STA4C04 – STATISTICAL INFERENCE AND QUALITY CONTROL

Module 1
ESTIMATION OF PARAMETERS

Statistical inference
The process of making inferences about the unknown aspects of the population
based on samples is called statistical inference.

Statistical inference can be classified into two

→Estimation of parameters
→Testing of hypothesis

Estimation of parameters
The process of estimating the unknown population parameters using sample
observations is called estimation.
There are two types of estimation
→Point estimation
→Interval estimation

Point estimation
Consider a population having a pdf f(x,θ) where θ is the unknown population
parameter. This parameter θ can take any value on a set Θ.
The set Θ which is the set of all possible values of θ is called parameter space.
So the pdf of the population can be written as f(x,θ) , θ є Θ.
Take a random sample X1, X2 ... Xn from the population and compute a
function of these sample observations. If we are finding a single value to
estimate the unknown population parameter using the sample observations ,
then the single value is called the point estimate and the process of finding this
point estimate is called point estimation.
Estimator and Estimate
The function of sample observations (statistic) used to estimate the unknown
population parameter is called the estimator of the population parameter and the
value taken by the estimator when applied to a numerical sample is called
estimate of the population parameter.
Example :- If we are using the sample mean X̅ = 25 to estimate the unknown
population mean μ , then X̅ is called the estimator of μ and 25 is called the
estimate of μ.

Desirable properties of a good estimator

1) Unbiasedness
2) Consistency
3) Efficiency
4) Sufficiency

Unbiasedness
Consider a population having a pdf f(x,θ) . Suppose the parameter θ is unknown
. Take a random sample X1, X2 ... Xn from the population and compute an
estimator t for θ . The estimator t is said to be unbiased for θ if E(t) = θ
If E(t) ≠ θ , then the estimator t is a biased estimator and the bias is given by
E(t) – θ .

Unbiased estimators – problems

1) Show that sample mean X̅ is an unbiased estimator of population mean when
samples are taken from a normal population N(μ , σ)
Solution :
When samples of taken from a normal population N(μ , σ) ,
The sample mean X̅ → N(μ , σ/√n)
➔ E(X̅) = μ
➔ X̅ is an unbiased estimator for μ.
2) Show that sample mean is an unbiased estimator of the population mean.
Solution
Let X1,X2, ....Xn be a random sample taken from a population having
population mean μ .
Then E(X1) = μ , E(X2) = μ ......E(Xn) = μ
𝛴𝑋ᵢ
The sample mean X̅ = 𝑛
𝛴𝑋ᵢ
E(X̅) = E( 𝑛 )
1
= 𝑛 𝐸(𝛴𝑋ᵢ)
1
= 𝑛 𝐸X1+ X2+.....+ Xn)
1
= 𝑛 [𝐸X1)+E(X2)+.....+E(Xn)]
1
= 𝑛 (𝜇 + 𝜇 + ⋯ + 𝜇)
1
= 𝑛 𝑛𝜇

E(X̅) = μ
Therefore the sample mean X̅ is an unbiased estimator of population mean μ.

3) Check whether the sample variance s² is an unbiased estimator of population

variance σ² when samples are taken from normal population N(μ , σ) . If not
find a no unbiased estimator for population variance σ².
Solution
𝑛𝑠 2
When samples are taken from in normal population N(μ , σ) , 𝜎2 →𝛘² (n-1)
𝑛𝑠 2
E( ) = n -1
𝜎2
𝑛
E(s 2 ) = n -1
𝜎2
𝑛−1
E(s²) = 𝜎2
𝑛

Therefore dissemble variance s² is not an unbiased estimator of the population

variance σ².
We know that
𝑛−1
E(s²) = 𝜎2
𝑛
𝑛
E(s²) = σ²
𝑛−1
 𝑛
E(𝑛−1 𝑠 2 ) = σ²
𝑛
Therefore 𝑠 2 is an unbiased estimator of σ².
𝑛−1

4 ) If unbiased estimator of θ. Show that

1) t² is a biased estimator of θ²
2) √t is a biased estimator of √θ
Solution
1) Given t is an unbiased estimator of θ.
Then E(t) = θ
Now consider V(t) ≥ 0 [since variance is a squared quantity]
ie, E(t²) – {E(t)}² ≥ 0
E(t²) - θ² ≥ 0 [since E(t) = θ]
E(t²) ≥ θ²
→ t² is a biased estimator of θ.

2) Consider V(√t) ≥ 0
E(√t)² - {E(√t)}² ≥ 0
E(t) - {E(√t)}² ≥ 0
θ -{E(√t)}² ≥ 0 [since E(t) = θ]
θ ≥ {E(√t)}²
{E(√t)}² ≤ θ
Taking square root
E(√t) ≤ √ θ
→√t is a biased estimator of θ.

5) Find an unbiased estimator of a parameter θ of a uniform distribution

(rectangular distribution) over the interval [0, θ] based on a random sample of
size n.
Solution
The pdf of a rectangular distribution over the interval [a,b] is given by
1
f(x) = 𝑏−𝑎 , a ≤ x ≤ b

The pdf of a rectangular distribution over the interval [0 , θ] is given by

1
f(x) = 𝜃 , 0 ≤x ≤ θ

E(X) = θ/2 [do yourself]

Take a random sample X1,X2, ....Xn from this population with mean θ/2
Then E(X1) = θ/2 , E(X2) = θ/2 ......E(Xn) = θ/2
The sample mean X̅ = ΣXᵢ/n
E(X̅) = E(ΣXᵢ/n)
= 1/n E(ΣXᵢ)
= 1/n E(X1+ X2+.....+ Xn)
= 1/n[EX1)+E(X2)+.....+E(Xn)]
= 1/n(θ/2+ θ/2+⋯+ θ/2)
1 𝑛𝜃
=𝑛 = θ/2
2

2E(X̅) = θ
E(2X̅) = θ
Therefore 2X̅ is an unbiased estimator of θ.

Consistency
Consider a population having the pdf f(x, θ). Suppose population parameter θ is
unknown. Then take a random sample X1, X2...Xn from the population and
compute an estimator t for θ. The estimator t is said to be consistent for θ , if for
any є>0, however small
P{|t - θ| > є}→ 0 , as n→ ∞
Or P{|t - θ| < є}→ 1 , as n→ ∞
That is t tends to θ in probability.
Sufficient condition for consistency
An estimator t is said to be consistent for the population parameter θ , if
1) E(t) → θ and
2) V(t) → 0 , as n →∞

Q) Check whether sample mean X̅ is consistent for population mean μ , when

samples are taken from N( μ , σ)
Solution
Let X1, X2...Xn be a random sample taken from N( μ , σ)
Then X̅ → N( μ , σ/√n)
➔ E(X̅)= μ and SD(X̅) = σ/√n
V(X̅) = σ²/n
As n→∞ , E(X̅) → μ and V(X̅) → 0
Therefore X̅ is consistent for μ when samples are taken from N( μ , σ)

Q) Check whether sample variance s² is consistent for population variance σ²

and samples are taken from N( μ , N( μ , σ))
Solution
𝑛𝑠 2
We know that → 𝛘²(n-1)
𝜎2

ns2
E( σ2 ) = n-1
n
E(s 2 )= n-1
σ2
n
E(s²) = n−1σ²
1
E(s²) = (1 − n) σ²
1
As n →∞ , 1 − n → ,0 , therefore E(s²) → σ²
𝑛𝑠 2
V( 𝜎2 ) = 2(n-1)
𝑛2
V(s²) =2(n-1)
𝜎4
2(𝑛−1)
V(s²) = σ⁴
𝑛2
 1 1
V(s²) = 2(𝑛 − 𝑛2 ) σ⁴
1 1
As n→∞ , − →0 , therefore V(s²) →0
𝑛 𝑛2

That is sample variance s² is a consistent estimator of population variance σ²

Note :- Sample variance s² is not an unbiased estimator of σ²

but is a consistent estimator of σ².

Consistent estimator - Problem

Q) If t is a consistent estimator of θ , show that t2 is a consistent estimator of θ2,
Solution
Given t is a consistent estimator of θ , then
P{|t – θ| < ε } → 1 as n →
P{- ε < t – θ < ε } → 1 as n →
Adding throughout by θ
P{θ - ε < t < θ +ε } → 1 as n →
Squaring
P{(θ – ε)2 < t2 < (θ +ε)2 } → 1 as n →
P{θ2 +ε2 -2θε < t2 < θ2 +ε2+ 2θε } → 1 as n →
Subtracting throughout by θ2
P{ ε2 - 2θε < t2 - θ2 < ε2 + 2θε } → 1 as n →
Adding -2ε2 (a very small number) to the left hand term of inequality
P{ ε2 - 2θε -2ε2 < t2 - θ2 < ε2 + 2θε } → 1 as n →
P{ -ε2 - 2θε < t2 - θ2 < ε2 + 2θε } → 1 as n →
P{ -(ε2 + 2θε) < t2 - θ2 < (ε2 + 2θε) } → 1 as n →
Put ε2 + 2θε = ε1
P{ -ε1 < t2 - θ2 < ε1} → 1 as n →
P{ | t2 - θ2|< ε1}→ 1 as n →
➔ t2 is consistent for θ2 .

Efficiency
Consider a population having the pdf f(x, θ) . Suppose the parameter θ is
unknown. Then we take a random sample X₁,X₂,....Xn from the population and
compute two unbiased estimators t1 and t2 for θ
If V(t1) < V(t2) then t1 is said to be more efficient than t2 .
The relative efficiency of is given by V(t1) / V(t2) .

