1

CHAPTER ONE

SAMPLING AND SAMPLING DISTRIBUTIONS

A sampling distribution is a probability distribution for the possible values of a sample

statistic, such as a sample mean.
NOTE: The normal probability distribution is used to determine probabilities for the
normally distributed individual measurements, given the mean and the
standard deviation. Symbolically, the variable is the measurement X, with the
population mean µ and population standard deviation δ. In contrast to such
distributions of individual measurements, a sampling distribution is a
probability distribution for the possible values of a sample statistic.
Basic Definitions:
 Parameter: Characteristic or measure obtained from a population.
 Statistic: Characteristic or measure obtained from a sample.
 Sampling: The process or method of sample selection from the population.
 Sampling unit: the ultimate unit to be sampled or elements of the population to
be sampled.
Examples:
 If somebody studies Scio-economic status of the households, households are the
sampling unit.
 If one studies performance of freshman students in some college, the student is
the sampling unit.
 Sampling frame: is the list of all elements in a population under study.
Examples:
- List of households.
- List of students in the registrar office.
 The Need for Sampling
- Reduced cost
- Greater speed
- Greater accuracy
- Greater scope
- More detailed information can be obtained.
 Errors in sample survey:
There are two types of errors
a) Sampling error:
- It is the discrepancy between the population value and sample value.
- May arise due to inappropriate sampling techniques applied
b) Non sampling errors: are errors due to procedure bias such as:
 2

- Due to incorrect responses

- Measurement
- Errors at different stages in processing the data.
 Types of Selections
There are two types of selections, with and without replacement.

i. Sampling with replacement is the method of selection of a unit from a

population in which the unit is returned to the population before the next selection

is made after recording the required characteristics. Therefore, the population size at

each selection remain the same (constant), which implies that in this method of

selection a unit can have a chance of selection more than once. Thus, under this

method there are Nn possible samples of size n. Note also that the probability at each

selection of a unit is the same.

ii. Sampling without replacement a unit selected from a population is not

returned or replaced to the population for the next selection. Therefore, the

population size reduces by one for each selection from the immediate preceding one.

Hence, there are NCn (N combination of n) possible samples under this method. The

probability of selection, therefore, is not constant at each selection, 1/N for the first

unit, 1/(N-1) for the next, 1/(N-2) for the third and so on

 The theoretical mechanism of drawing a sample could be with or without replacement.

In practice, however, sampling is almost always done without replacement. But to help

us understand some essential concepts, we will also consider here sampling with

replacement in which every elementary unit has an equal chance of being selected for

each successive observation. It is also evident that sample outcomes are statistically

independent when sampling with replacement and statistically dependent when

sampling without replacement

 3

Types of sampling techniques

 There are two types of sampling.
1. Random Sampling or probability sampling
- It is a method of sampling in which all elements in the population have a pre-assigned non-
zero probability to be included in to the sample.
Some of Random Sampling or probability sampling are:
 Simple random sampling
 Stratified random sampling
 Cluster sampling
 Systematic sampling
i. Simple Random Sampling:
- It is a method of selecting items from a population such that every possible sample of
specific size has an equal chance of being selected. In this case, sampling may be with
or without replacement. Or
- All elements in the population have the same pre-assigned non-zero probability to be
included in to the sample.
- Simple random sampling can be done either using the lottery method or table of
random numbers.
- In the lottery method: first give a unique identification code to each unit of the population.
Then, write down the codes on identical pieces of papers, mix them up in bowl and select the
units whose codes appear on the randomly selected pieces of papers.
- In the table of random numbers method: to select sample of size n, first make a list of the
population to be sampled and give a distinct code number to each unit of the population
then, choose the direction of selection randomly and finally take n units.

ii. Stratified Random Sampling:

- The population will be divided in to non-overlapping but exhaustive groups called
strata.
- Simple random samples will be chosen from each stratum.
- Elements in the same strata should be more or less homogeneous while different in
different strata.
- It is applied if the population is heterogeneous.
- Some of the criteria for dividing a population into strata are: Sex (male, female); Age
(under 18, 18 to 28, 29 to 39,); Occupation (blue-collar, professional, and other).
Proportional Allocation of sample size in stratified sampling
• The items are selected from each stratum in the same proportion as they exist in the
population.
 4

• The allocation of sample sizes is termed as proportional if the sample fraction, i.e., if the
ratio of the sample size to the population size, remains the same in all the strata.
Mathematically the principle of proportional allocation gives: = = =…
By the property of ratio and proportions, each of these ratios is equal to the ratio of the sum of
numerators to the sum of denominators,
∑
i.e., = = = = =∑ = = c, (constant )
since the total sample size n , and the population size N are fixed.
Hence, = ( ), = ( ), = ( ), = ( ), (i = 1, 2, 3,..., k )
Example: A stratified sample of size n =80 is to be taken form a population of size N=2000,
which consists of four strata for which N1 = 500, N2=1200, N3=200 and N4=100. If we use
proportional allocation, how large a sample must be taken from each stratum?
• Solution: In proportional allocation, we know that the sample size for the ith stratum is
given by
• = ( ), (i = 1, 2, 3,..., k ), Then
• = ( ) = 500( ) = 20 must be taken from the 1st stratum.
• = ( ) = 1200( ) = 48,
• = ( )= 200( )=8
• = ( )= 100( )=4

iii. Cluster Sampling:

- The population is divided in to non-overlapping groups called clusters.
- A simple random sample of groups or cluster of elements is chosen and all the sampling
units in the selected clusters will be surveyed.
- Clusters are formed in a way that elements within a cluster are heterogeneous, i.e.
observations in each cluster should be more or less dissimilar.
- Cluster sampling is useful when it is difficult or costly to generate a simple random
sample. For example, to estimate the average annual household income in a large city
we use cluster sampling, because to use simple random sampling we need a complete
list of households in the city from which to sample. To use stratified random sampling,
we would again need the list of households. A less expensive way is to let each block
within the city represent a cluster. A sample of clusters could then be randomly
selected, and every household within these clusters could be interviewed to find the
average annual household income.

iv. Systematic Sampling:

- A complete list of all elements within the population (sampling frame) is required.
- The procedure starts in determining the first element to be included in the sample.
- Then the technique is to take the kth item from the sampling frame.
N
- Let, N  population size, n  sample size, k   sampling int erval.
n
 5

- Choose any number between 1 and k . Suppose it is j (1  j  k ) .

- The j
th
unit is selected at first and then ( j  k )th , ( j  2k )th ,....etc until the
required sample size is reached.

2. Non Random Sampling or non-probability sampling

- It is a sampling technique in which the choice of individuals for a sample depends on the
basis of convenience, personal choice or interest.
Some of non-Random Sampling or non-probability sampling are:
 Judgment sampling.
 Convenience sampling
 Quota Sampling.
1. Judgment Sampling
- In this case, the person taking the sample has direct or indirect control over which
items are selected for the sample.
2. Convenience Sampling
- In this method, the decision maker selects a sample from the population in a manner
that is relatively easy and convenient.
3. Quota Sampling
- In this method, the decision maker requires the sample to contain a certain number of
items with a given characteristic. Many political polls are, in part, quota sampling.

Note: let N  population size, n  sample size.

1.Suppose simple random sampling is used
 We have N n possible samples if sampling is with replacement.
N
 We have   possible samples if sampling is without replacement.
n
2. After this onwards, we consider that samples are drawn from a given population
using simple random sampling.
 6

Sampling Distribution of the Mean

The sampling distribution of the mean is the probability distributions of the means, X
of all simple random samples of a given sample size n that can be drawn from the
population.
NB: the sampling distribution of the mean is not the sample distribution, which is the
distribution of the measured values of X in one random sample. Rather, the sampling
distribution of the mean is the probability distribution for X , the sample mean.

For any given sample size n taken from a population with mean µ and standard
deviation δ, the value of the sample mean X would vary from sample to sample if
several random samples were obtained from the population. This variability serves as
the basis for sampling distribution.

The sampling distribution of the mean is described by two parameters: the expected
value ( X ) = X , or mean of the sampling distribution of the mean, and the standard
deviation of the mean  x , the standard error of the mean.

Properties of the Sampling Distribution of Means

1. The mean of the sampling distribution of the means is equal to the population
mean. µ =  X = X .
2. the standard deviation of the sampling distribution of the means (standard error)
is equal to the population standard deviation divided by the square root of the
sample size:  x = δ/√n. This hold true if and only of n<0.05N and N is very large.
 N n N n
If N is finite and n≥ 0.05N,  x  * . The expression is called
n N 1 N 1
finite population correction factor/finite population multiplier. In the calculation
of the standard error of the mean, if the population standard deviation δ is
unknown, the standard error of the mean  x , can be estimated by using the
sample standard error of the mean S X which is calculated as follows:
S N n
SX  S or S X  * .
n n N 1

3. The sampling distribution of means is approximately normal for sufficiently

large sample sizes (n≥ 30).

Example:
 7

A population consists of the following ages: 10, 20, 30, 40, and 50. A random sample of
three is to be selected from this population and mean computed. Develop the sampling
distribution of the mean.

Solution:
The number of simple random samples of size n that can be drawn without replacement
from a population of size N is NCn. With N= 5 and n = 3, 5C3 = 10 samples can be drawn
from the population as:

Sampled items Sample means ( X )

10, 20, 30 20.00
10, 20, 40, 23.33
10, 20, 50 26.67
10, 30, 40 26.67
10, 30, 50 30.00
10, 40, 50 33.33
20, 30, 40 30.00
20, 30, 50 33.33
20, 40, 50 36.67
30, 40, 50 40.00
300.00

A systematic organization of the above figures gives the following:

Sample mean ( X ) Frequency Prob. (relative freq.) of X
20.00 1 0.1
23.33 1 0.1
26.67 2 0.2
30.00 2 0.2
33.33 2 0.2
36.67 1 0.1
40.00 1 0.1
10.00 1.00

Columns 1 and 2 show frequency distribution of sample means.

Columns 1 and 3 show sampling distribution of the mean.


 X   x  30, Regardless of the sample size   X .
N n
 8

 X 
2
X 1000
 
 14.142 .
i

N 5
 N  n 14.142 53
X  *  *  5.774
n N 1 3 5 1
2

  X X 
i
 333.4
   5.774
N 10
Since averaging reduces variability  x < δ except the cases where δ = 0 and n = 1.

Central Limit Theorem and the Sampling Distribution of the Mean

The Central Limit Theorem (CLT) states that:
1. If the population is normally distributed, the distribution of sample means is
normal regardless of the sample size.
2. If the population from which samples are taken is not normal, the distribution of
sample means will be approximately normal if the sample size (n) is sufficiently
large (n ≥ 30). The larger the sample size is used, the closer the sampling
distribution is to the normal curve.

The relationship between the shape of the population distribution and the shape of
the sampling distribution of the mean is called the Central Limit Theorem.

The significance of the Central Limit Theorem is that it permits us to use sample
statistics to make inference about population parameters with out knowing anything
about the shape of the frequency distribution of that population other than what we can
get from the sample. It also permits us to use the normal distribution (curve for
analyzing distributions whose shape is unknown. It creates the potential for applying
the normal distribution to many problems when the sample is sufficiently large.

Example:
1. The distribution of annual earnings of all bank tellers with five years of experience is
skewed negatively. This distribution has a mean of Birr 15,000 and a standard deviation
of Birr 2000. If we draw a random sample of 30 tellers, what is the probability that their
earnings will average more than Birr 15,750 annually?
Solution:
Steps:
1. Calculate µ and  x
µ = Birr 15,000
 x = δ/√n= 2000/√30 = Birr 365.15
 9

2. Calculate Z for X
X X X 
ZX  
X X
15,750  15,000
Z15,750    2.05
365

3. Find the area covered by the interval

P ( X > 15,750) = P (Z > +2.05)
= 0.5 - P (0 to +2.05)
= 0.5 – 0.47892
= 0.02018
4. Interpret the results
There is a 2.02% chance that the average earning being more than Birr 15, 750
annually in a group of 30 tellers.

2. Suppose that during any hour in a large department store, the average number of
shoppers is 448, with a standard deviation of 21 shoppers. What is the probability of
randomly selecting 49 different shopping hours, counting the shoppers, and having the
sample mean fall between 441 and 446 shoppers, inclusive?
Solution:
1. Calculate µ and  x
µ = 448 shoppers
 x = δ/√n= 21/√49 = 3

2. Calculate Z for X
X X X 
ZX  
X X
441  448 446  448
Z 441    2.33 Z 446    0.67
3 3
3. Find the area covered by the interval
P (441 ≤ X ≤ 446) = P (-2.33 ≤ Z≤ -0.67)
= P (0 to -2.33) - P (0 to - 0.67)
= 0.49010 – 0.24857
= 0.24153
4. Interpret the results
There is a 24.153% chance of randomly selecting 49 hourly periods for which the
sample mean falls between 441 and 446 shoppers.
 10

3. A production company’s 350 hourly employees average 37.6 year of age, with a
standard deviation of 8.3 years. If a random sample of 45 hourly employees is taken,
what is the probability that the sample will have an average age of less than 40 years?
Solution:
1. Calculate µ and  x
µ = 37.6 years n/N= 45/350 > 5%...... FPCF is needed
 N n 8.3 350  45
x  * = x  * 1.16
n N 1 45 350  1

2. Calculate Z for X
X X X 
ZX  
X X
40  37.6
Z 40    2.07
1.16

3. Find the area covered by the interval

P ( X < 40) = P (Z < +2.07)
= 0.5 + P (0 to +2.07)
= 0.5 + 0.48077
= 0.98077
4. Interpret the results
There is a 98.08% chance of randomly selecting 45 hourly employees and their mean
age be less than 40 years.

4. Suppose that a random sample size of 36 is being drawn from a population with a mean
of 278. If 86% of the time the sample mean is less than 280, what is the population
standard deviation?
Solution:
µ = 278 n = 36 P ( X < 280) = 0.86 δ =?
(Z/P=0.36) = +1.08
X  
ZX  X 
X n
280  278 
Z 280  1.85 
X 36
280  278 
 1.08  1.85 
X 6
 11

2   6 * 1.85
 1.08 
X 11 .1
2
X   1.85
1.08

5. A teacher gives a test to a class containing several hundred students. It is known that
the standard deviation of the scores is about 12 points. A random sample of 36 scores is
obtained.
a) What is the probability that the sample mean will differ from the population mean
by more than 6 points?
b) What is the probability that the sample mean will be within 6 points of the
population mean?
Solution:
a)
 12 12
n = 36 δ =12 X    2
n 36 6
P ( X > µ +6) + P ( X < µ - 6) =?
 6  6
Z 6   3 Z  6   3
2 2
P ( X > µ +6) + P (Z> µ - 6) = P (Z > 3) + P (Z < - 3)
= [0.5 – P (0 to +3)] + [0.5 – P (0 to -3)]
= (0.5 – 0.49865) + (0.5 – 0.49865)
= 0.00135(2) = 0.00270
b)
 12 12
n = 36 δ =12 X    2
n 36 6
P (µ - 6≤ X ≤ µ + 6) = P (- 3≤ Z ≤ 3)
= P (0 to 3)*2
= 0.49865*2
= 0.9973
If the population standard deviation is 12, in a random sample of 36 scores there is a
99.73% chance of getting a sample mean score to lie within 6 points of the population
mean.
 12

Sampling Distribution of Proportions ( P )

Some times in statistics it is important to know the proportion of a certain characteristic
in a population. That is, there are numerous problems in business where we want to
know the proportion of items in a population that possess a certain characteristic. For
example,
- A quality control engineer might want to know what proportions of
products of an assembly line are defective.
- A labor economist might want to know what proportion of the labor
force is unemployed.

Whereas the mean is computed by averaging a set of values, the sample proportion is
computed by dividing the frequency that a given characteristic occurs in a sample by
the number of items in the sample.
X
P
n
Where P = sample proportions
X = number of items in a sample that possess the characteristic
n = number of items in the sample
Like other probability distribution, sampling distribution of the proportion is described
by two parameters: the mean of the sample proportions, E ( P ) and the standard
deviation of the proportions,  P which is called the standard error of the proportion.

Properties of Sampling distribution of P

1. As the sampling distribution of the mean does, the population proportion, P, is

always equal to the mean of the sample proportion, i.e., P = E ( P ).
Pq
2. The standard error of the proportion is equal to:  P  , where P= population
n
proportion
q=1–P
n = sample size.
Or
N n
P 
Pq
* , where N  n = finite population correction factor.
n N 1 N 1

The finite population correction factor is not needed if n < 0.05N.

 13

Central Limit Theorem (CLT) and Sampling distribution of P

How does a researcher use the sample proportion in analysis?

Answer: By applying the Central Limit Theorem. The CLT states that normal
distribution approximates the shape of the distribution of sample proportions if np and
nq are greater than 5. Consequently we solve problems involving sample proportions
by using a normal distribution whose mean and standard deviation are:
Pq PP
P  P ,  P  and Z P 
n P

NB: The sampling distribution of p can be approximated by a normal distribution

whenever the sample size is large i.e., np and nq>5.

Example:
1. Suppose that 60% of the electrical contractors in a region use a particular brand of wire.
What is the probability of taking a random sample of size 120 from these electrical
contractors and finding that 0.5 or less use that brand of wire?
Solution:

n = 120 P = 0.6 q = 0.4 P ( p < 0.5) =?

Steps:
1. Check that np and nq > 5
120*0.6 = 120, and 120*0.4 = 48. Both are greater than 5.
2. Calculate  P
Pq 0.6 * 0.4
P  =   0.0477
n 120
3. Calculate Z for p
P P
Zp 
p
0.5  0.6
Z 0.5    2.24
0.0477
4. Find the area covered by the interval
P ( p < 0.5) = P (Z < -2.24)
= 0.5 - P (0 to -2.24)
= 0.5 – 0.48745
 14

= 0.01255

5. Interpret the results

The probability of finding 50% or less of the contractors to use this particular brand
of wire is very low (1.255%) if we take a random sample of 120 contractors.

2. If 10% of a population of parts is defective, what is the probability of randomly selecting

80 parts and finding that 12 or more are defective?
Solution:
0.10 * 0.90
n = 80 P   0.0335
80
P = 0.1
X = 12
p = X/n = 12/80 = 0.15
P ( p > 0.15) =?
0.15  0.1
Z 0.15    1.49
0.0335
P ( p > 0.15) = P (Z > + 1.49)
= 0.5 – P(0 to + 1.49)
= 0.5 – P (0 to + 1.49)
= 0.5 – 0.43189 = 0.06811

About 6.81% of the time, twelve or more defective parts would appear in a random
sample of eighty parts when the population proportion is 0.10.

3. Suppose that a population proportion is .40 and that 80% of the time you draw a
random sample from this population, you get a sample proportion of 0.35 or more.
How large a sample were you taking?
Solution:
P= 0.4
P ( p > 0.35) = 0.80
n =?
Pq
(Z/P= 0.30) – 0.84 P  ; squaring both sides
n
0.35  0.4  .4 * .6 
Z 0.35  0.0595 =  

p  n 
 15

2
 
-0.84 = -0.05/  P 0.0595   .4 * .6  0.0035 = 0.24/n
2

 n 
0.84  P = 0.05 0.0035 = 0.24/n
 P = 0.05/0.84 n = 0.24/0.0035
= 0.0595 n = 68

4. If a population proportion is 0.28 and if the sample size is 140, 30% of the time the
sample proportion will be less than what value if you are taking random samples?
Solution:
P= 0.28 n = 140 P ( p < X) = 0.30 X =?
(Z/P = 0.2) = - 0.52
PP Pq
ZP  P 
p n
P  0.28 0.28 * 0.72
 0.52  
0.0379 140
 0.0197  P  0.28  0.0379

P  0.26

Sampling Distribution of the Difference between Two Independent

Sample Means X1  X 2 

This distribution is concerned with finding the difference between sample means drawn
from two populations; it is interested in determining if the mean of one population is
equal to the mean of another.

For example we might want to know:

- Whether the mean life expectancy of females is equal to the mean life
expectancy for males
- Whether the mean productivity of women and men are equal or not
- Whether the mean CGPA for business students is equal to the mean CGPA
for social science students
- Whether the mean number of white blood cells in a droplet of blood is equal
to the mean number of red blood cells
- etc.
 16

In each case we have two different populations (p1 and p2). Population 1 has mean µ1
and variance,  , and population 2 has mean µ2 and variance  2 2 .
2
1

To make such comparisons we take independent random samples of n1, observations

from population one, P1 and n2 observations from population two, P2 and then calculate
the respective means, which are denoted by x1 and x 2 .

N.B: In sampling distribution of the difference between two means X 1  X 2 , we are  

actually concerned with five different probability distributions.

1st, Two distribution of the two populations, which have means and variances of
µ 1 µ 2 and  1 &  2 respectively.
2 2

2nd, Two sampling distributions of x1 and x 2 with µ 1 = x1 ,  x1 and µ 2 = x2 , x2 .

 
3rd, One sampling distribution of x1  x2 with µ 1 - µ2, and  X X .
1 2

The sampling distribution of the difference between two sample means X 1  X 2 is  

described by two parameters.
1. Mean of the difference between two sample means, µ 1 - µ 2

2. Standard error of the difference between two sample means,  X X =

1 2

 X2
1X2


Variance X 1  X 2 =   X2 1X2
=
 12
n1

 22
n2
 2
X1
 2
X2

[If X and Y are independent random variables: var (X-Y) = var (X) + var (Y)

Where,  12 = variance of population one

 22 = variance of population two
n1 = sample size drawn from population one
n2 = sample size drawn from population two

If more than 5% of the population is sampled without replacement, we must apply the
finite population correction factor and the formula becomes:
  12 N1  n1    22 N 2  n2 
 X2 1X2
=
 n *   
  n * 

 1 N1  1   2 N2 1 
 17

The Central Limit Theorem and the sampling distribution of X1  X 2 

The central limit theorem states that:
If n1 and n2 are greater than 30, the distribution of X 1  X 2 will be approximately  
normal no matter how the original populations are distributed
If the original populations are normally distributed, then the distribution of
 
X 1  X 2 is exactly normally distributed for any values of n1 and n2. This means the
sum or difference of independent normal variables is normally distributed.

