Chapter 5

Interval estimation
and testing

5.1 INTRODUCTION
As stated earlier, the reason for taking a sample is to obtain information
about the unknown characteristics of a population or process. Two types
of population characteristics are of special interest in business: the mean of
a variable, and the proportion of elements belonging to a category. After
the sample has been taken, these are estimated by the corresponding sample
mean and proportion. In the majority of studies in business, these estimates
(there could be many of them in a typical marketing survey) are all that is
required of the sample.
The key to good estimates lies in the design of the sample, and this
takes place before the sample is actually selected. A well-designed sample,
we have argued, should be: (a) randomly selected from the population or
process of interest; and (b) large enough so that the estimates will have the
desired degree of accuracy.
Interval estimates may be used in place of, or as a supplement to, the
“point” estimates we have encountered up to now. Rather than state, after
a sample is taken, that a population characteristic is estimated to be such-
and-such a number, it may on occasion be more informative to state that
the characteristic is estimated to be in such-and-such an interval. Not any
arbitrary interval, obviously, will do. If we are to make interval estimates, we
want assurance that our statement will be correct (that is, that the interval
will contain the population characteristic) with a given probability. Such
intervals are known as conÞdence intervals and are described in the following
two sections.
The remainder of the chapter deals with statistical tests. A statistical
test is a rule–a prescription, if you like–for deciding which of two state-
ments concerning an unknown population characteristic is true. No decision
rule (statistical or other) is infallible. The attractive feature of statisti-
cal tests is that they allow the desision-maker to control the probability of
making an error judged (by the decision-maker) to be the more serious.

c Peter Tryfos, 2001.

°
 2 Chapter 5: Interval estimation and testing

5.2 INTERVAL ESTIMATION

In our discussion so far, we have argued that it is reasonable to use the
sample mean (X̄) as an estimator of the population mean (µ) of a variable,
and the sample proportion (R) as an estimator of the population proportion
(π) of a category. When the sample is taken, the numerical values of X̄ and
R are the estimates of µ and π respectively.
We know well, of course, that in general it is unlikely these estimates
will equal µ or π exactly. Although we may say “we estimate µ and π to be
10.3 and 0.12,” we certainly do not mean we consider these to be necessarily
the true values of µ and π.
Instead of saying, for example, “µ is X̄,” we could say “µ is in the
interval from X̄ − c to X̄ + c.” This is an interval around X̄, usually
abbreviated as X̄ ± c, with c to be speciÞed. X̄ ± c is an interval estimator
of µ, and we may prefer it to the ordinary estimator X̄.
For any choice of c, the statement “µ is in the interval X̄ ± c” is some-
times correct (that is, the interval contains µ), sometimes not. Forming an
arbitrary interval is not at all difficult, but forming an interval having a
given probability of containing µ is not easy.
An interval estimator of a population characteristic which can be said
to contain the population characteristic with given probability is called a
conÞdence interval, and the given probability the conÞdence level. Approx-
imate conÞdence intervals for the population mean of a variable and for the
population proportion of a category, applicable when the population and
sample sizes are large, are described in the box that follows.
Note that the intervals (5.1) and (5.2) make c equal to Uα/2 SX̄ and
Uα/2 SR respectively, and can be calculated once the sample observations
are available. The probability 1 − α is speciÞed in advance. It could be large
(e.g., 99%, 90%) or small (e.g., 30%, 10%), as desired. We shall continue to
interpret “large N and n” to mean “larger than 200 and 100 respectively,”
according to the rule of thumb of the previous chapter. One should bear
in mind, however, that it is a rule of thumb; more accurately, we should
be saying that the probability of the interval containing the population
characteristic tends to 1 − α as n and N get larger.
To show that the intervals (5.1) and (5.2) contain π and µ respectively
with probability approximately 1 − α we need a property of large samples
described in the next section; the proof is also in that section.
Before commenting on the features of these intervals, let us illustrate
how they are calculated.
 5.2 Interval estimation 3

If the population (N ) and the sample size (n) are large, the prob-
ability is approximately 1 − α that the interval

R ± Uα/2 SR (5.1)

contains the population proportion (π). Likewise, the probability is

approximately 1 − α that the interval

X̄ ± Uα/2 SX̄ (5.2)

contains the population mean (µ). These intervals are known as 100(1−
α)% conÞdence intervals for π and µ respectively, and (1 − α) is known
as the conÞdence level. Similar conÞdence intervals for N π and N µ are

N (R ± Uα/2 SR ) and N (X̄ ± Uα/2 SX̄ ) (5.3)

In the above expressions, S 2 is the sample variance of the variable X:

n
1X
S2 = (Xi − X̄)2 , (5.4)
n i=1

and r r
S2 N − n R(1 − R) N − n
SX̄ = , SR = . (5.5)
n N −1 n N −1
Uα/2 for selected (1 − α) are given in Table 5.1.

Table 5.1
Uα/2 for selected 1 − α
1−α Uα/2 1−α Uα/2
0.99 2.576 0.80 1.282
0.95 1.960 0.60 0.842
0.90 1.645 0.50 0.674

Example 5.1 A random sample of n = 500 households from a population

4 Chapter 5: Interval estimation and testing

of N = 100, 000 was taken. 42% of the sampled households said they would
buy an experimental product. We calculate
r
(0.42)(1 − 0.42) 100000 − 500
SR = = 0.022.
500 100000 − 1
Suppose that a 90% conÞdence interval is desired for the proportion of
households in the population who intend to buy. Then, 1 − α = 0.90,
and Uα/2 = 1.645. Thus, Uα/2 SR = (1.645)(0.022) = 0.036. The desired
interval is from (0.420 − 0.036) to (0.420 + 0.036), or from 38.4% to 45.6%.
A 90% conÞdence interval for the number of households who intend to
buy is (100,000)(0.420 ± 0.036), or from about 38,400 to 45,600.

Example 5.2 A random sample of n = 800 households in a town revealed

that the average weekly household expenditure on food was X̄ = $95 with a
standard deviation of S = 34. (Just in case the meaning of S is not clear, 34
is the standard deviation of the 800 responses to the question, “How much
is your household’s weekly food expenditure?” The average of these 800
responses is, of course, 95.) There are N = 10, 000 households in the town.
A 95% conÞdence interval for the average weekly food expenditure by all
households in the town would be given by
r
(34)2 10000 − 800
(95) ± (1.96) .
800 10000 − 1
This is the interval from (95 − 2.26) to (95 + 2.26), or from $92.74 to $97.26.
A 95% conÞdence interval for the total household expenditures on food
in the city is (10,000)(95 ± 2.26), or from about $927,400 to $972,600.

One should not interpret a calculated conÞdence interval as containing

the population characteristic with the stated probability. For instance, it is
not correct to say in Example 5.1 that the probability is 90% that the pop-
ulation proportion is between 38.4 and 45.6%. The population proportion
is a given number, and either lies in that interval or does not–we do not
know which is true. It is the procedure by which the intervals are calculated
which is correct with probability 90%.
To understand this more clearly, imagine it is possible to select a large
number of samples, all of the same size n, from the given population. After
each sample is taken, imagine calculating the interval X̄ ± Uα/2 SX̄ and
stating that µ lies in this interval. Because X̄ and SX̄ will vary from sample
to sample, the location and width of the intervals will vary. See Figure 5.1.
The claim of Equation (5.2) is that, in the long run, 100(1 − α)% of
these intervals will contain (will “bracket”) µ. (In Figure 5.1, all but one of
 5.2 Interval estimation 5

Figure 5.1
Interval estimates

the Þve conÞdence intervals shown–the exception being that from Sample
No. 4–contain µ.) The statement that the population mean lies in the
interval X̄ ± Uα/2 SX̄ will therefore be correct in 100(1 − α)% of the samples
in the long run.
The calculated conÞdence interval, however, conveys some information
about the reliability of the sample estimate. The intervals (5.1) and (5.2),
it will be noted, are formed around the estimate of the population charac-
teristic, and their width depends on the size of the sample, the variability
of the observations, and the desired conÞdence level.
As an illustration, consider the width of the conÞdence interval for a
population mean: it is approximately twice the amount
r
1 1
c = Uα/2 S 2( − ).
n N

The smaller the conÞdence level, 1 − α, the smaller is Uα/2 . Other things
being equal, therefore, the larger the sample size, the smaller the variability
of the sample observations (as measured by S 2 ), and the lower the conÞdence
level, the narrower the interval–which is as one intuitively expects.
The reader should always bear in mind that a conÞdence interval is not
a necessary appendage to the sample estimate. In most business samples
designed to obtain estimates of a large number of population characteristics,
a routine reporting of all possible conÞdence intervals would most certainly
confuse rather than enlighten.*

* In addition to the conÞdence intervals based on large samples described

in this section, it is possible to construct a conÞdence interval for a popula-
 6 Chapter 5: Interval estimation and testing

5.3 FURTHER PROPERTIES OF LARGE SAMPLES

In order to understand the derivation of the conÞdence intervals in the pre-
vious section, and of the statistical tests described in the next section, we
must state and brießy explain two more properties of large samples.

If the population (N ) and the sample (n) are large, the probability
distributions of the random variables

X̄ − µ R−π
U1 = , U2 = , (5.6)
SX̄ SR

are approximately standard normal.

As usual, X̄ is the sample mean and µ the population mean of a variable

X, while R is the sample and π the population relative frequency of a
2
category. SX̄ , deÞned in Equations (5.5), is an estimator of the variance of
X̄,
σ2 N − n
V ar(X̄) = , (5.7)
n N −1
obtained by replacing the population variance (σ 2 ) with the sample variance
(S 2 ). Similarly, SR
2
, also deÞned in Equations (5.5), is an estimator of the
variance of R,
π(1 − π) N − n
V ar(R) = , (5.8)
n N −1
obtained by replacing the population proportion π by the sample proportion
R.
Each U is a “standardized” random variable, measuring the difference
between X̄ (or R) and its expected value µ (or π), with the difference ex-
pressed in estimated standard deviations of its distribution. Because X̄, R,
SX̄ , and SR will vary from sample to sample, so will the U ’s, which are
functions of these random variables.
Consider Example 4.1 of Chapter 4 by way of illustration. In this case,
N = 10, π = 0.7, and µ = 0.9. Refer to Table 4.1 of Chapter 4 and suppose

tion proportion in a sample of any size with or without replacement. Similar,

“any-sample-size” conÞdence intervals for the population mean of a variable,
however, can be formed only in the case where the population distribution
of the variable is of a special form (for example, normal, Poisson, etc.) and
the sample is with replacement.
 5.3 Further properties of large samples 7

that Outcome No. 3 occurs, that is, X1 = 0, X2 = 0, and X3 = 2. It follows

that X̄ = 2/3 = 0.667, R = 1/3 = 0.333, and the sample variance is

1X 1
S2 = (Xi − X̄)2 = [(0 − 0.667)2 + (0 − 0.667)2 + (2 − 0.667)2 ] = 0.889.
n 3

Therefore,
r r
0.889 10 − 3 (0.333)(1 − 0.333) 10 − 3
SX̄ = = 0.480, SR = = 0.240,
3 10 − 1 3 10 − 1

