0% found this document useful (0 votes)
15 viewstesting of hypothesis part -1
Testing of hypothesis
Uploaded by
biswalsubhalaxmi040Copyright
© © All Rights Reserved
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
15 viewstesting of hypothesis part -1
Testing of hypothesis
Uploaded by
biswalsubhalaxmi040Copyright
© © All Rights Reserved
Available Formats
Download as PDF or read online on Scribd
/ 47
CHAPTER 6
Test of Hypothesis ©
concerning the parameter sI
hhould be rejected or not
data, Our concem is tosee ifthe data support some
some preconceived idea regarding the parameter ¥
.e permissible arsenic level in ground water
's specification). Ifan environmental scientist
in a particular regi
safe or not, on the basis of different
the ground wat
water samples
content of the water, rather his concern will be to examine ifthe arsenic content
exceeds the permissible safety level.
Similarly, suppose that a pharmaceutical company is fermenting a vat of
antibiotic and they are interested in inferring about the mean potency fora
antibiotics in the vat, based on the sample collected from it with the purpose
examining if they meet the minimum government requirement. In such situation,
they are not concerned about the estimate of mean potency of the an
‘Their interest is to see if the mean potency exceeds the minimum governmeat
potency standards.
These types of problems involve the
wish to
we use the estimators ©
e drawn regarding the “ ,ypOTHESIS
pst oF wo
eee
‘we need to perform a
'297 mi then just on the basis ofthat esimate we
‘they underfill the bottles. Because, this may happen due to
"Another sample may give as an average of 299 or 300 or even
Sto al The difference between the sample average and the cetn My be
aa are to be the result of chance causes alone fits small. But theces
‘lowmuch difference can be considered tobe a small or nF50 A
ee how large the difference should be in order to make &
's soft drink bottles (vide Example 6.5)
“Many types of decision making problems can Be formulated as hypothesis
testing problems.This chapter will develop the ‘ee procedures fer several
jimportant situations. Now we discuss the basic concepts and terminology that
are applied in testing st hypotheses.
6.2 Basic Concepts and Terminology
This section introduces the basi terminolOBY 20° the
studying the tests of hypotheses.
-ompany underfill
concepts needed for
Statistical hypothesis: A
‘conjecture about the proba
It often involves one.
hypotheses are stated it
Since the eonject
hypotheses that are CO”
hypotheses.
Nultand alternative hypomneres.
‘A statement or conjuncture
seen aia 470__PROBABILITY AND STATISTICAL INFERENCE : Theory and Practice
as ‘H-naught
which is suppor
The wor:
ofthe paramete
A hypothesis agai
researcher's concern, The
sided alternative (i
and two-sided (no
Suppose that we wish to testa m
population mean j, given by
Hy: M= Hy
‘against alternative hypothesis
Hy: H> Ho
«, Hy: BSH
«. Hy: Wig.
Here H,and Hy are called one-sided alternative and H, is two-sided. To
bemore specific, H, is right.sided alternativeand H, isa left-sided alternative.
Here p19 s the value ofthe parameter specified by the null hypothesis. t
is called a hypothesized value. The value of jg is known.
Depending upon our objective or research interest we decide on the nature
of the alt i led or two-sided. Following
sided or left-sided or both-
hypothesis. (H,) about an unknown
hypothesis
Hy: w= 300
‘against lefi-sided alternative
Hy: < 300.
Here, his concern is to san content is 300 ml. or less (vide
Example 6,5), ofa particular brand ofa cigarette is
wean nicotine content forall the cigarettes
ular brand would probably be
mld then be worse for
agaist
eExample 6)
» According to 1991 census*, the percentage of female population in
sven Bengal was 47.84. A particular village in West Bengalis chosen and
Survey is eonducted to see ifp, the proportion of female population of
ith the state figure. The sample is collected to test
Hg: p= 0.4784
against H,: p #0784.
In the examples (i) and (ji) above we consider Hy tobe ofthe form b= Ho
weonable to state Hy as the inequality (say.
might have seemed
that village is same wit
Dut often it is more reai
11S Hy OF > Ho)-For example, inthe example (i) abo
esc nenstle to state the mull hypothesis, as the inequality. i.e» HoH 15,
‘The reason why still we go for the
‘quality is as fol
Ie is more easily comprehensible if
+ The development of a decision rul
there is only a single valucof 1.
«if sample data gives much more supporto the aterm Xe hypothesis
then j1> 1-5 would also be much more
H,w>1.Sthan to H=15)
plausible than $1.5.
+ Ifthe sample data sugges
of p> 1.5, we would cert
pols.
1 j= 1.5 should not be rejected in favour
ete oe oe
Simple and composi
Specified by the hyp‘
otherwise
* Consus of Inc
Econom ad le for deciding whether to reject or not
toreject Hp, on the basis of the sample data,
At the outset, we assume that the null hypothesis ( Hg) is true. Then we
‘want to see if the data contradict the assumption. We reject H if there is
enough convireing evidence against Hy.(H) is not rejected, otherwise. Two
possible conclusions of any test could be: ‘reject H” or “fail to reject Ho In
latter case, we say, ‘ Hp cannot be rejected’, but we should not say, “Hy is
‘accepted. The fact that * Hcannot be rejected” does not mean that *
‘accepted’. According tothe astronomer Carl Sagan, “absence of evidence is not
ence of absence”. In loose sense the term ‘acceptance’ is often used to
indicate ‘non-rejection’ of Hy
From the above discussion it is clear that we must be careful in setting up
the hypotheses for a test. A test is only capable of demonstrating a strong
support for the alternative hypothesis by rejection of Hy. But remember that
when Hy is not rejected, if does not mean a strong support for Hy; it only
indicates insuficient evidence or lack of strong evidence against Hy .Con:
the example ofcontents of softdrink bottles. There we test Hy
H,: <300.1f Hy is rejected, that indicates a strong support for *y < 300".
i.e, the mean contents less than 300 ml, whereas non-1
that there is insufficient evidence against Hy: = 300,
enough evidence to say that mean
later case we do not say that the
Errors in hypothesis test
In hypothesis rust OF HYPOTHESIS
ting the null hypothesis,
the null hypothesis,
lowing decision table shows all the possibilities:
Tounderstand the idea of type I and type Il errors consider the following
‘cxample of an accused person whois tried in a court of law being indicted for
committing a crime.
