474 PROBABILITY AND STATISTICAL INFERENCE: Theory and Procs:

Power ofatest: The probability of rejecting H, when it is false is known as the

power of atest. It is the power ofa test in detecting the falsity of H,:
Power ofa test =1-P (type Il error) =l-B.
We try to design the test rulc in such a way that the error probabilitics are
minimized. But it is not possible to reduce both the error probabilities, a and B.
simultaneusly. Therefore, we try to increase the power of the test subject to a
fixed level a, i.e., we try to minimize ß (i.e., maximize power) for afixed a.
The following examples illustrate how to compute probabilities of type Iand
type lIerrors in significance testing.
Example 6.1 The fraction of defective items in a large lot is p. To test the null
hypothesis H,:p=0.2, one considers the number of defectives Xin a sample of
8 items and fails to reject the null hypothesis if X<6 and rejectsthe hypothesis,
otherwise. What is the probability of type I error of this test? What is the
probability of type IIerror corresponding to p=0.01? What is the power of the
test for p=0.01?
Here number of defectivesin a sample of8 items, X, is bin(8, p).
We want to test Ho:p=0.2.
Test rule is :Reject H,ifobserved X> 6, do not reject H, otherwise.
P (typeI error) = P(rejecting H|H, is true)
=P(X>6|p=0.2)
=P(X= p= 0.2) + P(X =8p= 0.2),
since Xis bin(8, p =0.2), under H,
8
x(0.2)' x(0.8)+ x(0.2° x(0.8)°
=8.(0.2) x0.8+ (0.2)8
=0,00008448.
P(type IIerrorp = 0.1) = P(not rejecting HÍ lp=0.1)
= P(Xs6 |p =0.1)
=]-P(X= 7p=0.1)- P(X =8| p=0.1)
-1-}x(0.)' x(0.9)-( x(0.I³ x(0:9"
TEST OF HYPOTHESIS 475

=1-8x(0.1) x0.9-(0.1)*
=1-000000072- 000000001
-0,99999927.
The power ofthe test corresponding to p =0.I
=1-P (type ll error p= 0.)
= 0.00000073.

Example 6.2 Tocxamine the claim ofa reputed publisher that m, the average
number of misprints per page of a book, is 1, ifa particular page of that book
contains morethan 2 misprints, then the nullhypothesis that m= 1is rejected.
What is the probability of type Ierror ?Find the power of this test when the
alternative hypothesis is m=2. Given that e = 0.368.
LetXbe the number of misprintsper page of thebook. Here Xis Poisson (m).
We want to test the null hypothesis H,: m=l
The test rule is : reject H, if observed X> 2; do not reject H, otherwise.
Then

P(type Ierror) = P(rejecting H% | H, is true)

= P(X> 2 |m=1)
=1-P(X = 0\m = 1)-P(X =l|m=l)-P(X = 2|m = 1)
=1-e +
0! 1! 2!
=]-0.368 ×2.5
=1-0.92
=0,08.
by
Power of thetest when alternative hypothesis H,:m=2 is given
Power = P(rejecting H, |H, is false)
hypothesis is H,:m=2
= P(X > 2| m= 2) , since alternative
=]-P(=0|m=2)-P(X =l|m=2)-P(X =2| m=2)
2 22
21

=|-0.135424 x 5
=|-0.67712
= 0.32288.
 476 PROBABILITY AND STATISTICAL INFERENCE: Theory and Practi

Example 6.3 A coin is to be tested for unbiasedness. The hypothesis that itis
unbiased is rcjccted if 9 or more tosses of the coin out of 10 tosses result in
heads. Can we take 1% as level of significance? Can you suggest an appropriate
level of significance for this test rule?
Let X bc the number of heads resulted in 10 tosses of a coin. Let p be the
probability of obtaining a hcad for cach toss.
Then Xis bin(10,p). If the coinis unbiased, p=;.
1
We want to test HÍ :p=;
The test rule says that 'reject Ho if observed X>9.'
P (type lerror) =P (rejecting HÍ | Ho is true)
= P(X>9lp=)
= P(X =9) + P(X=10), when p =