0.667 − 0.9
U1 = = −0.485,
0.480
and
0.333 − 0.7
U2 = = −1.529.
0.240
For example, U1 = −0.485 indicates that the observed value of the sample
mean lies 0.485 estimated standard deviations to the left of its expected
value, and U2 = −1.529 shows that the observed value of R lies 1.529 esti-
mated standard deviations to the left of the mean of its distribution.
For each sample outcome listed in Table 4.1 of Chapter 4, there cor-
responds a value of U1 and U2 . The U ’s are random variables, and their
probability distributions can be constructed in the usual way. We shall not
do so, however, because both N and n are small, and no useful purpose will
be served by continuing the calculations.
Consider now Figure 5.2. The histograms depict the probability distri-
butions of U1 for samples of size 3, 10, 20, and 30 with replacement drawn
from the population of Example 4.1 of Chapter 4. Each of these probabil-
ity distributions was routinely constructed from a list of all possible sample
outcomes, their probabilities, and the associated values of U1 ; a computer
program was used to perform the long and tedious numerical calculations.
When S 2 = 0, which occurs when all the sample observations are identical,
U1 has an indeterminate value; the probability of an indeterminate value
is indicated by a separate bar in the histograms of Figure 5.2. Note that
the probability of an indeterminate value decreases as the sample size in-
creases. Compare the histograms with the superimposed standard normal
distribution. As the sample size increases, the standard normal distribution
provides an increasingly better approximation to the actual distribution of
U1 .
What is illustrated for U1 and sampling with replacement can be shown
to hold for U2 and for sampling without replacement: for large N and n
(n < N ), the actual probability distribution of any Ui may be approximated
by the standard normal distribution.
8 Chapter 5: Interval estimation and testing

Figure 5.2
Distribution of U1
 5.3 Further properties of large samples 9

It has not been made clear so far just how large N and n must be
for the normal approximations to be satisfactory. Unfortunately, no exact
guidance can be given. In statistical theory, the large-sample properties
are proven for inÞnitely large N and n. Empirical investigations, however,
show the approximation to be surprisingly good in some cases for samples
as low as 30 or smaller. On the other hand, there are other cases where the
approximation is poor even for very large samples.
Roughly speaking, the more symmetric the population distribution of
the variable X (in the case of U1 ) or the closer π is to 0.5 (in the case of
U2 ), the smaller is the sample size required.
As a rule of thumb, the reader would probably be safe in assuming that
for n ≥ 100 and N ≥ 200–conditions easily satisÞed in most samples used
in business–the normal approximation is satisfactory, and that the results
based on this approximation (to be described in the following sections) are
applicable.
We would now like to show that the probability that the interval (5.2) will
contain (“bracket”) µ is approximately 1 − α. The proof concerning the in-
terval (5.1) is very similar, and is left as an exercise for the reader.
¶ For large N and n, the distribution of the ratio U1 = (X̄ − µ)/SX̄ is ap-
proximately the standard normal. Let Uα/2 be a number such that the probability
that the standard normal variable will exceed that number is α/2. By the sym-
metry of the normal distribution, P r(−Uα/2 ≤ U1 ≤ Uα/2 ) = 1 − α. Substituting
(X̄ − µ)/SR for U1 , we have

X̄ − µ
P r(−Uα/2 ≤ ≤ Uα/2 ) = 1 − α. (5.9)
SX̄

Consider the expression within the parentheses. We apply the same two rules
concerning inequalities stated in Chapter 2.* Multiplying all three terms by (the
positive) SX̄ , we get −Uα/2 SX̄ ≤ X̄ − µ ≤ Uα/2 SX̄ . Multiplying these terms by
−1, we reverse the inequalities and get Uα/2 SX̄ ≥ µ − X̄ ≥ −Uα/2 SX̄ . Adding
X̄ to all three terms gives X̄ − Uα/2 SX̄ ≤ µ ≤ X̄ + Uα/2 SX̄ . The Þrst expression
implies the last. All we have done is write the original inequalities in a different
form. Therefore, Equation (5.9) implies that

P r(X̄ − Uα/2 SX̄ ≤ µ ≤ X̄ + Uα/2 SX̄ ) = 1 − α. (5.10)

In words, the probability is 1 − α that the interval from (X̄ − Uα/2 SX̄ ) to (X̄ +
Uα/2 SX̄ ) contains the population mean. This interval is therefore a 100(1 − α)%
conÞdence interval for µ, and the proof is complete. ¶

It may be noted that the type of conÞdence interval described in these

sections (two-sided, symmetric, and centered around the sample estimate) is

* If a ≤ X ≤ b then ca ≤ cX ≤ cb for c > 0, or ca ≥ cX ≥ cb for c < 0. Also,

if a ≤ X ≤ b then a + c ≤ X + c ≤ b + c for any c.
 10 Chapter 5: Interval estimation and testing

by no means unique. Instead of an interval of the form X̄−c ≤ µ ≤ X̄+c, one

can construct conÞdence intervals of the form µ ≤ c, µ ≥ c, or c1 ≤ µ ≤ c2
(where c, c1 , and c2 depend on the sample observations), all having the same
probability of containing µ.
For example, we may begin with the observation that the following
statements are approximately true when the random variable U has the
unit normal distribution:

P r(−1.28 ≤ U ≤ 1.28) = P r(−1.04 ≤ U ≤ 1.64) =

= P r(U ≤ 0.84) = P r(U ≥ −0.84) = 0.80.

Substituting (X̄ − µ)/SX̄ for U and following the same approach as above,
we Þnd that:

P r(X̄ − 1.28SX̄ ≤ µ ≤ X̄ + 1.28SX̄ ) =

= P r(X̄ − 1.64SX̄ ≤ µ ≤ X̄ + 1.04SX̄ ) =
= P r(µ ≤ X̄ + 0.84SX̄ ) =
= P r(µ ≥ X̄ − 0.84SX̄ ) = 0.80.

Therefore, the intervals:

(a) from X̄ − 1.28SX̄ to X̄ + 1.28SX̄ ,

(b) from X̄ − 1.64SX̄ to X̄ + 1.04SX̄ ,
(c) from X̄ − 0.84SX̄ to +∞,
(d) from −∞ to X̄ + 0.84SX̄

are all approximate 80% conÞdence intervals for µ. An inÞnite number of

type (b) intervals can be constructed by varying appropriately the limits
of the probability statement. The symmetric interval (a) is preferable to
any asymmetric one of type (b) because it is narrower: in our example, the
width of (a) is (2)(1.28)SX̄ = 2.56SX̄ , while that of (b) is 1.64SX̄ + 1.04SX̄
= 2.68SX̄ . Due to a property of the standard normal distribution, the same
is true for all asymmetric intervals (b). The choice between (a), (c), and
(d), however, is not obvious and depends on the speciÞc problem.

5.4 UNDERSTANDING STATISTICAL TESTS

In this section, we examine with some care two types of statistical tests.
As will soon be clear, there are many such tests. It is not our intention to
examine them all exhaustively, but to identify in the two tests of this section
the features and problems that most tests have in common.

Example 5.3 A certain brand of light bulbs has a “rated life of 1,000
hours.” This rating is assigned by the manufacturer. The small print on the
 5.4 Understanding statistical tests 11

package explains that the average life of this brand of bulbs is warranted
to be more than 1,000 hours. (It is understood that the life of individual
light bulbs varies; some of this variability could perhaps be removed by
better quality control, but some is inherent in the manufacturing process
and cannot be eliminated.)
A batch of 10,000 bulbs has been produced. Before it is shipped out,
a test must be made to determine if the quality of the batch is consistent
with the rating. At issue, therefore, is whether or not the average life of the
bulbs in the batch is more than 1,000 hours. Measuring the life of all 10,000
bulbs is obviously out of the question since the measurement is destructive:
life is measured by letting the bulb burn until it burns out. A sample must
be used.
In one sense, the problem is simple. The manufacturer could select a
random sample of light bulbs, measure their life duration, and calculate the
average life of the bulbs in the sample. If this sample average is greater
than 1,000 hours, the conclusion could be that the batch is Good (that
is, the quality rating is justiÞed); the batch would then be released. On
the other hand, if the sample average is less than or equal to 1,000 hours,
the conclusion could be that the batch is Bad (does not meet the quality
standard) and its release would be withheld.
Let µ be the (unknown) average life of the bulbs in the batch. The issue
then is whether or not µ is more than 1,000 hours. A statistician would say
that the problem involves two hypotheses concerning µ:

H1 : µ ≤ 1000 (Batch is Bad)

H2 : µ > 1000 (Batch is Good)

Obviously, if H1 is true, H2 is false, and vice versa.

A decision must be based on a sample. The previously suggested deci-
sion rule can be expressed as

Accept H1 (i.e., conclude batch is Bad) if X̄ ≤ 1000,

(5.11)
Reject H1 (i.e., conclude batch is Good) if X̄ > 1000,

where X̄ is the average life of n randomly selected light bulbs. (The terms
“accept” and “reject” are part of the established terminology but need not
be taken literally. By “accept H1 ” we mean “decide in favor of H1 ,” and by
“reject H1 ” “decide in favor of H2 .”)
This decision rule may be perfectly sensible, but it is not at all ßexible.
Let us therefore make it a little more general, as follows:

Accept H1 if X̄ ≤ c,
(5.12)
Reject H1 if X̄ > c,
12 Chapter 5: Interval estimation and testing

Table 5.2
Hypotheses, actions, and consequences
Acts
Events Accept H1 Reject H1
H1 is true No error Type I error
H1 is false Type II error No error

where c is a number to be determined (as a special case, it could be made

equal to 1,000). In this version, the decision rule recommends accepting H1
when X̄ is “small,” and rejecting H1 when it is “large”–with c distinguish-
ing small from large values of X̄.
Table 5.2 shows the two hypotheses, the two possible decisions, and the
associated consequences. If H1 is true and is accepted, or false and rejected,
no error is made. But if H1 is true and is rejected, or false and accepted, an
error is made. These are two different kinds of error, and are distinguished
in Table 5.2 as errors of Type I and II.
In general, by a Type I error we shall understand that of rejecting H1
when H1 is true–whatever the meaning of H1 happens to be. In our ex-
ample, a Type I error means that a Bad batch is declared Good. Possible
consequences include customer complaints, an erosion of the manufacturer’s
quality image, etc.
A Type II error, on the other hand, is that of accepting H1 when it is
false–again, whatever the meaning of H1 may be. In the present example,
a Type II error means that a Good batch is declared Bad. It could result in
the lot being scrapped or reworked.
The ideal decision rule should guarantee that the probabilities of both
types of error are zero. A little thought will lead one to the conclusion that
this is impossible, unless the sample is the entire population. It is possible,
however, to formulate a rule guaranteeing that the probability of one type
of error–whichever is the more serious–does not exceed a given number.
Let us suppose that a Type I error (Bad lot declared Good) is considered
more serious, and that the probability of such an error should not exceed
20%. The problem, therefore, is to Þnd a decision rule ensuring that the
probability of a Type I error does not exceed 20%.
Our problem is in fact a special case of that described in the box that
follows, and a decision rule having the required property is given by (5.14).
Before explaining the derivation of this decision rule, let us illustrate
its implementation.
In our problem, µ0 = 1, 000 and α = 0.20. Suppose that a random
sample of n = 100 light bulbs is taken, that the average life of the bulbs
in the sample is X̄ = 1, 010 hours, and their standard deviation S = 80.
 5.4 Understanding statistical tests 13

When N and n are large, the approximate decision rule for testing

H1 : µ ≤ µ0 ,
(5.13)
H2 : µ > µ0 ,

µ0 a given number, so that the probability of Type I error does not

exceed α and that of Type II error does not exceed 1 − α, is to

Accept H1 if X̄ ≤ µ0 + Uα SX̄ ,
(5.14)
Reject H1 if X̄ > µ0 + Uα SX̄ .

Selected values of Uα are given in Table 5.3. SX̄ is given in Equations

(5.5).

