Professional Documents
Culture Documents
7
Nonparametric Statistical Methods
105
DK598X_book.fm Page 106 Monday, January 24, 2005 11:11 PM
106 Nachlas
x j = t j (7.1)
108 Nachlas
k
FˆT (t) =
n k = { j|x j ≤ t, x j +1 > t} (7.2)
ˆ n −k
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
FT (t) =
n
where the caret, “^” above a quantity indicates an estimate.
Now, while these expressions seem intuitively logical, they
can be shown to be based on statistical reasoning and to be
one of two reasonable forms.
Assume an arbitrary time interval, say (t, t + ∆T) and
let pj denote the probability that the jth device failure observed
occurs during that interval. In principle, any of the failures
could occur during that interval so the probability that it is
the jth failure is a probability on the index of the failure time
that happens to be the one in the interval of interest. Next,
we ask how it could occur that the jth failure would fall in the
selected interval and the answer is that:
1. Each of the preceding j − 1 failures would have to
occur before the start of the interval, and this occurs
with probability FT(t) for each one.
2. Each of the succeeding n − j failures must occur after
the end of the interval, and this happens with prob-
ability FT (t + ∆t) for each one.
3. The jth failure must fall within the interval, and this
happens with probability dFT(t).
4. The number of ways the n copies of the device put on
test can be separated into sets of j − 1, 1, and n − j is
given by the multinomial coefficient. Therefore
n!
( ) (f )( )
j −1 n− j
pj = FT (t) (t)dt FT (t) (7.3)
( j − 1)!(1)!( n − j)!
T
DK598X_book.fm Page 110 Monday, January 24, 2005 11:11 PM
110 Nachlas
fraction failed is
j
FˆT (t) = (7.4)
n
while the median of the distribution on the jth order statistic
implies that an appropriate estimate for the fraction failed is
j − 0.3
FˆT (t) = (7.5)
n + 0.4
Although it may seem counterintuitive, both of these
estimates are reasonable, and each has advantages. The chief
advantage of the estimator based on the mean, Expression
7.4, is that it is unbiased. That is
but rather different ways. Specifically, one may base the com-
putation of the estimate of the reliability (or failure probabil-
ity) on the number of survivors (or observed failures) or on
the proportion of the test items that have survived (or failed).
Consider first the use of the number of survivors.
The number of survivors at the end of a fixed time inter-
val is a random variable for which the dispersion is best
modeled using the binomial distribution. That is, survival of
each copy of the component may be viewed as a Bernoulli
trial, so that the quantity n − j has a binomial distribution
with success probability FT (t). In this case
E[ n − j ] = nFT (t)
and
(
Var[ n − j ] = nFT (t) 1 − FT (t) ) (7.6)
n− j
E = FT (t)
n
and
(
n − j FT (t) 1 − FT (t) )
Var = (7.7)
n n
112 Nachlas
ˆ
FT (t) + zα
ˆ
( ˆ
FT (t) 1 − FT (t) ) ≤ F (t) ≤ Fˆ (t)
T T
2 n
(7.8)
ˆ
( ˆ
FT (t) 1 − FT (t) )
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
+ z1− α
2 n
is valid asymptotically. Note that a table of standard Normal
probabilities is provided in Appendix A along with numerical
approximation strategies for Normal probabilities.
Consider an example. For the ordered data set provided
in Table 7.2, the point estimates of the failure probability and
of the reliability at t = 0.5 time units are
j 15 ˆ
FˆT (0.5) = = = 0.30 FT (0.5) = 1 − FT (0.5) = 0.70
n 50
j − 0.3 14.7 ˆ
FˆT (0.5) = = = 0.292 FT (0.5) = 1 − FT (0.5) = 0.708
n + 0.4 50.4
Taking α = 0.05, so that zα = −1.96 , the first of the
2
estimates provides the confidence interval
Γ ( η + δ + 2)
fU (u, η, δ) = u η (1 − u)δ du (7.9)
Γ ( η + 1)Γ (δ + 1)
In fact, this beta distribution form also follows directly
from the distribution on the order statistics stated in Expres-
sion 7.3. With the definition u = FT(t) and the parameter
identities of η = j−1 and δ = n−j, replacing fT(t)dt by du yields
Expression 7.9. Keep in mind that the proportion u must be
in the interval (0, 1).
