Performance of parameter-estimates in step-stress accelerated life-tests with various sample-sizes

Ellen Overton McSorley, Jye Chyi Lu, Chin-Shang Li

Research output: Contribution to journalArticle

18 Citations (Scopus)

Abstract

In accelerated life test (ALT) studies, the maximum likelihood (ML) method is commonly used in estimating model parameters, and its asymptotic variance is the key quantity used in searching for the optimum design of ALT plans and in making statistical inferences. This paper uses simulation techniques to investigate the required sample size for using the large sample Gaussian approximation s-confidence interval and the properties of the ML estimators in the finite sample situation with different fitting models. Both the likelihood function and its second derivatives needed for calculating the asymptotic variance are very complicated. This paper shows that a sample size of 100 is needed in practice for using large-sample inference procedures. When the model is Weibull with a constant shape parameter, fitting exponential models can perform poorly in large-sample cases, and fitting Weibull models with a regression function of shape parameters can give undesirable results in small-sample situations. When small fractions of the product-life distribution are used to establish warranties and service polices, the interests of the producer and consumer should be balanced with the resources available for conducting life tests. From the standpoint of safety, an unrealistically long projected service life could harm the consumer. Establishing a short warranty period protects the producer but hurts revenue. An overly generous warranty period could cost the producer in terms of product replacement. To estimate life-distribution parameters well, use more than 100 samples in for LSCI and life-testing plans derived from the asymptotic theory. When the model becomes complicated, as in this paper, monitor the convergence of the parameter estimation algorithm, and do not trust the computer outputs blindly. In many applications, the likelihood ratio test inverted confidence interval performs better than the usual approximate confidence interval in small samples; thus its performance in the step-stress ALT studies should be in future research.

Original languageEnglish (US)
Pages (from-to)271-277
Number of pages7
JournalIEEE Transactions on Reliability
Volume51
Issue number3
DOIs
StatePublished - Sep 2002
Externally publishedYes

Fingerprint

Maximum likelihood
Law enforcement
Service life
Parameter estimation
Derivatives
Testing
Costs
Optimum design

Keywords

  • Confidence interval
  • Cumulative exposure model
  • Maximum likelihood estimation

ASJC Scopus subject areas

  • Electrical and Electronic Engineering
  • Hardware and Architecture
  • Computer Graphics and Computer-Aided Design
  • Software
  • Safety, Risk, Reliability and Quality

Cite this

Performance of parameter-estimates in step-stress accelerated life-tests with various sample-sizes. / McSorley, Ellen Overton; Lu, Jye Chyi; Li, Chin-Shang.

In: IEEE Transactions on Reliability, Vol. 51, No. 3, 09.2002, p. 271-277.

Research output: Contribution to journalArticle

@article{de0a941713c347a9bb762d1c16ba9d28,
title = "Performance of parameter-estimates in step-stress accelerated life-tests with various sample-sizes",
abstract = "In accelerated life test (ALT) studies, the maximum likelihood (ML) method is commonly used in estimating model parameters, and its asymptotic variance is the key quantity used in searching for the optimum design of ALT plans and in making statistical inferences. This paper uses simulation techniques to investigate the required sample size for using the large sample Gaussian approximation s-confidence interval and the properties of the ML estimators in the finite sample situation with different fitting models. Both the likelihood function and its second derivatives needed for calculating the asymptotic variance are very complicated. This paper shows that a sample size of 100 is needed in practice for using large-sample inference procedures. When the model is Weibull with a constant shape parameter, fitting exponential models can perform poorly in large-sample cases, and fitting Weibull models with a regression function of shape parameters can give undesirable results in small-sample situations. When small fractions of the product-life distribution are used to establish warranties and service polices, the interests of the producer and consumer should be balanced with the resources available for conducting life tests. From the standpoint of safety, an unrealistically long projected service life could harm the consumer. Establishing a short warranty period protects the producer but hurts revenue. An overly generous warranty period could cost the producer in terms of product replacement. To estimate life-distribution parameters well, use more than 100 samples in for LSCI and life-testing plans derived from the asymptotic theory. When the model becomes complicated, as in this paper, monitor the convergence of the parameter estimation algorithm, and do not trust the computer outputs blindly. In many applications, the likelihood ratio test inverted confidence interval performs better than the usual approximate confidence interval in small samples; thus its performance in the step-stress ALT studies should be in future research.",
keywords = "Confidence interval, Cumulative exposure model, Maximum likelihood estimation",
author = "McSorley, {Ellen Overton} and Lu, {Jye Chyi} and Chin-Shang Li",
year = "2002",
month = "9",
doi = "10.1109/TR.2002.802888",
language = "English (US)",
volume = "51",
pages = "271--277",
journal = "IEEE Transactions on Reliability",
issn = "0018-9529",
publisher = "Institute of Electrical and Electronics Engineers Inc.",
number = "3",