At the beginning of the tral the person is presumed not guilty. Here the
appropriate hypotheses are:
Hg: The accused person is innocent.
‘The person, actually, may or may not be guilt, whichis notknown. The null
hypothesis H, that the person is innocent should be rejected ifthere is enough
Level of significance
level of significance or the size 0
= 005 or 0.01
Actually, P(type Hero") 474 ~—s PROBABILITY ANY
‘power ofa tes. Its the power of test in detecting the falsity of Hg
Power ofa test =1—P (type I error) = 1B.
Wetry todesign
minimized. Bt it
simultaneusly. Therefore,
test subject to a
maximize power) fora fixed a,
ving examples illustrate how to compute probabilities oftype [and
type Il errors in significance testing,
Example 6.1 The fraction of defective items in a large lot isp. To test the null
hypothesis Hp: p= 0.2, one considers the number of defectives X ina sample of
items and fails to reject the null hypothesis if X < 6 and rejects the hypothesis,
‘otherwise. What is the probability of type I error of this test? What is the
probability of type II error corresponding to p = 0.01? What is the power ofthe
test for p= 0.017
‘Here number of defectives in a sample of 8 items, X, is bin(8,p).
‘We want to test Hy:p = 02
Test ruleis : Reject Hyifobserved X> 6, donot reject Hy otherwise.
P(typel error) =P (rejecting Ho | Hy is true)
= P(X >6|p=0.2)
= P(X = 1|p=0.2)+ P(X =8|p=0.2),
since Xis bin(8, p=0.2), under Hy
-(') «(0.2)" xos)+(°) x(0.2)° x (0.8)
=8.(0.2)7 x0.8+ (0.2)®
=0.00008448,
P (type I ervorp =0.1) = P (not rejecting Hy |p = 0.1)
= P(X <6|p=0a)
=1-P(X =1|p=0.1)- P(X =8| p= 0.0)
S 1-(8) (0.1) x. »)-(8) x(0.)* x (0.9)"
—— ee
restormnvommess _at
=1-8x(0.1)’ 0.9-(0.1)*
=1-0.0000072 -0,00000001
=0,99999927.
‘the power of the test corresponding to p = 0.1
=1-P (type error |p = 0.1)
= 0,00000073.
Example 6.2 To examine the claim ofa reputed publisher that m, the average
rember of misprints per page ofa book, is 1, if particular page ofthat book
rentains more than 2 misprints, then the null hypothesis that m= 1 is rejected.
‘SMhat is the probability of type I error ? Find the power of this test when the
alternative hypothesis is m= 2. Given that e~! = 0.368.
Let Xe the number of misprints per page ofthe book. Here Xis Poisson (m)
test the null hypothesis Hy : m=1
if observed X> 2; donot reject Hy otherwise.
We wan
The test rule is: reject Ho
Then
P (type error) = P (rejecting Ho | Ho is tue)
= P(X>2|m=1)
=I P(X =0|m =1)— P(X = Im =1)- P(X =2Im=1)
° 2
=1-0368x25
=1-092
=0.08.
Power of the test when
Power= P(rejecting Ho | Ho is false)
= PA > 2|m= 2) since alternative hypoukesis Hy:m=2
ternative hypothesis H, :m=2 is given by
=1-0135424%5
=1-067712
= 0,32288, k ~ ‘pumber of mispri
ontains more than 2 misprints then the
‘What is the probabil
smple 6.2 To examine the claim ofa reputed
publisher that m, the average
fa particular page of that book
rypothesis that m= 1 is rejected.
he power ofthis test when the
per page of a book,
of typeI error?
sltemative hypothesis is m= 2, Given that e! = 0.368.
‘Let Xbe the numberof misprints per page ofthe book. Here Xis Poisson
We want to test
The test ruleis
Then
P (ypelenor) —
Power ofthe test when alternative hypo
= PX
=1-P
the null hypothesis Hy: m=1
iffobserved X> 2; domot reject Hy otherwise
reject Hy
P(eajcting Hy | Hy steve)
=P(X>2\m=1)
= 1-P(X = 0]m=1)— P(X =Ilm=1)-P(X =2/m=1)
=1-0368%25
=1-092
=0.08.
>2|m=2)+
(x = 0jm=2)~PX = 476 PRO!
Example 6.3. Aco
reject Hy if observed X 29.
cting Hy | Hp istrue)
=PX29|p=4)
=P(X=9)+ P(X =10), when p=!
1y/1y | (10 1" sy
“(9 G 0X2) (2,
=H -o01074> 001.
Tora
Note that the size ofthe tet or P (type Terror) < c, the evel ofthe test.
Therefore, we cannot take 1% as the level of significance here. But we can
carryout the test at 5% level, since P(typel error) = 0.01074 <0.05.
‘Note: Actually we define our testrulein such a way that the level=a. Here the
testruleis such that thelevel cannot be 1%. Itis possible to take a = 005 here.
Test statistic: A statistic, a function of sample observations, used to test the null
hypothesis is known as atest statistic. A conclusion toreject or not to reject Hy
'is based on the value ofthe test statistic, observed in a sample selected
‘There are two appreaches for making a decision regarding rejection
rejection of Hy, viz,
* p-value approach
clusion. (vide Examples 6.4, 05.
will take the value as
erved, when Ha is true. It
extreme as or more extrem
isalso known as observed level o}
— ee ee
st oF HYPOTHESIS m
rest oF OT
tical value: The entire set of
jo two sets or regions. One region
hypothesis and leads to
ion, The other
the rejection of Hy is it ‘or rejection
of the values that support Hy is called the non-rejectic
ice region), The value of the test statistic wt
region and acceptance region is known as critical value. It
al region and acceptance region. The critical value depends
ince, 0, the risk of making an incorrect decision regarding,
rejection of truc Hy.
