampment. While the tents were fixing, an alarm was
given of wild boars: one of our party followed the scent, and, on his
return, said he had seen a lion, and near him seven gazelles. I could
not, however, find from the natives, that lions were ever seen here:
numerous other animals appeared to abound, and that confirmed the
opinion.
Feb. 7.—We moved for Woodie about eight, accompanied by two
Arabs of Boo-Saif. I left the kafila, and proceeded a little to the
westward, making a parallel movement with the camels. Birds of the
most beautiful plumage were perched on every tree. Guinea fowls
were in flocks of eighty or one hundred; and several monkeys
chattered at us so impudently, that, separating one from the rest, we
chased him for nearly half an hour: he did not run very fast, or
straight forward, but was constantly doubling and turning, with his
head over his shoulder, to see who was close to him. He was a
handsome fellow, of a light brown colour, and black about the
muzzle. About noon we came on a village of huts, called Barrah; and
although only three in number, the natives flew in all directions. On
our approaching the town, we beckoned to them, and got off our
horses, for the purpose of giving them confidence, and sat down
under the shade of a large tamarind tree. An old negro, who spoke a
little Arabic, was the first who ventured to approach: seeing that he
was not ill-treated, the others soon followed his example. I begged a
little leban (sour milk), a most refreshing beverage after a hot ride,
but none was to be found, until they were assured that I should pay
for it; and at the sight of the dollar they all jumped and skipped like
so many monkeys. Some biscuit, which I carried in my saddle-cloth
pocket, and now began to eat, created much astonishment, and the
first to whom I gave some, refused to eat it. One, rather bolder than
the rest, put a small piece into his mouth, and pronounced it good,
with such extravagant gestures, that my visitors all became so
clamorous, that my stock was speedily demolished. I refused for a
long time the man who had been suspicious at first, to the great
amusement of the rest, who seemed to like the joke amazingly.
I had promised the Arabs to share with them a sheep, provided
they did not help themselves, and now made signs of my wish to
purchase one. Two men went off to bring, as they said, a fat one.
After a short time had elapsed, during which they had been delighted
with the opening and shutting of my pocket-knife, a very miserable
sheep was brought to me, which they seriously endeavoured to
make me understand was a very fine one. The Arabs declared it to
be good for nothing; and, therefore, though unwilling to be
displeased, I quickly returned my dollar to my pocket, and made a
motion towards my horse. The whole tribe, to my great
astonishment, shouted out, and began to push about the vender of
the sheep, and dance round me. Another very fine fat sheep was
now brought forward from behind the crowd: offering the other first
seemed a trick, in order to try whether I should find out the lean from
the fat one; and although much sagacity was not required for this, it
appeared to have raised me very much in their estimation.
The little nest of thatched huts in which they lived was most
beautifully situated on a rising spot, in the midst of a rich and
luxuriant, though not thick forest, about three miles to the north-east
of Woodie; and the wells, which stand in a dell, thickly planted with
palms (the first we had seen on this side of the desert), had troughs
for more than a hundred and fifty cattle to drink at. One of the old
men accompanied us, while his son carried the carcass of the sheep
to Woodie, for which service he was rewarded by two coral beads,
and a little snuff.
Close to the town we found the tents. Our party had made about
fourteen miles, without leaving the banks of the lake at any great
distance. Two elephants were seen swimming in the lake this day;
and one, belonging to a drove at a distance, absolutely remained just
before the kafila. Hillman had gone on in front on his mule, suffering
sadly from weakness and fatigue, and had laid himself down in what
appeared a delightful shade, to await the arrival of the camels, not
expecting to see an elephant. He was absolutely reposing within a
dozen yards of a very large one, without being aware of it; and on an
Arab’s striking the animal with a spear, he roared out and moved off.
Poor Hillman’s alarm was extreme.
Feb. 8.—On walking to the shores of the lake this morning, soon
after sunrise, I was surprised to see how the water had encroached
since the day before. More than two miles of the wood was entirely
overflowed—the cotton plantations were covered with water. Were
the lands cleared of wood, which would not be a laborious task, as
the trees are mostly tulloh, and not large, almost any thing might be
produced.
Feb. 9.—The courier had been sent off a second time, after being
re-clothed and re-mounted, to receive the sheikh’s orders, and we
were not to proceed beyond Woodie until his pleasure was known.
