Learning Objectives for Chapter 7

Report
7
Sampling
Distributions and
Point Estimation of
Parameters
CHAPTER OUTLINE
7-1 Point Estimation
7-2 Sampling Distributions and the
Central Limit Theorem
7-3 General Concepts of Point
Estimation
7-3.1 Unbiased Estimators
7-3.2 Variance of a Point Estimator
7-3.3 Standard Error: Reporting a
Point Estimate
7-3.4 Mean Square Error of an
Estimator
7-4 Methods of Point Estimation
7-4.1 Method of Moments
7-4.2 Method of Maximum
Likelihood
7-4.3 Bayesian Estimation of
Parameters
Chapter 7 Title and Outline
1
Learning Objectives for Chapter 7
After careful study of this chapter, you should be able to do the
following:
1.
2.
3.
4.
5.
6.
7.
Explain the general concepts of estimating the parameters of a
population or a probability distribution.
Explain the important role of the normal distribution as a sampling
distribution.
Understand the central limit theorem.
Explain important properties of point estimators, including bias,
variances, and mean square error.
Know how to construct point estimators using the method of moments,
and the method of maximum likelihood.
Know how to compute and explain the precision with which a parameter
is estimated.
Know how to construct a point estimator using the Bayesian approach.
Chapter 7 Learning Objectives
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
2
Point Estimation
• A point estimate is a reasonable value of a
population parameter.
• Data collected, X1, X2,…, Xn are random
variables.
• Functions of these random variables, x-bar
and s2, are also random variables called
statistics.
• Statistics have their unique distributions that
are called sampling distributions.
Sec 7-1 Point Estimation
3
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
Point Estimator
A point estim ate of som e population para m eter θ
is a single num erical value  .
T he statistic  is called the p oint esti m ato r.
A s an exam ple, suppose the random variable X is norm ally distributed w ith
an unknow n m ean μ. T he sam ple m ean is a point estim ator of the unknow n
population m ean μ. T hat is, μ  X . A fter the sam ple ha s been selected,
the num erical value x is the point estim a te of μ.
T hus if x1  25, x 2  30, x 3  29, x 4  31, the point estim ate of μ is
x 
25  30  29  31
 28.75
4
Sec 7-1 Point Estimation
4
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
Some Parameters & Their Statistics
Parameter
μ
σ2
σ
p
μ1 - μ2
p1 - p2
Measure
Mean of a single population
Variance of a single population
Standard deviation of a single population
Proportion of a single population
Difference in means of two populations
Difference in proportions of two populations
Statistic
x-bar
s2
s
p -hat
x bar1 - x bar2
p hat1 - p hat2
• There could be choices for the point estimator of a parameter.
• To estimate the mean of a population, we could choose the:
– Sample mean.
– Sample median.
– Average of the largest & smallest observations of the sample.
• We need to develop criteria to compare estimates using statistical
properties.
Sec 7-1 Point Estimation
5
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
Some Definitions
• The random variables X1, X2,…,Xn are a random
sample of size n if:
a) The Xi are independent random variables.
b) Every Xi has the same probability distribution.
• A statistic is any function of the observations
in a random sample.
• The probability distribution of a statistic is
called a sampling distribution.
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
6
Sampling Distribution of the Sample Mean
• A random sample of size n is taken from a normal
population with mean μ and variance σ2.
• The observations, X1, X2,…,Xn, are normally and
independently distributed.
• A linear function (X-bar) of normal and independent
random variables is itself normally distributed.
X 
X 1  X 2  ...  X n
has a norm al distribution
n
w ith m ean  X 
    ...  
and variance  X 
2
 
n

2

2
n
 ...  
2
2
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
7
Central Limit Theorem
If X 1 , X 2 , ..., X n is a random sam ple of size n is
taken from a population (either finite o r infinite)
w ith m ean  and finite variance  , and
2
if X is the sam ple m ean,
then the lim iting form of the dist ribution of
Z 
X 

