0% found this document useful (0 votes)
43 views14 pages

Statistics 550 Notes 21

1. The document discusses the Neyman-Pearson lemma, which establishes that the most powerful size α test is a likelihood ratio test. It defines the likelihood ratio statistic and proves parts (a)-(c) of the lemma. 2. It provides two examples to illustrate likelihood ratio tests. The first example tests the mean of a normal distribution. The second deals with a discrete distribution. 3. It briefly notes the connection between likelihood ratio tests and Bayesian tests in terms of minimizing the 0-1 loss for hypothesis testing.

Uploaded by

Himadri Samanta
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
43 views14 pages

Statistics 550 Notes 21

1. The document discusses the Neyman-Pearson lemma, which establishes that the most powerful size α test is a likelihood ratio test. It defines the likelihood ratio statistic and proves parts (a)-(c) of the lemma. 2. It provides two examples to illustrate likelihood ratio tests. The first example tests the mean of a normal distribution. The second deals with a discrete distribution. 3. It briefly notes the connection between likelihood ratio tests and Bayesian tests in terms of minimizing the 0-1 loss for hypothesis testing.

Uploaded by

Himadri Samanta
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
You are on page 1/ 14

Statistics 550 Notes 21

Reading: Section 4.2


I. Neyman-Pearson Lemma (Section 4.2)
Deine t!e "i#e"i!ood ratio statistic $y
1
0 1
0
( % )
( & & )
( % )
p
L
p

X
X
X
&
'!ere
( % ) p x
is t!e (ro$a$i"ity mass )nction or
(ro$a$i"ity density )nction o t!e data X.
*!e statistic L ta#es on t!e +a")e

'!en
1 1
( % ) 0& ( % ) 0 p p > X X
and $y con+ention e,)a"s 0
'!en $ot!
1 1
( % ) 0& ( % ) 0 p p X X
.
-e ca""
k

a "i#e"i!ood ratio test i


0 1
0 1
0 1
1 i ( & & )
( ) i ( & & )
0 i ( & & )
k
L k
c L k
L k



>


'

<

x
x x
x
*!eorem 4.2.1 (Neyman-Pearson Lemma): .onsider
testing
0 0 0 0
: ( ( % ) ) H P X P
+s.
1 1 1 1
: ( ( % ) ) H P X P
(a) I 0 > and
k

is a si/e

"i#e"i!ood ratio test& t!en


k

is
a most (o'er)" "e+e"

test
1
($) 0or eac! 0 1 & t!ere e1ists a most (o'er)" si/e

"i#e"i!ood ratio test.


(c) I

is a most (o'er)" "e+e"

test& t!en it m)st $e a


"e+e"

"i#e"i!ood ratio test e1ce(t (er!a(s on a set A


satisying
0 1
( ) ( ) 0 P A P A

X X
21am("e 1:
1
& &
n
X X K
iid
( &1) N
.
0 0
: H
&
1 1
: H
& '!ere
0 1
<
.
*!e "i#e"i!ood ratio statistic is
2
1
1
0 1
2
0
1
2 2
1 0 1 0
1
2 2
1 0 1 0
1
1 1
e1( ( )
2
2
( & & )
1 1
e1( ( )
2
2
1
e1( ( ) ( )
2
e1( ( ) ( )
2
n
i
i
n
i
i
n
i
i
n
i
i
x
L
x
x
n
x



' ;



' ;



' ;



' ;

X
Re3ecting t!e n)"" !y(ot!esis or "arge +a")es o
0 1
( & & ) L X
is e,)i+a"ent to re3ecting t!e n)"" !y(ot!esis or
"arge +a")es o
1
n
i
i
X

()sing t!e act t!at


0 1
<
).
-!at s!o)"d t!e c)to $e4 *!e distri$)tion o
0
1
( )
n
i
i
X

)nder t!e n)"" !y(ot!esis is


(0& ) N n
so t!e
most (o'er)" "e+e"

tests re3ects or
2
0
1
( )
(1 )
n
i
i
X
n


>

'!ere is t!e .D0 o a


standard norma".
0or t!e data considered in Notes 20&
1 5
( & & ) X X K
(1.1054&
1.1556& -0.1502& 1.0747& -0.1089)& 'it!
0
: 0 H
+s.
1
: 1 H
& t!e most (o'er)" "e+e" 0.05 test re3ects or
1
( 0) : (1 0.05) 1.54
n
i
i
X n

