MTH 216 STATISTICAL INFERENCE 2 BRIEF COURSE OUTLINE.

1.0 Theory OF Estimation

2.0 Tests of Hypotheses
3.0 Goodness-of-fit Tests and Contingency Tables
4.0 Correlation Analyses
5.0 Simple Linear Regression Models

CHAPTER ONE
THEORY OF ESTIMATION

5.1 Introduction
Inferential statistics is basically concerned with techniques for drawing
conclusions about an entire population based on the results from a study of a sample
drawn from the population. The conclusion may be to predict the values of some
population parameters or to supply a range of values which has a known probability
of including the true value of the parameter, or to assess the probability of certain
kinds of results under certain population conditions.
A population is defined in statistics as the entire body of items about which we want
to obtain some information or reach an opinion. It is the totality of all units whose
attribute is under investigation. Sometimes, when a population is under study, it
might be impracticable to extract the required data from every item of that
population. We might be contented with just a part of the population selected
according to some rules. This fraction of the population which is selected for the
purpose of study, so as to make some statements of conclusion about the entire
population is called a sample.
Any descriptive measure obtained (by calculation or otherwise) from sample values
is called a statistic, while those obtained using the entire population data are called
parameters. Examples of sample statistics are sample mean, 𝑋̅ , sample variance, 𝑆 2 ,
sample standard deviation, 𝑆, sample median, sample size, n, and every other

1
measure calculated or obtained from sample data. Examples of population
parameters are population mean, population variance, 𝜎 2 , population standard
deviation, 𝜎, population correlation coefficient, 𝜌, population size N, and indeed every
other measure obtained by using full population data.
Any statistic T, derived from a random sample, and used to give information
about some unknown population parameter 𝜃, is called an estimator for 𝜃. If the
estimator T, for a population parameter 𝜃 is given by a single value, then the estimator
is called a point estimate. Interval estimation is concerned with locating a range of
values within which a population parameter is expected to lie with a given degree of
confidence or probability. This degree of confidence is usually expressed in
percentage or in probability form.

5.2 Qualities of a Good Estimator

Due to difficulties associated with census taking, sample values are often used to
estimate the parameters of a population. For such estimates to serve their useful
purposes, they must meet some basic criteria. Among them are Unbiasedness,
Efficiency, Consistency, and Sufficiency.

(a) Unbiasedness:
Let 𝜃 be an unknown population parameter, and T an estimator for 𝜃. Then T is said
to be an unbiased estimator for 𝜃 if E(T) = 𝜃, where E(T) stands for the mathematical
expectation (or average value of T). It should be noted that for any population
parameter 𝜃, it is possible to have more than one unbiased estimators for 𝜃.

Theorem 5.1
Let X1, X2, X3,……,Xn be a random sample of size n, drawn from a population having
unknown population mean μ and variance σ2. Then
(i) the sample mean 𝑋̅ is an unbiased estimator for the population mean μ

2
 𝑛
(ii) 𝑆 2 is an unbiased estimator for σ2
𝑛−1

Proof:
(i)
𝑋1 + 𝑋2 + 𝑋3 + ⋯ + 𝑋𝑛 ∑𝑋𝑖
𝑋̅ = =
𝑛 𝑛
∑𝑋
Now, E(𝑋̅ ) = 𝐸[ 𝑖 ] = 1⁄𝑛 ∑𝐸 (𝑋𝑖 ) = 𝜇
𝑛

Hence 𝑋̅ is an unbiased estimator for μ

(ii) Now the sample variance is given by 𝑆 2 = 1⁄𝑛 ∑(𝑋𝑖 − 𝑋̅ )2
𝑛 1
∴ 𝑆2 = ∑(𝑋𝑖 − 𝑋̅ )2 … … … … … … … .. (1)
𝑛−1 𝑛−1

But Var (Xi ) = E{( Xi – μ )2 = σ2 … .. … . (2)

∑𝑋𝑖 𝜎2
Also Var(𝑋̅ ) = Var ( ) =
𝑛 𝑛
𝜎2
i.e. E{(𝑋̅ - μ)2 } = …… … … … (3)
𝑛

But observe that ∑(Xi - 𝑋̅ )2 = ∑ ( Xi – μ + μ - 𝑋̅ )2

= ∑(X i − μ)2 − n (𝑋̅ − μ)2 … … … … (4)
𝑛 1
Hence, 𝐸 [ 𝑆2] = 𝐸 [ ∑(𝑋𝑖 − 𝑋̅ )2 ]
𝑛−1 𝑛−1

1
= 𝐸 ( ∑(𝑋𝑖 − 𝜇)2 − 𝑛(𝑋̅ − 𝜇)2 )
𝑛−1
1
= { ∑𝐸((𝑋𝑖 − 𝜇)2 ) − 𝑛 𝐸((𝑋̅ − 𝜇)2 )
𝑛−1
1 𝜎2 1
= [ 𝑛𝜎 2 − 𝑛 ] = (𝑛𝜎 2 − 𝜎 2 )
𝑛−1 𝑛 𝑛−1
1
= (𝑛 − 1)𝜎 2 = 𝜎2
𝑛−1
𝑛 𝑛
Hence, 𝐸 [ 𝑆2] = 𝜎 2 ⇒ 𝑆 2 is an unbiased estimator for σ2
𝑛−1 𝑛−1

Exercises
Show that ∑(𝑋𝑖 − 𝑋̅ )2 = ∑𝑋𝑖2 − 𝑛𝑋̅ 2 and that 𝑆 2 is not an unbiased estimator for σ2.

3
Example 5.1
The following random sample was obtained from a population:
12, 8, 11, 10, 8, 8, 13, 9, 11, 10.
Find 𝑋̅ and 𝑆 2 , and hence obtain unbiased estimates of μ and σ2.
Solution:
𝑛𝑆 2
Unbiased estimates of μ and σ2 are given by 𝑋̅ and , respectively.
𝑛−1

𝑋̅ = 1⁄𝑛 ∑𝑋𝑖 = 100⁄10 = 10

∴ 𝜇 = 10
Sample variance is 𝑆 2 = 1⁄𝑛 ∑(𝑋𝑖 − 𝑋̅ )2 = 28⁄10 = 2.8
Unbiased estimate for population variance from sample variance is

2
𝑛𝑆 2 10 𝑥 2.8
𝜎̂ = = = 3.11
𝑛−1 9
1
This is equivalent to 𝜎̂ 2 = ∑(𝑋𝑖 − 𝑋̅ )2
𝑛−1

NOTE
 Sample variance is 𝑆 2 = 1⁄𝑛 ∑(𝑋𝑖 − 𝑋̅ )2
 Unbiased estimate for population variance from sample data is
1
𝜎̂ 2 = ∑(𝑋𝑖 − 𝑋̅ )2
𝑛−1
1
 But the actual population variance is 𝜎 2 = ∑(𝑋𝑖 − 𝜇)2
𝑁

(b) Efficiency
Let T1 and T2 be distinct unbiased estimators for 𝜃, where 𝜃 is an unknown population
parameter, i.e. E(T1) = E (T1) = 𝜃.
Both T1 and T2 will have well defined distributions, and suppose their variances are
not equal, and if Var(T1) < Var(T2), then we would say that T1 is a more efficient
estimator of 𝜃 than T2. Efficiency is therefore concerned with the comparison of all
unbiased estimators of a parameter 𝜃. The one with the smallest variance is called the

4
most efficient estimator of 𝜃. This property is also called Minimum Variance Unbiased
Estimator (MVUE).

Theorem 5.2
Given any random sample of size n from a population with mean μ and variance σ 2.
(i) The most efficient estimator for μ is 𝑋̅
𝑛
(ii) The most efficient estimator for σ2 is 𝑆2
𝑛−1

Example 5.2
1
Given a random sample X1, X2, X3,…, Xn of size n show that both (X1+X2+X3) and
3
1
(X4+X5) are unbiased estimators for μ . Which of the two is more efficient? Hence
2

show that of all unbiased estimates derived from the sample, 𝑋̅ is the most efficient.

Solution:
1 1
Clearly E{ (X1+X2+X3)} = .3μ = μ
3 3

Also E{1/2 (X4+X5)} = ½ (μ + μ ) = μ

Hence both are unbiased estimators of μ
1 𝜎2
But Var{ (X1+X2+X3x4) } = 3 σ2 / 9 = (1)
3 3
𝜎2
Also Var{1/2 (X4 + X5)} = ¼ (σ2+σ2) =
2
𝜎2 𝜎2
Since σ2> 0, <
3 2
1
... (X1+X2+X3) is a more efficient estimator for μ than 1/2 (X4 + X5)
3
𝜎2
Also, Var(𝑋̅ ) =
𝑛

If n is the sample size, then obviously, Var(𝑋̅ ) is least.

Hence 𝑋̅ is the most efficient estimator of μ.

5
(c) Consistency
This is concerned with the behaviour of an estimator as the sample size n becomes
large. Given sample X1, X2, …………,Xn of size n from a population with unknown
parameter 𝜃. Supposed T is an unbiased estimator of 𝜃 derived from the sample.
If as n → ∞, Var(T) → 0 or lim 𝑃[|𝑇 − 𝜃|] → 0 , then T is called a consistent estimator
𝑛→∞

for 𝜃. Simply put as the sample size increases, a sufficient estimator becomes more
reliable.
𝜎2
For example Var{ ½ (X4+X4)} = which clearly is independent of n, as the
2

sample size increases, remains constant and does not tend to zero.
Hence, ½ (X4+X5) though, unbiased, is not a consistent estimator of μ

Theorem 5.3
i. 𝑋̅ is a consistent estimator for μ
𝑛
ii. 𝑆 2 is a consistent estimator for σ.
𝑛−1

Proof:
(i) Var (𝑋̅ ) = Var{1⁄𝑛 ∑𝑋𝑖 }
1 1 𝜎2
= 2
Σ𝑉𝑎𝑟 (𝑋𝑖 ) = 2
× 𝑛𝜎 2 =
𝑛 𝑛 𝑛
𝜎2
lim 𝑉𝑎𝑟(𝑋̅ ) = lim [ ]→0
𝑛→∞ 𝑛→∞ 𝑛

∴ 𝑋̅ is a consistent estimate for μ.

𝑛 1
(ii) Var [ 𝑆 2 ] = 𝑉𝑎𝑟{ ∑(𝑋𝑖 − 𝑋̅ )2 }
𝑛−1 𝑛−1
1
= 𝑉𝑎𝑟 {∑𝑋𝑖 − 𝑋̅ )2 }
(𝑛−1)2

1 𝜎2
= ∑{ 𝜎 2 − }
(𝑛−1)2 𝑛

Using theorem 5.1

1 𝜎2
= (𝑛𝜎 2 − 𝜎 2 ) =
(𝑛−1)2 𝑛−1

6
 𝑛 𝜎2
lim { 𝑆 2 } = lim [ ]→0
𝑛→∞ 𝑛−1 𝑛→∞ 𝑛−1
𝑛
∴ 𝑆 2 is a consistent estimator for σ2
𝑛−1

(d) Sufficiency
An estimator is said to be sufficient if it extracts from the sample every bit of available
information relative to the parameter.
Example 5.3
A random sample X1, X2, X3, ……..,Xn is drawn from a distribution with mean μ and
variance σ2 both assumed unknown. Consider the statistic
1
T= ∑𝑋𝑖 . Show that Var(T) < Var(𝑋̅ ) for all values of n. explain why this does not
𝑛+1

contradict the fact that 𝑋̅ is the most efficient estimator of μ.

Is T a consistent estimator for μ?