One-tailed and two-tailed texts: When the critical region is in one tail of the
isribution of test statistic, thetestis called a one-tailed test. If, in particular, itis in
the eft (right led aleft- tailed (right- tailed)
test. When the 9 the distribution, the testis called a
nwortailed test. 1) test corresponds to the left-sided (right-
sided) alternative and two-tailed test corresponds tothe two-sided alternative.
Randomized and non-randomized test: \estis called a non-randomized test
ifwe reject Hy ifand only ifthe observed value ofthe test statistic falls inthe
critical region.
‘A esti called a randomized tstif‘rejection of Hg depends on the outcome
ofarandom experiment. Forinstance, reject H ifand onlyifan even number
‘appears ina die-throwing experiment.
Inthis book we confine our discussion to the non-randomized tests only,
stcal test consists of the fe
1. Identification of appropriate popu!
for which the hypotheses are 0 be
wing steps:
distribution and the parameters Co
significance a for which
rhe pratveis
Hycan be reject C a type I error, if Hy is
rejected based 0” eee eee The p-value measures
Werreject Hy at level «. ifp-value < a.
thestren
‘Then we ca FOP cally significant at level a.
itical value: The entire set of
non-reje
10 two sets or regions. One region
Gritical regions
values of test
ng afthe values that support the alternative hypothesis and leads to
of Hy .d the critical region or rejection region. The other
cf the values that support Hy is called the man-rejection region
-eptance region). The value of the test statistic lies at the
ritical region and acceptance region is knowns 7
‘ical region and acceptance region. The I value depends
the risk ofmaking an incorrect decision regarding
{ofien called
boundary oft
separates the!
onthe level of significance, ot,
rejection of true Hy.
One-aied and two-tailed tests: When the critical region i in one tail of the
rrcbaton cf test statistic, the testis called a one-tailed est In particular, itis in
thelefttal (right tail) ofthe di led a left- tiled (right- tailed)
test. When the: egion i in both tails ofthe distribution the test's called a
poo-tailed test, Left-tailed (right-tailed) test ‘corresponds to the left-sided (right-
1 two-sided alternative.
sided) alternative and two-tailed test corresponds to the
ized test: A testis called a now-randomized est
1 value of the tes statistic falls in the
Randomized and non-random!
if we reject Hy ifand only ifthe observe
iitical region.
"Atestiscalled a randomized testif TejecHO of Hy” depends.on the outcome
ve reject Ho ifand only fan even number
ofa random experiment. For instan
ing experiment.
appears in a die-throwit
In this book we confine our di to the non-randomized tests only
cal test consists ofthe 1
iscussion
A statisti fpution and the parameters
distribution
msample 478. PROBABILITY AND STATISTICAL INFERENCE : Theory and Practice
mination of p-value
6 p-value approach : Deter
‘ation of critical values and critical
Critical value approach : Determini
region for the preassigned level «
7. Conchision
povalue approach: Reject Hy at level ot if p-value < a.
Critical value approach: Reject Ho at level o: ifthe observed value of
test statistic falls in critical region. Do not reject Hp, otherwise
Thus the main parts of a statistical test can be stated as. follows:
+ anulland an alternative hypotheses (//y and H,,)
* test statistic and its observed value
+ p-valucor critical value
conclusion relating to the rejection or non-rejection of H, ata specified
level «together with the practical significance of the conclusion
In general, the conditions for validity ofa test what we are going to discuss
now are as follows:
1. Thesample 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
Hy:0= 0
against = Hy 0 # 6
at level a.
ie aaa oy) ee ! seo a. G M% confidence interval for the parameter @ , where
n the sample data,
The null hypothesis Hy is to be rejected at level a in favour of the
alternative H, if the hypothesize i
a ypothesized value ls out = a)%
confidence interval (L, U) renee iS
Similarly, if we want to test
Hy :0= 0
against H,:0>0)
is ee anda 100(1—a)% confidence interval for @ is (-20,U), then Hy is
FA ce at level in favour of the alternative H, if the interval (-«0,U)
oes not contain @, the hypothesized value of 6.
— * TEST OF HYPOTHESIS -
--~-s _s
And if we want to test
Hy :0 = 0
against H, :0<%
atlevel «anda 100(1~a)% lower confidence interval for @ is (L,%), then
Ho is rejected at level « in favour of the alternative H, ifthe interval (L,~0)
does not contain O
In general we can say that the null hypothesis Ho should be rejected at
100% level if 100(1-c)% confidence interval does not contain the
hypothesized value of the parameter.
6.3 Test for Population Mean
Let x), ),..-,%, bea random sample of sizen drawn from the population
of the random variable X which follows a normal distribution with unknown
mean and variance o”. We are interested in testing the hypothesis
Ho: b= Ho
against alternative Hy: > Ho
or, Hy: 1 480 PROBABILITY AND STATIST Theory and Practcg
which is equivalent to
P(Z> zp), _ when alternative hypothesis is H,
P=}P(Z|zo|), when alternative hypothesis is 1,
TH. Conclusion: Reject Hp at love, er ifp-value < cdo not reject 1,
otherwise,
Critical value approach
I Compute the value of test statistic
¥-Ho
Zee o
olvn
for the given sample observations. We denote the observed value of Z
by =
TL Obtain critical value (from the standard normal probability table)
(@ =q, ifalterative hypothesis is one-sided, (i.e., H, or Hy )
Gi) 24,2, ifalternative hypothesis is two-sided, (ie., H).
IL. Conclusion: Reject Hy at level if
© 79> 2q, when alternative hypothesis is right-sided, (i.e., H,)
G) 2) <~Zq, when altemnative hypothesis is left-sided, (ie., Hy)
(iii) | 2o| > zay2> when alternative hypothesis is two-sided, (i.e., H)
The region of values of test statistic described by (i), ot (ii), ot (iii) above
defines the size-o. critical region for the test, as the case may be, since the
Probability of type J error for this test iso.
Usually a is taken to be 0.01 or 0.05, Note that the rejection of null
hypothesis H, at 1% level implies the rejection of Hg at 5% level. But the
Converse is not true. On the other hand, the non-rejection of Ay at 5% level
implies the non-rejection of Hy at 1% level. But the converse ig not true.