To standardize a difference between two sample means X 1  X 2   we use the

following formula:

ZX 
X 1 
 X 2   1   2 
1X2
X
1X2

Example:
1 A financial loan officer claims that the mean monthly payment for credit cards is Br 80
with a variance of 1400 for single females and Br 80 with a variance of 1320 for single
males. You take a random sample of 100 females (population 1) and an independent
random sample of 120 (population 2). What is the probability that the sample mean for
females will be at least Br 5 higher than the sample mean for males?
Solution:
1  80 2  80  12  1400  2 2  1320 n1 = 100 n2 = 120

p x1  x 2  5  ? 

 
Е X1  X 2 = µ 1 - µ 2  X X =
 12

 22
1 2
n1 n2
1400 1320
= 80 – 80 = 
100 120
=0 =5

Z X X 
X 1 
 X 2  1   2 
1 2
 X X1 2

50
Z5   1.0
5
 18

 
p x1  x 2  5  P( Z  1.0)
= 0.5 – P (0 to +1.00)
= 0.5 - 0.34134
= 0.15866

There is a 15.9% chance that the mean monthly payment for credit cards for single
females will be higher than that of single males by at least Birr 5

2 MOHA soft drinks factory produces two soft drinks: 7 up and Pepsi-cola. The daily
production of 7 up averages 1  15,000 bottles and is normally distributed with a
standard deviation 1  2,000 and 2  12,500 bottles and standard deviation  2  2500
bottles. A sample of five randomly selected daily production figures is taken from each
of the plants. What is the probability that the sample mean production for 7 up will be
less than or equal to the sample mean production for Pepsi-cola?
Solution:
1  15,000 bottles 2  12,500bottles 1  2,000  2  2,500bottles n1 = 5 n2 = 5

p x1  x 2  0  ? 

 
Е X1  X 2 = µ 1 - µ 2  X X =
 12

 22
1 2
n1 n2
= 15,000 – 12,500

= 2,500 =
2000 2 
2,500 2
5 5
= 1,431.78

Z X X 
X 1 
 X 2  1   2 
1 2
 X X1 2

0  2,500
Z0   1.75
1431.78
 
p x1  x 2  0  P( Z  1.75)
= 0.5 – P (0 to -1.75)
= 0.5 - 0.45994
= 0.04006

Thus, there is only a 4% chance that the mean productivity for 7UP will be smaller than
the mean productivity for Pepsi-cola. So, if the owner of the two plants found a smaller
 19

first sample mean, say x2  13,500 bottles, in independent random samples of five
randomly selected days from each plant, he would suspect that either the sampling was
faulty or that the difference in the plant’s mean daily outputs had changed.

3 X company claims that the mean annual repair bill for its rental cars is Br 290 and the
standard deviation is Br. 50. Y Company also claims its mean annual repair bill is Br
290 and the standard deviation is Br. 50. If independent random samples of 100 cars
from each company are obtained, what is the probability that x1  x2 exceeds Br 5?
Solution:
1  290  2  290  1  50  2  50 n1 = 100 n2 = 100

p X1  X 2  5  ? 
 
Е X1  X 2 = µ 1 - µ 2  X X =
 12

 22
1 2
n1 n2
= 290 – 290

=0 =
50 2 
50 2
100 100
= 7.071

Z X X 
X 1 
 X 2  1   2 
1 2
 X X1 2

50
Z5   0.711.75
7.071
 
p x1  x 2  0  P( Z  0.71) * 2
= [0.5 – P (0 to 0.71)]2
= [0.5 - 0.26115]2
= 0.4777

There is 47.77% chance that the difference between the mean annual repair bill for X
and Y companies exceed Br 5.

4 Two population of measurement are normally distributed with 1  57 and 2  25. the
two populations standard deviations are 1  12 and  2  6 . Two independent samples
of n1 = n2 = 36 are taken from the populations.
a. What is the expected value of the difference in sample means, x1  x2 ?
b. What is the standard deviation of the distribution of x1  x2 ?
 20

c. What is the shape of the distribution of x1  x2 ? How do you know?

Solution:
a. Expected value of the difference in same means   x1  x2  1   2  
= 57-25
= 32
b. The standard deviation of the distribution of x1  x2 = Standard error of the
difference between two sample means:
 X X =  12

 22
1 2
n1 n2

=
12 2 
( 6) 2
= 2.24
36 36

c. The shape of the distribution of x1  x2 is normal because of the central limit

theorem. The central limit theorem states that

If the populations from which the samples are drawn are normal in shape
then the distribution of x1  x2 will be normal in shape.
If the populations from which samples are drawn are not normal in shape,
then the distribution of x1  x2 will be approximately normal, owing to the
central limit effect, if the sample sizes n1 and n2 are both large.

5 Two production processes are, on the average, identical. Both use an average of
1  2  500 kg, of raw material per day. Both have the same standard deviation of
daily use, 1   2  9 kgs per day. Thus the daily use of material may vary for the two
processes, but on the average they are the same. Find the probability that differ by no
more than 1.0 kg.
Solution:
1  500  2  500 1  9 2 9 n1 = 81 n2 = 36
 
p X 1  X 2  1  P(1  X 1  X 2  1)  ?

 
Е X1  X 2 = µ 1 - µ 2  X X =
 12

 22
1 2
n1 n2
= 500 – 500

=0 =
9 2 
9
2

81 36
= 1.8028
 21

Z X X 
X 1 
 X 2  1   2 
1 2
 X X1 2

1 0
Z1   0.55
1.8028
 
p x1  x 2  1  P( Z  0.55) * 2
= [P (0 to 0.71)] 2
= [0.20884] 2
= 0.41768

SAMPLING DISTRIBUTION OF THE DIFFERENCE BETWEEN TWO

SAMPLE PROPORTIONS P1  P2  .

Frequently we are interested in determining if the proportion of items in one population

that possesses a certain characteristic is the same as the proportion possessing the
characteristic in another population. For example, a doctor who gives one type of
medicine to some patients and another medicine to others may want to determine if the
percentage of people cured by the first medicine is the same as the percentage of people
cured by the second. For this and other similar cases the sampling distribution of the
difference between sample proportions is used. It is also used to measure the
proportion of market share and proportion of vote.

Suppose we take independent samples of size n1 and n2 from two populations. Let p1
and p2 be the proportions of items in each population that possesses a certain
characteristic, and let q1  1  p1  and q2  1  p2  .

If n1p1, n1q1 are greater than 5 and n2p2, n2q2 are greater than 5, then the random variable
 
p1  p2 is approximately normally distributed with
 
Mean: Е p1  p2 = P1 – P2; and

 Pq

Pq
n2
n
N1
n
Variance: Var p1  p2 = 1 1  2 2 ; if 1  0.05 or 2 , finite population
n1 N2
correction factor is used.

To standardize a difference between two sample proportions p  p we

1 2 use the
following formula:
 22

Z P P 
P  P  P  P 
1 2 1 2
1 2
 P P
1 2

Example:
1 At Addis Ababa University there is a movement to re-establish the Students’ Union.
Approximately 90% of the entire students favor the reinstatement. A pro-union student
takes a random sample of 100 students. An anti-union student takes an independent
random sample of 100 students. Let p1 denote the proportion of student who favor
union in a sample taken by the pro-union student and p2 denote the proportion of
students who favor the union in the sample taken by an anti-union student. Calculate
the probability that p1 exceeds p2 by 0.1 or more.
Solution:
Pro-union Anti-union
P1 = 0.9 P2 = 0.9
q1 = 0.1 q2 = 0.1
n1 = 100 n2 = 100

P p1  p 2  0.10  ? 

 
Е p1  p2 = P 1 - P 2  p p =
p1 q1

p2 q2
1 2
n1 n2
= 0.9 – 0.9

=0 =
0.9 * 0.1  0.9 * 0.1
100 100
= 0.04243

Z p p 
 p  p   P  P 
1 2 1 2
1 2
 p p 1 2

0.1  0
Z 0.1   2.36
0.04243
 
P p1  p 2  0.10  P( Z  2.36)
= 0.5 - P (0 to +2.36)
= 0.5 – 0.49086
= 0.00914

2 A TV channel airs two talk shows: Talk-show 1 and Talk-show 2. On a Sunday

afternoon, a random sample of 400 people is taken to estimate p1 , the proportion of the
population that watched the Talk show 1 on the TV channel. On the following Sunday,
 23

an independent random sample of 400 people is taken to estimate p2 , the proportion of

the population who watched the Talk show 2 on the TV channel. If p1 = 0.6 and p2 = 0.5,
find the probability, that p2  p1 in our samples. That is, find p p1  p2  0 .  
Solution:
Talk-show 1 Talk-show2
P1 = 0.6 P2 = 0.5
q1 = 0.4 q2 = 0.5
n1 = 400 n2 = 400

 
P p1  p 2  0  ?

Е p  p  = P - P
1 2 2
 p p =
p1 q1

p2 q2
1 1 2
n1 n2
= 0.6 – 0.5

= 0.10 =
0.6 * 0.4  0.5 * 0.5
400 400
= 0.035

Z p p 
p
1 
 p 2  P1  P2 
1 2
 p p
1 2

0  0.1
Z 0.1   2.86
0.035
 
P p1  p 2  0  P( Z  2.86)
= 0.5 + P (0 to -2.86)
= 0.5 + 0.49788
= 0.99788
 24

CHAPTER Two

STATISTICAL ESTIMATION

Introduction

The sampling distribution of the mean shows how far sample means could be from a
known population mean. Similarly, the sampling distribution of the proportion shows
how far sample proportions could be from a known population proportion. In
estimation, our aim is to determine how far an unknown population mean could be
from the mean of a simple random sample selected from that population; or how far an
unknown population proportion could be from a sample proportion. Those are the
concerns of statistical inference, in which a statement about an unknown population
parameter is derived from information contained in a random sample selected from the
population.

Basic concepts:

Estimation: is the process of using statistics as estimates of parameters. It is any

procedure where sample information is used to estimate/ predict the numerical
value of some population measure (called a parameter).
Estimator- refers to any sample statistic that is used to estimate a population
parameter. E.g. x for  , p for p.
Estimate- is a specific numerical value of our estimator. E.g. x  9, 2, 5

x, p, s 2 , s <<<<<. Estimators
 , p, 2 , <<<<<<< items being estimated
1, 0.5, 9, 3 <<<<<<<... Estimates

Four Important Properties of Estimators

A number of different estimators are possible for the same population parameter, but
some estimators are better than others. To understand how, we need to look at four
important properties of estimators: unbiasedness, efficiency, consistency, and
sufficiency.

Unbiasedness: An estimator exhibits unbiasedness when the mean of the sampling

estimator is equal to the population parameter: E (ө) = Ө.
 25

In general, unbiasedness is a desirable property for an estimator. The sample mean is an

unbiased estimator of the population mean. Similarly, the sample variance is an
unbiased point estimator the population variance because the mean of the sampling
distribution of the sample variance is equal to the population variance. And the sample
proportion is an unbiased estimator of the population proportion. However, because
standard deviation is a nonlinear function of variance, the sample standard deviation is
not an unbiased estimator of population standard deviation. The bias of a point
estimator is: Bias = E (ө) = Ө. If there are a number of unbiased estimators to choose
from, there are three other criteria that could be used to select an estimator.

Efficiency: Efficiency is another standard that can be used to evaluate estimators.

Efficiency refers to the size of the standard error of the statistics. The most efficient
estimator is the one with the smallest variance. Thus, if there are two estimators for Ө
with var (ө1) and var (ө2), then the first estimator ө1 is said to be more efficient than the
second estimator ө2, if var(ө1) < var (ө2) although E(ө1) = E(ө2) = Ө.

Consistency: A third property of estimators, consistency, is related to their behavior as

the sample gets large. A statistic is a consistent estimator of a population parameter if,
as the sample size increases, it becomes almost certain that the value of the statistic
comes very close to the value of the population parameter.

An unbiased estimator is a consistent estimator if the variance approaches 0 as n

increases. For example, the sample mean is an unbiased and a consistent estimator of
population mean. Although the sample standard deviation is not an unbiased estimator
of population standard deviation, it is a consistent estimator of population standard
deviation.

Sufficiency: The last property of a good estimator is sufficiency. A sufficient statistic is

an estimator that utilizes all the information a sample contains about the parameter to
be estimated. For example, the sample mean is a sufficient estimator of the population
mean. This means that no other estimator of the population mean from the same sample
data, such as the sample median, can add any further information about the parameter
(population mean) that is being estimated.

Types of Estimates:
We can make two types of estimates about a population: a point estimate and an
interval estimate.
 26

A point estimate: - is a single number that is used to estimate an unknown population

parameter. It is a single value that is measured from a sample and used as an estimate
of the corresponding population parameter.
The most important point estimates (given that they are single values) are:


o Sample mean x for population mean   ;

o Sample proportion p for population proportion  p  ;
o Sample variance s 2  for population variance  2  and
o Sample standard deviation s  for population standard deviation  

An interval estimate - is a range of values used to estimate a population parameter. It

describes the range of values with in which a parameter might lie. Stated differently, an
interval estimate is a range of values with in which the analyst can declare with some
confidence that the population parameter will fall.

Example:
Suppose we have the sample 10,20,30,40 and 50 selected randomly from a population
whose mean  is unknown.

The sample mean, x ,  =

xi 10  20  30  40  50
 30 is a point estimate of  .
n 5

On the other hand, if we state that the mean,  , is between x  10 , the range of values
from 20 (30-10) to 40 (30+10) is an interval estimate.

INTERVAL ESTIMATION

Point estimators of population parameters, while useful, do not convey as much

information as interval estimators. Point estimation produces a single value as an
estimate of the unknown population parameter. The estimate may or may not be close
to the parameter value; in other words, the estimate may be incorrect. An interval
estimate, on the other hand, is a range of values that conveys the fact that estimation is
an uncertain process. The standard error of the point estimator is used in creating a
range of values; thus a measure of variability is incorporated into interval estimation.
Further, a measure of confidence in the interval estimator is provided; consequently,
interval estimates are also called confidence intervals. For these reasons, interval
estimators are considered more desirable than point estimators.
 27

Interval estimation for population mean, 

As a result of the Central Limit Theorem (discussed in Chapter III) the following z
formula for sample means can be used when sample sizes are large, regardless of the
shape of the population distribution or for smaller sizes if the population is normally
distributed.
X 
Z

n
Rearranging the formula:
  X  Z n
Because the sample mean can be greater than or less than the population mean, z can be
positive or negative. Thus, the preceding expression takes the form:
  X  Z n
The value of the population mean,  , lies somewhere within this range. Rewriting this
expression yields the confidence interval for population mean:
X  Z    X  Z
n n

The confidence interval for population mean is affected by:

1. The population distribution, i.e., whether the population is normally

distributed or not
2. The standard deviation, i.e., whether  is known or not.
3. The sample size, i.e., whether the sample size, n, is large or not.

Confidence internal estimate of  - Normal population,  known

A confidence interval estimate for  is an interval estimate together with a statement of
how confident we are that the interval estimate is correct.

When the population distribution is normal and at the same time  is known, we can
estimate  (regardless of the sample size) using the following formula1.

1
This formula works also for problems which involve large sample size (n>30) even though the
population is not normally distributed. And if n>.05N, finite population correction factor may be used.
 28

  X  Z / 2  n
Where:
X = sample mean
Z = value from the standard normal table reflecting confidence level
σ = population standard deviation
n = sample size
α = the proportion of incorrect statements (α = 1 – C)
 = unknown population mean

From the above formula we can learn that an interval estimate is constructed by adding
and subtracting the error term to and from the point estimate. That is, the point estimate
is found at the center of the confidence interval.

To find the interval estimate of population mean,  we have the following steps.
1. Compute the standard error of the mean  x 
2. Compute  2 from the confidence coefficient.
3. Find the Z value for the  2 from the table
4. Construct the confidence interval
5. Interpret the results

Example:
1. The vice president of operations for Ethiopian Tele Communication Corporation (ETC)
is in the process of developing a strategic management plan. He believes that the ability
to estimate the length of the average phone call on the system is important. He takes a
random sample of 60 calls from the company records and finds that the mean sample
length for a call is 4.26 minutes. Past history for these types of calls has shown that the
population standard deviation for call length is about 1.1 minutes. Assuming that the
population is normally distributed and he wants to have a 95% confidence, help him in
estimating the population mean.
Solution:
n= 60 calls X = 4.26 minutes σ = 1.1 minutes C= 0.95

iv.   X  Z / 2 
1.1
i.  X  = = 0.142
n 60 n
ii. α = 1 – C = 1- 0.95 = 0.05 = 4.26 ± 1.96(0.142)
 2 = 0.05/2 = 0.025 = 4.26 ± 0.28
iii. Z / 2  Z 0.025  1.96
3.98 ≤  ≤ 4.54
 29

The vice-president of ETC can be 95% confident that the average length of a call for the
population is between 3.98 and 4.54 minutes.
2. A survey conducted by “Addis Zemen Gazetta” found that the sample mean age of
men was 44 years and the sample mean age of women was 47 years. Altogether, 454
people from Addis were included in the reader poll –340 women and 114 men. Assume
that the population standard deviation of age for both men and women is 8 years.
a. Develop a 95% confidence interval estimate for the mean age of the population
men who read the gazetta.
b. Develop a 95% confidence interval estimate for the mean age of the population
women who read the gazetta.
c. Compare the widths of the two interval estimates form part (a) & (b) which one
has a better precision? Why?
Solution:
a.
n= 114 men X = 44 years σ = 8 years C= 0.95

iv.   X  Z / 2 
8
i.  X  = = 0.75
n 114 n
ii. α = 1 – C = 1- 0.95 = 0.05 = 44 ± 1.96(0.75)
 2 = 0.05/2 = 0.025 = 44 ± 1.47
iii. Z / 2  Z 0.025  1.96
42.53 ≤  ≤ 45.47
b.
n= 340 women X = 47 years σ = 8 years C= 0.95

iv.   X  Z / 2 
8
i. X  = = 0.434
n 340 n
ii. α = 1 – C = 1- 0.95 = 0.05 = 47 ± 1.96(0.434)
 2 = 0.05/2 = 0.025 = 47 ± 0.85
iii. Z / 2  Z 0.025  1.96
46.15 ≤  ≤ 47.85

c. Part b has a better precision because the sample size is larger as compared with part
a.

3. Time magazine reports information on the time required for caffeine from products
such as coffee and soft drinks to leave the body after consumption. Assume that the
99% confidence interval estimate of the population mean time for adults is 5.6 hrs to 6.4
hrs.
 30

a. What is the point estimate of the mean time for caffeine to leave the body after
consumption?
b. If the population standard deviation is 2 hrs, how large a sample was used to
provide the interval estimate?

Solution:
C = 0.99 Confidence interval: 5.6 ≤  ≤6.4

5.6  6.4
a. point estimate =  6 hours
2
Or;
5.6  X  Z / 2 
 n

6.4  X  Z / 2 
 n
12 = 2 X
X = 6 hours

b. 0.99 σ = 2 hours Confidence interval: 5.6 ≤  ≤6.4 n=?

α = 1- C = 1- 0.99 = 0.01 α/2 = 0.005 Z  / 2  Z 0.005  2.57

6.4  X  Z / 2 
n
6.4  6  2.57 2
n
0.4  5.14 ; rearranging the expression
n
5.14
n
0.4
n  12 .85 ; squaring both sides
n = 165

We state with 99% confidence that the mean time required for caffeine to leave the body
after consumption lies between 5.6 and 6.4 hrs.
Confidence interval estimate of  - Normal population,  unknown, n
large
If we know that the population is normal, and we know the population standard
deviation,  the confidence interval for  should be constructed in the manner already
shown, i.e.,   X  Z / 2  . If the population standard deviation is
n
unknown, it has to be estimated from the sample; i.e., when  is unknown, we use
 31

2
( X i  X )
sample standard deviation: S  . Then, the standard error of the mean,  X ,
n 1
is estimated by the sample standard error of the mean: S X  S .
n

Therefore, the confidence interval to estimate  when population standard deviation is

unknown, population normal and n is large is2
  X  Z / 2 S .
n

Example:
1. Suppose that a car rental firm in Addis wants to estimate the average number of miles
traveled by each of its cars rented. A random sample of 110 cars rented reveals that the
sample means travel distance per day is 85.5 miles, with a sample standard deviation of
19.3 miles. Compute a 99% confidence interval to estimate  .
Solution:
n= 110 rented cars X = 85.5 miles s = 19.3 miles C= 0.99
S 19.3
i. S X  = = 1.84 iv.   X  Z / 2 s
n 110 n
ii. α = 1 – C = 1- 0.99 = 0.01 = 85.5 ± 2.57(1.84)
 2 = 0.01/2 = 0.005 = 85.5 ± 4.73
iii. Z  / 2  Z 0.005  2.57
80.77 ≤  ≤ 90.23

We state with 99% confidence that the average distance traveled by rented cars lies
between 80.77 and 90.23 miles.
2. A study is being conducted in a company that has 800 engineers. A random sample of
50 of these engineers reveals that the average sample age is 34.3 years, and the sample
standard deviation is 8 years.Assuming normality, construct a 98% confidence interval
to estimate the average age of all engineers in this company.
Solution:
n= 50 engineers N = 800 engineers X = 34.3 years s = 8 years C= 0.98
S N n 3 8 800  50
i. S X  * = * = 1.10
n N 1 50 800  1

2
This formula also works for large sample size even though the parent population is not normally
distributed.

3
Since the sample size is greater than 5% of the population size, finite population multiplier is used to
calculate the sample standard error of the mean.
 32

ii. α = 1 – C = 1- 0.98 = 0.02

 2 = 0.02/2 = 0.01
iii. Z  / 2  Z 0.01  2.33
N n
iv.   X  Z / 2 s *
n N 1
= 34.3 ± 2.33(1.10)
= 34.3 ± 2.56

31.74 ≤  ≤ 36.86
We state with 98% confidence that the mean age of engineers lies between 31.74 and
36.86 years.

