14.5The Weibull Distribution in Life Testing 603
could conclude that
yi≈βlogx(i)+logα, i=1,...,n (14.5.4)
We could then chooseαandβto minimize the sum of the squared errors — that is,α
andβare chosen to
minimize
α,β
∑n
i= 1
(yi−βlogx(i)−logα)^2
Indeed, using Proposition 9.2.1 we obtain that the preceding minimum is attained when
α=ˆα,β=βˆwhere
βˆ=
∑n
i= 1
yilogx(i)−nlogxy ̄
∑n
i= 1
(logx(i))^2 −n(logx)^2
logαˆ= ̄y−βlogx
where
logx=
∑n
i= 1
(logx(i))
/
n, y ̄=
∑n
i= 1
yi
/
n
To utilize the foregoing, we need to be able to determine valuesyithat approximate log
log(1/[ 1 −F(x(i))]=log[−log(1−F(x(i)))),i=1,...,n. We now present two different
methods for doing this.
Method 1:This method uses the fact that
E[F(X(i))]=
i
(n+1)
(14.5.5)
and then approximatesF(x(i))byE[F(X(i))]. Thus, this method calls for using
yi=log{−log(1−E[F(X(i))])} (14.5.6)
=log
{
−log
(
1 −
i
(n+1)
)}
=log
{
−log
(
n+ 1 −i
n+ 1
)}