Вы находитесь на странице: 1из 23

Tuqaec Matablhtc kai Katanomc Pijanottwn

O stqoc mac enai ap na uposnolo parathrsewn (degma) na


bgloume sumpersmata gia lo ton plhjusm. PARADEIGMA:
Endiafermaste na mjoume gia tic dapnec diatrofc ap ta
Ellhnik noikokuri, , gia to mso eisdhma ap ergasa sthn
Ellda.

OIKONOMETRIA I:
Epanlhyh Basikn Statistikn Ennoin
Panagithc J. Kwnstantnou

Stlioc Fountc

Gr.

506 (Prgoc D), pkonstant@uom.gr

Gr.

310 (Prgoc D), sfountas@uom.gr

H upjesh pou knoume enai ti uprqei mia gnwsth diadikasa


pou pargei ta stoiqea pou qoume sth dijes mac statistik
perama kai ti aut h diadikasa mpore na perigrafe ap mia
katanom pijanottwn.
To stoiqeo thc oikonomac pou mac endiafrei, h metablht
{apotlesma}, Y enai mia tuqaa metablht. Mqri na gnei to
perama, den gnwrzoume me bebaithta thn tim pou ja lbei h Y .

Panepistmio Makedonac, Jessalonkh, 2007

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

1 / 91

Tuqaec Metablhtc

Basikc Statis tikc 'Ennoiec

PA. MAK.

2 / 91

Isqei ti:

pou ikanopoie ton periorism




s : X (s) x X 1 ([, x]) F , x RX
dhl. mia tuqaa metablht enai mia sunrthsh pou antistoiqe
arijmoc se la ta stoiqea tou deigmatiko qrou S me ttoio
trpo ste na diathretai h dom twn endeqomnwn tou qrou
endeqomnwn F .
O qroc (S, F , P ()) onomzetai qroc pijanthtac (afhrhmnh
nnoia). H t.m. apeikonzei ton deigmatik qro S se na pedo timn
RX , to qro endoqomnwn F 7 B (RX ) se na pedo Borel
(lgebra - )kai th sunrthsh
pijanthtac

1
P7PX P X () : B (RX ) 7 [0, 1] (qrsimo gia
upodeigmatopohsh).
(OE)

(OE)

Tuqaec Metablhtc

Mia tuqaa metablht X sto qro endeqomnwn F enai mia


sunrthsh
X () : S7RX

KWNSTANTINOU, FOUNTAS

KWNSTANTINOU, FOUNTAS

Basikc Statis tikc 'Ennoiec

PA. MAK.

3 / 91

X ()

(S, F , P ()) 7 (RX , B (RX ) , PX ()) 7=


(x;
f
)
,
x

R
,

| {z }

pdf

pou f (x; ) enai mia sunrthsh pijanthtac.




To snolo = f (x; ) , x RX , apotele na updeigma
pijanthtac, dhl. mia sullog ap sunartseic pijanthtac, pou
qrhsimopoie wc dekth na snolo paramtrwn, . Emperiqei mia
sunrthsh pijanthtac gia kje sto qro paramtrwn .

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

4 / 91

(Example: our interest is to learn about the food expenditures by UK families, or the
average income in the UK).
Our assumption is that there is an unknown process that generates the data we have
statistical experiment and that this process can be described by a probability
distribution. The aspect of the economy we are interested in, our outcome variable, Y ,
is a random variable. Until the experiment is performed it is uncertain what value Y

Ajroistik Sunrthsh Pijanthtac (Monometablht)

Sunrthsh (Puknthtac) Pijanthtac

will take. Typically, we will use capital letters for the name of a random variable and
lowercase letters for the values it takes.

turning to a formal denition of random variables, let us review some terminology


H pijanthta ti h Before
tuqaa
metablht X enai mikrterh sh ap x
from continuous distribution theory:
sumbolzetai me Fx (x).Cumulative
H Fx (x)
enaiFunction
h ajroistik
Distribution
(univariate): The sunrthsh
probability that a random variable X is less than or equal to x is denoted F (x). F (x) is the Cumulative
(puknthtac) pijanthtac
(CDF)
.
Distribution Function (CDF).