Q) Let X₁,X₂,X₃ be three independent observations from a population with mean

μ and variance σ² and if t₁ = X₁+X₂-X₃ and t₂ = 2X₁+3X₂-4X₃ are two
estimators of μ , compare their efficiencies.
Solution
Given E(X₁) = μ V(X₁) = σ²
E(X₂) = μ V(X₂) = σ²
E(X₃) = μ V(X₃) = σ²
Checking Unbiasedness
E(t₁) = E(X₁+X₂-X₃) = E(X₁)+E(X₂)-E(X₃) = μ+μ-μ = μ
E(t₂) = E(2X₁+3X₂-4X₃) = 2E(X₁)+3E(X₂)-4E(X₃) = 2μ+3μ-4μ = μ
➔ t₁ and t2 are unbiased estimators for μ
Checking efficiency
V(t₁) = V(X₁+X₂-X₃) = V(X₁)+V(X₂)+V(X₃) = σ²+ σ² + σ² = 3 σ²
V(t₂) = V(2X₁+3X₂-4X₃) = 4V(X₁)+9V(X₂)+16V(X₃) = 4σ²+9σ²+16σ² = 29 σ²
Here V(t₁) < V(t₂) ➔ t₁ is more efficient than t₂
Sufficiency
Consider a population having the pdf f(x,θ) . Suppose the parameter θ is
unknown . Then take a random sample X1 , X2 , ...Xn from the population and
compute a statistic t for θ . The estimator t is said to be sufficient for θ , if t has
all the information about θ that is contained in the sample.

Likelihood function
The likelihood function of n random variables X1 , X2 , ...Xn is defined as the
joint pdf/pmf of the random variables.
ie , L(x1 , x2 , ...xn ) = f(x1 , x2 , ...xn )
If X1 , X2 , ...Xn are n observations of a random sample then the likelyhood
function is given by
L(x1 , x2 , ...xn, θ ) = f(x1 ,θ)f(x2 ,θ) ...f(xn ,θ) since X1 , X2 , ...Xn sre independent
observations being from a random sample.

Criteria for sufficiency

The estimator t is said to be sufficient for θ if we can factorise the likelyhood
function of the sample in to product of two functions of which one is a function
of t and θ and other is a function independent of θ.
ie L(x1 , x2 , ...xn, θ ) = L1(t,θ) L2((x1 , x2 , ...xn )
This is called Fisher Neyman factorisation theorem.

Problem
Q) Show that the sample mean 𝑋̅ is sufficient for estimating the parameter λ in
a poisson population
Solution
𝑒 −𝜆 𝜆𝑥
The pmf of the poisson population is given by f(x,λ) = , x = 0,1,2...
𝑥!

As the parameter λ is unknown , take a random sample X1 , X2 , ...Xn from the

population and compute an estimator t = 𝑋̅ for λ.
The sample observations X1 , X2 , ...Xn follows a poisson population with
parameter λ.
𝑒 −𝜆 𝜆𝑥₁ 𝑒 −𝜆 𝜆𝑥 ₂ 𝑒 −𝜆 𝜆𝑥ₙ
ie f(x1,λ) = , f(x₂,λ) = ,..... f(xₙ,λ) =
𝑥₁! 𝑥₂! 𝑥ₙ!

The likelyhood function of the sample is given by

L(x1 , x2 , ...xn, λ) = f(x1,λ) . f(x₂,λ) ...... f(xₙ,λ)
𝑒 −𝜆 𝜆𝑥₁ 𝑒 −𝜆 𝜆𝑥 ₂ 𝑒 −𝜆 𝜆𝑥ₙ
= . ......
𝑥₁! 𝑥₂! 𝑥ₙ!

(𝑒 −𝜆 )𝑛 𝜆(𝑥₁+𝑥₂+⋯𝑥ₙ)
=
𝑥₁!𝑥₂!….𝑥ₙ!

𝑒 −𝑛𝜆 𝜆∑ 𝑥ᵢ
= 𝑥₁!𝑥₂!….𝑥ₙ!

𝑒 −𝑛𝜆 𝜆𝑛𝑥̅
= 𝑥₁!𝑥₂!….𝑥ₙ!
1
= (𝑒 −𝑛𝜆 𝜆𝑛𝑥̅ ) 𝑥₁!𝑥₂!….𝑥ₙ!

= L1(𝑥̅ , λ) . L2(x1 , x2 , ...xn)

⸫ 𝑋̅ is sufficient for λ .

Q) Show that 𝑋̅ is sufficient estimator for p , when samples are taken from a
binomial population with parameters (m.p)
Solution
The pmf of the binomial population with parameters m and p if given by
f(x,m,p) = mCx pxqm-x , x = 0,1,2...m , 0<p<1 , p+q=1
As the population parameter p is unknown take a random sample X1 ,X2 , ....Xn
from the binomial population.
Then f(x1,m,p) = mCx₁px₁qm-x₁
f(x2,m,p) = mCx₂ px₂qm-x₂
.................................
f(xn,m,p) = mCxₙ pxₙqm-xₙ
The likelyhood function of the sample is given by
L(x1,x2 .....xn , m ,p) = f(x1,m,p) f(x2,m,p) ..... f(xn,m,p)
= mCx₁px₁qm-x₁ m
Cx₂ px₂qm-x₂ ...... mCxₙ pxₙqm-xₙ
= [mCx₁ mCx₂ .... mCxₙ ] px₁ px₂.... pxₙ qm-x₁ qm-x₂ .... qm-xₙ
= [mCx₁ mCx₂ .... mCxₙ ] 𝑝 ∑𝑖 𝑥ᵢ 𝑞 𝑛𝑚−∑𝑖 𝑥ᵢ
= [mCx₁ mCx₂ .... mCxₙ ] [𝑝 ∑𝑖 𝑥ᵢ (1 − 𝑝)𝑛𝑚−∑𝑖 𝑥ᵢ ]
=[𝑝 ∑𝑖 𝑥ᵢ (1 − 𝑝)𝑛𝑚−∑𝑖 𝑥ᵢ ][ mCx₁ mCx₂ .... mCxₙ ]
̅̅̅̅ ̅̅̅̅ m
= [𝑝 𝑛𝑥 (1 − 𝑝)𝑛𝑚−𝑛𝑥 ][ Cx₁ mCx₂ .... mCxₙ ]
= L1(𝑥̅ ,p) L2(x1 ,x2 , ....xn)
⸫ 𝑋̅ is sufficient estimator for p

Minimum variance unbiased estimator(MVUE)

Let X1,X2.....Xn be a random sample taken from a population with pdf f(x,θ) .
Let t be an estimator of θ such that
1) t is an unbiased estimator of θ , ie. E(t) = θ
2) V(t) is less than the variance of any other unbiased estimator of θ , ie
V(t) < V(t’) where t’ is an unbiased estimator of θ.

Then t is said to be the minimum variance unbiased estimator (MVUE) of θ.

Cramer-Rao inequality
If t is an unbiased estimator of γ(θ) , a function of θ , then
2 2
[𝛾′ (𝜃)] [𝛾′ (𝜃)] 𝜕𝑙𝑜𝑔𝐿 2 𝜕 2 𝑙𝑜𝑔𝐿
V(t) ≥ 𝜕𝑙𝑜𝑔𝐿 2
Or V(t) ≥ − 2 since 𝐸 ( ) = −𝐸 ( )
𝜕𝜃 𝜕𝜃2
𝐸( 𝜕𝜃 ) 𝐸(𝜕 𝑙𝑜𝑔𝐿 )
𝜕𝜃2

Cramer –Rao inequality provides a lower bound for the variance of an unbiased
estimator t of a function of population parameter γ(θ).
2 2
[𝛾′ (𝜃)] [𝛾′ (𝜃)]
The Cramer –Rao lower bound is given by 𝜕𝑙𝑜𝑔𝐿 2
or − 2
𝐸( 𝜕𝜃 ) 𝐸(𝜕 𝑙𝑜𝑔𝐿
2 )
𝜕𝜃
Note :- If t is an unbiased estimator of θ , that is γ(θ) = θ , then γ’(θ)=1 and
Cramer-Rao inequality can be written as
1 1
V(t) ≥ 𝜕𝑙𝑜𝑔𝐿 2
Or V(t) ≥ − 2
𝐸( ) 𝜕 𝑙𝑜𝑔𝐿
𝜕𝜃 𝐸( )
𝜕𝜃2

Minimum variance bound estimator (MVBE)

An unbiased estimator t of γ(θ) for which Cramer-Rao lower bound is attained
is called a minimum variance bound estimator.
That is t is a minimum variance bound estimator if
[𝛾′ (𝜃)]2 [𝛾′ (𝜃)]
2
V(t) = 𝜕𝑙𝑜𝑔𝐿 2
or V(t) = − 2
𝐸( 𝜕𝜃 ) 𝜕 𝑙𝑜𝑔𝐿
𝐸( )
𝜕𝜃2

Condition for an unbiased estimator to attain Cramer-Rao lower bound

A necessary and sufficient condition for an unbiased estimator to attain Cramer-
𝜕𝑙𝑜𝑔𝐿
Rao lower bound is that should be proportional to t-γ(θ)
𝜕𝜃
𝜕𝑙𝑜𝑔𝐿
That is = [t-γ(θ)] A(θ)
𝜕𝜃

Q) Obtain the minimum variance bound (MVB) estimator for μ in normal

population N(μ ,σ) , when σ is known
Solution
𝜕𝑙𝑜𝑔𝐿
Note : To find a MVB estimator for γ(θ) we have to write in the form
𝜕𝜃
[t- γ(θ)] A(θ) . Then we can say that t is a MVB estimator of γ(θ)
Given the population distribution is normal N(μ ,σ)
The pdf of the normal population is given by
 (𝒙−𝝁)𝟐
𝟏 −
f(x,μ,σ) = 𝒆 𝟐𝝈𝟐 , -<x<
𝝈√𝟐𝝅

To compute an estimator for μ , take a random sample X1,X2.....Xn from the

normal population
(𝒙₁−𝝁)𝟐
𝟏 −
Then f(x₁,μ,σ) = 𝒆 𝟐𝝈𝟐
𝝈√𝟐𝝅
(𝒙₂−𝝁)𝟐
𝟏 −
f(x₂,μ,σ) = 𝒆 𝟐𝝈𝟐
𝝈√𝟐𝝅

.................
(𝒙ₙ−𝝁)𝟐
𝟏 −
f(xₙ,μ,σ) = 𝒆 𝟐𝝈𝟐
𝝈√𝟐𝝅

The likelyhood function of the sample is given by

L(x1,x2.....xn ,μ,σ) = f(x₁,μ,σ) f(x₂,μ,σ) ...... f(xₙ,μ,σ)
)𝟐 (𝒙₂−𝝁)𝟐 (𝒙ₙ−𝝁)𝟐
− 𝒙₁−𝝁𝟐
(
𝟏 𝟏 − 𝟏 −
=
𝝈√𝟐𝝅
𝒆 𝟐𝝈 𝒆 𝟐𝝈𝟐 ....... 𝝈√𝟐𝝅 𝒆 𝟐𝝈𝟐
𝝈√𝟐𝝅
)𝟐 )𝟐 )𝟐
𝒏 − 𝒙₁−𝝁𝟐 +− 𝒙₂−𝝁𝟐 + ….+ 𝒙ₙ−𝝁𝟐
( ( (
𝟏
=(
𝝈√𝟐𝝅
) 𝒆 𝟐𝝈 𝟐𝝈 𝟐𝝈

