Probability Models and Statistical

Methods in Reliability

Larry Leemis
Department of Mathematics
College of William and Mary
Williamsburg, VA 23187-8795
[email protected] 757-221-2034

Undergraduate Simulation, Modeling and Analysis

February 14, 2000

Outline
1. Introduction
2. Coherent Systems Analysis
3. Lifetime Distributions
4. Parametric Lifetime Models
5. Specialized Models
6. Repairable Systems
7. Lifetime Data Analysis
8. Fitting Parametric Models to Data
9. Parametric Estimation for Models with Covariates
10. Nonparametric Methods
11. Assessing Model Adequacy
1. Introduction

Motivation
• Space Shuttle Challenger accident
• Chernobyl and Three Mile Island accidents
• product liability
• customer goodwill
• corporate reputation

Three closely-related fields of study

• Actuarial science
• Biostatistics
• Reliability engineering

Terminology
The event at the end of a lifetime is called
• a failure by reliability engineers
• death by actuaries and biostatisticians
• an epoch by point process researchers
The object of a study is called
• a system, component or item by reliability
engineers
• an individual by actuaries
• an organism by biostatisticians
To avoid switching terms, failure of an item will be
used here.
1.1 A definition of reliability

Definition 1.1 The reliability of an item is the

probability that it will adequately perform its
specified purpose for a specified period of time
under specified environmental conditions.
Item
• Resolution
an item may be an interacting arrangement of
components or the component level of detail
in the model may not be of interest
• Level of detail
determine the level of detail to be modeled
• External boundary for the item
what is to be considered part of the item and
what is to be considered part of the
environment around the item

Probability
• Range
all reliabilities must be between 0 and 1
inclusive
• Spinoffs from the probability axioms
statistical independence
Adequate performance
• Must be stated unambiguously
• Standards
Example: a ball bearing has failed when its
diameter falls outside of 3 + 0.05 mm
• Binary models
the item is in either the functioning or failed
state (e.g., a fuse)

Purpose
• Intended use
Example: a drill may have one grade for a
handyman and another for a contractor

Time
• Units
must be specified (e.g., hours, years)
• Notation
many lifetime models use the random
variable T
• Time need not be taken literally
consider an automobile tire, light switch
• Time duration must be specified
Example: 1000 hour reliability is 0.8
• Continuous operation vs. on/off cycling
time alone may not be the only consideration
(e.g., motors, computers)
Environmental conditions
• Factors
temperature, humidity, and turning speed all
affect the lifetime of a machine tool
• Preventive maintenance
usually effective in prolonging the lifetime of
the item and hence increasing the reliability

Reliability vs. quality

• reliability incorporates the passage of time
• quality is a static descriptor of an item
Example 1: Two transistors of equal quality. One
used in a television set, the other in a cannon
launch environment. Identical quality, different
reliabilities.
Example 2: Two automobile tires, each of high
quality. One was produced in 1957, the other in
1994. Same purpose, different reliabilities due
to improved design (e.g., tread or steel belts),
components (e.g., rubber) or processes (e.g.,
manufacturing advances). Some quality
improvements (e.g., improved tread design)
improve the reliability of the tire, while others
(e.g., improved white wall design) will not.
1.2 Case study

Item under consideration: the O-rings on the solid

rocket motors on the Space Shuttle

Subsystems
• orbiter
• external liquid-fuel tank
• two solid rocket motors

Each assembled solid rocket motor contains three

field joints that must be sealed.

O-rings
• 37.5 feet in diameter
• 0.28 inches thick
• all six O-rings must operate to avoid having the
propellant escape causing potential failure, so
the O-rings form a six-component series system

Figure 1.1 A six-component series

arrangement of O-rings.
Redundancy: a technique to increase reliability
• Redundancy is highly effective if the components
are independent.
• In 1977, NASA discovered field joint rotation
indicating that the failure of the primary and
secondary O-rings may not be independent.
• Prior to the Challenger accident, the solid rocket
motors were recovered in 23 of the 24 shuttle
flights.
• There was concern that an environmental
variable, temperature at launch, might influence
the reliability of the field joints.
• There was a forecast of 31o F for the morning of
the launch of the Challenger, the coldest launch
temperature to date.

Figure 1.2 A 12-component

arrangement of O-rings.
 Failures
3.0 •

2.5

2.0 •

1.5

1.0 •• • •

0.5

0.0 • ••• • • •• •• • •• •• •
55 60 65 70 75 80 85
Temperature
Figure 1.3 Launch temperature versus
number of field joint failures.

Conclusion: temperature was indeed significant

2. Coherent Systems Analysis

• assume that an item (system) consists of n

 components
• two key modeling decisions
which elements of the system are included
the level of detail
• the first two sections: structural properties
• the next two sections: probabilistic properties
• outline
structure functions
minimal path and cut sets
reliability functions
reliability bounds

2.1 Structure functions

Definition 2.1 The state of component i, x i , is
 0 if component i has failed
xi = 
 1 if component i is functioning
for i = 1, 2, . . . , n.
The binary model
• n components form a system
• system state vector, x = (x 1 , x 2 , . . . , x n )
• the system state vector can assume 2n different
values
 n
• of these vectors correspond to exactly j
 j
functioning components, j = 0, 1, . . . , n
• the structure function, φ (x), maps the system
state vector x to 0 or 1, the system state
Definition 2.2 The structure function φ is
0 if the system has failed under x
φ (x) = 
1 if the system is functioning under x.

Example 2.1 A series system functions when all

of its components function. Thus φ (x) assumes
the value 1 when x 1 = x 2 = . . . = x n = 1, and 0
otherwise.
 0 if there exists an i such that x i = 0
φ (x) = 
 1 if x i = 1 for all i = 1, 2, . . . , n
= min {x 1 , x 2 , . . . , x n }
n
= Π xi .
i=1

1 2 3 ... n
Figure 2.1 A series system block diagram.

Example 2.2 A parallel system functions when

one or more of the components function. Thus
φ (x) assumes the value 0 when
x 1 = x 2 = . . . = x n = 0, and 1 otherwise.
  0 if x i = 0 for all i = 1, 2, . . . , n
φ (x) = 
 1 if there exists an i such that x i = 1
= max {x 1 , x 2 , . . . , x n }
n
=1− Π (1 − x i ).
i=1

3
..
.
n
Figure 2.2 A parallel system block diagram.

Applications
• kidneys
• brake system on an automobile with two brake
fluid reservoirs
Series and parallel systems are special cases of k-out-
of-n systems, where the system functions if k or more
of the n components function.
Applications
• suspension bridge (components: cables)
• an automobile engine (components: cylinders)
• a bicycle wheel (components: spokes)
Example 2.3 The structure function for a k-out-
of-n system is
 n
 0 if i Σ
=1
xi < k
φ (x) =  n
 1 if Σ x i ≥ k.
 i=1
The block diagram for a k-out-of-n system is
difficult to draw in general.