Solution:
𝜎 2
Now Var(𝑋̅ ) = ,
𝑛

1 1 𝑛𝜎 2
Var (T) = Var [ ∑𝑋𝑖 ] = ∑𝜎 2 =
𝑛+1 𝑛+1 (𝑛 + 1)2
𝑛𝜎2 𝜎2
To show that < ,
(𝑛+1)2 𝑛

Clearly for n> 1, 𝑛2 < (𝑛 + 1)2

𝑛2 (𝑛+1)2
Divide both sides by n (n+1)2 to get <
𝑛(𝑛+1)2 𝑛(𝑛+1)2
𝑛 1
i.e. <
(𝑛+1)2 𝑛

𝑛𝜎2 𝜎2
Multiply both sides by σ2 to get <
(𝑛+1)2 𝑛

∴ 𝑉𝑎𝑟(𝑇) < 𝑉𝑎𝑟(𝑋̅ )

Now for an estimator to be efficient, it must of necessity be unbiased.
1 1
E(𝑋̅ ) = μ (obvious), but 𝐸 [𝑇] = 𝐸 [ ∑𝑋𝑖 ] = ∑𝐸 (𝑋𝑖 )
𝑛+1 𝑛+1

7
 1 𝑛
= 𝑛𝜇 = 𝜇 ≠ 𝜇
𝑛+1 𝑛+1

Hence T is not an unbiased estimator for μ. Thus the fact that 𝑋̅ is the most efficient
unbiased estimator for μ has not been contradicted, and the biasedness of T precludes
it from being a consistent estimator for μ.

5.3 Pooled Estimates

We can also make point estimates of parameters based on results of more than one
random sample, taken however from the same population, and that is the essence of
the following theorem:
Theorem 5.4
Let 𝑋̅ and 𝑌̅ be the sample means of two random samples of respective sizes n and m,
drawn from a population with unknown mean μ and variance σ2Let𝑆𝑥2 𝑎𝑛𝑑 𝑆𝑦2 denote
also the respective sample variances for the two samples. Then.
𝑛𝑋̅ +𝑚𝑌̅
(a) an unbiased estimator for μ is given by 𝜇=
𝑛+𝑚
2 2
𝑛 𝑆𝑋 +𝑚𝑆𝑌
(b) an unbiased estimator for σ2 is given by σ2 =
𝑛+𝑚−2

Proof
{𝑛𝑋̅ +𝑚𝑌̅} 1 1
(a) 𝐸[ ]= 𝐸 (𝑛𝑋̅ + 𝑚𝑌̅) = (𝑛𝜇 + 𝑚𝜇) = 𝜇
𝑛+𝑚 𝑛+𝑚 𝑛+𝑚

(b) Using Theorem 5.1

2 2
{ 𝑛 𝑆𝑋 } { 𝑚 𝑆𝑌 }
𝐸 = 𝜎 2 𝑎𝑛𝑑 𝐸 = 𝜎2
𝑛−1 𝑚−1

Hence 𝐸 ( 𝑛 𝑆𝑋2 ) = 𝜎 2 ( 𝑛 − 1) and 𝐸 ( 𝑛 𝑆𝑌2 ) = 𝜎 2 (𝑚 − 1)

Add both to get 𝐸 ( 𝑛 𝑆𝑋2 ) + 𝐸 ( 𝑚 𝑆𝑌2 ) = 𝜎 2 ( 𝑚 + 𝑛 − 2 )
2 2
{ 𝑛 𝑆𝑋 + 𝑚 𝑆𝑌 }
𝐸 = 𝜎2
𝑚+𝑛−2

Exercise
Two random samples of size 8 and 7 respectively are drawn from a population as
follows:

8
X: 6, 8, 9, 10, 17, 14, 13, 11. and Y: 9, 10, 6, 12, 7, 11, 8.
From these two samples, calculate the unbiased estimates of the population mean and
variance. (Answer: 10.07, 8.92)

5.4 Interval Estimation

Sometimes, we might be interested in constructing a range of values within which a
population parameter is expected to lie with a given probability. Such a probability
range for a parameter is called a confidence interval, and the degree of confidence is
the probability that the parameter lies in that interval. Interval estimation is,
therefore, concerned with locating a range of values within which a population
parameter is expected to lie with a given degree of confidence or probability. This
degree of confidence is usually expressed in percentage or in probability form. A 100(
1- α )100% confidence interval for some unknown parameter θ is an interval
constructed based on the results of a random sample so that the probability that θ
lies in that interval is ( 1 - α).

5.5 Confidence Intervals for Mean

(a) When Population Variance is known:
Consider a random sample X1, X2, … ,Xn of size n, from a normal population having a
known variance σ2. We wish to construct a 100(1-α)% confidence interval for the
population mean, µ. Recall that from a normal population, the sample mean, 𝑋̅ , has
𝜎2
mean µ and variance, , hence by central limit theorem,
𝑛
𝜎2 (𝑋̅ −µ)
𝑋̅ ~ 𝑁 ( µ , ) 𝑎𝑛𝑑 ~ 𝑁(0,1) … … …(1)
𝑛 𝜎/√𝑛

Let 𝑍𝛼⁄2 be the value of the standard normal variable to the right of which is an area
of 𝛼⁄2under the density function (fig. 5.2). Then we can write.
𝑋̅ − 𝜇
𝑃(−𝑍∝/2 ≤ < 𝑍∝/2 = 1−∝ … … (2)
𝜎/√𝑛

9
Multiply each term in the inequality of (2) by σ √𝑛, subtract 𝑋̅ x from each term and
multiply through by -1, to get

𝑃 (𝑋̅ − 𝑍𝛼 𝜎⁄ < µ < 𝑋̅ + 𝑍𝛼 𝜎⁄ ) = 1− ∝

2 √𝑛 2 √𝑛
Thus the (1−∝)100% confidence interval for µ when σ2 is known is
𝑋̅ − 𝑍∝/2 𝜎⁄ < µ < 𝑋̅ + 𝑍∝/2 𝜎⁄
√𝑛 √𝑛

𝛼/2
𝛼/2

−𝑍𝛼/2 𝑍𝛼/2
Example 5.4
Suppose that from a random sample, n = 20, and 𝑋̅ = 64.3 . If the variance of the
population is known to be σ2 = 225, then to obtain a 95% Confidence Interval for the
mean of the population from which the sample was drawn.

Solution
𝜎 𝜎
A (1−∝)100%CI for µ is given by 𝑋̅ − 𝑍𝛼 < µ < 𝑋̅ + 𝑍𝛼
2 √𝑛 2 √𝑛

1−∝ = 0.95, ∝ = 0.05, 𝜎 = 15, 𝑍∝/2 = 1.96

Hence the CI is given by 64.3 − 1.95 × 15/√20 < µ < 64.3 + 1.95 × 15/√20
= 57.726 < µ < 70.874
Hence, the 95% C.I. for µ is (57.7, 70.9)

Example 5.5

10
Find a 95% C.I. for the true population mean µ if a random sample of 12 from the
population with variance 124 yielded 𝑋̅ = 51.2.
𝟏𝟐𝟒
Solution: CI is given by 𝟓𝟏. 𝟐 ± 𝟏. 𝟗𝟔 × √ = (44.9, 57.5)
𝟏𝟐

For small samples selected from population that do not satisfy the normality
assumption, we cannot expect our degree of freedom to be accurate. However for
samples of sizes n ≥ 30, good results are guaranteed by theory.

(b) When Population Variance is Unknown

For a normal population where the population variance is unknown, and it is
impossible to obtain a sample of size n > 30, confidence intervals can be constructed
by using the sampling distribution of the statistic
(𝑋̅ −µ)√𝑛
𝑇= …… … … … …(1)
𝑆

where S is the sample variance. The procedure is same as in (a) above except that we
use the t - distribution in place of the standard normal.

𝛼/2
𝛼/2
−𝑡𝛼 𝑡𝛼
2 2

Fig. 5.3
From the figure above (fig. 5.3)
𝑃(−𝑡𝛼 < 𝑇 < 𝑡𝛼 ) = 1−∝ … … … …(2)
2 2

Where 𝑡𝛼 is the t - value with n-1 degrees of freedom, to the right of which there is an
2

area of ∝/2. By symmetry, an equal area of ∝/2 will fall to the left of −𝑡𝛼 . Substituting
2

for T in (2), we get

11
 𝑋̅ −µ
𝑃(−𝑡𝛼⁄2 < ( ) √𝑛 < 𝑡𝛼⁄2 ) = 1−∝
𝑆

Appropriate algebra yields

𝑆 𝑆
𝑃(𝑋̅ − 𝑡𝛼⁄2 . < µ < 𝑋̅ + 𝑡𝛼⁄2 . ) = 1−∝
√𝑛 √𝑛

Thus for a small sample, where n < 30 and with σ2 unknown, a 100(1−∝)%
confidence interval for the mean µ is given by
𝑆 𝑆
𝑋̅ − 𝑡𝛼⁄2 . < µ < 𝑋̅ + 𝑡𝛼⁄2 .
√𝑛 √𝑛

However, for large samples (with size n > 30), a 100(1−∝)% confidence interval for
µ is given by
𝑆 𝑆
𝑋̅ − 𝑍𝛼⁄2 . < µ < 𝑋̅ + 𝑍𝛼⁄2 .
√𝑛 √𝑛

Where 𝑍𝛼⁄2 𝑍 ∝/2 is as defined earlier. This result is possible because for large values
of n, the T-distribution closely approximates that of a standard normal distribution.
Moreover when n is large, S2 is a good estimator for σ2 .
𝑛
We had that 𝑆 2 is an unbiased estimator for σ2. Thus, when n is large enough, we
𝑛−1
𝑛
shall have that E[S2 ] = σ2 , since → 1 𝑎𝑠 𝑛 → ∞ .
𝑛−1

Example 5.6
Suppose a paint maker wishes to determine the true average drying time of a new
paint. He tests twelve areas of equal sizes, and gets a mean of 66.3 minutes and a
standard deviation of 8.4 minutes. Based on this sample, construct a 95% CI for the
actual drying time of the paint.
Solution
∝ = 0.05, 𝑛 = 12, 𝑋̅ = 66.3, S = 8.4, 𝑡𝛼⁄2 = 2.201, since sample size is small.
𝑆 𝑆
Therefore CI for µ is given by 𝑋̅ − 𝑡𝛼⁄2 . < µ < 𝑋̅ + 𝑡𝛼⁄2 .
√𝑛 √𝑛

Which gives 66.3 ± 5.337 i,e. 60.963 < µ < 71.637 is the 95% C.I.
5.6 Error in Estimating the mean

12
The 100(1−∝)% 𝐶. 𝐼. provides an estimate of the accuracy of the point estimate 𝑋̅ .
We must mention that most of the time 𝑋̅ will not be exactly equally to µ, so the point
estimate 𝑋̅ will be in error. The size of this error will be the absolute difference
between 𝑋̅ and µ, and we can only be 100(1−∝)% confident that this difference will
𝜎
not exceed𝑍𝛼 . .
2 √𝑛

Furthermore we may wish to know how large a sample will be necessary to ensure
that the error in estimating µ will not exceed a specified quantity K. We can be
100(1−∝)% confident that the error will not exceed a specified amount K when the
sample size is
(𝑍𝛼⁄2 . 𝜎)2
n =
𝐾
Errors can really be estimated when σ2 is known or when n ≥ 30, otherwise, we may
not expect our level of confidence to be reliable.

