\documentstyle[12pt,draft]{article} %\documentclass[birk,11pt]{book} %\pagestyle{myheadings} %\setcounter{page}{1} %__________________LAYOUT____________________________ \oddsidemargin 0cm \textwidth 15.3cm \textheight 24.2cm \topmargin -1.5cm \def \R{I\!\!R} %space \def \N{I\!\!N} %the set of natural numbers \def \X{X\!\!\!\!\!\!X} %sample \def \Z{Z\!\!\!Z} %the set of whole numbers \def \F{I\!\!F} \def \iint{\mathop{\int\!\!\!\int}\limits} %double integral \def \idotsint{\mathop{\int\!...\!\int}\limits} %multiply integral %___roman____________________________________________ \def \cov{{\bf Cov\,}} %Covariance \def \var{{\bf Var\,}} %Variance %___roman + bold_____________________________________ \def \PP{{\bf P}} %Probability \def \EE{{\bf E}} %Expected Value \def \Zero{{\displaystyle\bf 0}} %0 bold (\displaystyle for \frac) \def \Un{{\displaystyle\bf 1}} %1 bold (\displaystyle for \frac) %___italic___________________________________________ \newcommand{\imin}{\mathop{\it min\/}\limits} \newcommand{\imax}{\mathop{\it max\/}\limits} \newcommand{\iinf}{\mathop{\it inf\/}\limits} \newcommand{\isup}{\mathop{\it sup\/}\limits} \newcommand{\ilim}{\mathop{\it lim\/}\limits} %___italic + bold ___________________________________ \def\xx{\mbox{\boldmath$x$}} \def\aa{\mbox{\boldmath$a$}} \def\XX{\mbox{\boldmath$X$}} \def\AA{\mbox{\boldmath$A$}} \def\yy{\mbox{\boldmath$y$}} \def\bb{\mbox{\boldmath$b$}} \def\YY{\mbox{\boldmath$Y$}} \def\BB{\mbox{\boldmath$B$}} \def\zz{\mbox{\boldmath$z$}} \def\cc{\mbox{\boldmath$c$}} \def\ZZ{\mbox{\boldmath$Z$}} \def\CC{\mbox{\boldmath$C$}} \def\ii{\mbox{\boldmath$i$}} \def\dd{\mbox{\boldmath$d$}} \def\II{\mbox{\boldmath$I$}} \def\DD{\mbox{\boldmath$D$}} \def\vv{\mbox{\boldmath$v$}} \def\uu{\mbox{\boldmath$u$}} \def\VV{\mbox{\boldmath$V$}} \def\UU{\mbox{\boldmath$U$}} \def\hh{\mbox{\boldmath$h$}} \def\rr{\mbox{\boldmath$r$}} \def\HH{\mbox{\boldmath$H$}} \def\RR{\mbox{\boldmath$R$}} \def\jj{\mbox{\boldmath$j$}} \def\ff{\mbox{\boldmath$f$}} \def\JJ{\mbox{\boldmath$J$}} \def\FF{\mbox{\boldmath$F$}} \def\TT{\mbox{\boldmath$T$}} \def\WW{\mbox{\boldmath$W$}} \def\ww{\mbox{\boldmath$w$}} \def\QQ{\mbox{\boldmath$Q$}} \def\qq{\mbox{\boldmath$q$}} \def\GG{\mbox{\boldmath$G$}} \def\KK{\mbox{\boldmath$K$}} \def\LL{\mbox{\boldmath$L$}} \def\MM{\mbox{\boldmath$M$}} \def\NN{\mbox{\boldmath$N$}} \def\SS{\mbox{\boldmath$S$}} \def\pp{\mbox{\boldmath$p$}} \def\ttt{\mbox{\boldmath$t$}} %\tt for TeleType font \def\sss{\mbox{\boldmath$s$}} %\ss for gotic beta \def\g{\mbox{\boldmath$g$}} %\gg for >> \def\ee{\mbox{\boldmath$e$}} %\ee_1=(1,0,0,...,0) %___math_____________________________________________ \def \beq{\begin{equation}} \def \eeq{\end{equation}} \def \lf{\left\{} %{ \def \rf{\right\}} %} \def \lq{\left[} %[ \def \rq{\right]} %] \def \lc{\left(} %( \def \rc{\right)} %) \def \tran{^{\rm T}} \def \suml{\sum\limits} \def \prodl{\prod\limits} \def \intl{\int\limits} \def\barXX{\bar{\XX}} \def\barYY{\bar{\YY}} %___Greek______________________________________________ \def \SIGMA{{\bf\Sigma}} \def \ssigma{\mbox{\boldmath$\sigma$}} \def \TH{{\bf\Theta}} \def \th{\mbox{\boldmath$\theta$}} \def \barth{\bar{\theta}} \def \OO{{\bf\Omega}} \def \LAM{{\bf\Lambda}} \def \lam{\mbox{\boldmath$\lambda$}} \def \PSI{{\bf\Psi}} \def \PHI{{\bf\Phi}} \def \mm{\mbox{\boldmath$\mu$}} \def \nn{\mbox{\boldmath$\nu$}} \def \ba{\mbox{\boldmath$\alpha$}} \def \be{\mbox{\boldmath$\beta$}} \def \btau{\mbox{\boldmath$\tau$}} \def \DELTA{{\bf\Delta}} \def \del{\mbox{\boldmath$\delta$}} \def \eeps{\mbox{\boldmath$\varepsilon$}} %__Expressions______________________________________________ %__"md_*.tex" \def\pdf{probability density function } \def\pdfs{probability density functions } %__"test_hyp.tex","app3.tex" \def\prob{probability } \def\probs{probabilities } \def\dist{distinguishing } \def\suf{sufficient } \def\stats{statistics} \def\hyp{hypothesis} \def\hyps{hypotheses} %__"tables.tex" \def\kn{\;{\rm known}} \def\wh{{\rm where}\quad} \def\cmpn{\;{\rm are}\;{\rm components}\;{\rm of}\;} %__"norm_ue.tex" \def\csf{complete sufficient for } \begin{document} \centerline{\large \bf Statistical Analysis of Semiparametric Models} \centerline{\large \bf in Accelerated Life Testing} \vspace{1cm} \leftline{\scriptsize VILIJANDAS BAGDONAVI\v{C}IUS } \leftline{\scriptsize \it Department of Statistics, Vilnius University, Lithuania} \bigskip \leftline{\scriptsize MIKHAIL S. NIKULIN } \leftline{\scriptsize \it Steklov Mathematical Institute, 27 