1 2

1 3

2 3
Figure 2.3 A 2-out-of-3 system
block diagram.
φ (x) = 1 − (1 − x 1 x 2 )(1 − x 1 x 3 )(1 − x 2 x 3 )
2.3 Reliability functions

Assumptions
• the binary model applies to components and
systems
• the n components must be nonrepairable
• the components are independent
Definition 2.9 The random variable denoting
the state of component i, X i , is
 0 if component i has failed
Xi = 
 1 if component i is functioning
for i = 1, 2, . . . , n.
Random component states
• these n values can be written as a random system
state vector X
• pi = P[X i = 1] is the reliability of the i th
component, i = 1, 2, . . . , n
• reliability vector p = ( p1 , p2 , . . . , p n )
• must specify the time to which the reliability
applies (e.g., 5000-hour reliability is 0.83)
• the system reliability, r, is defined by
r(p) = P[φ (X) = 1]
• r( p) used when all component reliabilities are
equal
Technique 1: Definition of r(p)
Example 2.12 Series system of n independent
components
n n n
r(p) = P[φ (X) = 1] = P[ Π X i = 1] = Π P[X i = 1] = Π pi
i=1 i=1 i=1
"Weakest link" for series systems
• system reliability less than smallest component
reliability
• improvement of weakest component most
effective
Special case: identical components r( p) = p n

r(p)
1.0
0.8
0.6 n=1
0.4
n=2
0.2 n=5

0.0 n = 10
p
0.0 0.2 0.4 0.6 0.8 1.0
Figure 2.12 Reliability of n-component series systems.
Technique 2: Expected value of φ (X)
P[φ (X) = 1] = E[φ (X)]
since φ (X) is a Bernoulli random variable.
Example 2.13 Parallel system of n independent
components.
n
r(p) = E[φ (X)] = E[1 − Π (1 − X i )]
i=1
n n
=1− Π E[1 − X i ] = 1 − Π (1 − pi )
i=1 i=1

Special case: identical components

r( p) = 1 − (1 − p)n

r(p)
1.0
n = 10
0.8 n=5
n=2
0.6
n=1
0.4
0.2
0.0 p
0.0 0.2 0.4 0.6 0.8 1.0
Figure 2.13 Reliability of n-component parallel systems.
"Law of diminishing returns" for parallel systems
• marginal gain in reliability decreases
dramatically as more components are added
• improvement of the strongest component is the
most effective
Notes on parallel systems
• standby system
• shared-parallel system
3. Lifetime Distributions

Motivation
Up to this point, reliability has only been considered
at one particular instance of time.
Outline
• lifetime distribution representations
• discrete distributions
• moments and fractiles
• system lifetime distributions
• distribution classes

3.1 Distribution representations

Five functions that define the distribution of T
• survivor function
• probability density function
• hazard function
 • cumulative hazard function
• mean residual lifetime function
Survivor function (reliability function)
S(t) = P[T ≥ t] t≥0
All survivor functions satisfy three conditions
S(0) = 1 lim S(t) = 0 S(t) is nonincreasing
t →∞
Interpretations
• S(t) is the probability that an individual item is
functioning at time t
• S(t) is the expected fraction of items surviving to
time t

S(t)
1.0
0.8
0.6 S1(t)
S2(t)
0.4
0.2
0.0 t
0.0 0.5 1.0 1.5 2.0
Figure 3.1 Two survivor functions.

Conditional survivor functions

P[T ≥ t and T ≥ a] P[T ≥ t] S(t)
S T | T ≥ a (t) = = =
P[T ≥ a] P[T ≥ a] S(a)
for all t ≥ a.
Probability density function
f (t) = − S′(t)
f (t)∆t = P[t ≤ T ≤ t + ∆t]
for small ∆t values.
b
P[a ≤ T ≤ b] = ∫ f (t) dt = S(a) − S(b)
a
All probability density functions satisfy
∞
∫ f (t) dt = 1 f (t) ≥ 0 for all t ≥ 0
0

f(t)
0.8

0.6

0.4
F(t0 )
0.2
S(t0 )
0.0 t0
t
0.0 0.5 1.0 1.5 2.0
Figure 3.3 The relationship between survivor and
cumulative distribution functions.
Hazard function (failure rate, force of mortality)
h(t) = f (t) / S(t) t≥0
h(t)∆t = P[t ≤ T ≤ t + ∆t | T ≥ t]
for small ∆t values. Units: failures per unit time.
Interpretations
• h(t) is the amount of risk an item is under at t
• h(t) is a special case of the intensity function for
a nonhomogeneous Poisson process
All hazard functions must satisfy
∞
∫ h(t) dt = ∞ h(t) ≥ 0 for all t ≥ 0
0

h(t)
5 IFR
4 DFR
3
2 BT
1
0 t
0.0 0.5 1.0 1.5 2.0
Figure 3.5 Common hazard function shapes.
Cumulative hazard function (integrated
hazard function and the renewal function)
t
H(t) = ∫ h(τ )dτ t≥0
0
All cumulative hazard functions satisfy
H(0) = 0 lim H(t) = ∞ H(t) is nondecreasing
t →∞
Applications
• variate generation in Monte Carlo simulation
• implementing certain procedures in statistical
inference
• defining certain distribution classes

Mean residual life function

1 ∞
L(t) = E[T − t | T ≥ t] = ∫
S(t) t
τ f (τ )d τ − t t≥0

All mean residual life functions satisfy

∞ dt
L(t) ≥ 0 L′(t) ≥ − 1 ∫ =∞
0
L(t)
 Example 3.2 Consider the exponential
distribution defined by the survivor function
S(t) = e− λ t t≥0
with positive scale parameter λ .

f (t) = λ e− λ t t≥0
The mean residual life function is
∞ 1
L(t) = e λt
∫ τ λ e− λτ d τ − t = t≥0
t
λ
by using integration by parts.

