0% found this document useful (0 votes)
10 views28 pages

Chapter4 Estimation

This document discusses estimation in inferential statistics, focusing on obtaining approximate values for statistical parameters from sample data. It covers different methods for calculating estimators, including the method of moments and maximum likelihood estimation (MLE), along with examples illustrating these concepts. Additionally, it outlines the main qualities of an estimator, such as convergence, unbiasedness, and admissibility.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
10 views28 pages

Chapter4 Estimation

This document discusses estimation in inferential statistics, focusing on obtaining approximate values for statistical parameters from sample data. It covers different methods for calculating estimators, including the method of moments and maximum likelihood estimation (MLE), along with examples illustrating these concepts. Additionally, it outlines the main qualities of an estimator, such as convergence, unbiasedness, and admissibility.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 28

Probability & Statistics II

Chapter4: Estimation
Presented by: Dr. BOUREDJA Sara

[email protected]
1- Introduction

The concept of estimation was introduced in inferential statistics. Its


essential mission is to obtain an approximate value for one or more
statistical parameters, for a 𝑅𝑉 in a population, from data observed in a
sample.

[email protected]
2-Estimator

Let 𝑋1 , 𝑋2 , … , 𝑋𝑛 a 𝑛 −sample from 𝑅𝑉. 𝑋 distributed according to the same


probability distribution 𝑓(𝑥, 𝜃), where 𝜃 is a parameter associated with this
probability distribution and its value is unknown.
An estimator of 𝜃 is a random variable 𝑇 as a function of 𝑋𝑖 .
𝑇 = 𝑓(𝑋1 , 𝑋2 , … , 𝑋𝑛 ).
When the sample has actually been achieved, the estimator takes on a
numerical value, called the estimator of the parameter 𝜃:
𝜃෠ = 𝑓(𝑥1 , 𝑥2 , … , 𝑥𝑛 )

Statistical model= Probability distribution

[email protected]
3-Estimator calculation methods

3-1-The method of moments


Let 𝜇𝑟 be the theoretical moment of order 𝑟 of the distribution, i.e.:
Continuous case 𝜇𝑟 = 𝐸 𝑋 𝑟 = ‫𝑥 𝑓 𝑟 𝑥 ׬‬, 𝜃 𝑑𝑥,
Discrete case 𝜇𝑟 = 𝐸(𝑋 𝑟 ) = σ𝑛𝑖=1 𝑥𝑖𝑟 𝑃 𝑥, 𝜃
1
Let 𝑚𝑟 the empirical moment of order 𝑟 of the sample, i.e.: 𝑚𝑟 = σ𝑛𝑖=1 𝑋𝑖𝑟
𝑛

෡ is obtained by solving the following equation:


The estimator 𝜽
1
Continuous case ‫𝑥 𝑓 𝑟 𝑥 ׬‬, 𝜃 𝑑𝑥 = 𝑛 σ𝑛𝑖=1 𝑋𝑖𝑟
1 𝑛
Discrete case σ𝑛𝑖=1 𝑥𝑖𝑟 𝑃 𝑥, 𝜃 = σ𝑖=1 𝑋𝑖𝑟
𝑛

[email protected]
Example 1
1 − 𝑥−𝜃
ቐ𝜃 𝑒 , 𝑥≥0
𝜃
Let 𝑋1 , 𝑋2 , … , 𝑋𝑛 a sample from the RV 𝑋, with density: 𝑓 𝑥, 𝜃 =
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
-Determine an estimator of 𝜃 by the method of moments.
Solution
1
The empirical moment: 𝑚1 = σ𝑛𝑖=1 𝑋𝑖 = 𝑋ത
𝑛
+∞
The theoretical moment: 𝜇1 = 𝐸 𝑋 = ‫׬‬−∞ 𝑥𝑓 𝑥, 𝜃 𝑑𝑥
+∞ 1 − 𝑥−𝜃
= ‫׬‬0 𝑥 𝜃 𝑒 𝜃 𝑑𝑥
𝑥−𝜃 +∞ 𝑥−𝜃
− 𝜃 +∞ − 𝜃
= − 𝑥𝑒 0 − ‫׬‬0 −𝑒 𝑑𝑥
𝑥−𝜃 +∞

= 𝜃 + −𝜃𝑒 𝜃 0
=𝜃+𝜃
= 2𝜃
So, the estimator 𝜃መ is obtained by solving the following equation: 2𝜃 = 𝑋ത
𝑋ഥ
𝜃መ =
2
[email protected]
3-2-Maximum Likelihood Method (MLE)