Confidence interval for    unknown, n-small, population normal

If the sample size is small (n<30), we can develop an interval estimate of a population
mean only if the population has a normal probability distribution. If the sample
standard deviation s is used as an estimator of the population standard deviation  and
if the population has a normal distribution, interval estimation of the population mean
can be based up on a probability distribution known as t-distribution.

Characteristics of t-distribution
1. The t-distribution is symmetric about its mean (0) and ranges from - ∞ to ∞.
2. The t-distribution is bell-shaped (unimodal) and has approximately the same
appearance as the standard normal distribution (Z- distribution).
3. The t-distribution depends on a parameter ν (Greek Nu) 4, called the degrees of
freedom of the distribution. Ν = n -1, where n is sample size. The degree of freedom,
ν, refers to the number of values we can choose freely.
4. The variance of the t-distribution is ν/ (ν-2) for ν>2.
5. The variance of the t-distribution always exceeds 1.

4
What are degrees of freedom? We can define them as the number of values we can choose freely. In
general, the degrees of freedom for a t statistic are the degrees of freedom associated with the sum of
squares used to obtain an estimate of the variance. The variance estimate depends on not only on the
sample size but also on how many parameters must be estimated with the sample:
Degrees of  Number of  Number of parameters that
freedom Observatio ns must be estimated beforehand
Here we calculate sample variance by using n observations and estimating one parameter (the mean).
Thus, there are (n – 1) degrees of freedom.
 33

6. As ν increases, the variance of the t-distribution approaches 1 and the shape

approaches that of the standard normal distribution.
7. Because the variance of the t-distribution exceeds 1.0 while the variance of the Z-
distribution equals 1, the t-distribution is slightly flatter in the middle than the Z-
distribution and has thicker tails.
8. The t-distribution is a family of distributions with a different density function
corresponding to each different value of the parameter ν. That is, there is a separate
t-distribution for each sample size. In proper statistical language, we would say,
“There is a different t-distribution for each of the possible degrees of freedom”.
9. The t formula for sample when  is unknown, the sample size is small, and the
population is normally distributed is: t  X    X   This formula is essentially
SX s
n
the same as the z-formula, but the distribution table values are not.

The confidence interval to estimate  becomes:

  X  t / 2,v s
n

Where: X = sample mean

α=1–C
ν = n – 1 (degrees of freedom)
s = sample standard deviation
n = sample size
 = unknown population mean
Steps:
1. Calculate degrees of freedom (v=n-1) and sample standard error of the mean.
2. Compute  2

3. Look up t / 2 ,V
4. Construct the confidence interval
5. Interpret results

Example:
1. If a random sample of 27 items produces x = 128.4 and s = 20.6. What is the 98%
confidence interval for  ? Assume that x is normally distributed for the population.
What is the point estimate?
Solution:
 34

The point estimate of the population mean is the sample mean, in this case 128.4 is the
point estimate.

n= 27 X = 128.4 s = 20.6 C= 0.98

S 20.6
i. S X  = = 3.96 ν = n – 1 = 27-1 = 26
n 27
ii. α = 1 – C = 1- 0.98 = 0.02
 2 = 0.02/2 = 0.01
iii. t / 2,v t 0.01, 26  2.479
iv.   X  t / 2,v s
n
= 128.4 ± 2.479(3.96)
= 128.4 ± 9.82

118.56 ≤  ≤ 138.22
We state with 98% confidence that the population mean lies between 118.56 and 138.23.

2. A sample of 20 cab fares in Bahir Dar city shows a sample mean of Br 2.50 and a sample
standard deviation of Br. 0.50. Develop a 90% confidence interval estimate of the mean
cab fares in Bahir Dar city. Assume the population of cab fares has a normal
distribution.

n= 20 X = Birr 2.50 s = Birr 0.50 C= 0.90

S 0.5
i. S X  = = 0.112 ν = n – 1 = 20-1 = 19
n 20
ii. α = 1 – C = 1- 0.90 = 0.10
 2 = 0.10/2 = 0.05
iii. t / 2,v t 0.05,19  1.729
iv.   X  t / 2,v s
n
= 2.50 ± 1.729(0.112)
= 2.50 ± 0.194

2.31 ≤  ≤ 2.69

We state with 90% confidence that the mean of cab fares in Bahir Dar city lies between
Birr 2.31 and 2.69.
 35

3. Sales personnel for X Company are required to submit weekly reports listing customer
contacts made during the week. A sample of 61 weekly contact reports showed a mean
of 22.4 customer contacts per week for the sales personnel. The sample standard
deviation was 5 contacts.

a. Develop a 95% confidence interval estimate for the mean number of weekly
customer contacts for the population of sales personnel.
b. Assume that the population of weekly contact data has a normal distribution.
Use the t distribution to develop a 95% confidence interval for the mean number
of weekly customer contacts.
c. Compare your answer for parts (a) and (b). What do you conclude from your
results?

Solution:
a.
n= 61 weekly contact reports5 X = 22.4 contacts s = 5 contacts C= 0.95
S 5
i. S X  = = 0.64
n 61
ii. α = 1 – C = 1- 0.95 = 0.05
 2 = 0.05/2 = 0.025
iii. Z  / 2  Z 0.025  1.96
iv.   X  Z / 2 s
n
= 22.4 ± 1.96(0.64)
= 22.4 ± 1.25

21.15 ≤  ≤ 23.65

I state with 95% confidence that the mean weekly contact lies between 21.15 and 23.65
contacts.

b.
n= 61 weekly contact reports X = 22.4 contacts s = 5 contacts C= 0.95
S 5
i. S X  = = 0.64 ν = n – 1 = 61 – 1 = 60
n 61
ii. α = 1 – C = 1- 0.95 = 0.05
 2 = 0.05/2 = 0.025

5
Since the sample size is large, we use the Z-distribution to construct the confidence interval.
 36

iii. t / 2,v t 0.025,60  2.00

iv.   X  t / 2,v s
n
= 22.4 ± 2.00 (0.64)
= 22.4 ± 1.28

21.12 ≤  ≤ 23.68

I state with 95% confidence that the mean weekly contact lies between 21.12 and 23.68
contacts.

c. As the sample size increases, the t-distribution and z (normal) distribution

approximate to be equal.

Confidence interval for   n small,  unknown, population not normal

This is solved by non-parametric tests, which do not require assumption about the
underlying form of the population data
Interval Estimation of the Population Proportion

We know that a sample proportion, p , is an unbiased estimator of a population

proportion P and if the sample size is large then, the sampling distribution of p is
PP PP
normal with Z   .
p Pq
n

However, here p is unknown and we want to estimate p by p and hence z becomes

PP pq
Z . That is,  p is substituted by S p  n
pq
n

Solving for P results in P  p  Z p q n and since Z can assume both positive and

negative values, it becomes P  p  Z p q n .

 37

P  p  Z / 2 pq
Since Z represents the confidence level we write it as n
 p  Z / 2 S p
Where: p = sample proportion
q =1- p
α=1–C
n = sample size
P = unknown population proportion
Example:
1. Recently, a study of 87 randomly selected companies with telemarketing operation was
completed. The study revealed that 39% of the sampled companies had used
telemarketing to assist them in order processing. Using this information estimate the
population proportion of telemarketing companies who use their telemarketing
operation to assist them in order processing taking a 95% confidence level.
Solution:
n= 87 p = 0.39 q = 0.61 C = 0.95

i. S p  pq 0.61 * 0.39
n = 87
= 0.0523
ii. α = 1 – C = 1- 0.95 = 0.05
 2 = 0.05/2 = 0.025
iii. Z  / 2  Z 0.025  1.96
iv. P  p  Z / 2 S p
= 0.39 ± 1.96(0.0523)
= 0.39 ± 0.1025

0.2875 ≤ P ≤ 0.4925
We state with 955 confidence that the proportion of companies which use telemarketing
to assist order processing lies between 0.2875 and

2. A fast food restaurant took a random sample of 400 customers to determine the
proportion of customers who are female. A confidence interval of .73 to .87 was
reported.
a. Find the number of females and the sample proportion
b. Find the level of confidence of this interval
Solution:
a.
n= 400 0.73 ≤ P ≤ 0.87 p =? Number of females=?
 38

0.73  0.87
Point estimate =  0.80
2
Or;
0.73  p  Z / 2 s p


0.87  p  Z / 2 s p

1.60 = 2 p
p = 0.8
Number of females (X) = n* p = 400*0.8 = 320
b. P  p  Z / 2 S p
0.87 = 0.8+ Z / 2 S p

0.07 = Z / 2 0.8 * 0.2

400
0.07 = Z  / 2 * 0.02
3.50 = Z  / 2
(P/Z=3.5) = 0.49977
C = 0.49977*2
= 99.954%
3. A random sample of 400 faculty members at AAU contained 120 people who believed
that the University should improve its library service. On the basis of this sample
information, an analyst calculated the confidence interval (.25, .35) for the population
proportion of faculty members favoring improvement. What is the level of confidence
of this interval?
Solution:
n= 400 X = 120 p = 0.30 Interval estimate 0.25 ≤ P ≤ 0.30 C =?
P  p  Z / 2 S p
0.25 = 0.30 - Z / 2 S p

0.05 = Z / 2 0.7 * 0.3

400
0.05 = Z  / 2 * 0.023
2.17 = Z  / 2
(P/Z=2.17) = 0.485
C = 0.485*2
= 97%
Interval Estimation of the Difference between two independent Means
 39

It is clear that the unbiased point estimate of the difference between the means of two
populations 1  2  is the difference between two sample means x1  x2 , where each  
sample is a random sample taken from the respective target population. The confidence
interval is constructed by adding the relevant standard error value which is called
standard error of the difference between means and the confidence level desired.

Interval Estimation of 1  2 - population normal,  known

If the two parent populations are normal, then the sampling distribution of the
difference between two means will be normally distributed regardless of n (sample
size). And we can estimate 1  2 (regardless of n1 & n2 using the following formula;
given that  1 &  2 are known.

 1   2  X 1  X 2  Z  / 2 X  X 6
1 2

 12  22
X   X2   X2  
1X2 1 2
n1 n2
When  1 and  2 are not known, the standard error between two sample means
 x  x 
1 2 is estimated by the sample standard error of the difference between two
S12 S 22
sample means, S X  S X2  S X2   , and the interval estimation takes the
1X2 1 2
n1 n2

following form: 1   2  X 1  X 2  Z / 2 S X  X , given that the sample sizes are

1 2

large.

Example:
1. In a sex discrimination case, an employee alleged that a large corporation paid men
more than women for comparable work. Let population 1 represents all male
employees performing certain jobs and population 2 represent all female employees
performing comparable jobs at the corporation. Independent samples are taken of
n1  100 males and n2  100 females; the sample means are x1  Birr 20,600 and
x 2  Birr 19,700 , and the sample standard deviations are s1  Birr 3,000 and
s 2  Birr 2,500 . Construct a 95% confidence interval for 1  2 . What do you conclude
from this?
Solution:
Male employee’s Female employees

6
This formula works also for problems which involve large sample sizes n1 & n2  30 even though the
parent population may not be normally distributed.
 40

n1  100 Males n2  100 females C= 0.95

x1  Birr 20,600 x 2  Birr 19,700
s1  Birr 3,000 s 2  Birr 2,500
Steps:
i. Calculate the (sample) standard error of the difference between two means
S12 S 22 (3,000) 2 (2,500) 2
SX      142,500  390.51
1X2
n1 n2 100 100
ii. Compute  2
α = 1-C = 1- 0.95 = 0.05
α/2 = 0.05/2 = 0.025
iii. Look up Z  / 2  Z 0.025  1.96
iv. Construct the confidence interval
1   2  X 1  X 2  Z / 2 S X  X
1 2

1   2  (20,600  19,700)  1.96(390.51) = 900 ±

765.40
134.60 ≤ 1  2 ≤ 1,665.40
We state with 95% confidence that the mean salary difference between the male and
female workers lies between Birr 134.60 and Birr 1665.40

Because this interval contains only positive values, we can be quite confident that
1  2  > 0. Thus, it reasonable to assume that the mean salary for males exceeds the
mean salary for females.

2. A farmer wants to determine if different types of feed can influence the mean member
of eggs that hens lay per month. In a random sample of 100 hens that ate feed 1, the
average member of eggs per month was x1  15.2 with variance 4. In a random sample
of 100 hens that ate feed2, the average number of eggs per month was x2  14 with
variance 4. Construct a 95% confidence interval for 1  2 . What do you conclude?
Solution:
Feed 1 Feed 2
n1  100 hens n2  100 hens C= 0.95
x1  15.2 eggs x 2  14 eggs
s  4 eggs
2
1 s 22  4 eggs
Steps:
i. Calculate the (sample) standard error of the difference between two means
S12 S 22 4 4
SX      0.08  0.283
1X2
n1 n2 100 100
 41

ii. Compute  2
α = 1-C = 1- 0.95 = 0.05
α/2 = 0.05/2 = 0.025
iii. Look up Z  / 2  Z 0.025  1.96
iv. Construct the confidence interval
1   2  X 1  X 2  Z / 2 S X  X 1 2

1  2  (15.2  14)  1.96(0.283)

= 1.2 ± 0.5547
0.6453 ≤ 1  2 ≤ 1.7547

We state with 95% confidence that the mean number of eggs laid by hens which ate the
two type of feeds lies between 0.6543 eggs and 1.7547 eggs.

Since the interval contains only positive values, then those hens which ate feed type 1
are more productive than those hens that ate feed type 2.

Interval estimation of 1  2 population normal, 1and 2 unknown,

n1 & n2  30.

When the sample sizes are small, the population standard deviations are unknown, and
the population distributions are normal, we use t-distribution to construct a confidence
interval for 1  2 . Moreover, to use a t-distribution we have to assume that the two
variances (standard deviations) are equal. In short, to use a t-distribution for
constructing confidence interval for 1  2 , we assume the following:

1. The population standard deviation  1 and  2 are not known.

2. The sample sizes are small ( n1 & n2  30. ).
3. The populations are assumed to be approximately normally distributed.
4.
2

The two (unknown) population variances are equal  2   2   2 .
1

Given these assumptions, the sampling distribution of x1  x2 is normally distributed
regardless of the sizes. Because of the equal variances assumption, the standard error of
2 2 1 1
the means is written as  X     2    .
1X2
n1 n2  n1 n 2 
If the variance  2 of the populations is known, 1   2  X 1  X 2  Z  / 2 X  X can be used
1 2

to develop the interval estimate of 1  2 .

 42

However, in most cases,  2 is unknown; thus the two sample variances s1 2 and s 22 must
2 2
be used to develop the estimate of  2 . Since  X   is based on the
1X2
n1 n2
assumption that  1   2   2 , we do not need a separate estimates of  12and 2 2 . In fact,
2 2

we can combine the data from the two samples to provide the best single estimate of  2 .
The process of combing the results of two independent simple random samples to
provide one estimate of  2 is referred to as pooling. The pooled estimator of variance,
 
 2 , denoted by s 2 s p is the weighted average of the two sample variances, s1 2 and s 22 ,
2

with the degrees of freedom associated with each sample being used as the weights.
The formula for the pooled estimator of  2 is:
n  1S12  n2  1S 22  X i1  X 1    X i 2  X 2 
2

S2  1 
n1  n2  2 n1  n 2  2
P

Where:
S P2 = pooled estimate of the variance
n1 = sample size drawn from population 1
n 2 = sample size drawn from population 2
S 12 = sample variance of the sample drawn from population 1
S 22 = sample variance of the sample drawn from population 2
n1+n2-2 = pooled degrees of freedom

Based on the assumption that the population standard deviations are equal, the
standard error of the difference between means is estimated by the sample standard
error of the difference between two sample means, S X  X , according to the following
1 2

1 1 S P2 S P2
equation: S X  S P2    = 
1X2
 n1 n 2  n1 n2

The confidence interval for 1  2 when the common standard deviations 1   2  

are not known is based on t-distribution, and is given by:
 1   2  X 1  X 2  t / 2 , v S X
1X2

Where:
1 1
SX  S P2   
1X2
 n1 n 2 
ν = pooled degrees of freedom (n1 + n2 – 2)
 43

Example:
1. Two manufacturing companies produce drill tips that are used to cut holes in steel
sheets. A customer wishing to know which drill tips have the longer site purchases,
independent samples of n1  20 drill tips from company 1 and n2  15 drill tips from
company 2. The mean lives of the drill tips are x1  78 minutes and x 2  84 minutes.
The population variances are unknown but assumed to be equal. The sample variances
are s1 2  41 and s 2 2  36 . Construct a 95% confidence interval for 1  2 assuming that
the two populations are normally distributed.

Solution:
Company One Company Two
n1  20 drill tips n2  15 drill tips C = 0.95
x1  78 minutes x 2  84 minutes
S 12 = 41 S 22 = 36

i. Calculate the sample standard error of the difference between two means and the
pooled degrees of freedom
S P2 S P2 n1  1S12  n2  1S 22  1 1

SX    n 
1X2
n1 n 2 n1  n 2  2  1 n 2 
20  141  15  136  1 1
=   
20  15  2  20 15 
1,283  1 1
=   
33  20 15 
= 2.13
ν = n1 +n2 -2
= 20 + 15 -2 = 33
ii. Compute  2 and look up t  / 2 , v
α = 1-C = 1- 0.95 = 0.05
α/2 = 0.05/2 = 0.025
t  / 2 , v = t 0.025,33 = 2.04
iii. Construct the confidence interval
 
 1   2  X 1  X 2  t / 2 , v S X
1X2

 78  84  2.042.13

  6  4.34
-10.34 ≤ 1  2 ≤ -1.66
 44

The 95% confidence interval is (-10.34 to –1.66). This interval contains only negative
values indicating that the drill tips made by company 1 do not last as long on average,
as those made by company 2.

2. Five year children were being studied to determine whether children whose parents are
college graduates watched more or less TV than children whose parents are not college
graduates. Independent random samples of 21 children were selected from each
population. The sample means and variances were x1  22hrs, s12  16, x 2  26, ands2 2  14.
The population variances are assumed to be equal and the populations are assumed to
be normal. Calculate the 95% confidence interval for the difference between the two
population means.
Solution:
College graduates’ children Non-college graduates’ children
n1  21 children n2  21 children C = 0.95
x1  22 hours x 2  26 hours
S12 = 16 S 22 = 14

i. Calculate the sample standard error of the difference between two means and the
pooled degrees of freedom
S P2 S P2 n1  1S12  n2  1S 22  1 1
SX
1X2
  
n1  n 2  2  n  n 
n1 n 2  1 2 

21  116  21  114  1 1

=   
21  21  2  21 21 
600  1 1
=   
40  21 21 
= 1.2
ν = n1 +n2 -2
= 21 + 21 -2 = 40
ii. Compute  2 and look up t  / 2 , v
α = 1-C = 1- 0.95 = 0.05
α/2 = 0.05/2 = 0.025
t  / 2 , v = t 0.025, 40 = 2.021
iii. Construct the confidence interval
 
 1   2  X 1  X 2  t / 2 , v S X 1X2

 22  26  2.0211.2

  6  2.43
 45

-6.43 ≤ 1  2 ≤ -1.58
We state with 95% confidence that the mean difference between the two populations
lies between –6.42 and –1.58
 Children whose parents are college graduates watched less TV than children
whose parents are not college graduates.

To use a t-test to construct a confidence interval we assume that: populations are

normal, population standard deviations are equal and sample sizes are less than 30.
However,  1 and  2 may not be equal. In such cases the estimated sample variances of
the difference between two sample means is calculated as:
S12 S 22
SX  
1X2
n1 n2
2
 S12 S 22 
  
However, the degrees of freedom is calculated as: df   n1 n2 
2 2
 S12   S 22 
   
n  n 
 1  2

n1  1 n2  1

Confidence interval for the difference between two population

proportions  p1  p2 

We know that the unbiased estimator of the difference between the proportions of two
populations  p1  p2  is the difference between two sample proportions  p1  p2  , where
each sample is a random sample taken from the respective target population.
Moreover, based on CLT, if n1 p1, n1 q1 and n 2 p 2, n 2 q 2 are greater than 5, the sampling

distribution of p1  p 2 is normal with Z 

P  P   P  P 
1 2 1 2

P1 q1 p 2 q 2

n1 n2

However, here p1andp2 are unknown, and we want to estimate p1andp2 by p 1 and p 2
respectively, and hence Z becomes:

Z 1
 
P  P2  P1  P2 
. That is,  p  p is substituted by S p  p
1 2 1 2
P1 q1 p 2 q 2

n1 n2
 46

Solving for p1  p2 results in:

p1 q1 p 2 q 2
P1  P2  p 1  p 2  Z  , and since Z can assume
n1 n2
both positive and negative values, it becomes:

p1 q1 p 2 q 2
P1  P2  p 1  p 2  Z 
n1 n2

Since z represents the confidence level we write it as

p1 q1 p 2 q 2
P1  P2  p 1  p 2  Z  / 2 
n1 n2
Where:
p1 = the sample proportion of success in the first sample
p 2 = the sample proportion of in the second sample
q 1 = 1- p1
q 2 = 1- p 2
n1 = sample size drawn from the first population
n 2 = sample size drawn from the second population
α=1-C

This formula holds true provided that n1 p1, n1 q1  5 and n 2 p 2 , n 2 q 2  5.

Example:
1. A TV executive is interested in determining if the proportion of people who watch a
late-night talk show is higher with the regular host or a guest host. In a random sample
of 400 people, 175 watch the show when the regular host is on. In an independent
random sample of 500 people, 185 watch the show a guest host is on. Calculate a 95%
confidence interval for p1  p2 . What do you conclude?