Knowing one of the five lifetime distribution

representations implies knowledge of the other four.
If the survivor function is known, for example, the
cumulative hazard function can be determined by
t t
f (τ )
H(t) = ∫ h(τ )d τ = ∫ S(τ )
d τ = − log S(t)
0 0
3.3 Moments and fractiles
Motivation
Moments and fractiles contain less information than a
lifetime distribution representation, but they are often
useful ways to summarize the distribution of a
random lifetime.
Examples
• the mean time to failure, E(T )
• the median, t 0.50
• the 95th percentile of a distribution, t 0.95

Assumption: random lifetime T is continuous

∞
E[u(T )] = ∫ u(t) f (t) dt
0
Mean (abbreviated by MTTF or MTBF)
∞ ∞
µ = E[T ] = ∫t f (t) dt = ∫ S(t) dt
0 0
Variance
σ 2 = V [T ] = E[(T − µ )2 ] = E[T 2 ] − (E[T ])2
Coefficient of variation
σ
γ =
µ
Skewness
 T − µ 3 
γ 3 = E
 σ  
 
Kurtosis
 T − µ 4 
γ 4 = E
 σ  
 
Fractiles: t p satisfies
F(t p ) = P[T ≤ t p ] = p or t p = F −1 ( p )
Example 3.5 The exponential distribution has
survivor function
S(t) = e− λ t t≥0
∞ ∞ 1
µ = E[T ] = ∫ S(t) dt = ∫ e− λ t dt =
0 0
λ
∞ ∞ 2
E[T ] =2
∫ t f (t) dt =
2
∫ t 2 λ e− λ t dt =
0 0
λ2
2 1 1
σ = E[T ] − (E[T ]) =
2 2 2
− =
λ2 λ2 λ2
γ 3 = λ 3 6 λ − 3 − 6 λ − 3 + 3λ − 3 − λ − 3  = 2
 
1
tp = − log (1 − p)
λ
4. Parametric Lifetime Models
Motivation
Survival patterns of a drill bit, a fuse, and an
automobile are vastly different.
Outline
• parameters
• exponential
• Weibull
• gamma
• other distributions

4.1 Parameters
Three types of parameters:
• location
• scale
• shape

Location (or shift) parameters

Shift a distribution along the time axis. If c 1
and c 2 are two values of a location parameter
for a lifetime distribution with survivor function
S(t; c), then there exists a constant α such that
S(t; c 1 ) = S(α + t; c 2 ).
Example Mean µ in the normal
distribution.
 Scale parameters
Used to expand or contract the time axis by a
factor of α . If λ 1 and λ 2 are two values for a
scale parameter for a lifetime distribution with
survivor function S(t; λ ), then there exists a
constant α such that S(α t; λ 1 ) = S(t; λ 2 ).
Example Exponential scale parameter λ .
Shape parameters
Affect the shape of the probability density
function.
Example Weibull shape parameter κ .

S(t)
1.0 O
O
O
O
O
OO
O
O
O

0.8 O

0.6
0.4
0.2
0.0 t
0 50 100 200 300
Figure 4.2 A mixed discrete-continuous
survivor function.
4.2 The exponential distribution
Motivation
The exponential distribution plays a central role in
reliability modeling since it is the only continuous
distribution with a constant hazard function.
f(t) S(t)
2.0
λ=2 0.8
1.0 λ=1
0.4
λ=1 λ=2
0.0 t 0.0 t
0.0 1.0 2.0 0.0 1.0 2.0
h(t) H(t)
2.0 4
λ=2
λ=2
1.0 2
λ=1 λ=1
0.0 t 0 t
0.0 1.0 2.0 0.0 1.0 2.0
L(t)
0.8 λ=1
0.4 λ=2
0.0 t
0.0 1.0 2.0

Figure 4.3 Lifetime distribution representations

for the exponential distribution.
 S(t) = e− λ t f (t) = λ e− λ t h(t) = λ
1
H(t) = λ t L(t) =
λ
Property 4.1 (Memoryless property) If
T ∼ exponential(λ ) then
P[T ≥ t] = P[T ≥ t + s | T ≥ s] t ≥ 0; s ≥ 0.

S(t)
1.0
0.8
0.6
0.4
0.2
0.0 t
0.0 0.5 1.0 1.5 2.0
Figure 4.4 The memoryless property
of the exponential distribution.
Property 4.2 The exponential distribution is
the only continuous distribution with the
memoryless property.

4.3 The Weibull distribution

Motivation
The exponential distribution’s constant failure rate
is often too restrictive or inappropriate.
− (λ t)κ κ κ − 1 − ( λ t)κ
S(t) = e f (t) = κ λ t e
h(t) = κ λ κ tκ − 1 H(t) = (λ t)κ
for all t ≥ 0.
Notes
• λ is a positive scale parameter
• κ is a positive shape parameter
• exponential distribution is a special case
(κ = 1)
• hazard function increases from 0 when κ > 1
(IFR)
• hazard function decreases from ∞ to 0 when
κ < 1 (DFR)
• κ = 2 known as the Rayleigh distribution
• when 3 < κ < 4 the probability density
function resembles that of a normal random
variable
• the mode and median of the distribution are
equal when κ ≈ 3. 26
• the characteristic life is a special fractile
1
defined by t c = ; all Weibull survivor
λ
1
functions pass through the point ( , e−1 )
λ
• since H(t) = − log S(t), all Weibull cumulative
hazard functions pass through the point
1
( , 1)
λ
• if T has the Weibull distribution, then
Y = log T has the extreme value distribution
• self-reproducing property: if
T i ∼ Weibull(λ i , κ ) for i = 1, 2, . . . , n, then
n
min {T 1 , T 2 , . . . , T n } ∼ Weibull( Σ λ i , κ )
i=1
• moments
1  1 1 1
µ = Γ 1+ = Γ
λ  κ  λκ  κ 
 2 
1 2 2 1 1  
σ = 2 Γ
2
− Γ  
λ κ κ  κ κ   
 
1/2
 2 
2 2 1 1  
 Γ  −  Γ   
κ κ κ κ  
γ =
 
1 1
Γ
κ κ 
 f(t) S(t)
κ = 0.5 1.0
1.2 κ=3
0.8
κ=2
0.8 0.6
κ=1 0.4 κ = 0.5
0.4 κ=1
0.2 κ=2
0.0 κ=3
t 0.0 t
0.0 1.0 0.0 1.0
h(t) H(t)
κ=3
6 κ=3 3
κ=2
4 κ = 0.5 2
κ=2 κ=1
2 1 κ = 0.5
κ=1
0 t
0 t
0.0 1.0 0.0 1.0
L(t)
κ = 0.5
4
3
2
1 κ=1
κ=2
0 κ=3
t
0.0 1.0

Figure 4.7 Lifetime distribution representations

for the Weibull distribution.
4.5 Other lifetime distributions
Table 4.4 Distribution classes.