5.7 CI for Difference Between Two Means

Consider two normal populations X and Y with means µ x and µy, and variance
𝜎𝑋2 𝑎𝑛𝑑 𝜎𝑌2 , respectively. Let 𝑋̅ and 𝑌̅ be the respective means of two samples of size
n and m drawn from the populations, respectively.
Now 𝑋 ~ 𝑁( µ𝑥 , 𝜎𝑥2 )and also, 𝑌 ~ 𝑁( µ𝑦 , 𝜎𝑦2 )
Obviously, it can be shown that
𝜎𝑦 2 2
𝜎
𝐸 (𝑋̅ − 𝑌̅) = µ𝑥 − µ𝑦 and that 𝑉𝑎𝑟(𝑋̅ − 𝑌̅) = 𝑥 −
𝑛 𝑚
{(𝑋̅ − 𝑌̅) − (µ𝑥 − µ𝑦 )}
Let 𝑍 =
2 𝜎 2
√𝜎𝑥 + 𝑦
𝑛 𝑚

that is, Z has standard normal distribution and 𝑃 (−𝑍𝛼 < 𝑍 < 𝑍𝛼 ) = 1− ∝
2 2

Which implies that µ𝑥 − µ𝑦 has probability of 1−∝ of being in the interval

𝜎 2 𝜎𝑦2
(𝑋̅ − 𝑌̅) ± 𝑍∝/2 √ 𝑥 +
𝑛 𝑚

13
Hence a 100(1−∝)% confidence interval for µ𝑥 − µ𝑦 when population variances are
known is given by
𝜎 2 𝜎𝑦2
(𝑋̅ − 𝑌̅) ± 𝑍∝/2 √ 𝑥 +
𝑛 𝑚

However, when population variances are unknown, and cannot be estimated from
large samples, then the sample equivalents will be used, resulting in the confidence
1 1
interval, (𝑋̅ − 𝑌̅) ± 𝑡∝/2 S ∗ √ + … …(2)
𝑛 𝑚

Where S*2 is the pooled estimator given by

(𝑛−1)𝑆𝑥2 +(𝑚−1)𝑆𝑦2
𝑆 ∗2 = … … … … … (3)
𝑛+𝑚−2

Therefore the CI is given by

(𝑛 − 1)𝑆𝑥2 + (𝑚 − 1)𝑆𝑦2 1 1
(𝑋̅ − 𝑌̅) ± 𝑡∝/2 √ ( + )
𝑛+𝑚−2 𝑛 𝑚

Example 5.7
A sample size of 15 from a normal population with mean µx and variance 60, yields a
sample mean of 𝑋̅ = 70.1; while an independent sample of size 8 from another normal
population with mean µy and variance 40 had a sample mean of 𝑌̅ = 75.3 . Find a
95% CI and another 90% CI for µx - µy .
Solution
σ2x = 60, σ2y = 40, 𝑋̅ = 70.1, 𝑌̅ = 75.3, n = 15, m = 8
𝜎 2 𝜎𝑦2
The appropriate CI is given by (𝑋̅ − 𝑌̅) ± 𝑍∝/2 √ 𝑥 +
𝑛 𝑚

For a 95% CI, 1−∝= 0.95, ∝ = 0.05, 𝑍∝/2 = 1.96

60 40
∴ 𝐶. 𝐼. = (70.1 − 75.3) ± +1.96√ +
15 8

= -5.2 ± 5.88 = (-11.08, 0.68)

Hence, the CI is -11.08 < µx - µy< 0.68

14
(ii) For a 90% CI, ∝= 0.1, 𝑍∝/2 = 1.645

60 40
C.I. = -5.2+1.645√ + ; Hence, -10.135 < µx - µy < - 0.265
15 8

5.8 CI For Proportions

To estimate a proportion or probability, we assume that we are sampling from a
binomial population with parameter θ. A point estimator of θ is given by the statistic
𝑇 = 𝑋/𝑛, where X represents the number of successes and n the number of trials.
𝑥 𝑛𝜃
𝐸 (𝑇 ) = 𝐸 [ ] = =𝜃
𝑛 𝑛
𝑋 1 1 𝜃(1−𝜃)
𝑉𝑎𝑟(𝑇) = 𝑉𝑎𝑟 [ ] = 𝑣𝑎𝑟(𝑋 ) = 𝑛𝜃 (1 − 𝜃) =
𝑛 𝑛2 𝑛2 𝑛

To construct a 100( 1 - α )% CI for θ, observe that for large n,

𝑋⁄ − 𝜃
𝑛
𝑍= has a standard normal distribution
√𝜃(1−𝜃)

∴ 𝑃(−𝑍∝/2 < 𝑍 < 𝑍∝/2 ) = 1 − 𝛼

Substitute for Z, and resolve to obtain
𝑋 𝜃(1 – 𝜃) 𝑥 𝜃(1−𝜃)
𝑃 [ − 𝑍𝛼 √ < 𝜃 < + 𝑍𝛼 √ ]=1−𝛼
𝑛 2 𝑛 𝑛 2 𝑛

It is difficult to manipulate the inequalities to obtain an interval whose end points are
𝑋
independent of θ. If n is large we use the point estimate for θ at the end points to get
𝑛

a 100(1 − 𝛼)% confidence interval as

X X
X (1 − )
± Zα √n n
n 2 n
Example 5.8
If a sample yielded 140 successes in 400 trials, and the assumptions underlying the
binomial distribution are met, construct a 95% CI for θ, the mean of the distribution.
Solution:
140
X = 140, n = 400, hence, θ = = 0.35, 𝑍𝛼/2 = 1.96
400

15
 0.35(0.65)
Therefore, a 95% CI for θ is given as: 0.35 ± 1.96√
400

Which gives 0.303 < θ < 0.397

Example 5.9
Before a bye-election for which there were two candidates, A and B, it was discovered
in a sample among 400 voters that 208 of them preferred candidate A to B. Construct
a 95% CI for the actual percentage of voters favourable to A. If in fact 55% of the
voters were in favour of B, what is the probability that a random sample of 400 voters
will contain at least as many in favour of A as there are for B?

Solution
X = 208, n = 400
Let θ represent the true population proportion who intend to vote for A.
208
A point estimator of θ is = 0.52and a 95% CI for θ is
400

(0.52)(0.48)
0.52 ± 1.96√
400

∴ 0.471< θ < 0.569 or 47.1% < θ < 56.9%

Case II
Here θ = 0.45, and our interest is in the distribution of θ for a sample of 400.
𝑋⁄ ~ 𝑁(𝜃, 𝜃(1−𝜃)
𝑛 ) , where θ = 0.45 and n = 400
𝑛

∴ 𝑋⁄𝑛 ~𝑁(0.45, 0.00062). Since 𝑋⁄𝑛 is the sample vote proportion for A,
the proportion for B will be 1 - 𝑋⁄𝑛. Hence for A to have at least half of the voters in
favour, we require P(𝑋⁄𝑛≥ 1- 𝑋⁄𝑛 ), which is equivalent to P(𝑋⁄𝑛≥ ½ )
0.5−0.45
Let 𝑍= (0.45)(0.55)
= 2.01
√
400

𝑃(𝑍𝛼 ≥ 2.01 = 1 − 𝑃(𝑍𝛼 < 2.01) = 1 - 0.4778 = 0.5222

If 𝑋⁄𝑛 is used as an estimate of θ then we can be 100(1 − 𝛼)% confident that

16
 X X
(1− )
(1) the error will not exceed: 𝑍𝛼 √n n
n

(ii) The error will not exceed a specified amount of K when the sample size is
2 X⁄n[ 1−X⁄n]
𝑍𝛼/2 .
n
𝑛 =
𝐾2
In the last expression we must note that
𝑥 𝑥 1
(1 − ) ≤
𝑛 𝑛 4

With this we can be at least 100(1 − 𝛼)% confident that the error will not exceed a
specified amount K when the sample size is
2
𝑍𝛼/2
𝑛=
4𝐾 2
When solving for the sample size, n, all fractional values are rounded up to the upper
integer. When we are to estimate the difference between two proportions, a
100(1 − 𝛼)% CI for the difference between two binomial parameters θ1 and θ2 is
given by
𝑋1 𝑋 𝑋2 𝑋
𝑋1 𝑋2 (1− 1 ) (1− 2 )
𝑛 𝑛 𝑚 𝑚
( − ) ± 𝑍𝛼/2 𝑆, where 𝑆 = +
𝑛 𝑚 𝑛 𝑚

and n is the sample size of the sample from the first binomial population with
parameter θ1 , and m is the sample size of the sample from second binomial
population with parameter θ2.

5.9 CI for Variance

Given a random sample of size n from a normal population we can obtain a
100(1 − 𝛼)% confidence interval for σ2 by using the statistic of the Chi-square
distribution given by
𝑛−1 2
Χ2 = 𝑆
𝜎2
With n-1 degrees of freedom, and S2 is the variance of the random sample. The X2
values cannot be negative, hence the curve is not symmetric about 0.

17
From the distribution, the probability that a random sample produces a X 2 value
greater than some specified value is equal to the area under the curve to the right of
this value.

𝛼/2
𝑋𝛼2

Fig. 5.4
𝑋𝛼2 is used to represent the X2 value to the right of which we find an area of 𝛼 (fig. 5.4).

𝛼/2 𝛼/2
2
𝑋1−𝛼/2 2
𝑋𝛼/2

Fig. 5.5
From this figure we see that
2 2
𝑃(𝑋1−𝛼/2 < 𝑋 2 < 𝑋𝛼/2 ) =1−𝛼 … … … …(2)
2 2
Where 𝑋1−𝛼/2 and 𝑋𝛼/2 are values of the chi-square distribution with n - 1 degrees of
freedom, leaving areas under the curve of 1 − 𝛼/2 and 𝛼/2 respectively, to the right.
From (1) and (2) we have

18
 2 (𝑛−1) 2
𝑃 (𝑋1−𝛼/2 < 𝑆 2 < 𝑋𝛼/2 )=1−𝛼
𝜎2

Resolving algebraically yields

(𝑛−1) (𝑛−1)
𝑃( 2 𝑆2 < 𝜎 2 < 2 𝑆2) = 1 − 𝛼
𝑋𝛼/2 𝑋1−𝛼/2

Thus a 100(1 − 𝛼)% confidence interval for σ2 is given by

𝑛−1 𝑛−1
2 𝑆2 < 𝜎 2 < 2 𝑆 2 ………………………………………(3)
𝑥𝛼/2 𝑥1−𝛼/2

With n -1 degrees of freedom for the chi-square values.

Example 5.10
Suppose in 16 test runs, gas consumption of an engine had s = 2.2 litre. Construct a
99% confidence interval for σ2 as a true indication of the variability of the gas
consumption.

Solution:
Assuming that the data comes from a normal population,
n = 16, s = 2.2, S2 = 4.84, 𝛼 = 0.01, 1 −α/2= 0.995,
2 (15)
𝑋𝛼/2 = 32.80, X21-α/2(15) = 4.60
15(4.84) 15(4.84)
The CI is given by ≤ 𝜎2 ≤ ie 2.213 ≤ 𝜎 2 ≤ 15.783
32.8 4.60

Example 5.11
A random sample of 8 from an approximately normal population gave the following
values: 12, 11, 12, 8, 12, 10, 13, 10. Construct a 90% confidence interval
for the variance of the population.

Solution
𝑋̅ = 11, 𝑆 2 = 2.57, 𝛼 = 0.1, 2
𝑋.05 2
= 14.067, 𝑋.95 = 2.167
The CI is given by
7×2.57 7×2.57
< 𝜎2 < ie 1.279 < σ2< 8.302.
14.067 2.17

19
Note that here S2 is the unbiased estimate of the population variance.
To estimate the ratio of two variances,𝜎𝑋2 and 𝜎𝑌2 , we use the point estimate which is
𝑆𝑋2 /𝑆𝑌2 which is the ratio of the sample variances. If 𝜎𝑋2 and 𝜎𝑌2 are the variances of two
normal populations, we can establish an interval estimate of 𝜎𝑋2 / 𝜎𝑋2 by using the
statistic
𝑆𝑥2 /𝜎𝑥2 𝜎𝑦2 𝑆𝑥2
𝐹= = … … .(5)
𝑆𝑦2 /𝜎𝑦2 𝜎𝑥2 𝑆𝑦2

Whose sampling distribution is the F- distribution. Theoretically therefore, the F-

statistic is the ratio of two independent X2 variables, each divided by their degrees of
freedom. The statistic F possesses an F- distribution with
n -1 and m -1 degrees of freedom, where n is the size of the first sample and m is the
size of the second sample. The degree of freedom of the sample variance in (5) is
always stated first. By appropriate manipulations, a 100(1 − 𝛼)% C.I. for 𝜎𝑋2 / 𝜎𝑌2 is
given by
2 2 2
𝑆𝑋 𝜎𝑋 𝑆𝑋 𝑓𝛼/2 (𝑉1 ,𝑉2 )
2𝑓 < <
𝑆𝑌 𝛼/2 (𝑉1 ,𝑉2 ) 𝜎𝑌2 2
𝑆𝑌

Where v1 = n - 1, v2 = m - 1 and 𝑓𝛼/2 (v1,v2) is the F-value with (v1, v2) degrees of
freedom, leaving an area of 𝛼/2 to the right.