𝟐
𝒏 ∑𝒏 (𝒙ᵢ−𝝁)
𝟏 − 𝒊=𝟏 𝟐
=( ) 𝒆 𝟐𝝈
𝝈√𝟐𝝅

Taking logarithm
𝟐
∑𝒏
𝒏 − 𝒊=𝟏(𝒙ᵢ−𝝁)
𝟏
logL = log [( ) 𝒆 𝟐𝝈𝟐 ]
𝝈√𝟐𝝅
𝟐
𝒏 ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
𝟏 −
= log [(𝝈√𝟐𝝅) ] + log[ 𝒆 𝟐𝝈𝟐 ]
𝟐
𝟏 ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
= nlog (𝝈√𝟐𝝅) + - 𝟐𝝈𝟐
logee
𝟐
𝟏 ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
= nlog (𝝈 )+ - (since logee =1)
√𝟐𝝅 𝟐𝝈𝟐

Differentiating with respect to μ

 𝟐
𝝏𝒍𝒐𝒈𝑳 𝝏 𝝏 ∑𝒏
𝟏 𝒊=𝟏(𝒙ᵢ−𝝁)
= ( n log ( )) +
𝝏𝝁 𝝏𝝁 𝝈√𝟐𝝅 𝝏𝝁 𝟐𝝈𝟐
𝟐
𝟏 𝝏 ∑𝒏 (𝒙ᵢ−𝝁)
=0 − 𝟐𝝈𝟐 𝒊=𝟏𝝏𝝁
𝟏
= − 𝟐𝝈𝟐 2 ∑𝒏𝒊=𝟏(𝒙ᵢ − 𝝁) ∗ −𝟏
∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
=
𝝈𝟐
∑𝒏 𝒏
𝒊=𝟏 𝒙ᵢ−∑𝒊=𝟏 𝝁
= 𝝈𝟐
∑𝒏
𝒊=𝟏 𝒙ᵢ−𝒏𝝁
=
𝝈𝟐
∑ 𝒏
𝒙ᵢ
𝒏[ 𝒊=𝟏 −𝝁]
𝒏
= 𝝈𝟐
𝒏
= ̅ –μ)
(𝒙
𝝈𝟐
𝝏𝒍𝒐𝒈𝑳 𝒏
̅ –μ) 𝟐
=(𝒙
𝝏𝝁 𝝈

̅
The RHS is of the form [t- γ(θ)] A(θ) where t = 𝒙

⸫𝒙
̅ is a MVB estimator of the parameter μ .

Q) A random sample X1,X2.....Xn is taken from a normal population with mean

∑𝒏 𝟐
𝒊=𝟏 𝒙ᵢ
0 and variance σ . Examine whether
2
is a MVB estimator for σ2 .
𝒏
Solution
Given the distribution of the population is N(0, σ2)
So the pdf of the population is given by
𝒙𝟐
𝟏 −
f(x, σ ) =
2
𝒆 𝟐𝝈𝟐 , -<x<
𝝈√𝟐𝝅

Put σ2 = θ
𝒙ᵢ𝟐
𝟏 − 𝟐𝜽
f(x, θ) = 𝒆
√𝟐𝝅𝜽
As the sample observations X1,X2.....Xn are taken from N(0, σ2) or N(0, θ)
𝒙₁𝟐 𝒙₂𝟐 𝒙ₙ𝟐
𝟏 − 𝟐𝜽 𝟏 − 𝟐𝜽 𝟏 − 𝟐𝜽
f(x₁, θ) = 𝒆 , f(x₂, θ) = 𝒆 ........ f(xₙ, θ) = 𝒆
√𝟐𝝅𝜽 √𝟐𝝅𝜽 √𝟐𝝅𝜽

The likelyhood function of the sample is given by

L(x1,x2.....xn ,θ) = f(x₁, θ). f(x₂, θ) ........ f(xₙ, θ)
𝒙₁𝟐 𝒙₂𝟐 𝒙ₙ𝟐
𝟏 − 𝟐𝜽 𝟏 − 𝟏 −
=
√𝟐𝝅𝜽
𝒆 . 𝒆 𝟐𝜽 ...... 𝒆 𝟐𝜽
√𝟐𝝅𝜽 √𝟐𝝅𝜽
𝒏 𝟐
𝒏 ∑ 𝒙ᵢ
𝟏 − 𝒊=𝟏
=( ) 𝒆 𝟐𝜽
√𝟐𝝅𝜽
∑𝒊=𝟏 𝒙ᵢ 𝒏 𝟐

= (𝟐𝝅𝜽)−𝒏/𝟐 𝒆− 𝟐𝜽
Taking logarithms
∑ 𝒙ᵢ𝒏 𝟐
− 𝒊=𝟏
log L = log [(𝟐𝝅𝜽)−𝒏/𝟐 𝒆 𝟐𝜽 ]
∑𝒏 𝒙ᵢ𝟐
− 𝒊=𝟏
= log(𝟐𝝅𝜽)−𝒏/𝟐 + log(𝒆 𝟐𝜽 )

𝑛 ∑𝒏
𝒊=𝟏 𝒙ᵢ
𝟐
= - 2 log(𝟐𝝅𝜽) + − logee
𝟐𝜽

𝑛 𝑛 ∑𝒏
𝒊=𝟏 𝒙ᵢ
𝟐
= − 2 log(𝟐𝝅) + − 2 log(𝜽) + − 𝟐𝜽
Differentiating wrt θ
𝝏𝒍𝒐𝒈𝑳 𝝏 𝑛 𝝏 𝑛 𝝏 ∑𝒏
𝒊=𝟏 𝒙ᵢ
𝟐
= 𝝏𝜽 (− 2 log(𝟐𝝅)) +𝝏𝜽 (− 2 log(𝜽) + (− )
𝝏𝜽 𝝏𝜽 𝟐𝜽

𝒏 ∑𝒏
𝒊=𝟏 𝒙ᵢ
𝟐 𝟏
= 0 − 𝟐𝜽 − (− )
𝟐 𝜽𝟐

𝒏 ∑𝒏
𝒊=𝟏 𝒙ᵢ
𝟐
= − 𝟐𝜽 + 𝟐𝜽𝟐
𝟏
= 𝟐𝜽𝟐[-nθ+ ∑𝒏𝒊=𝟏 𝒙ᵢ𝟐 ]

𝒏 ∑𝒏
𝒊=𝟏 𝒙ᵢ
𝟐
= [-θ+ ]
𝟐𝜽𝟐 𝒏
 ∑𝒏
𝒊=𝟏 𝒙ᵢ
𝟐 𝐧
=[ − 𝜽] 𝟐𝛉𝟐
𝒏

∑𝒏 𝟐
𝒊=𝟏 𝒙ᵢ
The RHS is of the form [t- γ(θ)] A(θ) where t =
𝒏

∑𝒏 𝟐
𝒊=𝟏 𝒙ᵢ
⸫ is a MVB estimator of the parameter θ=σ2 .
𝒏

Completeness
Consider a statistic t calculated using a random sample of size n taken from a
population with pdf f(x, θ) . The distribution of t will depend on θ
.Corresponding to t we will have a family of distributions { g(t,θ) , θ ϵΘ}.
Then the statistic t or more precisely the family of distributions { g(t,θ) , θ
ϵΘ}.is said to be complete for θ if
E[h(t)] =0 for all θ ➔ P[h(t)=0]=1 → h(t)=0 almost surely.

Methods of estimation
Different methods of finding estimators of unknown population parameter are
1) Method of moments
2) Method of maximum likelyhood
3) Method of least squares
4) Method of minimum variance
Method of moments
Consider a population having pdf f(x,θ) .Suppose k parameters of the
population are unknown . Take a random sample X1,X2.....Xn from the
population. To estimate the k parameters , equate the first k raw moments of the
population to the first k raw moments of the sample .
The rth raw moment of the population is given by
μr’ = E(Xr)
The rth raw moment of the sample is given by
∑𝒏
𝒊=𝟏 𝑿ᵢ
𝒓
m r’ =
𝒏

Problems
1) Find the estimator of the parameter p of the binomial distribution B(m,p)
using the method of moments.
Solution
As the parameter p of the binomial population is unknown , take a
random sample X1,X2.....Xn from B(m,p) and equate the first moment of
the population to the first moment of the sample.

First moment of the population μ1’ = E(X) =mp

∑𝒏
𝒊=𝟏 𝑿ᵢ
First moment of the sample m1’ = ̅
=𝑿
𝒏

̅
Equating μ1’ and m1’➔ mp =𝑿
̅
𝑿
𝑝̂ = 𝑚 is the moment estimator of p.

2) Obtain the moment estimator of the parameter θ of the exponential

𝑥
1
distribution whose pdf is given by f(x,θ) = 𝜃 𝑒 −𝜃 , x>0
Solution
 Take a random sample X1,X2.....Xn from the exponential population and
equate the first moment of the population to the first moment of the
sample.
First moment of the population μ1’ = E(X) = θ
∑𝒏
𝒊=𝟏 𝑿ᵢ
First moment of the sample m1 =’ ̅
=𝑿
𝒏

̅
Equating μ1’ and m1’➔ θ =𝑿
→ 𝜃̂ =𝑿
̅

3) Estimate the parameters μ and σ2 of the normal population N(μ,σ) using

the method of moments

Solution
Take a random sample X1,X2.....Xn from the normal population N(μ,σ).
As two parameters of the population are unknown , equate the first two
moments of the population to the first two moments of the sample.

