Stochastic Models of Failure
4.5 MULTIVARIATE ANALOGUES OF THE FAILURE RATE FUNCTION Multivariate analogues of the failure rate function have been considered, among others, by
Basu (1971), Cox (1972), Johnson and Kotz (1975), Marshall (1975) and by Puri and Rubin (1974). One such analogue is the ‘hazard gradient’ discussed by Marshall (1975); this will be considered first.
4.5.1 The Hazard Gradient
A good starting point for introducing the hazard gradient is the cumulative hazard function Ht= −lnF t= t
0
rudu t≥0 (4.11)
mentioned in section (4.3).
ConsiderRt1 tn=PT1≥t1 Tn≥tn, the joint survival function ofnlifelengths T1 Tn, where, for convenience, we have omitted the. Lett=t1 tnbe such that Rt1 tn >0. Then, analogous to (4.11), we define Ht, the multivariate cumulative hazard functionasHt= −lnRt1 tn. If then-dimensional functionHthas a gradient, sayrt=Ht, wherert=r1t rnt, with
rit=
tHt i=1 n (4.12)
thenrtis called thehazard gradientofRt1 tn.
Johnson and Kotz (1975) interpretritas the conditional failure rate of Ti evaluated at ti, were it to be so thatTj> tj, for allj=i. That is
rit= fitiTj> tj for allj=i
PTi> tiTj> tj for allj=i (4.13) wherefitiTj> tj, for allj=iis the conditional probability density function ofTi, given that Tj> tj, for allj=i.
From elementary calculus, it now follows that Ht= t 0
rudu (4.14)
MULTIVARIATE ANALOGUES OF THE FAILURE RATE FUNCTION 77
from which we have as a multivariate analogue of (4.9) the result that PT1≥t1 Tn≥tn=exp− t
0
rudu (4.15)
the role of the hazard gradient is now apparent. Equation (4.14) holds as long as ruexists almost everywhere and that along the path of integration0t Htis absolutely continuous.
The following (additive) decomposition of Htgiven by Marshall (1975) is noteworthy; it parallels the (additive) decomposition ofHtin the univariate case. Specifically,
Ht= t1 0
r1u10 0du1+ t2 0
r2t1 u20 0du2+ + tn 0
rnt1 tn−1 undun (4.16) where r1u10 0 is the failure rate of T1 at u1, and rit1 ti−1 ui0 0 is the conditional failure rate of Ti at ui, were it be such that T1≥t1 T2≥t2 Ti−1≥ti−1. An interpretation of the decomposition of (4.16) will be given in section 4.6.2; however, a conse- quence of this decomposition is the well-known (multiplicative) decomposition ofRt1 tn. Specifically,
PT1≥t1 Tn≥tn=PT1≥t1 PT2≥t2T1≥t1 PTn≥tnT1≥t1 Tn−1≥tn−1 see Marshall (1975).
The univariate analogues of the additive and the multiplicative decompositions given above appear in section 4.3, following Equation (4.10).
4.5.2 The Multivariate Failure Rate Function
A more natural, though perhaps less useful, multivariate analogue of the univariate failure rate function is the multivariate failure rate function introduced by Basu (1971). I give below its bivariate version, and state some of its interesting characteristics.
Suppose that an absolutely continuous bivariate distribution functionFt1 t2withF00=0 generates a probability density function ft1 t2 at t1 t2. Then the bivariate failure rate of Ft1 t2at some1 2is, forPT1≥1 T2≥2 >0, given as
r1 2= f1 2
PT1≥1 T2≥2 (4.17)
= f1 2
1+F1 2−F1−F 2
Clearly, when T1 andT2 are mutually independent,r1 2=r1r2, where r1and r2are the corresponding univariate failure rates at 1 and2, respectively (equation (4.7)).
The relationship of (4.17) can be extended to the multivariate case.
As seen before, in section 4.4, the only univariate distribution whose failure rate is a constant is the exponential. We now ask if there is an analog to this result when ther1 2of (4.17) is a constant, say >0. Basu (1971) shows that the only absolutely continuous bivariate distribution with marginal distributions that are exponential and whose bivariate failure rate function is a constant, is the one that is obtained whenT1andT2 are exponentially distributed.
The search for multivariate distributions with exponential marginals is motivated by the fact that univariate exponential distributions, because of their simplicity and ease of use, are very popular in engineering reliability. The attractiveness of absolute continuity stems from the fact
that statistical inference involving such distributions is relatively easy (cf. Proschan and Sullo, 1974). If Basu’s (1971) requirement of exponential marginals is relaxed, then according to a theorem by Puri and Rubin (1974), the only absolutely continuous multivariate distribution whose multivariate failure rate (in the sense of (4.17)) is a constant is the one given by a mixture of exponential distributions.
Thus to summarize, it appears that the main value of the multivariate analogue of the failure rate function considered in this subsection is a characterization of the multivariate failure rate function that is a constant.
4.5.3 The Conditional Failure Rate Functions
Another multivariate analogue of the univariate failure rate function is the one proposed by Cox (1972) in his classic paper, and under the subheading ‘bivariate life tables’. In the bivariate case involving lifelengthsT1andT2we define, analogous to (4.4) of section 4.3, the following four univariate conditional failure rate functions:
rp0t=lim
dt↓0
Pt≤Tp≤t+dtT1≥t T2≥t
dt forp=12
r21tu=lim
dt↓0
Pt≤T2≤t+dtT2≥t T1=u
dt foru < t and r12tu=lim
dt↓0
Pt≤T1≤t+dtT1≥t T2=u
dt foru < t (4.18) In the latter two expressions,Ti=u i=12, is to be interpreted asTi∈u u+du, where duis infinitesimal.
A motivation for introducing the conditional failure rate functions parallels that for introducing the failure rate function of (4.4) and (4.5). Namely, that a subject matter specialist can subjectively specify the functionsrp0t, forp=12, based on the aging characteristics of an item, and the functionsrijt u i j=12 based on the aging as well as the load sharing features of the surviving item; for example, Freund (1961). IfFt1 t2denotes the joint distribution function ofT1 andT2, and ft1 t2the probability density at t1 t2generated byFt1 t2, then using arguments that parallel to those leading to (4.10), it can be shown that fort1≤t2
ft1 t2=exp− t1 0
r10u+r20udu− t2 t1
r21ut1dur10t1r21t2t1 (4.19) analogously for the caset2≤t1.
Consequently, as was the case with the univariate failure rate function, a specification of the conditional failure rate functions of (4.18) enables us to obtain ft1 t2and hence Ft1 t2. Alternatively, if we are to knowRt1 t2=PT1≥t1 T2≥t2, then we may obtainr10tand rijtu i j=12, via relationships of the type:
r10t= − 1 Rt t
tRt u
u=t
and
r12tu= −2Rt u
t u /Rt u u
Finally, it is relatively easy to show thatT1is independent ofT2, if and only ifr12tu=r10t, andr21tu=r20t.
THE HAZARD POTENTIAL OF ITEMS AND INDIVIDUALS 79 To conclude this section, we note thatRt1 tncan be obtained through (4.15) via the hazard gradientruor through (4.19) via the conditional failure rate function. The former is of mathematical interest, especially for the additive decomposition of (4.16); the latter has an operational appeal since it upholds the original spirit of introducing the failure rate function.