(7-1)
n
as n   , is th e standard norm al distribut ion .
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
8
Sampling Distributions of Sample Means
Figure 7-1 Distributions of
average scores from
throwing dice. Mean = 3.5
F o rm u las
 
ba
2

2
X

 b  a  1
12
X X n
2
2
2
1
a)
b)
c)
d)
e)
n
std
dies var dev
1 2.9 1.7
2 1.5 1.2
3 1.0 1.0
5 0.6 0.8
10 0.3 0.5
a= 1
b= 6
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
9
Example 7-1: Resistors
An electronics company
manufactures resistors having a
mean resistance of 100 ohms
and a standard deviation of 10
ohms. The distribution of
resistance is normal. What is
the probability that a random
sample of n = 25 resistors will
have an average resistance of
less than 95 ohms?
Answer:
X 
X
n

10
Figure 7-2 Desired probability is
shaded
 2 .0
25
 X  
 95  200 







2



X

    2 .5   0 .0 0 6 2
0.0062 = NORMSDIST(-2.5)
A rare event at less than 1%.
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
10
Example 7-2: Central Limit Theorem
Suppose that a random variable X has a continuous
uniform distribution:
1 2 , 4  x  6
f x  
 0, otherw ise
Find the distribution of the sample mean of a random
sample of size n = 40.
D istribution is norm al by the C LT .
 
ba
64

2

2

2
b  a 
2

6  4
12
X 
2

 5.0
2
n

2
1 3
12
1 3
40

1
120
Figure 7-3 Distributions of
X and X-bar
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
11
Two Populations
We have two independent normal populations. What is
the distribution of the difference of the sample means?
T he sam pling distribution of X 1  X 2 is:
X
1
X2
  X   X  1   2
1
2
1
2
X
2
1
X2

2
X1

2
X2

n1
2
2

n2
T he distribution of X 1  X 2 is norm al if:
(1) n1 and n 2 are both greater than 30,
regardless of the distributions of X 1 and X 2 .
(2) n1 and n 2 are less than 30,
w hile the distributions are som ew hat norm al.
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
12
Sampling Distribution of a Difference in Sample Means
• If we have two independent populations with means μ1 and
μ2, and variances σ12 and σ22,
• And if X-bar1 and X-bar2 are the sample means of two
independent random samples of sizes n1 and n2 from these
populations:
• Then the sampling distribution of:
Z 
X
1
 X 2    1   2 

2
1
n1


(7-4)
2
2
n2
is approximately standard normal, if the conditions of the
central limit theorem apply.
• If the two populations are normal, then the sampling
distribution is exactly standard normal.
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
13
Example 7-3: Aircraft Engine Life
The effective life of a component
used in jet-turbine aircraft
engines is a normal-distributed
random variable with
parameters shown (old). The
engine manufacturer
introduces an improvement
into the manufacturing
process for this component
that changes the parameters
as shown (new).
Random samples are selected
from the “old” process and
“new” process as shown.
What is the probability the
difference in the two sample
means is at least 25 hours?
Figure 7-4 Sampling distribution of
the sample mean difference.
Process
Old (1) New (2) Diff (2-1)
x -bar = 5,000 5,050
50
s=
40
30
50
n=
16
25
Calculations
s / √n =
10
6
11.7
z=
-2.14
P(xbar2-xbar1 > 25) = P(Z > z) = 0.9840
= 1 - NORMSDIST(z)
Sec 7-2 Sampling Distributions and the Central Limit Theorem
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
14
General Concepts of Point Estimation
• We want point estimators that are:
– Are unbiased.
– Have a minimal variance.
• We use the standard error of the estimator to
calculate its mean square error.
Sec 7-3 General Concepts of Point Estimation
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
15
Unbiased Estimators Defined
T he point estim ator  is an u n biased estim ator
for the param eter θ if:
 
E  θ
(7-5)
If the estim ator is not unbiased, then t he diff e rence:
 
E  θ
(7-6)
is called the bi a s of the estim ator  .
T he m ean of the sam pling distribution of 
is equ al t o θ .
Sec 7-3.1 Unbiased Estimators
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
16
Example 7-4: Sample Man & Variance Are Unbiased-1
• X is a random variable with mean μ and variance σ2. Let X1,
X2,…,Xn be a random sample of size n.
• Show that the sample mean (X-bar) is an unbiased estimator
of μ.
EX