So jealous and so suspicious are these negro princes of the
encroachment of the Arabs, that divers were the speculations as to
whether the sheikh would, or would not, allow them to proceed with
us nearer his capital.
A weekly fsug, or market, was held about a mile from the town;
and the women flocking from the neighbouring negro villages,
mounted on bullocks, who have a thong of hide passed through the
cartilage of the nose when young, and are managed with great ease,
had a curious appearance: a skin is spread on the animal’s back,
upon which, after hanging the different articles they take for sale,
they mount themselves; milk, sour and sweet, a little honey, fowls,
gussub, gafooly, are amongst their wares, fat and meloheea (ochra),
a green herb, which, with bazeen, all negroes eat voraciously, and
indeed Christians too, as I afterwards found out. The men brought
oxen, sheep, goats, and slaves: the latter were few in number, and in
miserable condition.
Woodie is a capital, or as they say, Blad Kebir, and is governed by
a sheikh, who is a eunuch, and a man of considerable importance:
they appear to have all the necessaries of life in abundance, and are
the most indolent people I ever met with. The women spin a little
cotton, and weave it into a coarse cloth of about six inches’ width;
the men either lie idling in their huts all day, or in the shade of a
building, formed by four supporters and a thatched roof, which
stands in an open space amongst the huts: this is also the court of
justice and place of prayers. The men are considerably above the
common stature, and of an athletic make; but have an expression of
features particularly dull and heavy. The town stands about one mile
west of the Tchad, four short days’ march from Bornou. Game of all
descriptions comes to within a stone’s-throw of their doors, and the
lake abounds with fish and water-fowl; yet have they so little
exertion, that a few fish was almost the only produce of their labour
which was offered for sale.
The women, like the Tibboo, have a square piece of blue or white
cloth, tied over one shoulder, which forms their whole covering: their
hair is however curiously and laboriously trained, and I observed that
no one of tender years had any thing like a perfect head of hair.
From childhood the head is shaved, leaving only the top covered; the
hair from hence falls down quite round from the forehead to the pole
of the neck, and is then formed into one solid plait, which in front
lying quite flat just over the eyes, and behind being turned up with a
little curl, has just the appearance of an old-fashioned coachman’s
wig in England: some of them are, however, very pretty.
Feb. 10.—I this morning went to the eastward, in order to see the
extent of the forest, and also, if possible, to get a sight of the herd, of
upwards of one hundred and fifty elephants, which some of the
Arabs had seen the day before while their camels were feeding. I
was not disappointed. I found them about six miles from the town, in
the grounds annually overflowed by the waters of the lake, where the
coarse grass is twice the height of a man: they seemed to cover the
face of the country, and, I should think, exceeded the number I had
expected to see. When the waters flow over these their pasturages,
they are forced by hunger to approach the towns, and spread
devastation throughout their march; whole plantations, the hopes of
the inhabitants for the next year, are sometimes destroyed in a single
night. Nothing, however, more ferocious than large antelopes, with a
fox and wild hog or two, was to be seen, besides elephants, although
I beat every thicket. We had followed about half a dozen of these
antelopes for more than three hours, who merely changed their
place without ever getting out of sight, but never allowed us to get
near enough to hazard a shot. When quite fatigued, I determined on
making for some distant huts, and begging a little milk, sweet or
sour. No knowing landlady of a country inn ever scanned the
character of her customer more than did this untaught, though
cunning negro, whom we found there. He first denied that he had
any, notwithstanding the bowls were full scarcely ten paces behind
him; and then asked, what I had got to pay for it? I had really
nothing; and after offering my pocket-handkerchief, which was
returned to me as not worth any thing, I was about to depart, though
ten long miles from the tents, thirsty as I was, when the Arab pointed
to a needle, which was sticking in my jacket: for this and a white
bead which the Arab produced, we had a bowl of fine milk and a
basket of nuts, which refreshed us much; and we returned home by
the lake, where I shot two birds—one a very fine crane, and the
other of the woodpecker species, and saw a flock of at least five
hundred pelicans, but could not get near enough to fire at them.
The whole surface of the country for the last eighteen days had
been covered with a grass which produced a calyx so full of prickles
as to annoy us almost to misery: these prickles were of the finest