>

. -e !a+e
1
( 0) : 1.75 1.54
n
i
i
X n

<

so 'e acce(t (retain)


0
: 0 H
.
21am("e 2:
P(X;x)
0 1 2 7 4
0
P
0.1 0.1 0.1 0.2 0.5
1
P
0.7 0.7 0.2 0.1 0.1
0 1
( & & ) L x
7 7 2 0.5 0.2
*!e most (o'er)" "e+e" 0.2 test re3ects i and on"y i X;0
or 1.
*!ere are m)"ti("e most (o'er)" "e+e" 0.1 tests& e.g.& 1)
re3ect t!e n)"" !y(ot!esis i and on"y i X;0< 2) re3ect t!e
n)"" !y(ot!esis i and on"y i X;1< 7) "i( a coin to decide
'!et!er to re3ect t!e n)"" !y(ot!esis '!en X;0 or X;1.
Proo o Neyman-Pearson Lemma:
7
(a) -e (ro+e t!e res)"t !ere or contin)o)s random
+aria$"es X. *!e (roo or discrete random +aria$"es
o""o's $y re("acing integra"s 'it! s)ms.
Let
=
$e t!e test )nction o any ot!er "e+e"

test $esides
k

. >eca)se
=
is "e+e"

&
0
=
P
E
. -e 'ant to s!o'
t!at
=
1 1
( ) ( ) ( ) ( )
k
dP dP

x x x x
.
-e e1amine
=
1 0
( ( ) ( ))( ( ) ( ))
k
p kp d

x x x x x
and s!o'
t!at
=
1 0
( ( ) ( ))( ( ) ( )) 0
k
p kp d

x x x x x
. 0rom t!is&
'e conc")de t!at
= =
1 0
( ( ) ( )) ( ) ( ( ) ( )) ( ))
k k
p kp d

x x x x x x x
*!e "atter integra" is 0 $eca)se
=
0 0
( ) ( ) & ( ) ( )
k
p d p d

x x x x x x
?ence& 'e conc")de t!at
=
1
( ( ) ( )) ( ) 0
k
p

x x x
as
desired.
*o s!o' t!at
=
1 0
( ( ) ( ))( ( ) ( )) 0
k
p kp d

x x x x x
& "et
=
=
0 =
@ : ( ) ( )A
@ : ( ) ( )A
@ : ( ) ( )A
k
k
k
S
S
S



+

>
<

x x x
x x x
x x x
S)((ose S
+
x . *!is im("ies
( ) 0
k
> x
'!ic! im("ies
t!at
1 0
( ) ( ) p kp x x
. *!)s&
=
1 0
( ( ) ( ))( ( ) ( )) 0
k
S
p kp d
+

x x x x x
.
B"so& simi"ar"y&
4
=
1 0
( ( ) ( ))( ( ) ( )) 0
k
S
p kp d

x x x x x
and
0
=
1 0
( ( ) ( ))( ( ) ( )) 0
k
S
p kp d

x x x x x =
(since
1 0
( ) ( ) p kp x x
or
0
S x ).
*!)s&
=
1 0
( ( ) ( ))( ( ) ( )) 0
k
p kp d

x x x x x
and t!is s!o's t!at
=
1
( ( ) ( )) ( ) 0
k
p

x x x
as arg)ed a$o+e.
($) Let
0 1 0 0
( ) C ( ) ( )D 1 ( ) c P p cp F c > x x
'!ere
0
F
is t!e
cd o
1
0
( )
( )
p
p
x
x
)nder
0
P
. >y t!e (ro(erties o .D0s&
( ) c
is
nonincreasing in c and rig!t contin)o)s.
>y t!e rig!t contin)ity o
( ) c
& t!ere e1ists
0
c
s)c! t!at
0 0
( ) ( ) c c


. So deine
1
0
0
0 1
0
0 0 0
1
0
0
( )
1 i
( )
( ) ( )
( ) i
( ) ( ) ( )
( )
0 i
( )
p
c
p
c p
x c
c c p
p
c
p

