Abstract
In this article, we consider step-stress accelerated life testing (SSALT) models assuming that the time-to-event distribution belongs to the proportional hazard family and the underlying population consists of long-term survivors. Further, with an increase in stress levels, it is natural that the mean time to the event of interest gets shortened and hence a method of obtaining order-restricted maximum likelihood estimators (MLEs) of the model parameters is proposed based on expectation maximization (EM) algorithm coupled with the reparametrization technique. To illustrate the effectiveness of the proposed method, extensive simulation experiments are performed and a real-life data example is analyzed in detail.
Similar content being viewed by others
References
Balakrishnan N, Beutner E, Kateri M (2009) Order restricted inference for exponential step-stress models. IEEE Trans Reliab 58:132–142
Bhattacharyya GK, Soejoeti Z (1989) A tampered failure rate model for step-stress accelerated life test. Commun Stat Theory Methods 18:1627–1643
Farewell VT (1982) The use of mixture models for the analysis of survival data with long-term survivors. Biometrika, pages 1041–1046
Greven S, Bailer AJ, Kupper LL, Muller KE, Craft JL (2004) A parametric model for studying organism fitness using step-stress experiments. Biometrics 60(3):793–799
Hanin L, Huang SL (2014) Identifiability of cure models revisited. J Multivar Anal 130:261–274
Kannan N, Kundu D (2020) Simple step-stress models with a cure fraction. Brazilian J Probab Stat 34:2–17
Kateri M, Kamps U (2015) Inference in step-stress models based on failure rates. Stat Pap 56:639–660
Khamis IH, Higgins JJ (1998) A new model for step-stress testing. IEEE Trans Reliab 47:131–134
Kundu D, Ganguly A (2017) Analysis of step-stress models: existing methods and recent developments. Elsevier/Academic Press, Amsterdam, The Netherlands, 2017
Maller RA, Zhou X (1996) Survival analysis with long-term survivors. Wiley, New York
Pal A, Mitra S, Kundu D (2020) Order restricted classical inference of a Weibull multiple step-stress model. J Appl Stat. https://doi.org/10.1080/02664763.2020.1736526
Pal A, Samanta D, Mitra S, Kundu D (2021) A simple step-stress model for Lehmann family of distributions. In: Advances in statistics-theory and applications : honoring the contributions of Barry C. Arnold in Statistical Science, Part-VI: 315–343
Samanta D, Ganguly A, Gupta A, Kundu D (2019) On classical and Bayesian order restricted inference for multiple exponential step stress model. Statistics 53(1):177–195
Samanta D, Kundu D (2018) Order restricted inference of a multiple step-stress model. Comput Stat Data Anal 117:62–75
Samanta D, Kundu D (2021) Meta-analysis of a step-stress experiment under Weibull distribution. J Stat Comput Simul. https://doi.org/10.1080/00949655.2021.1873992
Sedyakin NM (1966) On one physical principle in reliability theory. Tech Cybern 3:80–87
Self SG, Liang KL (1987) Asymptotic properties of the maximum likelihood estimators and likelihood ratio test under non-standard conditions. J Am Stat Assoc 82:605–610
Wang RH, Fei HL (2003) Uniqueness of the maximum likelihood estimate of the weibull distribution tampered failure rate model. Commun Stat Theory Methods 32(12):2321–2338
Acknowledgements
The authors would like to thank two unknown reviewers for their constructive suggestions which have helped to improve the manuscript significantly.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors do not have any conflict of interests in preparation of the manuscript.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This article is part of the topical collection “Special Issue: AISC-2021 Special Collection” guest edited by Tahani Coolen-Maturi, Javid Shabbir, and Arman Sabbaghi.