Fontanka, 191011, Saint Petersburg, Russia } \bigskip \noindent {\bf Abstract:} Several accelerated life models are introduced. Their relationships are investigated. Semiparametric estimation procedures are proposed. Properties of estimators are discussed \bigskip \noindent {\bf Keywords and phrases:} Accelerated life models, additive accumulation of damages, generalized proportional hazards, semiparametric estimation \vspace*{30pt} \leftline{\bf 1. Models} \vspace{0.5cm} \par Suppose that stresses are deterministic time functions: $$ x(\cdot)=(x_1(\cdot),...,x_m(\cdot))^T\,:[0,\infty)\to B\in {\bf R^m}. $$ Let $\Omega$ be a population of items and suppose that the time-to-failure of items under the stress $x(\cdot)$ is defined by a non-negative absolutely continuous random variable $T_{x(\cdot)}=T_{x(\cdot)}(\omega),\omega \in \Omega$, with the survival function $S_{x(\cdot)}(t) = {\bf P}\{T_{x(\cdot)}>t\}$ and the cummulative distribution function $F_{x(\cdot)}(t)=1-S_{x(\cdot)}(t)$. If $x(\cdot)$ is constant in time, we'll write $x$ instead of $x(\cdot)$ in all formulae. We shall consider here only the deterministic stresses. The case when the stress is a stochastic process X(t),can be done by the same way as in Bagdonavi\v{c}ius and Nikulin (1996, 1997c, 1997d) for the generalized additive and additive-multiplicative semiparametric models with random covariates. \par The proportion $F_{x(\cdot)}(t)$ of items from $\Omega$ which fail until the moment $t$ under the stress $x(\cdot)$ is called the {\it resource used until the moment $t$}. \par Let $$ \alpha_{x(\cdot)}(t) =\lim_{h\downarrow 0}\, \frac{1}{h}\, \PP \{T_{x(\cdot)}\in (t,t+h] \mid T_{x(\cdot)}>t\}=-\frac{S^{\prime}_{x(\cdot)}(t)}{S_{x(\cdot)}(t)}, \, t \geq 0, $$ be the hazard rate function under the stress $x(\cdot)$. Denote by $$ A_{x(\cdot)}(t)=\int^t_0 \alpha_{x(\cdot)}(u)du=-ln\{S_{x(\cdot)}(t)\}, \, t \geq 0, $$ the accumulated hazard rate under $x(\cdot)$. The first idea which comes by modelling the influence of a stress on lifetime distribution is to suppose that the hazard rate at any moment $t$ depends on the value of the stress at this moment and the resource used until $t$: \par {\it The generalized Sedyakin's }(GS) model (Bagdonavi\v{c}ius (1978),Bagdonavi\v{c}ius and Nikulin (1998) ) holds on the set of stresses $E $, if for all $x(\cdot)\in E$ $$ \alpha_{x(\cdot)}(t) = g \left(x(t),A_{x(\cdot)}(t) \right). $$ \par Consider the following important particular case of the GS model. Suppose that the hazard rate at any moment $t$ is proportional to a function of the stress used until this moment and to a function of the resource used until $t$: \par {\it The additive accumulation of damages }(AAD) model (Bagdonavi\v{c}ius (1978)) holds on the set of stresses $E$, if for all $x(\cdot)\in E$ $$ \alpha_{x(\cdot)}(t) = r \{x(t)\}\,q\{A_{x(\cdot)}(t) \}. $$ \par {\bf Proposition 1}. {\it The AAD model holds on $E$ iff there exists a survival function $G$ such that} $$ S_{x(\cdot)}(t) = G\left(\int^{t}_{0} r\{x(\tau)\}\;d\tau\right). \eqno (1) $$ \par The AAD model written in the form (1) is also called the {\it accelerated failure time} (AFT) model (Cox and Oakes (1984)). \par Cox (1972) considered the model which means that the hazard rate at any moment $t$ is proportional to some function of a stress applied at this moment and to a baseline rate which doesn't depend on the stress: \par The {\it proportional hazards} (PH) model holds on $E$ if for all $x(\cdot) \in {E}$ $$ \alpha_{x(\cdot)}(t)=r\{x(t)\}\,\,\alpha_0(t). \eqno(2) $$ \par When the PH model is also the AAD model? The answer is given in the following two propositions. \par {\bf Proposition 2} (Cox and Oakes (1984)). {\it Suppose that the PH model holds on a set $E_0$ of constant stresses such that the set $r(E_0)$ has an interior point. \par The AAD model also holds on $E_0$ iff the time-to-failure distribution is Weibull for all} $x \in E_0$. \par Suppose that $E_0$ is the set of constant stresses defined in Proposition 2, $x_1,x_2\in E_0$ are two fixed constant stresses and a step-stress $x_s(\cdot)$ has the form $$ x_s(\tau)=\left\{\begin{array}{cc} x_1,& 0 \leq \tau < s,\\ x_2,& \tau \geq s, \end{array} \right. \eqno (3) $$ where $s$ is a fixed positive number. \par {\bf Proposition 3}. {\it Suppose that the PH model holds on the set $E$ including $ E_0$ and ${x_s(\cdot)}$. The AAD model also holds on ${E}$ iff the time to-failure is exponential for all $x\in {E}_0$.