Distribution IFR DFR BT UBT

Exponential YES YES NO NO
Muth YES NO NO NO
Weibull YES YES NO NO
Gamma YES YES NO NO
Uniform YES NO NO NO
Log normal NO NO NO YES
Log logistic NO YES NO YES
Inv. Gaussian NO NO NO YES
Expon. power YES NO YES NO
Pareto NO YES NO NO
Gompertz YES NO NO NO
Makeham YES NO NO NO
IDB YES YES YES NO
Gen. Pareto YES YES NO NO
5. Specialized Models
Motivation
There are several ways to combine and extend the
continuous lifetime models previously outlined.
Outline
• competing risks
• mixtures
• accelerated life
• proportional hazards

5.1 Competing risks

Notes
• causes of failure may be grouped into k
classes
• an item is subject to k competing risks (or
causes) C 1 , C 2 , . . . , C k
• can be thought of as a series system of
components
• origins of competing risks theory traced to a
study by Daniel Bernoulli in the 1700’s
concerning the impact of eliminating
smallpox
• a second and equally appealing use of
competing risks models is that they can be
used to combine component distributions to
 form more complicated models

h(t)
1.5
h T (t)
1.0
h X1(t)
0.5
h X2(t)
0.0 t
0.0 0.5 1.0 1.5 2.0

Figure 5.1 Hazard functions

for a competing risks model.
• notation (for j = 1, 2, . . . , k)
T lifetime
k number of risks
Xj net life for risk j
Yj crude life for risk j
πj probability risk j causes failure
• net lifetimes: causes C 1 , . . . , C k are viewed
individually
• crude lifetimes: lifetimes are considered in the
presence of all other risks
5.2 Mixtures
Mixture models are appropriate when items ar
drawn from one of several populations (finite
mixtures) or can be differentiated by a continuous
parameter.
Finite mixtures
m
f (t) = Σ pl f l (t | θ l )
l=1
m
where Σ pl = 1, pl ≥ 0 for l = 1, 2, . . . , m.
l=1

Continuous mixtures (stochastic parameters)

f (t) = ∫ f (t|θ ) p(θ )dθ
allθ
where θ is called the mix parameter and p(θ )
indicates the distribution of the mix parameter.
Example 5.4 If m = 2 facilities produce
items with exponential(1) and exponential(2)
lifetimes, respectively, and 1 / 3 of the items
come from facility 1 and 2 / 3 come from
facility 2, the probability density function of
the time to failure of an item whose
manufacturing site is unknown is
f (t) = p1 f 1 (t | λ 1 ) + p2 f 2 (t | λ 2 )
 1 − t 4 − 2t
=
e + e t≥0
3 3
which is a finite mixture of the two
populations. This model is a special case of
the hyperexponential distribution.

Combining competing risks and finite mixtures

 k − ∫ Σ h lj (τ )d τ 
t
k l

m
pl  Σ h lj (t) e 0 j = 1 
l
f (t) = Σ  j=1 
l=1
 
m
where m is the number of populations, Σ pl = 1,
l=1
k l is the number of risks acting within the l th
population, h lj (t) is the hazard function for the j th
risk within the l th population.

Application: casualty insurance

• m = 3 populations of dwellings
single family dwellings
condominiums
apartments
• k 1 = k 2 = k 3 = 5 risks
fire
flood
tornado
earthquake
burglary
5.3 Accelerated life
The accelerated life and proportional hazards
models are appropriate for including a vector of
covariates in a lifetime model.
The q × 1 vector z = (z 1 , z 2 , . . . , z q )′ contains q
covariates associated with a particular item.
Example Reliability
T : drill bit failure time
z 1 : turning speed
z 2 : feed rate
z 3 : hardness of the material

Example Biostatistics
T : patient survival time
z 1 : age
z 2 : gender
z 3 : cholesterol level

Example Recidivism
T : time to return to prison
z 1 : age
z 2 : time served
z 3 : number of previous convictions
Notation
z = (z 1 , z 2 , . . . , z q )′ covariates
β = ( β 1 , β 2 , . . . , β q )′ regression coefficients
ψ (z) link function
S 0 (t), f 0 (t), h0 (t), H 0 (t) baseline functions

How to link covariates to a lifetime distribution

• one lifetime model when z = 0 (often called
the baseline model)
• other models when z ≠ 0
The accelerated life model
S(t) = S 0 (t ψ (z)) t≥0
Notes
• S 0 is a baseline survivor function
• ψ (z) is a link function satisfying ψ (0) = 1 and
ψ (z) > 0 for all z
• a popular link function choice is the log-linear
form ψ (z) = e β ′z
• the covariates accelerate the rate at which the
item moves through time with respect to the
baseline case when ψ (z) > 1
• the covariates decelerate the rate at which the
item moves through time with respect to the
baseline case when ψ (z) < 1
• application: situations when testing items at
their operating environments is too time
 consuming

5.4 Proportional hazards

Whereas accelerated life models modify the rate
that the item moves through time based on the
values of the covariates, proportional hazards
models modify the hazard function by the factor
ψ (z).

The proportional hazards model can be defined by

h(t) = ψ (z) h0 (t).
Notes
• the covariates increase the risk when ψ (z) > 1
• the covariates decrease the risk when ψ (z) < 1
• the log-linear form ψ (z) = e β ′z is still an
appropriate choice for the link function
Table 5.1 Lifetime distribution representations
for regression models.
 Accelerated Life Proportional Hazards
S(t) S 0 (tψ (z)) [S 0 (t)]ψ (z)
f (t) ψ (z) f 0 (tψ (z)) f 0 (t)ψ (z)[S 0 (t)]ψ (z) − 1
h(t) ψ (z)h 0 (tψ (z)) ψ (z)h 0 (t)
H(t) H 0 (tψ (z)) ψ (z) H 0 (t)

Example 5.11 Consider the baseline hazard

function (Cox and Oakes, 1984)
1 0≤t<1
h0 (t) = 
t t≥1
Assumptions
• single binary covariate z
• when z = 0 (the control case), ψ (z) = 1
• when z = 1 (the treatment case), ψ (z) = 2
 h(t)
4 h AL(t) h PH(t)

2 h 0(t)

0 t
0.0 0.5 1.0 1.5 2.0

Figure 5.3 Hazard functions for a piecewise-continuous

baseline hazard function.
h PH (t) = ψ (z) h0 (t) t≥0
h AL (t) = ψ (z)h0 (tψ (z)) t≥0

6. Repairable Systems
Motivation
So far, only nonrepairable systems of components
have been considered. Most systems are
repairable.
Outline
• Introduction
• Point processes
• Availability
 • Birth-death processes

6.1 Introduction
A repairable item may be returned to an operating
condition after failure to perform a required
function by any method other than replacement of
the entire item.
Replacement models
• used when a nonrepairable item is replaced
with another item upon failure
• "socket models"
• unlimited spares
• redundancy allocation problem (optimal
number of spares)
• replacement policies
X X X X X X X t
0

Figure 6.1 Failure replacement policy.

X X O X O X X O X t
0 c c c

Figure 6.2 Age replacement policy.