First moment of the population μ1’ = E(X) = μ

∑𝒏
𝒊=𝟏 𝑿ᵢ
First moment of the sample m1’ = ̅
=𝑿
𝒏

̅
Equating μ1’ and m1’➔ μ =𝑿
̅
→ 𝜇̂ =𝑿
Second moment of the population μ2’ = E(X2)
We know that variance of the normal population N(μ,σ) is
V(X) = σ2
→E(X2) – {E(X)}2 = σ2
E(X2) – μ2 = σ2
E(X2) = σ2 + μ2
ie μ2’ = σ2 + μ2
𝟐
’ ∑𝒏
𝒊=𝟏 𝑿ᵢ
Second moment of the sample m2 = 𝒏
𝟐
∑𝒏
𝒊=𝟏 𝑿ᵢ
Equating μ2’ and m2’➔ σ2 + μ2 = 𝒏
 𝒏 𝟐
̂ ∑ 𝑿ᵢ
𝝈𝟐 = 𝒊=𝟏𝒏 - μ2
̅
As μ is also unknown replace μ by its estimator 𝑿
𝒏 𝟐
̂ ∑ 𝑿ᵢ
̅ 2 = s2 , the sample variance
𝝈𝟐 = 𝒊=𝟏𝒏 - 𝑿

Method of maximum likelihood

The estimator obtained by the method of maximum likelihood is called
maximum likelihood estimator (MLE).
MLE is an estimator computed from a random sample for which the likelihood
of the sample is maximum. That is we have to use the sample X1,X2.....Xn to
estimate the unknown population parameter θ ,if the likelihood function
L(x1,x2.....xn ,θ) is maximum .
For L to be maximum, the first derivative of L with respect to θ ,
𝝏𝑳 𝝏𝟐 𝑳
= 0 and the second derivative 𝝏𝜽𝟐 < 0
𝝏𝜽

L is maximum means logL is maximum which means

𝜕𝑙𝑜𝑔𝐿 𝜕 2 𝑙𝑜𝑔𝐿
= 0 and <0
𝜕𝜃 𝜕𝜃2

Problems
𝑥
1
Q) Obtain the MLE for θ for a population having the pdf f(x,θ) = 𝜃 𝑒 −𝜃 , x > 0

Solution
Let X1,X2.....Xn be a random sample taken from a population having pdf
𝑥
1
f(x,θ) = 𝜃 𝑒 −𝜃 , x > 0 [ Exponential distribution with parameter 1/θ ]
𝑥₁ 𝑥₂ 𝑥ₙ
1 −𝜃 1 −𝜃 1 −𝜃
Then f(x₁,θ) = 𝜃 𝑒 , f(x₂,θ) = 𝜃 𝑒 , ........ f(xₙ,θ) = 𝜃 𝑒

Likelyhood function of the sample is given by

L(x1,x2.....xn ,θ) = f(x₁,θ). f(x₂,θ) ...... f(xₙ,θ)
 𝑥₁ 𝑥₂ 𝑥ₙ
1 1 1
= 𝜃 𝑒 − 𝜃 . 𝜃 𝑒 − 𝜃 ...... 𝜃 𝑒 − 𝜃
∑𝑛
𝑖=1 𝑥ᵢ
1 𝑛 −
= (𝜃) 𝑒 𝜃

∑𝑛
𝑖=1 𝑥ᵢ
= (𝜃)−𝑛 − 𝑒 𝜃

Taking logarithms
∑𝑛
𝑖=1 𝑥ᵢ
−𝑛 −
logL = log((𝜃) 𝑒 𝜃 )
∑𝑛
𝑖=1 𝑥ᵢ
−𝑛 −
= log((𝜃) +log (𝑒 𝜃 ))

∑𝑛
𝑖=1 𝑥ᵢ
= -n log(θ) - 𝜃
logee
∑𝑛
𝑖=1 𝑥ᵢ
= -n log(θ) - 𝜃

Differentiating wrt θ
𝜕𝑙𝑜𝑔𝐿 𝑛 ∑𝑛
𝑖=1 𝑥ᵢ
= −𝜃 +
𝜕𝜃 𝜃2

𝜕𝑙𝑜𝑔𝐿 𝑛 ∑𝑛
𝑖=1 𝑥ᵢ
= 0 ➔ −𝜃 + =0
𝜕𝜃 𝜃2
∑𝑛
𝑖=1 𝑥ᵢ 𝑛
=𝜃
𝜃2
∑𝑛
𝑖=1 𝑥ᵢ
=𝑛
𝜃
𝑛
∑ 𝑥ᵢ
𝜃̂ = 𝑖=1 = 𝑥̅
𝑛

Q) Find the MLEs for μ and σ2 using samples drawn from a normal distribution
N(μ,σ)
Solution
Given the population distribution is normal N(μ ,σ)
The pdf of the normal population is given by
(𝒙−𝝁)𝟐
𝟏 −
f(x,μ,σ) = 𝝈 𝒆 𝟐𝝈𝟐 , -<x<
√𝟐𝝅
To compute an estimator for μ , take a random sample X1,X2.....Xn from the
normal population
(𝒙₁−𝝁)𝟐
𝟏 −
Then f(x₁,μ,σ) = 𝝈 𝒆 𝟐𝝈𝟐
√𝟐𝝅
(𝒙₂−𝝁)𝟐
𝟏 −
f(x₂,μ,σ) = 𝝈 𝒆 𝟐𝝈𝟐
√𝟐𝝅

.................
(𝒙ₙ−𝝁)𝟐
𝟏 −
f(xₙ,μ,σ) = 𝝈 𝒆 𝟐𝝈𝟐
√𝟐𝝅

The likelyhood function of the sample is given by

L(x1,x2.....xn ,μ,σ) = f(x₁,μ,σ) f(x₂,μ,σ) ...... f(xₙ,μ,σ)
(𝒙₁−𝝁)𝟐 (𝒙₂−𝝁)𝟐 (𝒙ₙ−𝝁)𝟐
𝟏 − 𝟏 − 𝟏 −
=𝝈 𝒆 𝟐𝝈𝟐 𝒆 𝟐𝝈𝟐 ....... 𝝈 𝒆 𝟐𝝈𝟐
√𝟐𝝅 𝝈√𝟐𝝅 √𝟐𝝅

𝒏 (𝒙₁−𝝁)𝟐 (𝒙₂−𝝁)𝟐 (𝒙ₙ−𝝁)𝟐

𝟏 − +− + ….+
=( ) 𝒆 𝟐𝝈𝟐 𝟐𝝈𝟐 𝟐𝝈𝟐
𝝈√𝟐𝝅
𝟐
𝒏 ∑𝒏 (𝒙ᵢ−𝝁)
𝟏 − 𝒊=𝟏 𝟐
= (𝝈 ) 𝒆 𝟐𝝈
√𝟐𝝅
𝟐
𝒏 ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
𝟏 −
=( ) 𝒆 𝟐𝝈𝟐
√𝟐𝝅𝝈𝟐
𝟐
𝒏 ∑𝒏 (𝒙ᵢ−𝝁)
−𝟐 − 𝒊=𝟏 𝟐
L = (𝟐𝝅𝝈𝟐 ) 𝒆 𝟐𝝈

Taking logarithms
𝟐
𝒏 ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
𝟐 )− 𝟐 −
logL = 𝐥𝐨𝐠[(𝟐𝝅𝝈 𝒆 𝟐𝝈𝟐 ]
𝟐
𝒏 ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
𝟐 )− 𝟐 −
= 𝐥𝐨𝐠[(𝟐𝝅𝝈 ] + 𝐥𝐨𝐠[ 𝒆 𝟐𝝈𝟐 ]
𝟐
∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
𝒏 −
= − 𝟐 𝐥𝐨𝐠(𝟐𝝅𝝈𝟐 ) + 𝐥𝐨𝐠[ 𝒆 𝟐𝝈𝟐 ]
𝟐
∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
𝒏 𝒏 −
= − 𝟐 𝐥𝐨𝐠(𝟐𝝅) + (− 𝟐) 𝐥𝐨𝐠(𝝈𝟐 ) + 𝐥𝐨𝐠[ 𝒆 𝟐𝝈𝟐 ]
𝟐
𝒏 𝒏 𝟐) − ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
= − 𝟐 𝐥𝐨𝐠(𝟐𝝅) + (− 𝟐) 𝐥𝐨𝐠(𝝈 + logee
𝟐𝝈𝟐
 𝟐
𝒏 𝒏 𝟐) − ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
logL = − 𝟐 𝐥𝐨𝐠(𝟐𝝅) + (− 𝟐) 𝐥𝐨𝐠(𝝈 + 𝟐𝝈𝟐

To find MLE for μ

Differentiate logL wrt to μ
𝜕𝑙𝑜𝑔𝐿 𝟏
= 0+0+ − 𝟐𝝈𝟐 2 ∑𝑛ᵢ=1(𝑥ᵢ − 𝜇) (−1)
𝜕𝜇

𝜕𝑙𝑜𝑔𝐿 𝟏
= 𝝈𝟐 ∑𝑛ᵢ=1(𝑥ᵢ − 𝜇)
𝜕𝜇

𝜕𝑙𝑜𝑔𝐿
Equating =0
𝜕𝜇

𝟏
∑𝑛ᵢ=1(𝑥ᵢ − 𝜇) = 0
𝝈𝟐

∑𝑛ᵢ=1(𝑥ᵢ − 𝜇) = 0
∑𝑛ᵢ=1 𝑥ᵢ − 𝑛𝜇 = 0
∑𝑛ᵢ=1 𝑥ᵢ = 𝑛𝜇
∑𝒏
𝒊=𝟏 𝒙ᵢ
⸫ 𝜇̂ = ̅ is the MLE for μ
=X
𝒏

To find MLE for σ2

We have
𝟐
𝒏 𝒏 − ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
logL = − 𝟐 𝐥𝐨𝐠(𝟐𝝅) + (− 𝟐) 𝐥𝐨𝐠(𝝈𝟐 ) + 𝟐𝝈𝟐

Differentiating wrt σ2
𝟐
𝜕𝑙𝑜𝑔𝐿 𝒏 𝟏 − ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
𝜕𝜎2
= 0+(− 𝟐) (𝝈𝟐 ) - 𝟐𝝈𝟒
𝜕𝑙𝑜𝑔𝐿
Now =0
𝜕𝜎2
𝟐
𝒏 𝟏 − ∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
(− 𝟐) (𝝈𝟐 ) - =0
𝟐𝝈𝟒
𝟐
∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁) 𝒏
= 𝟐𝝈𝟐
𝟐𝝈𝟒
𝟐
∑𝒏
𝒊=𝟏(𝒙ᵢ−𝝁)
=n
𝝈𝟐
 𝒏 𝟐
̂𝟐 = ∑𝒊=𝟏(𝒙ᵢ−𝝁)
⸫𝝈 𝒏