} \par Proof. The Proposition 2 implies that for all $x\in E_0$ $$ S_{x}(t)=e^{-(\frac{t}{\theta (x)})^{\alpha}}. \eqno (4) $$ Put $\theta_i=\theta(x_i),\,i=1,2$. Then $$ S_{x_i}(t)=e^{-(\frac{t}{\theta_i})^{\alpha}},\quad \alpha_{x_i}(t)=\frac{\alpha}{\theta_i^{\alpha}}t^{\alpha-1}. $$ The PH model implies that $$ \alpha_{x_s(\cdot)}(t)=\left\{\begin{array}{cc} \alpha_{x_1}(t),& 0 < \tau < s,\\ \alpha_{x_2}(t),& \tau \geq s, \end{array} \right. $$ and for all $t>s$ $$ S_{x_s(\cdot)}(t)=\exp\{-\int_0^t \alpha_{x_s(\cdot)}(u)du\} =\exp\{ -\int_0^s \alpha_{x_1}(u)du-\int_s^t \alpha_{x_2}(u)du\} = $$ $$ \exp \left\{ -\left(\frac{s}{\theta_1}\right)^{\alpha} -\left(\frac{t}{\theta_2}\right)^{\alpha}+ \left(\frac{s}{\theta_2}\right)^{\alpha} \right\}. \eqno (5) $$ \par 1) Suppose that both the PH and the AAD models hold on $E$. Then (1) and (4) imply that for all $t>s$ $$ S_{x_s(\cdot)}(t)=\exp \left\{ -\left(\frac{s}{\theta_1}+ \frac{t-s}{\theta_2}\right)^{\alpha} \right\} \eqno (6) $$ The equalities (5) and (6) imply that for all $t>s$ $$ \exp \left\{ -\left(\frac{s}{\theta_1}\right)^{\alpha}-\left(\frac{t}{\theta_2}\right)^{\alpha}+ \left(\frac{s}{\theta_2}\right)^{\alpha} \right\} = \exp \left\{ -\left(\frac{s}{\theta_1}+ \frac{t-s}{\theta_2}\right)^{\alpha} \right\} . \eqno (7) $$ If $\alpha=1$, this equality is verified. Suppose that $\alpha \neq 1$. For all $t>s$ put $$ g(t)= \exp\left\{-\left(\frac{s}{\theta_1}\right)^\alpha-\left(\frac{t}{\theta_2}\right)^\alpha +\left(\frac{s}{\theta_2}\right)^\alpha\right\}- \exp\left\{-\left(\frac{s}{\theta_1}+\frac{t-s}{\theta_2}\right)^\alpha\right\}. $$ This function is strictly monotone for all $t>s$ with fixed $\theta_1\neq \theta_2$ and $\alpha\neq 1$. So the function $g$ is not constant in $t$ which contradicts to the equality (7). The assumption $\alpha \neq 1$ was false. Thus $\alpha =1$, and the lifetime distribution under any $x \in E_0$ is exponential. \par 2) Suppose that the PH model holds on $E$ and the time-to-failure is exponential for all $x \in E_0$. The formula (2) implies that for all $x \in E_0$ $$ S_x(t) = \exp\{-r(x)\,A_0(t)\}, \quad A_0(t)=\int_0^t \alpha_0(v)dv. $$ Exponentiality of the times-to-failure under $x\in E_0$ and the last formula imply that $A_0(t)=c\,t$. The constant $c$ can be included in $r(x)$, so we have $A_0(t)=t$. The formula (2) implies that $$ S_{x(\cdot)}(t)=\exp{\left\{-\int^t_0r\{x(u)\}du\right\}}, $$ i.e. the AAD model holds on $E$. \par The proof is complete. \par The GS model is more general then the AAD model. When the PH model is also the GS model? It is given in the following proposition. \par {\bf Proposition 4}. {\it Suppose that the PH model holds on the set ${E}$ including $E_0$ and all the stresses of the form (3) with $s<\delta$, where $\delta$ is any positive number. The GS model also holds on ${E}$ iff the time to-failure is exponential for all $x\in {E}_0$.} \par Proof. The PH model implies that for all $s<\delta $ $$ \alpha_{x_s(\cdot)}(t)=\alpha_{x_2}(t),\; t > s. $$ \par 1) If the GS model also holds on ${E}$, then for all $s<\delta$ $$ \alpha_{x_s(\cdot)}(t)=\alpha_{x_2}(t-s+\varphi (s)),\; t > s, $$ where $$ \varphi (s)=A^{-1}_{x_2}(A_{x_1}(s)) $$ is an increasing function. It implies that if both the GS and PH models hold on $E$ then for all $s_1<\delta$ and $s_2<\delta$ $$ \alpha_{x_2}(t-s_1+\varphi (s_1))=\alpha_{x_2}(t-s_2+\varphi (s_2)),\; t>max(s_1,s_2). $$ Any function $\alpha_{x_2}(t)\equiv $const verifies this. Assume that the function $\alpha_{x_2}(t)$ is not constant. Then $\varphi (s)-s=c=$const for all $s>0$, because the function $\alpha_{x_2}(t)$ cannot be two or more-periodic. Note that $c\neq 0$, because $$ A_{x_2}(\varphi(s))=A_{x_1}(s)\neq A_{x_2}(s). $$ The equalities $$ \lim_{s\to 0} A_{x_2}(\varphi (s))=\lim_{s\to 0} A_{x_1}(s)=0 $$ and the monotonicity of $\varphi (s)$ imply that $\lim_{s\to 0} \varphi (s)=0.