 X XO X O X XO O X O X t
c c c c c
0
Figure 6.3 Block replacement policy.
• choice between these three replacement
policies depends on the lifetime distribution,
the cost of failure, administrative costs, etc.
• age and block replacement policies collapse to
a failure replacement policy as c → ∞
• expected number of items consumed (c fixed)
n f (t) ≤ n a (t) ≤ n b (t) t>0
6.2 Point processes
Hazard vs. intensity functions

h(t) h(t)
4 4
3 3
2 2
1 1
0 X t 0 X t
0.0 1.0 2.0 0.0 1.0 2.0

Figure 6.4 Hazard functions for an

item with a DFR and IFR distribution.

λ(t ) λ(t )
4 4
3 3
2 2
1 1
0 X XX X X X t 0 X XX XX XXXX t
0.0 1.0 2.0 0.0 1.0 2.0

Figure 6.5 Intensity functions for an improving

item and a deteriorating item.
 Table 6.1 Terminology for nonrepairable
and repairable items.

Status Nonrepairable Repairable

Gets better burn-in h′(t) ≤ 0 improving λ ′(t) ≤ 0
Gets worse wear out h′(t) ≥ 0 deteriorating λ ′(t) ≥ 0

Point process models

• Poisson processes
• renewal processes
• nonhomogeneous Poisson processes

Notation and assumptions

• failures occur at times T 1 , T 2 , . . .
• the time to replace or repair an item is
negligible
• the origin is defined to be T 0 = 0
• the times between the failures are X 1 , X 2 , . . .
• T k = X 1 + X 2 + . . . + X k , for k = 1, 2, . . .
• the counting function N (t) is the number of
failures that occur in (0, t]
N (t) = max { k | T k ≤ t}
for t > 0
• {N (t), t > 0} is often called a "counting
process"
* if t 1 < t 2 then N (t 1 ) ≤ N (t 2 )
 * if t 1 < t 2 then N (t 2 ) − N (t 1 ) is the number
of failures in the interval (t 1 , t 2 ]
• Λ(t) = E[N (t)] is the expected number of
failures that occur in the interval (0, t]
• λ (t) = Λ′(t) is the rate of occurrence of
failures
4 O

3 O O

2 Λ( t ) O O N(t)

1 O O

X1 X2 X3 X4
0 O
X X X X
t
T1 T2 T3 T4
0.0 0.5 1.0 1.5 2.0

Figure 6.6 A point process realization.

Two important properties
• independent increments: the number of
failures in mutually exclusive intervals are
independent

0 ( X X ] X ( ] X t
t1 t2 t3 t4
 Figure 6.7 Independent increments.
• stationarity: the distribution of the number of
failures in any time interval depends only on
the length of the time interval
Homogeneous Poisson process (HPP)
Definition 6.1 A counting process is a
Poisson process with parameter λ > 0 if
• N (0) = 0
• the process has independent increments
• the number of failures in any interval of
length t has the Poisson distribution
with parameter λ t.
Implications
• the distribution of the number of events in
(t 1 , t 2 ] has the Poisson distribution with
parameter λ (t 2 − t 1 ).
[λ (t 2 − t 1 )] x e− λ (t 2 − t 1 )
• P[N (t 2 ) − N (t 1 ) = x] =
x!
for x = 0, 1, 2, . . .
• N (t) has the Poisson distribution with mean
Λ(t) = E[N (t)] = λ t, where λ is often called
the rate of occurrence of failures
• the intensity function is λ (t) = Λ′(t) = λ
• if X 1 , X 2 , . . . are independent and identically
distributed exponential random variables,
 then N (t) corresponds to a Poisson process
• this model is sometimes called just a Poisson
process
Nonhomogeneous Poisson process (NHPP)
Four reasons to consider an NHPP
• the HPP is a special case of an NHPP
(stationarity assumption relaxed)
• the probabilistic model for an NHPP is
mathematically tractable
• the statistical methods for an NHPP are also
mathematically tractable
• the NHPP is capable of modeling improving
and deteriorating systems
Intensity function: λ (t)
t
Cumulative intensity function: Λ(t) = ∫ λ (τ )dτ
0

Definition 6.4 A counting process is a

nonhomogeneous Poisson process with
intensity function λ (t) ≥ 0 if
• N (0) = 0
• the process has independent increments
• the probability of exactly n events
occurring in the interval (a, b] is given
by
 b
b − ∫ λ (t)dt
[ ∫ λ (t)dt] en a

P[N (b) − N (a) = n] =

a
n!
for n = 0, 1, . . . .

6.3 Availability
Notation
• X i denotes the i th time to failure, i = 1, 2, . . .
• Ri denotes the i th time to repair, i = 1, 2, . . .

X O X O X O t
X1 R1 X2 R2 X3 R3
0

Figure 6.10 Failure and repair process

realization.
 Time to detect failure
Time to diagnose the problem
Time to obtain parts and labor
Repair time
Testing time

X O t
Ri

Figure 6.11 Partitioning the repair time.

7. Lifetime Data Analysis

Motivation
Parameters have been assumed to be known
constants. The rest of the tutorial considers
parameter estimation (e.g., component reliability,
distribution parameter values).
Outline
• point estimation
• interval estimators
• likelihood function
• asymptotic properties of the likelihood
function
• censoring

7.1 Point estimation

A point estimator is a statistic used to estimate a
population parameter.
Definition 7.1 The point estimator θˆ is an
unbiased estimator of θ if and only if
E[θˆ ] = θ .

Definition 7.2 Let θˆ1 and θˆ2 be two unbiased

point estimators of the parameter θ . Then
V (θˆ1 )
V (θˆ2 )
is the efficiency of θˆ1 relative to θˆ2 .

7.2 Interval estimation

Confidence intervals give bounds that contain a
population parameter with a prescribed probability
L ≤θ ≤U
Notes
• L and U are functions of
the sample size n
the lifetimes t 1 , . . . , t n
the nominal coverage of the interval 1 − α
• true value of the parameter θ is denoted by θ 0
• popular choices for α are 0.10 and 0.05
• confidence intervals: exact, approximate,
asymptotically exact
 ( O ) θ
( O ) θ
( O ) θ
( O ) θ
( O ) θ
( O ) θ
( O ) θ
( O ) θ
( O ) θ
( O ) θ
θ0
θ
0.0 0.5 1.0 1.5 2.0 2.5
Figure 7.2 Ten 90% confidence intervals for θ (n = 25).
7.3 Likelihood theory
Notation
• t 1 , t 2 , . . . , t n is a set of random lifetimes
• θ = (θ 1 , θ 2 , . . . , θ p )′ is a vector of unknown
parameters
• L(t, θ ) is the likelihood function
n
L(t, θ ) = Π f (t i , θ )
i=1
• log L(t, θ ) is the log likelihood function
n
log L(t, θ ) = Σ log f (t i , θ )
i=1
• θˆ = (θˆ1 , θˆ2 , . . . , θˆ p )′ is the maximum
likelihood estimator

f(t)
1.0
0.8
0.6
0.4
0.2
0.0 X X X X
t
0.0 0.5 1.0 1.5 2.0

Figure 7.7 Maximum likelihood estimation.