CHAPTER TWO
TESTS OF HYPOTHESES
6.1 Introduction
A statistical hypothesis is a statement or conjecture about a given population.
Hypothesis testing involves the formulation of a set of rules which will enable us to
make decision (reject or accept a statement) about the given population. In a simple
hypothesis, the functional form of the underlying distribution, as well as the values of
the parameters are stated, Whereas in composite hypothesis, the functional form of
the distribution may be stated without the exact value of the parameter. An example

20
of a simple hypothesis is the statement “the population is binomially distributed with
parameter θ=0.48; while that of a composite hypothesis is “the population is
binomially distributed with parameter θ > 0.48 .
Rejecting a hypothesis means that it is false on the basis of some evidence
provided by a test. Accepting a hypothesis means that we have no evidence to believe
otherwise. A null hypothesis is the statistical assumption we wish to verify or
possibly disprove. In general format, it assumes no deviation from the normal, it is
usually stated in null form with equality constraints, and most often denoted with Ho.
It is the actual focus of a statistical test. The alternative hypothesis, denoted with H 1
or HA, is that hypothesis that is automatically accepted on the rejection of H 0. It is a
negation of the null hypothesis in a statistical sense. It usually puts emphasis on a
range of values. To enable us test H0 against H1, we partition the sample space of
outcomes into two disjoint, mutually exclusive and exhaustive regions called the
acceptance region for H0, and the rejection region for H0.
The rejection region is also called the critical region. The values separating the
acceptance region from the rejection are called critical values. The size of a critical
region is the probability of obtaining an outcome which falls in that critical region.
A type I error is committed when we reject the null hypothesis whereas it is
true. A type II error is committed if we accept the null hypothesis, whereas it is false,
and mathematically, we write
P (type I error) = P(Rejecting H0 given that H0 is true) = α
P (type II error) = P(Accepting H0 given that H0 is false) = β
The value of 𝛼 is the level of significance of the test. The probability of occurrence of
both errors can be decreased by increasing the sample size. The power of any test is
an indication of how well that test will enable us to minimize type II error. A null
hypothesis concerning a population parameter will always be stated so as to specify
an exact value, whereas the alternative allows for several values of the parameter.

21
6.2 General Test Procedure
Performing classical statistical test will involve the following steps:
(i) Make a short summary of known facts, such as nature of the distribution,
population parameters, or sample statistics, etc.
(ii) Formulate the null hypothesis H0
(iii) Formulate the alternative hypothesis, H1, whose acceptance is implied by the
rejection of H0.
(iv) Choose or identify the level of significance (size of the CR)
(v) Determine the appropriate test statistic and the corresponding CR
(vi) Formulate a test criterion.
The decision is usually to Reject the null hypothesis if the value of the test statistic
falls in the CR, otherwise do not reject. A test of any statistical hypothesis where the
alternative is one sided is called a one-tailed test.
An example of such is H0: θ = θo versus H1: θ >θo
Or Ho: θ = θo versus H1: θ <θo .
The Critical Region for H1: θ >θo lies entirely in the right tail of the distribution, while
that of H1: θ <θo lies entirely in the left tail. A test of any statistical hypothesis where
the alternative is two sided is called a two tailed test. The Critical Region is split into
two equal parts, and located in each tail of the distribution of the test statistic. An
example is
Ho: θ = θo versus H1: θ ≠ θo
A test is said to be significant if Ho is rejected at α = 0.05, and highly significant if
H0 is rejected at α = 0.01.

6.3 Tests Concerning Population Means

(a) When population Variance is known.
Recall that by the central limit theorem, if a random variable X is normally distributed
with population mean µ and variance σ2 then 𝑋̅ is normally distributed with mean µ
2 𝑋̅ − 𝑢
and variance 𝜎 ⁄𝑛. Moreover, 𝑍 = is a standard normal variable (with mean,
𝜎/√𝑛
0 and variance, 1). To test for the population mean, under the null hypothesis, we

22
present below the appropriate hypotheses, test statistic, and decision criteria (or
critical region).
Case I: To test Ho : µ = µ0 Versus H1: µ < µ0
This is a left sided one-tailed whose Test statistic is
̅ − μ0 )
(X
𝑍= σ
⁄ n
√
For α level of significance, Reject H0 if Z < - Zα

Case II: To test H0: µ = µ0 Versus H1: µ > µ0

This is a right sided one-tailed test whose test statistic remains Z as defined in case 1
above. But for an α level of significance reject H0 if Z > Zα

AR AR

-Z Z
Fig. 6.1 LS one tailed test (case I) Fig. 6.2 RS one tailed test (case II)

Case III: To test H0: µ = µ Versus H1: µ ≠ µ0,

(𝑥−µ)√n
This is a two tailed test whose test statistic is 𝑍 =
σ2

and the decision rule is reject H0 if | Z | > Zα/2

CR CR
−𝑍 ∝⁄2 𝑍 ∝⁄2

Fig. 6.3 Two tailed test (case III)

Example 6.1

23
Suppose a certain type of 100 watts bulb has been standardized so that the mean life
of the bulbs is 1000hrs and standard deviation is 128 hrs. A sample of 16 of these
bulbs having mean µ was tested and found to have a mean of 967.8hrs. Test at both
1% and 5% levels of significance the hypothesis that the actual mean is less than
1000hrs.
Solution:
𝑋̅ = 967.8, µo = 1000, σ = 128, n = 16
H0: µ = 1000 H1: µ < 1000
(967.8−1000)√16
𝑍= = −1.00625
128

From standard normal table for 𝛼= 0.01, Z0.01 = 2.33

𝛼 =0.05, Z0.05 = 1.645
At 1% level of significance Z > - Z0.01, so we do not reject H0
Also at 5% level Z > - Z0.05 , so H0 cannot be rejected either.

Example 6.2
Suppose we know that the breaking strength of a certain type of steel bar has a normal
distribution with mean µ and variance 25. The manufacturing process is changed as
a result of research and an observed sample of breaking strength of 100 steel bars
had a mean of 77.8. However the people who were involved in the decision to change
the process have conjectured that µ is above 80. Test at 5% level of significance the
reliability of their proposition.
Solution:
H0: µ = 80, H1: µ ≠ 80, 𝑋̅ = 77.8, µ0 = 80, σ = 5
This is a two-tailed test, so the test statistic is
(𝑋̅ − μ0 )√n (77.8−80)√100
|𝑍|=| |=| | = 4.4
σ 5

At 5% level of significance Z0.025 = 1.96

Since Z >Z0.025 , we reject H0 and conclude that their conjecture may not be correct.

24
(b) When Population Variance is Unknown
(𝑋̅ − 𝜇0 )√n
The statistic 𝑇=
S

has a t-distribution with n-1 degrees of freedom. Hence for test concerning the mean
of a normal population where the population variance is unknown, the procedure is
as follows:
Case I
H0: µ = µ0 Versus H1: µ < µ0 , Reject H0 if T < -tα(n-1)
Case II
H0: µ = µ0 Versus H1: µ > µ0 , Reject H0 if T > tα(n-1)
Case III
H0: µ = µ0 versus H1: µ ≠ µ0 , Reject H0 if |T| > tα/2(n-1)

Example 6 .3
Eight different determination of alcoholic contents in a bottle of wine yielded a
sample mean of 𝑋̅ = 16.6% with s=0.06%. If µ is the population mean of the
determination, then test at both 10% and 5% levels of significance the following
hypotheses:
(a) H0: µ = 16.64% Versus H1: µ < 16.64%
(b) H0: µ = 16.64% Versus H1: µ ≠ 16.64%

Solution:
(16.6−16.64)√8
𝑇= = −1.886
0.06

(a) H0: µ = 16.64 VsH1: µ < 16.64

At 10% level, t0.1(7) = 1.415, hence we reject Ho
At 5% level t0.05(7) = 1.895, hence we do not reject Ho
(b) H0: = 16.64% VsH1:U = 16.64%
At 10% level, 𝛼/2 = 0., 05, 𝑡0.05 (7) =1.895, hence we do not reject Ho

25
At 5% level 𝛼/2 = 0.025, 𝑡0.025 (7) = 2.365, hence we do not reject Ho

6.4 Tests Concerning Difference Between Two Means

(a) When Population Variances are Known
Consider two normal populations X (with mean µX and variance 𝜎𝑥2 ) and Y
(with mean µY and variance 𝜎𝑥2 ). Suppose random samples of sizes n and m are drawn
from X and Y, respectively. In testing for difference between µ X and µY, the following
procedure will be followed for the various sided tests:
Case I
H0: µX - µY = 0 Versus H1: µX - µY < 0
Which can be equivalently stated as
H0 : µX = µY Versus H1: µX< µY
The test statistic is
(𝑋̅ −𝑌̅) −(𝜇𝑋 −𝜇𝑌 )
𝑍= 2 2
… … … … …(1)
√𝜎𝑥 + 𝜎𝑌
𝑛 𝑚

But under H0: µX - µY = 0. Hence the actual test statistic is

(𝑋̅ −𝑌̅)
𝑍= 2 2
… … … … … (2)
√𝜎𝑥 + 𝜎𝑌
𝑛 𝑚

Reject H0if Z< -Zα, where α is the level of significance

Case II: For H0: µx = µy Versus H1: µx> µy
Reject H0 if Z > Zα, .
Case III: For H0: µx - µy = 0 Versus H1: µx - µy ≠ 0
Reject H0 if | Z | > Zα/2 at α level of significance.

(b) When Population Variances are Unknown

When the population variances are unknown, but are assumed to be equal, we instead
make use of the sample variances and the appropriate t-test follows. The test Statistic
(𝑋̅ −𝑌̅) (𝑛−1)𝑆𝑥2 +(𝑚−1)𝑆𝑦2
becomes 𝑇= , where 𝑆 2 = .
1 1 𝑛+𝑚−2
𝑠√ +
𝑛 𝑚

26
Hence the decision criteria becomes (at n+m-2 degrees of freedom),
Case I : H0 : µX = µY Versus H1: µX< µY. Reject H0 if T < - tα,
Case II: H0: µx = µy Versus H1: µx> µy . Reject H0 if T > tα, .
Case III: H0: µx = µy Versus H1: µx ≠ µy . Reject H0 if |T | > tα/2 .

(c) When Testing for difference in means, it may sometimes happen that H1 may be
in the form:
H1: µx - µy < d, for the LS one tailed test, or
H1: µx - µy > d, for the RS one tailed test, or
H1: µx - µy ≠ d, for the two tailed test.
In that case, if 𝜎𝑥2 and 𝜎𝑌2 are both known, the test statistic will be
(𝑋̅ − 𝑌̅)−𝑑
𝑍= 2 2
…………………………………………(3)
√𝜎𝑥 + 𝜎𝑌
𝑛 𝑚

and decision will be as described in 6.4 (a) above.