The point estimate of Expression 7.4 is an appropriate
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
FU (ulower , j − 1, n − j ) = α (7.12)
2
and uupper is the solution to the equation
FU (ulower , j, n − j − 1) = 1 − α (7.13)
2
Clearly, these confidence limits must be computed
numerically, but the effort required to do this is not great.
For the same example case as the one treated above in which
n = 50 and j = 15 at t = 0.5, the point estimate for FT(0.5) is still
j 15
FˆT (0.5) = = = 0.30
n 50
DK598X_book.fm Page 114 Monday, January 24, 2005 11:11 PM
114 Nachlas
assumed form for the life distribution. They apply to any set
of test data, provided the data set is complete.
FT (τ γ ) = γ (7.14)
u = FT ( x j )
Pr[ τ γ ≥ x j ] ≥ 1 − α
(7.15b)
= 1 − B( j − 1, n, u)
where the notation B(k, n, p) represents the cumulative bino-
mial probability. Substituting this identity back into Equation
7.15a yields:
Pr[ τ γ ≥ x j ] = 1 − B( j − 1, n, FT (τ γ ))
(7.15c)
= 1 − B( j − 1, n, γ ) ≥ 1 − α
so equivalently, the tolerance bound is defined by
B( j − 1, n, γ ) ≤ α (7.15d)
Now, there are several ways in which we can use this
result.
First, consider the cases associated with j = 1. For this
case, Expressions 7.15b and 7.15d simplify to
DK598X_book.fm Page 116 Monday, January 24, 2005 11:11 PM
116 Nachlas
FU (u, 1, n − 1) = 1 − (1 − u) n = 1 − (1 − γ ) n ≥ 1 − α
so we find that for the example data set of Table 7.2, using α
= 0.05 and n = 50, we compute γ = 0.058, and we say that we
have 95% confidence that the reliability of the component
population is at least 1 − γ = 0.942 at a time of x1 = 0.006.
An alternate use of Expression 7.15d is to ask how many
items we should have tested in order for x1 to correspond to
a Type A design allowable. To answer this question, we set α
= 0.05 and γ = 0.10, and we compute n to be 29. For the
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
FT−1 ( u)
H F −1 (u) =
T ∫ 0
FT (w) dw (7.17)
=
(1 − e −0.75 λ
) = 0.467 = 0.556
λ λ
DK598X_book.fm Page 118 Monday, January 24, 2005 11:11 PM
118 Nachlas
FT−1 ( u=1.0 ) ∞
H F −1 (u = 1.0) =
T ∫0
FT (w) dw =
∫ 0
FT (w) dw = µ (7.18)
and the integral of the survivor function over the full range
of any random variable yields the mean value of that random
variable.
The second useful property of the transform is that its
derivative, when evaluated at any value of the cumulative
failure probability, equals the reciprocal of the corresponding
value of the hazard function. This is shown as follows:
FT−1 ( u)
∫
d d
H F −1 (u) = FT (w)dw
du T du 0
FT−1 ( u)
d −1
=
du
( d
FT (u) FT ( FT−1 (u)) −
)
0 FT (0) +
du
( ) ∫0
d
du
FT (w) dw
FT ( FT−1 (u))
FT−1 (u) is the time for which the cumulative failure prob-
ability equals u, so evaluating the survivor function at that
value yields 1 − u. That is,
FT ( FT−1 (u)) = 1 − u
To evaluate the derivative of the inverse function, we
proceed as follows:
DK598X_book.fm Page 119 Monday, January 24, 2005 11:11 PM
t = FT−1 (u)
so
dt d −1
= FT (u)
du du
Then
dt
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
= 1 = 1 = 1
du du d f (t)
FT (t)
dt dt
and therefore,
d −1 1
FT (u) =
du f ( FT−1 (u))
d 1− u FT (t) 1
H F −1 (u) = = = (7.19)
du T
f ( FT−1 (u)) u= F f (t) zT (t)
T (t)
H F −1 (u) H F −1 (u)
Θ F −1 (u) = T
= T
(7.20)
T H F −1 (1) µ
T
DK598X_book.fm Page 120 Monday, January 24, 2005 11:11 PM
120 Nachlas
d
H F −1 (u) 1
d du T zT ( FT−1 (u))
Θ F −1 (u) = = (7.21)
du T µ µ
Consider what the derivative of the scaled transform tells
us. Suppose the life distribution happens to be exponential.