 X 1  X 2  ...  X n 
 E

n



1
n

1
n
 E  X 1   E  X 2   ...  E  X n  

   ...    
n
 
n
Sec 7-3.1 Unbiased Estimators
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
17
Example 7-4: Sample Man & Variance Are Unbiased-2
Show that the sample variance (S2) is a unbiased estimator of σ2.
E S
2



 E



n

X  X

2
i 1
n 1


1
 n
2

E   X i  X
 n  1  i 1


2

 2 XX i 

  n
1  n
2
2 
2

E
X

nX

E
X
 nE  X

i 
  i


n  1   i 1
  n  1  i 1
1
 n
2





n  1  i 1
1

1
 n  2  n
n 1 
2
2
  n
2

2
2




n 

1
2
2
  n  1  2   
 n     

n 1 
Sec 7-3.1 Unbiased Estimators
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
2
18
Other Unbiased Estimators of the Population Mean
M ean = X 
110.4
 11.04
10
M edian = X 
10.3  11.6
 10.95
2
T rim m ed m ean =
110.04  8.5  14.1
 10.81
8
• All three statistics are unbiased.
– Do you see why?
• Which is best?
– We want the most reliable one.
Sec 7-3.1 Unbiased Estimators
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
i
xi
xi'
1
2
3
4
5
6
7
8
9
10
Σ
12.8
9.4
8.7
11.6
13.1
9.8
14.1
8.5
12.1
10.3
110.4
8.5
8.7
9.4
9.8
10.3
11.6
12.1
12.8
13.1
14.1
19
Choosing Among Unbiased Estimators
S uppose that  1 and  2 are unbiased estim ato rs of θ.
T he variance of  1 is less than the varian ce of  2 .
  1 is preferable.
Figure 7-5 The sampling distributions
of two unbiased estimators.
Sec 7-3.2 Variance of a Point Estimate
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
20
Minimum Variance Unbiased Estimators
• If we consider all unbiased estimators of θ, the
one with the smallest variance is called the
minimum variance unbiased estimator (MVUE).
• If X1, X2,…, Xn is a random sample of size n from a
normal distribution with mean μ and variance σ2,
then the sample X-bar is the MVUE for μ.
• The sample mean and a single observation are
unbiased estimators of μ. The variance of the:
– Sample mean is σ2/n
– Single observation is σ2
– Since σ2/n ≤ σ2, the sample mean is preferred.
Sec 7-3.2 Variance of a Point Estimate
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
21
Standard Error of an Estimator
T he s tandard error of an estim ator  is its standard de viation, given by
 
 
V  .
If the standard error involves unknow n p aram eters tha t can be estim ated ,
substitution of these values into  
produces an estim ated standar d error , denoted by   .
 
E quivalent notation:    s   se 
If the X i are ~ N   ,   , then X is norm ally di stributed,
and 
X


. If  is not kno w n, then 
n
X

s
.
n
Sec 7-3.3 Standard Error Reporting a Point Estimate
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
22
Example 7-5: Thermal Conductivity
• These observations are 10
measurements of thermal
conductivity of Armco iron.
• Since σ is not known, we use s to
calculate the standard error.
• Since the standard error is 0.2% of
the mean, the mean estimate is fairly
precise. We can be very confident
that the true population mean is
41.924 ± 2(0.0898).
xi
41.60
41.48
42.34
41.95
41.86
42.18
41.72
42.26
41.81
42.04
41.924 = Mean
0.284 = Std dev (s )
0.0898 = Std error
Sec 7-3.3 Standard Error Reporting a Point Estimate
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
23
Mean Squared Error
T he m ean squared error of an estim ator 
of the param eter θ is defined as:
 

M SE   E   θ

2
(7-7)
 
2
 
C an be rew ritten as  E    E     θ  E  




 
 V    bias 
2
2
Conclusion: The mean squared error (MSE) of the
estimator is equal to the variance of the
estimator plus the bias squared. It measures
both characteristics.
Sec 7-3.4 Mean Squared Error of an Estimator
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
24
Relative Efficiency
• The MSE is an important criterion for
comparing two estimators.