But if 𝜎𝑥2 and 𝜎𝑌2 are both unknown but equal, the sample equivalents yields the test
statistic
(𝑋̅ − 𝑌̅)−𝑑 (𝑛−1)𝑆𝑥2 +(𝑚−1)𝑆𝑦2
𝑇= , where 𝑆 2 = … … (4)
1 1 𝑛+𝑚−2
𝑆√ +
𝑛 𝑚

and decision proceeds as in 6.4 (b)

Finally, if𝜎𝑥2 and 𝜎𝑌2 are unknown and not equal, the test statistic will be
(𝑋̅ – 𝑌̅)– 𝑑
𝑇= 2 2
… … … … (5)
√𝑆𝑥 + 𝑆𝑌
𝑛 𝑚

But the degree of freedom will be the integer closest to

𝑆2 𝑆2
( 𝑥 + 𝑥)2
𝑛 𝑛
V= 𝑆2 𝑆2
… … … …(6)
𝑥 𝑦
+
𝑛2 (𝑛−1) 2
𝑚 (𝑚−1)

Decision is still as in 6.4 (b).

6.5 Tests Concerning Variances

(a) Testing for a Single Variance
To test whether a random sample of size n with sample variance S2, is drawn from a
normal population with variance σ2, we use the test statistic

27
 2
(𝑛 − 1)𝑆 2
𝑋 =
𝜎02
This is the chi-square value for testing H0: σ2 = 𝜎02 against a relevant alternative. The
statistic above has a chi-square distribution with n-1 degrees of freedom. It must be
mentioned that the chi-square distribution is not symmetric about any axis, however
the two tails are of equal probability.

𝐴𝑅 𝐶𝑅

𝑿𝟐∝ (𝒏 − 𝟏)
Fig. 6.4 chi-square distribution
The test procedure would be as follows
Case I: To test H0: σ2 =𝜎02 versus H1: σ2 <𝜎02 , Reject H0 if X2 <𝑋1−𝛼
2
(n-1) .
Case II :To test H0: σ2 <𝜎02 versus H1: σ2 >𝜎02 , Reject H0 if X2 >𝑋𝛼2 (n-1) .
Case III: To test H0: σ2 =𝜎02 versus H1: σ2 ≠ 𝜎02 ,
2 2
Reject H0 if X2 <𝑋1−𝛼/2 (n-1) or if X2 >𝑋𝛼/2 (n-1) .

CR
AR AR CR

Fig. 6.5 CR for 𝜎 2 < 𝜎02 Fig. 6.6 CR for 𝜎 2 < 𝜎02

28
 AR CR
2
𝑋1−𝛼/2 2
𝑋𝛼/2
Fig. 6.7 CR for 𝜎 2 ≠ 𝜎02

Example 6.4
A sample of size 9 has variance of 8.01. Test at 5% whether the sample is likely to
have been drawn from a normal population with variance of 9.0

Solution
This is a two tailed test with H0: σ2 = 9,H1: σ2 ≠ 9, n=9, S2 = 8.01.
(𝑛−1)𝑆 2 8×8.01
Test statistic is 𝑋2 = = = 7.12
𝜎02 9
2 2
𝜶 = 0.5 𝑋0.025 (8) = 17.53, 𝑋0.975 (8) = 2.18
X2 0.975 = 2.18, x20.025 (8) = 17.53
Since 2.18 < 7.12< 17.53, we do not reject Ho and hence, we conclude that it is likely
that the sample was drawn from the normal population.

Example 6.5
A soft drink dispensing machine is said to be out of control if the variance of the
contents exceeds 1.15 litres. If a random sample of 25 drinks from this machine has a
variance of 2.03 litres, does this indicate at 5% level of significance, that the machine
is out of control? (Assume that the contents are approximately normally distributed).

Solution
σ2 = 1.15, S2 = 2.03, n = 25, H0 : σ2 = 1.15 (not out of control)
H1: σ2 > 1.15 (the machine is out of control)

29
 (𝑛−1)𝑆 2 24 ×2.03
Test statistic 𝑋2 = = = 42.365
𝜎02 1.15
2 (
The critical value is 𝑋0.05 24) = 35.415. Since X2> 𝑋𝛼2 , we reject H0 , and conclude
that the machine is out of control.

(c) Tests for equality of two Variances

Finally, we shall consider the problem of testing for the equality of the variances 𝜎𝑥2
and 𝜎𝑌2 of two populations X and Y, assumed to be normal, on the basis of samples of
sizes n and m draw respectively from X and Y. The null hypothesis, as usual, will be

H0: 𝜎𝑥2 = 𝜎𝑌2 , versus H1: 𝜎𝑥2 <𝜎𝑌2 or 𝜎𝑥2 > 𝜎𝑌2 or 𝜎𝑥2 ≠ 𝜎𝑌2 .
The test statistic for this test is 𝐹 = 𝑠𝑥2 /𝑆𝑦2 , where 𝑆𝑥2 and 𝑆𝑦2 are variances computed
from the two samples. F so defined is a value of the F – distribution with n-1 and m-1
degrees of freedom. Decision criteria for α level of significance will be
(i) To test H0 against H1: 𝜎𝑥2 < 𝜎𝑦2 , reject H0 if F < F1-α(n-1, m-1)
(ii) To test H0: against H1: 𝜎𝑥2 > 𝜎𝑦2 , reject H0 if F > Fα(n-1, m-1)
(iii) To test H0 against H1: 𝜎𝑥2 ≠ 𝜎𝑦2 , reject H0 ,
if 𝐹 < 𝐹1−𝛼/2 (𝑛 − 1, 𝑚 − 1) 𝑜𝑟 𝐹 > 𝐹𝛼/2 (𝑛 − 1, 𝑚 − 1).

Example 6.6
A large automobile manufacturing company is trying to decide whether to purchase
brand X or brand Y tyres for its new models. To help arrive at a decision, an
experiment is conducted using 12 of brand X and 15 of brand Y. The tyres are run
until they wear out. The results are 𝑋̅ x = 37,900km, Sx = 5100km, 𝑌̅ =
39,800km , and Sy = 5900km. At 5% level of significance, and assuming that the
population are approximately normally distributed, test the hypothesis that
(i) there is no difference in the mean life of the tyres.
(ii) both brands of tyres have same variance.

30
Solution
(i) 𝛼 = 0.05, 𝑋̅ = 37,900, Sx = 5100, n = 12, m = 15, 𝑌̅ = 39,800, and Sy =
5900km. H0: µx = µy H1: µx ≠ µy
Observe that population variances are not known for both populations and H 0
assumes equal variances. Hence test statistic will be, as in 6.4 (c)
| 37900−39800 | 11×5100+14×5950
𝑇= , 𝑆=√ = 74.485
1 1 25
𝑆√ +
12 13

1900
𝑇 = = 65.863
74.485 × 0.387

But t0.025 (25) = 2.060, hence we reject H0.

(ii) H0:𝜎𝑥2 = 𝜎𝑥2 H1: 𝜎𝑥2 ≠ 𝜎𝑥2
𝑆𝑥2 51002
Test statistic is F = = = 0.7472
𝑆𝑦2 59002

𝐹0.025 (11, 14) = 3. 15and𝐹0.975 (11, 14) = 0.3003.

Since 0.3003 < F < 3.15, we do not reject H0 .

6.6 Tests Concerning Proportions

Tests of hypotheses concerning proportions may be required in such areas as:
(i) manufacturing, where firms are concerned about the proportion of defectives
in a given package.
(ii) games of chance, where the player depends upon a knowledge of the
proportion of outcomes that he considers favourable.
(iii) politics, where contestants are interested in knowing the fraction of the voters
that are favourably disposed to him in a given election.
We shall assume the problem to be that of testing the hypothesis that the proportion
of successes in a binomial experiment equals some specified value. That is to test H 0:
θ = θ0 against some appropriate alternatives. A summary of the tests are given below
for a CR of size α .

31
Case I: To test H0: θ = θ0 versus H1: θ < θ0
The critical region of size 𝛼 is given by 𝑋 < 𝐾𝛼 , where 𝐾𝛼 is the largest integer for
which 𝑃( X<𝐾𝛼 / θ = 𝜃0 ) = 𝛴𝑏(𝑋, 𝑛, 𝜃0 ) ≤ ∝
Case II: For H0: θ = 𝜃0 versus H1: θ > θ0,
the critical region of size ∝ is given by X >𝐾𝛼′ , where 𝐾𝛼′ is the smallest integer for
which 𝑃( X >𝐾𝛼′ / θ = θo ) = 𝛴𝑏(𝑋, 𝑛, 𝜃0 ) ≤∝
Case III: For H0: θ = θ0versus H1: θ ≠ θ0, the critical region of size ∝ is given by
′
X≤ 𝐾𝛼/2 and X > 𝐾𝛼/2 , where K and K’ are as defined in case I and case II
above. In all three, X is the number of successes. The decision is that H 0 should be
rejected whenever X falls in the critical region.

Example 6.7
An official of the Delta State University students Union claims that at least 60% of the
student population prefer campus hostel accommodation to off campus. What
conclusion would you draw, if only 11 in a sample of 20 students preferred campus
hostel? (use α = 0.05 level of significance).
Solution
H0: θ = 0.6 H1: θ < 0.6, ∝ = 0.05, n = 20, X = 11.
Critical Region is X< Kα , From table, K = 7 (binomial prob)
∴ CR is X< 7. But X = 11, hence we do not reject H0.
In the case when n, the sample size is large, the normal approximation with
parameters µ = nθ0 and σ2 = nθ0(1- θ0) is used and it provides an accurate test
provided θ0 is not too close to zero or 1. The normal approximation gives
𝑋 − 𝑛𝜃
𝑍=
𝑛𝜃0 (1 − 𝜃)
Which is a value of the standard normal variable Z .
For H0: θ = θ0 versus H1: θ < θ0, reject H0 if Z < - Zα

32
For H0: θ = θ0 versus H1: θ > θ0, reject H0 if Z > Zα
and for H0: θ = θ0 versus H1: θ ≠ θ0 , reject H0 if |Z| > Zα/2

Example 6.8
A union official of Delta State University Students Union claims that at least 60% of
the students prefers campus hostel accommodation to off campus. If in a sample of
200 students, 110 of them preferred campus hostel, test at 5% level of significance if
this claim is exaggerated.

Solution
𝜃0 = 0.6, 𝑛 = 200 (𝑙𝑎𝑟𝑔𝑒 ), 𝑋 = 110, ∝= 0.05,
H0: θ = 0.6 H1: θ< 0.6
Since n is large the normal approximation gives
110 − 200 × 0.6
𝑍= = −1.443
200 × 0.6 × 0.4
Table value is 𝑍0.05 = 1.645, −𝑍0.05 = -1.645
Since 𝑍 > −𝑍0.05 we do not reject the claim, and we conclude that it is not an
exaggeration.

6.7 Tests for Difference Between two Proportions

Situations may arise in which we wish to test the hypothesis that two proportions are
equal. Basic to this are two binomial populations x and y, with proportion parameters
θx and θy respectively. The statistic in which we base our decision criterion is the
random variable θx – θy . Independent samples of sizes n and m are selected from X
and Y, and the proportion of successes θx and θy are computed for both. In
general we wish to test
H0: θx = θy = θ against any suitable alternative.

33
 𝜃𝑥 – θy
When n is large, 𝑍=
θ (1−θy )
√θx (1−θx )+ y
n m

When H0 is true
𝜃𝑥 − 𝜃𝑦
𝑍=
√{𝜃(1 − 𝜃)(1⁄𝑛 + 1⁄𝑚)}

This is a value of the standard normal variable Z.