We call the likelihood function of 𝜃 a realization 𝑥1 , 𝑥2 , … , 𝑥𝑛 of a sample 𝑋1 , 𝑋2 , … , 𝑋𝑛 the


function of 𝜃 :
Continuous case 𝐿 𝑥1 , 𝑥2 , … , 𝑥𝑛 , 𝜃 = ς𝑛𝑖=1 𝑓(𝑥𝑖 , 𝜃)
Discrete case 𝐿 𝑥1 , 𝑥2 , … , 𝑥𝑛 , 𝜃 = ς𝑛𝑖=1 𝑃(𝑋 = 𝑥𝑖 )

The method of estimating 𝜃 by the value that maximises the likelihood 𝐿, is called the
maximum likelihood method (𝑀𝐿𝐸).
𝜕𝑙𝑛 𝐿(𝑥1 , 𝑥2 , … , 𝑥𝑛 ; 𝜃)
=0
𝜕𝜃
𝜕 2 ln𝐿(𝑥1 , 𝑥2 , … , 𝑥𝑛 ; 𝜃)
෠ <0
(𝜃)
2
𝜕 𝜃

[email protected]
Example2
We want to estimate the parameter 𝜆 of a Poisson distribution from a sample.
𝑥
−𝜆 𝜆
We have: 𝑃 𝑋 = 𝑥 = 𝑒 𝑥!
The likelihood function
𝑛
: 𝑛 𝑛 𝑛
−𝜆
𝜆 𝑥𝑖 −𝜆𝑛
𝜆 𝑥𝑖 −𝜆𝑛
𝜆σ𝑖=1 𝑥𝑖
𝐿(𝑥1 , 𝑥2 , … , 𝑥𝑛 ; 𝜆) = ෑ 𝑃 𝑋 = 𝑥𝑖 = ෑ 𝑒 =𝑒 ෑ =𝑒
𝑥𝑖 ! 𝑥𝑖 ! ς𝑛𝑖=1 𝑥𝑖 !
𝑖=1 𝑖=1 𝑖=1

𝑛
σ𝑖=1 𝑥𝑖 𝑛 𝑛 𝑛
𝜆 𝑛
𝑙𝑛 𝐿(𝑥1 , 𝑥2 , … , 𝑥𝑛 ; 𝜆) = 𝑙𝑛 𝑒 −𝜆𝑛 𝑛 = 𝑙𝑛𝑒 −𝜆𝑛 + 𝑙𝑛𝜆σ𝑖=1 𝑥𝑖 + 𝑙𝑛 ෑ 𝑥𝑖 ! = −𝜆𝑛 + 𝑙𝑛 𝜆 ෍ 𝑥𝑖 − ෍ 𝑙𝑛 𝑥𝑖 !
ς𝑖=1 𝑥𝑖 !
𝑖=1 𝑖=1 𝑖=1
𝜕𝑙𝑛 𝐿(𝑥1 ,𝑥2 ,…,𝑥𝑛 ;𝜆) σ𝑛
𝑖=1 𝑥𝑖
The first derivative : 𝜕𝜆
= −𝑛 +
𝜆
σ𝑛
𝑖=1 𝑥𝑖
Annuls for 𝜆መ = 𝑛
𝜕2 𝑙𝑛 𝐿(𝑥1 ,𝑥2 ,…,𝑥𝑛 ;𝜆) σ𝑛
𝑖=1 𝑥𝑖
The second derivative : 𝜕2 𝜆
=−
𝜆2
Is negative.
σ𝑛 𝑖=1 𝑥𝑖
Thus, the estimator given 𝜆መ = 𝑛
= 𝑋ത leads to a maximum likelihood estimator equal to
𝜆መ 𝑀𝐿𝐸 = 𝑥ҧ

[email protected]
4- Main qualities of an estimator
❖Convergent estimator: 𝜃መ is said to be convergent if: lim 𝑅 𝜃;
መ 𝜃 =0
𝑛→∞
መ 𝜃 = 𝐸 𝑙 𝜃,
Risk of an estimator: 𝑅 𝜃; መ 𝜃
መ 𝜃) = |𝜃መ − 𝜃|2
Quadratic loss: 𝑙(𝜃,
Note1 in the case of quadratic loss, the risk will then be:
𝑅 𝜃;መ 𝜃 = 𝐸 |𝜃,መ 𝜃|2

❖Unbiased estimator: 𝜃መ is unbiased if 𝐸 𝜃መ = 𝜃.