M SE  
M SE 1
R elative efficie n cy
2
• If the relative efficiency is less than 1, we
conclude that the 1st estimator is superior to
the 2nd estimator.
Sec 7-3.4 Mean Squared Error of an Estimator
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
25
Optimal Estimator
• A biased estimator can be
preferred to an unbiased
estimator if it has a
smaller MSE.
• Biased estimators are
occasionally used in linear
regression.
• An estimator whose MSE
is smaller than that of any
other estimator is called
an optimal estimator.
Figure 7-6 A biased estimator has a
smaller variance than the unbiased
estimator.
Sec 7-3.4 Mean Squared Error of an Estimator
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
26
Methods of Point Estimation
• There are three methodologies to create point
estimates of a population parameter.
– Method of moments
– Method of maximum likelihood
– Bayesian estimation of parameters
• Each approach can be used to create
estimators with varying degrees of biasedness
and relative MSE efficiencies.
Sec 7-4 Methods of Point Estimation
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
27
Method of Moments
• A “moment” is a kind of an expected value of
a random variable.
• A population moment relates to the entire
population or its representative function.
• A sample moment is calculated like its
associated population moments.
Sec 7-4.1 Method of Moments
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
28
Moments Defined
• Let X1, X2,…,Xn be a random sample from the
probability f(x), where f(x) can be either a:
– Discrete probability mass function, or
– Continuous probability density function
• The kth population moment (or distribution
moment) is E(Xk), k = 1, 2, ….
• The kth sample moment is (1/n)ΣXk, k = 1, 2, ….
• If k = 1 (called the first moment), then:
– Population moment is μ.
– Sample moment is x-bar.
• The sample mean is the moment estimator of the
population mean.
Sec 7-4.1 Method of Moments
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
29
Moment Estimators
L et X 1 , X 2 , ..., X n be a random sam ple from eithe r
a probability m ass function or a probabi lity density function
w ith m unknow n param eters θ 1 , θ 2 , ..., θ m .
T he m om ent estim at ors  1 ,  2 , ...,  m are found
by e qu ating the first m population m om ents
to the first m sam ple m om ents and
solving the resulting sim ultaneous equat ions
for the unknow n param eters.
Sec 7-4.1 Method of Moments
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
30
Example 7-6: Exponential Moment Estimator-1
• Suppose that X1, X2, …, Xn is a random sample
from an exponential distribution with
parameter λ.
• There is only one parameter to estimate, so
equating population and sample first
moments, we have E(X) = X-bar.
• E(X) = 1/λ = x-bar
• λ = 1/x-bar is the moment estimator.
Sec 7-4.1 Method of Moments
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
31
Example 7-6: Exponential Moment Estimator-2
• As an example, the time to
failure of an electronic module
is exponentially distributed.
• Eight units are randomly
selected and tested. Their
times to failure are shown.
• The moment estimate of the λ
parameter is 0.04620.
xi
11.96
5.03
67.40
16.07
31.50
7.73
11.10
22.38
21.646 = Mean
0.04620 = λ est
Sec 7-4.1 Method of Moments
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
32
Example 7-7: Normal Moment Estimators
Suppose that X1, X2, …, Xn is a random sample from a
normal distribution with parameter μ and σ2. So E(X)
= μ and E(X2) = μ2 + σ2.
  X 
1
n

n
Xi
and
 
2
2

i 1


1
n

n

i 1
Xi  X
2

n
2
Xi
i 1
n
2
n
1
2

i 1
1
2
Xi  n
n
n
2
n


 
 n
 Xi  
1
 i 1
 
2
  X i 



n i 1
n




n

i 1

Xi 

n
X
i
i 1
 X
2

2
(biased)
n
Sec 7-4.1 Method of Moments
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
33
Example 7-8: Gamma Moment Estimators-1
P aram eters = S tatistics
r

r

2
r  r  1

2
 EX

 EX
2
 EX 
 EX
2

r 
X is th e m ean
2
is th e varian ce o r
an d n o w so lvin g fo r r an d  :
2
X
n
1 / n  
Xi  X
2
2
i 1
 