Calculate
r
(80)2 10000 − 100
SX̄ = = 7.96.
100 10000 − 1

Uα = U0.20 = 0.842, and µ0 + Uα SX̄ = (1000) + (0.842)(7.96) = 1,006.70.

Since X̄ = 1010 > 1006.7, the decision rule (5.14) requires that H1 be
rejected–in other words, that the batch be declared Good.

It is clear that the decision rule (5.14) implies a c in (5.12) equal to

µ0 + Uα SX̄ . Note that when α = 0.50 (a value appropriate when the two
types of error are considered equally serious), Uα = 0, and the rule becomes:
Accept H1 if X̄ < µ0 . In our example, the decision rule (5.12) is therefore
appropriate in just such a situation.

¶ We would now like to show that the decision rule (5.14) is indeed such that
the probability of Type I error does not exceed α, and that of Type II error does
not exceed 1 − α.
By deÞnition, a Type I error occurs when µ ≤ µ0 , and a Type II error when
µ > µ0 . The proof, then, is in two steps, as follows.
First, suppose that the true value of µ is less than or equal to µ0 , i.e., H1 is
14 Chapter 5: Interval estimation and testing

true. The probability of a Type I error is the probability of rejecting H1 , i.e.,

¡ X̄ − µ0 ¢
P r(X̄ > µ0 + Uα SX̄ ) = P r > Uα
SX̄
¡ X̄ − µ + µ − µ0 ¢
= Pr > Uα
SX̄
¡ X̄ − µ µ − µ0 ¢
= Pr > Uα −
SX̄ SX̄
= P r(U1 > Uα − e)
= P r(U1 > U 00 ),
where U1 , for large N and n, has approximately the standard normal distribution
and e is less than or equal to zero. Therefore, U 00 ≥ Uα ; as illustrated in Figure
5.3, P r(U1 > U 00 ) ≤ α. We conclude that the probability of a Type I error does
not exceed α.

Figure 5.3
Standard normal distribution

For the second step of the proof, suppose that µ is greater than µ0 , i.e., H1
is false. The probability of Type II error is the probability of accepting H1 , or
¡ X̄ − µ0 ¢
P r(X̄ ≤ µ0 + Uα SX̄ ) = P r ≤ Uα
SX̄
¡ X̄ − µ + µ − µ0 ¢
= Pr ≤ Uα
SX̄
¡ X̄ − µ µ − µ0 ¢
= Pr ≤ Uα −
SX̄ SX̄
= P r(U1 ≤ Uα − e)
= P r(U1 ≤ U 0 )
where e is positive and U 0 < Uα . As illustrated in Figure 5.3, P r(U1 ≤ U 0 ) < 1−α,
that is, the probability of Type II error cannot exceed 1 − α, and the proof is
complete. ¶
 5.4 Understanding statistical tests 15

The choice of α, hence also of Uα , should depend on the relative seri-

ousness of the two types of error. The more serious the consequences of a
Type I error in relation to those of a Type II error, the lower α should be,
and vice versa. α can be any number between 0 and 1.
If it is desired to restrict the probability of a Type II error in a given
situation to be, say, no more than 0.10, this can be accomplished by setting
α = 1 − 0.10 = 0.90. If the two errors are equally serious, set α = 0.50, in
which case the decision rule is simpliÞed as we have already seen.
A Þnal note. The ≤ in the statement of H1 appears to suggest that
this hypothesis is somewhat “heavier” than H2 . In fact, the case µ = µ0 ,
say, µ = 1 means precise equality (e.g., µ = 1 means precisely 1, and not
0.999999 or some such number close to 1). Unless it happens that µ is
discrete, it makes no difference which hypothesis the case µ = µ0 is assigned
to or whether it is omitted entirely.
A similar test concerning a population proportion can be developed
along the same lines, and is summarized by expressions (5.15) and (5.16) in
the box that follows.

For large N and n, the approximate decision rule for testing

H1 : π ≤ π0 ,
(5.15)
H2 : π > π0 ,

(π0 a given number) so that the probability of Type I error does not
exceed α and that of Type II error does not exceed 1 − α is to

Accept H1 if R ≤ π0 + Uα SR
(5.16)
Reject H1 if R > π0 + Uα SR .

Selected values of Uα are given in Table 5.3. SR is given in Equations

(5.5).

Example 5.4 Suppose that by a “rating of 1,000 hours,” the light bulb
manufacturer means “relatively few bulbs last less than 1,000 hours.” (Many
quality standards in manufacturing have this type of interpretation.) More
speciÞcally, suppose that “relatively few” means “5% or less.” Let π rep-
resent the proportion of light bulbs in the batch which last less than 1,000
 16 Chapter 5: Interval estimation and testing

hours. We wish to test

H1 : π ≤ 0.05 (Batch is Good),
H2 : π > 0.05 (Batch is Bad).

In this case, a Type I error occurs when a Good lot is declared Bad, and
a Type II error when a Bad lot is declared Good. Suppose it is the latter
error which is considered more serious, and that the probability of Type II
error should not exceed 10%. We set α = 1 − 0.10 = 0.90, which implies
Uα = −1.282.
A random sample without replacement of size n = 100 is taken from
the batch of N = 10,000 light bulbs. Four of the 100 bulbs (R = 0.04) are
found to last less than 1,000 hours. We calculate
r
(0.04)(1 − 0.04) 10000 − 100
SR = = 0.019,
100 10000 − 1
and then π0 + Uα SR = (0.050) − (1.282)(0.019) = 0.026. Since R > 0.026,
H1 is rejected and the batch is declared Bad.
This somewhat surprising conclusion is dictated by the choice of α
which guards against the occurrence of a Type II error. If the two errors
were considered equally serious, or if the Type I error were thought to be
more serious, the opposite conclusion would have been reached.

5.5 THE STRUCTURE OF STATISTICAL TESTS

The two problems described in the previous section have several features in
common.
There are two possible acts (release or withhold the batch). The con-
sequences of these acts depend on the unknown value of a population or
process characteristic (µ or π). The possible values of this characteristic
are partitioned into two mutually exclusive and exhaustive sets or intervals
(e.g., µ ≤ 1000, µ > 1000), so that one act is optimal if the characteristic
is in the Þrst interval, and the other is optimal if it is in the second one.
No information is available concerning the likelihood of the possible values
of the population characteristic; a sample is taken to obtain some such in-
formation. There is no additional information concerning the consequences
of the two acts beyond the fact that each is optimal for certain values of
the population characteristic, and that perhaps some consequences are more
serious than others.
These are indeed the common features of situations for which most
statistical tests are designed.
Let us call the population characteristic θ (it could be µ, π, or something
else), and the two sets or intervals S1 and S2 . Denote the two possible acts
 5.5 The structure of statistical tests 17

a1 and a2 ; assume a1 is preferable if θ is in S1 , and a2 is preferable if θ is in

S2 .
Attention, therefore, focusses on two questions: Is θ in S1 ? Is it in
S2 ? These questions correspond to the hypotheses: H1 : θ in S1 , and
H2 : θ in S2 . The two acts can be labelled a1 : Accept H1 , and a2 :
Reject H1 . And, since the consequences cannot be made more precise,
they are indicated broadly as “No error,” “Type I error,” “Type II error,”
as shown in Table 5.4.

Table 5.4
Events, acts, and consequences
Acts
Events a1 : Accept H1 a2 : Reject H1
H1 : θ in S1 No error Type I error
H2 : θ in S2 Type II error No error

A test is a decision rule prescribing when to choose a1 and when a2 .

Clearly, there are many possible decision rules for any given hypotheses. At
a minimum, it is reasonable to require that the rule be based on the sample
(that is, not to be arbitrary as, for instance, the rule: Reject H1 if the moon
is full), and that it allow control of the probabilities of the two types of
error.
All the tests of this chapter have the property that the probability of a
Type I error will not exceed α, and that of a Type II error will not exceed
1 − α, no matter what the value of θ happens to be. (Such a test, however,
may not be unique; that is, there may be several tests with this property
for given hypotheses.)
There are, obviously, many possible hypotheses, and, consequently,
many statistical tests. There are, to begin with, many population character-
istics which conceivably could matter in a given situation (e.g., the median,
the mode, or the variance of a variable, to name just three). There are many
types of hypotheses which can be formulated (one pair that comes quickly
to mind: H1 : µ1 ≤ µ ≤ µ2 , H2 : µ > µ2 or µ < µ1 , where µ1 and µ2 are
given numbers). It is possible to formulate hypotheses involving jointly two
or more population characteristics (e.g., that the population mean and vari-
ance are in a speciÞed region), or more than two sets or intervals of values of
a population characteristic (although multiple-action problems rapidly be-
come quite complicated). And, Þnally, there may be several ways of testing
the given hypotheses.
Not all hypotheses are testable, however, and not all tests are designed
with business requirements in mind. Consider, for example, the oft-cited
 18 Chapter 5: Interval estimation and testing

hypotheses H1 : µ = µ0 vs. H2 : µ 6= µ0 . µ is the population mean of a

variable, and µ0 a given number–say, 10. A test of H1 : µ = 10 is precisely
that–a test of the hypothesis that µ equals 10 precisely (not 10.00000001, or
9.9999999, or some other such number). Unless µ takes integer values only,
it is very unlikely in typical business problems that the unknown population
mean of a variable equals a given number exactly, and thus we know in
advance that H1 is false. H1 will be rejected if the sample size is large
enough. The non-rejection of H1 is simply a consequence of a small sample.
(This shortcoming is shared by some of the more frequently quoted tests, as
will soon be explained.)
In the remainder of this chapter, we describe (but do not derive) a
number of frequently cited tests. Their implementation–despite a Þrst ap-
pearance of complexity–is quite easy. When implementing a test, however,
it is worthwhile to heed three broad admonitions.
• Two-action tests are designed so that the probability of a Type I error
does not exceed α, and that of a Type II error does not exceed 1 − α.
A Type I error, by deÞnition, is the rejection of H1 when it is true; a
Type II error is the acceptance of H1 when it is false. The real meaning
and consequences of these errors, however, depend on the situation.
When applying a given test, therefore, one should understand what
these errors imply and choose α or 1 − α appropriately. There is an
unfortunate tendency in practice (occasionally reinforced by the manner
in which tables of critical values are presented) to select a low value of
α automatically and indiscriminately, without regard to the situation.
• With the exception of Section 5.6, all the tests of this chapter are ap-
proximate. Although the calculations are carried out to three decimal
places in order to avoid ambiguity, in practice one should not lose sight
of the fact that the conditions “large N and n,” as well as the value of
α, are not and cannot be speciÞed precisely.
• A statistical test is a device for controlling the probability of one type
of error, and not–as many erroneously believe–a means of proving or
disproving one or the other hypothesis. It is quite possible that two
parties, presented with the same information, will reach diametrically
opposite conclusions if they guard against different kinds of error.

5.6 INFERENCE FOR SPECIAL POPULATION DISTRIBUTIONS

When the form of the distribution of a variable in the population or process

is mathematically tractable, it is sometimes possible to construct conÞdence
intervals or tests that do not require the sample size to be large–they apply
to small as well as large samples. Perhaps the simplest and most elegant
results are obtained when this distribution is normal.
 5.7 Tests concerning the form of the population distribution 19

If the distribution of the measurement X of an independent process is

normal with mean µ and standard deviation σ, the probability distribution
of the ratio
X̄ − µ
U1 = ,
SX̄
in samples of any size, n, is the t distribution with parameter ν = n − 1.
The t distribution is deÞned in Appendix 2. It is symmetric, centered
at 0, and looks much like the standard normal. As usual, X̄ is the sample
mean, and SX̄ is deÞned by Equations (5.5).
From the table of the t distribution in Appendix 4G, it is possible
to determine a number–call it Tα/2 –such that P r(−Tα/2 ≤ U1 ≤ Tα/2 ) =
1−α. (For example, if n = 5 and 1−α = 0.90, then ν = 4 and T0.05 = 2.132.)
Proceeding exactly as in Section 5.3, we Þnd that

P r(X̄ − Tα/2 SX̄ ≤ µ ≤ X̄ + Tα/2 SX̄ ).