መ 𝜃 = 𝑉 𝜃መ
Note2 If the estimator is unbiased, then: 𝑅 𝜃;

❖ Asymptotically unbiased estimator: 𝜃መ is an asymptotically unbiased estimator of 𝜃 if: lim 𝐸( 𝜃)


መ =𝜃
𝑛→∞
❖Best estimator:
Let 𝜃መ1 and 𝜃መ2 be two estimators of 𝜃, we say that 𝜃መ1 is a better estimator than 𝜃መ2 if: 𝑅 𝜃መ1 ; 𝜃 < 𝑅 𝜃መ2 ; 𝜃 , ∀ 𝜃.
❖Admissible estimator: An estimator is said to be admissible if there is no better estimator than it.

[email protected]
Example 3: Let 𝑋 be a random variable with uniform distribution on [0, 𝜃] et 𝑋1 , 𝑋2 , … , 𝑋𝑛 an n-sample from 𝑋.
- Find the 𝑀𝐿𝐸 de 𝜃.
Solution:
1
10≤𝑥≤𝜃 ,
The density of 𝑋 is written: 𝑓 𝑥 = ቐ𝜃
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
1 1 1
The likelihood is therefore: 𝐿 𝑥 ; 𝜃 = 10≤𝑥1 ≤𝜃 × 10≤𝑥2 ≤𝜃 × ⋯× 10≤𝑥𝑛 ≤𝜃
𝜃 𝜃 𝜃
1
= 1 1 .
𝜃𝑛 {0≤min 𝑥𝑖 } {max 𝑥𝑖 ≤𝜃}
Since the domain of 𝑓 𝑥 ; 𝜃 depends on 𝜃, the maximum likelihood of 𝜃 cannot be obtained by deriving with
respect to 𝜃, but by choosing 𝜃 as large as possible. Since 𝜃 ≥ 𝑥, in particular 𝜃 ≥ max 𝑥𝑖 , the likelihood
𝐿 𝑥1 , 𝑥2 , … , 𝑥𝑛 ; 𝜃 is then written:
1
, 𝑓𝑜𝑟 𝜃 ≥ max 𝑥𝑖 ;
𝐿 𝑥; 𝜃 = 𝐿 𝑥1 , 𝑥2 , … , 𝑥𝑛 ; 𝜃 = ቐ𝜃 𝑛
0, 𝑓𝑜𝑟 0 ≤ 𝜃 < max 𝑥𝑖 .

1
𝐿 𝑥1 , 𝑥2 , … , 𝑥𝑛 ; 𝜃 cannot be larger than 𝜃𝑛 . Therefore, the MLE of 𝜃
is 𝜃෠ = max 𝑥𝑖 .
[email protected] Figure1.Graph of 𝐿 𝑥; 𝜃 as a function of 𝜃
5-Sufficient statistics (exhaustive)

Let 𝑋1 , 𝑋2 , … , 𝑋𝑛 be an n-sample, the likelihood of the sample is given by the function


𝐿 𝑥; 𝜃 .
Let 𝑇 be a function statistic of 𝑋1 , 𝑋2 , … , 𝑋𝑛 with distribution 𝑔(𝑡; 𝜃).
𝑇 Is said to be exhaustive if the likelihood 𝐿 𝑥; 𝜃 is written in the following form:
𝐿 𝑥; 𝜃 = 𝑔 𝑡; 𝜃 ℎ 𝑥
This is the Neyman-Fisher factorisation principle.
Example 4
Let 𝑋1 , 𝑋2 , … , 𝑋𝑛 be an n-sample from a RV 𝑋 with the Poisson distribution 𝑃 𝜃 defined
𝜃𝑥
by: 𝑃 𝑋 = 𝑥 = 𝑒 −𝜃
𝑥!
- Is the statistic 𝑇 = σ𝑛𝑖=1 𝑋𝑖 sufficient for the parameter 𝜃?