X
n
1 / n  
Xi  X
2
2
i 1
Sec 7-4.1 Method of Moments
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
34
Example 7-8: Gamma Moment Estimators-2
xi
Using the exponential example
data shown, we can estimate
the parameters of the gamma
distribution.
x-bar =
21.646
11.96
5.03
67.40
16.07
31.50
7.73
11.10
22.38
ΣX 2 = 6645.4247
2
X
r 

n
1 / n  
Xi  X
2
2
21.646
2
xi
143.0416
25.3009
4542.7600
258.2449
992.2500
59.7529
123.2100
500.8644
2
1 8  6645.4247  21.646
2
 1.29
i 1
X
 

n
1 / n   X i  X
2
2
21.646
1 8  6645.4247  21.646
2
 0.0598
i 1
Sec 7-4.1 Method of Moments
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
35
Maximum Likelihood Estimators
• Suppose that X is a random variable with probability
distribution f(x:θ), where θ is a single unknown
parameter. Let x1, x2, …, xn be the observed values in a
random sample of size n. Then the likelihood function of
the sample is:
L(θ) = f(x1: θ) ∙ f(x2; θ) ∙…∙ f(xn: θ)
(7-9)
• Note that the likelihood function is now a function of
only the unknown parameter θ. The maximum likelihood
estimator (MLE) of θ is the value of θ that maximizes the
likelihood function L(θ).
• If X is a discrete random variable, then L(θ) is the
probability of obtaining those sample values. The MLE is
the θ that maximizes that probability.
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
36
Example 7-9: Bernoulli MLE
Let X be a Bernoulli random variable. The probability mass
function is f(x;p) = px(1-p)1-x, x = 0, 1 where P is the parameter
to be estimated. The likelihood function of a random sample
of size n is:
L p  p
x1
1  p 
1  x1
p
x2
1  p 
1 x2
 ...  p
xn
1  p 
1 xn
n
n


p
xi
1  p 
1  xi
 xi
 p i 1
n
1  p 
n
 xi
i 1
i 1
 n


ln L  p     x i  ln p   n 
 i 1 

n

i 1

x i  ln  1  p 

n


n

x
xi 
 i
d ln L  p  
i 1


 i 1

0
dp
p
1  p 
n
n
p
x
i
i 1
n
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
37
Example 7-10: Normal MLE for μ
Let X be a normal random variable with unknown mean μ and
known variance σ2. The likelihood function of a random
n
  x     2 
1
sample of size n is:
L    
e
2
2
i
i 1

2
1

1
 2 
2
ln L    
d ln L   
d

n
2
ln  2 
n
1

n 2
2
e
 x
i
2
2
2
n
  xi   
2
i 1

n
1
2
2

 xi   
2
i 1
 0
i 1
n
 
x
i 1
i
 X (sam e as m om ent estim ator)
n
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
38
Example 7-11: Exponential MLE
Let X be a exponential random variable with parameter λ. The
likelihood function of a random sample of size n is:
n
L   
n
 e
  xi

 e
n
 xi
i 1
i 1
n
ln L     n ln       x i
i 1
d ln L   
d

n

n

x
i
0
i 1
n
 n
x
i
1 X
(sam e as m om ent estim ator)
i 1
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
39
Why Does MLE Work?
• From Examples 7-6 & 11 using the 8 data observations, the plot of
the ln L(λ) function maximizes at λ = 0.0462. The curve is flat near
max indicating estimator not precise.
• As the sample size increases, while maintaining the same x-bar, the
curve maximums are the same, but sharper and more precise.
• Large samples are better
Figure 7-7 Log likelihood for exponential distribution. (a) n = 8, (b) n = 8, 20, 40.
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
40
Example 7-12: Normal MLEs for μ & σ2
Let X be a normal random variable with both unknown mean μ
and variance σ2. The likelihood function of a random sample
of size n is:
1

 

L   ,   
e
n
i 1

1
 2 
2
ln L   , 
 ln L   , 
2
2



 ln L   , 
 
2

2



n
n
 x
2
e
2
2
n
  xi   
i 1
Sec 7-4.2 Method of Maximum Likelihood
n
1
  2   x
2
2
i

i
2
 0
2

n
1
2
4
  xi   
and 
2
0
i 1
n
  X
2
i 1
n
2
2
i 1
1

n 2
ln  2 
2
2
2
1

2
 xi  
2
2


 xi  X

2
i 1
n
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
41
Properties of an MLE
U nder very general and non-restrictive c onditions,
w hen the sam ple size n is large and if  is the M LE of the param eter ,
 