Solution:
Regular host Guest Host
n1 = 400 p1 = 0.4375 n 2 = 500 p 2 = 0.37
X1 = 175 q 1 = 0.5625 X2 = 185 q 2 = 0.63
C = 0.95
 47

i. Calculate the sample standard error of the diff. between two proportions
p1 q1 p 2 q 2 0.4375 * 0.5625 0.37 * 0.63
S p p      0.033
1 2
n1 n2 400 500
ii. Compute  2
α = 1-C = 1- 0.95 = 0.05
α/2 = 0.05/2 = 0.025
iii. Look up Z  / 2  Z 0.025  1.96
iv. Construct the confidence interval
p1 q1 p 2 q 2
P1  P2  p 1  p 2  Z  / 2 
n1 n2
 0.4375  037  1.96(0.033)
= 0.0675 ± 0.065

0.0025 ≤ p1  p2 ≤ 0.1325

We state with 95% confidence that the true difference between p1  p2 is between 0.0025
and 0.1325. Since this interval contains only positive value it is reasonable to say that
the proportion of people who watch TV when the regular host is on is greater than
when the guest host is on.

2. A city planner claims that home owner tend to have closer ties to their community than
do renters. Thus, home owners are more willing to pay for good schools and
recreational facilities than are renters. In a random sample of 120 home owners, 51
stated that the local tax rates were too high and 69 stated that tax rates were “about
right” In an independent random sample of 200 renters 70 stated that the tax rates were
too high and 130 thought they were “about right” .
a. Find a 99% confidence interval for the difference in proportions who think that
taxes are too high.
b. Do the data support the city planners claim?

Solution:
Home Owners Renters
n1 = 120 p1 = 0.425 n 2 = 200 p 2 = 0.35
X1 = 51 q 1 = 0.575 X2 = 70 q 2 = 0.65
C = 0.99

i. Calculate the sample standard error of the diff. between two proportions
 48

p1 q1 p 2 q 2 0.425 * 0.575 0.35 * 0.65

S p p      0.056
1 2
n1 n2 120 200
ii. Compute  2
α = 1-C = 1- 0.99 = 0.01
α/2 = 0.01/2 = 0.005
iii. Look up Z  / 2  Z 0.005  2.57
iv. Construct the confidence interval
p1 q1 p 2 q 2
P1  P2  p 1  p 2  Z  / 2 
n1 n2
 0.425  035  2.57(0.056)
= 0.075 ± 0.144

- 0.069 ≤ p1  p2 ≤ 0.219
We state with 99% confidence that the difference between the proportion of home
owners and renters who said that the tax rates are too high lies between –0.069 and
0.219.

 Since the confidence interval contains positive, zero, and positive values, we can not
certainly state that home owners are more willing to pay for good schools and
recreational facilities than are renters. Hence, the data do not necessarily support
the city planner’s claim.

Determination of Sample Size (n)

The reason for taking a sample from a population is that it would be too costly to gather
data for the whole population. But collecting sample data also costs money; and the
larger the sample, the higher the cost. To hold cost down, we want to use as small a
sample as possible. On the other hand, we want a sample to be large enough to provide
“good” approximation/estimates of population parameters. Consequently, the question
is “How large should the sample be?”

The answer depends on three factors:

1) How precise (narrow) do we want a confidence interval to be?
2) How confident do we want to be that the interval estimate is correct?
3) How variable is the population being sampled?
 49

Sample size for estimating population mean, 

The confidence interval for  is   X  Z / 2  .
n

From the above expression Z  / 2 is called error of estimation (e). That is, the
n
difference between x and  which results from the sampling process. So

e = Z / 2
n
2 Z 2 / 2 2
Squaring both sides results in e  Z  / 2
2 2
. Solving for n results in, n 
n e2


2
Z
n    / 2 
 e 
Example:
1. A gasoline service station shows a standard deviation of Birr 6.25 for the changes made
by the credit card customers. Assume that the station’s management would like to
estimate the population mean gasoline bill for its credit card customers to be with in ±
Birr 1.00. For a 95% confidence level, how large a sample would be necessary?
Solution:
e = Birr 1.00 σ = Birr 6.25 C = 0.95 Z  / 2  Z 0.025  1.96
Z  
2

n    / 2 
 e 
2
 1.96 * 6.25 
n    7
 1 
150.06  151

2. The National Travel and Tour Organization (NTO) would like to estimate the mean
amount of money spent by a tourist to be with in Birr 100 with 95% confidence. If the
amount of money spent by tourist is considered to be normally distributed with a
standard deviation of Br 200, what sample size would be necessary for the NTO to meet
their objective in estimating this mean amount?
Solution:
e = Birr 100 σ = Birr 200 C = 0.95 Z  / 2  Z 0.025  1.96

7
It a procedure for determining sample size produces a non-integer value, always round to the next
larger integer.
 50

Z  
2

n    / 2 
 e 
2
 1.96 * 200 
n   
 100 
15.37  16

If population standard deviation,  , is unknown we have to make an educated guess or

take a pilot sample and estimate it.
H L
- The rough approximation is   because 95.4% of the total population
4
falls with in  2 .
  1 4 range

Relationship between the error term and sample size

Reducing error term in estimation of an interval estimate to 1/a of the original amount,
while holding the confidence level constant requires a sample size of a2 times the
original sample size.

Sample size for estimating population proportion, p.

pq pq
The confidence interval for p is P  p  Z  / 2 . The expression Z  / 2 is called the
n n
error term (e). That is,
pq
e  Z / 2 , squaring both sides
n
pq
e 2  Z 2 / 2 , solving for n
n
Z 2 / 2 p q
np 
e2
Since we are trying to determine n, we cannot have p and q . Instead, we should have p
2
 Z / 2 
n 
and q. so it becomes p   pq
 e 
Example
1. Suppose that a production facility purchases a particular component parts in large lots
from a supplier. The production manager wants to estimate the proportion of defective
parts received from this supplier. She believes that the proportion of defects is no more
 51

than 0.2 and wants to be with in 0.02 of the true proportion of defects with a 90% level
of confidence. How large a sample should she take?
Solution:
e = 0.02 p = 0.2 q =0.8 C = 0.90 Z  / 2  Z 0.05  1.64
2
Z 
n p    / 2  pq
 e 
2
 1.64 
np    0 .2 * 0 .8
 0.02 
 1075.84  1076

2. What is the largest sample size that would be needed in estimating a population
proportion to with in ± 0.02, with a confidence coefficient of 0.95?
Solution:
e = 0.02 C = 0.95 Z  / 2  Z 0.025  1.96

The largest sample size would be obtained when p = 0.5. So,

2
Z 
n p    / 2  pq
 e 
2
 1.96 
np    0.5 * 0.5
 0.02 
 2401
If p is unknown and there is no possibility of estimating it, use 0.5 as the value of p
because it will generate the greatest possible sample size as compared with other
values.

Determining Sample Size When Estimating 1  2

When taking two random samples and using the difference in sample means to estimate
the difference in population means, a researcher should have an idea of how large the

sample sizes need to be solving for n form the formula Z 

 
X 1  X 2   1   2 
does
 12  22

n1 n2
not look promising, because the equation has nine variables including two different
values of n. However making some assumptions can generate a workable sample size
formula.
1. Variances of the two populations are the same:  1   2   2
2 2

2. The sample size for each sample is the same: n1  n2  n

 52

The difference between x1  x 2 and 1  2 is the error of estimation. Or

 
e  x1  x 2  1  2 .

Incorporating these assumptions into the z-formula yields

e e e e
Z = = =
2 2 2 2 2 2

2


n n n n n

Solving for n produces the sample size:

2 * Z 2 / 2 2 Z  
2

n  2  / 2 
e2  e 
The above formula suggests that the necessary sample sizes for comparing two sample
means are each twice as large as the required sample size for estimating single sample
means. It is clear that the larger the sample, the more it costs. Thus sample size
formulas can be effective aids in ensuring that a research project’s goals are met and
that the cost of sampling is minimized.

Example:
1. A college admissions officer wants to estimate the difference in the average GMAT
scores of men and women. She plans to take a random sample of men and women who
have taken the GMAT at the same time. She wants to be with in 10 points of the true
difference in the mean scores of men and women and 95% confident of her results. Past
GMAT test results indicate that the standard deviation of GMAT test scores is about 105
points. How large the sample sizes be?
Solution:
e = 10 points σ = 105 points C = 0.95 Z  / 2  Z 0.025  1.96 n=?
Z  
2

n  2  / 2 
 e 
2
 1.96 *105 
n  2 
 10 
 2( 421.54)
 847.10  848

2. A researcher wants to estimate the difference between the average price of a 21-inch
black and white TV and the average price of a 21-inch color TV set. He believes that the
standard deviation of the price of a 21-inch TV set is about Birr 100. He wants to be 99%
 53

confident of his results and with in Birr 20 of the true difference. How large a sample
should he take for each type of television set?
Solution:
e = Birr 20 σ = Birr 100 C = 0.99 Z  / 2  Z 0.005  2.57 n=?

 Z  / 2 
2

n  2 
 e 

.
 54

CHAPTER Three

HYPOTHESIS TESTING

Introduction
In Chapter II Estimation, we used the information obtained in a simple random sample
to construct a confidence interval estimate of the unknown value of a population
parameter. In this chapter, hypothesis testing, we start with an assumed value of a
population parameter: then we shall use sample evidence to decide wither the assumed
value is unreasonable and should be rejected or whether it should be accepted.

The assumptions we make about the values of population parameters are called
hypotheses. Sample evidence is used to test the reasonableness of hypotheses; hence,
the statistical inferences made in this chapter are referred to as hypothesis testing. A
procedure based on sample evidence and probability theory to determine whether the
hypothesis is a reasonable statement is called hypothesis testing.

In hypotheses testing we begin by making a tentative assumption about a population

parameter. This tentative assumption is called the null hypothesis, and is denoted by
Ho - it is the assumption we wish to test. We then define another hypothesis, called the
alternative hypothesis, which is the opposite of what is stated in the null hypothesis.
This alternative hypothesis specifies all possible values of the population parameter that
are not specified in the null hypothesis, and in denoted by Ha. The hypothesis testing
procedure involves using data from a sample to test the two competing statements
indicated by Ho and Ha. Ho and Ha are mutually exclusive and collectively
exhaustive.

In the process of hypothesis testing, the null hypothesis is initially assumed to be true.
The data are gathered and examined to determine whether the evidence is strong
enough away from the null hypothesis to reject it when the researcher in testing an
industry standard or a widely accepted values, the standard or accepted value is
assumed to be true in the null hypothesis. Null in this sense means that nothing is new,
or there in no new value or standard. The burden is then placed on the researcher to
demonstrate through gathered data that the null hypothesis is false.

Ha, Hi = research hypothesis a statement that in accepted if the sample data provide
enough evidence that the Ho is false.
 55

The situation encountered in hypothesis testing is similar to the one encountered in a

criminal trial. In a criminal trial the assumption is that the defendant is innocent. Thus,
the null hypothesis is one of innocence.
The opposite of the null hypothesis is the alternative hypothesis that the defendant is
guilty. Thus, the hypothesis far a criminal trial would be written

Ho: The defendant is innocent

Ha: The defendant is guilty

To test these competing statements, or hypotheses, a trial is held. The testimony and
evidence obtained during the trial provide the sample information. If the sample
information is not inconsistent with the assumption of innocence, the null hypothesis
that the defendant is innocent cannot be rejected. However, if the sample information is
inconsistent with the assumption of innocence, the null hypothesis will be rejected. In
this case, action will be taken based upon the alternative hypothesis that the defendant
is guilty.

Example

1. The manager of a hotel has stated that the mean guest bill for a weekend is Birr 400 or
less. A member of the hotel’s accounting staff has noticed that the total charges for
guest bills have been increasing in recent months. The accountant will use a sample of
weekend guest bills to test the manager’s claim.

Required: State the null and alternative hypotheses

Solution:

Ho: µ  Birr 400

Ha:   Birr 400

2. Production workers at XY Company have been trained in their jobs by using two
different training programs. The company training director would like to know
whether there is a difference in mean productivity for workers trained in the two
programs.
Required: Develop the null and alternative hypotheses.
Solution
Ho: 1 = 2 or 1 - 2 = 0
Ha: 1  2 1 - 2  0
 56

3. The manager at a drugstore claims that the company’s employees are honest. However,
there have been many shortages from the cash register lately.

Required: Specify the null and alternative hypothesis

Solution:
Ho: Employees are honest
Ha: Employees are dishonest

“In many situations, the choice of Ho and Ha is not obvious; in such cases, judgment on
the part of the user is needed to select the proper farm of Ho and Ha. However, the
equality part of the expression (either =, or ) always appears in the null hypothesis.

Type I and Type II Errors

There are four possible outcomes of any hypothesis test, two of which are correct and
two of which are incorrect. The incorrect ones are called type I and Type II errors.

Type I Error
In hypothesis testing sample evidence is used to test the null hypothesis Ho.
Occasionally the sample data gathered in research process lead to a decision to reject a
null hypothesis when actually it is true. A type I error is committed when a true null
hypothesis is rejected. In short, rejecting a true Ho is called Type I error. The
possibility of committing a Type I error is represented by Alpha (), or the level of
significance. Alpha is sometimes referred to as the amount of risk taken in an
experiment. Alpha represents the proportion of the area of the curve occupied by the
rejection region. The most commonly used values of  are 0.001, 0.01, 0.05 and 0.10.
The larger the area of the rejection region, the greater is the risk of committing Type I
error.

Type II Error
A Type II error is committed by failing to reject a false null hypothesis. That is to say
that, accepting a null hypothesis when it is false is called a Type II error. The
probability of committing a Type II error is represented by beta ().

Alpha () is determined before the experiment, however, Beta () is computed using
alpha, the hypothesized parameter, and various theoretical alternatives to the null
hypothesis.

(Null Hypothesis)
 57

State of Nature
Decision Ho True Ho False
Accept Ho Correct Decision Type II Error
Reject Ho Type I Error Correct Decision

There is a tradeoff between alpha and beta (Type I and Type II errors). The probability
of making one type of error can be reduced only if we are willing to increase the
probability of making the other type of error. However, this does not mean that 1;
rather it means that the smaller  is the larger will be , and the larger  in the smaller 
will be.

One – Tailed Vs Two – Tailed Tests

These are three possible null hypotheses along with their corresponding alternative
hypotheses.

(1) Ho:  (2) Ho:  (3) Ho: 

Ha:  Ha:  Ha: 

  

Leads to two – tailed test leads to a right – tailed test Leads to a left tests

Two – Tailed test

A two – tailed test of a hypothesis will reject the null hypothesis if the sample statistic is
significantly higher than or lower than a hypothesized population parameter. Thus, in
a two – tailed test there are two rejection regions. A two tailed is appropriate when the
null hypothesis is equal to some specified value (e.g. Ho:   ) and the alternative
hypothesis is different from (not equal to) some specified value (e.g. Ha:  . The
dividing point between the region where the null hypothesis is rejected and accepted is
called critical value.

One Tailed test

A one tailed test is one in which the alternative hypothesis is directional; unlike a two –
tailed test which does not specify direction.
One – tailed test can be:

 A right – tailed / upper – tailed test.

 A left – tailed / lower – tailed test.
 58

Aright – tailed test will reject the null hypothesis if the sample statistic is significantly
higher than the hypothesized population parameter.

A left – tailed test will reject the null hypothesis if the sample statistic is significantly
lower than the hypothesized population parameter.

STEPS IN HYPOTHESIS TESTING

A summary of the steps that can be applied to any hypothesis test are:

1. Determine the null and alternative hypotheses.

E.g. Ho:   
Ha:   

2. Select the test statistic that will be used to decide whether or not to reject the null
hypothesis
E.g. Z – distribution, t – distribution, F- dist, x2 – distribution

3. Select the level of significance to determine the critical values and develop the
rejection rule that indicates the values of the test statistic that will lead to the
rejection of Ho.
E.g.  = 0.05 Z025 = 1.96  Reject Ho if /Sample Z/  1.96

4. Collect the sample data, and compute the value the test statistic. A test statistic is a
random variable whose value is used to determine whether we reject the null
hypothesis.
E.g. Sample Z=2.0

5. Compare the value of the test statistic to the critical value(s) and make the decision
(either reject Ho or accept HO /do not reject).

HYPOTHESIS TEST ABOUT A POPULATION MEAN:

Population - Normal, Standard Deviation - Known
In hypothesis testing if the population in normal and standard deviation is known, we

use Z-Value to test the hypothesis; regardless of the sample size, n. It is also applicable

when n  3 regardless of the pop distribution.

Example:
 59

1. Matador-Addis Tyre Share Company claims that its tires have a mean life of 35,000
miles. A random sample of 16 of these tires is tested if the sample mean in 33,000 miles.
Assume that the population standard deviation is 3000 miles and the lives of tires are
approximately normally distributed. Test the share company’s claim using a 5% level
of significance.
Solution
1. Ho:  = 35,000 miles 2. Z – distribution, two tailed test
Ha:  ≠ 35,000 miles
3.  = 0.05 4. X = 33,000 miles
/2 = 0.025  = 3,000 miles
Z0.025 = ± 1.96 n = 16 tires
Reject Ho if /Sample Z/ > 1.96 Sample Z =?
33,000  35,000
Z 33, 000    2.67
3,000
16
5. Reject Ho because /-2.67/ > 1.96

Alternative Methods of Testing Hypotheses

Two other methods of testing hypotheses are the critical value method and the
probability (p-value) method. All three methods of testing lead to the same conclusion
with regard to rejection or failure to reject the null hypothesis. Each of these other two
methods is explained below.

Critical Value Method

One alternative method of testing hypotheses is the critical value method. In the
preceding example, the null hypothesis was rejected, because the computed Z value did
fall in the rejection region. What mean life of tyres would it take to cause the calculated
z value to fall into the rejection region? The critical value method determines the critical
mean value required for z to be in the rejection region and uses it to test the hypothesis.

x
Similar to that of the first method, the critical value method uses the formula Z 

n
x
or Z depending on the knowledge of the population standard deviation.

n
However, instead of a calculated z, a critical X value, X c , is determined. The critical
value of Z c is inserted into the formula, along with µ and σ. Thus,
 60

Xc 
Zc  .

n

Substituting values from the preceding example gives

X  35,000
 1.96  c Or
3,000
16
3,000
X c  35,000  1.96 =35,000 ± 750. Therefore, X c =35,750 or 34,250.
16

With the critical value method, most of the computational work is done ahead of time.
In this case, before the sample means are computed, it is known that a sample mean
value of less than 34,250 miles but greater 35,750 than miles must be attained in order to
reject the population mean. Because the sample mean for this problem is 33,000 miles,
we reject the null hypothesis. This method is particularly attractive in industrial settings
where standards can be set ahead of time and then quality control technicians can
gather data and compare actual measurements of products to specifications.

The probability (p-value) Method

A third way to make decisions in hypothesis testing problems is by using the
probability (p-value) method. In this method, in stead of comparing the computed
statistic (z value) with a table value to determine whether the value falls into the
rejection region, we determine a probability for the computed value and compare that
probability to the level of significance (alpha).

For example, in the mean life of Matador-Addis Tyre case, the computed value of z was
-2.67. The Z table lists the probability of a value this large or larger (this small or
smaller) occurring by chance as 0.00379 (0.5 - 0.49621). As this probability is smaller
than α/2, the null hypothesis is rejected.

In order to reject the null hypothesis with the probability method, the probability of
the computed value must be less than α for a one tailed test or less than α/2 for a two
tailed test.

2. A Teachers’ union is on strike for higher wages. The union claims that the mean salary
for teachers is at most Birr 8,400 per year. The legislator does not want to reject the
 61

union’s claim, however, unless the evidence is very strong against if. Assume that
salaries follow a normal distribution and the population standard deviation is known to
be Birr 3000. A random sample of 64 teachers is obtained, and the sample mean is Birr,
9,400. Test if the state legislator accepts the unions’ claim or not at 1% significance level.

Solution:
1. Ho:  ≤ Birr 8,400 2. Z – distribution, Right – tailed test
Ha:   Birr 8,400
3.  = 0.01
Z = Z0.01 = 2.33
Reject Ho if sample Z  + 2.33
4. n = 64
X = Birr 9,400
 = Birr 3,000
Sample Z =?
9,400  8,400
Z 9, 400   2.67
3,000
64
5. Reject Ho because + 2.67  2.33

3. A fertilizer company claims that the use of its product will result in a yield of at least 35
quintals of wheat per hectare, on average, Application at the fertilizer to a randomly
selected 36 sample hectares resulted in a yield of 34quintals per hectare. Assume the
population standard deviation is 5 quintals and yields per hectare are normally
distributed. Test the company’s claim at 1% level of significance.
Solution
1. Ho:  ≥ 35 quintals
Ha:   35 quintals
2. Z – distribution, left – tailed test
3.  = 0.01
Z= Z0.01 = 2.33
 Reject Ho if sample Z  -2.33
4. X = 34 quintals
n = 36
 = 5 quintals
Sample Z =?
34  35
Z 34   1.20
5
36
 62

4. Do not reject H0, because -1.20 > -2.33

4. A survey of college graduates showed that the average yearly cash income for these
graduates in at least Birr 12,000. In Addis where you live this average does not seem
possible, so you decide to test this claim. You randomly select 48 graduates who are
marking. The sample average income for these working graduates is Birr 11,400 with a
standard deviation of Birr 2,280. Is there enough evidence from this sample data to
reject the national claim for your area as being too high? Use  = 0.10.
Solution
1. Ho :   Birr 12,000
Ha:   Birr 12,000
2. Z – distribution, Left tailed test
3.  = 0.1
Z = Z0.1 = -1.28
Reject Ho if sample Z < -1.28
4. X = Birr 11,400
S = Birr 2,280
n = 48
Sample Z =?
11,400  12,000
Z 11, 400   1.82
2,280
48
= -1.82
5. Reject Ho because –1.82 < -1.28

HYPOTHESIS TESTING about: Population Normal,  Unknown, n

Small

As explained in chapter 4, Interval Estimation, use of the t-distribution is appropriate

when  is unknown, the sample size is small (n30), and population is normally
distributed. When we use the t-distribution, it replaces Z as test statistic.
X  S
t SX 
SX n
Example:
 63

1. A contractor assumes that construction workers are idle for 75 minutes or less per day.
A random sample of 25 construction workers was taken and the mean idle time was
found to be 84 minutes per day with a sample standard deviation of 20 minutes.
Assume that the population is approximately normally distributed, use a 5% level of
significance to test the contractor’s assumption.

