a4
INTRODUCTION
If an engineer has to decide on the basis of sample data whether the true average
lifetime of a certain kind of tire is at least 42,000 miles, if an agronomist has to
decide on the basis of experiments whether one kind of fertilizer produces a higher
yield of soybeans than another, and if a manufacturer of pharmaceutical products
hhas to decide on the basis of samples whether 90 percent of all patients given a new
‘medication will recover from a certain disease, these problems can all be translated
into the language of statistical tests of hypotheses. In the first case we might say that
‘the engineer has to test the hypothesis that 0, the parameter of an exponential pop-
ulation, is at least 42,000; in the second case we might say that the agronomist has to
decide whether 1 > 42, where jz) and 12 are the means of two normal populations;
land in the third case we might say that the manufacturer has to decide whether 0, the
parameter of a binomial population, equals 0.90. In each ease it must be assumed, of
‘course, that the chosen distribution correctly describes the experimental conditions;
that is, the distribution provides the correct statistical model
‘As in the preceding examples, most tests of statistical hypotheses concern the
parameters of distributions, but sometimes they also concern the type, or nature,
Of the distributions themselves. For instance, in the first of our three examples the
engineer may also have to decide whether he is actually dealing with a sample
from an exponential population or whether his data are values of random variables
having, say. the Weibull distribution of Exercise 623.
Desmrrion 12.1. A statistical hypothesis is an assertion or conjecture about
the dstribution of one or more random variables. If a statistical hypothesis
competely specifies the distribution, itis referred to as a simple hypothesis: if
not, i is referred to as a composite hypothesis,
375
376 Chapter 12. Hypothesis Testing
A simple hypothesis must therefore specify not only the functional form of the
underlying distribution, but also the values of all parameters. Thus, in the third of
the above examples, the one dealing with the effectiveness of the new medication,
the hypothesis @ = 0.90 is simple, assuming, of course, that we specify the sample
size and that the population is binomial. However, in the first of the preceding
examples the hypothesis is composite since @ 2 42,000 does not assign a specific
value to the parameter 8.
To be able to construct suitable criteria for testing statistical hypotheses, it is
necessary that we also formulate alternative hypotheses, For instance, inthe example
dealing with the lifetimes of the tires, we might formulate the alternative hypothesis,
that the parameter 4 of the exponential population is less than 42,000; in the example
dealing with the two kinds of fertilizer, we might formulate the alternative hypothesis.
i = 42; and in the example dealing with the new medication, we might formulate
the alternative hypothesis that the parameter 6 of the given binomial population is
‘only 0.60, which is the disease’s recovery rate without the new medication.
‘The concept of simple and composite hypotheses applies also to alternative
hypotheses, and inthe first example we can now say that we are testing the composite
hypothesis 9 2 42,000 against the composite alternative @ < 42,000, where @ is the
parameter of an exponential population. Similarly, in the second example we are
testing the composite hypothesis 4: > 2 against the composite alternative 1 = 12,
where 11 and j are the means of two normal populations, and in the third example
‘we are testing the simple hypothesis @ = 0.90 against the simple alternative 0 = 0.60,
where @ is the parameter of a binomial population for which n is given.Frequently, statisticians formulate as their hypotheses the exact opposite of
what they may want to show. For instance, if we Want to show that the students
in one school have a higher average 1.Q. than those in another school, we might
formulate the hypothesis that there is no difference: the hypothesis ya = 12.
With this hypothesis we know what to expect, but this would not be the case if
wwe formulated the hypothesis 4; > 12, at least not unless we specify the actual
difference between ji and 42.
Similarly, if we want to show that one kind of ore has a higher percentage
‘content of uranium than another kind of ore, we might formulate the hypothesis that
the two percentages are the same; and if we Want to show that there isa greater vari-
ability in the quality of one product than there is in the quality of another, we might
formulate the hypothesis that there is no difference; that is, o1 = o2. In view of the
assumptions of “no difference,” hypotheses such as these led to the termnll hypoth
sis, but nowadays this term is applied to any hypothesis that we may want to test
Symbolically, we shall use the symbol Ho for the null hypothesis that we want
to test and H or Ha for the alternative hypothesis. Problems involving more than
two hypotheses, that is, problems involving several alternative hypotheses, tend to
bbe quite complicated, and they will not be studied in this book.
12.2 TESTING A STATISTICAL HYPOTHESIS
‘The testing of a statistical hypothesis is the application of an explicit set of rules
for deciding whether to accept the null hypothesis or to reject it in favor of the
alternative hypothesis. Suppose. for example, that a statistician wants to test the
Section 12.2 Testing a Statistical Hypothesis 377
ull hypothesis @ = @ against the alternative hypothesis @ = 6}. In order to make a
choice, he will generate sample data by conducting an experiment and then compute
the value of a test statistic, which will tell him what action to take for each possible
‘outcome of the sample space. The test procedure, therefore, partitions the possible
values of the test statistic into two subsets: an acceptance region for Ho and a
rejection region for Ho.
‘The procedure just described can lead to two kinds of errors. For instance, if
the true value of the parameter @ is and the statistician incorrectly concludes that
6 = 6, he is committing an error referred to as a type I error. On the other hand, if
the true value of the parameter @ is #1 and the statistician incorrectly concludes that
@ = é, he is committing a second kind of error referred to as a type II error.
Derirrion 122.
1 Rejection of the null hypothesis when it is true is called a type I error;
the probability of committing a type I error is denoted by a.
2. Acceptance of the null hypothesis when itis false is called a type HI error;
the probability of committing a type Il error is denoted by 8.
It is customary to refer to the rejection region for Hy as the critical region of
the test and to the probability of obtaining a value of the test statistic inside the
critical region when Ho is true as the size of the critical region. Thus, the size of a
critical region is just the probability a of committing a type I error. This probability
is also called the level of significance of the test (see the discussion on page 389).EXAMPLE 12.1
Suppose that the manufacturer a of a new medication wants to test the null
hypothesis # = 0.90 against the alternative hypothesis 0 = 0.60. His test statistic is.
X, the observed number of successes (recoveries) in 20 trials, and he will accept the
null hypothesis ifx > 14; otherwise, he will reject it. Find a and p.
Solurion The acceptance region for the null hypothesis is x = 15, 16,17, 18,19, and"
20, and, correspondingly, the rejection region (or critical region) isx = 0,1,2,..., 14
‘Therefore, from Table I,
@ = P(X S14;6 = 0.90) = 0.0114
and
B= P(X > 14,0 = 0.60) = 0.1255 .
‘A good test procedure is one in which both « and f are small, thereby giving
tus a good chance of making the correct decision. The probability of a type II error
in Example 12.1 is rather high, but this can be reduced by appropriately changing
the critical region. For instance. if we use the acceptance region x > 15 in this
example so that the critical region is x S 15. it can easily be checked that this would
make « = 0.0433 and f = 0.0509, Thus, although the probability of a type Il error
is reduced. the probability of a type I error has become larger. The only way in
which we can reduce the probabiliti
378 Chapter 12___Hypothesis Testing
of the sample, but as long as 7 is held fixed, this inverse relationship between the
probabilities of type I and type II errors is typical of statistical decision procedures.
In other words, if the probability of one type of error is reduced, that of the other
type of error is increased.
EXAMPLE 12.2
‘Suppose that we want to test the null hypothesis that the mean of a normal population
with o? = 1 is wo against the alternative hypothesis that it is 4, where 1 > 0-
Find the value of K such that ¥ > K provides a critical region of size = 0.05 for a
random sample of size n.
Solution Referring to Figure 12.1 and Table II, we find that z = 1.685 corresponds
to.an entry of 0.4500 and hence that
K-40
1645 = X= Ho
dn
Itfollows that
k= m+ SS .
ve
‘of both types of errors i to increase the sizeEXAMPLE 12.3
With reference to Example 12.2, determine the minimum sample size needed to test
the null hypothesis yo = 10 against the alternative hypothesis 4; = 11 with 6 S 0.06.
Solution Since f is given by the area of the ruled region of Figure 12.1, we get
™” 1.645
p=r(x<10+*S.4=n)
(10+ 488) —1
=Pi|Z< vn
Iva
= P(Z < - Vii + 1.645)
FIGURE 12.1: Diagram for Examples 12.2 and 12.3.
Section 123 Losses and Risks 379
and since z = 1.555 corresponds to an entry of 0.5000 — 0.06 = 0.4400 in Table IIL,
wwe set — J+ 1.645 equal to ~1.555. It follows that fi = 1.645 + 1.555 = 3.200 and
= 10.24, or 11 rounded up to the nearest integer. .