F (x)

x
The probabilitythc
associated
the continuous
random
X taking
any
Hpargwgoc
CDFwith
fx (x)
x
x variable
RX , enai
h sunrthsh
particular value x is zero (Pr(X = x) = 0)
(puknthtac) pijanthtac PDF
Continuous Random
Variableorzetai
has a smooth,
Hsunrthsh
pijanthtac
katnon-decreasing
trpo ste:CDF.

f (x) 0, x R
@F (x)
f (x) dx = 1
f (x) dx =
Distribution
RX xFunction, f (x) @x x ; is the Probability Density Function (PDF): The
(a sox that
(a x < b) = Pr (a < x b) =
b)
= Pr0 and
p.d.f. isPr
dened
f (x)
Rb
b
Pr (a < xPr(a
< b) X= ab) =
fx R(u)
du == F
f (x)dx
F x(b)(b)F
(a)Fx (b)
0
1

X
Rx
R +
Probability
Density Function
(univariate): The derivative of the Cumulative
2

F (1) = Pr(X < 1) = 1


F ( 1) = 0

If x > y then F (x)

F (x)

R1
Grafik, aut
shmanei
f (x)dx = 1ti h pijanthta mia tuqaa metablht na
1
lambnei
timc metax a kai b enai h skiazmenh perioq ktw ap
Graphically this implies that the probability that the random variable takes
thnvalues
PDFbetween
.
a and b is the highlighted area under the p.d.f.

F (y)

F (x)

Rx

1
(u) du, fx (u) 0.
Rb
Pr (a x b) = a fx (u) du = Fx (b) Fx (a) =
Rb
Ra
(u)
f
f (u) du
du

Fx (x) =

f
x

0 Fx (x) 1
En z < y, tte Fx (z) Fx (y)
Fx (+) = Pr (x < +) = 1
Fx () = 0.

KWNSTANTINOU, FOUNTAS

(OE)

y = f (x)

Basikc Statis tikc 'Ennoiec

PA. MAK.

5 / 91

KWNSTANTINOU, FOUNTAS

Algebraically, we note
f (x)

Sunrthsh (Puknthtac) Pijanthtac

(OE)

Basikc Statis tikc 'Ennoiec

@F (x) denition of derivative(s)


F (x)
=
lim F (x+dx)
,
@x
dx
dx!0

PA. MAK.

6 / 91

where

Ropc Tuqawn Metablhtn


F (x + dx)

F (x) = Pr(X

x + dx)

Pr(X

x)

= Pr(X 2 [x; x + dx]) since Pr(X = x) = 0

Orismc

or f (x) = lim
0
Majhmatik Elpda
MtradxKentrikc
Tshc
dx!0
with dx ! 0
) Pr(X 2 [x; x + dx]) = f (x)
dx
}|
{
Zz}|{ z
pdf width of the interval
E [x]
xfx (x) dx =
Pr(X2[x;x+dx])

Algebrik parathrste ti: fx (x)

Fx (x)
x

= limdx0

F(x+dx)F(x)
dx

F (x + dx) F (x) = Pr (X x + dx) Pr (X x) =


Pr (X [x, x + dx])
fx (x) = limdx0

Pr(X [x,x+dx])
dx

bution theory). If X is a vector of random variables (X1 ; ::; Xn ), we need to use multi-

dx
|{z}

m
: Dimesoc
(median):
Pr (x m) 1/2 Pr (x m) 1/2,
variate
distribution
theory.
Epikratosa tim (mode) max fx (x). Gia mia sunrthsh tou x , g(x) h
majhmatik elpda orzetai wc 2
Z
E [g (x)]
g (x) fx (x) dx

pltoc tou diastmatoc

RX

Pr (X [x, x + dx]) = fx (x)


|{z}
PDF

RX random variable X (univariate distriSo far only considered the p.d.f. and c.d.f. of one

Gia pardeigma, stw g (x) = a + bx

E [g (x)] = E [a + bx] = a + bE (x) = a + b.


KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

7 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

8 / 91

Ropc Tuqawn Metablhtn

Ropc Tuqawn Metablhtn


'Enac genikteroc orismc twn ropn enai o akloujoc

Orismc
Diakmansh

Mtro Diasporc

Orismc
Z
i
h i
2
2
2
2
x Var [x] E (x ) = E x (E [x]) =
h

(x )2 fx (x) dx > 0
RX

Gia mia sunrthsh tou x , g(x) h diakmansh orzetai wc


Z
(g (x) E [g (x)])2 fx (x) dx > 0
Var [g (x)] =

Rx

Gia pardeigma, o msoc (majhmatik elpda) enai h rop 1hc txhc


per thn arq (mhdn), dhl. = 0.

RX

Gia pardeigma, en g (x) = a + bx , tte

H diakmansh enai h rop 2hc txhc per to mso, dhl. = .

Var [g (x)] = Var [a + bx] = b 2 Var [x] dhl. Var [a] = 0.

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

9 / 91

Shmantikc Anisthtec

Pr [ k x + k ] 1

1
k2

Basikc Statis tikc 'Ennoiec

PA. MAK.