[email protected]
Solution
First we determine the distribution of 𝑇:
𝑥
−𝜃 𝜃
𝑋 ↝ 𝑃(𝜃) ⟶ 𝑃 𝑋 = 𝑥 = 𝑒
𝑥!
Since the Poisson distribution is stable, then 𝑇 = 𝑡σ𝑛𝑖=1 𝑋𝑖 ↝ 𝑃(𝑛𝜃). Then the distribution of the
(𝑛𝜃)
statistic 𝑇 will be: 𝑔 𝑡, 𝜃 = 𝑃 𝑇 = 𝑡 = 𝑒 −𝑛𝜃 , avec 𝑡 = σ𝑛𝑖=1 𝑥𝑖 .
𝑛 𝑡!

𝐿(𝑥1 , 𝑥2 , … , 𝑥𝑛 ; 𝜃) = ෑ 𝑃 𝑥𝑖 , 𝜃
𝑖=1
𝑛 −𝜃 𝜃 𝑥𝑖
= ς𝑖=1 𝑒
𝑥𝑖 !
σ𝑛
−𝜃𝑛 𝜃 𝑖=1 𝑥𝑖
= 𝑒 ς𝑛
𝑖=1 𝑥𝑖!
−𝜃𝑛 𝜃𝑡
= 𝑒 ς𝑛
𝑖=1 𝑥𝑖!
𝑒 −𝜃𝑛 (𝑛𝜃)𝑡 𝑡!
= ×
𝑡! 𝑛𝑡 (ς𝑛
𝑖=1 𝑥𝑖 !)
= 𝑔 𝑡; 𝜃 ℎ 𝑥
According to the factorisation principle, 𝑇 is exhaustive for 𝜃.
[email protected]
6- Fisher information

The amount of Fisher information 𝐼𝑛 𝜃 , provided by a sample on the parameter 𝜃 is


expressed as follows:
2
𝜕𝑙𝑛 𝐿 𝑥; 𝜃
𝐼𝑛 𝜃 = 𝐼𝑥1 ,…,𝑥𝑛 𝜃 = 𝐸
𝜕𝜃

Theorem 3
If the domain of definition of the random variable 𝑋 does not depend on 𝜃 then:
𝜕 2 𝑙𝑛 𝐿 𝑥; 𝜃
𝐼𝑛 𝜃 = −𝐸
𝜕𝜃 2

[email protected]
Example 5
let 𝑋1 , 𝑋2 , … , 𝑋𝑛 an n-sample from a RV 𝑋 with the normal distribution 𝑁(𝜇, 𝜎 2 ).
-Find the Fisher information provided by the sample on the parameter 𝜇?
Solution :
𝜕2 ln 𝐿 𝑥;𝜇,𝜎 2
The domain of 𝑋 does not depend on 𝜇, so: 𝐼𝑛 𝜇 = −𝐸
𝜕𝜇 2
1 1
2 − 2 σ𝑛𝑖=1(𝑥𝑖 −𝜇)2
𝐿 𝑥; 𝜇, 𝜎 = ( )𝑒 2𝜎
𝜎 2𝜋 𝑛
1
𝑙𝑛 𝐿 𝑥; 𝜇, 𝜎 = −𝑛 𝑙𝑛 𝜎 2𝜋 − 2 ෍(𝑥𝑖2 − 2𝑥𝑖 𝜇 + 𝜇2 )
2
2𝜎
𝑛 𝑖=1
𝜕𝑙𝑛 𝐿 𝑥; 𝜇, 𝜎 2 −1
= 2 ෍(−2𝑥𝑖 + 2𝜇)
𝜕𝜇 2𝜎
2 2𝑖=1
𝜕 𝑙𝑛 𝐿 𝑥; 𝜇, 𝜎 −𝑛
2
= 2
𝜕𝜇 𝜎
−𝑛 𝑛
𝐼𝑛 (𝜇) = −𝐸 2 = 2
𝜎 𝜎
[email protected]
Property 1