𝑋+𝑌
To compute θ we use the pooled estimate 𝜃=
𝑛+𝑚

Where X and Y are the number of successes in each of the two samples.
Therefore in testing H0: θx= θy, the Z value becomes
𝜃𝑥 −𝜃𝑦
𝑍=
√{𝜃(1−𝜃)(1⁄𝑛+1⁄𝑚)}

and decision criteria are as follows:

(a) for H1:θx <θy, reject H0 if 𝑍 < −𝑍 ∝
(b) for H1: θx>θy, reject H0 if 𝑍 > 𝑍 ∝
(c) for H1: θx ≠ θy, reject H0 if |𝑍| < 𝑍𝛼/2

Example 6.9
An opinion poll was conducted among secondary school students in a certain state to
determine whether to continue with the November / December GCE or not. If 120 of
200 female students prefer May / June SSCE, and 240 of 500 male students prefer
May/June SSCE to November / December GCE, would you agree that the proportion
of female students who favour the scrapping of Nov./Dec. GCE is higher than the
proportion of male students who favour same? (Use α = 0.025 level of significance).
Solution
Let X and Y represent the populations of female and male students respectively.
θx = 0.6, θx = 0.48, n = 200, m = 500, ∝ = 0.025
360
H0: θx = θy, H1: θx >θy , and θ = = 0.514
700

34
 0.6−0.48
𝑍= 1 1
= 2.87, and 𝑍0.025 = 1.96
√{(0.514)(0.486)( + )}
200 500

Since Z > Z0.025. we reject H0 and conclude that the proportion of female students
who prefer May/June SSCE is higher than the proportion of male students who prefer
same.

CHAPTER THREE
GOODNESS – OF - FIT TESTS
7.1 Introduction
Goodness of fit tests are tests that determine if a population has a specified theoretical
distribution. It measures how good a fit we have between the frequency of occurrence
of observations in an observed sample and the expected frequencies obtained from
the hypothesized distribution.
Under the general goodness of fit test, the test statistics is

(𝑂−𝐸)2
𝑋2 = 𝛴
𝐸

Where O stands for the observed frequency, while E represents the expected
frequency under H0. By this, the expected frequency is thus computed through a
theoretical distribution based on H0. If the observed frequencies differ considerably
from the expected frequencies, X2 value will be large, and the fit poor. A good fit (small
X2 value) leads to the rejection of H0. Therefore, the critical region of H0 will fall in the
right tail of the chi-square distribution. Thus for α level of significance reject H0 if X2
>𝑋𝛼2 . Underlying the above, each expected frequency must be at least 5. Such
frequencies that are less than 5 should be combined with adjacent cells, resulting in
reduction of number of degrees of freedom. The number of degrees of freedom
associated with the chi-square distribution on goodness of fit is equal to the number
of cells minus the number of quantities (or parameters) obtained from the observed
data, which are used in the calculation of the expected frequency.

35
Example 7.1
A die was thrown 120 times and the following frequency distribution was obtained.
We wish to test at α = 0.05, whether or not the die is biased.
Face 1 2 3 4 5 6
Frequency 15 12 20 18 30 25

Solution
H0: The die is not biased. From H0 the theoretical distribution implies equal frequency
should be expected. Thus we have:
Face 1 2 3 4 5 6
Obs. Freq. 15 12 20 18 30 25
Exp. Freq. 20 20 20 20 20 20
𝑂−𝐸 -5 -8 0 -2 10 5
(𝑂 − 𝐸 ) 2 25 64 0 4 100 25
(O-E)2/E 1.25 3.2 0 0.2 5 1.25

2
∑ (𝑂 − 𝐸 ) 2
𝑋 = = 1.25 + 3.2 + 0 + 0.02 + 5 + 1.25 = 10.9
𝐸
There are 6 cells and 1 restriction (total frequency)
... degree of freedom = 6 – 1 = 5 , 𝑋𝛼2 = 11.070
2
Since X2 <𝑋0.05 (5) , we do not reject H0, and we conclude that the die is not biased.

Example 7.2
No. of faults 0 1 2 3 4 5 6 7
No. of pieces 28 25 12 8 6 2 1 0
The number of minor faults in a steel plate produced by a machine were observed as
above. Under normal conditions, the expected distribution of faults based on two
restrictions are as follows:
No. of faults 0 1 2 3 4 5 6 7
No of pieces 26 22 15 8 5 3 2 1
Using an appropriate test, say whether the two distributions are same at 5% level of
significance.

36
Solution:
Faults 0 1 2 3 4 5 or more
E 26 22 15 8 5 6
O 28 25 12 8 6 3
(𝑂 − 𝐸 ) 2 4 9 9 0 1 9
(O-E)2 / E 0.1538 0.4091 0.6 0 0.2 1.5

X2 = 0.1538 + 0.4091 + 0.6 + 0+0.2 +1.5 = 2.8629

2
No of degree of freedom 7 - 2 = 5, 𝑋0.05 (5) = 11.070.
2
Since X2 <𝑋0.05 (5), this indicates that the test result is not significant. We conclude
therefore that both distributions are essentially the same.

7.2 Goodness of Fit Test for Binomial Distribution

The objective here is to see whether the frequencies of an observed sample are
in such a way that we might consider the distribution as binomial. The procedure is
to use a sample statistic based on the null binomial distribution to generate a
theoretical binomial distribution. The two setsof frequencies are then compared,
using a chi – square test. To generate the theoretical distribution, we need know
i. the number of trials for the observed distribution
ii. the probability of success, p
However, when n is known and p unknown, we use the approximation p = 𝑋̅ /n, where
𝑋̅ is the sample mean. In this case, there will be two restrictions, on the mean, and on
the total frequency. But when n and p are known, we need only the sum of observed
frequencies in order to derive the expected frequencies. Hence there will be only one
restriction.

Example 7.3

37
Test whether the sampling distribution given below can be considered as binomial at
5% level of significance.
Score 1 2 3 4 5 6
Frequency 12 15 10 14 5 4

Solution:
2.95
∑ƒ = 60, ∑ 𝑓𝑋 = 177, 𝑋̅ = 2.95, n = 6, 𝑝 = = 0.4917
6

Ho: the sampling distribution is binomial.

H1: the sampling distribution is binomial.
𝑛
Under Ho, we use the binomial formula 𝑃(𝑋 = 𝑟) = ( ) 𝑝𝑟 (1 − 𝑝)𝑛−𝑟 to generate
𝑟
a theoretical distribution or expected frequencies.
Score X 1 2 3 4 5 6
Obs Frequency 12 15 10 14 5 4
Exp Frequency 6.01 14.53 18.73 13.59 5.26 0.85
The expected frequency for the last cell is less than 5, therefore we make an
adjustment by combining the last two cells to get the new table below:

Score X 1 2 3 4 5 or 6
Obs Frequency 12 15 10 14 9
Exp Frequency 6.01 14.53 18.73 13.59 6.11
( O – E )2 35.88 0.2209 76.2129 0.1681 8.3521
(O – E )2 / E 5.970 0.0152 4.069 0.0124 1.367

∑(𝑂−𝐸)2
Hence, 𝑋 2 = = 11.434.
𝐸

2
Degree of freedom is 5 – 2 = 3 and Critical value is 𝑋0.05 (3) = 7.815
2
Comparing the test statistic X2 and the critical value 𝑋0.05 (5), we conclude that the
2 ( )
test is significant since X2>𝑋0.05 5 . Hence, we reject the hypothesis that the given
sampling distribution is binomial.

38
NOTE:
The expected frequencies were obtained as follows:
6
Ef(1) = N × P(X = 1) = 60 x ( ) 𝑝1 (1 − 𝑝)5 = 6.01,
1
6
Ef(2) = N × P(X = 2) = 60 x ( ) 𝑝2 (1 − 𝑝)4 = 14.53
2
6
Ef(3) = N × P(X = 3) = 60 x ( ) 𝑝3 (1 − 𝑝)3 = 18.73
3
6
Ef(4) = N × P(X = 4) = 60 x ( ) 𝑝4 (1 − 𝑝)2 = 13.59
4
6
Ef(5) = N × P(X = 5) = 60 x ( ) 𝑝5 (1 − 𝑝)1 = 5.26
5
6
Ef(6) = N × P(X = 6) = 60 x ( ) 𝑝6 (1 − 𝑝)0 = 0.85
6

Example 7.4
Four coins were tossed 200 times with the following results.
No. of Heads 0 1 2 3 4
No. of Times 9 42 73 61 15

Using X2 goodness of fit test at 5% level of significance, decide whether the coins were
biased.

Solution:
H0: The coins are unbiased . H0: The coins are not unbiased

Hence under H0 p(H) = ½. This means that there is only one restriction, and it is on
the total frequency.
Expected frequency for:
4
X = 0 is 200 ( ) (1/2)0 (1/2)4 = 12.5
0
4
X = 1 is 200 ( ) (1/2)1 (1/2)3 = 50
1

39
 4
X = 2 is 200 ( ) (1/2)2 (1/2)2 = 75
2
4
X = 3 is 200 ( ) (1/2)3 (1/2)1 = 50
3
4
X = 4 is 200 ( ) (1/2)4 (1/2)0 = 12.5
4
Thus we have
No. of heads 0 1 2 3 4
Observe freq 9 42 73 61 15
Expect. Freq 12.5 50 75 50 12.5
(O-E)2/E 0.98 1.28 0.053 2.42 0.5

2 ( )
X2 = 5.233, df = 5 – 1 = 4, 𝑋0.05 4 = 9.49
We conclude that the test is not significant at 5% and that the binomial distribution
with p = ½, gives a good fit to the given sampling distribution. Hence, the coins are
unbiased.

Exercise
Test whether the sampling distribution given below can be considered to be binomial
at either 5% or 10% level of significance.
X 0 1 2 3 4 5
Frequency 1 6 14 33 31 15

7.3 Goodness of Fit Test for a Poison Distribution

The theoretical distribution here is generated using the poison formula
𝑒 𝜇 𝜇𝑟
𝑃 (𝑋 = 𝑟 ) = , 𝑟 = 0, 1, 2, …
𝑟!

The parameter 𝜇 and total frequency are required for the generation of the theoretical
distribution. The parameter µ is usually estimated by the mean of the given
distribution. Hence, there are usually two restrictions on the choice of the expected
frequency values. Thus the degree of freedom for this test is n-2.

40
Example 7.5
Test whether a good fit is given by a poison distribution to the following frequency
distribution at 5% level of significance.
X 0 1 2 3 4 5
F 19 26 27 13 11 4

Solution:
H0: the distribution is Poisson. H0: the distribution is not Poisson.
𝑋̅ = 183/100 = 1.83, hence, µ = 1.83
Using the formula above, we compute the expected frequencies as follows:

Exp freq(X = r ) = 100 x P(X = r) = 16.04

Hence, we have 100 x P(X = 0) = 16.04
Similarly, 100 x P(X = 1) = 29.36, 100 x P(X = 2) = 26.86
100 x P(X = 3) = 16.38, 100 x P(X = 4) = 7.50
And 100 x P(X = 5) = 2.74.

Since the last expected frequency is less than 5, we combine the last two cells to have
the table below.
X 0 1 2 3 4 or 5
𝑂 19 26 27 13 15
E 16.04 29.36 26.86 16.38 10.24
(O-E)2/E 0.546 0.385 0.0007 0.697 2.213

2
X2 = 3.496, degree of freedom is 3, 𝑋0.05 (3) = 7.81
2
We do not reject H0 since X2<𝑋0.05 (3). We thus conclude that the Poisson distribution
gives a good fit to the sampling distribution above.

7.4 Goodness of Fit Test for a Normal Distribution

41
In testing for normality of a given sampling distribution, two sample parameters 𝑋̅
and S, along with the total frequency will be needed. This implies that there will be
three restirctions on the choice of the expected values. The procedure is as follows:
1
(a) Compute 𝑋̅ and S from the given sample [ 𝑆 = √ ∑(𝑋 − 𝑋̅ )2 ]
𝑛

(b) Use 𝑋̅ and S as estimates µ and σ, and the total to set up a theoretical normal
distribution.
(c) Compute the observed and expected frequencies as usual, using the X2 – test
statistic with three restrictions. Whenever µ and σ are known for the
theoretical distribution, then the estimates in (a) and (b) above will not be
needed.

Example 7.6
To test whether a good fit is provided by the normal distribution to the following
frequency distribution.