Solution
1. Ho:  ≤ 75 minutes
Ha:   75 minutes
2. t – distribution, Right – tailed test
3.  = 0.05
n = 25
 = n – 1 = 25 – 1 = 24
t, = t0.05,24 = 1.711
 Reject Ho if sample t  1.711
4. n = 25
X = 84 minutes
S = 20 minutes
Sample t =?
84  75
t 84   2.25
20
25
5. Reject Ho; because + 2.25  1.711. Workers are idle for more than 75 minutes per day.
2. A director of a secretarial school claims that its graduates can type at least 50 words per
minute on average. Suppose you want to hire some of these graduates if the director’s
claim is true; and you test the typing speed of 18 of the graduates and obtain a mean of
40 wards per minute with a sample variance of 720. Assuming the typing speed for the
graduates of the secretarial school is normally distributed, test the director’s claim and
decide whether to hire the graduates or not, using a 5% level of significance.
Solution
1. Ho:   50 words
Ha:  50 words
2. t – distribution, Left – tailed test
3.  = 0.05
n = 18
=n–1
= 18 – 1 = 17
t, = t0.05, 17 = 1.74
 Reject Ho if sample t  -1.74
4. X = 40 words
 64

n = 18
s2 = 720
Sample t =?
40  50
t 40   1.58 5. Do not reject Ho because –1.58 > -1.74
720
18
HYPOTHESIS TESTING ABOUT A POPULATION PROPORTION (P)

A proportion is a value between 0 and 1 that expresses the part of the whole that
possesses a given characteristic.

The formula (methods) for proportions based on the central limit theorem make
possible the testing of hypotheses about the population proportion in a manner similar
to that of the formula used to test sample means. Similar to that of hypothesis testing
about a population mean, hypothesis testing about a population proportion has three
terms.

1. Ho: P = y 2. Ho: P  y 3. Ho: P  y

Ha: P  y Ha: Py Ha: P y

The first form is a two – tailed test, where as the second and third forms are one – tailed
tests. The specific form used depends up on the application of interest.

Hypothesis testing about a population proportion is based on the difference between

the sample proportion p and the hypothesized value P.

The central Limit Theorem applied to the sample proportions states that p values are
 pq 
normally distributed, with a mean of P and a standard deviation of  p   , when np

 n 
and nq are greater than or equal to 5. If np and nq are greater than or equal to 5, a Z -
test is used to test hypothesis about P.

p p
Z
p
Example:
 65

1. A magazine claims that 25% of its readers are college students. A random sample of
200 readers is taken. It is found that 42 of these readers are college students. Use a 10%
level of significance and test the magazine’s claim.
Solution
1. Ho: P = 0.25
Ha: P  0.25
2. Z – distribution; two tailed test
3.  = 0.1 /2 = 0.05
Z/2 = Z0.05 = 1.64
Reject Ho if /sample Z/  1.64
4. n = 200
x = 42
p = 0.21
Sample Z =?
0.21  0.25
Z 0.21   1.31
0.25 * 0.75
200
5. Do not reject Ho because / -1.31/ < 1.64

2. An Economist states that more than 35% of Addis’s labor force in unemployed. You
don’t know if the economist’s estimate is too high or too low. Thus, you want to test the
economist’s claim using a 5% level of significance. You obtain a random sample of 400
people in the labor force, of whom 128 are unemployed. Would you reject the
economist’s claim?
Solution
1. Ho: P 0.35
Ha: P < 0.35
2. Z – distribution , Left tailed test
3.  = 0.05
Z 0.05 = 1.64
Reject Ho if sample Z < -1.64
4. n = 400
x = 128
p = 0.32
Sample Z =?
0.32  0.35
Z 0.32   1.26
0.35 * 0.65
400
 66

5. Do not reject Ho because -1.31 > -1.64

3. A survey of the morning beverage market has shown that the primary breakfast
beverage for 60% of Ethiopian town and city dwellers is tea. Ethiopian coffee and Tea
Authority believes that the figure is higher for Addis. To test this idea, one of the
employees of Ethiopian coffee and Tea Authority contacts a random sample of 500
residents in Addis and asks which primary beverage they consumed for breakfast that
day. Suppose 325 replied that tea was the primary beverage. Using a 0.01 level of
significance, test the idea that the tea figure is higher for Addis.
Solution
1. Ho: P = 0.60
Ha: P 0.60
2. Z – distribution, – two tailed test
3.  = 0.01
Z0.01 = 2.33
 Reject Ho if sample Z  2.33
4. n = 500
X = 325
P = 0.45
Sample Z =?
0.65  0.60
Z 0.65   2.28
0.65 * 0.4
500
5. Do not reject Ho because 2.28 < 2.33

HYPOTHESIS TESTING INVOLVING TWO POPULATIONS

Hypothesis Testing About the Difference between Two Means

According to the central limit theorem, the difference in two sample means in normally
distributed for large sample sizes (n1, n2  30) and if the parent populations are normally
distributed regardless of the sample size the difference in two sample means is
normally distributed.

Large sample case

As is true is hypothesis testing about a population mean or population proportion,
hypothesis testing about the difference between two means has three forms. That is, the
 67

null hypothesis can take three forms along with the corresponding alternative
hypothesis.

1. Ho: 1-2 = 0 2. Ho: 1-2  0 3. Ho: 1-2  0

Ha: 1-2  0 Ha: 1-2  0 Ha: 1-2  0
The first form leads to a two-tailed test, while the later two lead to a one-tailed test.
Based on the central limit theorem, the mean of the sampling distribution sample means
in the difference between two population means; i.e.,  X 1  X 2  = 1-2; and the
 12  22
standard error is  X 1  X   .
2
n1 n2

For large sample sizes the sampling distribution of the difference between two sample
means is normally distributed with a Z – test statistic.

Z 1

X  X 2   1   2  
 12  22

n1 n2
And whenever n1 and n2  30, we can use S 12 and S 22 as estimates of  12 and  22 to
compute Z if  12 and  22 are unknown, and Z will be computed as

Z
X 1 
 X 2   1   2 
S 12 S 22

n1 n 2
In most of the hypothesis tests about two means, the hypothesized difference is zero.
Example:

1. Is there any difference between the average salary of a legal secretary and a medical
secretary? In an effort to answer that question a researcher takes a random sample of 33
legal secretaries across a region, resulting in a sample average annual salary of Birr
20,000 with a standard deviation of Birr 1,550. The researcher then takes a random
sample of 35 medical secretaries a class the region, which yields an average annual
salary of Birr 18,500 with a standard deviation of Birr 2,100. Use =0.01 to test this
question
Solution
1. Ho: 1-2 = 0
Ha: 1-2  0
2. Z – distribution , two tailed test
3.  = 0.01, /2 = 0.005
Z/ = Z0.005 = 2.57
 68

 Reject Ho if /sample Z/  2.57

4. Legal Secretary
X 1 = Birr 20,000
n1 = 33
S1 = Birr 1,550
Sample Z =?
Medical Secretary
X 2 = Birr 18,500
n2 = 35
S2 = Birr 2,100

Z 1
 
X  X 2   1   2 
S 12 S 22

n1 n 2
20,000  18,500  0  3.36
Z X
1X2   Z 20,00018,500
1,550 2 2,100 2

33 35
5. Reject Ho; because /sample Z/ = 3.36  2.57. There is a difference in the average
annual salary of legal and medical secretary.

2. A firm is studying the delivery times for two raw material suppliers. The firm in
basically satisfied with supplier A and is prepared to stay with this supplier provided
that the mean delivery time is the same or less than that of supplier B. However if a
firm finds that the mean delivery time from supplier B is less than that of supplier A, it
will begin making raw material purchases from supplier B.
a. What are the null and alternative hypotheses for this situation?
b. Assume that independent samples show the following delivery time x is for the two
suppliers.

Supplier A Supplier B
n1 = 50 n2 = 30
X 1 = 14 days X 2 = 12.5 days
S1 = 3 days S2 = 2 days

Using  = 0.05, what is your conclusion for the hypotheses from part (a)? What action
do you recommend in terms of supplier selection?
Solution
1. Ho: 1-2  0
Ha: 1-2  0
2. Z – distribution , Right - tailed test
3.  = 0.05
Z = Z0.05 = 1.64
Reject Ho if sample Z  1.64
 69

4. Supplier A Supplier B
n1 = 50 n2 = 30
X 1 = 14 days X 2 = 12.5 days
S1 = 3 days S2 = 2 days
Sample Z =?
14  12.5  0  2.68
Z X
1X2   Z 1412.5 
32 2 2

50 30

5. Reject Ho; because 2.68  1.64

This shows that the mean delivery time for supplier A is greater than supplier B. So the
firm should shift to supplier B.

3. In a wage discrimination case involving male and female employees, it is assumed that
male employees have a mean salary less than or equal to that of female employees. To
justify this, independent random samples of male and female employees were taken
and the following result obtained.
Male Employees Female Employees
n1 = 100 n2 = 100
X 1 = Birr 20,600 X 2 = Birr 19,700
S1 = 3,000 S2 = Birr 2,500

Test the hypothesis with  = 0.025. Does wage discrimination appear to exist in this
case?

Solution
1. Ho: 1-2  0
Ha: 1-2  0
2. Z – distribution , Right - tailed test
3.  = 0.025
Z = Z0.025 = 1.96
Reject Ho if sample Z  1.96
4. Male employees Female employees
n1 = 100 n2 = 100
X 1 = Birr 20,600 X 2 = Birr 19,700
S1 = Birr 3,000 S2 = Birr 2,500
Sample Z =?
 70

20,600  19,700  0  2.30

Z X
1X2   Z 20,60019,700
3,000 2 2,500 2

100 100
5. Reject Ho; because 2.30  1.96
Wage discrimination appears to exist
Small Sample Case
In testing of hypothesis for the difference between two means, given that the sample
sizes taken are small (n1, n2  30), we can use either Z or t – distribution as a test statistic
assuming that the parent populations are normally distributed.

1. Population normal, and 1 and 2 known – Z distribution is used

2. Population normal, and 1 , and 2 unknown - t distribution is used

t – Distribution: used when population normal 1, and 2 unknown, and n1, and/or n2 
30. The unknown population standard deviations are approximated by sample
standard deviations as:
S 12 n1  1  S 22 n 2  1 1 1
X  SX   ; Where df = n1 + n2 - 2
1X2 1X2
n1  n 2  2 n1 n 2
This approximation is based on the assumption that the two population standard
deviations are equal.

Once S X  X is known t is computed as: t 

X 1  X 2   1   2   
S 12 n1  1  S 22 n 2  1 1
1 2
1

n1  n 2  2 n1 n 2
Or
S p2  PooledVariance
X 
 X 2   1   2 
t S12 n1  1  S 22 n 2  2
1
; Where
S 2
S 2

p

p
n1  n 2  2
n1 n2

If 1 and 2 are not equal, the sample standard error of the difference between two

means and the degrees of freedom are calculated as:

2
 S 12 S 22 
  
S 12 S 22  n1 n 2 
SX   ; V  df 
1X2 2 2
n1 n 2  S 12   S 22 
   
n  n 
 1   2 

n1  1 n2 1
 71

Example:
1. A marketing research firm wishes to know if the mean number of his of TV viewing per

week is the same for teenage boys and teenage girls using a 5% level of significance.

The unknown population variances are assumed to be equal. The following data were

obtained is an attempt to test the equality of 1 and 2.

Teenage Boys Teenage Girls

n1 = 20 n2 = 12
X 1 = 24.5 hrs X 2 = 28.7 hrs
S21 = 64 hrs S22 = 71 hrs
Test the above hypothesis.
Solution
1. Ho: 1  2 or 1-2 = 0
Ha: 1  2 or 1-2  0
2. t – distribution , Two - tailed test
3.  = 0.05 v = n1 + n2 - 2
/2 = 0.025 = 12 + 20 –2 = 30
t/2, V = t0.25, 30 = 2.042
Reject Ho if /sample t/  2.042
4. Teenage Boys Teenage Girls
n1 = 20 n2 = 12
X 1 = 24.5 hrs X 2 = 28.7 hrs
S 1 = 64 hrs
2
S22 = 71 hrs
Sample t =?

t X X  1

X  X 2   1   2 
= t 24.5 28.7 
24.5  28.7   0  1.41
6420  1  7112  1 1
1 2 2 2
Sp Sp 1
 
n1 n 2 20  12  2 20 12

5. Do not Reject Ho; There is no significant difference

2. A time-and-motion study is conducted to test whether the mean length of time required
to perform a certain task is lesser for employees on the day shift than for employees on
the night shift. The data are as follows.
Day shift Night Shift
n1 = 10 n2 = 8
X 1 = 20 hrs X 2 = 29 hrs
S1 = 64 hrs
2
S22 = 50 hrs
 72

Use a 1% level of significance to test the hypothesis. Assume the populations are
approximately normal, the population variances are equal, and the samples are
independent.

Solution
1. Ho: 1 - 2 < 0
Ha: 1 - 2 > 0
2. t – distribution , Right - tailed test
3.  = 0.01 v = 10 +8-2 = 16
4. t, v = t0.01,16 = 2.583
Reject Ho if /sample t/ > 2.583
Sample t =?
t
20  29  0  2.49
6410  1  508  1 1 1

10  8  2 10 8

5. Do not reject Ho because /–2.49/ <2.58

The mean length of time required to perform a certain task for day-shift employees is
greater than or equal to for night shift employees.

Hypothesis testing about the Difference between two population

proportions (P1-P2)

Frequently we are interested in testing the hypothesis that the proportions of

individuals who possess a certain characteristic in population 1 is the same as that in
population 2. E.g. The proportion of defectives in lot A and Lot B

To test hypothesis about the difference between two population proportions, we obtain
independent random sample of n1 items from the first population and n2 items from the
X X
second; and calculate P1  1 and P2  2 . Once we obtain P1 and P2 , we use a test
n1 n2
based on the standard normal distribution.

In testing hypothesis about the difference in two population proportions, particular

values of the population proportions are not usually known or assumed; unlike
hypothesis testing about population proportion. Rather, the hypotheses are about the
difference in the two population proportions, (P1-P2). Hypothesis testing about the
difference between two population proportions has three forms:
 73

1. Ho: P1-P2 = 0 2. Ho: P1-P2  0 3. Ho: P1-P2  0

Ha: P1-P2  0 Ha: P1-P2  0 Ha: P1-P2  0

The first form leads to a two-tailed test while the later two lead to a one-tailed test.

The central Limit Theorem applied to the difference between two population
proportions states that P1  P2 values are normally distributed, with a mean of P1-P2 a

standard deviation of
P1 (1  P1 ) P2 (1  P2 )
 , and Z  1
 
P  P2  P1  P2 
.
n1 n2 P1 q1 P2 q 2

n1 n2
However, since the standard error is unknown, it has to be estimated from the sample
data. While we may be tempted to use P1 and P2 as we did with the interval estimation
procedure, in hypothesis testing we often adjust it to a slightly different form. For the
special case where the hypotheses involve no difference between the population
proportions (i.e. either Ho: P1-P2 = 0 Ho: P1-P2  0, or Ho: P1-P2  0) is modified to reflect
the fact that when we assume Ho to be true at the equality, we are assuming P1 = P2.
When this occurs, we combine or pool the two sample proportions to provide one
n p  n2 p2 X 1  X 2
estimate. This pooled estimator, denoted by P , is as P  1 1  , and the
n1  n 2 n1  n 2
standard deviation  p p is estimated by S p  p , which is calculated as
1 2 1 2

1 1  1 1  PQ PQ
S p  p  P(1  P)   = PQ   =  and Z is calculated as
 n1 n 2   n1 n 2  n1 n2
1 2

Z
P  P  P  P  . And when P =P
1 2 1 2
1 2, Z becomes Z  .
P  P 
1 2

1 1  1 1 
PQ   PQ  
 n1 n 2   n1 n 2 
Example:
1. In a sample of 400 products produced by machine I, 200 were defective, and in a sample
of 400 products produced by machine II, 170 were defective. Using  = 0.05, test the
hypothesis that the rate of defect is the same for both machine 1 and machine 2.
Solution
1. Ho: P1-P2 = 0
Ha: P1-P2  0
2. Z – distribution, two – tailed test
3.  = 0.05,
/2 = 0.025
Z/2 = Z0.025 = 1.96
 74

Reject Ho if /Sample Z/  1.96

4. Machine I Machine II
n1 = 400 n2 = 400
x1 = 200 x2 = 170
X  X 2 170  200
P 1   0.4625
n1  n 2 400  400

Sample Z 
P  P  P  P  =
1 2 1 2 0.5  0.425  0  2.13
1 1   1 1 
PQ   0.4625 * 0.5375  
 n1 n 2   400 400 
5. Reject Ho

2. To test the effectiveness of the approach and layout of two direct mail brochures, a
marketing manager of SELAM Inc mailed out 150 copies of each brochure and recorded
the number of responses penetrated by each. There were 30 responses generated by the
first brochure and 15 generated by the second. Can the marketing manager conclude
that the first brochure is more effective? Use = 0.05.
Solution
1. Ho: P1-P2 > 0
Ha: P1-P2 0
2. Z – distribution, Left – tailed test
3.  = 0.05,
Z = Z0.05 = 1.64
Reject Ho if Sample Z < -1.64
4. Brochure I Brochure II
n1 = 150 n2= 150
x1 = 30 x2 = 15
X  X2 30  15
P 1   0.15
n1  n 2 150  150

Sample Z 
P  P   P  P  =
1 2 1 2 0.20  0.10  0  2.43
1 1   1 1 
PQ   0.15 * 0.85  
 n1 n 2   150 150 

5. Do not Reject Ho. YES, the marketing manager can conclude that the first brochure
is more effective than the second.

3. Random samples of Video cassette Recorders assembled in ALEM assembly plant

included 50 assembled during the first shift and 50 assembled during the second shift.
 75

Of the Video Cassette Recorders assembled during the first shift 10 were defective; and

20 were defective from the second shift. From the data, would the production foreman

reject the hypothesis that the proportion of defectives assembled by the first shift is

greater than or equal to that for the second shift? Use a 0.05 level of significance.

Solution
1. Ho: P1-P2  0
Ha: P1-P2  0
2. Z distribution, Left-tailed test
3.  = 0.05,
Z = Z0.05 = 1.64
Reject Ho if Sample Z  1.64
4. Shift 1 Shift 2
n1 = 50 n2 = 50
x1 = 10 x2 = 20
X  X 2 10  20
P 1   0.30
n1  n 2 50  50

Sample Z 
P  P   P  P  =
1 2 1 2 0.20  0.40  0  2.18
1 1   1 1 
PQ   0.30 * 0.70  
 n1 n 2   50 50 
5. Reject Ho.
The proportion of defectives assembled by the second shift is greater than that for the
first shift.

Solving for Type II Errors (Computation of Beta, β)

Determining the probability of committing a Type II error is more complex than finding
the probability of committing a Type I error. The probability of committing a Type I
error either is given in a problem or is stated by the researcher before proceeding with
the study. A Type II error, β, varies with the possible values of the alternative
parameter.

For example, suppose that a researcher is conducting a statistical test on the following
hypothesis:
Ho:   12 oz
Ha:   12 oz
 76

A Type II error can be committed only when the researcher fails to reject a false null
hypothesis. In these hypotheses, if the null hypothesis,   12 oz is false, what is the true
value for the population mean? Is the mean really 11.99 oz or 11.90 oz, or 11.50 oz or 10
oz? For each of possible values of the population mean, the researcher can compute the
probability of committing a Type II error. Often, when the null hypothesis is false, the
value of the alternative mean is unknown, so the researcher will compute the
probability of committing Type II errors for several possible values. How can the
probability of committing a Type II error be computed for a specific alternative value of
mean?

Suppose that, in testing the hypotheses above, a sample of 60 cans of beverage yields a
sample mean of 11.985 0z, with a standard deviation of 0.10 oz. For α = 0.05 and a one-
tailed test, the table Z value is -1.64. The calculated Z value is

11.985  12.00
Z  1.16
0.10
60
From this calculated value of z, the researcher determines not to reject the null
hypothesis. By not rejecting the null hypothesis, the researcher either made a correct
decision or committed a Type II error. What is the probability of committing a Type II
error in this problem if the population mean actually is 11.99?

The first step in determining the probability of a Type II error is to calculate the critical
value for the mean, X . This value is used as a cutoff point for the acceptance region in
testing the null hypothesis. For any sample mean obtained that is less than X (or greater
for an upper tail rejection region), the null hypothesis is rejected. Any sample mean
greater than X (or less for an upper tail rejection region) causes the researcher to accept
the null hypothesis. Solving for the critical value of the mean gives
X 
Z
S
n
X  12.00
 1.64  ;
0.10
60

X = 11.979
From the above computation, we can learn that the null hypothesis will be rejected for a
sample mean vale of less than 11.979 oz. Assume that the alternative hypothesis is a =
11.99 oz. How often will the researcher accept the  = 12 as true, when, in reality,  =
11.99 is true? If the null hypothesis is false, the null hypothesis will be incorrectly
 77

accepted whenever  falls in the acceptance region, X  11.979 oz. If  actually equals
11.99 oz, what is the probability of failing to reject  = 12 oz when 11.979 oz is the critical
value? The researcher calculates this probability by extending the critical value ( X =
11.979 oz) and finding the area to the right of 11.979.
X   11.979  11.99
Z   0.85
S 0.10
n 60

This value of Z yields an area of 0.3023. The probability of committing a Type II error is
all the area to the right of X = 11.979, or 0.3023+0.5000 = 0.8023. Hence, there is an
80.23% chance of committing a Type II error if the alternative mean is 11.990z.