11
1024
=0.01074> 0.01.
Note that the size of the test or P (type I error) < a, the level of the test.
Therefore, we cannot take 1% as the level of significance here. But we can
carry out the test at 5% level, since P(typeIerror) = 0.01074<0.05.
Note: Actually we define our test rule in such away that the level =a. Here the
test rule is such that the level cannot be 1%. It is possible totake a =0.05 here.
Test statistic: Astatistic, a function of sample observations, used to test the null
hypothesis is known as a test statistic. Aconclusion to reject or not to reject HÍ
is bascd on the value of the test statistic, observed in a samplc selected.
There are two approaches for making adecision regarding rejection or non
rejection of H, viz.,
p-value approach
critical value approach.
Both the approaches lead to the same conclusion. (vide Examples 6.4, 6.5,
6.6). Let usfirst define p-value and critical value.
p-value: p-value is the probability that a test statistic will take the value as
Cxtreme as or more extreme than the one actually observed, when H, is true. lt
is also known as observed level of significance.
TEST OF HYPOTHESIS 477

The p-valuc is thesmallest valuc of the levcl of significance a for which

H, can be rejccted. It is the actual risk of committing a type I error, if H, is
reiccted bascd on the observed value of the test statistic. The p-value measures
the strength of the evidence against H,. We reject H, at level a ifp-value [ a.
Then we can report that the results are statistically significant at level a.
Critical region, non-rejection region and critical value: The entire set of
pOSsible valucs of test statistic isdivided intotwo sets or regions. One region
consisting of the values that support the alternative hypothesis and leads to
the rejcction of H is called the critical region or rejection region. The other
consisting of the values that support H, is called the non-rejection region
(often called acceptance region). The value of the test statistic which lies at the
boundary of critical region and acceptance region is known as critical value. It
separates thecritical region and acceptance region. The critical value depends
on the level of significance, o, the risk ofmaking an incorrect decision regarding
rejection of true Ho:
One-tailed and two-tailed tests: When the critical region is in one tail of the
distribution of test statistic, the test is called a one-tailed test. If, in particular, it is in
the left tail (right tail) of thedistribution, then it is called a lefi-tailed (right- tailed)
is called a
test. When the critical region is in both tails ofthe distribution, the test
two-tailed test. Left-tailed (right-tailed) test corresponds to the left-sided (right
alternative.
sided) alternative and two-tailed test corresponds to the two-sided
non-randomized test
Randomnized and non-randomized test: A test is called a
test statistic falls in the
if we reject H, if and only if the observed value of the
critical region.
on the outcome
Atest is called a randomized test if'rejection of H,'depends
only if an even number
of arandom experiment. For instance, reject H ifand
appears in a die-throwing experiment.
non-randomized tests only.
In this book we con fine our discussion to the
A statistical test consists of the followingsteps:
1 ldentification ofappropriate population distribution and the parameters
for which the hypothesesare to be tested.
2 Formulation of nulland alternative hypotheses.
3. Statement of level of significance.
distribution
4 Selection of test-statisticand determinationofits sampling
under H, (known as nu.ldistribution of the test statistic).
sanmple
5. Determination of observed value of test statistic for the random
drawn.
478 PROBABILITY AND STATISTICAL INFERENCE: Theory and Practiee

6 p-value approach :Determination ofp-value.