Appendix
Appendix
1.1 Proof of Lemma 1
For fixed \(\alpha \), we have \(m_1^{k+1}(p)=\bar{n}_{m+1} \ln p + \displaystyle \sum _{i \in J_{m+2}}(1-\pi _{i}^{(k)})\ln (1-p)+ \displaystyle \sum _{J_{m+2}}\pi _{i}^{(k)}\ln (p)+c,\) where c does not depend on p. Moreover,
and
Now, since \(m_1^{k+1''}(p)<0\), it immediately follows that \(m_1^{k+1}(p)\) is a concave function. The result follows immediately by considering the fact that \(m_1^{k+1}(p) \rightarrow -\infty \) as \(p \rightarrow 0\) or \(p \rightarrow 1.\)
1.2 Proof of Theorem 2
We provide the proof for four stress levels for illustration purposes. One can proceed along the similar lines for the general case. For \(m=3,\) the log-likelihood function at the \((k+1)-\) th step (see (10)) becomes
From \(k_1(p|\varvec{\theta ^{(k)}}),\) one can obtain the MLE of p explicitly using (13). Also, for fixed \(\alpha ,\) we have
where \(E_i(\alpha )\) is as defined in (8). The function (16) has a unique maximum at
\((\widehat{\beta }_1^{(\alpha )},\,\widehat{\beta }_2^{(\alpha )},\,\widehat{\beta }_3^{(\alpha )},\,\widehat{\lambda }_{4}^{(\alpha )}),\) where
Moreover, the function (16) does not have any other local maximum. It is to observe that for fixed \(\alpha \) and given \(\beta _1,\,\beta _2,\,\beta _3,\) the function (16) attains its maximum when
Plugging in \( \widehat{\lambda }_{4}^{(\alpha )}(\beta _1,\beta _2,\beta _3)\) in (16) and ignoring the additive constants (depending on \(\alpha \)), the profile log-likelihood function of \(\beta _1,\,\beta _2,\,\beta _3\) can be obtained as
Hence,
From (18), it is observed that the function \(s^{(\alpha )}(\beta _1,\beta _2,\beta _3)\) has a unique maximum at \((\widehat{\beta }_1^{(\alpha )},\,\widehat{\beta }_2^{(\alpha )},\,\widehat{\beta }_3^{(\alpha )}),\) where \(\widehat{\beta }_1^{(\alpha )},\,\widehat{\beta }_2^{(\alpha )}\,\,\text {and}\,\,\widehat{\beta }_3^{(\alpha )},\) are same as defined in (17), and the function (18) does not have any other local maximum.
Proposition 1
If \(\widehat{\beta }_1^{(\alpha )}>1,\) then
Proof
We prove it by contradiction. If the proposition is not true, then there must exist \(0<\widetilde{\beta }_1^{(\alpha )}<1,\,\widetilde{\beta }_2^{(\alpha )}>0,\,\widetilde{\beta }_3^{(\alpha )}>0,\) such that
Clearly, it implies that \((\widetilde{\beta }_1^{(\alpha )},\,\widetilde{\beta }_2^{(\alpha )},\,\widetilde{\beta }_3^{(\alpha )}) \ne (\widehat{\beta }_1^{(\alpha )},\,\widehat{\beta }_2^{(\alpha )},\,\widehat{\beta }_3^{(\alpha )}) \) is a local maximum of \(s^{(\alpha )}(\beta _1,\beta _2,\beta _3)\) as \(s^{(\alpha )}(\beta _1,\beta _2,\beta _3) \rightarrow -\infty \) as \(\beta _2 \rightarrow \infty \) and as \(s^{(\alpha )}(\beta _1,\beta _2,\beta _3) \rightarrow -\infty \) as \(\beta _3 \rightarrow \infty .\) \(\square \)
Along the same lines, it can be proved that if \(\widehat{\beta }_2^{(\alpha )}>1,\) then
if \(\widehat{\beta }_3^{(\alpha )}>1,\) then
Combining (19) and (20), we can obtain if \(\widetilde{\beta }_1^{(\alpha )}>1\) and \(\widetilde{\beta }_2^{(\alpha )}>1,\) then
Similarly from (19) and (21), it can be noted that if \(\widetilde{\beta }_1^{(\alpha )}>1\) and \(\widetilde{\beta }_3^{(\alpha )}>1,\) then
and if \(\widetilde{\beta }_2^{(\alpha )}>1\) and \(\widetilde{\beta }_3^{(\alpha )}>1,\) then
Finally, it is to note that if \(\widetilde{\beta }_1^{(\alpha )}>1\), \(\widetilde{\beta }_2^{(\alpha )}>1\) and \(\widetilde{\beta }_3^{(\alpha )}>1,\) then
Additionally, we can observe that
for all \(i\ne j\ne k\) and \(1\le i,j,k \le 3.\) Now let us consider the different cases.
Case 1: \(\widetilde{\beta }_1^{(\alpha )}>1,\,\widetilde{\beta }_2^{(\alpha )}>1\,\widetilde{\beta }_3^{(\alpha )}>1\)
The MLEs of \(\beta _1,\) \(\beta _2,\) and \(\beta _3\) are 1,1, and 1, respectively.