and most penetrating sharpness that can be imagined; they attached
to every part of our dress; and so small were the points, that it was
impossible to extract them without their breaking and leaving a part
behind: if we walked, at every step we were obliged to clear them
from our feet—mats, blankets, trowsers, were filled with these
irritating annoyances, so that there was no getting rid of them, by
day or night; in short, no part of the body was free from them. The
seed from this grass is called kashcia, and is eaten[21].
Feb. 11.—Two of the sheikh’s officers arrived last night, with
letters, and a present of goroo nuts of Soudan: they have a pleasant
bitter taste, and are much esteemed by all the Tripoli people. After
eating these nuts, water has a grateful flavour, be it ever so bad: the
Arabs call them the coffee of the black country. These letters
pressed Boo Khaloom to continue his march towards Kouka, with all
his people—a very great proof of his confidence in the peaceable
disposition of our chief. The men were clothed with a bornouse each,
a turban, and a red cap; and after giving us fifteen bullocks, six
sheep, and seventeen kail of gussub, returned home, promising that
fresh supplies should be prepared for his people at Yeou, two days’
march nearer Bornou. It was nearly dark when we reached a town
called Burwha. We had travelled at a considerable distance from the
lake after the first four miles of our journey, which here sweeps off
greatly to the east.
Burwha is a walled town, and the first negro one we had seen: it
may be called in this country a place of some strength; in proof of
which the inhabitants have always defied the Tuarick marauders,
who never entered the town: the walls may be about thirteen or
fourteen feet high, and have a dry ditch, which runs quite round
them. The town probably covers an extent equal to three square
miles, and contains five or six thousand inhabitants. There is a
covered-way, from which the defenders lance their spears at the
besiegers, and instantly conceal themselves: there are but two
gates, which are nearly east and west; and these being the most
vulnerable parts for an enemy to attack, are defended by mounds of
earth thrown up on each side, and carried out at least twenty yards
in front of the gate, and have nearly perpendicular faces. These
advanced posts are always thickly manned, and they conceive them
to be a great defence to their walls: they cannot, however, calculate
upon their being abandoned, as an enemy once in possession of
them would so completely command the town, that from thence
every part of it may be seen. Nevertheless, Burwha is a strong place,
considering the means of attack which the Arabs have; and we were
much struck with its appearance.
Feb. 12.—I rode through the town early this morning, previous to
our move. All the principal huts had their little inclosure, with a cow
or two, some goats and fowls; and I saw a very fine fish, apparently
roasted, or broiled, carried into one of them, on which I could have
breakfasted with great pleasure. Gussub, in large baskets, and in the
straw, was every where to be seen, and the women were spinning at
the doors of most of the huts.
I rode nearly the whole of this day with Min Ali Tahar, the
Gundowy Tibboo sheikh, who was accompanying us to Bornou: he
had some little difference with the sheikh, of whom he was perfectly
independent, and Boo Khaloom, ever politic, undertook to make up
the misunderstanding; thereby not only showing his influence, but
securing, in a manner, the future friendship of Tahar, whose district
was always considered as the most dangerous part of the Tibboo
country on the road to Mourzuk. Tahar was a sharp, intelligent fellow,
spoke a little Arabic, and had often asked me many questions about
my country, and my sultan; but to-day he was more inquisitive than
usual.—“Rais Khaleel,” said he, “what would your sultan do to Min
Ali, if he was to go to England? Would he kill me, or keep me there a
prisoner? I should like to be there for about a month.” I answered,
“Certainly neither one nor the other: he would be much more inclined
to make you a handsome present, and send you back again.” Min
Ali. “Oh! I should take him something; but what could I give him?
nothing but the skins of a dozen ostriches, some elephants’ teeth,
and a lion’s skin.” Ans. “The value of the present could be of no
importance to my sultan; he would look at the intention: befriend his
people—remember the Inglezi that you have seen; and should any
more ever find their way to your tents, give them milk and sheep,
and put them in the road they are going: promise me to do this, and I
can almost promise you, that my sultan shall send you a sword, such
a one as Hateeta had on my return, without your going to England,
or giving him any thing.” Min Ali. “Is he such a man? Barak Allah!
What’s his name?” “George.” “George! Health to George; much of it!
Salam Ali: George Yassur. Tell him, Min Ali Tahar wishes him all
health and happiness; that he is a Tibboo who can command a