(1)  is an approxim ately unbiased estim ator for θ, i.e.,  E   θ 


(2) T he var iance of  is nearly as sm all as the vari ance
that could be obtained w ith any other es tim ator, and
(3)  has an approxim ate norm al distribu tion.
Notes:
• Mathematical statisticians will often prefer MLEs because of these
properties. Properties (1) and (2) state that MLEs are MVUEs.
• To use MLEs, the distribution of the population must be known or
assumed.
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
42
Importance of Large Sample Sizes
• Consider the MLE for σ2 shown in Example 7-12:
n
 
E 
2

 x
i
 X

2
i 1

n 1
n

2
n
T hen the bias is:
 
E 
2
2

n 1
n
 
2
2


2
n
• Since the bias is negative, the MLE underestimates
the true variance σ2.
• The MLE is an asymptotically (large sample) unbiased
estimator. The bias approaches zero as n increases.
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
43
Invariance Property
Let  1 ,  2 , ...,  k be the m axim um likelihood est im ators (M LE s)
of the param eters θ 1 , θ 2 , ..., θ k .
T hen the M LE s for any function h  θ 1 , θ 2 , ..., θ k

is the sam e function h  1 ,  2 , ...,  k

of these param eters
 of the estim ators  , 
1
2
, ...,  k
This property is illustrated in Example 7-13.
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
44
Example 7-13: Invariance
For the normal distribution, the MLEs were:
n
  X
and 
2

 x
 X
i

2
i 1
n
T o obtain the M LE of the function h   , 
substitute the estim ators  and 
n
 

2

 x
i
 X

2
2



2
,
into the function h :
2
i 1
n
w hich is not the sam ple standard deviation s.
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
45
Complications of the MLE Method
The method of maximum likelihood is an
excellent technique, however there are two
complications:
1. It may not be easy to maximize the likelihood
function because the derivative function set
to zero may be difficult to solve algebraically.
2. The likelihood function may be impossible to
solve, so numerical methods must be used.
The following two examples illustrate.
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
46
Example 7-14: Uniform Distribution MLE
Let X be uniformly distributed on the interval 0 to a.
f  x   1 a for 0  x  a
L a 
n

i 1
dL  a 
da

1
a
n
a
n 1

1
a
n
  na
a
n
for 0  x i  a
  n 1
a  m ax  x i 
Figure 7-8 The likelihood function for
this uniform distribution
Calculus methods don’t work here because L(a) is maximized at
the discontinuity.
Clearly, a cannot be smaller than max(xi), thus the MLE is max(xi).
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
47
Example 7-15: Gamma Distribution MLE-1
Let X1, X2, …, Xn be a random sample from a gamma
distribution. The log of the likelihood function is:
 n  r x ir  1 e   xi
ln L  r ,    ln  

 r
 i 1



n
n
i 1
i 1
 nr ln      r  1   ln  x i   n ln    r      x i
 ln L  r ,  
r
 ln  x   n
 n ln    
 ' r 
i
i 1
 ln L  r ,  

n

 
nr

r
x
 r
 0
n

x
i
 0
i 1
 ' r 
   ln  x   n   r 
n
and
n ln  
i
i 1
T here is no closed solution for r and  .
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
48
Example 7-15: Gamma Distribution MLE-2
Figure 7-9 Log likelihood for the gamma distribution using the failure time data
(n=8). (a) is the log likelihood surface. (b) is the contour plot. The log likelihood
function is maximized at r = 1.75, λ = 0.08 using numerical methods. Note the
imprecision of the MLEs inferred by the flat top of the function.
Sec 7-4.2 Method of Maximum Likelihood
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
49
Bayesian Estimation of Parameters-1
• The moment and likelihood methods interpret probabilities
as relative frequencies and are called objective frequencies.
• The Bayesian method combines sample information with
prior information.
• The random variable X has a probability distribution of
parameter θ called f(x|θ). θ could be determined by
classical methods.
• Additional information about θ can be expressed as f(θ),
the prior distribution, with mean μ0 and variance σ02, with
θ as the random variable. Probabilities associated with f(θ)
are subjective probabilities.
• The joint distribution is f(x1, x2, …, xn, θ)
• The posterior distribution is f(θ|x1, x2, …, xn) is our degree
of belief regarding θ after gathering data
7-4.3 Bayesian Estimation of Parameters
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
50
Bayesian Estimation of Parameters-2
• Now putting these together, the joint is:
– f(x1, x2, …, xn, θ) = f(x1, x2, …, xn |θ) ∙ f(θ)
• The marginal is:
f
 x1 , x 2 , ..., x n 
  f  x1 , x 2 , ..., x n , θ  , for θ discrete
 θ