̅
If μ is unknown replace μ by its MLE X
𝒏 𝟐
̂𝟐 = ∑𝒊=𝟏(𝒙ᵢ−X̅) =s2 , the sample variance
Then MLE for σ is given by 𝝈
2
𝒏

Q) Find the moment estimator and MLE for the parameter θ if the population
has the pdf f(x,θ) = (𝜃 + 1)𝑥 𝜃 , 0<x<1 .
Also find the values of the moment estimator and MLE when the sample
observations are given by
X : 0.2, 0.4, 0.8, 0.5, 0.7, 0.9, 0.8, 0.9
Solution
Given the population parameter θ is unknown .Then take a random sample X1,
X2 ,.... Xn from the population.
Here the random sample is given as X : 0.2, 0.4, 0.8, 0.5, 0.7, 0.9, 0.8, 0.9
1 5.2
𝑋̅ = ∑𝑛𝑖=1 𝑋ᵢ = = 0.65
𝑛 8

Moment estimator
To find a moment estimator for the parameter θ , we have to equate the first
moment of the population to the firts monet of the sample.
First moment of the population is given by
μ1’ = E(X) = ∫ 𝑥𝑓(𝑥)𝑑𝑥
1
= ∫0 𝑥(𝜃 + 1)𝑥 𝜃 𝑑𝑥
1
= (𝜃 + 1) ∫0 𝑥 𝜃+1 𝑑𝑥
𝑥 𝜃+2
= (𝜃 + 1) ]¹0
𝜃+2
𝜃+1
μ1’ = 𝜃+2
First moment of the sample is given by
1
m1’ = 𝑛 ∑𝑛𝑖=1 𝑋ᵢ = 𝑋̅

Equating the moment estimators

μ1’ = m1
𝜃+1
= 𝑋̅
𝜃+2

𝜃 + 1 = (𝜃 + 2) 𝑋̅
𝜃 + 1 = 𝜃𝑋̅ + 2𝑋̅
𝜃 − 𝜃𝑋̅ = 2𝑋̅ − 1
𝜃(1 − ̅̅̅
𝑋) = 2𝑋̅ − 1
2𝑋̅ −1 2∗0.65−1 0.3
𝜃̂ = (1−𝑋)̅̅̅̅ = = = 0.857
(1−0.65) 0.35

MLE
Given the pdf of the population f(x,θ) = (𝜃 + 1)𝑥 𝜃 , 0<x<1 .
Then f(x₁,θ) = (𝜃 + 1)𝑥₁𝜃 , 0<x<1 .
f(x₂,θ) = (𝜃 + 1)𝑥₂𝜃 , 0<x<1 .
............................................
f(xₙ,θ) = (𝜃 + 1)𝑥ₙ𝜃 , 0<x<1 .
Likelyhood function ,
L(x1, x2 ,.... xn ,. θ ) = f(x₁,θ) f(x₂,θ) ........ f(xₙ,θ)
= (𝜃 + 1)𝑥₁𝜃 . (𝜃 + 1)𝑥₂𝜃 ....... (𝜃 + 1)𝑥₂𝜃
= (𝜃 + 1)𝑛 (𝑥₁. 𝑥₂ … . 𝑥ₙ)𝜃
logL= 𝑙𝑜𝑔(𝜃 + 1)𝑛 + 𝑙𝑜𝑔(𝑥₁. 𝑥₂ … . 𝑥ₙ)𝜃
= 𝑛𝑙𝑜𝑔(𝜃 + 1) + 𝜃log (𝑥₁. 𝑥₂ … . 𝑥ₙ)
= 𝑛𝑙𝑜𝑔(𝜃 + 1) + 𝜃[log x₁ + 𝑙𝑜𝑔𝑥₂ + ⋯ + 𝑙𝑜𝑔𝑥ₙ]
 = 𝑛𝑙𝑜𝑔(𝜃 + 1) + 𝜃 ∑𝑛𝑖=1 𝑙𝑜𝑔𝑥ᵢ
Differentiating wrt θ
𝜕𝑙𝑜𝑔𝐿 𝑛
= 𝜃+1 + ∑𝑛𝑖=1 𝑙𝑜𝑔𝑥ᵢ
𝜕𝜃
𝜕𝑙𝑜𝑔𝐿
Put =0
𝜕𝜃
𝑛
+ ∑𝑛𝑖=1 𝑙𝑜𝑔𝑥ᵢ = 0
𝜃+1
𝑛
= − ∑𝑛𝑖=1 𝑙𝑜𝑔𝑥ᵢ
𝜃+1
𝜃+1 1
= − ∑𝑛
𝑛 𝑖=1 𝑙𝑜𝑔𝑥ᵢ

𝑛
θ+1 = − ∑𝑛
𝑖=1 𝑙𝑜𝑔𝑥ᵢ

𝑛
𝜃̂ = − ∑𝑛 −1
𝑖=1 𝑙𝑜𝑔𝑥ᵢ

when applied to the given numerical sample 𝜃̂ = 0.89

Q) Obtain the MLEs for the parameters a and b of the uniform distribution
(rectangular distribution)
Solution
The pdf of rectangular distribution with parameters a and b is given by
1
f(x) = 𝑏−𝑎 , a ≤ x ≤ b

To find the MLEs of a and b take a random sample X1,X2.....Xn from the
rectangular population.
1 1 1
Then f(x1) = 𝑏−𝑎 , f(x2) = 𝑏−𝑎 , .......... f(xn) = 𝑏−𝑎

Likelyhood function ,
L(x1,x2.....xn) = f(x1). f(x2) ....... f(xn)
1 1 1
= 𝑏−𝑎 . 𝑏−𝑎 ..... 𝑏−𝑎
1 𝑛
= (𝑏−𝑎)

As the likelyhood function doesn’t contain any sample observations here the
method of calculus fails to give MLEs for a and b.
MLE is the estimator which maximises the likelyhood function.
1 𝑛
For the likelyhood function L = (𝑏−𝑎) to be maximum , (b-a)n should be
minimum or (b-a ) should be minimum.
For (b-a) to be minimum, b should be minimum and a should be maximum.
Let the sample observations X1,X2.....Xn be such that a < X1< X2 <.....< Xn < b
Then the maximum value of a is X1 , the smallest observation in the sample and
the minimum value of b is Xn , the largest observation in the sample.
That means the MLE of a is X1 and MLE of b is Xn .

Properties of moment estimators

1) Moment estimators are asymptotically unbiased ( ie unbiased when the
sample size n →∞)
2) Moment estimators are consistent.
3) Under fairy general conditions the distribution of moment estimators are
asymptotically normal.

Properties of maximum likelyhood estimators (MLEs)

1) MLEs are consistent
2) MLEs are most efficient
3) MLEs are sufficient , when sufficient estimators exist
4) MLEs are not necessarily unbiased
5) The distribution of MLEs tends to noramality as n→∞

Interval Estimation
Consider a population having the pdf f(x,θ) . Suppose the parameter θ is
unknown . Then take a random sample X1, X2,.....Xn from the population.
Instead of finding a single value as an estimate of θ , if we are finding an
interval say (t1,t2) within which θ may lie with certain probability then this
interval is called the interval estimator of θ [ confidence interval of θ] . The
process of finding this interval estimator is called the interval estimation.
Note:
If P(t1 ≤ θ ≤ t2) = 95/100 = 0.95 then (t1,t2) is the 95% confidence interval for
the parameter θ .
In general if P(t1 ≤ θ ≤ t2) = 1 - α then (t1,t2) is the 100(1 - α )% confidence
interval for the parameter θ . [Here α can be 0.1 ,0.05 ,0.02, 0.01 etc]
1 - α is called the confidence coefficient . It is the probability that a confidence
interval will contain the true value of the parameter.
Higher the value of confidence coefficient , the more certain that the values of
confidence interval are accurate.

Interval estimation of population mean

Case 1 :- When samples are taken from normal population N(μ , σ) , σ known
Here the sample mean 𝑋̅ → N((μ , σ/√n)
𝑋̅ −μ
Then Z = σ/√n → N(0,1)

To find 95% confidence interval for population mean μ

By area property of standard normal distribution ,
P(-1.96 ≤ Z ≤ +1.96) = 0.95
𝑋̅ −μ
P(-1.96 ≤ σ/√n ≤ +1.96) = 0.95

Multiplying throughout by σ/√n

P(-1.96 σ/√n ≤ ̅𝑋 − μ ≤ +1.96 σ/√n) = 0.95
Adding throughout by - ̅𝑋
P( - ̅𝑋-1.96 σ/√n ≤ - μ ≤ - ̅𝑋 +1.96 σ/√n) = 0.95
Multiplying throughout by -1
P( 𝑋̅ +1.96 σ/√n ≥ μ ≥ ̅𝑋 - 1.96 σ/√n) = 0.95
P( 𝑋̅ -1.96 σ/√n ≤ μ ≤ ̅𝑋 + 1.96 σ/√n) = 0.95
⸫ The 95% confidence interval for μ is (𝑋̅ -1.96 σ/√n , ̅𝑋 + 1.96 σ/√n )
The 90% confidence interval for μ is (𝑋̅ -1.645 σ/√n , ̅𝑋 + 1.645 σ/√n )
The 98% confidence interval for μ is (𝑋̅ -2.326 σ/√n , ̅𝑋 + 2,326 σ/√n )
The 99% confidence interval for μ is (𝑋̅ -2.58 σ/√n , ̅𝑋 + 2.58 σ/√n )
In general
The 100(1-α)% confidence interval for μ is (𝑋̅ -Zα/2 σ/√n , ̅𝑋 + Zα/2 σ/√n )
where Zα/2 is the table value .