>



'

<

x
x
x
x
x
x

*!en&
5
0
0 1 1
0 0 0 0
0 0 0 0
0
0 0 0
0 0
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( ) ( )
( )
( ) C ( ) ( )D
( ) ( )

P
c p p
E P c P c
p c c p
c
c c c
c c

> +

x x
x
x x
So 'e can ta#e k to $e
0
c
.
(c) Let
=
$e t!e test )nction or any most (o'er)" "e+e"

test. >y (arts (a) and ($)& a "i#e"i!ood ratio test


k

'it!
si/e

can $e o)nd t!at is most (o'er)". Since


=
and
k

are $ot! most (o'er)"& it o""o's t!at


=
1
( ( ) ( )) ( ) 0
k
p

x x x
0o""o'ing t!e (roo in (art (a)& im("ies t!at
=
1 0
( ( ) ( ))( ( ) ( )) 0
k
p kp d

x x x x x
'!ic! can $e t!e case i and on"y i
=
( ) ( ) 0
k
x x
'!en
1 0
( ) ( ) p kp < x x
(i.e.&
0 1
( & & ) L k < X
) and
=
( ) ( ) 1
k
x x
'!en
1 0
( ) ( ) p kp > x x
(i.e.&
0 1
( & & ) L k > X
) e1ce(t on a set A satisying
0 1
( ) ( ) 0 P A P A

X X
.
.onnection o "i#e"i!ood ratio tests to >ayes tests: .onsider
t!e 0-1 "oss or !y(ot!esis testing. *!e >ayes test c!ooses
5
1 1
: H
o+er
0 0
: H
i
1
0
( % )
1
( % )
p
p

>
X
X
& '!ic! is
e,)i+a"ent to
0
0 1
0
( )
( & & ) 1
1 ( )
L



>

x
. *!e "i#e"i!ood ratio
test
0 1
0 1
0 1
1 i ( & & )
( ) i ( & & )
0 i ( & & )
k
L k
c L k
L k



>


'

<

x
x x
x
is a >ayes test or t!e (rior
0
1
( )
1 k

+
& $eca)se or t!is
(rior
1
0 1
0
( % ) 1
( & & )
( % )
p
L
p k


X
x
X
is greater t!an 1 i and on"y
i
0 1
( & & ) L k > x
.
*!e dierence $et'een t!e >ayes a((roac! and t!e
Neyman-Pearson a((roac! is t!at t!e >ayes a((roac! starts
'it! a (rior and t!is determines t!e c)to k or t!e
"i#e"i!ood ratio test& '!i"e t!e Neyman-Pearson a((roac!
starts 'it! a signiicance "e+e" (a ma1im)m acce(ta$"e
*y(e I error rate) and t!is determines t!e c)to k or t!e
"i#e"i!ood ratio test.
II. Eniorm"y Fost Po'er)" *ests (Section 4.7)
-!en t!e a"ternati+e !y(ot!esis is com(osite&
1 1
: H
&
t!en t!e (o'er can $e dierent or dierent a"ternati+es.
0or eac! (artic)"ar a"ternati+e
1

& a test is t!e most


8
(o'er)" "e+e"

test or t!e a"ternati+e


1

i t!e test is most


(o'er)" or t!e sim("e a"ternati+e
1 1
: H
.
I a (artic)"ar test )nction
=
( ) x is t!e most (o'er)" "e+e"

test or a"" a"ternati+es


1

& t!en 'e say t!at
=
( ) x is
a uniformly most powerful (EFP) "e+e"

test and 'e


s!o)"d c"ear"y )se
=
( ) x as o)r test )nction )nder t!e
Neyman-Pearson (aradigm.
In notation& "et
( & )
denote t!e (o'er o a test

at t!e
a"ternati+e . B "e+e"

test
=
is EFP i
=
1
( & ) ( & ) or a""
or any ot!er "e+e"

test

.
21am("e o )niorm"y most (o'er)" test:
Let
1
& &
n
X X K
$e iid
( &1) N
and s)((ose 'e 'ant to test
0 0
: H
+ers)s
1 0
: H >
. 0or eac!
1 0
>
& t!e most
(o'er)" "e+e"

test o
0 0
: H
+ers)s
1 1
: H

re3ects t!e n)"" !y(ot!esis or
0
1
( )
(1 )
n
i
i
X
n


>

.
Since t!is same test )nction is most (o'er)" or eac!
1 0
>
& t!is test )nction is EFP.
>)t s)((ose 'e consider t!e a"ternati+e !y(ot!esis
1 0
: H
. *!en t!ere is no EFP test. *!e most
(o'er)" test or eac!
1 1
: H
& '!ere
1 0
>
& re3ects
6
t!e n)"" !y(ot!esis or
0
1
( )
(1 )
n
i
i
X
n