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
1
Θ −1 (u) = λ = 1
d
du FT 1
λ
When the life distribution has constant hazard, the
scaled TTT transform has a slope equal to one. In that case,
the scaled transform plots as a straight line. Suppose the
hazard is not constant. For the Weibull distribution, the mean
is θ Γ(1 + 1/β), and the hazard function is
β −1
β t
zT (t) =
θ θ
(a) TTT
1
0.8
0.6
0.4
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
0.2
CDF
0.2 0.4 0.6 0.8 1
(b) TTT
1
0.8
0.6
0.4
0.2
CDF
0.2 0.4 0.6 0.8 1
Figure 7.1 (a) TTT transform for increasing hazard functions. (b)
TTT transform for decreasing hazard functions.
τ( x j ) = ∑x
k=1
k + (n − j) x j (7.22)
where the xj are our ordered failure times, so τ(xj) is the total
amount of testing time that is accumulated by the time of the
jth failure. This may be seen as follows:
DK598X_book.fm Page 122 Monday, January 24, 2005 11:11 PM
122 Nachlas
• • •
• • •
(xn−1, xn) 1 (xn − xn−1)
τ( x j ) = ∑ (n − k + 1)(x
k=1
k − xk−1 )
0 0 ≤ t < x1
1n x1 ≤ t < x2
FX n (t) = (7.23)
j x j ≤ t < x j +1
n
1 xn ≤ t < ∞
0 0 ≤ t < 0.006
0.02 0.006 ≤ t < 0.019
0.04 0.019 ≤ t < 0.026
0.06 0.026 ≤ t < 0.038
FX n (t) =
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
0.98 4.631 ≤ t < 5.292
1.00 5.292 ≤ t < ∞
A few values of the corresponding inverse function are
FX−1n (0.04) = 0.019 , FX−1n (0.30) = 0.465, and FX−1n (0.99) = 4.631 . For
the same data set, note that τ(x1) = 50(0.006) = 0.30, τ(x2) =
50(0.006) + 49(0.013) = 0.937, and τ(x3) = 50(0.006) + 49(0.013)
+ 48(0.007) = 1.273.
The use of the TTT transform to analyze test data pro-
ceeds by applying the transform to FX n (t) in the same manner
as for FT(t). That is,
FX−1 ( u)
H F −1 (u) =
∫
n
FX n (w) dw (7.25)
Xn 0
FX−1 ( j ) FX−1 ( 1 )
H F −1 ( j ) =
∫ ∫
n n
FX n (w) dw =
n n
FX n (w) dw
Xn n 0 0
FX−1 ( 2 ) FX−1 ( j )
∫ ∫
n n
+ FX n (w) dw + +
n n
FX n (w) dw
FX−1 ( 1 ) FX−1 ( j − 1 )
n n n n
DK598X_book.fm Page 124 Monday, January 24, 2005 11:11 PM
124 Nachlas
x1 x2
n −1 xj
n − j +1
∫ ∫ ∫
n
= dw + dw + + dw
0 n x1 n x j −1 n
1
= τ( x j )
n
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
FX−1 ( j )
lim H F −1 ( j ) = lim
∫
n n
FX n (w) dw
n →∞ Xn n n →∞ 0
j →s j →s
n n
FT ( s)
=
∫ 0
FT (w) dw = H F −1 (s)
T
1
H F −1 ( j ) τ( x j )
j n n τ( x j )
Θ F −1 = = =
Xn
( ) (7.26)
T n H F −1 (1) 1 τ( xn )
Xn τ( xn )
n
The application to sample data is implied by the above
discussion, and the result is that the quantities τ(xj) and τ(xn)
are computed, and their ratio is plotted against j/n. If the
DK598X_book.fm Page 125 Monday, January 24, 2005 11:11 PM
126 Nachlas
TTT
1
0.8
0.6
0.4
0.2
j
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
near zero and near 1.0, the entire plot lies above the 45° line.
The appropriate interpretation is that the devices that produced
the test data appear to display increasing hazard. The observed
behavior is confirmed by an inspection of the transform values,
which lie well above the corresponding values of j/n.
Both of the data analyses displayed so far used the entire
complete data set. As discussed earlier, it is often impractical
to run a test until all test units have failed. When the test is
terminated early and a censored data set is obtained, it is
still possible to use the TTT transform. The adjustments nec-
essary to do this are treated in the following section.