• the i th element of the score vector is
 ∂ log L(t, θ )
U i (θ ) = i = 1, 2, . . . , p
∂θi
• the score vector components have expectation
E [U i (θ ) ] = 0 i = 1, 2, . . . , p
and variance-covariance matrix
I (θ ) = E[U (θ ) U′ (θ )]
• this variance-covariance matrix is called the
Fisher information matrix with components

 − ∂2 log L(t, θ ) 
Cov(U i (θ ), U j (θ )) = E  
 ∂ θ i ∂ θ j 
for i = 1, 2, . . . , p and j = 1, 2, . . . , p.
• the observed information matrix has
components O(θˆ ) is
 − ∂2 log L(t, θ )  i = 1, 2, . . . , p
 
 ∂ θ i ∂ θ j θ = θˆ j = 1, 2, . . . , p
Example 7.7 Collect t 1 , t 2 , . . . , t n from an
exponential population with a single
parameter θ
1 − t /θ
f (t; θ ) = e t>0
θ
The likelihood function is
n
L (t, θ ) = Π f (t i , θ )
i=1
n 1
= Π e− t i / θ
i=1 θ
n

−n
− Σ ti / θ
=θ e i=1

The log likelihood function is

n
log L(t, θ ) = − n log θ − Σ ti / θ
i=1
The score vector is
n

n iΣ
ti
∂ log L(t, θ ) =1
U(θ ) = =− +
∂θ θ θ2
The maximum likelihood estimator is
1 n
θˆ = Σ ti
n i=1
The derivative of the score vector is
n

∂ log L(t, θ )
2
n
2 Σ ti
i=1
= −
∂θ 2 θ 2 θ3
The information matrix is
 − ∂2 log L(t, θ ) 
I (θ ) = E  
 ∂ θ 2

 n

 n 2 Σ i t
i=1 
= E − 2 + 
 θ θ 3

 
n 2  n 
= − 2 + 3 E  Σ ti 
θ θ  i=1 
n
= 2
θ
The observed information matrix is
 − ∂2 log L(t, θ ) 
O(θˆ ) =  
 ∂ θ 2
θ = θˆ
n
= 2
θˆ
7.5 Censoring
A censored observation occurs when only a bound
is known on the time of failure.
Notation
• n: number of items on test
• r: number of observed failures
• c: censoring time
A data set where all failure times are known is
called a complete data set.
X
X
X
X
X

Figure 7.9 A complete data set with n = 5.

A data set containing one or more censored
observations is called a censored data set. The
most common type of censoring is right censoring.
X
X
O
O
X

Figure 7.10 A single Type II right-censored

data set with n = 5 and r = 3.
In single Type II censoring, the time to complete
the test is random. The second special case is
single Type I or time censoring.
X
X
X
O
X

Figure 7.11 A single Type I right-censored

data set with n = 5 and r = 4.
In single Type I censoring, the number of failures
is random.
Random censoring occurs when individual items
are withdrawn from the test at any time during the
study. It is usually assumed that the i th lifetime t i
and the i th censoring time c i are independent
random variables.
X
O
O
X
O

Figure 7.12 A randomly right-censored data set

with n = 5 and r = 2.

8. Fitting Parametric Models to Data

Motivation
Find point and interval estimators for the
exponential and Weibull distributions for sample
data sets.
Outline
• sample data sets
• exponential distribution
• Weibull distribution

8.1 Sample data sets

Example 8.1 A complete data set of n = 23

ball bearing failure times to test the
endurance of deep groove ball bearings (in
106 revolutions)
17.88 28.92 33.00 41.52 42.12 45.60
48.48 51.84 51.96 54.12 55.56 67.80
68.64 68.64 68.88 84.12 93.12 98.64
105.12 105.84 127.92 128.04 173.40

Example 8.2 A Type II right-censored data

set of n = 15 automotive a/c switches with
r = 5. The observed failure times measured
in number of cycles are
1410 1872 3138 4218 6971
Example 8.3 Determine the effect of 6-MP
(6-mercaptopurine) on leukemia remission
times. A sample of n = 21 patients were
treated with 6-MP, and r = 9 remission times
were observed. The survival times (in
weeks) are
6 6 6 6* 7 9* 10 10* 11* 13 16 17*
19* 20* 22 23 25* 32* 32* 34* 35*
Control group: 21 other leukemia patients
1 1 2 2 3 4 4 5 5 8 8
8 8 11 11 12 12 15 17 22 23

Example 8.4 Forty motorettes were placed

on test at 150o C, 170o C, 190o C and 220o C
(ten motorettes at each temperature level and
Type I censoring). The failure times (in
hours) are
150o C: 8064* 8064* 8064* 8064*
8064* 8064* 8064* 8064* 8064* 8064*
170o C: 1764 2772 3444 3542 3780
4860 5196 5448* 5448* 5448*
190o C: 408 408 1344 1344 1440 1680*
1680* 1680* 1680* 1680*
220o C: 408 408 504 504 504 528*
528* 528* 528* 528*
Failure times are the midpoint of an
inspection period. Operating temperature:
130o C.
8.2 The exponential distribution
Goal: find point and interval estimates for the
p = 1 parameter λ .
The exponential distribution can be parameterized
by either its failure rate λ or its mean
µ = θ = 1 / λ.
S(t, λ ) = e− λ t f (t, λ ) = λ e− λ t
h(t, λ ) = λ H(t, λ ) = λ t
for all t ≥ 0.
Complete data sets
A complete data set consists of failure times
t1, t2, . . . , t n .
n
L(λ ) = Π f (t i , λ )
i=1
The log likelihood function is
n
log L(λ ) = Σ [log h(t i , λ ) − H(t i , λ )]
i=1
or
n n
log L(λ ) = Σ [log λ − λ t i ] = n log λ − λ Σ ti
i=1 i=1
The single element score vector is
∂ log L(λ ) n n
U(λ ) = = − Σ ti
∂λ λ i=1
n
The MLE is λˆ = n
Σ ti
i=1
Exact confidence intervals for λ
λˆ χ 2n,
2
1−α /2 λˆ χ 2n,
2
α /2
<λ<
2n 2n
Example 8.5 Consider the n = 23 ball
bearing failure times. The maximum
likelihood estimator is
n 23
λ = n
ˆ = = 0. 0138
1661. 16
tΣ i
i=1
failures per 106 revolutions. An exact 95%
confidence interval is
0. 00878 < λ < 0. 0201

S(t)
1.0
0.8
0.6
0.4
0.2
0.0 t
0 50 100 150

Figure 8.1 Empirical and exponential fitted

survivor functions for the ball bearing
 data set.