Class 10-14 15-19 20-24 25-29 30-34 35-39

Frequency 3 7 15 20 9 6

Solution:
The first thing is to find 𝑋̅ and S as estimates of µ and σ.
X f f.X di fdi2
12 3 36 -13.58 553.2492
17 7 119 -8.58 515.3148
22 15 330 -3.58 192.2460
27 20 540 1.42 40.3280
32 9 288 6.42 370.9476
37 6 222 11.42 782.4984
60 1535 2454.5840

1535 2454.5840
𝑋̅ = = 25.58, 𝑆 = = 6.45 .
60 59

42
Next we standardize the upper bounds as follows:
14.5−25.58 19.5 − 25.58
𝑧1 = = −1.718,𝑧2 = = −0.943 ,
6.45 6.45
24.5−25.58 29.5− 25.58
𝑧3 = = −0.167, 𝑧4 = = 0.608,
6.45 6.45
34.5 − 25.58
𝑧5 = = 1.383, 𝑧5 = ∞.
6.45

Class Upper bd Stdub ΦZ) P Exp. Freq

10 – 14 14.5 -1.718 0.0427 0.0427 2.56
15 – 19 19.5 - 0.943 0.1736 0.1309 7.85
20 – 24 24.5 - 0.167 0.4325 0.2589 15.53
25 – 29 29.5 0.608 0.7291 0.2966 17.80
30 – 34 34.5 1.383 0.9162 0.1871 11.23
35 – 39 39.5 ∞ 1.0000 0.0838 5.03

We combine the first and second cells to bring the expected frequency level to at least
5. Thus, we now have
Class 10-19 20 – 24 25 - 29 30 - 34 35- 39
Obs freq 10 15 20 9 6
Exp freq. 10.41 15.53 17.80 11.23 5.03
(O-E)2 / E 0.016 0.018 0.272 0.443 0.187
X2 = 0.936, degree of freedom = 5 - 3 = 2. At 5% level of significance, X20.05(2) =
5.991. Hence, the test is not significant, so we conclude that the given sampling
distribution appears normal.

7.5 Contingency Tables and Tests for Independence

Here populations are classified according to two factors of interest, each factor
consisting of multiple levels. The objective of the test is to find out whether or not the
two attributes (or factors) are independent of each other.
The observed information comes in the form of a contingency table, whose general
form is specified below:

43
 Factor A
A1 A2 An Totals
B1 O11 O12 …. O1n R1
B2 O21 O22 …. O2n R2
Factor B
…. …. …. …. …. …
Bm Om1 Om1 …. Omn Rm
Totals C1 C2 …. Cn T

Where A1, A2, … ,An and B1, B2, …, Bm are the respective levels of factors A and B.
Deciding on the level of a factor depends on what the researcher considers important.
Oij denotes the number of observations that possess the attributes B i and Aj
simultaneously, Ri and Cj are the row and column totals respectively, while T is the
grand total of the observations.
To each entry Oijof the contingency table we can compute the expected frequency
thus:
𝑅𝑖 𝑥 𝑐𝑗
𝐸𝑖𝑗 =
𝑇
∑ ∑(𝑂−𝐸)2
Test statistic remains X2 = , the decision rule is that at α level of significance,
𝐸
reject H0 if X2 >𝑋𝛼2 {(m - 1)(n - 1)}.

Example 7.7
A group of students were tested first in Language, and then in mathematics. The
results were graded into three categories A, B, C, for each test. A summary of the
distribution of performance is given below:
Language
A B C TOTALS
A 55 72 12
Maths

B 48 162 38
C 14 42 85

Using a chi-square test at 5% level of significance, decide whether students

performance in language is related to students performance in mathematics.
Solution:
H0: Student’s performance in language is independent of student’s performance in
mathematics.

44
First, we compute the row totals, column totals and grand total. Then, under H0, we
draw up a table of expected frequency using
𝑇𝑖. × 𝑇.𝑗
𝐸𝑖𝑗 = to get the following table of expected frequency:
𝑇..
Language
A B C TOTALS
A 30.8 72.7 35.5 139
Maths

B 55.0 129.6 63.4 248

C 31.2 73.7 36.1 141
Totals 117 276 135 528

2
(55 − 30.8)2 (72 − 72.7)2 (12 − 35.5)2 (48 − 55)2 (162 − 129.6)2
𝑋 = + + + +
30.8 72.7 35.5 55 129.6
(38 − 63.4)2 (14 − 31.2)2 (42 − 73.7)2 (85 − 36.1)2
+ + + +
63.4 31.2 73.7 36.1
= 19.01 + 0.01 + 15.56 + 0.89 + 8.1 + 10.18 + 9.48 + 13.63 + 66.24
= 143.10 . Degree of freedom if 2 x 2 = 4
2
But 𝑋0.05 (4) = 9.49
The result is highly significant, and hence we reject H0 .

CHAPTER FOUR
CORRELATION ANALYSIS

4.7 The Linear Correlation Coefficient

The coefficient of correlation is another measure of closeness (relationship) between
two sets of values. It is a study of the degree of interdependence between two
variables. Correlation can be positive, negative or zero.
Correlation coefficient r is the parameter for measuring correlation and its value lies
between -1 and 1, i.e. 1≤ r ≤ 1.

45
Two variables are said to be positively correlated if they tend to increase or decrease
together in the same direction, and for this type 0 < r ≤ 1. Two variables X and Y
are said to be negatively correlated if X and Y change in opposite direction; that is,
when X increases, Y decreases. Here the value of r will satisfy -1 < r <0 . Two variables
are said to be un-correlated when they tend to change with no definite pattern
regarding each other. This is also called zero correlation. Here r = 0. We present
here a summary of the Properties and the interpretation of the value of a Correlation
Coefficient, r.
The correlation coefficient usually denoted by ‘ 𝑟 ’ satisfies the following conditions:

i. −1 ≤ 𝑟 ≤ 1 i.e. the value of r must be between −1 and +1; it can never be

greater that 1 or less than −1.
ii. If 𝑟 = 0, then, there is no correlation between 𝑋 and 𝑌.
iii. If 𝑟 = 1, then, there is perfect positive correlation between 𝑋 and Y.
iv. If 𝑟 = −1, there is perfect negative correlation between 𝑋 and 𝑌.
v. If 0.5 ≤ 𝑟 < 1, then, there is strong positive correlation
between𝑋and 𝑌.
vi. If 0 < 𝑟 < 0.5, then, there is weak positive correlation
between X and 𝑌.
vii. If −0.5 < 𝑟 < 0, then, there is weak negative correlation
between𝑋 and 𝑌.
viii. If −1 < 𝑟 ≤ 0.5, then, there is strong negative correlation
between𝑋 and 𝑌.

The diagrams below display these types of correlation

46
We will consider two common methods for calculating the coefficient of correlation.
These are:
(a) The Pearson’s Product Moment Correlation Coefficient and
(b) The Spearman’s Rank Correlation Coefficient.

4.8 Pearson’s Product – Moment Method

This method is executed by using the raw scores (actual observations) and the mean
deviations. The steps in the computation are as follows.
(i) Create a table for the observations with the following columns
X, Y, X2, Y2, XY
(ii) Calculate the mean of each set of variable, and obtain the deviations.

To derive r,
Let the variables be X and Y. Find their respective means 𝑋̅ 𝑎𝑛𝑑 𝑌̅. Obtain the mean
deviation of each Xi and Yi
Let dXi = Xi - 𝑋̅ and dYi = Yi - 𝑌̅

47
𝑇ℎ𝑒𝑛 ∑(𝑋𝑖 − 𝑌̅)(𝑌𝑖 − 𝑌̅) = ∑𝑑𝑋𝑖 𝑑𝑌𝑖 … …

∑𝑑𝑋𝑖 𝑑𝑌𝑖
∴ = 𝑆𝑥𝑦
𝑛
This is nthe covariance between X and Y. Now, divide 𝑆𝑥𝑦 by the standard deviations
of X and Y, the result is the sample correlation coefficient r, given by
∑𝑑𝑋𝑖 𝑑𝑌𝑖 𝑆𝑥𝑦
𝑟 = =
𝑛𝑆𝑥 𝑆𝑦 𝑆𝑥 𝑆𝑦
where

∑(𝑋𝑖 − 𝑋̅ )2 ∑(𝑌𝑖 − 𝑌̅)2

𝑆𝑥 = √ 𝑎𝑛𝑑 𝑆𝑦 = √
𝑛 𝑛

Now substitute for values of 𝑆𝑥 , 𝑆𝑦 , 𝑎𝑛𝑑 𝑆𝑥𝑦 to get

𝑛∑𝑋𝑌 − (∑𝑋)(∑𝑌)
𝑟=
√[𝑛∑𝑋 2 − (∑𝑋)2 ][𝑛∑𝑌 2 − (∑𝑌)2 ]

∑ 𝑋𝑌 − 𝑛 𝑋̅ 𝑌̅
𝑟=
√[∑𝑋 2 − 𝑛 𝑋̅ 2 ][∑𝑌 2 − 𝑛 𝑌̅ 2 ]
∑(𝑋 − 𝑋̅ )(𝑌 − 𝑌̅)
𝑜𝑟 𝑟=
√∑ (𝑋 − 𝑋̅ )2 ∑ (𝑌 − 𝑌̅)2

Example 4.7
Calculate the Pearson’s Product Moment correlation coefficient for the data below:
X 2 3 6 4 7
Y 9 6 8 2 5

Solution:
X Y XY X2 Y2
2 9 18 4 81
3 6 18 9 36

48
 6 8 48 36 64
4 2 8 16 4
7 5 35 49 25
22 30 127 114 210

5 × 127 − 22 × 30 −25
𝑟 = = = −0.2201
√[5 × 114 − 222 ][5 × 210 − 302 ] √86 × 150
Or
127 − 5 × 4.4 × 6 −5
𝑟= = = −0.2201
√[114 − 5 × 4.42 ][210 − 5 × 62 ] √17.2 × 30

Example 4.8
Use the Pearson’s product moment method to obtain the linear correlation coefficient
between price (X) and quantity (Y).
Time (n) 1 2 3 4 5 6 7 8 9 10
Quantity(Y) 10 20 50 40 50 60 80 90 90 120
Price(X) 2 4 6 8 10 12 14 16 18 20
Solution
N Y X XY X2 Y2
1 10 2 20 4 100
2 20 4 80 16 400
3 50 6 300 36 2500
4 40 8 320 64 1600
5 50 10 500 100 2500
6 60 12 720 144 3600
7 80 14 1120 196 6400
8 90 16 1440 256 8100
9 90 18 1620 324 8100
10 120 20 2400 400 14400
Total 610 110 8520 1540 47700
Now
𝑛∑𝑋𝑌 – (∑𝑋)( ∑𝑌)
𝑟=
√{𝑛∑𝑋 2 − (∑𝑋)2 } {𝑛∑𝑦 2 − (∑𝑋)2 }
From the table above,
49
 10(8520)−(110)(610)
𝑟=
√{(10)(1540)−(110)2 }{(10)(47700)−(610)2

18100
= = 0.97282
√3300 × 104900
Hence , r = 0.973, which shows a very strong positive correlation between
quantity and price.

An alternative method is the mean deviation method.

From the table,
∑𝑋𝑖 110 ∑𝑌𝑖 610
𝑋̅ = = = 11 𝑎𝑛𝑑 𝑌̅ = = = 61
𝑛 10 𝑛 10
N Xi Yi Xi−𝑿 ̅ Yi−𝑿 ̅ (Xi−𝑿 ̅ )2 (Yi−𝒀̅ )2 ̅ ) (Yi−𝒀
(Xi−𝑿 ̅)
1 2 10 -9 -51 81 2601 459
2 4 20 -7 -41 49 1681 287
3 6 50 -5 -11 25 121 55
4 8 40 -3 -21 9 441 63
5 10 50 -1 -11 1 121 11
6 12 60 1 -1 1 1 -1
7 14 80 3 19 9 361 57
8 16 90 5 29 25 841 145
9 18 90 7 29 49 841 203
10 20 120 9 59 81 3481 531
110 610 0 0 330 10490 1810

Using
∑(𝑋 − 𝑋̅ )(𝑌 − 𝑌̅) 1810
𝑟= = = 0.97282
√∑ (𝑋 − 𝑋̅ )2 ∑ (𝑌 − 𝑌̅)2 √330 × 10490

4.9 Spearman’s Rank Correlation Coefficient

The Spearman’s Rank Correlation Coefficient is one of the easiest to compute. It takes
the following sets. Given two sets of values for variables X and Y respectively, which

50
are paired observations, rank each variable independent of the other and find the
difference in ranks of each pair, then square these difference in rank and use the
formula
6∑𝑑𝑖2
𝑟 =1−
𝑛(𝑛2 − 1)
Where n is the number of pairs of observations, di is the difference in ranks of the ith
pair of observation, and r is the coefficient of correlation.
In using the rank correlation coefficient, observations are ranked and the ranks used
in the computation instead of the actual observations. The observations are ranked
in a specific sequence e.g. in ascending or descending order of attributes.