Case 2 When samples are taken from normal population N(μ , σ) , σ unknown,
n large
Here sample mean 𝑋̅ → N((μ , s/√n) , where s is the sample standard deviation
𝑋̅ −μ
Then Z = s/√n → N(0,1)

The 95% confidence interval for μ is (𝑋̅ -1.96 s/√n , ̅𝑋 + 1.96 s/√n )
The 90% confidence interval for μ is (𝑋̅ -1.645 s/√n , ̅𝑋 + 1.645 s/√n )
The 98% confidence interval for μ is (𝑋̅ -2.326 s/√n , ̅𝑋 + 2,326 s/√n )
The 99% confidence interval for μ is (𝑋̅ -2.58 s/√n , ̅𝑋 + 2.58 s/√n )
The 100(1-α)% confidence interval for μ is (𝑋̅ -Zα/2 s/√n , ̅𝑋 + Zα/2 s/√n ) where
Zα/2 is the table value .

Case3 When samples are taken from normal population N(μ , σ) , σ unknown,
n small
𝑋̅ −μ
Here t = s/ → t(n-1)
√n−1

To find 95% confidence interval for μ

By referring the two tailed t tables for n-1 df and probability 0.05 we can find a
table value tα/2 such that
P( - tα/2 ≤ t ≤ tα/2) = 0.05
𝑋̅ −μ
P( - tα/2 ≤ s/ ≤ tα/2) = 0.05
√n−1

Multiplying throughout by s/√n − 1

P( - tα/2 s/√n − 1 ≤ 𝑋̅ − μ ≤ tα/2 s/√n − 1) = 0.05
Adding throughout by - 𝑋̅
P(-𝑋̅- tα/2 s/√n − 1 ≤ −μ ≤ -𝑋̅ + tα/2 s/√n − 1) = 0.05
Multiplying throughout by -1
P(𝑋̅ + tα/2 s/√n − 1 ≥ μ ≥ 𝑋̅ - tα/2 s/√n − 1) = 0.05
P(𝑋̅ - tα/2 s/√n − 1 ≤ μ ≤ 𝑋̅ + tα/2 s/√n − 1) = 0.05
⸫ The 95% CI for μ is given by (𝑋̅ - tα/2 s/√n − 1 , 𝑋̅ + tα/2 s/√n − 1)

Similarly to find the 98% CI for μ , tα/2 is obtained by referring the two tailed t
tables for n-1 df and probability 0.02 and to find the 99% CI for μ , tα/2 is
obtained by referring the two tailed t tables for n-1 df and probability 0.01

Case 4 : When large samples are taken from any large population , σ known
Here the sample mean 𝑋̅ → N((μ , σ/√n)
And the 95% confidence interval for μ is (𝑋̅ -1.96 σ/√n , ̅𝑋 + 1.96 σ/√n )
Case 5 : When large samples are taken from any large population , σ unknown
Here the sample mean 𝑋̅ → N((μ , s/√n)
And the 95% confidence interval for μ is (𝑋̅ -1.96 s/√n , ̅𝑋 + 1.96 s/√n )

Note :
The general format of a confidence interval of a parameter is given by
Point estimator ± Table value * Standard Error
ie., PE ± TV *SE
 1) The mean and standard deviation of a sample of size 60 are found to be
145 and 40 respectively. Construct a 95% confidence interval for
population mean,

Solution
Given n=60 (large n)
Sample mean 𝑋̅ =145
Sample SD , s = 40 (σ unknown)

Here the 95% confidence interval of the population mean μ is given by

s s
(𝑋̅ − 1.96 n , ̅𝑋 +1.96 n )
√ √
= (145- 1.96 * 40/√60 , 145+1.96 * 40/√60 )
= (134.88 , 155.12)

2) A random sample of 20 bullets produced by a machine show an average

diameter of 3.5mm and a SD of 0.2mm . Assuming that the diameter
measurements follows N(μ ,σ) , obtain the 95% confidence interval for
mean μ

Solution
Given n= 20 (small)
Sample mean 𝑋̅ =3.5
Sample SD , s = 0.2 (σ unknown)

Here the 95% confidence interval of the population mean μ is given by

s s
(𝑋̅ − tα/2 n−1 , ̅𝑋 +tα/2 n−1 ) where tα/2 is obtained by referring
√ √
the two tailed student’s t table for (n-1)df = 19df probability α=0.05
From tables tα/2 = 2.093
⸫ 95% confidence interval of μ is
= (3.5 – 2.093* 0.2/√19 , 3.5 + 2.093* 0.2/√19)
= (3.403 , 3.596)

Interval estimation of difference of two population means

Case 1 :- When samples are taken from two normal populations N(μ1 , σ1) and
N(μ2 , σ2) , σ1 , σ2 known
 𝜎₁2 𝜎₂2
̅̅̅̅-𝑋₂
Here the difference of sample means 𝑋₁ ̅̅̅ → N(μ1 – μ2 , √ + )
𝑛₁ 𝑛₂

̅̅̅̅̅−𝑋₂)
(𝑋₁ ̅̅̅̅̅ −(μ₁ – μ₂)
Then Z = 2 2
→ N(0,1)
√𝜎₁ +𝜎₂
𝑛₁ 𝑛₂

To find 95% confidence interval for difference of population means μ1- μ2

By area property of standard normal distribution ,
P(-1.96 ≤ Z ≤ +1.96) = 0.95
̅̅̅̅̅−𝑋₂)
(𝑋₁ ̅̅̅̅̅ −(μ₁ – μ₂)
P(-1.96 ≤ = ≤ +1.96) = 0.95
𝜎₁2 𝜎₂2
√ +
𝑛₁ 𝑛₂

𝜎₁2 𝜎₂2
Multiplying throughout by √ 𝑛₁ + 𝑛₂

𝜎₁2 𝜎₂2 2 2
P(-1.96 √ 𝑛₁ + ≤ (𝑋₁ ̅̅̅̅̅ − (μ₁ – μ₂) ≤ +1.96 √𝜎₁ + 𝜎₂ ) = 0.95
̅̅̅̅̅ − 𝑋₂)
𝑛₂ 𝑛₁ 𝑛₂

̅̅̅̅̅ − ̅̅̅̅̅
Adding throughout by –(𝑋₁ 𝑋₂)
𝜎₁ 𝜎₂ 2 2
P(- ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅ ̅̅̅̅̅ − ̅̅̅̅̅
𝑋₂) -1.96 √ 𝑛₁ + 𝑛₂ ≤ - (μ₁ – μ₂) ≤ −(𝑋₁ 𝑋₂) +1.96
𝜎₁2 𝜎₂2
√ + ) = 0.95
𝑛₁ 𝑛₂

Multiplying throughout by -1
2 2 2 2
̅̅̅̅̅ +1.96 √𝜎₁ + 𝜎₂ ≥ (μ₁ – μ₂) ≥ (𝑋₁
̅̅̅̅̅ − 𝑋₂)
P((𝑋₁ ̅̅̅̅̅ - 1.96 √𝜎₁ + 𝜎₂ )
̅̅̅̅̅ − 𝑋₂)
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂
= 0.95
𝜎₁ 𝜎₂ 2 2 𝜎₁ 𝜎₂ 2 2
P( ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) -1.96 √ + ≤ (μ₁ – μ₂) ≤ ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +1.96 √ + ) =
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂
0.95
⸫ The 95% confidence interval for (μ₁ – μ₂) is
𝜎₁ 𝜎₂ 2 2
𝜎₁ 𝜎₂ 2 2
̅̅̅̅̅ − ̅̅̅̅̅
((𝑋₁ 𝑋₂) -1.96 √ 𝑛₁ + 𝑛₂ , ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +1.96 √ 𝑛₁ + 𝑛₂ )

The 90% confidence interval for μ is

 2 2 2 2
̅̅̅̅̅ -1.645 √𝜎₁ + 𝜎₂ , (𝑋₁
̅̅̅̅̅ − 𝑋₂)
((𝑋₁ ̅̅̅̅̅ +1.645 √𝜎₁ + 𝜎₂ )
̅̅̅̅̅ − 𝑋₂)
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

The 98% confidence interval for μ is

𝜎₁ 𝜎₂ 2 2 𝜎₁ 𝜎₂ 2 2
̅̅̅̅̅ − ̅̅̅̅̅
((𝑋₁ 𝑋₂) -2.326 √ 𝑛₁ + 𝑛₂ , ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +2.326 √ 𝑛₁ + 𝑛₂ )

The 99% confidence interval for μ is

𝜎₁ 𝜎₂2 2 𝜎₁ 𝜎₂ 2 2
̅̅̅̅̅ − ̅̅̅̅̅
((𝑋₁ 𝑋₂) -2.58 √ 𝑛₁ + 𝑛₂ , ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +2,58 √ 𝑛₁ + 𝑛₂ )

In general
The 100(1-α)% confidence interval for μ is
2 2 2 2
̅̅̅̅̅ -Zα/2 √𝜎₁ + 𝜎₂ , (𝑋₁
̅̅̅̅̅ − 𝑋₂)
((𝑋₁ ̅̅̅̅̅ +Zα/2 √𝜎₁ + 𝜎₂ )
̅̅̅̅̅ − 𝑋₂)
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

where Zα/2 is the table value .

Case 2 When samples are taken from two normal populations N(μ 1 , σ1) and
N(μ2 , σ2) , σ1 , σ2 unknown , n1 ,n2 large
2 2
Here the difference of sample means 𝑋₁ ̅̅̅ → N(μ1 – μ2 , √𝑠₁ + 𝑠₂ )
̅̅̅̅-𝑋₂ , where
𝑛₁ 𝑛₂
s is the sample standard deviation
̅̅̅̅̅−𝑋₂)
(𝑋₁ ̅̅̅̅̅ −(μ₁ – μ₂)
Then Z = → N(0,1)
𝜎₁2 𝜎₂2
√ + 𝑛₂
𝑛₁

⸫ The 95% confidence interval for (μ₁ – μ₂) is

𝑠₁ 𝑠₂2 2 𝑠₁ 2
𝑠₂ 2
̅̅̅̅̅ − ̅̅̅̅̅
((𝑋₁ 𝑋₂) -1.96 √ 𝑛₁ + 𝑛₂ , ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +1.96 √ 𝑛₁ + 𝑛₂ )

The 90% confidence interval for μ is

 2 2 2 2
̅̅̅̅̅ -1.645 √𝑠₁ + 𝑠₂ , (𝑋₁
̅̅̅̅̅ − 𝑋₂)
((𝑋₁ ̅̅̅̅̅ +1.645 √𝑠₁ + 𝑠₂ )
̅̅̅̅̅ − 𝑋₂)
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

The 98% confidence interval for μ is

𝑠₁ 𝑠₂ 2 2 𝑠₁ 𝑠₂ 2 2
̅̅̅̅̅ − ̅̅̅̅̅
((𝑋₁ 𝑋₂) -2.326 √ 𝑛₁ + 𝑛₂ , ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +2.326 √ 𝑛₁ + 𝑛₂ )

The 99% confidence interval for μ is

𝑠₁ 𝑠₂ 2 2 𝑠₁ 𝑠₂ 2 2
̅̅̅̅̅ − ̅̅̅̅̅
((𝑋₁ 𝑋₂) -2.58 √ 𝑛₁ + 𝑛₂ , ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +2,58 √ 𝑛₁ + 𝑛₂ )

In general
The 100(1-α)% confidence interval for μ is
2 2 2 2
̅̅̅̅̅ -Zα/2 √𝑠₁ + 𝑠₂ , (𝑋₁
̅̅̅̅̅ − 𝑋₂)
((𝑋₁ ̅̅̅̅̅ +Zα/2 √𝑠₁ + 𝑠₂ )
̅̅̅̅̅ − 𝑋₂)
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

where Zα/2 is the table value .