$ So exists $\delta_0 \in (0,\delta)$ such that $\mid \varphi (s)-s \mid<\mid c \mid $ when $00$. It means that the assumption, that $\alpha_{x_2}(t)$ is not constant, was false. So $\alpha_{x_2}(t)=\alpha=$const which implies that $S_{x_2}(t)=e^{-\alpha_2 t}$. The PH model implies that for all $x\in E_0$ $$ S_{x}(t)=S_0(t)^{r(x)}=e^{-r(x)t}, $$ i.e. the time-to-failure distribution is exponential for all $x\in E_0$. \par 2) Suppose that for all $x\in E_0$ the time-to-failure distribution is exponential and the PH model holds. The proof of the Proposition 2.8 implies that the AAD model and consequently the GS model also holds on $E$. \par The proof is complete. \par The AAD and the PH models are rather restrictive. In the case of the AAD model the stress changes locally only the scale. \par Under the PH model the hazard rate under the stress $x(\cdot)$ at the moment $t$ {\it doesn't depend on the resource used until $t$}. It is not very natural if items are aging. Indeed, let ${E}_0 \subset {E}$ be a set of constant in time stresses, $x_0$ be an usual stress, $x_1$ be an accelerated with respect to $x_0$ stress, $x_0,x_1 \in {E}_0$, i.e. $S_{x_0}(t) > S_{x_1}(t)$ for all $t \geq 0$, and $E_1$ be a set of simple step-stresses of the form $$ x(t)=\left\{ \begin{array}{ll} x_1,& 0 \leq t < t_1,\\ x_0,& t \geq t_1. \end{array} \right. $$ If the PH model holds on $E_0\cup E_1$ then for all $ t_1 > 0$, $t > t_1$ we have $\alpha_{x(\cdot)}(t)=\alpha_{x_0}(t).$ \par If two groups of items are tested under the usual stress $x_0$ and the accelerated stress $x_1$, respectively, until a moment $t_1$ and after this moment both groups are observed under the same usual stress $x_0$, the failure rate after the moment $t_1$ is the same for both groups under the PH model. It is natural only for aging items. \par A generalization of AAD and PH models is obtained by supposing that the hazard rate at any moment $t$ is proportional to some function of the stress applied at this moment, to some function of the resource, used until $t$, and to some baseline rate. This is formalized by the following definition: \par {\it The first generalized proportional hazards} (GPH1) model ( Bagdonavi\v{c}ius and Nikulin (1999c)) holds on $E$ if for all $x(\cdot) \in {E}$ $$ \alpha_{x(\cdot)}(t)=r\{x(t)\}\;q\{A_{x(\cdot)}(t)\}\;\alpha_{0}(t). \eqno(8) $$ \par The particular cases of the GPH1 model are the PH model ($q(u)\equiv 1$) and the AAD model ($\alpha_0(t) \equiv \alpha_0=const$). \par A generalization of the GS and PH models is the following model: \par {\it The second generalized proportional hazards} (GPH2) model holds on $E$ if for all $x(\cdot) \in {E}$ $$ \alpha_{x(\cdot)}(t)=u\{x(t),A_{x(\cdot)}(t)\}\,\alpha_{0}(t). \eqno(9) $$ We have inclusions $$ AAD\subset GS \subset GPH1 \subset GPH2, \quad PH\subset GPH1 \subset GPH2. $$ If $AAD=PH$ on $E_0$ then the lifetime distribution is Weibull under any stress $x \in E_0$. If $AAD=PH$ on $E_1$ then the lifetime distribution is exponential under any stress $x \in E_0$ \par Particular cases of the GPH models give important accelerated life models. Numerous examples of real data show that taking two constant in time stresses, say $x_1$ and $x_2$, the ratio $\alpha_{x_2}(t)/\alpha_{x_1}(t)$ (which is constant under the PH model) can be increasing or decreasing in time and even a cross-effect of hazard rates can be observed, see Bagdonavi\v{c}ius and Nikulin (1999e). Such data can be modelled by submodels of the GPH1 or of more general the GPH2 model. \par {\bf Example 1.} Consider the following parametrization of $q$ in the GPH1 model: $q(u)=(1+u)^{\gamma+1}, $ where $\gamma \in I\!\!R$ is an unknown scalar parameter. We have the model $$ \alpha_{x(\cdot)}(t)=r\{x(t)\}(1+A_{x(\cdot)}(t))^{\gamma+1}\alpha_{0}(t). \eqno(10) $$ If $\gamma=-1$, it is the PH model. \par Suppose that $\gamma<0$ and $c_0=r(x_2)/r(x_1)$. Then $$ \alpha_{x_2}(t)/\alpha_{x_1}(t) =c_0 \left\{ \frac{1-\gamma\, r(x_2)\, A_0(t)}{1-\gamma\, r(x_1)\, A_0(t)}\right\}^{-1-\frac{1}{\gamma}}. $$ The ratio $\alpha_{x_2}(t)/\alpha_{x_1}(t)$ has the following properties: \par a) if $-1<\gamma<0$, then the ratio $\alpha_{x_2}(t)/\alpha_{x_1}(t)$ increases from the value $c_0$ until the value $c_{\infty}=\lim_{t\to \infty}\alpha_{x_2}(t)/\alpha_{x_1}(t)$, where the constant $c_{\infty}$ can take any value in the interval $(c_0, \infty)$; \par b) if $\gamma=-1$ (PH model), the ratio $\alpha_{x_2}(t)/\alpha_{x_1}(t)$ is constant in time. \par c) if $\gamma<-1$, then the ratio $\alpha_{x_2}(t)/\alpha_{x_1}(t)$ decreases from the value $c_0$ until the value $c_{\infty}\in (1, c_0)$. \par {\bf Example 2.} To obtain a cross-effect of hazard rates consider the following submodel of GPH2: $$ \alpha_{x(\cdot)}(t)=r( x(t))(1+A_{x(\cdot)}(t))^{\gamma^T x(t)+1}\, \alpha_{0}(t).\eqno(11) $$ Suppose that $c_0=r(x_2)/r(x_1)>1$ and $\gamma^T x_2<\gamma^T x_1<0$. Then $$ \alpha_{x_2}(t)/ \alpha_{x_1}(t)=c_0 \frac{(1-\gamma^T x_2\, r( x_2) A_0(t))^{-1-\frac{1}{\gamma^T x_2}}} {(1-\gamma^T x_1\, r(x_1)\, A_0(t))^{-1-\frac{1}{\gamma^T x_1}}} $$ and $$ \alpha_{x_2}(0)/ \alpha_{x_1}(0)=c_0>1,\;\; \lim_{t\to \infty}\alpha_{x_2}(t)/\alpha_{x_1}(t)=0. $$ So we have the cross-effect of the hazard rates.