Now with two-sided alternative, we will show how the above test rule is
developed
We want to test
Hy H= hy
against H,: yt 14,
atlevel a.
a est OF HYPOTHESIS
eT OE Ee eee EEE Ea
0
(@)
FosZo) Z>
0.025 4 0.025
aw
$3 3 2
vw
ry
°
Critical regions for (a) right-sided, (b) left-sided and (c) two-sided alternatives
Since x, >, ....x, bearandom sample drawn from V(j,, 6)-population,
31% -+-5%q ated, as NK, 0”). Hence, the sample mean ¥ ~ N(j1, 2)
Under Hp (ie., if Hp: =u istrue), F~ N (uo, aay
x
~Ho
Hence Z==—52 ~ 1 (0,1), under Hy.
ap Oe
Thus when Ho: =o is true, we may write
¥-Ho
Pagltapa $y Sten) =18 M
ie, Py, (Zap $ZS Zap) =1-@
which imples Py (Z <—zaj) ot Z> Zaj2) = % A
Here we are performing all tests at level a which means
P (type terror) «
oF, P (Hy is rejected | Hy istrue) = Pi (Mo isrejected)= a (ii)
egior e test Ss
Now, from (ii) and (ili), the size-at critical region for the testi
(wy
Zy <~Zu/2 oF 70
We fail toreject Hy at level « if ‘
— Zap S20 $2uj2
: 1 statistic Zander Hy
Where z, is the observed value of tes
31 ; : io — FETE
482 PROBABILITY AND STATISTICAL INFERENCE : Theory and Practice
critical | non-rejection | critical
region“ region ~{~ region
7a
seo . critical critical oe ;
<€non-rejection regio! region © region Aron rejection region-+
Similarly, we can develop the test rule for one-sided alternatives also.
For right-sided alternative, H,: > i,we can have the size -ar critical
region Z>z,from Py, (Z>Zq) = @. And for left-sided altemative H: 11 <4,
we can have the size-ar critical region Z <-zq from Py,(Z <-zg)=@.
From the following diagram (showing the distribution of test statistic under
H, (aul distribution)) it will be evident why p-value approach and critical-
value approach are equivalent and lead to the same conclusion:
Consider the problem of testing Hy: 4 = Hy against Hy: 4.> M-
OMe eis tezey
critical
value
Reject Hy if 2
(cy = observed valuc uf
the statistic)
Reject Hy if
p-value = P(Z > 29) $@
which is possible only if 482 PROBABILITY AND STATISTICAL INFERENCE : Theory and Practice
critical | non-rejection || critical
region region region>
=z, 0
a
Sai Z critical critical Pee ,
aoa ecLLony resion-| e7iteal region A MOn-rejection region->
Similarly, we can develop the test rule for one-sided alternatives also.
For right-sided alternative, H,: 11> 1,we can have the size -c critical
region Z>2,from Py, (Z > 2,)= a. And for left-sided altemative Hl: 11< Up,
we amhave the size-cr critical region Z<—z, from Py, (Z <—
gq) =a.
a
From the following diagram (showing the distribution of test statistic under
Ho (ull distribution) it will be evident why p-value approach and critical
value approach are equivalent and lead to the same conclusion:
‘Consider the problem of testing Ho: = 1 against Ay: u> ty.
0 24 zy Z>
critical
value
Reject Hy if zy >2,
(59 = observed valu uf
_the statistic)
Reject Ho if
p-value = P(Z > 29) Fa -¢ OF HYPOTHESIS
TES 483
similarly, this’ equivalence can be shown for: left-sided, and both sided tests too,
Now we show the probability of type I ili
sen (B) diagraratialy. pe L error (cz) and the probability of type
Intesting Ho: =Ho against H,: 11> py, we have seen that
= P(type L error) = P(Hy is rejected| H, is true)
= P(Z > zy |H= Ho)
: ~ (Zif>sain=r6]
=°(F>u+texZiu=m)
a
Ho Motte
In
non-rejection | critical
region region
=?(2> 19+ x2)
on:
Critical region for this test is
c
F > My tig X=.
ota XE
| Now, B= P (type Il error) = P ( Ho isnot rejected| Ho is false)
Thus
Ee o
peau=)=P(# Ho) -
X>
BL
Bo tage
< nonerejection region fe eritical region +
% confidence interval and a size-ct test for \
at the same decision regarding rejection
) % confidence interval for
Relationship between a 100.0
mention thi
In this connection we must
or non-rejection of Hy can be had from the 100(1 - a
H also. |
—a)y idence interval for
ae veered at tevel a Ho
should be rej
es not contain the hypothesized
should not be rejected,
Value pig, then Ho i
otherwise. :
“efi in
484 PROBABILITY AND STATIS! ENCE : Theory and Practice
The explanation is given below:
Inchapter 5, we have seen that a 100(1 ~ «t) % two-sided confidence interval
for p is
o . o
F-tap XP» pa (i)
which is obtained from
Fm
of-ian < air Stap2 }
which is, again, equivalent to
Hu :
P| -z, S24). |=1-a,under Hp: “= Ap.
( a a /: a | 0 0:
On the other hand, we know that the non-rejection region (for testing
Ho: =o against H,: 1 # Wo at level a) is
F=Ho
=2Zqp SA <2,
Peg
which can be rewritten as
= o =a o
E~Faja XPS My SE Zapp “oR (ii)
Thus from (vi) and (vii), we can say if 100(1-a@)% confidence interval
= cm c PB
(Freak, + Zep contains 44, Hy should not be rejected at
level a.
Similarly, we can establish the relationship between 100(1- a) % one-sided
confidence interval and the one-sided test of size @,
Now the question may arise, if we can reach the decision regarding rejection
ornon-rejection of Hy from the confidence interval, then why we need to study
the procedure of hypothses testing also. The reason is, there are many situations
(such as Pearsonian ?-test) where confidence intervals cannot help in making
such decision,
In this discussion of hypothesis testing, our basic assumption was that the
Sample is drawn from a normal distribution.