thousand spears, and fears no man. Is he liberal? Is his heart large?
Gulba Kebir. Does he give presents to his people?” “Very much so,
indeed; some of his people think him too generous.” Min Ali. “By the
head of my father! Raas el Booe! they are wrong; the sultan of a
great people should have a large heart, or he is unworthy of them.
Who will succeed him when he dies?” “His brother.” “What is his
name?” “Frederick.” “Barak Allah! I hope he will be like George,
matlook, liberal. Salem Ali! Frederick. Health to Frederick! How many
wives have they?” “No Englishman has more than one,” said I. “A
gieb! a gieb! Wonderful! wonderful! why they should have a
hundred.” “No! no! we think that a sin,” replied I. “Wallah! really
(literally, by G—!) why I have four now; and I have had more than
sixty. Her I like best, however, always says, one would be more
lawful: she may be right. You say she is. You are a great people. I
see you are a great people, and know every thing. I, a Tibboo, am
little better than a gazelle.”
The road to-day was thickly scattered with trees—saw flocks of
red cattle, and killed a wild hog. The hyænas came so close to the
tents last night, that a camel, which lay about a hundred yards from
the enclosure, was found nearly half-eaten. A lion first made a meal
on the poor animal; when the hyænas came down upon what he had
left. We had fires the whole night; and notwithstanding the continued
howlings which these animals kept up until daylight, our rest was but
little disturbed.—Halted near a water, called Chugelarem. We had
now about eleven miles to make, previous to arriving at Yeou.
13th.—Chugelarem, though said to be a branch of the Tchad, was
merely a still water, increased considerably by the overflowings of
that lake in the wet season: the bottom was muddy, and nowhere
deeper than two feet. The camels, horses, and followers of the kafila,
waded through it without being much above their knees: it takes a
zigzag direction, going first to the east, then to the north, and then to
the east again.
We proceeded south, passing several very neat negro villages;
and after about eleven miles, came to a very considerable stream,
called the Yeou, in some parts more than fifty yards wide, with a fine
hard sandy bottom, and banks nearly perpendicular, and with a
strong current running three miles and a half in an hour to the
eastward. As I expected, every one of the Arabs said this was the
Nile, and that it ran into the great water the Tchad. A town of the
same name stands on the south side of the river, which the
inhabitants were unanimous in saying came from Soudan. It is at
times double the width, and considerably deeper, and two canoes
now lay upon the sand, in which the goods and passengers of kafilas
passing in the wet season are conveyed across. The camels and
horses swim with their heads made fast to the canoes. These
canoes were of the rudest manufacture, and were formed of planks,
rudely shaped by a small hatchet, and strongly fastened together by
cords passed through holes bored in them, and a wisp of straw
between, which they say effectually keeps out the water: they have
high poops like the Grecian boats, and would hold twenty or thirty
persons. The air from a running stream of clear water, and the
freshness it imparted to all around, was such a relief after a march
through sandy deserts, that both man and beast were in a manner
renovated by its effects. The men, and even the women, bathed and
washed, and the negroes swam all the horses. We here received ten
bullocks by the sheikh’s order, to make up the fifteen which he had
directed to be given to Boo Khaloom, and the remainder of the
seventeen kail (loads) of gussub which was to accompany them.
Feb. 14.—Visited Yeou, which is a neat town of huts, walled, but
not above half the size of Burwha, and proceeded fourteen miles,
when we came to a well. Here we should have remained with our
tired camels and horses; but the numerous negro parties, with from
two to twelve laden oxen, all said another well was grieb (near). Boo
Khaloom, therefore, determined on proceeding to the next maten, or
halting-place: some of the group were picturesque in the extreme;
the women all laden with some purchase at the market, and the
naked black children mounted on the tops of the loaded bullocks;
and after twelve additional miles, an hour and a half after sunset, we
came to a halt, but without arriving at the well. The branches of the
trees hung so much over the road, and impeded the movements of
the camels so greatly, that it was past ten o’clock at night before
some of them came up.
Feb. 15.—We found the well, kalielwa, just off the road, nearly
four miles nearer Bornou, and we were to push the camels on as far
as possible, in order that the day after we might enter Kouka, the
residence of the sheikh, in Arab form, and at an early hour. The road
branched off in two directions: the one to the west led towards
Kouka. Soon after this we came to a well and small town, and after
sunset another; near the latter of which a Fezzaneer in the service of