Case 2: \(\widetilde{\beta }_1^{(\alpha )}>1,\,\widetilde{\beta }_2^{(\alpha )}>1\,\widetilde{\beta }_3^{(\alpha )}\le 1\)
The MLEs of \(\beta _1\) and \(\beta _2\) are 1 and 1, respectively, and the MLEs of \(\beta _3\) can be obtained as \(\mathop {\textrm{arg}\,\textrm{max}}\limits _{ 0\le \beta _3 \le 1} s^{(\alpha )}(1,1,\beta _3).\) Since, \(s^{(\alpha )}(1,1,\beta _3)\) is an unimodal function, it has a unique maximum. Additionally,
owing to (26).
Case 3: \(\widetilde{\beta }_1^{(\alpha )}>1,\,\widetilde{\beta }_2^{(\alpha )} \le 1\,\widetilde{\beta }_3^{(\alpha )}\le 1\)
The MLE of \(\beta _1\) is 1 and the MLEs of \(\beta _2\) and \(\beta _3\) can be obtained as \(\mathop {\textrm{arg}\,\textrm{max}}\nolimits _{ 0\le \beta _{2},\beta _3 \le 1} s^{(\alpha )}(1,\beta _2,\beta _3).\) The function \(s^{(\alpha )}(1,\beta _2,\beta _3)\) has a unique maximum and we can repeat the same argument as before.
1.3 Proof of Theorem 3
We will provide the proof mainly for five stress level, i.e., for \(m=4\), although the same proof holds for the general case also. In case \(n_1=0\), the MLEs of all the parameters will not exist and therefore we will not consider this case. Note that the order-restricted MLEs still exist even if there is no failure at the last stress level. We will consider this case too in our proof.
Case 1: Exactly one internal stress level with zero failure.
Here we consider \(m=4\) and without loss of generality let \({n}_3=0\). The form of \(k_2^*(\varvec{\eta } \vert \varvec{\theta ^{(k)}})\) without the additive constant is given by
For a fixed \(\beta _2\), the function (28) is maximized when
Therefore, the profile log-likelihood of \(\beta _2\) without the additive constant is given by
Since,
the profile log-likelihood of \(\beta _2\) is an increasing function of \(\beta _2\) \((0< \beta _2 \le 1)\). Hence, the maximum occurs at \(\beta _2=1\). Therefore,
Case 2: Zero failure at two consecutive stress levels including the last stress level.
Here we have assumed that \({n}_4=0\) and \({n}_5=0.\) The form of \(k_2^*(\varvec{\eta } \vert \varvec{\theta ^{(k)}})\) without the additive constant is given by
For a fixed \(\beta _3\) and \(\beta _4\), the function (29) is maximized when
Therefore, the profile log-likelihood of \(\beta _3\) and \(\beta _4\) without the additive constant is given by
Since,
the profile log-likelihood is an increasing function of \(\beta _3\) \((0< \beta _3 \le 1)\). Similarly it can also be shown that the profile log-likelihood is an increasing function of \(\beta _4\) \((0< \beta _4 \le 1)\). Hence, the maximum occurs at \(\beta _3=1\) and \(\beta _4=1\). Therefore,
Other cases can also be done using the method proposed by [13]. Hence, the result follows. In general, it can be shown that the MLE of \(\beta _{k-1}=1\) if \(n_{k}=0\) for \(k=2, \ldots m+1\).
1.4 Elements of Fisher Information Matrix
Suppose \(\varvec{\theta }=(\theta _1,\theta _2,\theta _3,\theta _4,\theta _5)=(p,\alpha ,\lambda _{1},\lambda _{2},\lambda _{3}),\) and \( I_3(\varvec{\theta }) = \Bigg (\Bigg (\dfrac{\partial ^2 l(\varvec{\theta }|\mathcal{D})}{\partial \theta _i \partial \theta _j }\Bigg )\Bigg ),\,i,j=1,2,\ldots ,5,\) be the observed Fisher information matrix for three stress levels. Note that here \(l=l(\varvec{\theta }|\mathcal {D})\) is the log-likelihood function as defined in (3) for \(m=2.\) The elements of the observed Fisher information matrix can be expressed as follows:
where
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Pal, A., Samanta, D. & Kundu, D. Cure Rate-Based Step-Stress Model. J Stat Theory Pract 17, 15 (2023). https://doi.org/10.1007/s42519-022-00313-4
Accepted:
Published:
DOI: https://doi.org/10.1007/s42519-022-00313-4