The test is no longer valid if the assumption is not satisfied. But if the
sample size is large (m2 30), then normality assumption is not required, as the?
by CLT, the distribution can be approximated by a normal distribution no matter
if the population is normal or non-normal. esr oF HYPOTHESIS
sO ts
as
4 Arandom samp ,
xample 64 Arandom sample of 25 workers is
amperage monthly Wage isRs.1,400 or not. The pone a factory to test if
fined to be normal with a standard deviation of Re20G¢ qs
i ly wage based onthe 25 observations comes out tobe 200. The average
rat level ose ifthe data suppor the Rs.1,500. Perform
I value approach, (c) confidence interv
Tatu be the average monthly wage of all workers
ers
‘We want to test Ly
1 Null and alternative hypotheses:
We want to test
Hy: »=1400
against —H, : w#1400
at 5% level of significance.
IL. Test statistic:
The observed value of test statistic is
nee
ie
1500-1400
- (here ¥ =1500; np =1400, = 200, n=25
00/25 a »
=25
IL. (@) p-value approach:
For this approach we need toeglulatep-vale forthe test.
p-value = 2P(Z >| 70D
=2P(Z>2.5)
= 21-025)
= 2(1- 0.9938), rom Table (in Appendis)
=0.0124
© Critical value approach: ee:
og this approach we need to find crial YANG Slt
= 1.96.
From TableA, 40/2 = 70%
NV. Conclusion: proach: Since pvalue=00124<005, Hy should
(a) Using p-value oa
° Te rejected at 5% level in YON fH, ICAL INFERENCE : Theory and Pray
eA Poche
Since 29 =2:5> 1.9625, 4,
level of significance, in favour of
486 PROBABILITY AND STATI
at 5%
Interpretation: Here rejection of Hy implies that the data do no
the claim that the mean monthly wage of the workers ofthe factory is
wz)
0.0062
25-196 0 1.9625 77
‘Note: If, instead of population s.d. the sample s.d. is considered here, the tes
‘procedure will remain the same, since the sample size is large
(©) Here 100(1-a) =100(1-0.05) = 95% confidence interval (two-sided)
for y is
= © 200
BF Zapp XZ = 1500 29095 Xe
ean 05 Bs
=15003196% 2%
= (1421.60, 1578.40),
Since 95% confidence interval for } (Rs.1421.60, Rs.1578.40) does 01
Contain the hypothesized value of which is Rs. 1400, H, should berejectedst
5% level of significance, in favour of the alterna
Example 6.5 A soft-drink company claims thi
drink bottle is 300 ml, as advert
underfilled. He measures the con
297, 298.1, 2988,
2953, 3015,
‘machinery. Assume that thecont
a standard deviation of 3.5
approach and cr
for the population mean and hence comment on the d —
est OF HYPOTHESIS lan
es
Lat ube the mean content ofthe soft-drink bottles and be the sample
wan.
Here sample size n= 10, population standard deviation @ = 3.5 ml
= _ 2991
‘sample mean, ¥ = a = 299.19 ml
Test rte:
Here we test Ho: 1=300 against H,:1<300 at « = 0.05.
Observed value of test sta
s
=, - ¥oHe _ 299.19-300
Gdn 35/0 =-073. 0.2327
+ pevalue approach: =
p-value = P(Z < z9), Z~N(, 1) 073 0 =
= P(Z <-0.73) = 0.2327.
Since p-value, 0.2327 > 0.05, the level of significance, H, cannot be rejected,
‘which indicates that there is no convincing evidence that the mean content is,
below 300 ml.
Critical vatue approach:
For a= 0.05, critical value zq = Zp.5 = 1.645.
Here 2 =—0,73>-1.645=~z95. Since zp >—295,Hy cannot be
rejected at 5% level.
=1,645 073 0 =
tof the soft-drink bottles
‘Thus we are not convinced
isless than 300 ml.
Note that
Here, a 95% lower eonfiden:
interval for
o
2005 *
35
Sx
ie, (299.191
re (297.37, )
‘This interval contains the hype
the null hypothesis Ho :# = 30
cannot conclude that the me:
300ml.
prample 6 Tew
supposed to be 1.5mg Per
EI
are measured. The sample meat
‘The population standard
evidence that the average
‘using p-value approach and value approach. Also draw conclusion
from the appropriate confidence interval.
‘Let Xe the nicotine content ofa randomly selected cigarette of this brand
“The mean and standard deviation of Xare yt and o. Here o =0.5mg. Since the
sample size, n=40, the central limit theorem can be applied. Hence the sampling
distribution ofthe sample mean ¥ will be approximately n
jormal
‘Here we want to test
Ho:u=15
against = Hy >1.5.
at 5% level
+ The observed value of test statistic is
E-Wy _1.62-15
7 Gidn 05/140
=1.52.
+ povalus =P(Z>%), Z~ NO,
= PZ >152)
= 0.0643,
© Conclusion: Since p-value = 0.0643 > a = 0.05, we
Since 2 <2y9s, wecannot reject Hy at —— ateelines
rorHEts
a
re. 0 95% upper confidence interval for yi
(-m, ¥ +2005 a)
eres
- Tao)
x, (21-75) mg.
Since this interval contains the hypothesized value 1.5mg, Hy should not
te rejected at 5% level
Example 6.7 The daily yield aa chemical plant, recorded for 50 days, possesses
-samplemean of 84 tons. The population standard deviation is known tobe42
tons Do the data provide enough evidence to indicate that the average dsily
tld of the chemical plant is 900 tens per day? Test at 1% level of significance.
amy oat the test using p-value approach and critical value approach. Also
make a decision using an appropriate confidence interval for average daily yield
ofthe chemical plant.
Let ube the mean daily yield ofthe chemical plant. Here sample size
n=50, sample mean % = 884 tons, and population s.d. 0 = 42 tons.
Since n is large, the central limit theorem can be here, ‘Thus the
sampling distribution of can besssumed tobe approximatelynormal
y= 500 against H,: n#900at level a= 001
+ Wewant to test Hy
«The observed value of test statistic is
a2oHoe get) = -2.69.