With two-tailed tests, both tails of the distribution contain rejection regions. If the null
hypothesis is false, obtaining a calculated statistic falling in the tails results in the
correct decision: to reject the null hypothesis. In this case, the probability of committing
a Type II error exists only for the area between the two critical values (the acceptance
region). However, the right critical value is so far away from the alternative mean that
the area between the right critical value and the mean essentially is 0.5000. Had there
been any area past the upper critical value of Pc (0.46), it would have been subtracted
from 0.5000, slightly reducing the value of 0.7454.

Example:
1. Suppose that you are conducting a two-tailed hypothesis test of proportions. The null
hypothesis is that the population proportion is 0.40. The alternative hypothesis is that
the population proportion is not 0.40. A random sample of 250 produces a sample
proportion of 0.44. Using alpha of 0.05 and assuming that the alternative population
proportions really is 0.36, what is the probability of committing a Type II error?

Solution:
For an alpha value of 0.05, the table Z value for α/2 is 1.96. Using 1.96, solve for the
critical value of the proportion.
pP
Z
Pq
n
p  0.40
 1.96 
0 . 4 * 0 .6
250
P  0.40  0.06
 78

The critical values are 0.34 on the lower end and 0.46 on the upper end. The alternative
population proportion is 0.36.

Solving the area between p = 0.34 and p = 0.36 yields

0.34  0.36
Z  0.66
0.36 * 0.64
250

The area associated with Z= 0.66 is 0.2454. The probability of committing a Type II error
is 0.5000 + 0.2454 = 0.7454.

2. Suppose that the null hypothesis is that the population mean is greater than or equal to
100. Suppose further that a random sample of 48 items is taken and the sample standard
deviation is 14. For each of the following alpha values, compute the probability of
committing a Type II error if the population mean actually is 99.
a. α = 0.01
b. α = 0.05
c. α = 0.10
d. Based on the answers to (a), (b), and (c), what happens to the value of β as α gets
larger?

Solution:
a. For an alpha value of 0.01, the table Z value for α is -2.33. Using -2.33, solve for the
critical value of the mean.
X 
Z c
s
n
X  100
 2.33 
14
48
X c  95.292

The critical value is 95.292 on the lower end. The alternative population mean is 99.

Solving the area between X c = 95.292 and µ = 99 yields

95.292  99
Z  1.83
14
48
 79

The area associated with Z= -1.83 is 0.46638. The probability of committing a Type II
error is 0.5000 + 0.46638 = 0.96638.

b. For an alpha value of 0.05, the table Z value for α is -1.64. Using -1.64, solve for the
critical value of the mean.
X 
Z c
s
n
X  100
 1.64 
14
48
X c  96.686

The critical value is 96.686 on the lower end. The alternative population mean is 99.

Solving the area between X c = 96.686 and µ = 99 yields

96.686  99
Z  1.14
14
48
The area associated with Z= -1.14 is 0.39251. The probability of committing a Type II
error is 0.5000 + 0.39251 = 0.89251.

c. For an alpha value of 0.10, the table Z value for α is -1.28. Using -1.28, solve for the
critical value of the mean.
X 
Z c
s
n
X  100
 1.28 
14
48
X c  97.413

The critical value is 97.413 on the lower end. The alternative population mean is 99.

Solving the area between X c = 97.413 and µ = 99 yields

97.413  99
Z  0.79
14
48
The area associated with Z= -0.79 is 0.28524. The probability of committing a Type II
error is 0.5000 + 0.28524 = 0.78524.
 80

d. Based on the answers to (a), (b), and (c), the value of β gets smaller as α gets larger.
3. For exercise 2 above, use α = 0.05 a solve the probability of committing a Type II error
for the following possible true alternative means.
a.  a = 98.5
b.  a = 98
c.  a = 97
d.  a = 96
e. What happens to the probability of committing a Type II error as the
alternative value of the mean gets farther from the null hypothesized value of
100?
Solution:
a. For an alpha value of 0.05, the table Z value for α is -1.64. Using -1.64, solve for the
critical value of the mean.
X 
Z c
s
n
X  100
 1.64 
14
48
X c  96.686
The critical value is 96.686 on the lower end. The alternative population mean is 98.5.
Solving the area between X c = 96.686 and µ = 98.5 yields
96.686  98.5
Z  90
14
48
The area associated with Z= -0.90 is 0.31594. The probability of committing a Type II
error is 0.5000 + 0.31594 = 0.81594.
b. For an alpha value of 0.05, the table Z value for α is -1.64. Using -1.64, solve for the
critical value of the mean.
X 
Z c
s
n
X  100
 1.64 
14
48
X c  96.686
The critical value is 96.686 on the lower end. The alternative population mean is 98.
Solving the area between X c = 96.686 and µ = 98 yields
 81

96.686  98
Z  0.65
14
48
The area associated with Z= -0.65 is 0.24215. The probability of committing a Type II
error is 0.5000 + 0.24215 = 0.74215.

c. For an alpha value of 0.05, the table Z value for α is -1.64. Using -1.64, solve for the
critical value of the mean.
X 
Z c
s
n
X  100
 1.64 
14
48
X c  96.686
The critical value is 96.686 on the lower end. The alternative population mean is 97.
Solving the area between X c = 96.686 and µ = 97 yields
96.686  97
Z  0.16
14
48
The area associated with Z= -0.16 is 0.06356. The probability of committing a Type II
error is 0.5000 + 0.06356 = 0.56356.
d. For an alpha value of 0.05, the table Z value for α is -1.64. Using -1.64, solve for the
critical value of the mean.
X 
Z c
s
n
X  100
 1.64 
14
48
X c  96.686
The critical value is 96.686 on the lower end. The alternative population mean is 96.
Solving the area between X c = 96.686 and µ = 96 yields
96.686  96
Z  0.34
14
48
The area associated with Z= +0.34 is 0.13307. The probability of committing a Type II
error is 0.5000 - 0.13307 = 0.36693.
 82

e. The probability of committing a Type II error decreases as the alternative value of

the mean gets farther from the null hypothesized value (of 100).
 83

CHAPTER FOUR
CHI-SQUARE DISTRIBUTIONS

A Chi-square (x2) distribution is a continuous distribution ordinarily derived as the

sampling distribution of a sum of squares of independent standard normal variables.
Characteristics of the square distributions
1. It is a continuous distribution
2. The X2 distribution has a single parameter; the degree of freedom, ν
3. The mean of the chi-square distribution is ν
4. The variance of the chi-square distribution is 2ν. Thus, the mean and Variance
depend on the degree of freedom.
5. It is based on a comparison of the sample of observed data (results) with the
expected results under the assumption that the null hypothesis is true.
6. It is a skewed distribution and only non-negative values of the variable X2 are
possible. The skewness decreases as ν increases; and when V increases without limit
it approaches a normal distribution. It extends indefinitely in the positive direction
7. The area under the curve is 1.0

Having the above characteristics, X2 dist has the following areas of application:
1. Testing for the equality of several proportions
2. Test for independence between two variables
3. Goodness of fit tests (Binomial, Normal, and Poisson )

TEST FOR THE INDEPENDENCE BETWEEN TWO VARIABLES

A X2 test of independence is used to analyze the frequencies of two variables with

multiple categories to determine whether the two variables are independent. That is,
the Chi-square distribution involves using sample data to test for the independence of
two variables. The sample data are given in to a two way table called a contingency
table. Because the X2 test of independence uses a contingency table, the test is
sometimes referred to as CONTINGENCY ANALYSIS (Contingency table test). The X 2
test is used to analyze, for example, the following cases:
 Whether employee absenteeism is independent of job classification
 Whether beer preference is independent of sex (gender)

 Whether favorite sport is independent of nationality.

 Whether type of financial investment is independent of geographic region.

The steps and procedures are similar with hypothesis testing.
 84

Example:
1. A company planning a TV advertising campaign wants to determine which TV shows
its target audience watches and thereby to know whether the choice of TV program an
individual watches is independent of the individuals income. The table supporting this
is shown below. Use a 5% level of significance and the null hypothesis.

Income Type of Show

Basketball Movie News Total
Low 143 70 37 250

Medium 90 67 43 200
High 17 13 20 50

Total 250 150 100 500

Solution
1. Ho: Choice of TV program an individual watches is independent of the individuals
income
Ha: Income and Choice of TV program are not independent
2. Decision rule
 = 0.05
ν = (R-1) (C-1)8*
= (3-1) (3-1)
=4
X , ν = X20.05, 4 = 9.49
2

Reject Ho if sample X2 is greater than 9.49

3. Compute the test statistic
In computing the test statistic our first task is to estimate the expected frequencies (eij
= rioj/n); where
ri = Observed freq total for row i.
Cj = observed freq total for column j
n = sample size

e11 = 250x250/500 = 125 e21 = 200 x 250/500 = 100 e31 = 50 x 250/500 =25

8
For the RxC contingency table, the degrees of freedom are calculated as (R-1) (C-1). The degrees of
freedom refers to the number of expected frequencies that can be chosen freely provided the row and
column totals of expected frequencies are identical to the row and column totals of the observed
frequency table.
 85

e12= 250x150/500 = 75 e22 = 200 x 150/500=60 e32 = 50 x 150/500=15

e13 = 250x100/500 =50 e23 = 200x100/500 =40 e33 = 50x100/500 =10

A test of the null hypothesis that variables are independent of one another is based on
the magnitudes of the differences between the observed frequencies and the expected
frequencies. Large differences between oij and eij provide evidence that the null
hypothesis is false. The test is based on the following Chi-square test statistic.
Oij  eij2  f  f e 2
 
2
Or  2   o
e ij fe
Where:
Oij (fo) = observed frequency for contingency table category in row i and column j.
Eij (fe) = expected frequency for contingency table in row i and column j.

2 
143  125 2  70  75 2  37  50 2  90  100 2  67  60 2  17  25 2  13  15 2
125 75 50 100 60 25 15


20  10  2

43  40  2
 21 .174
10 40

4. Reject the null hypothesis that choice of TV program is independent from income
level.

2. A human resource manager at EAGLE Inc. was interested in knowing whether the
voluntary absence behavior of the firm’s employees was independent of marital status.
The employee files contained data on marital status and on voluntary absenteeism
behavior for a sample of 500 employees is shown below.

Marital Status
Absence Married Divorced Widowed Single Total
behavior
Often absent 36 16 14 34 100
Seldom absent 64 34 20 82 200
Never absent 50 50 16 84 200
Total 150 100 50 200 500

Test the hypothesis that absence behavior is independent of marital status at a

significance level of 1%.

Solution
1. Ho: Voluntary absence behavior is independent of marital status
 86

Ha: Voluntary absence behavior and marital status are dependent

2.  = 0.01
V = (R-1) (C-1)
= (3-1) (4-1) = 6
X ,ν= X2 0.01,6 = 16.81
2

Reject Ho if sample X2 > 16.81

3. Sample X2
Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe) fe
36 30 36 1.200
64 60 16 0.267
50 60 100 1.667
16 20 16 0.800
34 40 36 0.900
50 40 100 2.500
14 10 16 1.600
20 20 0 0.000
16 20 16 0.800
34 40 36 0.900
82 80 4 0.050
84 80 16 0.200
 f o  f e 2 10.883
 fe

4. Do not reject Ho; because 10.883 is less than 16.81.

Voluntary absence and marital status are independent.

3. The personnel administrator of XYZ Company provided the following data as an

example of selection among 40 male and 40 female applicants for 12 open positions.

Applicant Status
Selected Not selected Total
Male 7 33 40
Female 5 35 40
Total 12 68 80
 87

a. The X2 test of independence was suggested as a way of determining if the decision to

hire 7 malls and females should be interpreted as having a selection bias in favor of
males. Conduct the test of independence using = 0.10. What is your conclusion?
b. Using the same test, would the decision to hire 8 malls and 4 females suggest
concern for a selection bias?
c. How many males could be hired for the 12 open positions before the procedure
would concern for a selection bias?

Solution
a.
1. Ho: There is no selection bias in favor of males. (Selection status and gender of the
applicant are independent).
Ha: There is selection bias in favor of males. (Selection status and gender of the
applicant are not independent).
2.  = 0.1
V = (R-1) (C-1)
= (2-1) (2-1) = 1
X2 ,ν= X2 0.1,1 = 2.71
Reject Ho if sample X2 > 2.71
3. Sample X2
Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe) fe
7 6 1 0.1667
33 34 1 0.0294
5 6 1 0.1667
35 34 1 0.0294
 f o  f e 2 0.3922
 fe

4. Do not reject Ho; because 0.392 is less than 2.71.

There is no selection bias in favor of male applicants.
b.
1. Ho: There is no selection bias in favor of males. (Selection status and gender of the
applicant are independent).
Ha: There is selection bias in favor of males. (Selection status and gender of the
applicant are not independent).
2.  = 0.1
V = (R-1) (C-1)
= (2-1) (2-1) = 1
 88

X2 ,ν= X2 0.1,1 = 2.71

Reject Ho if sample X2 > 2.71
3. Sample X2
Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe) fe
8 6 4 0.6667
32 34 4 0.1176
4 6 4 0.6667
36 34 4 0.1176
 f o  f e 2 1.5686
 fe

4. Do not reject Ho; because 1.569 is less than 2.71.

There is no selection bias in favor of male applicants.

c. There is no shortcut method to answer this question. Therefore, lets try by increasing the
number of male applicants who are accepted and decreasing the number of female
applicants who are females.
1. Ho: There is no selection bias in favor of males. (Selection status and gender of the
applicant are independent).
Ha: There is selection bias in favor of males. (Selection status and gender of the
applicant are not independent).
2.  = 0.1
V = (R-1) (C-1)
= (2-1) (2-1) = 1
X ,ν= X2 0.1,1 = 2.71
2

Reject Ho if sample X2 > 2.71

3. Sample X2
Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe) fe
9 6 9 1.5000
31 34 9 0.2647
3 6 9 1.5000
37 34 9 0.2647
 f o  f e 2 3.5294
 fe

4. Reject Ho; because 3.5294 is less than 2.71.

 89

Therefore, 8 male and 4 female applicants must be hired for the 12 open positions so as
to avoid selection bias in favor of males.
The Chi-square test for independence is useful in helping to determine whether a
relationship exists between two variables, but it does not enable us to estimate or
predict the values of one variable based on the value of the other. If it is determined
that a dependence does exist between two quantitative variables, then the techniques of
regression analysis are useful in helping to find a mathematical formula that expresses
the nature of mathematical relationship.

Small expected frequencies can lead to inordinately large chi-square values with the chi-
square test of independence. Hence contingency tables should not be used with
expected cell values of less than 5. One way to avoid small expected values is to
combine columns or rows whenever possible and whenever doing so makes sense.

TESTING FOR THE EQUALITY OF SEVERAL PROPORTIONS

Testing for the equality of several proportions emphasizes on whether several

proportions are equal or not; and hence the null hypothesis takes the following form:

Ho: P1 = P10; P2 = P2O; P3 = P30; --- Pk = PkO; and the alternative hypothesis takes the
following form:

Ha: The population proportions are not equal to the hypothesized values

The degree of freedom is determined as V= K-1; where K refers to the number of

proportions and all expected cell values must be greater than or equal to 5.

Example:
1. In the business credit institution industry the accounts receivable for companies are
classified as being “current,” “moderately late,” “very late,” and “uncollectible.”
Industry figure show that the ratio of these four classes is 9: 3: 3: 1. ENDURANCE firm
has 800 accounts receivable, with 439, 168, 133, and 60 falling in each class. Are these
proportions in agreement with the industry ratio? Let =0.05.
Solution
1. Ho: P1 = 9/16; P2 =3/16; P3 = 3/16; P4 = 1/16
Ha: One or more of the proportions are not equal to the proportions given in the null
hypothesis.
2.  = 0.05
ν =K - 1 = 4-1 = 3
 90

X2, ν = X20.05, 3 = 7.81

Reject Ho if sample X2> 7.81
3. Test Statistics (Sample χ2)
Class Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe = npi) fe
Current 439 450 121 0.269
Moderately late 168 150 324 2.160
Very late 133 150 289 1.927
Uncollectible 60 50 100 2.000
 f o  f e 2 6.356
 f
e

4. Do net reject Ho.

2. ETHIO Plastic Factory sells its products in three primary colors: Red, blue, and yellow.
The marketing manager feels that customers have no color preference for the product.
To test this hypothesis the manager set up a test in which 120 purchases were given
equal opportunity to buy the product in each of the three colors. The results were that
60 bought red, 20 bought blue, and 40 bought yellow. Test the marketing manager’s
null hypothesis, using =0.05.
Solution
1. Ho: People have no color preference with this product; P1 = P2 = P3 = 1/3
Ha: People have color preference with this product
2.  = 0.05
V= K-1 = 3 -1=2
X2,ν = X2 0.05,2 = 5.99
Reject Ho if sample X2 is greater than 5.99.
3. Sample χ2
Class Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe = npi); pi = 1/39 fe
Red 60 40 400 10.00
Blue 20 40 400 10.00
Yellow 40 40 0 0.00
f  f  2
20.00
 of e
e

9
Since the null hypothesis states that there is no color preference, each of the three
colors is preferred by one third of the purchases.
 91

4. Reject Ho; because 20 > 5.99. This means that customers do have color preference. It
appears that red is the most popular color and blue is the least popular.
3. Rating sciences, Inc., a TV program – rating service, surveyed 600 families where the
television was turned on during the prime time on week nights. They found the
following numbers of people turned to the various networks.

Name of the network Type Number of

viewers
NBC 210
CBS Commercial 170
ABC 165
PBS Noncommercial 55
600

a) Test the hypothesis that all four networks have the same proportion of viewers
during this prime time period. Use  = 0.05
b) Eliminate the results for PBS and repeat the test of hypothesis for the three
commercial networks, using  = 0.05
c) Test the hypothesis that each of the three major networks has 30% of the weeknight
prime time market and PBS has 10% using  = 0.005
Solution
a.
1. Ho: All of the four networks do have equal number of viewers; P1 = P2 = P3 = P4 = 1/4.
Ha: All of the four networks do not have equal number of viewers.
2.  = 0.05
V= K-1 = 4 -1= 3
X2,ν = X2 0.05,3 = 7.81
Reject Ho if sample X2 is greater than 7.81
3. Sample χ2
Class Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe = npi); pi = 1/410 fe
NBC 210 150 3,600 24.0000
CBS 170 150 400 2.6667
ABC 165 150 225 1.5000
PBS 55 150 9,025 60.1667

10
Since equal number of viewers is expected to watch each network, each of the four
networks is watched by one fourth of the viewers.
 92

 f o  f e 2 88.3334
 fe
4. Reject Ho; because 88.34 > 7.81.
b.
1. Ho: All of the three commercial networks do have equal number of viewers; P1 = P2 =
P3 = 1/3.
Ha: All of the three commercial networks do not have equal number of viewers.
2.  = 0.05
V= K-1 = 3 -1= 2
X2, ν = X2 0.05,2 = 5.99
Reject Ho if sample X2 is greater than 5.99.
3. Sample χ2
Class Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe = npi); pi = 1/3 fe
NBC 210 181.67 802.60 4.4179
CBS 170 181.67 136.20 0.7497
ABC 165 181.67 277.90 1.5270
f  f  2
6.6946
 of e
e

4. Reject Ho; because 6.70 > 5.99

c.
1. Ho: P1 = P2 = P3 = 0.30; P4 = 0.10
Ha: One or more of the proportions are not equal to the proportions given in the null
hypothesis.
2.  = 0.005
V= K-1 = 4 -1= 3
X2,ν = X2 0.05,3 = 7.81
Reject Ho if sample X2 is greater than 7.81
3. Sample χ2
Class Observed freq Expected Freq (fo-fe)2  f o  f e 2
(fo) (fe = npi) fe
NBC 210 180 900 5.00
CBS 170 180 100 0.55
ABC 165 180 225 1.25
PBS 55 60 25 0.42
 f o  f e 2 7.22
 f
e

4. Do not Reject Ho; because 7.22 < 7.81.

 93

4. Suppose that three companies, A, B, and C, have recently conducted aggressive

advertising campaigns in order to maintain and possibly increase their respective
shares of the market for a particular product. The market shares prior to the campaigns
were P1=0.45 for company A, P2 = 0.40 for company B, P3= 0.13 for company C, and P4 =
0.02 for other competitors. To determine if these market shares changed after the
advertising campaigns, a marketing analyst solicited the preferences of a random
sample of 200 customers of this product. Of these 200 customers 95 indicated a
preference for company A’s product, 85 preferred company B’s precut, 18 preferred
company C’s product, and the remainder preferred one or another of the products
distributed by the competitors conduct a test, at the 5% level of significance. If the
market shares have changed from the levels what they were at before the advertising
campaigns.
Solution
1. Ho: P1 = 0.45; P2 = 0.40; P3 = 0.13; P4 = 0.02
Ha: At least one Pi is not equal to its specified value
2.  = 0.05
V= K-1 = 3 – 1 = 211
χ2, ν = χ2 0.05,2 = 5.99
Reject Ho if sample X2 > 5.99.
Because of the above change Ho is restated as:
Ho: P1 = 0.45, P2=0.40, P3 = 0.15
Ha: At least one pi is not equal to its specified value

3. Sample X2

Company Observed freq Expected freq (fo-fe)2  f o  f e 2

(fo) (fe) fe
A 95 90 25 0.2778
B 85 80 25 0.125
Others 20 30 100 3.3333
f  f  2
3.9236
 of e
e

4. Do not reject Ho

11
Expected frequency value for P4 is less than 5 (200*0.02 = 4). So, we have to combine P 4 with one of other
expected frequencies, say P3, to obtain a combined expected frequency of 30 (200*0.15). it can also be combined
with other expected frequency values.
 94

There is no sufficient evidence at the 5% level of significance to conclude that the

market shares have changed from the levels they were at before the advertising
campaign.