DK598X_book.fm Page 127 Monday, January 24, 2005 11:11 PM
TTT
1
0.8
0.6
0.4
0.2
j
0.2 0.4 0.6 0.8 1 n
128 Nachlas
τ( xr ) = ∑x
k=1
k + ( n − j ) xr (7.27)
(a) TTT
1
0.8
0.6
0.4
0.2
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
j
0.2 0.4 0.6 0.8 1 n
(b) TTT
1
0.8
0.6
0.4
0.2
j
0.2 0.4 0.6 0.8 1 n
Figure 7.4 (a) Plot of scaled TTT transform for censored data with
r = 12. (b) Plot of scaled TTT transform for censored data with r = 20.
and this quantity is plotted vs. j/r. For the example data set
of Table 7.2, the computed values for r = 12 and for r = 20
are listed in Table 7.6. The corresponding plots of the scaled
transform are shown in Figure 7.4a and Figure 7.4b, respec-
tively. These plots serve to illustrate the facts that the trans-
form may be applied to censored data and that our ability to
interpret the plots is significantly influenced by the degree of
censoring. For the case in which r = 20, we can be reasonably
confident that the hazard is constant. For the plot correspond-
ing to the data censored after r = 12 observations, our con-
clusion of constant hazard is rather more tenuous.
DK598X_book.fm Page 130 Monday, January 24, 2005 11:11 PM
130 Nachlas
(a) TTT
1
0.8
0.6
0.4
0.2
j
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
0.8
0.6
0.4
0.2
j
0.2 0.4 0.6 0.8 1 n
Figure 7.5 (a) Plot of scaled TTT transform for censored data with
r = 12. (b) Plot of scaled TTT transform for censored data with r = 20.
132 Nachlas
7.6 EXERCISES
1. The following data set was obtained from a life test
of n = 50 copies of a component.
i ti i ti i ti i ti i ti
1 551.881 11 297.883 21 122.750 31 539.933 41 141.582
2 964.448 12 78.966 22 119.677 32 175.578 42 329.841
3 687.943 13 526.061 23 568.533 33 465.506 43 570.971
4 206.215 14 558.106 24 453.852 34 208.198 44 929.433
5 844.059 15 484.969 25 267.140 35 326.713 45 67.964
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
For this data, identify the values of x1, x5, x10, and x50.
2. For the data set of Problem 1, compute the point
estimates of the reliability at x1, x5, x10, and x50 with
both the mean and the median based estimation
equations. How do these estimates compare?
3. Using the data in Problem 1, compute 95% confidence
intervals for the failure probability at each of x1, x5,
x10, and x50 using each of Expressions 7.8 and 7.11.
How do these intervals compare?
4. Using the data of Problem 1, compute a 95% tolerance
bound on the reliability at x10 and x15. Then compute
the level of confidence the data provides that the
reliability at x20 exceeds 0.55.
5. For the following data set obtained from a life test,
order the data, and then plot the TTT transform.
Indicate what form of the hazard function is sug-
gested by the plot.
DK598X_book.fm Page 133 Monday, January 24, 2005 11:11 PM
i ti i ti i ti i ti i ti
1 635.655 11 456.731 21 335.464 31 282.015 41 170.998
2 369.012 12 459.482 22 100.790 32 172.954 42 465.023
3 312.489 13 420.944 23 453.539 33 216.222 43 254.582
4 196.092 14 306.064 24 82.843 34 204.064 44 319.789
5 72.393 15 216.330 25 356.053 35 228.195 45 285.048
6 22.150 16 180.638 26 255.021 36 528.971 46 307.34
7 302.257 17 137.704 27 302.217 37 270.25 47 318.541
8 114.434 18 159.855 28 181.568 38 117.524 48 242.783
9 68.381 19 231.442 29 93.694 39 70.280 49 458.005
10 200.899 20 203.094 30 314.594 40 93.151 50 130.900
Downloaded by [Engineers Australia ] at 01:26 15 May 2014
i ti i ti i ti i ti
1 29.835 11 1048.13 21 126.097 31 154.884
2 1262.860 12 641.953 22 434.761 32 103.444
3 804.623 13 762.882 23 170.046 33 176.225
4 691.363 14 206.062 24 1880.470 34 252.424
5 654.951 15 593.040 25 1058.727 35 333.961
6 409.087 16 224.793 26 957.271 36 1989.75
7 1615.690 17 203.809 27 2.970 37 1646.63
8 470.408 18 309.879 28 75.239 38 344.135
9 918.823 19 3094.740 29 346.605 39 48.831
10 68.348 20 55.854 30 801.645 40 131.215
DK598X_book.fm Page 134 Monday, January 24, 2005 11:11 PM
134 Nachlas