Nn SO
+ pevalue =2.P(Z >| 20D» 2 ~NO,)
=2.P(Z > 2.69)
= 20,0036
= 0.0072,
which is less than 0.01, the preassigned level ofsigoificance
since p-value is less tha
* Conclusion: Since P-"
“Jicates that the data
fers from 900 tons
ne eritical value approneh
value would PE
27905 2575
zap = 70005 PO OO
490 PROBABILITY AND STATISTICAL INFERENCE : Theory and Pry
critical value, we rej
Since | zp 2.69 > 2.575
I
is piven by
2
Here 99% confidence inter
8845 2.575%
8.71, 899,29) tons,
Since this }ot contain the hypothesized value 900 tons, 4
Note: For your convenience, you may remember the following criti
(e-values), since they are frequently needed in practice:
Case 2: unknown
Since ois unknown, we estimate o by the sample standard deviation »
‘where sample variance 3? is
variance 0” unbiasedly.
@ isreplaced by s, then the sampling
will be a distrib
the test rl
against
at level «., will be as HYPOTHESIS
est OF 9
—
roach
value apP?
L Compute the observed value of test statistic:
mee)
svn (62)
%
1 Obtain p-value as
when alternative hypothesis is H/,
=(PU>%0), when altemative hypothesis is #,
PUt| 9 |), when alternative hypothesis is Hy.
ILL Conclusion: Reject Hy
at level a if pevalue < a do not re
otherwise “
Gritical value approach
L Compute the value of the test statistic:
(fiom t-table in Appendix)
ve hypothesis is one-sided, ie., H; or, H.
ative hypothesis is two-sided, ic., Hy.
0
IL Conclusion: Reject Ho atlevel a if
@ > fa:n-1, when alternative hypothesis is),
ive hypothesis is Hy,
int ofa tistribution with df = n—1
GB [40> Fajen, when al
where fain-1 istheupper 1000%
We cannot reject Hp, otherwise.
Example 6.8 The weights of s
Process are measured to evaluat
(in carat) ae as follows:
002, 045,
yonds generated by a new
the process. The weights
the
Selected diamond generated by theP
- fing ion (a). Do the above data provide
in weight of the diamonds produced
using c= 0.05 by p-value approach
unknown mean (1) and stan
sufficient evidence to indicate th:
by this process exceeds 0.52 carat
and critical valueapproach.
Heren=7.
“The sample mean (¥) and sample standard devia
y-26 20537 carat
5 =0.061 carat
‘+ Here we want to test
Hy: w= 052
Hy: w>052
are as follows:
against
atlevel = 0.05.
«Observed value ofthe test statistic is
F— by _ 0537-052
a Te “067 O74.
+ pralue=P (120.74), follows distribution with af = n—1~6
From stable, corresponding tothe df= 6, we have
P(e> 1.440) =
Thus p-value, P('> 0.74) > P(t> 1.440) =0.1> 0.05 = a.
Hence Hy should not be rejected at 3% level indicating thatthe data do mot
provide sufficient evidence to indicate that the mean weight of the diamons
produced by the new process exceeds 0.52 carat
0 0.74 1.4404
For critical value approach, we have to compare the observed value of te
statistic (9) with the critical value
os; = 1.943 (from stable in Appendix)
fae
Here fp =,0.74<1.943= fo,
Hence Hg should not be rejected at
favour of the fact that the mean weight.
jcating lack of evidence i?
jonds exceeds 0.52 carat >» e
493
Means of Two Independent Populations
1 xj figeeig, Bea random sample of size m drawn from
Se normal random variable X; having unknown mean y and
Tome bet HoTnzw-r¥am bea random sample of size m drawn from
i another normal random variable X; having. unknown mean
spe poration of
and variance 93. Therandom variables X, and X> are independent. Let
teiky be the respective sample means
‘We wish to test the null hypothesis
Ho: y-#2 =
aesinst one ofthe possible alternative hypotheses
Hy: sha 20
Hy: #2 <0
Hy: 4-H #0
at preassigned level of significance
Equivalently, we consider the test of
aan > He
* Hy 0
Fy: My ba <0
Hy: ~My #0
atpreassigned level of significance a.
Equivalently, we consider the test of
Ho: y= #2
against Hy: My? Ha
or, Hy: wy < Po
or, Hy: yh
atlevel a.
Three cases will be discussed here:
1. of, 03 areknown
2
are unknown, but equal
2 yo}
3. of, 03 areunknown, and 402
Case 1: o},03 known
Here x, ~N
Xy~NG 8)
X,and X, are indet
spendent.
be PROBABILITY AND STATISTICAL INFERENCE : Theory and Proce
494
SS (S608
5, and ¥, are also independent.
Hence:
@&-%)~ af =H), SE
m
The standard error of (% -%) is
Thus the test-statistic
=SB=) -H2)
2 of
ie standard normal distribution. Under H (i.e., when the null
follows N((
hypothesis H: 4-12 =0 is true),
2th
camer 63)
ot
mom
(Note: 16 instead of testing Hg: 4, —p =0, we test Ho: 4 —Hy =6.i0
Beneral, then, under 1, the test statistic would be
G-%)-8
z= ]
a Gh
SL, 2
mm
The test rule is based on the observed value of the Z-statistic. The steps 1°
as follows:
p-value approwh
1. Compate the observed value of the test-sta ; TEST oF HYPOT
495
not reject Hy
Critical value approach
L_ Compute the observed value of test-st
under Hy
=e
2
si.
IL Obtain critical value of from standard normal probability table Table A,
in Appendix)
2q., ifalternative hypothesis is one-sided, ic., Hy or Ha
4/2, ifalternative hypothesis is two sided, ie, Hs
UL Conclusion: Reject Hy atlevel a if
() 29>, in favour of alternative 1,
ar, (ii) %<-zg, in favour ofalternative Ha
x, (iii) |zo]>Za/2 ,im favour of alternative H3,
as the case may be.
The region of z-values described by
critical region for the test.
above defines the size-a
Note: If g?,a3 are unknown, but sample sizes are large enough (i.e.,
n, 2 30,n, 2 30), then same z-test can be performed, just by replacing unknown
7, G3 by the sample variances 5? ands}, respect the expression of
the test-statistic.