Derivation of the formula for Spearman’s Rank Correlation Coefficient

Let R1, R2, …,Rn be the respective ranks of the observations X1, X2… ,Xnand T1, T2, …,Tn
the respective ranks of the observations Y1, Y2 ,….,Yn.
𝑛(𝑛+1)
We recall that 1 + 2 + ⋯ + 𝑛 = … … … … … … …. (𝑖)
2
𝑛(𝑛+1)(2𝑛+1)
And that 12 + 22 + ⋯ + 𝑛2 = … … … …. (𝑖𝑖)
6

From the above, it is clear that

𝑛 𝑛
𝑛 (𝑛 + 1) 𝑛+1
∑ 𝑅𝑖 = ∑ 𝑇𝑖 = 𝑎𝑛𝑑 𝑅̅ = 𝑇̅ =
2 2
𝑖=1 𝑖=1

and
𝑛 𝑛
𝑛(𝑛 + 1)(2𝑛 + 1)
∑ 𝑅𝑖2 = ∑ 𝑇𝑖2 =
6
𝑖=1 𝑖=1

Now
∑(𝑅𝑖 − 𝑅̅ )2 = ∑ 𝑅𝑖2 − 𝑛𝑅̅2
𝑛(𝑛+1)(2𝑛+1) 𝑛(𝑛+1)2 𝑛(𝑛2 −1)
= − =
6 4 12
𝑛(𝑛2 −1)
Similarly, ∑(𝑇𝑖 − 𝑇̅)2 = ∑ 𝑇𝑖2 − 𝑛𝑇 2 =
12

If the difference in rank for the ith pair of observations is denoted by di, then
51
 𝑑𝑖 = 𝑅𝑖 − 𝑇𝑖 = (𝑅𝑖 − 𝑅̅) − (𝑇𝑖 − 𝑇̅)
𝑑𝑖2 = (𝑅𝑖 − 𝑅̅)2 − 2 (𝑅𝑖 − 𝑅̅ )(𝑇𝑖 − 𝑇̅) + (𝑇𝑖 − 𝑇̅)2
∑𝑑𝑖2 = ∑(𝑅𝑖 − 𝑅̅ )2 − 2 ∑(𝑅𝑖 − 𝑅̅) (𝑇𝑖 − 𝑇̅) + ∑(𝑇𝑖 − 𝑇̅)2
𝑛(𝑛2 −1) 𝑛(𝑛2 −1)
∑𝑑𝑖2 = − 2∑(𝑅𝑖 − 𝑅̅)(𝑇𝑖 − 𝑇̅) +
12 12
2𝑛(𝑛2 −1)
∑𝑑𝑖2 = − 2∑(𝑅𝑖 − 𝑅̅)(𝑇𝑖 − 𝑇̅)
12

𝑛 ( 𝑛 2 − 1) 1 2
∑(𝑅𝑖 − 𝑅̅)(𝑇𝑖 − 𝑇̅) = − ∑𝑑𝑖
12 2
But
𝑆𝑋𝑌
𝑟 =
𝑆𝑋 𝑆𝑌

𝑛(𝑛2 −1) 1 𝑛(𝑛2 −1) 6∑𝑑𝑖2

𝑟={ − ∑𝑑𝑖2 } ÷ =1−
12 2 12 𝑛(𝑛2 −1)

6∑𝑑𝑖2
∴𝑟 =1−
𝑛(𝑛2 −1)

Steps in computation of Spearman’s r

(i) Rank the observations in a specified order, but independently for each set of
the observations. For a tie in rank, take the average position as rank for all scores
involved in the tie.
(ii) Obtain the difference in the ranks of pairs of observations.
(iii) Square the differences obtained in (ii) above and take their sum.
(iv) Substitute the values into the formula to get r

Example 4.9
Calculate the Spearman’s rank correlation coefficient for the data below.
X 2 3 6 4 7
Y 9 6 8 2 5

Solution
52
 X Y 𝑅𝑖 𝑇𝑖 𝑑𝑖 𝑑𝑖2
2 9 5 1 4 16
3 6 4 3 1 1
6 8 2 2 0 0
4 2 3 5 -2 4
7 5 1 4 -3 9
30

6∑𝑑 2
𝑛 = 5, ∑𝑑𝑖2 = 30, ℎ𝑒𝑛𝑐𝑒 𝑢𝑠𝑖𝑛𝑔 𝑟 = 1 −
𝑛 (𝑛 2 − 1)
6 × 30
𝑤𝑒 𝑔𝑒𝑡 𝑟 = 1− = −0.5
5 × 24
This implies a strong negative correlation between X and Y

Example 4.10
There are ten finalists in a competition for which there are two judges X and Y. The
Final scores by the judges are as follows:
Competitor A B C D E F G H I J
Judge X 31 20 55 30 60 38 37 24 27 41
Judge Y 50 30 28 20 36 52 26 38 47 47
Calculate the rank correlation coefficient between the scores awarded by the judges.

Solution:
Rank the observation x and y respectively to get the following table.
Competitor X Rank X Y Rank Y 𝑑𝑖 𝑑𝑖2
A 31 6 50 2 4 16
B 20 10 30 7 3 9
C 55 2 28 8 -6 36
D 30 7 20 10 -3 9
E 60 1 36 6 -5 25
F 38 4 52 1 3 9
G 37 5 26 9 -4 16
H 24 9 38 5 4 16
I 27 8 47 3.5 4.5 20.25
J 41 3 47 3.5 -0.5 0.25

53
 156.5

6∑𝑑2
Now, n = 10 and 𝑟 =1−
𝑛 (𝑛2 −1)
6 𝑥 156.5
𝑟 =1−
10 (102 −1)
972
𝑟 =1− = 0.05152
990
There is very low correlation between the scores of the judges.

Exercise:
The scores of ten students in a class cumulative test and their examination scores are
as in the table below. Calculate the coefficient of correlation using both the Pearson’s
Product moment and the Spearman’s Rank Correlation method. In each case
comment on your result.
Student A B C D E F G H I J
Cumulative x 12 14 8 8 7 6 4 5 8 2
Cumulative y 73 65 55 60 50 49 50 48 51 30

Example 4.11
𝑋 3 1 2 7 4 8
𝑌 4 5 2 9 6 7
Find the Spearman’s Rank Correlation Coefficient for the data above.

𝑛 = 6, ∑𝒅𝟐𝒊 = 8
6×8
𝑟 =1− = 0.7714
6(62 − 1)
∴ 𝑟 = 0.77which implies that there is strong positive correlation between 𝑋 and 𝑌.

Example 4.12
𝑋 2 6 4 8 5 10
𝑌 20 11 18 6 14 5
Find the Spearman’s Rank Correlation Coefficient using the data above

6×70
𝑟 =1− = −1, there is perfect negative correlation between 𝑋 and 𝑌.
6(62 −1)

Example 4.13

54
 𝑋 3 1 6 3 5 4 9 7 8 5
𝑌 9 10 2 8 6 5 3 8 4 8
Find the Spearman’s Rank Correlation Coefficient

6 × 283.0 6 × 283
𝑟 =1− = 1 − = −0.715
10(102 − 1) 990
There is strong negative correlation between 𝑋 and 𝑌.

Example 4.14
𝑋 2 3 4 5 6 8
𝑌 11 8 9 5 6 3
To Find the Spearman’s Rank Correlation Coefficient
6 × 66 396
𝑟 =1− = 1 − = −𝑜. 8857
6(62 − 1) 210
There is strong negative correlation between 𝑋 and 𝑌.

Example 4.15
𝑋 1 3 4 6 8 9 11 14
𝑌 1 2 4 4 5 7 8 9
Find the Spearman’s Rank Correlation Coefficient using the data above.

6 × 0.5
𝑛 = 6, ∑𝑑 2 = 0.5, 𝑟 = 1 − = 1 − 0.005952
8(82 − 1)
𝐻𝑒𝑛𝑐𝑒, 𝑟 = 0.994, there is strong positive correlation between 𝑋 and 𝑌.

Pearson’s Correlation Coefficient

This is simply the ratio of the covariance term to the product of the individual
standard deviations of the two variables.
The Formula is
𝑛 ∑ 𝑋𝑌 − (∑ 𝑋 )(∑ 𝑌)
𝑟=
√[𝑛 ∑ 𝑋 2 − (∑ 𝑋 )2 ][𝑛 ∑ 𝑌 2 − (∑ 𝑌)2 ]

Example 4.16
Compute the correlation coefficient for the data below using the Pearson’s formula.
𝑋 3 5 6 8 9
𝑌 2 3 6 5 4

55
Solution
𝑋 𝑌 𝑋𝑌 𝑋2 𝑌2
3 2 6 9 4
5 3 15 25 9
6 6 36 36 36
8 5 40 64 25
9 4 36 81 16
𝟑𝟏 𝟐𝟎 𝟏𝟑𝟑 𝟐𝟏𝟓 𝟗𝟎

5 × 133 − 31 × 20 45
𝑟= 2 2
= = 0.596
[5 × 215 − (31) ][5 × 90 − (20) ] √114 × 50
This implies a strong positive correlation.

Example 4.17
Compute the correlation coefficient for the data below using both Spearman’s
and Pearson’s methods.
𝑋 1 3 5 2 5
𝑌 4 8 7 6 10

Solution
𝑿 𝒀 𝑹𝑿 𝑹𝒀 𝒅𝟐𝒊 𝑿𝒀 𝑿𝟐 𝒀𝟐
1 4 5 5 0 4 1 16
3 8 3 2 1 24 9 64
5 7 1.5 3 2.25 35 25 49
2 6 4 4 0 12 4 36
5 10 1.5 1 0.25 50 25 100
𝟏𝟔 𝟑𝟓 𝟑. 𝟓0 𝟏𝟐𝟓 𝟔𝟒 𝟐𝟔𝟓

Spearman’s Method
6 ∑ 𝒅𝟐𝒊
𝑟 =1−
𝑛 (𝑛 2 − 1)
6 × 3.50
𝑟 =1− = 0.825
5(25 − 1)

Pearson’s Method

56
 𝑛 ∑ 𝑋𝑌 − (∑ 𝑋 )(∑ 𝑌)
𝑟=
√[𝑛 ∑ 𝑋 2 − (∑ 𝑋 )2 ][𝑛 ∑ 𝑌 2 − (∑ 𝑌)2 ]

5 × 125 − 16 × 35
𝑟=
√(5 × 64 − 162 )(5 × 265 − 352 )
65
𝑟= = 0.8125
√64 × 100

In both cases, there is strong positive correlation between 𝑋and 𝑌.

NOTE
Spearman’s correlation coefficient and Pearson’s correlation coefficient do not give
exactly the same results, except when r = +1 or -1. Spearman’s method is an
approximation and a quick guess. It belongs to the class of measures called non
parametric statistics. It is calculated from ranks instead of the actual observations.
Pearson’s method is more reliable because it is obtained from the actual observations.

CHAPTER FIVE
SIMPLE LINEAR REGRESSION

CHAPTER SIX
ONE – WAY ANALYSIS OF VARIANCE

57