\\ \par {\bf Example 3.} Taking parametrization $q(u)=e^{\gamma u}, \; \gamma \in {\bf R},$ we obtain the model $$ \alpha_{x(\cdot)}(t)=r({x}(t))\,e^{\gamma A_{{x}(\cdot)}(t)}\,\alpha_0(t). \eqno(12) $$ This model is a generalization of the gamma frailty model (Vaupel et al. (1979)). If $\gamma =0$, it becomes the usual PH model. \par {\bf Example 4}. Taking $q(u)=1/(1+\gamma u), \quad \gamma > 0$, we obtain the model $$ \alpha_{x(\cdot)}(t)=r({x}(t))\,\frac{1}{1+\gamma A_{{x}(\cdot)}(t)}\,\alpha_0(t). $$ \par Consider now an important model which doesn't lie in the class of the GPH models but includes the AAD model as the particular case: \par {\it Locally changing shape and scale} (LCSS) model (\, Bagdonavi\v{c}ius and Nikulin (1998, 1999a)) holds on $E$ if $$ S_{x(\cdot)}(t)=G\left(\int^t_0r\{x(\tau)\}\tau^{\nu(x(\tau))-1}d\tau\right). \eqno(13) $$ Schabe and Viertl (1995) considered an axiomatic approach to model building. \par {\bf Proposition 4} (Schabe and Viertl (1995)). {\it Suppose that exists a functional $a:E\times E\times [0, \infty )\rightarrow [0, \infty )$ such that for any $x_1(\cdot),x_2(\cdot) \in E$ it is differentiable and increasing in $t$, $a(x_1(\cdot),x_2(\cdot),0)=0$ and $T_{x_2(\cdot)} \sim a(x_1(\cdot),x_2(\cdot),T_{x_1(\cdot)})$, where $\sim $ denotes equality in distribution. \par For any differentiable on $[0, \infty )$ survival function $S$ exists a functional} $b:E\times [0, \infty ) \rightarrow [0, \infty )$ {\it such that for all} $x(\cdot)\in E$ $$ S_{x(\cdot)}(t)=S\left(\int^t_0b\{x(\cdot),u\}\,du\right). $$ All above considered models are particular cases of this general Schabe-Viertl model. See also Bagdonavi\v{c}ius and Nikulin (1999d). \par The GS and AAD models are not appropriate when stresses are periodic with quick change of their values. Greater is the number of stress cycles, shorter is the life of items. So the effect of cycling must be included in the model. \par Suppose that a periodic stress is differentiable. Then the number of cycles in the interval $[0,t]$ is $$ n(t)=\int^t_0 \mid d \Un \{x'(u)> 0\}\mid . $$ Generalization of the GPH1 model has the form $$ S_{x(\cdot)}(t)=G\left\{\int_0^t r_1\{x(u)\}\,dH(S_0(u))+\int_0^t r_2\{x(u)\}\mid d \Un \{x'(u)> 0\}\right\}. $$ The second integral includes the effect of cycling on resource usage. \par Taking $H=S_0^{-1}$, we obtain a generalization of the AAD model which includes the effect of cycling. The GS and AAD models are also not appropriate if $x(\cdot)$ is a step stress with many switch on's and switch off's which shorten the life of items. In this case the following model can be considered: $$ S_{x(\cdot)}(t)=\int_0^t r_1\{x(u)\}\,dH(S_0(u))+\int_0^t r_2\{x(u)\} \Un (\Delta x(u)>0)\frac{\mid d x(u)\mid}{\mid \Delta x(u)\mid } $$ $$ +\int_0^t r_3\{x(u)\} \Un (\Delta x(u)<0)\frac{\mid d x(u)\mid}{\mid \Delta x(u)\mid } . $$ The second and the third terms include the effect of switch-on's and switch-off's on resource usage. Taking $H=S_0^{-1}$, we obtain a generalization of the AAD model which includes this effect. \vspace*{30pt} \leftline{\bf 2. Semiparametric estimation} \vspace{0.5cm} Semiparametric estimation for the \, AAD \, model was considered by Basu and Ebrahimi (1982), Sethuraman and Singpurwalla (1982), Shaked and Singpurwalla (1983), Schmoyer (1986, 1991), Robin and Tsiatis (1992), Lin and Ying (1995), Bagdonavi\v{c}ius and Nikulin (1999a,b). \par Estimation for the PH model was developed by Cox (1972), Tsiatis (1981), Andersen and Gill (1982). \par Particular cases of the GPH1 model with parametrization of $q$ was considered by Andersen, Borgan, Gill and Keiding (1993), with completely known $q$ by Dabrowska and Doksum (1988a,b), Cheng and others (1995), Murphy and others (1997), Bagdonavi\v{c}ius and Nikulin (1997a,b). \par Estimation in LCSS model was discussed by Bagdonavi\v{c}ius and Nikulin (1998, 1999a,b,c). \par We consider the most general GPH2 model $$ \alpha_{x(\cdot)}(t)=u\{x(t),A_{x(\cdot)}(t),\theta\} \,\alpha_{0}(t). \eqno(14) $$ with any specified parametrization $u=u(x,s,\theta)$ via the parameters $\theta$ and unknown baseline function $\alpha_{0}(t)$. \par Suppose that $m$ groups of items are tested. The ith group of $k_i$ items is tested under the accelerated stress $x_i(\cdot)$. Denote by $T_{ij}$ and $C_{ij}$ the failure and censoring times for the jth item of the ith group and let $$ X_{ij}=T_{ij}\wedge C_{ij}, \, \delta_{ij}=I_{\{T_{ij} \leq C_{ij}\}}, \, N_{ij}(t)=I_{\{T_{ij}\leq t, \delta_{ij}=1\}}, \, Y_{ij}(t)=I_{\{X_{ij} \geq t\}}, $$ where $I_A$ denotes the indicator of the event $A$. Then $$ N_i(t)=\sum^{k_i}_{j=1}N_{ij}(t)\quad \mbox{ and}\quad Y_i(t)=\sum^{k_i}_{i=1}Y_{ij}(t) $$ are the numbers of observed failures in the interval $[0,t]$ and items "at risk" just prior to the moment $t$, respectively, for the ith group of items. \par We suppose that times-to-failures $T_{ij}$ are absolutely continuous random variables and censoring is independent right censoring. \par Suppose that maximal time given for experiment is $\tau \in (0,\infty)$ and all items which did not fail and were not censored before $\tau$, are censored at this moment. \par The partial likelihood function (see Andersen and others (1993)) $$ L(\theta)=\prod_{i=1}^m \prod _{j=1}^{k_i}\left(\int_0^{\tau}\frac{u\{x_i(v),A_i(v),\theta\}}{\sum_{l=1}^m Y_l(v)\,u\{x_l(v),A_l(v),\theta\}}dN_{ij}(v)\right)^{\delta_{ij}} \eqno(15) $$ depend on unknown accumulated hazard rates $A_i(t)=\int_0^t \alpha_{x_i}(u)du.$ \par Modify the partial likelihood function in the following manner: replace the functions $A_i(t)$ in (15) by the Nelson-Aalen estimators: $$ \hat A_i(t)=\int_0^t \frac{dN_i(s)}{Y_i(s)}. $$ The modified partial likelihood function $$ \tilde L(\theta)=\prod_{i=1}^m \prod _{j=1}^{k_i}\left(\int_0^{\tau}\frac{u\{x_i(v),\hat A_i(v),\theta\}}{\sum_{l=1}^m Y_l(v)\,u\{x_l(v),\hat A_l(v),\theta\}}dN_{ij}(v)\right)^{\delta_{ij}} \eqno(16) $$ Put $g(x,t,\theta)=\frac{\partial}{\partial \theta} u(x,t,\theta)$. The score function has the form $$ U(\theta)=\sum_{i=1}^m\int_0^{\tau}\left(\frac{g\{x_i(v),\hat A_i(v),\theta\}}{u\{x_i(v),\hat A_i(v),\theta\}}-\frac {\sum_{l=1}^m Y_l(v)\,g\{x_l(v),\hat A_l(v),\theta\}}{\sum_{l=1}^m Y_l(v)\,u\{x_l(v),\hat A_l(v),\theta\}}\right)dN_i(v).\eqno(17) $$ If $k_i$ are small, the Nelson-Aalen estimators can be not very efficient and the functions $A_i$ should be replaced by some other estimator which is a function of all data. This can be done in the following manner. \par The model (14) implies that $$ dA_i(t)=u\{x_{i}(t),A_i(t),\theta\}\,dA_{0}(t). $$ Using considerations similar to considerations used to obtain the Nelson-Aalen estimator for the PH model we obtain a pseudoestimator $\tilde A_0(t,\theta)$ (depending on $\theta$ ) for $A_0(t)$: $$ \tilde A_0(t,\theta)= \int^t_0 \frac{dN(u)}{\sum^m_{i=1} u\{x_i(u),\tilde {A}_i(u-;\theta),\theta\}\,Y_i(u)}. $$ The last two equalities give the recurrent equations for a pseudoestimator $\tilde {A}_i(t;\theta)$ of ${A}_i(t)$: $$ \tilde {A}_i(t;\theta)= \int^t_0 \frac{u\{x_i(u),\tilde{A}_i(u-,\theta),\theta\}\,dN(u)}{\sum^m_{l=1} u\{x_l(u),\tilde {A}_l(u-;\theta),\theta\}\,Y_l(u)},\,\,\tilde {A}_i(0;\theta)=0. \eqno(18) $$ \par Modify the score function (17) by considering the score function $$ \tilde U(\theta)= \sum_{i=1}^m\int_0^{\tau}\left(\frac{g\{x_i(v),\tilde A_i(v,\theta),\theta\}}{u\{x_i(v),\tilde A_i(v,\theta),\theta\}}-\frac {\sum_{l=1}^m Y_l(v)\,g\{x_l(v),\tilde A_l(v,\theta),\theta\}}{\sum_{l=1}^m Y_l(v)\,u\{x_l(v),\tilde A_l(v,\theta),\theta\}}\right).\eqno(19) $$ Note that to find an initial estimator $\hat \theta^{(0)}$ verifying the equations (19) you need initial estimators $\tilde{A}^{(0)}_i(t,\theta)$ and {\it vice versa}: to find initial estimators $\tilde{A}^{(0)}_i(t,\theta)$ you need an initial estimator $\hat \theta^{(0)}$. This problem can be solved by taking the initial estimator $\hat \theta^{(0)}$ as the solution of the equations $U(\theta)=0$, where the function $U(\theta)$ is defined by the formula (17). Then the initial estimators $\tilde{A}^{(0)}_i(t,\theta^{(0)})$ are obtained recurrently from the equation (18). The estimator $\hat \theta^{(1)}$ is obtained by solving the equations $\tilde U(\theta)=0$, where $\tilde U(\theta)$ is given by (19) with $\tilde{A}^{(0)}_i(t,\theta^{(0)})$ taking place of $\tilde{A}_i(t,\theta)$. And so on. Given the estimator $\hat \theta^{(l)}$, the estimator $\tilde{A}^{(l)}_i(t,\theta^{(l)})$ is right-continuous step-function with jumps at the observed failure moments $T_{(j)}$ ! ! (whi ch are the jump points of $N(t)$). \par The estimator of the survival function $S_{x_0}(t),\,t\leq \tau$ under the usual stress is $\hat S_{x_0}(t)=\exp \{-\hat A_{x_0}(t)\}$, where $\hat A_{x_0}(t)$ is found recurrently