Example 6.9 A dietitian has developed
tandom samples of size 15 of obese
Soup is placed o 496. PROBABILITY AND STATISTICAL INFERENCE : Theory and Prucee
eS
new low-fat diet is more effective in reducing weights compared tothe existing
cone. Take & = 0.05
‘New low- fat diet Existing reducing diet
m=15
%=23 kg
‘Assume that the weight losses for the two dicts, X, and X5 (X, for the
new diet and X; for the existing diet), are independently normally distributed
‘with standard deviations 3.4 kg and 2.7 kg, respectively.
Let sx, and 2) be the mean of the population of weight losses for obese
people who could be placed on the new low-fat diet and the existing weight
reducing diet, respectively. We wish to test
‘Ho: Hy = My (indicating that both the diets are equally effective in
reducing weights)
against H.: 14 > Mz (indicating that thenew low-fat diet is more effectivein
reducing weights)
at 5% level
‘The observed value of the test-statistic, under Ho,
5% 45-23
From Table A, we get the critical value, 29.95 =1.645.
Since zp > zoqs , wereject Hy at 5 % level in favour of alternative Hq»
and conclude that the new low-fat diet is more effective in reducing weights.
Example 6.10 A survey was conducted to sce if the smokers have high pulse
smokers and non-smokers are selected; a
are 90 and 88 per minutes for the smokers and non-smokers, respect
sample standard deviations are $ and 6, respectively. Test at a =0.05 if
smokers have higher pusle rates.
Let 11, and 1 be the population means of pulse rates of smokers and 108-
os: syPOTHESIS
est OF
sors respectively We want otst
Host =a
aginst H,
atlevel a = 005,
Here, the sample sizes m7, , re large. =100, ny = 100. The population
sariances, 07 and 03, are unknown, But we ean
thesample variances s? =25 and 3 =36, respectively.
Thus the observeé value of test
sp =H = 0-885,
eae eo
mm, V100 100
The critical value, 2995 =1.645.
Since 2 > Z9 95. WeTeject Hy at 5% level in favour of H,,, and conclude
that there is enough evidence that the pulse rates are higher for the smokers.
Case? = 6},63 unknown, but equal
Let the populations be homoscedastic (having equal variance),
i.e? =03 = 0? (unknown).
Here we estimate 6” by a pooled estimator
im st +m, -DS
where sf and 33 are the sample variances given by
2 OF HYPOTHESIS
Fe mokers, respectively. We want to test
wi =H
against H,
atlevel a = 0.05,
sample sizes m,n, ,are large. n, =100, = 100, The population
variances, a7 and 3, are unknown. But we can estimate them unbiasedly by
the sample variances 57 = 25 and 53 = 36, respectively.
Thus the observed value of test
-Ack
2?
st 3
mm
The critical value, zp 95 = 1.645.
Since 29 > Z905, wereject Hy at 5% level in favour of H., and conclude
that there is enough evidence that the pulse rates are higher for the smokers.
1,03 unknown, but equal
Let the populations be homoscedastic (having equal variance),
ie,o7 =} =o? (unknown).
Here we estimate 07 by a pooled estimator
2 _ (m—Dst +m — DF
Note that both 5? and 53 are the unbiased estimators of 6” , the common
Variance,
Here the test-statistic (known as Fishe
¥ Fi) =-(
t-statistic)
‘lows ¢-distribution (Fi
2
sj s
where y=! and v, ==.
™ Mm
This approximation gives good results when m, and) both are 5 or larger.
As we mentioned before (in chapter 5), this procedure is known as
Satterthwaite's approximation or Welch's approximation method.
Under Hp, the test-statistic becomes
Xi =%
t= (63)
Fe
1 2
mm
1 Compute the observed value of test-statistic, under Hy :
%)
nm
U. Obtain the critical value ¢, from table ‘corresponding to the d.f
level = a as
@ fa:v, ifalternative is directional, ie. Hy ot H
or, (ii) 44/2, v» alternative is non-directional, ie., Hy
I. Conclusion: Reject Hy atlevel a if
@ fo7la:y, in favour ofalternative A,
or, (ii) 19 <~la;
or, (iii) \tol >4ar2; »» in favour ofalternative H,,
v» in favour of alternative Hy
as the case may be, Do not reject Hy, otherwise.
Ef PROBABILITY AND STATISTICAL INFERENCE : Theory and Practice
follows a distribution approximately with d.f. v given by
2
(s? 33)
Se eee
2
mM ( +¥2)
Va Chea aS ie ETE
(gy (3)
Gel
yl
i 3
where ==! and vy) ==2.
n ny
This approximation gives good results when 7, and, both are 5 or larger.
As we mentioned before (in chapter 5), this procedure is known as
Satterthwaite’s approximation or Welch's approximation method.
Under Hp, the test-statistic becomes
(6.5)
1 Compute the observed value of test-statistic, under Hy:
(X — X3
tp = ZB)
st
m Nn
I Obtain the critical value 4, from t-table corresponding to the df= vand
Jevel = a as
@ fa:v, ifalternative is directional, i.c., H, or Hy
or, (ii) faz; »» if alternative is non-directional, i.e., H,.
Il. Conclusion: Reject Hy at level a if
@ fo > la:v, in favour of alternative 1,
or, (i) 4) <~la; > in favour of alternative Hy
or, (ii) Mo] >4ai2; y» in favour of alternative H,,
as the case may be. Do not reject Hy, otherwise. est OF HYPOTHESIS 501
Ti pepulations having unknown varia
aeee rens-Fisher problem.
as.
In this context it may be noted that iftwo Population variances ar:
one is known and one unknown, then there will be
ls ssions of test statistic and in degrees of freedom,
a
Population means of two independent
nces (unequal) is, in general, known
‘e unequal,
a slight change in the
Ifo is known and 6} is unknown, then, under Hy, the test statistic
i
becomes
Note:
1. 6} #63 (both unknown)
For obtaining the null distribution of the test-statistic
A=
2
2 fo
a a)/(tesh),
m mii \m Mm
? distribution, if of # 03. However,
have an approximate % -distribution, if oj # 03
But it does not
ethod for approximating the x7-
it 946)* proposed a mi
Satterthwaite (1941,1
distribution.