Critical value approach :Determination of critical values andcritical
region for the preassigned level a.
7. Conclusion:
p-value approach: Reject H at level aif p-value s a.
Critical value approach: Reject H, at level a if the observed value of
test statistic falls in criticalregion. Do not reject HÍ, otherwise.
Thus the main parts of a statistical test can be stated as follows:
a nulland an alternative hypotheses (Ho and H,)
test statistic and its observed value
p-value or critical value
conclusion relating to the rejection or non-rejection of H, at a specified
levela
together with the practical significance of the conclusion.
In general,the conditions for validity of a test what we are
now areas follows:
going to discuss
1. The sample drawn from the population must be a random
sample.
2 The population distribution must be normal. Note that the condition of
normality can be relaxed if the sample size is large, due to the central
limit theorem.
Relationship between statistical test and confidence interval
Suppose that we want to test
HÍ :0=
against H,:0 o
at level a.
Let (L, U) be a 100(1-a)% conidence interval for the parameter 0, where
L and U are calculated based on the sample data.
The null hypothesis H, is to be rejected at level a in favour of the
alternative H, ifthe hypothesized value falls outside the 100(1-a
confidence interval (L, U).
Similarly, if we want to test
HÍ:0=
against H, :0> o
at level a, and a 100(1 - a)% confidence interval for is (-o, U), then H, iS
rejected at level a in favour of the alternative H, if the interval (-o,U)
does not contain On,the hypothesized value of 0.
 H,:6=6, where sarrple
Imay such Case Case where 18-10
H,:8e,.
From
By 16-5-2.Optimum
be that (Hence
b) <1-a
1-Pe(W)
(ii) (i) ie,
Also
To and
Proof.
go If Powerprove
a
vector. of
z11-Pe i-Pe isk
(") is ((x))
Neyman
size
6e 0<k<1.
MP Pa determined br
and given B,. s lÍk2 that
P,
UMPCR (W) Nevman (a)
("), the
is
Letfrom
n (W) Socritical (W) (W) 2a
of
W
W N Since
by:marginal Lo,8) unbiased
Wi
CR.
s Pe 1, W
we
Pearson Ta > we 0<k<1, If then sk(1 = =1-P 2 is =-Ir
W Regions
be a,have of region (W)z ka, unbiased, soi: Pearson
get populationa V6e
k[1
sufficient size from tk>0 that W
=(x: Lemma, = is kaz -
a), - Lo (xe
P(x:xe ie., (rLL(r is
distribution , unbiased. a then (f), the the
L(x, f, l and then Pa, a k>
dx W
= we size
8,) 8,Lemma,
)2
kP iH;) Pe <
Sutficient
statistic with
* from we 0 (W ) have tL tL MPCR
9,) the 6) also (W) get WI (x:xe of
=
8o Wis
(iv), = 2a the (x. (x. we
2 of p.m.f. the > Ho)] P [On to dx
Lo
= 8,)) FUNDAMENTALS
MATHEMATICA
8,) of OF
kl MPCR the (t for unbiased (xe show test have
(*)). Statistics. above a we =k W'I = size
(x, statisitc 6. or W, is (x (x:L,
’get: ;
O)), for h(x) Thenp.d.f. (1-a) H,) W'I L,z that:
ie.. Ly
: for a
proof Wis H,) for
testing T= byfx, Let
CR. kLy <k2kks0,
k>
=L and Lot, Latesting
l
(x). Factorization 0), x, holds unbiased ;<kll
[:OnW,
0 whereXy, Using
H, if H,:
: for
:0=, nanthe
6,
C.R dx (i)]

aga
Iheoren we
 Since and Thus and Here () test.the observed ond
interval
Solution.
Here
obtain

the
=1,
Example
ß
under
power
(ii) the Similarly, Solution.
(i) you
that distributiodenfined Hence INFERENCE-
Example STATISTICAL I
on .. value in
values the B(x:1sxs1-5) afunction = sizes
W 0:5
are if
termsof T=
basis18-2. Power =P{xe := = Ho of Here Sx, of 18-1. X1,
of =1- P(xe type B W W
(11)What
testing
x. fx , e)
X2, f(x)
type of Given of
Function
the Ifx 1, of I =P{xe = = (x = = we
a W and
I
single 21 TAx,
5

WI9=1}
WI fx , -La,
the and 05 P{0.5 P{xe
:xs05) x:0-5 want SxS1:5 w0uld 1 th e -
,Xp th
sufficient
e is
W
0) test type nul l
elsewlhere0,
the =x:84,=
=type f(x, is Sxs1 WI Sx}= marginal (x:go,
=Size
(x:x1) observation the 1-ß = 9) =2) 0,= = WIH}} to hypothesis be freq1uency
0SxsO ,
Il 0) critical Jo - = for =| 1- ll e) HÍ} test asthe
of errors. = =z1-P(xE
dx errors }lo-z | (x:x the statistic((x)) (t
TypeI 1 1 1-5
ß= e= = Ho: sizes
exp 1- sxs
0-75 are dx =P 1) P{x critical distribution
function (x)).
and
from region 0-75 1 = lAx =1, of HÍ
error W= x),
(- 1:5. = (xs 0-5 = 205 0-5) the := for k.> h
(x)
the for
0-25
= WI0=15 respectively j0-dx25 = 05
e)
Jo
0:541 against regions type : 9 8a,
{x:x<1) = population, 0testing -1 | fx, 0=1)
| 1 then((x), >
Sx<a, 0-75 =
0
against 1 of k.
[xe P 2) dx = HË:0=2. ?and ga,
Ho =0, 2) 0)J= =
Al so Tthe (t(x))
=k>0
WIand :
a
= P{0-5
H,:0=
type t(x),
=2 dx obtain MPCR
HJ HÍ
0:5 II h(«)}, .
:0=2, and < errors, 2, rather for
= against xse|0= the by k>
P[x B 1. pover means
if the
21 = dx you than
H the 0-25 = test 0
|=2] :0= alternative 0-5 1} function cho0se of the may
a 18-11
1. single joint
of th e be