10 / 91

g (x) ' g (x0 ) + g 0 (x0 ) (x x0 )


= [g (x0 ) g 0 (x0 ) x0 ] + g 0 (x0 )x
|
{z
} | {z }

E [g (x)]
Pr [g (x) ] 1
.
g ()

H anisthta Jensen . 'Estw () : R 7 R, mia kurt sunrthsh,


dhl. (x1 ) + (1 ) (x2 ) (x1 + (1 ) x2 ) x1 , x2 R,
(0, 1). Upojtontac ti E (|x|) < +, isqei
[E (x)] E[ (x)].
Antstoiqa, en h () : R 7 R, enai kolh sunrthsh, tte
(E[x]) E[g(x)].
(OE)

(OE)

'Estw ti jloume na upologsoume thc diakmansh Var [g (x)] miac


mh-grammikc sunrthshc g() tou x , en dh gnwrzoume thn
diakmansh tou x , Var[x].
Arqik mporome na proseggsoume th sunrthsh grw ap mia tim
x0 me th qrsh tou jewrmatoc Taylor . 'Eqoume:

kai

KWNSTANTINOU, FOUNTAS

KWNSTANTINOU, FOUNTAS

Prosggish Diakmanshc Genikc Sunrthshc

H anisthta Chebyshev :

E [g (x)]
Pr [g (x) ]

g ()

H rop (tou plhjusmo) r txhc per enc shmeou, , orzetai wc:


h
i Z
r
(x )r fx (x) dx
E (x ) =

Basikc Statis tikc 'Ennoiec

PA. MAK.

11 / 91

= 0 + 1 x
En jsoume x0 = = E [x], tte

g (x) ' [g () g 0 () ] + g 0 () x
kai
KWNSTANTINOU, FOUNTAS

Var [g (x)] ' [g 0 ()]2 Var [x]


(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

12 / 91

Paradegmata

Polumetablhtc Katanomc
'Ewc ed qoume parousisei thn PDF kai CDF miac tuqaac
metablhtc x (jewra monometablhtn katanomn). En x enai na
dinusma tuqawn metablhtn [x1 , x2 , , xn ]0 , qreiazmaste na
qrhsimopoisoume thn jewra polumetablhtn katanomn.
Ap koino Ajroistik Sunrthsh Pijanthtac (Joint Distribution
Function)

Pardeigma
(
fx =

1/x 2 x 1
= Fx (x) =
0
x<1



1 1/x 2 x 1
0
x<1

Pr (X1 < x1 , X2 < x2 , ..., Xn < xn ) = F (x1 , ..., xn ) = F (xx ) .

Pardeigma
Ekjetik katanom fx (x) =

e x ,

> 0, x R+

Fx (x) = 1 e
1
E [x] =

 2
1
Var [x] =

Ap koino Sunrthsh (Puknthtac) Pijanthtac (Joint Density


Function)

, > 0, x R+

f (xx ) = f (x1 , x2 , ..., xn ) =


1

fR (xx )R 0

Rxn
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

13 / 91

Perijwriak / Up-Sunjkh Sunrthsh (Puknthtac)


Pijanthtac

Rxn

f (x, y)
f (x, y)
kai f ( y| x) =
.
fy (y)
fx (x)

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

(OE)

f (xx ) dx1 dx2 dxn = 1


Basikc Statis tikc 'Ennoiec

PA. MAK.

14 / 91

'Ena perama sunstatai sthn epanlhyh anexrthtwn, tautshma


katanemhmnwn (IID) lyewn pollc forc.
Anexarthsa
Anexarthsa: Do tuqaec metablhtc X kai Y enai statistik
anexrthtec en kai mno en h ap-koino sunrthsh pijanthtc
touc isotai me to ginmeno twn perijwriakn katanomn touc, dhl.

f (x1 , ..., xn ; ) = f1 (x1 ; )f2 (x2 ; )...fn (xn ; ) =

n
Y

fi (xi ; ).

t=1

Tautshma Katanemhmnec
Katanemhmnec: Do tuqaec metablhtc X kai Y enai
tautshma katanemhmnec en qoun thn dia sunrthsh
(puknthtac) pijanthtac PDF.

f ( y| x) y x.
f (x, y) = f ( x| y) fy (y) = f ( y| x) fx (x)
.. X Y , f ( y| x) = fy (y)
f ( x| y) = fx (x) .
KWNSTANTINOU, FOUNTAS

KWNSTANTINOU, FOUNTAS

Rx1

f (x, y) = fx (x) fy (y) x y .


X1 , X2 , ..., Xn

'Estw mia dimetablht katanom (do t.m., l.q. X kai Y ). H


up-sunjkh sunartseic pijanthtac enai

f ( x| y) =

Rxn 1

Anexrthtec Tautshma Katanemhmnec Tuqaec


Metablhtc IID

H perijwriak sunrthsh pijanthtac thc tuqaac metablhtc X1


ap thn ap koino PDF f (x1 , ..., xn ; ) orzetai wc:
Z
Z
f1 (x1 ; ) =
...
f (x1 , ..., xn ; )dx2 ...dxn
Rx2

n F (x1 , x2 , ..., xn )
x1 x2 ...xn

15 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

16 / 91

Ropc Tuqawn Dianusmtwn

Ropc Tuqawn Dianusmtwn


H mtra diakmanshc - sundiakmanshc enai h polumetablht rop
2hc txhc per ton mso.
Gia tic tuqaec metablhtc x1 , ..., xn enai mia mtra n n me ta
stoiqea thc na dnontai ap:
h

h ii
Cov(xi , xj ) = ij = E (xi E [xi ]) xj E xj , i, j = 1, 2, ..., n.