from the equation $$ \hat A_{x_0}(t)= \int^t_0 \frac{u\{x_0,\hat{A}_{x_0}(u-),\hat \theta\}\,dN(u)}{\sum^m_{l=1} u\{x_0,\tilde {A}_l(u-;\hat \theta),\hat \theta\}\,Y_l(u)},\,\,\hat A_{x_0}(0)=0. $$ \vspace*{30pt} \leftline{\bf 3. Properties of the estimators} \vspace{0.5cm} Suppose that the covariates $x_i(\tau)\equiv x_i$ are constant in time. Denote by $\theta_0$ the true value of $\theta$ under the GPH2 model, $\|A\|=\sup_{i,j}|a_{ij}|$ - the norm of the matrix $A=(a_{ij})$, $A^{\otimes 2}$ the product $A{A}^T$, $$ w_j(v;\theta)=\frac{g\{x_j,A_j(v),\theta\}}{u\{x_j,A_j(v),\theta\}} \quad S^{(0)}(v;\theta)=\sum^m_{j=1}Y_j(v)\,u\{x_j,A_j(v),\theta\}, $$ $$ S^{(1)}(v;\theta)=\sum^m_{j=1}Y_j(v)\,g\{x_j,A_j(v),\theta\}, \quad E(v;\theta)=\frac{S^{(1)}(v;\theta)}{S^{(0)}(v;\theta)}, $$ $$ S^{(0)}_*(v;\theta)=\sum^n_{i=1}Y_i(v)\,u\{x_j,A_j(v),\theta\}u_2\{x_j,A_j(v),\theta\}, $$ $$ S^{(1)}_*(v;\theta)=\sum^n_{j=1}g\{x_j,A_j(v),\theta\}\,u_2\{x_j,A_j(v),\theta\}, $$ $$ S^{(2)}(v;\theta)=\sum^n_{j=1}\frac{\partial w_j(v;\theta)}{\partial \theta}Y_j(v)\,u\{x_j,A_j(v),\theta\}, $$ $$ S^{(2)}_*(v;\theta)=\sum^n_{j=1}u\{x_j,A_j(v),\theta\}\,g_2\{x_j,A_j(v),\theta\}, $$ where $u_2$ denotes the partial derivative of $u$ with respect to the second argument. \par {\bf Assumptions A:}\\ \noindent a). Suppose that exist a neighbourhood $\theta$ of $\theta_0$ and continuous on $\theta$ uniformly in $t \in [0,\tau]$ and bounded on $\theta \times [0,\tau]$ functions $\, s^{(i)}(v; \theta)$, $s^{(i)}_*(v;\theta)$, such that $ s^{(0)}(v;\theta_0)>0$ on $[0,\tau]$ and $$ \sup_{\theta \in \theta,\, v\in [0,\tau]}\|\frac{1}{n}S^{(i)}(v;\theta)-s^{(i)}(v;\theta)\| \rightarrow 0 \quad \mbox{as} \quad n \rightarrow \infty,, $$ $$ \sup_{\theta \in \theta,\, v\in [0,\tau]}\|\frac{1}{n}S^{(i)}_*(v;\theta)-s^{(i)}_*(v;\theta)\| \rightarrow 0 \quad \mbox{as} \quad n \rightarrow \infty, \,\, (i=0,1,2), $$ $$ \sup_{\theta \in \theta, \, v \in [0,\tau]} \left\|\frac{\partial E(v;\theta)}{\partial \theta}-\frac{\partial e(v;\theta)}{\partial \theta}\right\| \stackrel{{\bf P}}{\rightarrow}0,\quad e(v;\theta)=\frac{s^{(1)}(v;\theta)}{s^{(0)}(v;\theta)}. $$ \noindent b). $A_0(\tau) < \infty$. \par Put $e_*(v;\theta)=s_*^{(0)}(v;\theta)/s^{(0)}(v;\theta)$, $h(t;\theta)=exp\{ \int^t_0e_*(v;\theta)dA_0(v)\}$ , $$ h_1(v;\theta)=\frac{s^{(1)}(v;\theta) s^{(0)}_*(v;\theta)-s^{(0)}(v;\theta) s^{(1)}_*(v;\theta)}{s^{(0)}(v;\theta)}-s^{(2)}_*(v;\theta), $$ $$ w(v;\theta)=e(v;\theta)-\frac{1}{h(v;\theta)s^{(0)}(v;\theta)}\int^{\tau}_vh_1(s;\theta)h(s;\theta)dA_0(s). $$ \noindent c). $\frac{1}{n}\sum^n_{i=1}\int^{\tau}_0J(v) (w_i(v;\theta)-w(v;\theta)^{\otimes 2}u\{x_j,A_j(v),\theta\} Y_i(v)dA_0(u)\stackrel{{\bf P}}{\rightarrow} \SIGMA(\theta). $\\ \noindent d).The matrix $$ \SIGMA_1(\theta_0)=-\int^t_0\left(s^{(2)}(u;\theta_0)- \frac{\partial e(u;\theta_0)}{\partial \theta}s^{(0)}(u;\theta_0)\right)dA_0(u) $$ is positive definite. \par {\bf Theorem 1}. {\it Under Assumptions A} $$ n^{1/2}(\hat{\theta}-\theta_0)\stackrel{{\cal D}}{\rightarrow} N(0,\SIGMA_1^{-1}(\theta_0) \SIGMA(\theta_0)\left(\SIGMA_1^{-1}(\theta_0)\right)^{T}). $$ \par Consider the asymptotic distribution of the survival function $S_{x_0}(t)$ under the usual stress $x_0$. Put $$ C(\theta_0)=-\int^t_0\frac{\partial}{\partial \theta}\ln{S^{(0)}(v;\theta_0)}dA_0(v), $$ $$ H_i(v;\theta_0)=J(v)\{C(\theta_0)\SIGMA_1^{-1}(\theta_0)(w_i(v;\theta_0)-w(v;\theta_0))+\frac{1}{s^{(0)} (v;\theta_0)}\}, $$ \par {\bf Assumptions B}. $$ \frac{1}{n}\sum^m_{i=1}\int^t_0H_{i}^2(v;\theta_0)\,Y_i(v)\,u\{x_i,A_i(v),\theta\}\,dA_0(v)\stackrel{{\PP}}{\rightarrow} \sigma^2(t). $$ Put $$ \sigma_{x_0}^2(t) = S^2_{x_0}(t)\,u^2\{x_0,A_{x_0}(t),\theta\}\,\sigma^2(t). $$ \par {\bf Theorem 2}. {\it Under Assumptions A and B} $$ \sqrt{n}(\hat{S}_{x_0}(t)-S_{x_0}(t))\stackrel{\cal D}{\rightarrow}N(0,\sigma_{x_0}^2(t)). $$ \vspace{0.5cm} \leftline{\bf References} \vspace{0.5cm} \begin{enumerate} \item Andersen, P.K. and Gill, R.D. (1982). Cox's regression model for counting processes: A large sample study, {\it Ann. Statist}, {\bf 10}, 1100--1120. \item Andersen, P.K., Borgan, O.,Gill, R.D. and Keiding, N. (1993). {\it Statistical Models Based on Counting Processes}, New York: Springer-Verlag. \item Bagdonavi\v{c}ius, V. (1978). Testing the hyphothesis of the additive accumulation of damages, {\it Probab. Theory and its Appl.