Case3 When samples are taken from two normal populations N(μ1 , σ1) and
N(μ2 , σ2) , σ1 = σ2 unknown , n1 ,n2 small
̅̅̅̅̅−𝑋₂)
(𝑋₁ ̅̅̅̅̅ −(μ₁ – μ₂)
Here t = → t(n1 + n2 -2)
n₁s₁2 +n₂s₂2 1 1
√[ (n₁+n₂)]
n₁+n₂−2

To find 95% confidence interval for μ₁ – μ₂

By referring the two tailed t tables for (n1 + n2 -2) df and probability 0.05 we can
find a table value tα/2 such that
P( - tα/2 ≤ t ≤ tα/2) = 0.05
̅̅̅̅̅−𝑋₂)
(𝑋₁ ̅̅̅̅̅ −(μ₁ – μ₂)
P( - tα/2 ≤ ≤ tα/2) = 0.05
n₁s₁2 +n₂s₂2 1 1
√[ (n₁+n₂)]
n₁+n₂−2

n₁s₁2 +n₂s₂2 1 1
Multiplying throughout by √[ (n₁ + n₂)]
n₁+n₂−2

n₁s₁2 +n₂s₂2 1 1
P( - tα/2 √[ (n₁ + n₂)] ≤ ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) − (μ₁ – μ₂) ≤
n₁+n₂−2

n₁s₁2 +n₂s₂2 1 1
tα/2 √[ (n₁ + n₂)]) = 0.05
n₁+n₂−2
Adding throughout by – ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂)
n₁s₁ 2 +n₂s₂2 1 1
P(– ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂)- tα/2 √[ ( + )] ≤ −(μ₁ – μ₂) ≤ – ̅̅̅̅̅
(𝑋₁ −
n₁+n₂−2 n₁ n₂
2
n₁s₁ +n₂s₂ 1 1 2
̅̅̅̅̅
𝑋₂) + tα/2 √[ n₁+n₂−2 (n₁ + n₂)]) = 0.05

Multiplying throughout by -1
n₁s₁ +n₂s₂ 1 1 2 2
̅̅̅̅̅ − ̅̅̅̅̅
P((𝑋₁ 𝑋₂) + tα/2 √[ n₁+n₂−2 (n₁ + n₂)] ≥ (μ₁ – μ₂) ≥ ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) -
n₁s₁2 +n₂s₂2 1 1
tα/2 √[ n₁+n₂−2
(n₁ + n₂)]) = 0.05

n₁s₁ +n₂s₂ 1 1 2 2
P( ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅̅
𝑋₂) - tα/2 √[ n₁+n₂−2 (n₁ + n₂)] ≤ (μ₁ – μ₂) ≤ ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +
n₁s₁2 +n₂s₂2 1 1
tα/2 √[ (n₁ + n₂)]) = 0.05
n₁+n₂−2

⸫ The 95% CI for μ is given by

n₁s₁ +n₂s₂ 1 12 2
̅̅̅̅̅ − ̅̅̅̅̅̅
((𝑋₁ 𝑋₂) - tα/2 √[ n₁+n₂−2 (n₁ + n₂)] , ̅̅̅̅̅
(𝑋₁ −
2
n₁s₁ +n₂s₂ 1 1 2
̅̅̅̅̅
𝑋₂)+tα/2 √[ n₁+n₂−2 (n₁ + n₂)])

Similarly to find the 98% CI for μ₁ – μ₂ , tα/2 is obtained by referring the two
tailed t tables for (n1 + n2 -2) df and probability 0.02 and to find the 99% CI for
μ₁ – μ₂ , tα/2 is obtained by referring the two tailed t tables for (n1 + n2 -2) df
and probability 0.01

Case 4 : When large samples are taken from any two large populations , σ1 , σ2
known
2 2
̅̅̅ → N(μ1 – μ2 , √𝜎₁ + 𝜎₂ )
̅̅̅̅-𝑋₂
Here the difference of sample means 𝑋₁ 𝑛₁ 𝑛₂

̅̅̅̅̅−𝑋₂)
(𝑋₁ ̅̅̅̅̅ −(μ₁ – μ₂)
Then Z = 2 2
→ N(0,1)
√𝜎₁ +𝜎₂
𝑛₁ 𝑛₂
And The 95% confidence interval for (μ₁ – μ₂) is
𝜎₁ 𝜎₂ 2 2 𝜎₁ 𝜎₂ 2 2
̅̅̅̅̅ − ̅̅̅̅̅
((𝑋₁ 𝑋₂) -1.96 √ + , ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +1.96 √ + )
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

Case 5 : When large samples are taken from any two large populations , σ1 , σ2
unknown
𝑠₁2 𝑠₂2
̅̅̅̅-𝑋₂
Here the difference of sample means 𝑋₁ ̅̅̅ → N(μ1 – μ2 , √ + ) , where
𝑛₁ 𝑛₂
s is the sample standard deviation
̅̅̅̅̅−𝑋₂)
(𝑋₁ ̅̅̅̅̅ −(μ₁ – μ₂)
Then Z = → N(0,1)
𝜎₁2 𝜎₂2
√ +
𝑛₁ 𝑛₂

⸫ The 95% confidence interval for (μ₁ – μ₂) is

𝑠₁ 𝑠₂ 2 2 𝑠₁ 𝑠₂ 2 2
̅̅̅̅̅ − ̅̅̅̅̅
((𝑋₁ 𝑋₂) -1.96 √ 𝑛₁ + 𝑛₂ , ̅̅̅̅̅
(𝑋₁ − ̅̅̅̅̅
𝑋₂) +1.96 √ 𝑛₁ + 𝑛₂ )

Interval estimation of population proportion of success p

𝑝𝑞
Sample proportion of success p’ →N(p ,√ 𝑛 ) where p is the population
proportion of success and q = 1-p and n is the sample size
𝑝′ −𝑝
Then Z = → N(0,1)
𝑝𝑞
√𝑛

95% confidence interval for p

By area property of standard normal distribution
P( -1.96 ≤ Z ≤ +1.96) = 0.95
𝑝′ −𝑝
P( -1.96 ≤ ≤ +1.96) = 0.95
𝑝𝑞
√𝑛

𝑝𝑞
Multiplying throughout by √ 𝑛

𝑝𝑞 𝑝𝑞
P( -1.96√ ≤ 𝑝 ′ − 𝑝 ≤ +1.96√ ) = 0.95
𝑛 𝑛

Adding throughout by –p’

 𝑝𝑞 𝑝𝑞
P(−𝑝 ′-1.96√ 𝑛 ≤ −𝑝 ≤ −𝑝′ +1.96√ 𝑛 ) = 0.95

Multiplying throughout by -1
𝑝𝑞 𝑝𝑞
P(𝑝 ′+1.96√ 𝑛 ≥ 𝑝 ≥ 𝑝 ′-1.96√ 𝑛 ) = 0.95

𝑝𝑞 𝑝𝑞
P(𝑝 ′-1.96√ ≤ 𝑝 ≤ 𝑝 ′ +1.96√ ) = 0.9
𝑛 𝑛

As the population proportion of success is unknown replace p by its estimator p’

𝑝𝑞
and q by q’ in √
𝑛

⸫ The 95% confidence interval for population proportion of success p is

𝑝′𝑞′ 𝑝′𝑞′
(𝑝 ′ -1.96√ , 𝑝′+1.96√ )
𝑛 𝑛

Note
The 90% confidence interval for population proportion of success p is
𝑝′𝑞′ 𝑝′𝑞′
(𝑝 ′ -1.645√ , 𝑝′+1. 645√ )
𝑛 𝑛

The 98% confidence interval for population proportion of success p is

𝑝′𝑞′ 𝑝′𝑞′
(𝑝 ′ -2.326√ , 𝑝′ +2.326√ )
𝑛 𝑛

The 99% confidence interval for population proportion of success p is

𝑝′𝑞′ 𝑝′𝑞′
(𝑝 ′ -2.58√ , 𝑝′+2.58√ )
𝑛 𝑛

Problem
Q) A random sample of 500 pineapples are taken from a large consignment and
65 were found to be bad . Find the 95% confidence interval for the population
proportion of bad pineapples.
Solution
Given sample size n =500
x =65
Sample proportion of success p’ =x/n = 65/500 = 0.13 and q’ = 1-p’ = 0.87
The 95% confidence interval for population proportion of success p is
𝑝′𝑞′ 𝑝′𝑞′
(𝑝 ′ -1.96√ , 𝑝′+1.96√ )
𝑛 𝑛

0.13∗0.87 0.13∗0.87
= (0.13 - 1.96√ 500
, 0.13 + 1.96√ 500
) = (0.0911 , 0.169)

Interval estimation of difference of two population proportions

Let p1 = population proportion of the first population
p2 = population proportion of the second population
p1’ = sample proportion of the sample from first population
p2’ = sample proportion of the sample from second population
𝑝₁𝑞₁ 𝑝₂𝑞₂
Then p1’→N( p1 , √ ) and p2’→N( p2 , √ )
𝑛₁ 𝑛₂