If the domain of defining the random variable 𝑋 does not depend on 𝜃 then:
𝐼𝑛 𝜃 = 𝑛 𝐼𝑋 (𝜃)
Example 6
let 𝑋1 , 𝑋2 , … , 𝑋𝑛 an n-sample from a RV 𝑋 with the normal distribution 𝑁(𝜇, 𝜎 2 ).
-Find the Fisher information provided by random variable 𝑋 on the parameter 𝜇?
-Find the Fisher information provided by the sample on the parameter 𝜇?
Solution
𝜕2 𝑙𝑛 𝑓 𝑥;𝜇,𝜎2
𝐼𝑋 𝜇 = −𝐸
𝜕𝜇2
(𝑥−𝜇)2
𝑙𝑛 𝑓 𝑥; 𝜇, 𝜎 2 = − 𝑙𝑛 𝜎 2𝜋 −
2𝜎2
𝜕𝑙𝑛 𝑓 𝑥;𝜇,𝜎2 −1 𝑥−𝜇
= 2 −2𝑥𝑖 + 2𝜇 = 2
𝜕𝜇 2𝜎 𝜎
2
𝜕 𝑙𝑛 𝑓 𝑥;𝜇,𝜎 2 −1
=
𝜕𝜇2 𝜎2
−1 1
𝐼𝑋 (𝜇) = −𝐸 2 = 2
𝜎 𝜎
1 𝑛
The Fisher information provided by the sample is then: 𝐼𝑛 𝜇 = 𝑛 𝐼𝑋 (𝜇) = 𝑛 × =
𝜎2 𝜎2
[email protected]
Note 4
Let 𝐼𝑛 𝜃 be the Fisher information contributed by the sample and 𝐼𝑇 𝜃 by the
statistic.
If 𝐼𝑛 𝜃 = 𝐼𝑇 𝜃 ⟹ 𝑇 is sufficient for 𝜃.

Example 7
let 𝑋 a RV with uniform distribution on 0, 𝜃 , and 𝑋1 , 𝑋2 , … , 𝑋𝑛 an n-sample from 𝑋
1 𝑥
with density 𝑓 𝑥 = , 𝑥 ∈ [0, 𝜃], and distribution function 𝐹 𝑥 = , 𝑥 ∈ [0, 𝜃].
𝜃 𝜃
1-Find the Fisher information provided by an n-sample on the parameter 𝜃?
2-Let the order statistic 𝑇 = 𝑠𝑢𝑝 𝑋𝑖 , calculate the Fisher information carried by 𝑇 .
3-Is the statistic 𝑇 sufficient for 𝜃.

[email protected]
Solution:
1-The Fisher information provided by an n-sample on the parameter 𝜃:
𝜕ln 𝐿 𝑥;𝜃 2
The domain of 𝑋 depends on the parameter 𝜃. Therefore: 𝐼𝑛 𝜃 = 𝐸
𝜕𝜃
𝑋 ↷⊔[0,𝜃]
1
𝑓 𝑥; 𝜃 = , 𝑥 ∈ [0, 𝜃].
𝜃

Therefore, the likelihood is:


1
𝐿 𝑥 ; 𝜃 = 𝜃𝑛, 𝑥 ∈ [0, 𝜃],
ln 𝐿 𝑥 ; 𝜃 = −𝑛 ln 𝜃 , 𝑥 ∈ [0, 𝜃],
𝜕 ln 𝐿 𝑥 ;𝜃 −𝑛
= , 𝑥 ∈ [0, 𝜃],
𝜕𝜃 𝜃

Therefore,
−𝑛 2 −𝑛 2
𝐼𝑛 𝜃 = 𝐸 =𝐸
𝜃 𝜃
𝜃 𝜃 𝜃 𝜃 𝜃
2 2
𝑛 2 𝑛 1 𝑛 1 𝑛2 1 𝑛 𝑛2
𝐼𝑛 𝜃 = න … න 𝐿(𝑥, 𝜃) 𝑑𝑥1 … 𝑑𝑥𝑛 = න … න 2 𝑛 𝑑𝑥1 … 𝑑𝑥𝑛 = 2 𝑛 න 1 𝑑𝑥1 … 𝑑𝑥𝑛 = 2 𝑛 𝜃 = 2
𝜃 𝜃 𝜃 𝜃 𝜃 𝜃 𝜃 𝜃
0 0 0 0 0
[email protected]
2-The Fisher information carried by 𝑇 = 𝑠𝑢𝑝 𝑋𝑖 :
Calculating the distribution 𝑔 𝑡; 𝜃 of the statistic 𝑇 = 𝑠𝑢𝑝 𝑋𝑖 :
𝐺 𝑡; 𝜃 = 𝑃(𝑇 ≤ 𝑡) = 𝑃(𝑠𝑢𝑝 𝑋𝑖 ≤ 𝑡) = (𝐹𝑋 (𝑡))𝑛
𝑡 𝑛−1 1 𝑛 𝑛−1
𝑔 𝑡; 𝜃 = 𝑛(𝐹𝑋 (𝑡))𝑛−1 × 𝑓𝑋 (𝑡) =𝑛 = 𝑡
𝜃 𝑛−1 𝜃 𝜃𝑛
𝑛 𝑛−1
𝑙𝑛 𝑔 𝑡; 𝜃 = ln 𝑛 𝑡 = 𝑙𝑛 𝑛 − 𝑛 𝑙𝑛 𝜃 + 𝑛 − 1 𝑙𝑛 𝑡
𝜃
𝜕𝑙𝑛𝑔 𝑡 ; 𝜃 −𝑛
=
𝜕𝜃 𝜃