4), “Sythe of Variance”, Pechomenrita, Yo. 6,pp.309-316
* Satterthwaite, FE: “ es “An Approximate Distribution of Estimates of Variance
ite, FE. (1946), he
Sete Bulletin, Vol.6, pp. 110-114,
Components”, itt
502 PROBABILITY AND STATISTICAL INFERENCE : Theory and Prac
ctice
33
nm
PK ~ 16, approximate;
(zi . =) a Ys
The method says,
2
nn
where v isobtained by the method of moments. By equating variances we get
a ore 22) rer 2)
1 2
where vy is obtained by matching the variances as follows: Pr
‘TEST OF HYPOTHESIS
503
SS
(3-4
v’ 20}
4
a =2v, since Var(s}) =
a) n(n, -1) {
204
mot
422
ny
Example 6.12
brands of | coffee, 504 __ PROBABILITY AND STATISTICAL INFERENCE : Theory and Prac
iii: —
450-gm jar witha standard deviation of 0.67 mg, Ass
of caffeine contents of coffee of two brands a
variances, test at 5% level whether the caffeine cont
the two brands.
Let 11 and 13 be the population mean caffeine contents of the coffe of
‘nwo brands (per $0-gm jar).
Here nm, =10, ny =15
¥=9.2 mg, % =81 mg
5, =0.52 mg, 5, = 0.67 mg
We want to test
Ho: =H2
against Agi by tha
atlevel a= 0.05.
‘The observed value of test-statistic is
~ Goh)
;
——— ih By
V0.027+0.030 0.239
From table, we have to obtain the critical value of f corresponding to the
probability % = 0.025 and degrees of freedom v,
i
2 2 2
where y= Ui") Si 3
a
ye ™ my
(0.52)
ye =0.
70 770.027
(0.67)
y=! oz = 0.030.
Hence the critical value,
apr; v = o0ns; 22 = 2.074 ‘ST OF HYPOTHESIS
sos
Since fo > faons. 22 We reject My at 5% level in favour of alternative H
samples.
For example, we may want to make inferences about the mean weight loss
of a group of persons who have gone through an exercise programme for a
certain period of time. To do so, we record their weights before and after the
Programme. Here the data are collected from the same group of persons who
have joined the programme, once before and once after the programme (such
an experiment is also known as ‘before-afier experimer
taking the examination for the second time. He may compare the MAT scores.
before and after the training to study the effectiveness of his training.
‘When the samples are dependent a
for dependent means is used. This test utilizes the difference in values of the
‘matched pairs. Let d, = X,—¥,-Furtherlet wy and 7 be the mean and variance
of the differences d,, respectively, in the population, n be the number of
difference (d,) observations, 4 be the sample mean of the differences d's,
and s, be the standard deviation of the differences in the sample.
12
test known as paired t-test
Here a-1S4, and s3 = 1, a)? are used to develop the
can en
The objective is to test Hg: y-H2 =0 (ies Ho! = #2) against any
‘ne of the possible alternatives
Hyith> Ha)
Ayi ay Ha >
, Hai py He <0 Ges
Hsia *
=0 is equivalent to the
Itis clear that the hypoth
SMpothesis Hg: py = 0. ABILITY AND STATISTICAL INFERENCE : Theory and Practice
506 PROB.
Hence we test the hypothesis
Hy: ba =9
against any one of the possible alternatives
Hyg > 9
or Hyitg <0
or Hy: Hy #0.
Appropriate test statistic is_
=a
salle
which follows ¢-distribution with d.f. = 1-1, and we follow the standard
test rule.
‘We compute the observed value of the test statistic ¢, i.e.,
menus =f =—2_— , and compare the value with the tabulated value fa»,
ent 0M il
or fa/2,n-1,aSthe case may be, where d is the sample mean of the difference
series (X;-¥%) and sy isitss.d. d and sz are given by
Test procedures are as follows:
@ Totest Hy: 44 =0 against Hy: yy >0,
we compute
d
re
sy/Vn
and reject Hg at 100 a % level of significance if to > 1 ;- Otherwise, Ho 8
not rejected. a
Gi) Totest Ho: 4, =O against H,: uy <0, we compute
d
ilMn
andreject Hp at 100 « % level of significance if ty <1, , Otherwise, Ho iS
not rejected. : !
fo rest OF HYPOTHESIS
507
(i To test Ho: Hy =O against Ag: by #0, we compute
a
sulNn
andreject Hy at 100a.% level of significance if || >%.). Omens
eae [10] >tar2,nt- Otherwise, H,
Example 6.13 A dictitian wants to see if a new diet based on oat bran can
in mg per dl), Six individuals were
a ten-week period. The data on
cholesterol level before and after the use of the new diet are given below. Can
it be concluded from the data that the cholesterol level has been reduced at
a=0.05?
Individual 1 2 3 4 5 6
Before(X) | 210 235 208 ~=190 «172,24
After (Y) 190 170 210 «188173228
Ifthe new diet is effective then ‘before diet’ cholesterol levels should more
than the ‘after diet’ cholesterol level.
Here let d, = X,-¥,, 1=1,2,...,6 and j1, be the population mean of the
differences d;’s. (mean change in cholesterol level due to new diet for the
Population). Set Ho: Wy =0 and H,: uy >0.
If the alternative hypothesis H,, is rejected, then that will imply that the
new diet has failed to reduce the cholesterol level. The following calculations
are needed to carry out the test:
Before diet After diet
cholesterol | cholesterol | 4)=Xi-¥ a?
level (X) level (Y)
210 190 20 400
235 170 65 4225
208 210 2 4
190 188, 2 4
172 173 al 1 508 PROBABILITY AND STATISTICAL INFERENCE : Theory and Practice
Be eeiog)
6
1 ago oo)?
(4890 a
= +(4890- 1666.67)
5
BS = 644.67.
o, — 5y = 25.39.
The observed value of the /-statistic is obtained as
d 16.7
& ey amass =1611.
Next the upper-tail critical value off for 5 degrees of freedom and a = 0.05
is obtained from the /-table as 2.015, i.e., fogs, 5 = 2.015.
Since 19