  f  x1 , x 2 , ..., x n , θ  d θ, for θ continuous
  
• The desired posterior distribution is:
f  θ | x1 , x 2 , ..., x n  
f  x1 , x 2 , ..., x n , θ 
f  x1 , x 2 , ..., x n 
• And the Bayesian estimator of θ is the expected
value of the posterior distribution
7-4.3 Bayesian Estimation of Parameters
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
51
Example 7-16: Bayes Estimator for a Normal Mean-1
Let X1, X2, …, Xn be a random sample from a normal distribution
unknown mean μ and known variance σ2. Assume that the
prior distribution for μ is:
f μ  
1
2  0
e
   0 
2
2 0
2
1

2 
e

   2  0   0
2
2

2 0
2
2
0
The joint distribution of the sample is:
f
 x1 , x 2 , ..., x n |   


 1 2
1
 2 
2
n 2
 2 
2
n 2
2
i 1

e
 ( x   )
i
e
 1 2
1
n
2
2




n

i 1
n
2
xi
2

i 1
7-4.3 Bayesian Estimation of Parameters
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.

2
xi  n  


52
Example 7-16: Bayes Estimator for a Normal Mean-2
Now the joint distribution of the sample and μ is:
f
 x1 , x 2 , ..., x n ,   

f
 x1 , x 2 , ..., x n |    f  μ 
1
 2 
2
n 2
 1   2
w here u  
 
 2  
 h1    e
u
2
 1
n 
 2  2   2
0  
  0  xi
 2 
2


 0

 1
 
1 
x 
2
 1 2    

2  0 


 2  2 n 
 2  2 n 

 0

 0
 
 h2    e
f   | x1 , x 2 , ..., x n   h3    e
2  0
e




2
2
  xi
0 
 2
 
2

 0 

& com pleting the square
2



2
 1
x 0
1  2   n 0

 1 2  





  2  2 n 
  2   2 n  2   2 n  
0
 0

 0
 

2
2



 1
1   2   n  0  0 x 
 1 2  

  
2
2
  2  2 n 

 
 0  n
 0


 

2
is the posterior distribution
hi     functi on to collect unneeded com ponents (not  )
7-4.3 Bayesian Estimation of Parameters
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
53
Example 7-16: Bayes Estimator for a Normal Mean-3
• After all that algebra, the bottom line is:
E     

2
n 0   0 x
2
0 
2
 1
1 
V     2  2 
0  n 
2
n
1
 0 
2

0 
2
n
2
2
n
• Observations:
– Estimator is a weighted average of μ0 and x-bar.
– x-bar is the MLE for μ.
– The importance of μ0 decreases as n increases.
7-4.3 Bayesian Estimation of Parameters
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
54
Example 7-16: Bayes Estimator for a Normal Mean-4
To illustrate:
– The prior parameters: μ0 = 0, σ02= 1
– Sample: n = 10, x-bar = 0.75, σ2 = 4
 


2
n 0   0 x
2
0 
2
4
2
n
10  0  1  0.75 
1   4 10 
 0.536
7-4.3 Bayesian Estimation of Parameters
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.
55
Important Terms & Concepts of Chapter 7
Bayes estimator
Bias in parameter estimation
Central limit theorem
Estimator vs. estimate
Likelihood function
Maximum likelihood estimator
Mean square error of an estimator
Minimum variance unbiased
estimator
Moment estimator
Normal distribution as the
sampling distribution of the:
– sample mean
– difference in two sample
means
Parameter estimation
Point estimator
Population or distribution
moments
Posterior distribution
Prior distribution
Sample moments
Sampling distribution
An estimator has a:
– Standard error
– Estimated standard error
Statistic
Statistical inference
Unbiased estimator
Chapter 7 Summary
56
© John Wiley & Sons, Inc. Applied Statistics and Probability for Engineers, by Montgomery and Runger.

similar documents