GOODNESS-OF-FIT TESTS (BINOMIAL, NORMAL, POISSON)

The chi-square test is widely used for a variety of analyses. One of the more important
uses of Chi-Square is the goodness-of-fit test. That is, it can be used to decide whether a
particular probability distribution, such as the binomial, Poisson or normal, is the
appropriate distribution. This is an important ability, because as decision makers using
statistics, we will need to choose a certain probability distribution to represent the
distribution of the data we happen to be considering.

In tests of hypothesis (Chapter 5), we assumed that the population was normal and
tested the hypothesis =o, p = Po, etc. But what if we want to check on the assumption
of normality itself? The multinomial χ2 goodness–of–fit test can be applied.

The null hypothesis for a goodness-off it test in that the distribution of the population
from which a sample is taken is the one specified. The alternative hypothesis is that the
actual distribution is not the specified distribution. Generally, a researcher specifies
only the name of distribution and uses the sample data to estimate the particular
parameters of the distribution. In this situation one degree of freedom is lest for each
parameter that has to be estimated. However, if the research completely specifies the
distribution including parameter values, then no additional degrees of freedom is lost.

Null hypothesis Parameters to Degrees of

be estimated freedom lost
Ho: Population is normal , 2
Ho: Population is normal with   x  1
Ho: Population is normal with  = y  1
Ho: Population is normal with   x,  = y None 0
Ho: Population is Poisson λ 1
Ho: Population is Poisson with λ=Z None 0
Ho: Population is binomial with p = b None 0

Example (Binomial)
1. Mrs. Tsion, Saleswoman for MOON Paper Company, has five accounts to visit per day.
It is suggested that sales by Mrs. Tsion May be described by the binomial distribution,
 95

with the probability of selling each account being 0.4. Given the following frequency
distribution of Mrs. Tsion’s number of sales per day, can we conclude that the data do
in fact follow the binomial distribution? Use the 0.05 significance level.

No. of sales day 0 1 2 3 4 5

Frequency 10 41 60 20 6 3
Solution
1. Ho: The frequency distribution is Binomial with n = 5 and P = 0.4
Ha: The frequency distribution is not binomial with n = 5 and P = 0.4
2.  = 0.05
K-1 –m = 5-1-0 = 4
X2, ν = X2 0.05,4 = 9.49
Reject Ho if sample x2 is greater than 9.49
3. Sample χ2.
No. of sales Prob. with n= Observed Expected Freq (fo-fe)2  f o  f e 2
per day 5, p = 0.4 freq (fo) (fe = npi) fe
0 .0778 10 10.892 0.7957 0.0731
1 .2592 41 36.288 22.2029 0.6119
2 .3456 60 48.384 134.9315 2.7888
3 .2304 20 32.256 150.2095 4.6567
4&5 .0870 9 12.18 10.1124 0.8302
 f o  f e 2 8.9607
 fe
4. Do not reject Ho. The data are well described by the binomial distribution with n=5
and P=0.4.
2. A professional baseball player, Philippos, was at bat five times in each of 100 games.
Philippos claims that he has a probability of 0.4 of getting a hit each time he goes to bat.
Test his claim at the 0.05 level by seeing if the following data are distributed binomially.

No. of hits / game 0 1 2 3 4 5

No. of games with that number of hits 12 38 27 17 5 1

Solution
1. Ho: The freq. Distribution can be best described by binomial distribution with n=5,
P=0.4
Ha: The freq. Distribution can’t be best described by binomial distribution with n=5,
P=0.4
2.  = 0.05
 96

V = K-1 –m = 5-1-0 = 4
X2,ν = X2 0.05,4 = 9.49
Reject Ho if sample χ2 > 9.49
3. Sample χ2
No. of hits No. of games with Prob. with Expected (fo-fe)2  f o  f e 2
per game that no. of hit (fo) n=5, P=0.4 freq (fe = npi) fe
0 12 .0778 7.78 17.8084 2.2890
1 38 .2592 25.92 145.9264 5.6249
2 27 .3456 34.56 57.1536 1.6538
3 17 .2304 23.04 36.4816 1.5834
4&5 6 .0870 8.70 4.2900 0.8379
 f o  f e 2 11.9940
 f
e

4. Reject Ho. The # of hit over the same in not binomially distributed

3. The Ethiopian postal service is interested in modeling the “mangled letter” problem. It
has been suggested that any letter sent to a certain area has a 0.15 chance of being
mangled. Since the post office is so big, it can be assumed that two letters chances of
being mangled are independent. A sample of 310 people was selected, and two test
letters were mailed to each of them. The number of people receiving zero, one, or two
mangled letters was 260, 40, and 10, respectively. At the 0.10 level of significance, is it
reasonable to conclude that the number of mangled letters received by people follows a
binomial distribution with P = 0.15?
Solution
1. Ho: The number of mangled letters received by people follows a binomial
distribution with n = 2, p = 0.15.
Ha: The number of mangled letters received by people doesn’t follow a binomial
distribution. With n =2, P = 0.15.
2.  = 0.1
V = K-1 – m = 3-1-0 = 2
X2, ν = X2 0.1,2 = 4.61
Reject Ho if sample x2 > 4.61
Sample χ2
No. of mangled Observed Prob. with Expected (fo-fe)2  f o  f e 2
letters freq. (fo) n=2 P=0.15 freq (fe = npi) fe
0 260 0.7225 223.9750 1297.8006 5.7944
1 40 0.2550 79.0500 1524.9025 19.2904
 97

2 10 0.0225 6.9750 9.1506 1.3119

 fo  fe  2
26.3967
 fe
3. Reject Ho. The number of hit over the game is not binomially distributed with n = 2
and P = 0.15.

Example (Poisson)
1. It is hypothesized that the number of breakdowns per month of a computer system at a
major university follows a Poisson distribution with µ = 2. The data below show the
observed number of breakdowns per month during a sample of 100 months. Use a 5%
level of significance and test the null hypothesis.

Breakdowns 0 1 2 3 4 5 and above

Observed freq. 14 20 34 22 5 3
Solution
1. Ho: The population distribution of breakdowns is Poisson with µ = 2.
Ha: The population distribution of breakdowns is not Poisson with µ = 2.
2.  = 0.05
V = K-1 – m = 6-1-0 = 5
X2, ν = X2 0.05,5 = 11.07
Reject Ho if sample x2 > 11.07
3. Sample χ2
Breakdowns Observed Prob. with Expected (fo-fe)2  f o  f e 2
freq. (fo) λ=2 freq (fe = npi) fe
0 14 0.1353 13.53 0.2209 0.0163
1 20 0.2707 27.07 49.9849 1.8465
2 34 0.2707 27.07 48.0249 1.7741
3 22 0.1804 18.04 15.6816 0.8693
4 5 0.0902 9.02 16.1604 1.7916
5 or more 5 0.0527 5.27 0.0729 0.0138
 f o  f e 2 6.3117
 f
e

4. Do not Reject Ho. The number of breakdowns per month of a computer system at
the university follows a Poisson distribution with µ = 2.

2. Suppose that a teller supervisor believes that the distribution of random arrivals at a
local bank is Poisson and sets out to test this hypothesis by gathering information. The
following data represent a distribution of frequency of arrivals during one minute
 98

intervals at a bank. Use α = 0.05 to test these data in an effort to determine whether they
are Poisson distributed.

No. of arrivals 0 1 2 3 4 5 and above

Observed freq. 7 18 25 17 12 5

Solution
Before we solve the question, first we have to compute the arrival rate per minute, and
hence one degree of freedom is lost.
 number of arrivals *
  observed frequency  0 * 7  18 *1  25 * 2  17 * 3  12 * 4  5 * 5 192
  
   2.3 cust / min
 Observed frequency 84 84

1. Ho: The arrival of customers at a bank is Poisson distributed with λ = 2.3

Ha: The arrival of customers at a bank is not Poisson distributed with λ = 2.3
2.  = 0.05
V = K-1 – m = 6-1-1 = 4
X2, ν = X2 0.05,4 = 11.07
Reject Ho if sample χ2 > 9.488
3. Sample χ2
Number of Observed Prob. with Expected (fo-fe)2  f o  f e 2
arrivals freq. (fo) λ=2.3 freq (fe = npi) fe
0 7 0.1003 8.4252 2.0312 0.2411
1 18 0.2306 19.3704 1.8778 0.0969
2 25 0.2652 22.2768 7.4158 0.3329
3 17 0.2033 17.0772 0.0060 0.0003
4 12 0.1169 9.8196 4.7541 0.4841
5 or more 5 0.0837 7.0308 4.1241 0.5866
 f o  f e 2 1.795
 f
e

4. Do not Reject Ho. The arrival of customers at a bank follows a Poisson distribution
with λ = 2.3.

3. The number of automobile accidents occurring per day in a particular city is believed to
have a poisson distribution. A sample of 80 days during the past year gives the data
shown below. Do the data support the belief that the number of accidents per day has a
poisson distribution? Use α = 0.05.
 99

No. of accidents 0 1 2 3 4
Observed freq. (days) 34 25 11 7 3

Solution
Before we solve the question, first we have to compute the occurrence rate per day, and
hence one degree of freedom is lost.
 number of accidents *
  observed frequency 
    0 * 34  25 * 1  11 * 2  7 * 3  3 * 4  80  1accident / day
 Observed frequency 80 80

1. Ho: The occurrence of accidents per day follows a poisson distribution with λ = 1.0
Ha: The occurrence of accidents per day does not follow a poisson distribution with
λ = 1.0
2.  = 0.05
V = K-1 – m = 4-1-1 = 2
X2, ν = X2 0.05,2 = 5.99
Reject Ho if sample χ2 > 5.99
3. Sample χ2
Number of Observed Prob. with Expected (fo-fe)2  f o  f e 2
accidents freq. (fo) λ=1.0 freq (fe = npi) fe
0 34 0.3679 29.4320 20.8666 0.7090
1 25 0.3679 29.4320 19.6426 0.6674
2 11 0.1839 14.7120 13.7789 0.9366
3 or more 10 0.0803 6.4240 12.7878 1.9906
 f o  f e 2 4.3036
 f
e

4. Do not Reject Ho. The occurrence of accidents per day follows a poisson distribution
with λ = 1.0

Example (Normal)
1. Suppose that Ato Paulos developed an overall attitude scale to determine how his

company’s employees feel toward their company. In theory the scores can vary from 0

to 50. Ato Paulos pretests his measurement instrument on a randomly selected group of

100 employees. He tallies the scores and summarizes them into six categories as shown
 100

below. Are these pretest scores approximately normally distributed with µ = 24.9 and σ

= 7.194? Use α = 0.05.

Score category 10-15 15-20 20-25 25-30 30-35 35-40

Frequency 11 14 24 28 13 10
Solution
1. Ho: The attitude scores are normally distributed with µ = 24.9 and σ = 7.194
Ha: The attitude scores are not normally distributed with µ = 24.9 and σ = 7.194
2.  = 0.05
V = K-1 – m = 6-1-0 = 5
X2, ν = X2 0.05,5 = 11.07
Reject Ho if sample χ2 > 11.07
3. Sample χ2
X 
With Z  , the expected probability of each category can be obtained as follows:


For category 10-15 Probability

10  24.9 0.48077
z10   2.07
7.194
15  24.9 - 0.41621
z15   1.38
7.194
Expected probability 0.06456

For category 15-20 Probability

15  24.9 0.41621
z15   1.38
7.194
20  24.9 - 0.25175
z 20   0.68
7.194
Expected probability 0.16446

For category 20-25 Probability

20  24.9 0.25175
z 20   0.68
7.194
25  24.9 +0.00399
z 25   0.01
7.194
Expected probability 0.25574

For category 25-30 Probability

 101

25  24.9 0.00399
z 25   0.01
7.194
30  24.9 +0.26115
z 30   0.71
7.194
Expected probability 0.25716

For category 30-35 Probability

30  24.9 0.26115
z 30   0.71
7.194
35  24.9 0.41924
z 35   1.40
7.194
Expected probability 0.25716

For category 30-35 Probability

35  24.9 0.41924
z 35   1.40
7.194
40  24.9 0.48214
z 40   2.10
7.194
Expected probability 0.06290

The six probabilities do not sum to 1.00. Even though observed frequencies were
obtained only for these six categories, getting a score less than 10 or greater than 40 was
also possible. Because 0.5 of the probabilities lie in each half of a normal distribution
and utilizing the sum of expected probabilities on each side of the mean, 24.9, we can
obtain a probability of the < 10 category: 0.5 – (0.06456 + 0.16446 + 0.25175) = 0.01923.
Similarly, we can obtain the probability of >40 category: 0.5 – (0.00399 + 0.25716 +
0.15809 + 0.06290) = 0.01786. expected frequencies can then be obtained by multiplying
each expected probability by the total frequency (100), as shown below.

Score category Probability Expected freq

(fe = npi)
< 10 0.01923 1.923
10 – 15 0.06456 6.456 8.379
15 – 20 0.16446 16.446 16.446
20 – 25 0.25574 25.574 25.574
25 - 30 0.25716 25.716 25.716
30 -35 0.15809 15.809 15.809
35 – 40 0.06290 6.290 8.076
> 40 0.01786 1.786
 102

As the < 10 and > 40 categories have values of less than 5, each must be combined with

the adjacent category. As a result, the < 10 category becomes part of the 10 – 15 category

and the > 40 category becomes part of the 35 – 40 category.

Expected freq
Score category Probability (fe = npi)
10 – 15 0.08379 8.379
15 – 20 0.16446 16.446
20 – 25 0.25574 25.574
25 - 30 0.25716 25.716
30 -35 0.15809 15.809
35 – 40 0.08076 8.076

The value of the chi-square can then be computed.

Score Observed Expected  f o  f e 2

category freq. (fo) Probability freq (fe = npi) (fo-fe)2 fe
10 – 15 11 0.08379 8.379 6.8696 0.8199
15 – 20 14 0.16446 16.446 5.9829 0.3638
20 – 25 24 0.25574 25.574 2.4775 0.0964
25 - 30 28 0.25716 25.716 5.2167 0.2029
30 -35 13 0.15809 15.809 7.8905 0.4991
35 – 40 10 0.08076 8.076 3.7018 0.4584
 f o  f e 2 2.4409
 f
e

4. Do not Reject Ho. The attitude score are normally distributed with mean 24.9 and
standard deviation 7.194.
2. The director of a major soccer team believes that the ages of purchasers of game tickets
are normally distributed. If the following data represent the distribution of ages for a
sample of observed purchasers of major soccer game tickets, use the chi-square
goodness-of-fit test to determine whether this distribution is significantly different from
the normal distribution. Assume that α = 0.05.

Age of purchaser 10-20 20-30 30-40 40-50 50-60 60-70

Frequency 16 44 61 56 35 19

Solution
1. Ho: The ages of purchasers of soccer game tickets are normally distributed.
 103

Ha: The ages of purchasers of soccer game tickets aren’t normally distributed
2.  = 0.05
V = K-1 – 2 = 6-1-2 = 3
X2, ν = X2 0.05,3 = 7.81
Reject Ho if sample χ2 > 7.81
3. Sample χ2
Age category Observed freq Mid point (M) fm fm2
10-20 16 15 240 3,600
20-30 44 25 1,100 27,500
30-40 61 35 2,135 74,725
40-50 56 45 2,520 113,400
50-60 35 55 1,925 105,875
60-70 19 65 1,235 80,275
231  fm = 9, 155  fm = 405,375
2

X 
 fm  9,155  39.63
n 231
 fm 2
9,1552
 fm 2

n
405,375 
231
S   13.60
n 1 231  1

X 
With Z  , the expected probability of each category can be obtained as follows:


For category 10-20 Probability

10  39.63 0.48537
z10   2.18
13.6
20  39.63 - 0.42507
z 20   1.44
13.6
Expected probability 0.06030

For category 20-30 Probability

20  39.63 0.42507
z 20   1.44
13.6
30  39.63 - 0.26115
z 30   0.71
13.6
Expected probability 0.16392

For category 30-45 Probability

 104

30  39.63 0.26115
z 30   0.71
13.6
40  39.63 + 0.01197
z 40   0.03
13.6
Expected probability 0.27312

For category 40-50 Probability

50  39.63 0.27637
z 50   0.76
13.6
40  39.63 - 0.01197
z 40   0.03
13.6
Expected probability 0.26440

For category 50-60 Probability

60  39.63 0.43319
z 60   1.50
13.6
50  39.63 - 0.27637
z 50   0.76
13.6
Expected probability 0.15682

For category 60-70 Probability

70  39.63 0.48713
z 70   2.23
13.6
60  39.63 - 0.43319
z 60   1.50
13.6
Expected probability 0.05394

The six probabilities do not sum to 1.00. Even though observed frequencies were
obtained only for these six categories, getting a score less than 10 or greater than 70 is
also possible.

For < 10:

Probability between 10 and the mean = 0.06030 + 0.16392 + 0.26115 = 0.48537.
Probability < 10 = 0.5 – 0.48537 = 0.01463

For > 70
Probability between 70 and the mean = 0.05394 + 0.15682 + 0.2640 + 0.01197 = 0.48713
Probability > 70 = 0.5 – 0.48713 = 0.01287

Then, the expected frequencies can be obtained by multiplying each expected

probability by the total frequency (231) as follows:
 105

Age category Probability Expected freq

(fe = npi)
< 10 0.01463 3.380
10 – 20 0.06030 13.929 17.309
20 – 30 0.16392 37.866 37.866
30 – 40 0.27312 63.091 63.091
40 - 50 0.26440 61.076 61.076
50 – 60 0.15682 36.225 36.225
60 – 70 0.05394 12.460
> 70 0.01287 2.973 15.433

Since the < 10 and > 70 categories have values of less than 5, each must be combined
with the adjacent category. As a result, the < 10 category becomes part of the 10 – 20
category and the > 70 category becomes part of the 60 – 70 category.

Age category Probability Expected freq

(fe = npi)
10 – 20 0.07493 17.309
20 – 30 0.16392 37.866
30 – 40 0.27312 63.091
40 - 50 0.26440 61.076
50 – 60 0.15682 36.225
60 – 70 0.06681 15.433

The value of the chi-square can then be computed.

Score Observed Expected  f o  f e 2

category freq. (fo) Probability freq (fe = npi) (fo-fe)2 fe
10 – 20 16 0.07493 17.309 1.7135 0.0990
20 – 30 44 0.16392 37.866 37.6260 0.9937
30 – 40 61 0.27312 63.091 4.3723 0.0693
40 - 50 56 0.26440 61.076 25.7658 0.4219
50 – 60 35 0.15682 36.225 1.5006 0.0414
60 – 70 19 0.06681 15.433 12.7235 0.8244
 f o  f e 2 2.4497
 f
e

4. Do not Reject Ho. The age of purchasers of soccer game tickets are normally
distributed.
 106

3. The instructor for Introductory Statistics course attempts to construct the final
examination so that the grades are normally distributed with a mean of 65. From the
sample of grades appearing in the accompanying frequency distribution table, can you
conclude that they have achieved his objective? Use α = 0.05.

Grade 30-40 40-50 50-60 60-70 70-80 80-90

Frequency 4 17 29 49 33 18

Solution
1. Ho: The grades of students are normally distributed with a mean of 65.
Ha: The grades of students are not normally distributed with a mean of 65.
2.  = 0.05
V = K-1 – 1 = 5-1-1 = 3
X2, ν = X2 0.05,3 = 7.81
Reject Ho if sample χ2 > 7.81
3. Sample χ2
Grade Observed freq Mid point (M) fm fm2
30-40 4 35 140 4,900
40-50 17 45 765 34,425
50-60 29 55 1,595 87,725
60-70 49 65 3,185 207,025
70-80 33 75 2,475 185,625
80-90 18 85 1,530 130,050
150  fm = 9, 690  fm = 649,750
2

 fm 2
9,690
2

 fm 2

n
649,750 
150
S   12.63
n 1 150  1
X 
With Z  , the expected probability of each category can be obtained as follows:


For category 30-40 Probability

30  65 0.49720
z30   2.77
12.63
40  65 - 0.47615
z40   1.98
12.63
Expected probability 0.02105
 107

For category 40-50 Probability

40  65 0.47615
z40   1.98
12.63
50  65 - 0.38298
z50   1.19
12.63
Expected probability 0.09317

For category 50-60 Probability

50  65 0.38298
z50   1.19
12.63
60  65 - 0.15542
z60   0.40
12.63
Expected probability 0.22756

For category 60-70 Probability

60  65 0.15542
z60   0.40
12.63
70  65 + 0.15542
z70   0.40
12.63
Expected probability 0.31084

For category 70-80 Probability

80  65 0.38298
z80   1.19
12.63
70  65 - 0.15542
z70   0.40
12.63
Expected probability 0. 22756

For category 80-90 Probability

90  65 0.47615
z90   1.98
12.63
80  65 - 0.38298
z80   1.19
12.63
Expected probability 0.09317

The six probabilities do not sum to 1.00. Even though observed frequencies were
obtained only for these six categories, getting a score less than 30 or greater than 90 is
also possible.

For < 30:

Probability between 30 and the mean = 0.02105 + 0.09317 + 0.22756 + 0.15542 = 0.49720.
 108

Probability < 10 = 0.5 – 0.49720 = 0.00280

For > 90
Probability between 90 and the mean = 0.15542 + 0.22756 + 0.09317 = 0.47615
Probability > 90 = 0.5 – 0.47615 = 0.02385

Then, the expected frequencies can be obtained by multiplying each expected

probability by the total frequency (150) as follows:

Age category Probability Expected freq

(fe = npi)
< 30 0.00280 0.42
30-40 0.02105 3.1575 17.553
40-50 0.09317 13.9755
50-60 0.22756 34.134 34.134
60-70 0.31084 46.626 46.626
70-80 0.22756 34.134 34.134
80-90 0.09317 13.9755
> 90 0.02385 3.5775 17.553

Since the < 30, 30-40 and > 90 categories have values of less than 5, they must be
combined with the adjacent categories. As a result, the < 30 and 30-40 categories become
part of the 40 – 50 category; and the > 90 category becomes part of 80-90 category.
Age category Probability Expected freq
(fe = npi)
40-50 0.11702 17.553
50-60 0.22756 52.5664
60-70 0.31084 71.8040
70-80 0.22756 52.5664
80-90 0.11702 17.553

The value of the chi-square can then be computed.