8.3 The Weibull distribution

Goal: find point and interval estimates for the
p = 2 parameters λ and κ .
The hazard and cumulative hazard functions are
−1
h(t, λ , κ ) = κ λ (λ t)κ and H(t, λ , κ ) = (λ t)κ
for t ≥ 0.
Notation
• n is the number of items on test
• r is the number of observed failures
• t 1 , t 2 , . . . , t n are the failure times
• c 1 , c 2 , . . . , c n are the censoring times
• x i = min {t i , c i } for i = 1, 2, . . . , n
Assuming random censoring
n
log L(λ , κ ) = Σ log h(x i , λ , κ ) − Σ H(x i , λ , κ )
i ∈U i=1
n
= r log κ + κ r log λ + (κ − 1) Σ log x i − λ κ
Σ xκi
i ∈U i=1
The 2 × 1 score vector has elements
∂ log L(λ , κ ) κ r −1
n
U 1 (λ , κ ) =
∂λ
=
λ
− κ λκ Σ xκi
i=1
 ∂ log L(λ , κ )
U 2 (λ , κ ) =
∂κ
r n
= + r log λ + Σ log x i − Σ (λ x i )κ log λ x i
κ i ∈U i=1
There is not a closed form solution for λˆ and κ̂ .
κr κ −1
n
−κλ
λ
Σ xκi = 0
i=1
r n

κ
+ r log λ + Σ log x i − Σ (λ x i )κ log λ x i = 0
i ∈U i=1
The first equation can be solved for λ in terms of κ
1 /κ
 
r 
λ = n
 κ 
Σ x
i=1 i 
The 2 × 2 information matrices are based on
− ∂2 log L(λ , κ ) κ r −2
n

∂λ 2
=
λ2
+ κ (κ − 1) λ κ
Σ xκi
i=1

− ∂2 log L(λ , κ ) r n

∂κ 2
= + Σ
κ 2 i=1
( λ x i )κ
(log λ x i ) 2

− ∂2 log L(λ , κ )
=
∂λ ∂κ
r κ −1
 n 
n
− +λ
λ
κ Σ x i log x i + (1 + κ log λ ) Σ x i 
κ κ
i=1 i=1
 
Information matrices
• the expected values of these quantities are not
tractable
• use λˆ and κ̂ to obtain the observed information
matrix
Example 8.15 Ball bearing data set. The
fitted Weibull distribution: λˆ = 0. 0122 and
κ̂ = 2. 10.

S(t)
1.0
0.8
Weibull
0.6
0.4
0.2 Exponential

0.0 t
0 50 100 150

Figure 8.8 Exponential and Weibull

fits to the ball bearing data.
The log likelihood function at the MLEs is
log L(λˆ, κ̂ ) = − 113. 691
The observed information matrix is
 681, 000 875 
O(λ , κ̂ ) = 
ˆ 
 875 10. 4 
Using the fact that the likelihood ratio
statistic, 2[log L(λˆ, κ̂ ) − log L(λ , κ )], is
asymptotically χ 2 (2), a 95% confidence
region is
2[− 113. 691 − log L(λ , κ )] < 5. 99
since χ 2,2 0.05 = 5. 99.

λ
0.020

0.015
.
0.010

0.005

0.0 κ
0 1 2 3 4
Figure 8.9 Confidence region for
λ and κ (α = 0. 05).
Inverse of the observed information matrix:
 −1
0. 00000165 − 0. 000139
O (λ , κ̂ ) = 
ˆ 
 − 0. 000139 0. 108 
The standard errors of the parameter
estimators are the square roots of the
diagonal elements
σ̂ λˆ = 0. 00128 σ̂ κ̂ = 0. 329
An asymptotic 95% confidence interval for κ
is
2. 10 − (1. 96) (0. 329) < κ < 2. 10 + (1. 96) (0. 329)
or 1. 46 < κ < 2. 74.

9. Parametric Estimation for Models

with Covariates
Motivation
Estimate parameters for the accelerated life and
proportional hazards models.
Outline
• model formulation
• accelerated life
• proportional hazards

9.1 Model formulation

Goal: estimate the vector of regression
coefficients β = ( β 1 , β 2 , . . . , β q )′
Applications
• determine which covariates significantly
impact survival
• determine the impact of changing the values of
covariates
The accelerated life model
S(t, z) = S 0 (t ψ (z))
The proportional hazards model
h(t, z) = ψ (z) h0 (t)
Notation (for i = 1, 2, . . . , n; j = 1, 2, . . . , q)
• x i = min {t i , c i }
• δ i is a censoring indicator variable
• zi = (z i1 , z i2 , . . . , z iq )′
• z ij is the value of covariate j for item i
• extra parameters: S(t, z, θ , β )
Matrix formulation
 x1   δ 1  z 11 z 12 . . . z 1q 
 x2   δ 2  z 21 z 22 . . . z 2q 
     
 .   .   . . . . 
x= δ = and Z =
 .  .  . . . . 
 .  .  . . . . 
     
 xn  δ n  z n1 z n2 . . . z nq 

L(θ , β ) = Π f (x i , zi , θ , β ) Π S(x i , zi , θ , β )
i ∈U i ∈C
The log likelihood function is
log L(θ , β ) = Σ log f (x i , zi , θ , β ) + Σ log S(x i , zi , θ , β )
i ∈U i ∈C
or
n
log L(θ , β ) = Σ log h(x i , zi , θ , β ) − Σ H(x i , zi , θ , β )
i ∈U i=1
Notes
• the maximum likelihood estimators for θ and
β cannot be expressed in closed form
• the number of unique covariate vectors and n
determine whether to use regression models
9.3 Proportional hazards

Example 9.2 A set of n = 3 light bulbs are

placed on test. The first and second bulbs are
100-watt bulbs and the third bulb is a 60-watt
bulb. A single (q = 1) covariate z 1 assumes
the value 0 for a 60-watt bulb and 1 for a
100-watt bulb. Does the wattage have any
influence on the survival distribution of the
bulbs?