𝑛2
3- 𝐼𝑛 𝜃 = 𝐼𝑇 𝜃 = ⟹ the statistic is sufficient
[email protected] for 𝜃.
𝜃2
Fréchet-Darmois-Cramér-Rao (FDCR) bound

Theorem4
If the domain of 𝑋 does not depend on the parameter to be estimated for any
unbiased estimator 𝑇 of 𝜃 we have:
1
𝑉𝑎𝑟 𝜃 𝑇 ≥ = 𝐵𝐹 (𝜃)
𝐼𝑛 𝜃
The quantity 𝐵𝐹 𝜃 is the Fréchet-Darmois-Cramer-Rao (FDCR) lower bound.

Efficient estimator
An unbiased estimator 𝑇 is said to be efficient if its variance is equal to the lower
bound of 𝐹𝐷𝐶𝑅 :
1
𝑉𝑎𝑟 𝜃 𝑇 = 𝐵𝐹 𝜃 =
𝐼𝑛 𝜃

[email protected]
Example 8
Let 𝑋1 , 𝑋2 , … , 𝑋𝑛 an n-sample from a RV 𝑋 with the normal distribution 𝑁(𝜇, 𝜎 2 ).
1
- The statistic 𝑇 = σ𝑛𝑖=1 𝑋𝑖 an estimator of 𝜇. Is 𝑇 efficient ?
𝑛
Solution 𝑛
1 𝜎2
𝑉𝑎𝑟 𝑇 = 𝑉𝑎𝑟 ෍ 𝑋𝑖 =
𝑛 𝑛
𝑖=1
𝑛
𝐼𝑛 (𝜇) = 2
𝜎
1 𝜎2
𝐵𝐹 𝜇 = = = 𝑉𝑎𝑟 𝑇 ⟹ 𝑇 is efficient for 𝜇.
𝐼𝑛 𝜇 𝑛

[email protected]
7- Complete Statistics
Exponential family
The distribution of 𝑋 is said to be in the exponential family if its density 𝑓(𝑥, 𝜃) is
written in the form :
𝑓 𝑥, 𝜃 = 𝐶(𝜃)𝑒 𝑞 𝜃 𝑇(𝑥) ℎ(𝑥)
Example 9
Let 𝑋 ↷ 𝑃(𝜃)
−𝜃 𝜃𝑥
𝑃 𝑋=𝑥 =𝑒
𝑥!
𝑥 1
= 𝑒 −𝜃 𝑒 𝑙𝑛(𝜃 )
𝑥!
1
= 𝑒 −𝜃 𝑒 𝑥𝑙𝑛(𝜃)
𝑥!
1
𝐶 𝜃 =𝑒 −𝜃
, 𝑞 𝜃 = 𝑙𝑛 𝜃 , 𝑇 𝑥 = 𝑥 and ℎ 𝑥 = .
𝑥!
The Poisson distribution belongs to the exponential family.

[email protected]
Theorem5 (DARMOIS)
If the domain of 𝑋 does not depend on 𝜃 and the distribution of 𝑋 belongs to the
exponential family, then the statistic 𝑇 = σ𝑛𝑖=1 T(𝑥𝑖 ) is sufficient and complete statistic
for 𝜃.

Example 10
Let 𝑋 ↷ 𝑃 𝜃 .
- Is 𝑇 = σ𝑛𝑖=1 𝑥𝑖 a complete statistic for 𝜃?
Solution
The Poisson distribution belongs to the exponential family (see Example 9)
The domain of 𝑋 does not depend on 𝜃 and its distribution belongs to the exponontiel
family,
So according to DARMOIS' theorem 𝑇 = σ𝑛𝑖=1 T(𝑥𝑖 ) = σ𝑛𝑖=1 𝑥𝑖 is sufficient and complete
statistic for 𝜃.