>

& $)t t!e


most (o'er)" test or eac!
1 1
: H
& '!ere
1 0
<
&
re3ects or
0
1
( )
( )
n
i
i
X
n


<

. Note t!at t!e test t!at


re3ects t!e n)"" !y(ot!esis or
0
1
( )
(1 )
n
i
i
X
n


>


cannot $e most (o'er)" or an a"ternati+e
1 0
<
$y (art
(c) (necessity) o t!e Neyman-Pearson Lemma since it is
not a "i#e"i!ood ratio test or
1 0
<
.
It is rare or EFP tests to e1ist. ?o'e+er& or one-sided
a"ternati+e& t!ey e1ist in some (ro$"ems.
B condition )nder '!ic! EFP tests e1ist is '!en t!e
ami"y o distri$)tions $eing considered (ossesses a
(ro(erty ca""ed monotone likelihood ratio.
Deinition: *!e one-(arameter ami"y o distri$)tions
@ ( % ) : A p x
is said to $e a monotone "i#e"i!ood ratio
ami"y in t!e one-dimensiona" statistic
( ) T x
i or
0 1
<
(a)
0 1
( % ) ( % ) p p x x
or a"" x (identiia$i"ity)
($)
1
0
( )
( )
p
p

x |
x |
is an increasing )nction o
( ) T x
.
21am("es o ami"ies 'it! monotone "i#e"i!ood ratio:
(a) 0or
1
& &
n
X X K
iid 21(onentia" ( )&
9
1
0
1
1 1 1
1 0
1
0 0
0
1
( )
e1( ( )
( )
i
i
n
X
n
n
i
i
n
X
i
i
e
p
X
p
e

_
_



,
,

x |
x |
0or
1 0
>
&
1
0
( )
( )
p
p

x |
x |
is an increasing )nction o
1
n
i
i
X

so t!e ami"y !as monotone "i#e"i!ood ratio in


1
( )
n
i
i
T X

x
.
($) .onsider t!e one-(arameter e1(onentia" ami"y mode"
( ) ( ) e1(@ ( ) ( ) ( )A p h T x | x x
I
( )
is strict"y increasing in & t!en t!e ami"y is
monotone "i#e"i!ood ratio in
( ) T x
. I
( )
is strict"y
decreasing in & t!en t!e ami"y is monotone
"i#e"i!ood ratio in
( ) T x
.
21am("e 2: Let
G >inomia"( & )& 0 1 X n < <
.
( % ) (1 )
e1(C "og "og(1 )D
1
x n x
n
p x
x
n
x n
x

_


,
_
_
+

,
,

*!is is a one-(arameter e1(onentia" ami"y 'it!
10
( ) "og & ( ) "og(1 )& ( ) & ( )
1
n
n T x x h x
x

_
_

,
,
Since
( )
is strict"y increasing in & t!e ami"y is
monotone "i#e"i!ood ratio in
( ) T x x
.
*!eorem (4.7.1 and .oro""ary): I t!e one-(arameter ami"y
o distri$)tions
@ ( % ) : A p x
!as monotone "i#e"i!ood
ratio in
( ) T x
& t!en t!ere e1ists a EFP "e+e"

test or
testing
0 0
: H
+ers)s
1 0
: H >
and it is gi+en $y
=
1 i ( )
( ) i ( )
0 i ( )
T c
T c
T c

>


'