100 w 1 X
t1 z11 = 1

2 X t2 z21 = 1
100 w

3 X t3 z31 = 0
60 w
t
20 50 80
t(1) t(2) t(3)
Figure 9.2 Proportional hazards
parameter estimation notation.
  80  1 1
x =  20  δ =1 Z=1
     
 50  1 0
Example 9.8 North Carolina collected
recidivism data on n = 1540 prisoners
in 1978 (Schmidt and Witte, 1988). T
is the time of release until the time of
return to prison. The purpose of the
study is to assess the impact of the
q = 15 covariates.
Table 9.2 North Carolina recidivism model.
 βˆ
zi Covar. βˆ 
√ V̂ [ βˆ ] p-value

√ V̂ [ βˆ ]
z2 AGE -3.34 0.52 -6.43 0.000
z3 PRIORS 0.84 0.14 6.10 0.000
z1 TSERV 1.17 0.20 5.96 0.000
z6 WHITE -0.44 0.09 -5.07 0.000
z8 ALCHY 0.43 0.10 4.11 0.000
z 13 FELON -0.58 0.16 -3.54 0.000
z9 JUNKY 0.28 0.10 2.91 0.002
z7 MALE 0.67 0.24 2.78 0.003
z 15 PROPTY 0.39 0.16 2.47 0.007
z4 RULE 3.08 1.69 1.82 0.034
z 10 MARRY -0.15 0.11 -1.42 0.077
z5 SCHOOL -0.25 0.19 -1.30 0.097
z 12 WORK 0.09 0.09 0.96 0.169
z 14 PERSON 0.07 0.24 0.30 0.381
z 11 SUPER -0.01 0.10 -0.09 0.464

10. Nonparametric Methods

Motivation
Let the data "speak for itself", rather than
approximating the lifetime distribution by
one of the parametric models.
Outline
 • nonparametric estimates of the survivor
function
• life tables

10.1 Survivor function estimation

S(t)
1.0
0.8 Population survivor function

0.6
0.4 Exponential fit
0.2
Nonparametric estimator
0.0 t
0 50 100 150

Figure 10.1 Parametric vs. nonparametric

survivor function estimates.
CASE I: complete data set of n lifetimes
Notation
• R(t), known as the risk set, contains the
indices of all items at risk just prior to
time t
• n(t) = |R(t)| is the cardinality, or number
of elements in R(t)
A nonparametric estimate for the survivor
function is
n(t)
Ŝ(t) = t≥0
n
Notes
• often referred to as the empirical survivor
function
1
• has a downward step of at each
n
observed lifetime if there are no ties
d
• takes a downward step of if there are d
n
tied observations at a particular time
value
An asymptotically valid 100(1 − α )%
confidence interval for S(t) is

Ŝ(t) ± zα / 2

√Ŝ(t)(1 − Ŝ(t))
n
Example 10.1 For the ball bearing data
set, find a nonparametric survivor
function estimator and a 95%
confidence interval for the probability
that a ball bearing will last 50,000,000
cycles.

S(t)
1.0
0.8
0.6
0.4
0.2
0.0 t
0 50 100 150

Figure 10.2 Ball bearing lifetime

survivor function estimate.
Note
• the downward steps in Ŝ(t) have
been connected by vertical lines
• some authors connect the survivor
function estimates at the failure
times with lines
• the survivor function takes a
downward step of 1 / 23 at each
data value except 68.64, where it
takes a downward step of 2 / 23
 A point estimate for S(50) is
16
Ŝ(50) = = 0. 696, and a 95%
23
confidence interval for S(50) is

or
Ŝ(50) ± 1. 96

√Ŝ(50)(1 − Ŝ(50))
23

0. 508 < S(50) < 0. 884

Case II: Right censoring

Notation
• let y 1 < y 2 < . . . < y k be the k distinct
failure times
• let d j denote the number of observed
failures at y j , j = 1, 2, . . . , k
• let n j = n(y j ) denote the number of
items on test just before time y j ,
j = 1, 2, . . . , k and it is customary to
include any values that are censored at
y j in this count
• R(y j ) is the set of all indices of items
that are at risk just before time y j ,
j = 1, 2, . . . , k
Kaplan-Meier (product-limit) estimator
 dj
Ŝ(t) = Π [1 − ]
j ∈ R(t)′ nj

11. Model Adequacy

Motivation
Once a distribution has been fitted to a
sample data set, the adequacy of the model
should be accessed.
Outline
• all parameters known
• parameters estimated from data

11.1 All parameters known

Kolmogorov-Smirnov test
H 0 : F(t) = F 0 (t)
H 1 : F(t) ≠ F 0 (t)
where F(t) is the underlying population
cumulative distribution function. For a
complete data set, the test statistic is
D n = sup | F̂(t) − F 0 (t) |
t
Computational formulas
+ i 
D n = max − F 0 (t (i) )
i = 1, 2, ..., n  n 
  i − 1
D−n = max F 0 (t (i) ) −
i = 1, 2, ..., n  n 
D n = max {D+n , D−n }

S(t)
1.0
0.8 D23 Nonparametric estimator
0.6
0.4
Exponential fit
0.2
0.0 t
0 50 100 150

Figure 11.1 Empirical and fitted survivor functions.

F(t)
1.0
0.8
Exponential fit
0.6
0.4
0.2 D23 Nonparametric estimator
0.0 t
0 50 100 150

Figure 11.2 Empirical and fitted cdfs.

D23 = 0. 301 occurs just to the left of
t (4) = 41. 52.
Table 11.1 Approximate K-S critical
values (all parameters known).

1−α
n 0.80 0.90 0.95 0.99
1 0.900 0.950 0.975 0.995
2 0.683 0.775 0.841 0.929
3 0.565 0.636 0.708 0.829
4 0.493 0.565 0.624 0.734
5 0.447 0.510 0.564 0.668
6 0.410 0.468 0.519 0.615
7 0.381 0.435 0.483 0.576
8 0.358 0.410 0.455 0.543
9 0.339 0.387 0.430 0.513
10 0.323 0.369 0.409 0.490
15 0.266 0.304 0.338 0.405
20 0.231 0.264 0.294 0.352
23 0.216 0.248 0.275 0.330
25 0.208 0.237 0.264 0.317
30 0.190 0.217 0.242 0.290
40 0.166 0.189 0.210 0.252
50 0.148 0.170 0.188 0.226
Table 11.1 gives estimates of the 1 − α
fractiles of the distribution of D n under H 0
determined by Monte Carlo simulation
(500,000 replications).
Example 11.1 Use the K-S test to
determine whether the ball bearing data
set was drawn from a Weibull
population with λ = 0. 01 and κ = 2.
Run the test at α = 0. 10.
The goodness-of-fit test is
H 0 : F(t) = 1 − e− (0.01t)
2

H 1 : F(t) ≠ 1 − e− (0.01t)
2

The test statistic is D23 = 0. 274. At

α = 0. 10, the critical value is 0. 248, so
H 0 is rejected.

F(t)
1.0 Nonparametric estimator
0.8
0.6 D23
0.4
0.2 Hypothesized Weibull dist.

0.0 t
0 50 100 150

Figure 11.6 Empirical and fitted

Weibull cumulative
distribution functions for the ball
bearings.