𝑝₁𝑞₁ 𝑝₂𝑞₂
p1’- p2’ → N( p1 - p2 , √ + )
𝑛₁ 𝑛₂

(𝑝₁’− 𝑝₂’)−( 𝑝₁ − 𝑝₂)

then Z = 𝑝₁𝑞₁ 𝑝₂𝑞₂
→ N(0,1)
√ 𝑛₁ + 𝑛₂

95% confidence interval

By area property of standard normal distribution
P(-1.96 ≤ Z ≤ +1.96) = 0.95
(𝑝₁’− 𝑝₂’)−( 𝑝₁ − 𝑝₂)
P(-1.96 ≤ 𝑝₁𝑞₁ 𝑝₂𝑞₂
≤ +1.96) = 0.95
√ 𝑛₁ + 𝑛₂

𝑝₁𝑞₁ 𝑝₂𝑞₂
Multiplying throughout by √ +
𝑛₁ 𝑛₂
 𝑝₁𝑞₁ 𝑝₂𝑞₂ 𝑝₁𝑞₁ 𝑝₂𝑞₂
P(-1.96√ + ≤ (𝑝₁’ − 𝑝₂’) − ( 𝑝₁ − 𝑝₂)≤ +1.96√ + ) =0.95
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

Adding throughout by −(𝑝₁’ − 𝑝₂’)

𝑝₁𝑞₁ 𝑝₂𝑞₂
P(−(𝑝₁’ − 𝑝₂’)-1.96√ + 𝑛₂ ≤ −( 𝑝1 − 𝑝2) ≤ −(𝑝₁’ −
𝑛₁

𝑝₁𝑞₁ 𝑝₂𝑞₂
𝑝₂’)+1.96√ + ) =0.95
𝑛₁ 𝑛₂

Multiplying throughout by -1
𝑝₁𝑞₁ 𝑝₂𝑞₂
P((𝑝₁’ − 𝑝₂’) +1.96√
𝑛₁
+ 𝑛₂
≥ ( 𝑝₁ − 𝑝₂) ≥ (𝑝₁’ −

𝑝₁𝑞₁ 𝑝₂𝑞₂
𝑝₂’) −1.96√
𝑛₁
+ 𝑛₂
) =0.95

𝑝₁𝑞₁ 𝑝₂𝑞₂
P((𝑝₁’ − 𝑝₂’) −1.96√ + ≤ ( 𝑝₁ − 𝑝₂) ≤ (𝑝₁’ −
𝑛₁ 𝑛₂

𝑝₁𝑞₁ 𝑝₂𝑞₂
𝑝₂’)+1.96√ + ) =0.95
𝑛₁ 𝑛₂

Replacing p1 and p2 by their point estimates p1’ and p2’

𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′
P((𝑝₁’ − 𝑝₂’) −1.96√ + ≤ ( 𝑝₁ − 𝑝₂) ≤ (𝑝₁’ −
𝑛₁ 𝑛₂

𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′
𝑝₂’)+1.96√ + ) =0.95
𝑛₁ 𝑛₂

⸫ 95% confidence interval for p1 – p2 is given by

𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′ 𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′
[(𝑝₁’ − 𝑝₂’) −1.96√ + , (𝑝₁’ − 𝑝₂’)+1.96√ + ]
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

Similarly 90% confidence interval for p1 – p2 is given by

𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′ 𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′

[(𝑝₁’ − 𝑝₂’) −1.645√ 𝑛₁ + 𝑛₂ , (𝑝₁’ − 𝑝₂’)+1.645√ 𝑛₁ + 𝑛₂ ]

98% confidence interval for p1 – p2 is given by

 𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′ 𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′
[(𝑝₁’ − 𝑝₂’) −2.326√ + , (𝑝₁’ − 𝑝₂’)+2.326√ + ]
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

99% confidence interval for p1 – p2 is given by

𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′ 𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′

[(𝑝₁’ − 𝑝₂’) −2.58√ + , (𝑝₁’ − 𝑝₂’) + 2.58√ + ]
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

Problem

In a survey of 400 people from a city , 188 preferred brand A soap to all others
and in a sample of 500 people from another city, 210 preferred same soap.
Prepare the 95% confidence interval for (p1 –p2) where p1 is the proportion
preferring brand A soap in the first city and p2 is the proportion preferring
brand A soap in the second city.

Solution
Given n1= 400 , x1 = 188 and n2= 500 , x2 = 210
Then p1’ = 188/400 = 0.47 and q1’ = 1- p1’ = 1-0.47 = 0.53
p2’ = 210/500 = 0.42 and q2’ = 1- p2’ = 1-0.42= 0.58
95% confidence interval for p1 – p2 is given by
𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′ 𝑝₁′𝑞₁′ 𝑝₂′𝑞₂′
[(𝑝₁’ − 𝑝₂’) −1.96√ + , (𝑝₁’ − 𝑝₂’)+1.96√ + ]
𝑛₁ 𝑛₂ 𝑛₁ 𝑛₂

0.47∗0.53 0.42∗0.58
= [(0.47 − 0.42) −1.96√ + , (0.47 −
400 500

0.47∗0.53 0.42∗0.58
0.42) +1.96√ + ]
400 500

= ( - 0.015 , 0.115)
Interval estimation of population variance σ 2
𝑛𝑠 2
We know that → χ2(n-1) where s2 is the sample variance
𝜎2
95% confidence interval
By referring the χ2 tables for (n-1) df and probabilities 1-α/2 and α/2 we can find
two table values χ21-α/2 and χ2 α/2 such that
P(χ21-α/2 ≤ χ2 ≤ χ2 α/2 ) = 0.95
[for 95% , α =0 .05 . so α/2 = 0.025 and 1- α/2 = 0.0975)
𝑛𝑠 2
P(χ 2
1-α/2 ≤ ≤ χ2 α/2 ) = 0.95
𝜎2
Taking reciprocals
𝜎2 2
P(1/χ21-α/2 ≥ ≥ 1/ χ2 α/2 ) = 0.95
𝑛𝑠 2
Multiplying throughout by ns2
P(ns2/χ21-α/2 ≥ σ 2 ≥ ns2/ χ2 α/2 ) = 0.95
P(ns2/ χ2 α/2 ≤ σ 2 ≤ ns2/χ21-α/2) = 0.95
⸫ 95% confidence interval for σ 2 is given by (ns2/ χ2 α/2 , ns2/χ21-α/2) )
Note : For finding 90% confidence interval take α =0 .1 , for 98% , take α =0
.02 and for 99% , take α =0 .01

Problems
Q ) A random sample of size 15 from a normal population has sample variance
s2 = 4.28. Determine the 90% confidence interval for σ2.
Solution
Given n= 15 , s2 = 4.28
90% confidence interval for σ 2 is given by (ns2/ χ2 α/2 , ns2/χ21-α/2) ) where
α=0.1 , α/2= 0.05 and 1- α/2 = 0.95
By referring the χ2 tables for (n-1)df = 14df and α=0.1 we can get
χ2 0.05 =28.6848 and χ20.95 = 6.5706
90% CI is given by (15*4.28/28.6848 , 15*4.28 /6.5706) = (2.71 , 9.77)
Q) If 8.6, 7.9, 8.3, 6.4, 8.4, 9.8, 7.2, 7.8, 7.6 are the observed values of a
random variable of size 9 from N(μ,σ) , construct a 95% CI for σ2
Solution
Here n=9
1
s2 = 𝑛 ∑𝑛𝑖=1(𝑥ᵢ − 𝑥̅ )2

xi ̅)
(xi - 𝒙 ̅)2
(xi - 𝒙
8.6 0.6 0.36
7.9 -0.1 0.01
8.3 0.3 0.09
6.4 -1.6 2.56
8.4 0.4 0.16
9.8 1.8 3.24
7.2 -0.8 0.64
7.8 -0.2 0.04
7.6 -0.4 0.16
∑ 𝑥ᵢ = 72 ∑(𝒙𝒊 − 𝒙 ̅)𝟐 = 7.1
1
𝑥̅ = ∑ 𝑥ᵢ = 72/9 =8
𝑛
1
s2 = 𝑛 ∑𝑛𝑖=1(𝑥ᵢ − 𝑥̅ )2 = 7.1/9 = 0.789

95% confidence interval for σ 2 is given by (ns2/ χ2 α/2 , ns2/χ21-α/2) )

Here α = 0.05 , α/2 = 0.025 and 1- α/2 =0.975
Referring the χ2 tables for (n-1)df = 8df , χ2 0.05 = 17.5345 and χ2 0.975 = 2.1797
95% CI is (9*0.789/ 17.5345 , 9*0.789/2.1797) = ( 0.404 , 3.25)
Interval estimation of population correlation coefficient ρ
Let ρ = population correlation coefficient
r = sample correlation coefficient
𝑟−𝜌
then t = 2
→t(n-2)
√1−𝑟
𝑛−2

95% CI
By referring the two tailed t tables for (n-2)df and probability 0.05 we can find
a table value tα/2 such that
P(-tα/2 ≤ t ≤ + tα/2 ) = 0.95
𝑟−𝜌
P(-tα/2 ≤ 2
≤ + tα/2 ) = 0.95
√1−𝑟
𝑛−2

1−𝑟 2
Multiplying throughout by √
𝑛−2

1−𝑟 2 1−𝑟 2
P(-tα/2 √ ≤ 𝑟 − 𝜌 ≤ + tα/2√ ) = 0.95
𝑛−2 𝑛−2

Adding throughout by –r

1−𝑟 2 1−𝑟 2
P(–r -tα/2 √ ≤ −𝜌 ≤ –r + tα/2√ ) = 0.95
𝑛−2 𝑛−2

Multiplying throughout by -1

1−𝑟 2 1−𝑟 2
P(r + tα/2 √ ≥ 𝜌 ≥ r - tα/2√ ) = 0.95
𝑛−2 𝑛−2

1−𝑟 2 1−𝑟 2
P(r -tα/2 √ ≤ 𝜌 ≤ r + tα/2√ ) = 0.95
𝑛−2 𝑛−2

⸫ 95% confidence interval for population correlation coefficient ρ is given by

1−𝑟 2 1−𝑟 2
[r -tα/2 √ 𝑛−2 , r + tα/2√ 𝑛−2 ]