<

x
x x
x
'!ere c and

are determined so t!at


0
=
( ) E

x
.
Proo: 0i1 an a"ternati+e
1 0
>
. >eca)se t!e ami"y is
monotone "i#e"i!ood ratio in
( ) T x
& t!e "i#e"i!ood ratio
statistic
1
0 1
0
( % )
( & & )
( % )
p
L
p

x
x
x
is an increasing )nction o
( ) T x
. *!e Neyman-Pearson
"emma gi+es t!at
=
( ) x is a most (o'er)" test or testing
0 0
: H
+ers)s
1 1
: H
. Since t!is !o"ds or a""
1 0
>
& 'e !a+e t!at
=
( ) x is EFP.
Note: Not a"" one-(arameter ami"ies !a+e monotone
"i#e"i!ood ratio. 0or e1am("e& t!e .a)c!y distri$)tion is
not monotone "i#e"i!ood ratio in x:
11
2
1 1
( % )
1 ( )
p x
x

+
.
0or 0 > &
2
2
( % ) 1
( % 0) 1 ( )
p x x
p x x

+
&
'!ic! is not increasing in x.
No' consider
0 0
: H
+ers)s
1 0
: H >
. Is t!e a$o+e
test
=
( ) x sti"" EFP4
0act: *!e test
=
( ) x is "e+e"

or
0 0
: H
(i.e.& or
e+ery
0

& t!e (ro$a$i"ity o re3ection is at most

).
>eore (ro+ing t!is act& 'e (ro+e .oro""ary 4.2.1.
.oro""ary 4.2.1: I
=
is a most (o'er)" "e+e"

test o
0 0
: H
+s.
1 1
: H
& t!en t!e (o'er o
=
at
1

is
greater t!an or e,)a" to

'it! e,)a"ity i and on"y i


0 1
( % ) ( % ) p p x x
'it! (ro$a$i"ity one ()nder $ot!
0
H

and
1
H
).
Proo: *!e test
( ) x
or a"" x (i.e.& re3ect 'it!
(ro$a$i"ity

regard"ess o x) !as "e+e"

and (o'er

.
>y t!e necessity (art o t!e Neyman-Pearson "emma& i
( ) x
'ere most (o'er)"& t!en
1
0
( % )
( % )
p
k
p


X
X
'it!
12
(ro$a$i"ity one. *!ereore&
1 0
( % ) ( % ) p kp X X
'it!
(ro$a$i"ity one '!ic! im("ies t!at k;1 and conse,)ent"y
t!at
1 0
( % ) ( % ) p p X X
'it! (ro$a$i"ity one.
Proo o act: Let
= =
( & ) C ( )D E

x
. -e 'ant to s!o'
=
( & ) or
0

. 0i1
2 0
<
. *!en
( ) x
is t!e
most (o'er)" test or testing
0 2
: H
+s.
1 0
: H
at
"e+e"
=
2
( & )
. >)t rom .oro""ary 4.2.1&
= =
0 2
( & ) ( & )
.
>ac# to t!e ,)estion o is
=
( ) x EFP "e+e"

or testing
0 0
: H
+ers)s
1 0
: H >
4
Hes& it is EFP. Proo: .onsider a "e+e"

test

or
0 0
: H
.

m)st a"so $e a "e+e"

test or
0 0
: H
. *!en& $eca)se
=
( ) x is EFP "e+e"

or testing
0 0
: H
+ers)s
1 0
: H >
&
=
0
( & ) ( & ) or a""
.
S)mmary statement o *!eorem 4.7.1 and .oro""ary:
I t!e one-(arameter ami"y o distri$)tions
@ ( % ) : A p x
!as monotone "i#e"i!ood ratio in
( ) T x
&
t!en t!ere e1ists a EFP "e+e"

test or testing
0 0
: H

+ers)s
1 0
: H >
and it is gi+en $y
17
=
1 i ( )
( ) i ( )
0 i ( )
T c
T c
T c

>


'

<

x
x x
x
'!ere c and

are determined so t!at


0
=
( ) E

x
.
21am("e 2 contin)ed: 0or
G >inomia"( & ) X n
and testing
0 0
: H
+s.
1 0
: H >
& t!e EFP "e+e"

test is
=
1 i
( ) i
0 i
x c
x x c
x c

>


'

<

'!ere
c
is t!e sma""est integer s)c! t!at
0
( ) P X c

> <
and
0 0
C ( )D : ( ) P X c P X c

>
.
14

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy