Testing of Hypothesis

Introduction
As already mentioned carlier, in all statistical investigations
the objective of the enquirer is to gather information about some
characteristics of the population, but most often he will have
access only to a part of the population, i.e. a sample. So the
cnquirer makes u guess about the population characteristics on
tie basis of knowledge acquircd from the sample. This process
of going from the known sample to the unknown population is
called statistical inference.
The two principal areas of statistical inference are the areas of
estimation of paramcters and of tests of statistical hypotheses.
In testing of hypothesis we are concerncd with examining the
validity of some statement about the unknown parameters or
about the formof a distribution, in the light of the sample.gi
Statistical Hypothesis
Statistical hypothesis : Astatistical hypothesis is an assertion
about the distribution of one or. more randon variables which
we want to verify on the basis of a sample. Thus, the assertion
that a random variable possesses a nornmal distribution, or that
themean of aPoisson distribution is 2, is a statistical hypothesis.
Test of Hypothesis

est of a hypothesis : Atest of astatistical hypothesis is arule

which, when the sampBe vales have been obiained. leads to a
decision to ccept or o reject the hypothesis under considera
tion.
In deciding whet:r a coin is fair, u test may be rejecting the
hypothesis of fairaess if in 5tosses of the coin i head or 5 heáds
ure chined, and ncceotin the nypothesis otiherwise.e
Simple and Composite Hypotheses

Sinple hypothesis and composite Iypothesis : A statistical

hypothesis may be cither siinple or composite. If a hypothesis
completeiyspecifies the distribution or distributions under study,
called
it is calleda simple hypoi:esis; but if it fails to do so, it is
composite hypottesis.
Ua randomvarubie x is known to follow the Poisson distri
buison with unknown mean , d:en the hypothesis =3is a
simpie bypothesis, but the fiypothesis 3 is a composite hy
pothesis with Idegree of Ireeiom, (The degree of freedom of a
hy pothesis is the numbe: of mspccificd parameters.)
 Null and Alternative Hypotheses

Nul! hypothesis and alternative lhypothesis : A hypothesis

which we set up and procecd to test vith the help of the sample
observations is called a null hypothesis. According to Fisher, a
hpothesis, which is tested for plausibie rejection under the as
sumption that it is true, is a null hypotlesis.
al
k Any hypothesis that contradicts the null hypothesis is an
ternative 1ypothesis.
9i Suppose the mean of u Poisson distribution is unknown.
Some authority suggests that 2 is likely to be 3. Then the null
hypothesis is 2 = 3. Here the alternative hypothesis may be >
3. Generally, anull hypothesis is denoted by H, and an alterna
tive hypothesis by H.
Sample Space

Sample space : Let a random sample E : (x Xy wi ) of

size n be drawn from.a population. If we draw another sample of
the same size from the same population, the sample observa
tions will generally be different. Considering un n-dimensional
space, if we plot such sample observations, we will get different
points for different samples. The collection of such points core
sponding to all possible samples is called a sumple space. A sam
ple spacc may be continuous or discrete.
Critical Region
Critical region : Let w be a subspace of the sample spuce W
such that we shall reject H, if the observed sample point E:(X.
Ky, ..x)falls in w and accept Ho ifE falls in the complemen
tary region (W - w). Then the region w is called the critical re
gion or the region of rejcction and (Ww) is called the region
of acceptance.
 FCritical Region
Test Statistic

W
1.645 2.5

Critical Critical
Value Value

Rejection
Nonrejection Rejection
Region
Region Region
Test Statistic

Test-criterion (or test-statistic) : It is a statistic (that is, a func

tion of the sumple values) which is used for
In practice, critical region is defined in formulating a test.
rnon. The set of values of this staistic that
terms of the test-crite
H, is thc critical region of the test.
leads to the rejection of
I may be noted that
sample space W, we have corresponding to different points E in the
different values of the test criterion.
Those values of the test-criterion that correspond to the points in
the subspace wconstitute the
critical region. *
General Procedure of Testing
General procedure : Let us consider the null hypothesis H
Any test for H, is nothing but a rule for rejecting or accepting
H depending on the nature of the sample observations E:(
X, . , . x). Atest would specify a subspace w in the sample
space W, called the critical region, and would require rejection
of H, if Elies is w and acceptance of H, if Elies in W-w.Io
cach test of H, there corresponds, a unique critical region and,
conversely, to cach critical region there corresponds a unigue
test.
Now, while choosing a test (or a critical region) from all avail
able tests (or critical regions), one must keep in mind the two
types of errors that one may commit in using the test.t
Remark : By acceptance of H, we do not mean that it is proved
to be true. All it means is that, based upon the data at hand, we
find no reason to suspect its validity. Similarly, rejection of H,
does not mean a disproof of H I means simply that, in the light
of the given sample, H, does not appear to be a plausible hy
pothesis.
Type -I Error and Type -I| Error

in
WType-l error (or error of first kind) : The error committed
is true is called
rejecting the null hypothesiswhen actually it and w the critical
type-l error. If E denotes the sample observations
c r r o r e y at a i
Probability of type I
Probability of rcecting B, when t Is true t Ba
Probability that E lies in wwhen H, is true
PEEWI,)
 error commit
MType-1l error (or error of second kind ):The is false is
accepting the null hypotiesis when actually it
ted in
called type-lI errOr
Probability of (ype-Il error
false
Probability of accepting Ho vhen it is
false, ie. alter
Probability that Elies in (W- w) when H,is
native hypothesis H is true
P(EeW-wlH)
=I-P (E ¬ wlH)

The position H, true H,false

Decision made
ype error Correct decision
Reject H, 7
Correct decision Type-Il error
Accept H
Power of a Test

complement of probabiliy of type-I error is called power

The rejecting H, when H, is false.
of the test. It is the probability of
corresponding to the alternative hypoth
The power of the test
esis H is P(EewH).)
clearly have the probabilities of both types
Agood test should
of error as smallas possible. However, for a sample of fired size.
bit is not possible to minimise both the probabilities at the same
time. They are so peculiarly related that as one decreases. ithe
other increases. As such, in choosing a test procedure, we first
require that the probability of type-I error should not exceed a
specificd value and then, from among all the tests fulfilling this
eriterion, we are to choose the one which has the highest powe:
(or lowest probability of type-Ilerror):
Level of Significance

wThe upper bound of the probability of type-I error of a test is

called the level of significance of the test. In the case of asimple
hypothesis, the level of significance and probability of type-I
signiticance
error are samc. In most practical cases, as levcl of
we take 0.01 or 0,05.
Unbiased Test
be unbiased
A good test should be unbiased. A test is said to
of
if the probability of type-l error does not exceed the power
unbiased
the test. Thus, the test based on the critical region w is
if
P(E e wlH) S P(E ew|H).
For a biased test,
P(E ewlH)> P(E e w]H),
so that in a biased test H, is more often rejected when it is true
than when it is false.
One and TwoSided Tests
statistical test may be aone-sided or a two-sided test de
A
Suppose we are testing
pending on the alternative hypothesis.
alternative hypothesis may
the null hypothesis H, :0= The
be onc of the following :
)H,:0= 0,(>,) ut
si (ii) H,:0=0, (<0,)
(ii) H, :0>,
(iv) H,:0<o s a

(v) H , : 0 , t
deviations are only positive or onlynegative [cases (i)
Whenthe one-tail) testS and when
tests aure called one-sided (or
to (iv)], the (case (v)], the tests
be positive or negative
the deviations may
are called two-sided (or twO-tail) tests.
Problem:

In order to test whether a coin is perfect, the coin is tossed5

if and only
times. The'null hypothesis of perfectness is rejectedprobability
the of
if more than 4 heads are obtained. What is
error when the cor
type-I eror? Find the probability of type-II (H.S. '82)
responding probability of head is 0.2.
obtained in 5 tosses of the
Let X denote the number of heads
coin and p denote the probability of head in a toss. Then the
p.m.f. ofX is
P(X = x) =( ) p -P)5,x=0(1)S.
the critical region
Here our null hypothesis is Hp=/2. and
is X> 4, i.e. X= 5.
So the probability of type-1 error
P(X=5|H) = P(X =5|p= /2)
 0.2)B()
Probability of type-Ilerror = PX S4Ip =
9tE =|-P(X=5|p:

|-(0.2) - |
it
3124
3125
Problem:

Given fx)=7.0Sxs0
8
= 0, elsewhere.
against the alternative hypoth
to test the hypothesis H, :0=
of the single observation x. Find the
csis H:0 =2 on the basis
critical region
error if the
probabilities of type-I error and type-II
be x > 0.5,
hypothesis is H,:0= l and the alternative hy
Here our null
pothesis is H :0= 2. So the p.d.f. of x under oS t s

aletatag ts kai Tsrsbae

0. elsewhere
and the p.d.f. of x under H is
tilidsag s s a q aio
f(x)= J/2,0SxS 2
=0, elsewhere.
fy)X
isx>o.s-
The critical region
 type-I error iedhoqd lhe uo s
So the probability of
=P(x > 0.S |H)= P(x > 0.5 |0 l) tiidong
='
05
dx =xs
= 0.5,
of type-II erroroileliig st# srat
and the probability
POx s 0.5 |H) =PX S0.| 2 h t : ÌHeÂarng e s s t

dx
Problem:
that it is
Acoin is tested for unbiasedness. The hypothesis
of 10
unbiased is rejected if 9 or more tosses of the coin out
significance?
tosses result in head. Can we take 1% as level of
tosses of
Let X denote the number of heads obtained in 10
the coin and p be the probability of head in a toss.

:. P(X =x) =((1-p)0-,x=0(0)10.

The null hypothesis is H,: p=/2, and the critical region is X
9and 10. i Saats
So the probability of type-I error
= P(X=9, 10|H)
= P(X =9, 10 |p = /2)

-(") g10
 =,10 1024
0.0107 > 0.01
that level of significance is the upper bound of the
We know type-I error
probability of type-I error. Here the probability of
0.01, we cannot take 1% as level of signifi
being greater than
cance.