the sheikh met us, with a request that we would pitch our tents near
a dead water called Dowergoo, a few miles further on, and remain
the next day, as the huts that had been preparing were not ready.
About eight we came to this piece of still water, abounding with wild
fowl, having a village near it, called Gurdawa.
Feb. 16.—Halted. Our visitors here were not very numerous,
although we were not above one hour’s journey from the sheikh’s
residence, Kouka. Various were the reports as to the opinion the
sheikh formed of the force which accompanied Boo-Khaloom: all
agreed, however, that we were to be received at some distance from
the town, by a considerable body of troops; both as a compliment to
the bashaw, and to show his representative how well prepared he
was against any attempt of those who chose to be his enemies.
One of the Arabs brought to me this day a Balearic crane; it
measured thirteen feet from wing to wing.
Feb. 17.—This was to us a momentous day, and it seemed to be
equally so to our conductors. Notwithstanding all the difficulties that
had presented themselves at the various stages of our journey, we
were at last within a few short miles of our destination; were about to
become acquainted with a people who had never seen, or scarcely
heard of, a European; and to tread on ground, the knowledge and
true situation of which had hitherto been wholly unknown. These
ideas of course excited no common sensations; and could scarcely
be unaccompanied by strong hopes of our labours being beneficial
to the race amongst whom we were shortly to mix; of our laying the
first stone of a work which might lead to their civilization, if not their
emancipation from all their prejudices and ignorance, and probably,
at the same time, open a field of commerce to our own country,
which might increase its wealth and prosperity. Our accounts had
been so contradictory of the state of this country, that no opinion
could be formed as to the real condition or the numbers of its
inhabitants. We had been told that the sheikh’s soldiers were a few
ragged negroes armed with spears, who lived upon the plunder of
the Black Kaffir countries, by which he was surrounded, and which
he was enabled to subdue by the assistance of a few Arabs who
were in his service; and, again, we had been assured that his forces
were not only numerous, but to a certain degree well trained. The
degree of credit which might be attached to these reports was nearly
balanced in the scales of probability; and we advanced towards the
town of Kouka in a most interesting state of uncertainty, whether we
should find its chief at the head of thousands, or be received by him
under a tree, surrounded by a few naked slaves.
These doubts, however, were quickly removed. I had ridden on a
short distance in front of Boo-Khaloom, with his train of Arabs, all
mounted, and dressed out in their best apparel; and, from the
thickness of the trees, soon lost sight of them, fancying that the road
could not be mistaken. I rode still onwards, and on approaching a
spot less thickly planted, was not a little surprised to see in front of
me a body of several thousand cavalry drawn up in line, and
extending right and left quite as far as I could see; and, checking my
horse, I awaited the arrival of my party, under the shade of a wide-
spreading acacia. The Bornou troops remained quite steady, without
noise or confusion; and a few horsemen, who were moving about in
front giving directions, were the only persons out of the ranks. On the
Arabs appearing in sight, a shout, or yell, was given by the sheikh’s
people, which rent the air: a blast was blown from their rude
instruments of music equally loud, and they moved on to meet Boo-
Khaloom and his Arabs. There was an appearance of tact and
management in their movements which astonished me: three
separate small bodies, from the centre and each flank, kept charging
rapidly towards us, to within a few feet of our horses’ heads, without
checking the speed of their own until the moment of their halt, while
the whole body moved onwards. These parties were mounted on
small but very perfect horses, who stopped, and wheeled from their
utmost speed with great precision and expertness, shaking their
spears over their heads, exclaiming, “Barca! barca! Alla hiakkum
cha, alla cheraga!—Blessing! blessing! Sons of your country! Sons
of your country!” and returning quickly to the front of the body, in
order to repeat the charge. While all this was going on, they closed
in their right and left flanks, and surrounded the little body of Arab
warriors so completely, as to give the compliment of welcoming them
very much the appearance of a declaration of their contempt for their
weakness. I am quite sure this was premeditated; we were all so
closely pressed as to be nearly smothered, and in some danger from
the crowding of the horses and clashing of the spears. Moving on
was impossible; and we therefore came to a full stop: our chief was
much enraged, but it was all to no purpose, he was only answered

Accelerated Life Testing of One Shot Devices Data Collection and Analysis Narayanaswamy Balakrishnan Full Chapter

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Accelerated Life Testing of One Shot Devices Data Collection and Analysis Narayanaswamy Balakrishnan Full Chapter

Uploaded by

Copyright:

Available Formats

Accelerated Life Testing of One-shot

Devices: Data Collection and Analysis

Data Collection And Analysis

Limit of Liability/Disclaimer of Warranty

Library of Congress Cataloging-in-Publication Data

Cover Design: Wiley

Set in 9.5/12.5pt STIXTwoText by SPi Global, Chennai, India

Sarah and Julia Balakrishnan, and Colleen Cutler NB

1 One-Shot Device Testing Data 1

4 Model Mis-Speciﬁcation Analysis and Model Selection 65

6 Semi-Parametric Models and Inference 95

7 Optimal Design of Tests 105

7.3 Optimal Design with Budget Constraints 106

8 Design of Simple Step-Stress Accelerated Life-Tests 119

9 Competing-Risks Models 141

10 One-Shot Devices with Dependent Components 173

10.3.3 Frank Copula 177

11 Conclusions and Future Directions 187

Appendix A Derivation of Hi (a, b) 191

Appendix B Observed Information Matrix 193

Appendix C Non-Identiﬁable Parameters for SSALTs Under Weibull

Appendix D Optimal Design Under Weibull Distributions with Fixed

Appendix E Conditional Expectations for Competing Risks Model Under

Appendix F Kendall’s Tau for Frank Copula 205

July, 2020 Narayanaswamy Balakrishnan

About the Companion Website

This book is accompanied by a companion website:

One-Shot Device Testing Data

1.1 Brief Overview

1.2 One-Shot Devices

considered electro-explosive devices in military applications, which induct a

surveillance. The periodic inspection process with nondestructive evaluation

1.3 Accelerated Life-Tests

1.4 Examples in Reliability and Survival Studies

Table 1.1 Failure records on electro-explosive devices under CSALTs

Test Inspection Number of Number of

Source: Fan et al. (2009).

1.4.2 Glass Capacitors Data

1.4.3 Solder Joints Data

Test Inspection Number of Number of

1 450 443 200 8 1

Source: Zelen (1959).

Test Inspection Number of Number of

1 300 293 Yes 10 4

Source: Lau et al. (1988).

1.4.4 Grease-Based Magnetorheological Fluids Data

1.4.5 Mice Tumor Toxicological Data

1.4.6 ED01 Experiment Data

1.4.7 Serial Sacriﬁce Data

Inspection Number of Number of

Source: Zheng et al. (2018).

Source: Kodell and Nelson (1980).

Source: Lindsey and Ryan (1993).

Source: Berlin et al. (1979).

1.5 Recent Developments in One-Shot Device Testing

2.1 Brief Overview

2.2 Under CSALTs and Different Lifetime Distributions

1 (x11 , x12 , … , x1J ) 𝜏1 K1 n1

𝚿i = 𝚿(xi ; 𝜽). The observed likelihood function is then given by

where R(𝜏i ; 𝚿i ) = 1 − F(𝜏i ; 𝚿i ) is the reliability at inspection time 𝜏i and C is the

Suppose the lifetimes of devices follow a distribution with probability density

2.3.1 Exponential Distribution

FE (t; 𝜆i ) = 1 − exp(−𝜆i t), t > 0, (2.7)