[email protected]
8-MINIMUM VARIANCE UNBIASED estimator (MVUE)

Theorem6 (Rao-Blackwell-Lehman-Scheffe)
Let 𝑋 be a random variable of probability density 𝑓(𝑥, 𝜃). Let 𝑇 be a sufficient and complete
statistic for 𝜃, then for every estimable 𝛾 𝜃 there exists a unique MVUE of 𝛾 𝜃 based on 𝑇.
Example 11
Let 𝑋1 , 𝑋2 , … , 𝑋𝑛 an n-sample from 𝑋 ↷ 𝑃 𝜃 .
1-Find a sufficient and complete statistic for 𝜃.
2-Find the MVUE of 𝜃.
Solution
1- Sufficient and complete statistic for 𝜃 :
From previous example 𝑇 = σ𝑛𝑖=1 𝑥𝑖 is a complete and sufficient statistic.
2-MVUE of 𝜃:
𝑇 is a complete and sufficient statistic for 𝜃, it still remains to check that 𝑇 is unbiased.
We have 𝑇 = σ𝑛𝑖=1 𝑥𝑖 ,
𝐸 𝑇 = 𝐸 σ𝑛𝑖=1 𝑥𝑖 = 𝑛 𝐸 𝑥 = 𝑛𝜃 ⟹ 𝑇 is with bias.
1 σ𝑛
𝑖=1 𝑥𝑖
𝑇′ = 𝑇 = is the 𝑀𝑉𝑈𝐸 of 𝜃.
𝑛 𝑛
[email protected]
9-Estimation by confidence interval
Confidence interval estimation of a parameter 𝜃 consists of associating with a sample a random
interval 𝐼 chosen in such a way that the probability of it containing the unknown value of the
parameter is equal to a number fixed in advance, as large as desired. We write: 𝑃 𝜃 ∈ 𝐼 = 1 − 𝛼
1 − 𝛼: is the probability associated with the interval of containing the true value of the
parameter; this is the confidence level or quasi-certainty.
9-1-Construction principal
Let 𝑇 be an estimator of 𝜃 (best possible estimator) whose distribution as a function of 𝜃 is
known. We then determine a probability interval of level 1 − 𝛼 for 𝑇 i.e.:
𝑃(𝑡1 𝜃 ≤ 𝑇 ≤ 𝑡2 𝜃 ) = 1 − 𝛼
We then need to invert this interval for 𝑇, whose bounds depend on 𝜃 to obtain an interval for 𝜃,
i.e.:
𝑃(𝑇1 ≤ 𝜃 ≤ 𝑇2 ) = 1 − 𝛼
𝛼
𝑃(𝜃 < 𝑇1 ) =
2
𝛼
𝑃(𝜃 > 𝑇2 ) =
2

[email protected]
9-2-Principle of construction Confidence interval for the mean of a normal
distribution

❑𝝈𝟐 is known
Example 12. 𝑋1 , 𝑋2 , … , 𝑋𝑛 an n-sample drawn from a normal distibution 𝑋 ↷ 𝑁 𝜇, 𝜎 2 ,
with the variance 𝜎 2 is known.
- Find a confidence interval for the mean 𝜇 .
• We know that 𝑇 = 𝑋ത is the best estimator of 𝜇.
1 𝑛 1
𝐸 𝑇 =𝐸 σ 𝑋 = 𝑛 𝐸 𝑋𝑖 =𝜇
𝑛 𝑖=1 𝑖 𝑛
1 𝑛 1 𝜎2
𝑉 𝑇 =𝑉 σ 𝑋 = 2 𝑛 𝑉 𝑋𝑖 =
𝑛 𝑖=1 𝑖 𝑛 𝑛
𝛼
𝑃 𝑇 < 𝑇1 = 𝑃 𝑋ത < 𝑇1 =
2