}

TY - JOUR

T1 - Performance of parameter-estimates in step-stress accelerated life-tests with various sample-sizes

AU - McSorley, Ellen Overton

AU - Lu, Jye Chyi

AU - Li, Chin-Shang

PY - 2002/9

Y1 - 2002/9

N2 - In accelerated life test (ALT) studies, the maximum likelihood (ML) method is commonly used in estimating model parameters, and its asymptotic variance is the key quantity used in searching for the optimum design of ALT plans and in making statistical inferences. This paper uses simulation techniques to investigate the required sample size for using the large sample Gaussian approximation s-confidence interval and the properties of the ML estimators in the finite sample situation with different fitting models. Both the likelihood function and its second derivatives needed for calculating the asymptotic variance are very complicated. This paper shows that a sample size of 100 is needed in practice for using large-sample inference procedures. When the model is Weibull with a constant shape parameter, fitting exponential models can perform poorly in large-sample cases, and fitting Weibull models with a regression function of shape parameters can give undesirable results in small-sample situations. When small fractions of the product-life distribution are used to establish warranties and service polices, the interests of the producer and consumer should be balanced with the resources available for conducting life tests. From the standpoint of safety, an unrealistically long projected service life could harm the consumer. Establishing a short warranty period protects the producer but hurts revenue. An overly generous warranty period could cost the producer in terms of product replacement. To estimate life-distribution parameters well, use more than 100 samples in for LSCI and life-testing plans derived from the asymptotic theory. When the model becomes complicated, as in this paper, monitor the convergence of the parameter estimation algorithm, and do not trust the computer outputs blindly. In many applications, the likelihood ratio test inverted confidence interval performs better than the usual approximate confidence interval in small samples; thus its performance in the step-stress ALT studies should be in future research.

AB - In accelerated life test (ALT) studies, the maximum likelihood (ML) method is commonly used in estimating model parameters, and its asymptotic variance is the key quantity used in searching for the optimum design of ALT plans and in making statistical inferences. This paper uses simulation techniques to investigate the required sample size for using the large sample Gaussian approximation s-confidence interval and the properties of the ML estimators in the finite sample situation with different fitting models. Both the likelihood function and its second derivatives needed for calculating the asymptotic variance are very complicated. This paper shows that a sample size of 100 is needed in practice for using large-sample inference procedures. When the model is Weibull with a constant shape parameter, fitting exponential models can perform poorly in large-sample cases, and fitting Weibull models with a regression function of shape parameters can give undesirable results in small-sample situations. When small fractions of the product-life distribution are used to establish warranties and service polices, the interests of the producer and consumer should be balanced with the resources available for conducting life tests. From the standpoint of safety, an unrealistically long projected service life could harm the consumer. Establishing a short warranty period protects the producer but hurts revenue. An overly generous warranty period could cost the producer in terms of product replacement. To estimate life-distribution parameters well, use more than 100 samples in for LSCI and life-testing plans derived from the asymptotic theory. When the model becomes complicated, as in this paper, monitor the convergence of the parameter estimation algorithm, and do not trust the computer outputs blindly. In many applications, the likelihood ratio test inverted confidence interval performs better than the usual approximate confidence interval in small samples; thus its performance in the step-stress ALT studies should be in future research.

KW - Confidence interval

KW - Cumulative exposure model

KW - Maximum likelihood estimation

UR - http://www.scopus.com/inward/record.url?scp=0036733183&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0036733183&partnerID=8YFLogxK

U2 - 10.1109/TR.2002.802888

DO - 10.1109/TR.2002.802888

M3 - Article

AN - SCOPUS:0036733183

VL - 51

SP - 271

EP - 277

JO - IEEE Transactions on Reliability

JF - IEEE Transactions on Reliability

SN - 0018-9529

IS - 3

ER -