Score Observed Expected freq  f o  f e 2

category freq. (fo) Probability (fe = npi) (fo-fe)2 fe
40-50 21 0.11702 17.5530 11.8818 0.6769
50-60 29 0.22756 34.1340 26.3580 0.7722
60-70 49 0.31084 46.6260 5.6359 0.1209
70-80 33 0.22756 34.1340 1.2860 0.0377
80-90 18 0.11702 17.5530 0.1998 0.0114
 109

 f o  f e 2
 fe 1.6190
4. Do not Reject Ho. YES. The grades of students are normally distributed with a mean
of 65.
 110

CHAPTER FIVE

ANALYSIS OF VARIANCE

Analysis of variance (ANOVA) is a procedure to test the hypothesis that several

populations have the same mean; i.e., it is used to test the equality of several means.
The name ANOVA stems from the somewhat surprising fact that a set of computations
of several variances is used to test the equality of several means.

When testing for differences in mans of more than two populations, we usually do not
proceed by considering all combinations of two populations at a time and testing for
differences in each pair.
1. Such an approach would require several tests rather than just one.
2. If each individual test were conducted using a level of significance of say α =
0.05, then the overall level of significance would be higher than 0.05. For
example, if Ho: µ1 = µ2 = µ3, α (the probability of rejecting a true null hypothesis)
= 0.143 (1-0.953).

Thus, we want to test simultaneously for differences among the means of all the
populations, and we want the joint level of significance of the test to be α. To perform
this test we make use of the F-distribution and use a method called ANOVA.

In order to use ANOVA, we assume the following:

1. All the samples were randomly selected and are independent of one another.
2. The populations from which the samples were drawn are normally distributed. If
however, the sample sizes are large enough, we do not need the assumption of
normality.
3. All the population variances are equal.

ANOVA is based on a comparison of two different estimates of the variances, σ 2, of

overall population.
1. The variance obtained by calculating the variation within the samples
themselves – Mean Square within (MSW).
2. The variance obtained by calculating the variation among sample means – Mean
Square between (MSB).

Since both are estimates of σ2, they should be approximately equal in value when the
null hypothesis is true. If the null hypothesis is not true, these two estimates will differ
considerably. The three steps in ANOVA, then, are:
 111

1. Determine one estimate of the population variance from the variation among
sample means
2. Determine a second estimate of the population variance from the variation
within the samples
3. Compare these two estimates. If they are approximately equal in value, accept
the null hypothesis.

Calculating the Variance among the Sample Means – MSB

The variance among the sample means is called Between Column Variance or Mean
Square between (MSB).
 X  X  .
2

Sample variance = S 2

n 1

Now, because we are working with sample means and the grand mean, let’s substitute
X for X, X for X , and K (number of samples) for n to get the formula for the variance
among the sample means:
2
 X  X 
   .
Variance among sample means  S X2 
K 1
In sampling distribution of the mean we have calculated the standard error of the mean

as  X  . Cross multiplying the terms    X n . Squaring both sides  2   X2 n .
n

In ANOVA, we do not have all the information needed to use the above equation to
find σ2. Specifically, we do not know  X2 . We could, however, calculate the variance
2

  X  X 
among the sample means, S X , using S X 
2 2
. So, why not substitute S X2 for  X2
K 1
and calculate an estimate of the population variance? This will give us:
2 2

2  n X  X  n  X  X 
      If n , n ,......n are equal.
 x  S X2 * n  
K 1 K 1
1 2 k

Which sample size to use?

There is a slight difficulty in using this equation as it stands. n represents the sample
size, but which sample size should we use when different samples have different sizes?
 112

2
We solve this problem by multiplying  X j  X  by it’s won appropriate nj, and hence
 
2
 X becomes:
2

2  n  X
j  X 
j
 .
MSB =  
K 1

Where:
2
 = First estimate of the population variance based on the variation among sample
means (the Between Column Variance – MSB)
nj = the size of the jth sample
X j = the sample mean of the jth sample

X = the grand mean

K = the number of samples
K-1 = the degrees of freedom associated with SSB.

Calculating the Variance With In the Samples (MSW)12

It is based on the variation of the sample observations within each sample. It is called
the within column variance or Mean Square Within (MSW). We calculate the sample
 X  X 
2

variance for each sample as S . 2


n 1
Since we have assumed that the variances of the populations from which samples have
been drawn are equal, we could use any one of the sample variances as the second
estimate of the population variance. Statistically, we can get a better estimate of the
population variance by using a weighted average of all sample variances. The general
formula for this second estimate of  2 is:

 n  1S 2j
k

j
2
MSW =   i 1

nT  k
k
n  1 S 2j
If n1, n2, -----, nk are equal MSW = 
2
 i 1 .
k n  1
Where:

12MSW is based on the variation within each of the samples; it is not influenced by whether or not the
null hypothesis is true. Thus, MSW always provides an unbiased estimate of the population variance.
 113

2
 = Second estimate of the population variance based on the variation within the
samples (the Within Column Variance – MSB)
nj = the size of the jth sample
nj-1 = degree of freedom in each sample
nT – k = degrees of freedom associated with SSB
S 2j  The sample variance of jth sample
K = the number of samples
nT = Σnj = the total sample size = n1 + n2 + <<.. + nk.

The estimate of population variance based on variation that exists between sample
means (MSB) is somewhat suspect because it is based on the notion that all the
populations have the same mean. That is, the estimate MSB is a good estimate of the σ 2
only if Ho is true and all the populations’ means are equal: µ1 = µ2 = µ3 = ------ = µk.

If the unknown population means are not equal, and probably are radically different
from one another, then the sample means ( X j ) will most likely be radically different
from each other too. This difference will have a marked effect on MSB. That is to say,
2
the X j values will vary a great deal and the  X j  X  terms will be large. Thus, if the
 
population means are not all equal, then the MSB estimate will be large relative to the
MSW estimate. That is, is the MSB is large relative to the MSW, and then the hypothesis
that all the population means are equal is not likely to be true.

The important question is, of course, How large is “large?” also, how do we measure
the relative sizes of the two variance estimates? The answer to these questions is given
by the F-distribution.

If k samples of nj (j = 1, 2< k) items of each are taken from k normal populations that
have equal variances and for which the hypothesis Ho: µ1 = µ2 = <= µk is true, then the
ratio of the MSB to the MSW is an F-value that follows an F-probability distribution.
MSB
F
MSW
 114

THE F-DISTRIBUTION
Characteristics of F-distribution

1. It is a continuous probability distribution

2. It is unimodal
3. It has two parameters; pair of degrees of freedom, ν1 and ν2
ν1 = the number of degrees of freedom in the numerator of F-ratio; ν1 = k – 1
ν2 = the number of degrees of freedom in the denominator of F-ratio; ν2 = nT - k
4. It is a positively skewed distribution, and tends to get more symmetrical as the
degrees of freedom in the numerator and denominator increase.
2
5. The mean for an F-distribution is , for ν2 > 2; and the standard deviation is
2  2
2 22  1   2  2

 1  2  22  2  4for ν2 > 4.

Example
1. The training director of a company is trying to evaluate three different methods of
training new employees. The first method assigns each to an experienced employee for
individual help in the factory. The second method puts all new employees in a training
room separate from the factory, and the third method uses training films and
programmed learning materials. The training director chooses 18 new employees
assigned at random to the three training methods and records their daily production
after they complete the programs. Below are productivity measures for individuals
trained by each method.

Method 1 Method 2 Method 3

45 59 41
40 43 37
50 47 43
39 51 40
53 39 52
44 49 37
271 288 250
X 1 = 45.17 X 2 = 48.00 X 3 = 41.67 X = 44.94
2 2 2
S = 30.17
1 S = 47.60
2 S = 31.07
3

At the 0.05 level of significance, do the three training methods lead to different levels of
productivity?
 115

Solution
1. Ho: µ1 = µ2 = µ3
µ1, µ2, and µ3 are not all equal
2. α = 0.05

ν1 = K - 1 ν2 = nT - k F0.05, 2,15 = 3.68

=3-1=2 = 18 – 3 = 15
Reject Ho if sample F > 3.68

3. Sample F
2

 n j  X j  X  6 45.17  44.942  48.00  44.942  41.67  44.942

 
MSB = 
K 1 3 1
120.66
  60.33
2

MSW =
 n j  1S12

530 .17  47 .60  31 .07  108 .84
  36 .28
nT  K 15 3
MSB 60.33
F   1.663
MSW 36.28

4. Do not reject Ho.

There are no differences in the effects of the three training programs (methods) on
employee productivity.

2. A department store chain is considering building a new store at one of the four different
sites. One of the important factors in the decision is the annual household income of the
residents of the four areas. Suppose that, in a preliminary study, various residents in
each area are asked what their annual household incomes are. The results are shown in
the accompanying table below. Is there sufficient evidence to conclude that differences
exist in the average annual household incomes among the four communities? Use α =
0.01.

Area 1 Area 2 Area 3 Area 4

25 32 27 18
27 35 32 23
21 30 48 29
17 46 25 26
29 32 20 42
30 22 12
 116

19 18
51
27
159 294 182 138
X 1 = 26.50 X 2 = 32.67 X 3 = 26.00 X 4 = 27.60 X = 28.63
2 2 2 2
S = 26.30
1 S = 107.5
2 S = 136.33
3 S = 81.30
4

Solution
1. Ho: µ1 = µ2 = µ3 = µ4
µ1, µ2, µ3 and µ4 are not all equal
2. α = 0.01

ν1 = K - 1 ν2 = nT - k F0.01, 3,23 = 4.76

=4-1=3 = 27 – 4 = 23
Reject Ho if sample F > 4.76

3. Sample F
2

 n  X
j  X 
j
  626.5  28.63  932.67  28.63  726.00  28.63  527.60  28.63
2 2 2 2
MSB =
K 1 4 1
227.84
  75.95
3

MSW =
 n j  1S12

526 .3  8107 .5  6136 .33   481 .3 2134 .68
  92 .81
nT  K 27  4 23
MSB 75.95
F   0.82
MSW 92.81

4. Do not reject Ho.

No difference exists in the average annual household incomes among the four
communities.
 117

CHAPTER SIX
SIMPLE LINEAR REGRESSION AND CORRELATION

Linear regression and correlation is studying and measuring the linear relationship among two or
more variables. When only two variables are involved, the analysis is referred to as simple
correlation and simple linear regression analysis, and when there are more than two variables the
term multiple regression and partial correlation is used.

Regression Analysis: is a statistical technique that can be used to develop a mathematical

equation showing how variables are related.

Correlation Analysis: deals with the measurement of the closeness of the relationship which are
described in the regression equation.
We say there is correlation if the two series of items vary together directly or inversely.

Simple Correlation: Suppose we have two variables X  ( X 1 , X 2 ,...X n ) and

Y  (Y1 , Y2 ,...Yn )
 When higher values of X are associated with higher values of Y and lower values of X
are associated with lower values of Y, then the correlation is said to be positive or direct.
Examples:
- Income and expenditure
- Number of hours spent in studying and the score obtained
- Height and weight
- Distance covered and fuel consumed by car.
 When higher values of X are associated with lower values of Y and lower values of X are
associated with higher values of Y, then the correlation is said to be negative or inverse.
Examples:
- Demand and supply
- Income and the proportion of income spent on food.
The correlation between X and Y may be one of the following
1. Perfect positive (slope=1)
2. Positive (slope between 0 and 1)
3. No correlation (slope=0)
4. Negative (slope between -1 and 0)
5. Perfect negative (slope=-1)

The presence of correlation between two variables may be due to three reasons:
1.One variable being the cause of the other. The cause is called “subject” or
“independent” variable, while the effect is called “dependent” variable.
2.Both variables being the result of a common cause. That is, the correlation that exists
between two variables is due to their being related to some third force.
 118

Example:
Let X1= ESLCE result
Y1= rate of surviving in the University
Y2= the rate of getting a scholar ship.

Both X1&Y1 and X1&Y2 have high positive correlation, likewiseY1 & Y2 have positive
correlation but they are not directly related, but they are related to each other via X1.

3.Chance: The correlation that arises by chance is called spurious correlation.

Examples:
 Price of teff in Addis Ababa and grade of students in USA.
 Weight of individuals in Ethiopia and income of individuals in Kenya.

Therefore, while interpreting correlation coefficient, it is necessary to see if there is any

likelihood of any relationship existing between variables under study.
The correlation coefficient between X and Y denoted by r is given by

r
 ( X i  X )(Yi  Y ) and the short cut formula is
 ( X i  X )  (Yi  Y )
2 2

n XY  ( X )( Y )
r
[n X 2  ( X ) 2 ] [n Y 2  ( Y ) 2
r
 XY  nXY
[ X 2  nX 2 ] [ Y 2  nY 2 ]
Remark: Always this r lies between -1 and 1 inclusively and it is also symmetric.
Interpretation of r
1.Perfect positive linear relationship ( if r  1)
2.Some Positive linear relationship ( if r is between 0 and 1)
3.No linear relationship ( if r  0)
4.Some Negative linear relationship ( if r is between -1 and 0)
5.Perfect negative linear relationship ( if r  1)

Examples:

1. Calculate the simple correlation between mid-semester and final exam scores of 10 students
(both out of 50)
 119

Student Mid Sem.Exam Final Sem.Exam

(X) (Y)
1 31 31
2 23 29
3 41 34
4 32 35
5 29 25
6 33 35
7 28 33
8 31 42
9 31 31
10 33 34
Solution:
n  10, X  31.2, Y  32.9, X 2  973.4, Y 2  1082.4
 XY  10331, X 2
 9920, Y 2
 11003

r
 XY  nXY
[ X 2  nX 2 ] [ Y 2  nY 2 ]
10331  10(31.2)(32.9)

(9920  10(973.4)) (11003  10(1082.4))
66.2
  0.363
182.5
This means mid semester exam and final exam scores have a slightly positive correlation.

Exercise The following data were collected from a certain household on the monthly income (X)
and consumption (Y) for the past 10 months. Compute the simple correlation coefficient.

X: 650 654 720 456 536 853 735 650 536 666
Y: 450 523 235 398 500 632 500 635 450 360

 The above formula and procedure is only applicable on quantitative data, but when we have
qualitative data like efficiency, honesty, intelligence, etc we calculate what is called
Spearman’s rank correlation coefficient as follows:

Steps
i. Rank the different items in X and Y.
ii. Find the difference of the ranks in a pair , denote them by Di
iii. Use the following formula
 120

6 Di
2
rs  1 
n(n 2  1)
Where rs  coefficient of rank correlation
D  the difference between paired ranks
n  the number of pairs
Example:
Aster and Almaz were asked to rank 7 different types of lipsticks, see if there is correlation
between the tests of the ladies.

Lipstick types A B C D E F G
Aster 2 1 4 3 5 7 6
Almaz 1 3 2 4 5 6 7

Solution:
D2
X Y R1-R2
6 Di
2
6(12)
(R1) (R2) (D)  rs  1   1   0.786
2 1 1 1 n(n  1)
2
7 ( 48)
1 3 -2 4
4 2 2 4
3 4 -1 1
5 5 0 0 Yes, there is positive correlation.
7 6 1 1
6 7 -1 1 Simple Linear Regression
Total 12 - Simple linear regression refers to the linear
relationship between two variables
- We usually denote the dependent variable by Y and the independent variable by X.
- A simple regression line is the line fitted to the points plotted in the scatter diagram, which
would describe the average relationship between the two variables. Therefore, to see the type
of relationship, it is advisable to prepare scatter plot before fitting the model.
Y    X  
Where :Y  Dependent var iable
- The linear model is: X  independent var iable
  Re gression cons tan t
  regression slope
  random disturbanc e term
Y ~ N (   X ,  2 )
 ~ N (0,  2 )

- To estimate the parameters (  and  ) we have several methods:

 121

 The free hand method

 The semi-average method
 The least square method
 The maximum likelihood method
 The method of moments
 Bayesian estimation technique.

- The above model is estimated by: Yˆ  a  bX

Where a is a constant which gives the value of Y when X=0 .It is called the Y-intercept. b is
a constant indicating the slope of the regression line, and it gives a measure of the change in Y
for a unit change in X. It is also regression coefficient of Y on X.
- a and b are found by minimizing SSE      (Yi  Yˆi )
2 2

Where : Yi  observed value

Yˆi  estimated value  a  bX i
And this method is known as OLS (ordinary least square)
- Minimizing SSE    gives
2

b
 ( X i  X )(Yi  Y )   XY  nXY
 ( X i  X )2  X 2  nX 2
a  Y  bX

Example 1: The following data shows the score of 12 students for Accounting and Statistics
Examinations.

a) Calculate a simple correlation coefficient

b) Fit a regression equation of Statistics on Accounting using least square estimates.
c) Predict the score of Statistics if the score of accounting is 85.

Accounting Statistics 2
X Y2 XY
X Y
1 74.00 81.00 5476.00 6561.00 5994.00
2 93.00 86.00 8649.00 7396.00 7998.00
3 55.00 67.00 3025.00 4489.00 3685.00
4 41.00 35.00 1681.00 1225.00 1435.00
 122

5 23.00 30.00 529.00 900.00 690.00

6 92.00 100.00 8464.00 10000.00 9200.00
7 64.00 55.00 4096.00 3025.00 3520.00
8 40.00 52.00 1600.00 2704.00 2080.00
9 71.00 76.00 5041.00 5776.00 5396.00
10 33.00 24.00 1089.00 576.00 792.00
11 30.00 48.00 900.00 2304.00 1440.00
12 71.00 87.00 5041.00 7569.00 6177.00
Total 687.00 741.00 45591.00 52525.00 48407.00
Mean 57.25 61.75
a)

The Coefficient of Correlation (r) has a value of 0.92. This indicates that the two variables are
positively correlated (Y increases as X increases).

b)

where:

 Yˆ  7.0194  0.9560 X is the estimated regression line.

 123

c) Insert X=85 in the estimated regression line.

Yˆ  7.0194  0.9560 X
 7.0194  0.9560(85)  88.28
Exercise: A car rental agency is interested in studying the relationship between the distance
driven in kilometer (Y) and the maintenance cost for their cars (X in birr). The following
summarized information is given based on samples of size 5.
2
i 1 X i  147,000,000 i 1Yi  314
5 5 2

i 1 X i  23,000 , i 1Yi  36 , i 1 X i Yi  212, 000

5 5 5

a) Find the least squares regression equation of Y on X

b) Compute the correlation coefficient and interpret it.
c) Estimate the maintenance cost of a car which has been driven for 6 km

- To know how far the regression equation has been able to explain the variation in Y we use a
2
measure called coefficient of determination ( r )

 (Yˆ  Y ) 2
i.e r  2

 (Y  Y ) 2
Where r  the simple correlation coefficient.

- r 2 gives the proportion of the variation in Y explained by the regression of Y on X.

- 1  r 2 gives the unexplained proportion and is called coefficient of indetermination.

Example: For the above problem (example 1): r  0.9194

 r 2  0.8453  84.53% of the variation in Y is explained and only 15.47% remains

unexplained and it will be accounted by the random term.

o Covariance of X and Y measures the co-variability of X and Y together. It is denoted by

S XY and given by

SX Y 
 ( X i  X )(Yi  Y )   XY  nXY
n 1 n 1
 124

o Next we will see the relationship between the coefficients.

2
S S
i. r  XY  r 2  X2 Y 2
S X SY S X SY
bS rS
ii. r X b Y
SY SX
o When we fit the regression of X on Y , we interchange X and Y in all formulas, i.e. we fit

Xˆ  a1  b1Y

b1 
 XY  nXY
 Y 2  nY 2
b1SY
a1  X  b1Y , r
SX
Here X is dependent and Y is independent.

Choice of Dependent and Independent variable

- In correlation analysis there is no need of identifying the dependent and independent
variable, because r is symmetric. But in regression analysis
If bYX is the regression coefficient of Y on X
bXY is the regression coefficient of X on Y
bYX S X bXY SY
Then r   r 2  bYX * bXY
SY SX
- Moreover, bYX and bX Y are completely different numerically as well as conceptually.

- Let us consider three cases concerning these coefficients.

1. If the correlation is perfect positive, i.e. r  1 then the b values reciprocals of each other.
2. If S X  SY , then irrespective of the value of r the b values are equal, i.e.
r  bYX  bXY ( but this is unlikely case)
3. The most important case is when S X  SY and r  1, here the b values are not equal
or reciprocals to each other, but rather the two lines differ , intersecting at the common
point ( X , Y )
 Thus to determine if a regression equation is X on Y or Y on X , we have to use
the formula r 2  bYX * bXY
 125

 If r [1,1] , then our assumption is correct

 If r [1,1] , then our assumption is wrong

Example: The regression line between height (X) in inches and weight (Y) in lbs of male
students are:
4Y  15 X  530  0 and
20 X  3Y  975  0
Determine which is regression of Y on X and X on Y

Solution
We will assume one of the equation as regression of X on Y and the other as Y on X and
calculate r

Assume 4Y  15 X  530  0 is regression of X on Y

20 X  3Y  975  0 is regression of Y on X
Then write these in the standard form.
530 4 4
4Y  15 X  530  0  X   Y  bXY 
15 15 15
 975 20 20
20 X  3Y  975  0  Y   X  bYX 
3 3 3
 4  20 
 r 2  bXY * bYX      1.78  1 ,
 15  3 
This is impossible (contradiction). Hence our assumption is not correct. Thus
4Y  15 X  530  0 is regression of Y on X
20 X  3Y  975  0 is regression of X on Y
To verify:
 530 15 15
4Y  15 X  530  0  Y   X  bYX 
4 4 4
975 3 3
20 X  3Y  975  0  X   Y  bXY 
20 20 20

 15  3  9
 r 2  bYX * bXY       0,1
 4  20  16