[email protected]
As the distribution function of the normal distribution is only tabulated for the standardized
𝜎2 𝑇−𝜇
variable, we will therefore standardize the variable 𝑇. We obtain: 𝑇 ↷ 𝑁(𝜇, ) ⟹ 𝜎 ↷ 𝑁(0,1)
𝑛
𝑛
𝑇−𝜇 𝑇1 −𝜇 𝛼 𝑇1 −𝜇 𝛼 𝑇1 −𝜇 𝛼 𝑇1 −𝑋ത 𝛼
𝑃 𝜎 < 𝜎 = ⟹𝑃 𝑍 < 𝜎 = ⟹∅ 𝜎 = ⟹ 𝜎 = ∅−1
2 2 2 2
𝑛 𝑛 𝑛 𝑛 𝑛
𝑇1 −𝑋ത 𝛼
⟹ 𝜎 = −∅−1 1 −
2
𝑛
𝜎 𝛼
⟹ 𝑇1 = 𝑋ത − ∅−1 1 −
𝑛 2
𝛼 𝛼 𝑇−𝜇 𝑇2 −𝜇 𝛼 𝑇2 −𝑋ത 𝛼
𝑃 𝑇 > 𝑇2 = ⟹ 𝑃 𝑇 < 𝑇2 = 1 − ⟹𝑃 𝜎 < 𝜎 =1 − ⟹ 𝑃 𝑍< 𝜎 =1−
2 2 2 2
𝑛 𝑛 𝑛

𝑇2 −𝑋ത 𝛼
⟹∅ 𝜎 =1−
2
𝑛
𝑇2 −𝑋ത 𝛼
⟹ 𝜎 = ∅−1 1 −
2
𝑛
𝜎 𝛼
⟹ 𝑇2 = 𝑋ത + ∅−1 1 −
𝑛 2
𝜎 −1 𝛼 𝜎 −1 𝛼

𝐶𝐼1−𝛼 (𝜇) = 𝑋 − ∅ 1− ത
;𝑋 + ∅ 1−
2
𝑛 [email protected] 𝑛 2
❑𝝈𝟐 is Unknown
Example 13 𝑋1 , 𝑋2 , … , 𝑋𝑛 an n-sample drawn from a normal distibution 𝑋 ↷ 𝑁 𝜇, 𝜎 2 , with the variance
𝜎 2 is unknown.
- Find a confidence interval for the mean 𝜇 .

1
X = σni=1 Xi
• The mean of a sample: ഥ
n
1
• The unbiased variance of the sample: 𝑆 ′2 = σ𝑛𝑖=1(𝑋𝑖 ത 2
− 𝑋)
𝑛−1

1
Note 5 we use the biased empirical variance S = 𝑛 σ𝑛𝑖=1(𝑋𝑖 − 𝑋)
ത 2 in theory and mathematical models,
1
and the unbiased empirical variance 𝑆 ′2 = 𝑛−1 σ𝑛𝑖=1(𝑋𝑖 − 𝑋)
ത 2 in estimation from sample data.


𝑋−𝜇
• Change of variable: 𝑆′
↷ 𝑡𝑛−1
𝑛
• Because:
𝜎2
The mean of a sample: 𝑋ത ↷ 𝑁 𝜇,
n
The unbiased variance of the sample: 𝑆 ′2 ↷ 𝜒 2
⟹The standardized ratio of these two independent quantities gives a student distribution.
[email protected]

𝑋−𝜇
𝑃 −𝑡 ≤ 𝑆′
≤𝑡 =1−𝛼
𝑛
𝑆′ 𝑆′
𝑃 𝑋ത − 𝑡 ≤ 𝜇 ≤ 𝑋ത + 𝑡 =1−𝛼
𝑛 𝑛

Where: 𝑡 = 𝑡 𝛼
𝑛−1,1−
⟹𝑡=𝑡 𝑛−1,
𝛼
2 2
𝑡=𝑡 𝑛−1,1− 2
𝛼 ⟹𝑡=𝑡 𝑛−1, 2
𝛼 𝑤𝑖𝑙𝑙 𝑏𝑒 𝑡𝑎𝑘𝑒𝑛 𝑓𝑟𝑜𝑚 𝑡ℎ𝑒 𝑆𝑡𝑢𝑑𝑒𝑛𝑡′𝑠 𝑑𝑖𝑠𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛 𝑡𝑎𝑏𝑙𝑒.

Note 6 The symmetry of STUDENT's distribution gives the following property: 𝑡(𝑛,1−𝛼) = −𝑡 𝑛,𝛼 ).

𝑆′
𝑇1 = 𝑋ത − 𝑡 𝛼
𝑛 𝑛−1, 2
𝑆′
𝑇2 = 𝑋ത + 𝑡 𝛼
𝑛 𝑛−1, 2
𝑆′ 𝑆′
𝐶𝐼1−𝛼 (𝜇) = 𝑋ത − 𝑡 𝛼
𝑛−1, 2
; 𝑋ത + 𝑡 𝛼
𝑛−1, 2
𝑛 𝑛
[email protected]
[email protected]

You might also like