It follows that the interval

X̄ ± Tα/2 SX̄
is a 100(1 − α)% conÞdence interval for the process mean of measurement
X, µ, valid for any sample size.
When n > 30, Tα/2 is approximately equal to Uα/2 , and this special
result merges with the more general ones of Section 5.2.
Following a procedure similar to that of Section 5.4, it is also fairly
easy to show that the decision rule for testing H1 : µ ≤ µo vs. H2 : µ > µo ,
where µ is the mean of an independent normal process, is given by (5.14),
except that Uα is replaced by Tα . Tα is a number such that the probability
that a variable having the t distribution with parameter ν will exceed that
number is α; see Appendix 4G.
The same results apply if the sample is random and with replacement
from a Þnite population in which the distribution of the variable X is normal.
The reader should be warned against uncritically accepting the nor-
mal assumption and making inferences with small samples on the basis of
this assumption alone. Because the normal distribution is mathematically
tractable and lends itself to many elegant results, the tendency of textbooks
and researchers alike is to give this distribution a more prominent role than
it deserves: in business practice, few population or process distributions
are normal. The test of the next section may be used to determine if the
normality assumption is valid in a given case.

5.7 TESTS CONCERNING THE FORM OF THE POPULATION

DISTRIBUTION
Suppose that the elements of a population or process are classiÞed into s
mutually exclusive and collectively exhaustive categories C1 , C2 , . . ., Cs . We
20 Chapter 5: Interval estimation and testing

wish to test the hypothesis that the population relative frequencies are equal
to speciÞed numbers π1o , π2o , . . . , πso , against the alternative hypothesis that
at least one of the πi is not equal to the speciÞed number.

Example 5.5 A casino is testing a die for fairness. The die is fair if the
six faces show up with equal relative frequencies (πi ) in the long run.

H1 : π1 = 1/6, π2 = 1/6, . . . , π6 = 1/6 (Die is fair)

H2 : πi 6= 1/6 for at least one πi (Die not fair)

The die will be rolled a number of times, and the relative frequencies with
which the six faces show up observed. How should one decide whether the
die is fair or not?

An approximate decision rule for this situation is described in the fol-

lowing box.

For a large random sample with replacement or sample from an

independent process, the approximate decision rule for testing

H1 : π1 = π1o , π2 = π2o , . . . , πs = πso ,

(5.17)
H2 : At least one πi 6= πio ,

so that the probability of a Type I error equals α and that of a Type

II error does not exceed 1 − α, is to:

Accept H1 if V ≤ Vα;s−1 ,
(5.18)
Reject H1 if V > Vα;s−1 ,

where
s
X (Ri − πio )2
V =n . (5.19)
i=1
πio

The πio are given numbers, Ri is the relative frequency of category Ci

in the sample, n is the sample size, and Vα;s−1 is as given in Appendix
4H.

Before illustrating the calculations, let us sketch the derivation of this

test. If H1 is true, that is, if the πi equal πio , the Ri will tend to be close
 5.7 Tests concerning the form of the population distribution 21

to the πio , and the value of V will tend to be close to zero. If, on the other
hand, the true πi deviate from the hypothesized values under H1 , πio , the
Ri will tend to deviate from the πio , and V will tend to be greater than
zero. We want to accept H1 when V is small, and to reject it when V is
large:
Accept H1 if V ≤ c,
Reject H1 if V > c.
c, the “critical value” of this test, distinguishes “small” from “large” V
values. As always, we would like to determine c so that the probability of a
Type I error does not exceed α, and that of a Type II error 1 − α.
In mathematical statistics, it is shown that if H1 is true and the sam-
ple large and with replacement, the probability distribution of V is approxi-
mately chi-square with parameter λ = s− 1. (The deÞnition of this distribu-
tion can be found in Appendix 2.) Denote by Vα;s−1 the number such that
the probability of a chi-square random variable with parameter λ = s − 1
exceeding that number is α. These numbers are tabulated in Appendix 4H.
It follows that if c is made equal to this number, the probability of a Type I
error will not exceed α (in fact, will equal α). Intuitively, it should be clear
that the probability of a Type II error is greatest when the πi are very close
to the πio , at which point the probability of a Type II error is 1 − α.
For calculations by hand, any one of the following versions of Equation
(5.19) may be used:
s
X Ps s
(Fi − nπio )2 ¡ i=1 Ri2 ¢ X Fi2
V = =n −1 = − n. (5.20)
i=1
nπio πio i=1
nπio

Example 5.5 (Continued) The die is rolled 60 times. The frequencies

with which the six faces of the die showed up are shown in Table 5.5. The
question, once again, is: Is the die fair?

Table 5.5
Observations, Example 5.5
Face (Ci ) Frequency (Fi ) Relative frequency (Ri )
1 11 0.183
2 9 0.150
3 12 0.200
4 8 0.133
5 9 0.150
6 11 0.183
1.000
22 Chapter 5: Interval estimation and testing

Let us examine the consequences of the two types of error. A Type

I error in this case is associated with rejecting a fair die; a Type II error
means accepting a bad die. We may assume the consequences to be far
more serious in the latter case, and we should want 1 − α small, meaning
α must be large. For the sake of this illustration, let us suppose α = 0.99.
Since s = 6, we get from Appendix 4H V0.99;5 = 0.554. The V statistic is
calculated as follows:
X s
Fi2
V = −n
i=1
nπ io

1
= [(11)2 + (9)2 + · · · + (11)2 ] − 60
(60)(1/6)
= 1.20.
Since V > 0.554, the hypothesis that the die is fair is rejected.

Related to the above is the test of the hypothesis that the population
distribution has a given mathematical form with certain (unspeciÞed) pa-
rameter values. An example would be the test of the hypothesis that the
distribution of the measurement of an independent process is normal with
some values of the parameters µ and σ. (If the parameter values as well
as the form of the distribution are speciÞed–for example, by the hypothe-
sis that the population distribution is normal with parameter µ = 2.0 and
σ = 0.3–the appropriate test is the earlier one in this section.)
In advanced mathematical statistics texts, it is shown that, when the
random sample is large and with replacement, or large and from an indepen-
dent process, the decision rule for this test is given by (5.18), except that the
critical value is Vα;s−k−1 , where k is the number of estimated parameters of
the hypothesized distribution. In calculating the V statistic, πio are proba-
bilities determined under the assumption that the form of the distribution
is that speciÞed by the hypothesis, with appropriately estimated parameter
values.

Example 5.6 Five hundred items produced by an independent manufac-

turing process were selected and the number of defects in each determined.
The results are shown in columns (1), (2), and (3) of Table 5.6. Column (4)
will be explained shortly.
The question is: Can it be supposed that the distribution of the number
of defects per item in the process is Poisson?
If the distribution of the number of defects is Poisson with parameter
m, the proportions of items with x = 0, 1, 2, . . . defects in the long run would
be given by
mx e−x
p(x) = . (5.21)
x!
 5.7 Tests concerning the form of the population distribution 23

Table 5.6
Distribution of number of defects, Example 5.6
Number of Number of Proportion of Proportion
defects items, Fi items, Ri under H1 , πio
(1) (2) (3) (4)
0 463 0.926 0.9213
1 34 0.068 0.0755
2 2 0.004 0.0031
3 1 0.002 0.0001
500 1.000 1.0000

The single parameter of the Poisson distribution, m, happens to be

equal to the mean of the distribution (see Table 2.3 of Chapter 2). It makes
sense, therefore, to estimate m by the sample mean–in this case, the average
number of defects per item, X̄:

X̄ = (0)(0.926) + (1)(0.068) + (2)(0.004) + (3)(0.002) = 0.082.

After substituting m = 0.082 in Equation (5.21), we can calculate p(x)

for x = 0, 1, 2, . . . . A computer program was used to generate the Þgures
shown in column (4) of Table 5.6. The last entry in column (4) is the sum
of p(x) for x ≥ 3.
To test the hypotheses:

H1 : Process distribution of defects is Poisson,

H2 : Process distribution of defects is not Poisson,

we Þrst calculate the V statistic applying (5.19):

£ (0.9260 − 0.9213)2 (0.0020 − 0.0001)2

V = 500 + ··· + = 18.57.
0.9213 0.0001
Let us examine the possible consequences of the two types of error. A Type
I error is associated with rejecting the hypothesis that the distribution is
Poisson when in fact it is; a Type II error is that associated with accepting
the Poisson hypothesis when it is false. Now, we did not explain why we
wanted to test this hypothesis in the Þrst place, but the reader may be
willing to accept without elaboration that certain elegant and useful results
in statistical quality control are applicable in the case where the process
distribution is Poisson. Applying these results in the belief that H1 is true
when in fact it is not is probably the more serious error. For this illustration,
let us suppose that the probability of a Type II error should not exceed 0.25,
implying α = 0.75.
 24 Chapter 5: Interval estimation and testing

One parameter (m) was estimated under H1 . Therefore, k = 1, s − k −

1 = 4 − 1 − 1 = 2, and Vα;s−k−1 = V0.75;2 = 0.575. Since V > Vα;s−k−1 , H1
is rejected.

The Þrst of the tests of this section, it should be noted, speciÞes H1

precisely. Although it apppears appropriate for the example used to illus-
trate it (testing a die for fairness), it should not be applied in cases where
it is unlikely that the population or process satisÞes H1 exactly. In such a
case, there would be no point carrying out the test–it would either reject
H1 , as expected, or would conÞrm that the sample size is not large enough
to reject it.
The second test of this section, though less restrictive than the Þrst,
shares the same problem. For it may well be that the issue is not whether the
process is exactly–say–Poisson (a feature which few real-world processes
possess precisely), but whether it approximates the Poisson well enough for
the purposes of the study.

5.8 TESTING THE INDEPENDENCE OF TWO ATTRIBUTES

OR VARIABLES
Imagine the elements of a population classiÞed into s categories, values, or
intervals A1 , A2 , . . ., As , according to one attribute or variable, and into t
categories or intervals B1 , B2 , . . ., Bt , according to a second. Let πij be the
proportion of elements in the population which fall into categories Ai and
Bj . Table 5.7 illustrates the notation.

Table 5.7
Population joint distribution
First Second attribute
attribute · · · Bj ··· Total
··· ··· ··· ··· ···
Ai · · · πij ··· πi.
··· ··· ··· ··· ···
Total · · · π.j ··· 1.0

The marginal relative frequency of category Ai is denoted by πi. , while

that of Bj by π.j .
Two attributes or variables are independent if all joint relative frequen-
cies equal the product of the corresponding marginals, that is, if πij = πi. π.j
for all i and j. (The notation is different, but the deÞnition of independence
is identical to that in Chapters 1 and 2.)
 5.8 Testing the independence of two attributes or variables 25

Instead of a population of Þnite size, we may have an independent

process, the elements of which are characterized by two measurements (for
example, the length and width of rectangular metal plates produced by a
stamping machine). By analogy with the one-measurement case, we shall
say that a two-measurement process is independent if the joint distribution
of measurements does not vary from element to element, and if one element’s
joint measurements are not related to those of any other element.
How, then, are we to determine whether or not the attributes or mea-
surements are independent?
Suppose that a sample is drawn from such a population or process. Let
Rij be the proportion of the sample elements falling into categories Ai and
Bj . Also, let Ri. and R.j be the marginal relative frequencies of categories
Ai and Bj respectively, as illustrated in Table 5.8.

Table 5.8
Sample joint distribution
First Second attribute
attribute · · · Bj ··· Total
··· ··· ··· ··· ···
Ai · · · Rij ··· Ri.
··· ··· ··· ··· ···
Total · · · R.j ··· 1.0

Before stating the decision rule for this case, let us illustrate the nota-
tion with an example.

Example 5.7 It has been argued that the main difference between younger
and older drivers is the tendency of the former to have relatively more acci-
dents and claims. However, again according to this argument, the amount
of the claim is determined largely by the circumstances of the accident and
should be unrelated to the age of the driver. On the other hand, if younger
drivers tend to drive larger and more expensive cars faster, the severity of
any accident in which they are involved will tend to be greater, and the
claim amount should be related to age.
The most recent 500 claims received by an automobile insurance com-
pany were analyzed, and the joint relative frequency distribution of claim
amount and age of the insured was obtained, as shown in Table 5.9.
For example, in 1.4% of the 500 selected Þles the insured was under 30
and the amount of the claim was over $10,000.
The question is: Is the amount of the claim independent of the age of
the insured?
26 Chapter 5: Interval estimation and testing

Table 5.9
Distribution of age and claim amount,
Example 5.7
Claim amount ($000)
Age Under 1 1 to 10 Over 10 Total
Under 30 0.362 0.016 0.014 0.392
30 to 50 0.318 0.024 0.012 0.354
Over 50 0.240 0.008 0.006 0.254
Total 0.920 0.048 0.032 1.000

The receipt of claims by the insurance company could be assumed to

form an independent random process (why?) characterized by several mea-
surements: the amount of the claim, the type of claim, the sex and age of
the insured, etc.
The hypotheses to be examined are:

H1 : Age and claim amount are independent

H2 : They are not

Note that the hypotheses refer to the process from which the Þles are
selected. A simple calculation will show that the deÞnition of independence
is not satisÞed for the sample.

In the general case, if the attributes or measurements are independent,

the sample joint relative frequencies, Rij , will tend to be equal to the product
of the corresponding marginals, Ri. R.j , in which case the statistic

Xs X t
(Rij − Ri. R.j )2
V =n (5.22)
i=1 j=1
Ri. R.j

will tend to be close to zero. If, on the other hand, the attributes are not
independent, the Rij will tend to deviate from the Ri. R.j , and the value of
V will tend to be large.
A reasonable decision rule, then, is to accept the hypothesis of inde-
pendence when V is small, and to reject it when V is large. This decision
rule can be written as

Accept H1 if V ≤ c,
Reject H1 if V > c,

where c distinguishes small from large values of V .

 5.8 Testing the independence of two attributes or variables 27

For a large sample from an independent process or with replace-

ment, the approximate decision rule for testing

H1 : πij = πi. π.j , for all i and j,

(5.23)
H2 : πij 6= πi. π.j , for some i and j,

so that the probability of a Type I error equals α and that of a Type

II error does not exceed 1 − α, is to

Accept H1 if V ≤ Vα;(s−1)(t−1) ,
(5.24)
Reject H1 if V > Vα;(s−1)(t−1) ,

where V is given by (5.22), s is the number of rows and t the number

of columns of Tables 5.7 and 5.8, and Vα;(s−1)(t−1) is obtained from
Appendix 4H.

In mathematical statistics, it is shown that if H1 is true, the sample

large, and either random and with replacement from a Þnite population, or
selected from an independent process, the probability distribution of V is
approximately chi-square with parameter λ equal to (s − 1)(t − 1). The
probability of a Type I error will be equal to α, if c is set to Vα;(s−1)(t−1) .
This decision rule is summarized in the box.
Two comments should be made before illustrating this test. First, ob-
serve that the V statistic is closely related to the coefficient of association
(P ) of Chapter 1, which measures the strength of the relationship between
two attributes. In fact, V = n(q − 1)P , where q is the smaller of the number
of rows and columns of Table 5.8.
Second, since joint relative frequencies are related to frequencies, Rij =
Fij /n, V can also be written as

X X (nRij − nRi. R.j )2 X X (Fij − Eij )2

V = = , (5.25)
i j
nRi. R.j i j
Eij

where, in the second term, Eij = nRi. R.j .

Example 5.7 (Continued) The V statistic is calculated from the data in

 28 Chapter 5: Interval estimation and testing

Table 5.9, as follows:

£ (0.362 − 0.920 × 0.392)2 (0.006 − 0.032 × 0.254)2 ¤

V = 500 + ··· +
| 0.920 × 0.392 {z 0.032 × 0.254 }
9 terms
= 2.938.

Now, a Type I error in this case is the rejection of the hypothesis of

independence when it is in fact true. A Type II error is the conclusion that
age and claim amount are unrelated when in fact they are not.
If we happen to advocate that age and claim amount are independent,
and prefer to guard against a Type I error by making α small–say, α = 0.10,
we would Þnd in Appendix 4H that V0.10;4 = 7.779, and would accept H1 .
If we happen to advocate the opposite theory and like to guard against
a Type II error, we would prefer 1 − α low–say, 1 − α = 0.10. This implies
α = 0.90, and V0.90;4 = 1.064. The decision then would be to reject the
hypothesis of independence.
The two conclusions are based on the same evidence. The opposing
interests lead to an opposite interpretation of this evidence.

It must be kept in mind that the hypothesis of independence is very

strict. For H1 in (5.23) to be true, all πij must equal exactly πi. π.j (to
the tenth, thousandth, . . . decimal place). In the real world, it is very
rare that two variables or attributes satisfy precisely this deÞnition. In
most situations, in other words, the test of this section serves no useful
purpose because it is applied to a hypothesis that is known to be false;
any failure to reject the hypothesis is simply a consequence of not having a
large enough sample. In such cases, the question should not be whether or
not the variables or attributes are related, but whether the relationship is
strong enough for practical purposes. In the last example, for instance, an
“acceptance” of the hypothesis that age and claim amount are independent
is in practical terms “non-rejection.” It should be interpreted to mean that
the sample is not large enough to reject the hypothesis of independence. If
the test rejects the hypothesis of independence, one should examine whether
or not the relationship is too weak to matter.

5.9 COMPARING TWO OR MORE POPULATIONS

Suppose that a random sample is drawn from each of t populations.
The elements of each sample are classiÞed into s common categories, values,
or intervals C1 , C2 , . . . , Cs , according to the same attribute or variable. Let
nj be the size of the jth sample, Fij the number, and Rij the proportion
 5.9 Comparing two or more populations 29

Table 5.10
Frequency distributions of t samples
Aggregate
Category, Sample Sample Sample relative
Ci 1 2 ··· t Total frequencies, Ri.
C1 F11 F12 ··· F1t F1. R1. = F1. /n
C2 F21 F22 ··· F2t F2. R2. = F2. /n
··· ··· ··· ··· ··· ··· ···
Cs Fs1 Fs2 ··· Fst Fs. Rs. = Fs. /n
Total n1 n2 ··· nt n 1

of elements of the jth sample that fall into category Ci . The notation is
illustrated in Table 5.10.
The problem is: Are the populations from which the samples were
drawn identical with respect to the distribution of the variable or attribute?

Example 5.8 A consumer product testing organization investigated the

durability of four brands of alkaline D batteries. The organization instructed
each of its shoppers (widely scattered across the country) to purchase a few
batteries at a large store. The batteries were then shipped to the central
laboratory, where they were tested for durability under uniform conditions.
In all, 100 Brand A, 120 Brand B, 80 Brand C, and 200 brand D batteries
were tested, with the results shown in Table 5.11.

Table 5.11
Four brands of batteries compared
Life (hours) Brand A Brand B Brand C Brand D Total Ri.
Under 19 hr 11 10 8 21 50 0.100
19 to 20 hr 29 35 25 63 152 0.304
20 to 21 hr 42 50 30 77 199 0.398
Over 21 hr 18 25 17 39 99 0.198
Total 100 120 80 200 500 1.000

For example, of the 200 Brand D batteries tested, 21 lasted under 19

hours, 63 lasted between 19 and 20 hours, and so on. Note that the sample
sizes are not (and need not be) the same. The question is: Are the batteries
30 Chapter 5: Interval estimation and testing

of the four brands identical in terms of durability? In different words: Are

the population distributions of battery life the same for all brands?
The sample, it will be noted, is not random in the strict sense. There is
obviously no list of batteries from which a random sample could be selected,
and physical randomization is impossible. On the other hand, it could be
argued that the organization’s method produces an essentially random sam-
ple, in that the batteries actually selected had the same chance of appearing
in the sample as any other batteries. This will be assumed here, but it is
instructive for the reader to consider reasons why the sample should not be
considered an essentially random one.

Returning to the general case, the problem is to test the hypothesis

that the populations from which the t samples are drawn have identical
distributions; that is, that the proportions of elements of each population
which fall into categories C1 , C2 , . . . , Cs are the same–and equal to, say,
π1 , π2 , . . . , πs .
If the populations are identical, the t separate samples may as well be
combined into one, assumed drawn from the pooled populations. Three
consequences may be expected to follow: (a) the aggregate relative frequen-
cies of categories C1 , C2 , . . ., Cs in the combined sample (R1. = F1. /n,
R2. = F2. /n, . . ., Rs. = Fs. /n) will tend to be close to the population rela-
tive frequencies π1 , π2 , . . ., πs of these categories; (b) the expected frequency
of category i in sample j, nj πi , will tend to be close to Eij = nj Ri. , and
this, in turn, will tend to be close to the observed sample frequency, Fij ;
Þnally, (c) the value of the statistic

Xs X t
(Fij − Eij )2
V = (5.26)
i=1 j=1
Eij

will tend to be small.

On the other hand, if the populations are not identical, the sample
frequencies will tend to deviate from the Eij , and V will tend to be large.
It is reasonable, therefore, to accept H1 when V is small, and to reject it
when V is large.
In mathematical statistics, it is shown that when the populations are
identical and the samples are large and with replacement, the probability
distribution of the statistic V is approximately chi-square with parameter
equal to (s−1)(t−1). The decision rule summarized in the box that follows is
based on this last result. In the box, πij denotes the proportion of elements
of population j that belong to category Ci .
The test of (5.27) is identical to that of the previous section. It should
be realized, however, that Tables 5.8 and 5.10 represent two different sit-
uations. Table 5.8 shows a single sample of size n drawn from a single
 5.9 Comparing two or more populations 31

If each of t random samples with replacement or from independent

processes is large, the approximate decision rule for testing

 π11 = π12 = · · · = π1t = π1

π21 = π22 = · · · = π2t = π2
H1 : (5.27)

 ···
πs1 = πs2 = ··· πst = πs

against H2 that at least one of the above equalities does not hold, so
that the probability of a Type I error equals α and that of a Type II
error does not exceed 1 − α, is to

Accept H1 if V ≤ Vα;(s−1)(t−1) ,
(5.28)
Reject H1 if V > Vα;(s−1)(t−1) .

V is given by (5.26), and Vα;(s−1)(t−1) is tabulated in Appendix 4H.

population, classiÞed jointly according to two attributes. Table 5.10, on the

other hand, shows the results of t random samples, each drawn from a dif-
ferent population, but all commonly classiÞed into s categories according to
a single attribute.

Example 5.8 (Continued) If the four samples come from populations hav-
ing identical distributions, the estimates of the common population relative
frequencies of the four categories are shown in the last column of Table 5.11:
R1. = 0.100, R2. = 0.304, R3. = 0.398, and R4. = 0.198. The estimated
expected frequencies under H1 , Eij = nj Ri. , are shown in Table 5.12.

Table 5.12
Estimated expected frequencies, Eij , Example 5.8
Category Brand A Brand B Brand C Brand D
Under 19 hr 10.0 12.00 8.00 20.0
19 to 20 hr 30.4 36.48 24.32 60.8
20 to 21 hr 39.8 47.76 31.84 79.6
Over 21 hr 19.8 23.76 15.84 39.6
Total 100.0 120.00 80.00 300.0
 32 Chapter 5: Interval estimation and testing

If the samples come from identical populations, the estimate of the

proportion of batteries that last, say, between 20 and 21 hours is 0.398.
Therefore, 39.8 of the 100 Brand A batteries, 47.76 of the 120 B batteries,
31.84 of the 80 C batteries, and 79.6 of the 200 D batteries can be expected
to last between 20 and 21 hours.
We now calculate the V statistic (the Fij come from Table 5.11, and
the Eij from Table 5.12):

X4 X 4
(Fij − Eij )2
V =
i=1 j=1
Eij
(11 − 10)2 (29 − 30.4)2 (39 − 39.6)2
= + + ··· +
| 10 30.4 {z 39.6 }
16 terms
= 1.447.

Let us examine the consequences of the two types of error. If the

distribution of battery life is identical for all four brands, and the consumer
organization reaches the opposite conclusion, a Type I error is made. A Type
II error, in this case, is the conclusion that the distributions are identical
when in fact they are not. Which error is more serious? It is difficult to
say, is it not? Let us assume that the two errors are thought to be equally
serious, in which case α may be set at 0.50.
Since s = 4 and t = 4, (s − 1)(t − 1) = 9. For α = 0.50, Appendix 4H
gives V0.50;9 = 8.343.
The value of the test statistic is smaller than the critical value of the
test, and we accept the hypothesis that the population distributions of the
four brands are identical in terms of durability.

Once again note that H1 is a very precise statement. To assert that the
population distributions are identical means, strictly speaking, that there is
absolutely no difference among the relative frequencies of any category. In
this example, the acceptance of H1 is best interpreted as non-rejection: the
samples are not large enough to determine conclusively that the populations
are not identical.

5.10 COMPARING TWO POPULATION MEANS

The test described in the last section is that of the hypothesis that the
populations from which the samples are drawn have identical distributions.
Occasionally, however, it may be useful to examine whether or not a certain
characteristic only (such as the mean or the variance) of these population
distributions is the same. Obviously, these are different questions. For ex-
ample, if two distributions have equal means but different variances, their
 5.10 Comparing two population means 33

distributions are different. On the other hand, if two distributions are iden-
tical, their means, variances, and all other characteristics are the same.
We shall describe here a test of the hypothesis that two population
means are equal. The Þrst population consists of N1 , and the second of N2
elements. Let µ1 and µ2 be the means of a certain variable X in the two
populations. At issue is whether or not µ1 = µ2 (alternatively, whether or
not µ1 − µ2 = 0).
A random sample is drawn from each of the two populations. Let ni ,
X̄i , and Si2 be the size of the sample, and the mean and variance of variable
X in the sample drawn from population i (i = 1, 2).
A statistic on which the test can be based is the difference between the
two sample means, X̄1 − X̄2 . If the two population means are equal, the
difference (X̄1 − X̄2 ) will tend to be close to 0; if not, (X̄1 − X̄2 ) will tend to
deviate from 0. We shall want to accept H1 when (X̄1 − X̄2 ) is close to 0,
and to reject it when (X̄1 − X̄2 ) is not close, in either direction, to 0. This
decision rule can be written as

Accept H1 if |X̄1 − X̄2 | ≤ c,

Reject H1 if |X̄1 − X̄2 | > c,

where c, the “critical value” of the test, is a number distinguishing “close

to 0” from “not close enough to 0” values. As usual, the problem is to
determine c so that the probability of a Type I error does not exceed α,
while that of a Type II does not exceed 1 − α. It can be shown that, when
the population and sample sizes are large, the approximate decision rule is
that described in the box that follows.

Example 5.9 The consumer product testing organization of Example

5.8 also investigated the durability of two major brands of heavy-duty zinc
chloride batteries. A total of 100 Brand A and 120 Brand B batteries were
purchased throughout the country and tested under uniform conditions with
the following results.

Brand A Brand B
Sample size n1 = 100 n2 = 120
Average life (hours) X̄1 = 14.25 X̄2 = 10.63
Standard deviation S1 = 1.92 S2 = 1.81

At issue is whether or not the average life of all Brand A batteries is

identical to that of all Brand B batteries.
Assuming, for the reasons stated in Example 5.8, that the samples can
be considered essentially random, we calculate Þrst:

|X̄1 − X̄2 | = 3.62.

34 Chapter 5: Interval estimation and testing

For random samples without replacement, and large n1 , N1 , n2 ,

and N2 , the approximate decision rule for testing

H 1 : µ1 = µ2 ,
(5.29)
H2 : µ1 6= µ2 ,

so that P r(Type I error) = α and P r(Type II error) ≤ 1 − α, is to

Accept H1 if |X̄1 − X̄2 | ≤ c,

(5.30)
Reject H1 if |X̄1 − X̄2 | > c,

where s
S12 N1 − n1 S 2 N2 − n2
c = Uα/2 + 2 . (5.31)
n1 N1 − 1 n2 N2 − 1

Selected values of Uα/2 are listed in Table 5.1.

If, as we assumed in Example 5.8, the two types of error are thought to
be equally serious and α = 0.50, then Uα/2 = 0.674. Next, we calculate c.
Since the Ni are very large, we may set (Ni − ni )/(Ni − 1) ≈ 1, in which
case
r
(1.92)2 (1.81)2 √
c = 0.674 + = 0.674 0.064 = (0.674)(0.253) = 0.171.
100 120
Since 3.62 > 0.171, we reject the hypothesis that the average life of the two
brands is the same.

Once again, the strictness of H1 should be kept in mind: for H1 to be

true, µ1 must equal µ2 precisely. There is no point in applying this test in
situations where this equality cannot possibly be true.
It is sometimes useful to construct a conÞdence interval for the differ-
ence between two population means. In advanced statistics, it is shown that
for large n1 , N1 , n2 , and N2 , a 100(1 − α)% conÞdence interval for µ1 − µ2
is given by
(X̄1 − X̄2 ) ± c, (5.32)
with c given by Equation (5.31) above. The meaning is similar to that of all
conÞdence intervals: the probability is 1−α that the interval (5.32) contains
the true difference µ1 − µ2 .
 5.11 A non-parametric test 35

Example 5.9 (Continued) A 95% conÞdence interval for the difference in

the average life of the two brands of batteries is

(3.62) ± (1.96)(0.253),

or from about 3.12 to 4.12 hours.

5.11 A NON-PARAMETRIC TEST

Let us consider once again the problem of comparing the durability of two
brands of batteries. Let us suppose that samples of Þve Brand A and
6 Brand B batteries were obtained in the manner described in Example
5.8 and tested, with the results shown in columns (1) and (2) of Table 5.13.
(The samples are deliberately small to illustrate simply the calculations that
follow.)

Table 5.13
Samples of two brands of batteries
Life (hours) Ranks
Brand A Brand B Brand A Brand B
(1) (2) (3) (4)
8.9 1
9.9 2
10.1 3
10.2 4
10.5 5
12.4 6
12.7 7
13.6 8
14.2 9
16.8 10
17.9 11
n1 =5 n2 = 6 R1 = 39 R2 = 27

Let us combine the two samples, arrange the observations in increasing

order of magnitude, and assign to them the ranks shown in columns (3) and
(4) of Table 5.13.
If the two populations from which the samples were drawn are identical
(that is, if the distributions of battery life are the same for the two brands),
36 Chapter 5: Interval estimation and testing

the ranks 1 to 11 would tend to be evenly distributed between the two

brands; neither column (3) nor column (4) should show a concentration of
low or high ranks. In fact, if the populations are identical, the sum of the
ranks in column (3) should behave like the sum of Þve integers selected
at random and without replacement from among the Þrst 11 (and that of
column (4) like the sum of 6 such integers).
In general, assume the sample from the Þrst population consists of n1
and that from the second of n2 elements. Let R1 denote the sum of ranks
assigned to the Þrst sample. If the two populations are identical, it can be
shown that the expected value and variance of R1 in random samples of size
n1 and n2 with replacement are:

n1 (n1 + n2 + 1)
E(R1 ) = ,
2
n1 n2 (n1 + n2 + 1)
V ar(R1 ) = .
12
It can also be shown, again if the two populations are identical, that the
probability distribution of R1 is approximately normal for large n1 and
n2 . (It does not matter which population and sample is called “the Þrst”
and which “the second.” R1 could refer to either, but n1 should be the
corresponding sample size.)
If the populations are identical, the observed R1 should tend to be
close to E(R1 ); if not, R1 should tend to deviate from E(R1 ). Therefore, we
should want to reject the hypothesis that the two populations are identical
when R1 deviates substantially from E(R1 ). The terms “close” and “sub-
stantially” are given precise meaning in the decision rule shown in the box
that follows.
To illustrate this test, let us return to our example and assume–as
we did in Example 5.8–that α = 0.50. Type I and II errors are assumed
equally serious, and Uα/2 = 0.674. Calculate

(5)(5 + 6 + 1) (5)(6)(5 + 6 + 1)
E(R1 ) = = 30, V ar(R1 ) = = 30.
2 12
p √
E(R1 ) ± Uα/2 V ar(R1 ) is (30) ± (0.674) 30, or from about 23.61 to 33.69.
Since the observed R1 = 39 lies outside this interval, H1 is rejected.
This decision rule is known as the Mann-Whitney or Wilcoxon (MWW)
test. Three points are worth noting.
First, the MWW test is an alternative to the chi-square test of Section
5.9. As we noted in Section 5.5, there may be more than one test of given
hypotheses. This raises the question of how to determine which of the alter-
native tests is better–a question that has received considerable attention
in the statistical literature, but which we shall not pursue here.
 Problems 37

If sampling is with replacement or from an independent process,

and the sample sizes n1 and n2 are large, the approximate decision rule
for testing

H1 : Two populations are identical

H2 : Two populations are not identical

so that the probability of a Type I error equals α and that of a Type

II error does not exceed 1 − α, is to
p
Accept H1 if E(R1 ) − Uα/2 V ar(R1 ) ≤ R1
p
≤ E(R1 ) + Uα/2 V ar(R1 ),
Reject H1 if otherwise.

Selected values of Uα/2 are given in Table 5.1.

The second point is that there is a MWW test applicable for any–not
only large–n1 and n2 ; the critical values for this test are obtained from
special tables, but the test itself is based on R1 . The noteworthy feature
of this test is that it does not require the two population distributions to
have a particular form. As we remarked earlier, decision rules applicable to
samples of any size usually make this requirement; they are sometimes called
parametric tests, in the sense that the hypotheses deal with parameters of
population distributions of a given type. By contrast, the MWW is one of
many non-parametric tests.
The Þnal point worth noting is that H1 , once again, is a precise state-
ment. What is being tested is the hypothesis that the two population distri-
butions are identical. This non-parametric test, therefore, shares the main
shortcoming of the chi-square test of Section 5.9. Since it is very rare that
two population distributions in the business world are exactly alike, care
should be taken that rejection of H1 be followed by an appraisal of the
magnitude of the differences between the two populations.

PROBLEMS
5.1 A random sample of size n = 200 was drawn without replacement from a
population of size N = 1, 000. The sample mean of a variable is X̄ = 150, and the
sample variance is S 2 = 280.
(a) Calculate the 99%, 95%, 90%, and 50% symmetric conÞdence intervals
for the population mean of the variable, µ. Brießy interpret these intervals.
38 Chapter 5: Interval estimation and testing

(b) Calculate the 99%, 95%, 90%, and 50% symmetric conÞdence for the
total value of the variable in the population. Brießy interpret these intervals.
(c) Do (a) and (b) under the assumption that the sample is with replacement.
5.2 A random sample of size n = 200 was drawn without replacement from a
population of size N = 1, 000. The proportion of elements in the sample falling
into a certain category is R = 0.43.
(a) Calculate the 99%, 95%, 90%, and 50% symmetric conÞdence intervals for
the proportion of elements in the population that fall into this category. Brießy
interpret these intervals.
(b) Calculate the 99%, 95%, 90%, and 50% symmetric conÞdence intervals
for the number of elements in the population that belong to this category. Brießy
interpret these intervals.
(c) Do (a) and (b) under the assumption that the sample is with replacement.
5.3 In the manner described in Section 5.3, construct (a) a two-sided symmetric,
(b) a two-sided asymmetric, and (c) two one-sided 90% conÞdence intervals for
(i) the population mean of a variable, and (ii) a population proportion. Brießy
discuss the differences among these intervals. Which type of interval is preferable?
5.4 Following complaints that parking meters were malfunctioning, the Depart-
ment of Consumer Affairs selected at random and without replacement 10% of the
1,850 parking meters in a metropolitan area. Of the meters tested, 105 gave the
correct reading, 75 gave more time that was paid for, and 5 gave less time than
was paid for. “Parking meters,” the Department’s news release concluded, “may
be one of the few bargains left.”
What is your estimate of the proportion of all parking meters giving the cor-
rect time? Construct an interval estimate of this proportion; the interval estimate
should contain the true proportion with probability 95%. Do you need any addi-
tional information in order to determine if parking meters are indeed a bargain?
If so, what?
5.5 The management of a supermarket is concerned about the average waiting
time of its customers at the checkout counters on Saturdays. Fifty customers
“were randomly selected” one Saturday, and the time they spent waiting in line
before being served was recorded. Their average waiting time was 5.2 minutes and
the standard deviation of waiting times was 1.7 minutes.
(a) Assuming that the selected customers constitute a random sample with
replacement from the population of all customers, test the hypothesis that the
mean waiting time of all Saturday customers is less than or equal to 5 minutes,
against the alternative hypothesis that it is greater than 5 minutes. Assume that
the probability of a Type I error should not exceed 10%.
(b) Explain the exact meaning of Type I and II errors in this case, and their
likely consequences. Which error is the more serious in this case?
(c) Redo the test in (a) under the assumption that the probability of a Type
II error should not exceed 10%.
(d) How would you select a random sample of Saturday customers? Would
any n customers, no matter how selected, constitute a sample to which the test in
(a) could be applied?
5.6 The manager of a department store wished to estimate the proportion of time
that the sales clerks are idle. She divided the week into three periods of about
equal business volume (weekdays to 5 p.m., weekdays after 5 p.m., Saturdays).
Over a period of one month, a number of checks were made in each period at
 Problems 39

times selected at random. A particular clerk was checked a total of 70 times

during one of the three periods and in 12 of these times he was found to be idle.
(a) Assuming that the observations form a random sample from an inÞnite
population of potential timings, test the hypothesis that the true proportion of
time that the clerk is idle is less than or equal to 15%, against the alternative
hypothesis that it is greater than 15%. The probability of a Type I error should
not exceed 5%.
(b) Explain the exact meaning of Type I and II errors in this case, and their
likely consequences. Which error is the more serious in this case?
(c) Redo the test in (a) under the assumption that the probability of a Type
II error should not exceed 5%.
(d) How would you select the observation times so that the test in (a) can be
applied?
5.7 The Plastics Packaging Division (PPD) manufactures plastic containers, such
as cold-drink cups, yogurt cups, and creamers, by a high-speed extrusion and
vacuum thermo-forming process. One of these products is an 8-oz yogurt container.
The quality control department has established through tests the minimum wall
thickness required to ensure that the container does not crack when Þlled. This
minimum thickness is 11 mills (1 mill = 0.001 inches). A container with a wall
thickness of less than 11 mills is considered defective, and one with wall thickness
greater than or equal to 11 mills is considered good. These containers are produced
at the rate of about 100,000 per hour, and are packaged for shipment to the
customer in lots of about 40,000. Because of random variation in the quality of
the raw material and the manufacturing process, it is impossible to ensure that
all containers in a lot will be good. The Division and the customer agree that a
lot is acceptable if no more than 4% of the containers are defective.
PPD’s current quality control procedure is to take a random sample of 30
containers from each lot, and to reject the lot if the average wall thickness of the
containers in the sample is less than 11 mills.
Comment on the current quality control procedure, and compare it critically
with other possible procedures.
5.8 Cigarettes are manufactured by high-speed machines, some of which pro-
duce at a rate in excess of 5,000 per minute. These machines blend three main
ingredients (tobacco lamina, tobacco stem, and synthetic or reprocessed tobacco),
measure the quantity of blend that goes into each cigarette, and roll and wrap the
cigarette with the proper paper.
A certain amount of variation in the Þnished weight of the cigarette is in-
evitable and tolerated, but the weight should be neither too low (because cus-
tomers will Þnd it unsatisfactory) nor too high (since this would tend to increase
raw material costs). The ideal weight and tolerance for a particular type of
cigarette is 1,000 ± 100 mg per cigarette. In other words, an acceptable cigarette
should have a weight between 900 and 1,100 mg.
Brink Tobacco Company has purchased several high-speed machines. When
a machine is functioning properly (“under control”), no less than 96% of the
cigarettes produced are within the speciÞcation limits.
Occasionally, the machine goes “out of control.” As a result, the proportion
of cigarettes within the speciÞcation limits declines, while that of light and heavy
cigarettes increases. The problem for Brink Tobacco is how to detect whether or
not a machine is under control, and to be able to do so frequently, quickly, and
economically. Obviously, it would be impossible to weigh every cigarette produced;
whatever inspection system is used must be based on a sample.
40 Chapter 5: Interval estimation and testing

It is proposed that a sample of 10 cigarettes be taken randomly from the

machine’s output at the stroke of every minute. Each cigarette in the sample will
be weighed. Depending on weights observed, one of two decisions will be made:
allow the machine to run, or shut it down in order to repair and adjust it.
(a) You are asked to formulate a plan such that the probability of shutting
down a machine when it is under control does not exceed 20%. Without doing
any calculations, describe how such a plan could be determined.
(b) Determine the plan in (a).
(c) Same as (b), but the sample size is large–say 200.
5.9 In order for a set of numbers to qualify as random numbers, it is necessary
that the ten digits 0 to 9 appear with equal relative frequency (1/10) in the long
run. Of course, the actual relative frequencies of a Þnite number of such random
numbers will not always equal the theoretical relative frequencies.
One hundred numbers produced by a computer program yielded the frequency
distribution shown in Table 5.14.

Table 5.14
Data, Problem 5.9
Number Frequency
0 8
1 11
2 9
3 10
4 10
5 12
6 9
7 11
8 12
9 8
100

(a) Assuming that the above numbers can be treated as a random sample
from an inÞnite population of numbers that could be generated by this program,
test the equal frequency hypothesis. The probability of a Type I error should not
exceed 1%.
(b) Which other requirement must these numbers satisfy to qualify as random
numbers? How could this be tested? Describe only, do not calculate.
(c) Determine the exact meaning of Type I and II errors in this case, and
their likely consequences. Which error is more serious?
(d) Redo the test in (a) under the condition that the probability of a Type
II error should not exceed 1%.
(e) Would any n numbers produced by the program qualify as a sample to
which the tests in (a) or (d) apply?
5.10 A study was made of the time that elapsed between 150 successive telephone
calls to a certain exchange. The results were as shown in Table 5.15.
(a) Assuming that the observations constitute a random sample from an inÞ-
nite population of calls, test the hypothesis that the distribution of time between
calls is exponential. The probability of a Type I error should not exceed 5%.
 Problems 41

Table 5.15
Data, Problem 5.10
Time from previous call Number of
(minutes) calls
0.0 to 0.5 93
0.5 to 1.0 36
1.0 to 1.5 12
1.5 to 2.0 6
2.0 to 2.5 3
Total 150

Table 5.16
Wage distribution, Problem 5.11
Wage interval
(dollars) Number of workers
120 to 125 2
125 to 130 7
130 to 135 10
135 to 140 15
140 to 145 20
145 to 150 25
150 to 155 19
155 to 160 17
160 to 165 11
165 to 170 8
170 to 175 3
175 to 180 2
180 to 185 1
140

Hints: Estimate the parameter λ of this distribution by the inverse of the average
observed time between calls, using the midpoints of the time intervals above; that
is, set λ = 1/X̄ (why?). The probability that a variable X, having an exponential
distribution with parameter λ, will be in the interval from a to b (a < b) can be
shown to be equal to

P r(a ≤ X ≤ b) = e−λa − e−λb .

(b) Determine the exact meaning of Type I and II errors in this case, and
their likely consequences. Which error is more serious?
(c) Redo (a) under the condition that the probability of a Type II error should
not exceed 5%.
(d) Can the observations be assumed to form a sample to which the tests
above may be applied?
5.11 The distribution of daily wages in a random sample without replacement of
140 workers in a large factory employing about 15,000 workers is shown in Table
5.16.
42 Chapter 5: Interval estimation and testing

(a) Assuming that the sample can be treated as essentially one with replace-
ment (because of the large population size), test the hypothesis that the distribu-
tion of wages in the factory is normal. The probability of a Type I error should not
exceed 10%. Hint: Estimate the parameters µ and σ of the normal distribution by
the sample mean and standard deviation of wages calculated using the midpoints
of the intervals above.
(b) Determine the exact meaning of Type I and II errors in this case, and
their likely consequences. Which error is more serious?
(c) Redo the test in (a) under the condition that the probability of a Type II
error should not exceed 10%.
5.12 A study was made of the feasibility of constructing short-term storage
facilities for liquid products (animal, vegetable, and marine oils, chemicals, etc.,
but excluding petroleum products) exported from or imported to a Great Lakes
port. Table 5.17 shows the frequency distribution of the time between successive
arrivals of all liquid-carrying vessels at the port in the most recent summer season.

Table 5.17
Great Lakes study, Problem 5.12
Interarrival time Number of
(days) vessels
0 to 5 23
5 to 11 10
11 to 15 6
Over 15 1∗
40
∗
24 days

Can it be assumed that the distribution of interarrival times at the port is

exponential? Note the Hints of Problem 5.10, but otherwise answer this question
as you consider appropriate.
5.13 One hundred thirty-eight students were enrolled in Administration 531 and
532 (Quantitative Methods, I and II) at an MBA program. Of these students, 79
received an A or B grade in both 531 and 532. The actual distribution of grades
was as follows:

ADM532
ADM531 A B Total
A 18 16 34
B 31 14 45
Total 49 30 79

For example, 31 students had a B in ADM531 and an A in ADM532, etc.

As a person knowledgeable in statistical methods, you are invited to comment
on these numbers.
5.14 Among the questions included in a survey of apartment tenants were the
following:
 Problems 43

A: Do you prefer living in an apartment? Yes No

B: Do you have children? Yes No
The following table summarizes the responses of a random sample of 100
tenants.