}, {\bf 23}, 403--408. \item Bagdonavi\v{c}ius, V., Nikulin, M., (1996). Analysis of generalized additive semiparametric models. {\it Comptes Rendus, Academie des Sciences de Paris}, {\bf 323}, 9, Serie I,1079-1084. \item Bagdonavi\v{c}ius, V., Nikulin, M., (1997a). Transfer functionals and semiparametric regression models. {\it Biometrika}, {\bf 84} , 365-378. \item Bagdonavi\v{c}ius, V., Nikulin, M., (1997b). Asymptotic analysis of semiparametric models in survival analysis and accelerated life testing. {\it Statistics}, {\bf 29}, 261-283. \item Bagdonavi\v{c}ius, V., Nikulin, M., (1997c). Statistical analysis of generalized additive semiparametric models with random covariates. {\it Questiio}, {\bf 21}, 1-2, 279-299. \item Bagdonavi\v{c}ius, V., Nikulin, M., (1997d). Analysis of generalized semiparametric models with random covariates. {\it Romanian Journal of Pure and Applied Mathematics}, {\bf 42}, 5-6, 351-369. \item Bagdonavi\v{c}ius, V., Nikulin, M., (1998). {\it Additive and Multiplicative Semiparametric Models in Accelerated Life testing and Survival Analysis}, Queen's Papers in Pure and Applied Mathematics, {\bf 108}, Kingston, Canada. \item Bagdonavi\v{c}ius, V. and Nikulin, M. (1999a). On semiparametric estimation of reliability from accelerated data, In: {\it Statistical and Probabilistic Models in Reliability}, Boston: Birkhauser (ed. Ionescu, D.C. and Limnios, N.), 75--89. \item Bagdonavi\v{c}ius, V. and Nikulin, M. (1999b). On nonparametric estimation in accelerated experiments with step-stresses, {\it Statistics}, {\bf 31}, to appear. \item Bagdonavi\v{c}ius, V. and Nikulin, M. (1999c). Generalized Proportional Hazards Models Based on Modified Partial Likelihood, {\it Lifetime Data Analysis}, {\bf 5}, 323-344. \item Bagdonavi\v{c}ius, V. and Nikulin, M. (1999d). Models building in accelerated experiments, In: {\it Statistical and Probabilistic Models in Reliability}, \, Boston: Birkhauser (eds. Ionescu, D.C. and Limnios, N.), 51-73. \item Bagdonavi\v{c}ius, V. and Nikulin, M. (1999e). On goodness-of-fit tests for the linear transformation and frailty models, {\it Statistics and Probability Letters}, {\bf 45}, to appear. \item Basu, A.P. and Ebrahimi, N., (1982). Nonparametric accelerated life testing, {\it IEEE Trans. on Reliability }, {\bf 31}, 432--435. \item Cheng, S.C., Wei, L.J. and Ying, Z. (1995). Analysis of tranformation models with censored data, {\it Biometrika}, {\bf 82}, 835--846. \item Cox, D.R. (1972). Regression models and life tables, {\it J.Roy.Statist. Soc.}, B, {\bf 34}, 187--220. \item Cox, D.R. and Oakes, D. (1984). {\it Analysis of Survival Date}, London: Chapman and Hall. \item Dabrowska, D.M. and Doksum, K.A. (1988a). Estimation and Testing in a Two-Sample Generalized Odds-Rate Model, {\it JASA}, {\bf 83} 744--749. \item Dabrowska, D.M. and Doksum, K.A. (1988b). Partial likelihood in transformation model with censored data, {\it Scand. J. Statist.}, {\bf 15}, 1--23. \item Lin, D.Y. and Ying, Z., (1995). Semiparametric inference for the accelerated life model with time dependent covariates, {\it Journal of Statistical Planning and Inference}, {\bf 44}, 47--63. \item Murphy, S.A., Rossini, A.J. and van der Vaart, A.W. (1997). Maximum likelihood estimation in the proportional odds model, {\it JASA.}, {\bf 92}, 968--976. \item Robins, J.M. and Tsiatis, A.A., (1992). Semiparametric estimation of an accelerated failure time model with time dependent covariates, {\it Biometrika}, {\bf 79}, 311--319. \item Schabe,H., Viertl, R., (1995). An axiomatic approach to models of accelerating life testing, {\it Engineering Fracture Mechanics}, {\bf 30}, 203-217. \item Schmoyer, R. (1986). An exact distribution-free analysis for accelerated life testing at several levels of a single stress, {\it Technometrics}, {\bf 28}, 2, 165--175. \item Schmoyer, R. ( 1991). Nonparametric analysis for two-level single-stress accelerated life tests, {\it Technometrics}, {\bf 33}, 175-186. \item Sethuraman, J. and Singpurwalla, N.D. (1982). Testing of hypotheses for distributions in accelerated life tests, {\it JASA},{\bf 77}, 204-208. \item Shaked, M. and Singpurwalla, N.D. (1983). Inference for step-stress accelerated life tests. J. Statist. Plann. Inference, {\bf 7}, 295-306. \item Tsiatis, A.A. (1981). A large sample study of Cox's regression model, {\it Ann. Statist.,} {\bf 9}, 93--108. \item Vaupel, J.W., Manton, K.G. and Stallard, E. (1979). The impact of heterogeneity in individual frailty on the dynamics of mortality, {\it Demography}, {\bf 16} 439--454. \end{enumerate} \end{document}