B
A Yes No Total
Yes 15 35 50
No 25 25 50
Total 40 60 100

(a) Assuming that the population of tenants is very large so that the sample
is essentially one with replacement, test the hypothesis that a tenant’s preference
is independent of the presence of children. The probability of a Type I error should
not exceed 10%.
(b) Determine the meaning and likely consequences of the two types of error
in this case. Which error is the more serious?
(c) Redo the test in (a) under the assumption that the two errors are equally
serious.
5.15 A sample of 280 households in a large metropolitan area was selected in
order to investigate the usage of XL White, a brand of laundry bleach. The results
of the study are in part as shown in Table 5.18.

Table 5.18
Results of bleach study, Problem 5.15
Household income Non-users Light users Heavy users Total
Under 15,000 28 16 10 54
15,001 to 20,000 32 15 8 55
20,001 to 30,000 27 14 12 53
30,001 to 40,000 31 17 11 59
Over 40,000 32 18 9 59
Total 150 80 50 280
V = 1.826
Family size Non-users Light users Heavy users Total
Under 3 60 26 10 96
3 to 4 52 33 15 100
Over 4 38 21 25 84
Total 150 80 50 280
V = 13.799

(a) Test the two hypotheses that usage is independent of (i) income, and (ii)
family size. The probability of a Type I error should not exceed 10%.
(b) Examine the meaning and likely consequences of the two types of error
in this case. Which error is the more serious? Redo the tests under the condition
that the probability of a Type II error should not exceed 10%.
(c) How would you select a simple random of households? Would any sample
do for the purpose of applying the above tests?
(d) What are the implications of the test results?
44 Chapter 5: Interval estimation and testing

5.16 A study was undertaken to determine the factors inßuencing handling

time for metal plates used in a punch press. The weight of the metal piece was
thought to be one of the determining factors. Accordingly, the weight category
(light, medium, heavy) and the handling time (classiÞed into short and long) were
recorded for a random sample of 15 metal plates with the results shown in Table
5.19.

Table 5.19
Weight and handling time of
metal plates, Problem 5.16
Plate No.: Weight Handling time
1 Medium Long
2 Light Short
3 Medium Short
4 Heavy Short
5 Light Short
6 Heavy Long
7 Medium Short
8 Heavy Long
9 Light Short
10 Medium Long
11 Light Long
12 Medium Short
13 Heavy Long
14 Medium Long
15 Heavy Long

Does the weight of a metal piece inßuence the handling time? If yes, what is
the nature of the relationship? In answering these questions treat the sample as
if it were large. Why is this assumption necessary? Explain and justify any other
assumptions you are forced to make.
5.17 A product testing laboratory was asked to evaluate the durability of four
brands of tires. The durability tests were made under normal city driving condi-
tions with the assistance of a Þrm operating a ßeet of taxicabs. In all, 140 taxis
were employed and each was Þtted with four new tires, one from each brand. These
tires were randomly selected from retail outlets. Each day, the tires were rotated
in a predetermined manner to ensure uniform exposure to wear, unrelated to their
original location on the car. Also daily, the tires were inspected to determine if
they had reached the end of their useful life. The test results are summarized in
Table 5.20.
(a) Assuming that the observations can be considered random samples from
inÞnitely large populations, do the four samples of tires come from populations
having identical distributions of life? The probability of a Type I error should not
exceed 5%.
(b) Interpret the meaning of the two types of error in this case, and their
likely consequences. Which error is more serious?
(c) Redo the test in (a) under the condition that the probability of a Type II
error should not exceed 5%.
 Problems 45

Table 5.20
Tire test results, Problem 5.17
Life Brands
(000 miles) A B C D
Under 30 25 30 25 26
30 to 31 40 45 45 43
31 to 32 60 55 50 55
Over 32 15 10 20 16
Total 140 140 140 140

(d) Comment on the method used by the laboratory to measure the durability
of tires. Does the method produce samples to which the tests in (a) and (c) may
be applied?
5.18 In a sample of 200 beer consumers in city A, the average monthly beer
consumption was 26.5 oz, and the sample standard deviation was 2.14. In a
sample of 150 beer consumers in city B, the average monthly beer consumption
was 29.8 oz, and the sample standard deviation was 3.87.
(a) Assuming that the samples are random and without replacement from
very large populations, test the hypothesis that the average beer consumption of
all beer drinkers in city A is the same as that in city B. The probability of a Type
I error should not exceed 20%.
(b) Interpret the meaning and the likely consequences of the two types of
error in this case. Which error is the more serious?
(c) Redo the test in (a) assuming the probability of a Type II error should
not exceed 20%.
(d) How would you select a random sample of beer consumers in a city?
5.19 Surveys of potential buyers can provide useful information about the likeli-
hood of success of planned (that is, as yet not manufactured or offered) products
and services. In the case described here, the product in question was an electric
car which, at the time the survey was conducted, was still at the prototype stage.*
Data were collected through personal interviews with 1,229 randomly selected
shoppers at designated shopping centers in three cities. The respondents were
given a detailed description of the planned electric car, which included its price,
passenger and luggage capacity, size, speed, cost of operation, and safety features.
The respondents were asked to rate their intention to buy the electric car when
it became available on an 11-point scale, ranging from 0 (absolutely no chance
of buying) to 10 (almost certain of buying). They were also asked to rate the
importance they attributed to each of a number of factors inßuencing their choice
of car on a 10-point scale, ranging from 0 (low) to 9 (high importance). The results
of the study are shown in part in Tables 5.21, 5.22, and 5.23.
In Table 5.23, the importance groups in column (2) are formed as follows: I
= low importance (0 to 3 on the 9-point scale); II = medium importance (4 to 6);
and III = high importance (7 to 9). The intention groups are as follows: None =
0 or 1, Low = 2 to 4, Medium = 5 to 7, and High = 8 to 10 on the 11-point scale.

* This case is based on G. M. Naidu, G. Tesar, and G. Udell, “Determinants

of buying intentions of the electric car,” 1973 Proceedings of the Business and
Economic Statistics Section of the American Statistical Association, pp. 515-20.
46 Chapter 5: Interval estimation and testing

Table 5.21
Distribution of buying intention,
Problem 5.19
Frequency of response
Rating City A City B City C
0 94 53 33
1 35 12 16
2 32 12 31
3 43 12 38
4 28 16 37
5 101 65 63
6 37 30 29
7 58 36 35
8 53 32 26
9 23 31 12
10 52 33 21
Total 556 332 341
Mean 4.72 5.30 4.69
Std. dev. 3.22 3.21 2.78

Table 5.22
Mean and standard deviation
of ratings for factors, Problem 5.19
Ratings
Factor Mean Std. deviation
Cost of operation 7.00 2.30
Ease of maintenance 6.24 2.49
Cost of maintenance 6.81 2.36
Luggage capacity 4.56 2.51
Passenger capacity 4.92 2.41
Size 5.18 2.58
Mileage 6.85 2.34
Price 6.91 2.33
Speed 4.34 2.43
Acceleration 4.69 2.58
Safety 7.24 2.36
Pollution 6.28 2.80

Column (7) shows the mean intention-to-buy rating for each importance group.
Column (8) is the V statistic for testing the independence of factor and intention
to buy.
Assume you are the marketing manager of the company intending to produce
this electric car. Interpret these Þndings.
5.20 An issue of some importance in marketing is the extent to which consumers
are conscious of the price of an article at the time of its purchase. One theory
is that consumers watch prices and carefully adjust their purchases from different
 Problems 47

Table 5.23
Association between importance of factor
and intention to buy, Problem 5.19
Imp. Intention group Mean
Factor group None Low Medium High intention V
(1) (2) (3) (4) (5) (6) (7) (8)
Cost I 43 18 36 25 3.73
of II 49 83 92 50 4.18 67.19
operation III 87 210 326 208 5.14
Ease I 47 40 78 38 4.18
of II 47 114 120 65 4.45 32.82
maintenance III 85 157 256 180 5.13
Cost I 39 20 46 30 4.12
of II 49 85 101 56 4.29 37.12
maintenance III 91 206 307 197 4.98
I 62 91 168 104 4.91
Luggage II 51 148 186 114 4.96 26.60
capacity III 66 72 100 65 4.33
I 54 73 130 87 4.92
Passenger II 56 154 213 127 5.02 23.28
capacity III 69 84 111 69 4.27
I 50 29 36 17 3.03
Mileage II 41 86 102 49 4.41 79.70
III 88 196 316 217 5.19
I 35 30 37 20 3.59
Price II 49 96 95 59 4.40 29.99
III 95 185 322 204 5.11
I 78 93 148 133 4.93
Speed II 59 144 218 103 4.68 31.76
III 42 74 88 47 4.41
I 53 60 72 34 3.80
Pollution II 36 98 115 50 4.60 47.36
III 90 153 267 199 5.18

outlets so as always to minimize the total cost of the articles they buy. Opponents
of this theory argue that consumers cannot remember accurately the prices of the
hundreds of commodities they normally buy; they are often concerned more with
whether or not they can afford to buy an article rather than with its exact price.
To investigate this issue, 640 housewives were approached at randomly se-
lected addresses in a city.* “Housewife” was interpreted broadly to mean the
person, male or female, responsible for current purchases of provisions for the
household, but will be referred to as a “she” in this case. The questionnaire
contained questions about recent purchases of Þfteen selected commodities and
about certain aspects of the household. For each commodity, the housewife was

* Adapted from A. Gabor and C. W. J. Granger, “On the price consciousness

of consumers,” Applied Statistics, Vol. 10, No. 3, pp. 170-88.
48 Chapter 5: Interval estimation and testing

asked when she bought it last. The interviewers were instructed not to ask for
further information if the last purchase was more than a week ago. However, if
the housewife had purchased the commodity within the last week, she was asked
to state the brand or type of the commodity, whether or not she recalled the price
paid for it, and, if so, how much she paid. The interviewers were asked to supply
their personal estimate of the social group to which the housewife belonged. Five
social groups were distinguished: A (the well-to-do), B (the professional middle
class), C (the lower middle class), D (the working class), and E (the poor). The
recognizable characteristics of these groups were described in detail in the written
instructions to the interviewers.
Of the Þfteen commodities listed in the questionnaire, eight were sold at such
a variety of prices that it was not practicable to check the answers. The remaining
seven, however, could be checked and the price named by the housewife could be
compared with the list price obtained from industry sources. If there was any
departure from the list price, the answer was classiÞed as incorrect. The Don’t
Know category includes both those who, for some reason, refused to name the
price of the purchase, and those who admitted that the price named was more or
less a guess. Altogether, 422 housewives reported 1,888 recent purchases of one or
more of the seven commodities the price of which could be checked. The results
by commodity are shown in Table 5.24.

Table 5.24
Results by commodity, Problem 5.20
Stated Commodity
price was: Tea Coffee Sugar Jam Margarine Flour Cereal All
Correct 283 109 266 117 116 100 85 1,076
Wrong 56 35 53 46 110 82 98 480
Don’t know 18 16 78 34 26 99 61 332
All purchases 357 160 397 197 252 281 244 1,888

The relationship between the percentage of correct answers and social group
is shown in Table 5.25.

Table 5.25
Results by social group, Problem 5.20
Number of
Social Number of Number of prices named Percentage
group housewives purchases correctly correct
A 9 42 19 45.2
B 28 116 54 46.6
C 118 544 309 56.8
D 229 1,052 616 58.6
E 38 134 78 58.2
All groups 422 1,888 1,076 57.0
 Problems 49

The percentage correct is the ratio of the number of correct prices named to
the number of items bought.
What are the implications of this study for the price-awareness theory?