H majhmatik elpda enc tuqaou diansmatoc (diansmatoc


tuqawn metablhtn) x diastsewc n (dhl. n tuqawn metablhtn)
enai:

E [xx ] E

KWNSTANTINOU, FOUNTAS

(OE)


E [x1 ]

E [x2 ]

..


.

xn
E [xn ]
x1
x2
..
.

1
2
..
.

Basikc Statis tikc 'Ennoiec

ij xi xj .
ij > 0 (< 0) xi xj
() .

=
x

PA. MAK.

H kria diagnioc thc mtrac diakmanshc - sundiakmanshc


periqei tic diakumnseic ii = 2i .
Ta stoiqea ektc thc kurac diagwnou periqoun tic
sundiakumnseic ij .
H mtra diakmanshc - sundiakmanshc enai summetrik: ij = ji
H mtra diakmanshc - sundiakmanshc enai jetikc hmiorismnh.
17 / 91

Ropc Tuqawn Dianusmtwn

(OE)

Basikc Statis tikc 'Ennoiec

Basikc Statis tikc 'Ennoiec

PA. MAK.

18 / 91

(sunqeia mtrac diakmanshc)

h
i
Var [xx ] E (xx x ) (xx x )0 = E [xx x 0 ] x 0x

x1 1

x h

2
2

=E

n
n
.
1
1
2
2

xn n

KWNSTANTINOU, FOUNTAS

(OE)

Ropc Tuqawn Dianusmtwn

Gia na tuqao dinusma x epomnwc qoume:

(x1 1 )2
(x1 1 ) (x2 2 )

(x2 2 ) (x1 1 )
(x2 2 )2

= E
..
..

.
.

(xn n ) (x1 1 ) (xn n ) (x2 2 )

KWNSTANTINOU, FOUNTAS

(x1 1 ) (xn n )
(x2 2 ) (xn n )
..
..
.
.

(xn n )2

PA. MAK.

19 / 91

Cov [x1 , x2 ] Cov [x1 , xn ]


Var [x1 ]

Cov [x1 , x2 ]
Var [x2 ]
Cov [x2 , xn ]
=
..
..
..
..

.
.
.
.

Cov [x1 , xn ] Cov [x2 , xn ]


Var [xn ]
2

1 12 1n

12 22 2n

= .
..
.. x
..
..
.
.
.

1n 2n 2n

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

20 / 91

Ropc Tuqawn Dianusmtwn

Ropc Tuqawn Dianusmtwn

H mtra susqetsewn sundetai sten me th mtra diakmanshc sundiakmanshc:


h
i
h
i
h
i
Cov xi , xj
Cov xi , xj
h i
Corr xi , xj = ij = q
h i=
[x
]
SD
SD
xj
i
Var [xi ] Var xj
ij
ij
=
[1, 1]
=
ii jj
i j

1 12 13 1n

12 1 23 2n

3n
Corr [xx ] 13 23 1
..
..
..
..
.
..
.
.
.
.

1n 2n 3n 1

H kria diagnioc thc mtrac susqetsewn periqei 1. Mia t.m. enai


pnta tleia susqetizmenh me ton eaut thc.
H mtra susqetsewn enai summetrik.
H susqtish enai hanexrthth twn mondwn
mtrhshc
i
i twn
h
metablhtn: Corr ai + bi xi , aj + bj xj = Corr xi , xj .

R
x

H susqtish enai na mtro thc grammikc exrthshc/sqshc


metax do t.m.
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

21 / 91

Anexarthsa kai Mh-Susqtish

, Cov [X , Y ] = E [XY ] x y = 0.

'OMWS
Mhdenik sundiakmansh ; Anexarthsa, (exaresh apotelon t.m.
pou katanmontai kanonik.)

Basikc Statis tikc 'Ennoiec

Basikc Statis tikc 'Ennoiec

PA. MAK.

22 / 91

Jerhma

h

i
Cov [X , Y ] = E [(X E [X ]) (Y E (Y ))] = E (X x ) Y y =
E [XY ] x y
E [XY ] =
Z Z
Z Z
Anexarthsa
xyf (x, y) dxdy
=
xyfx (x) fy (y) dxdy
Z
Z
=
xfx (x) dx
yfy (y) dy = x y = E [X ] E [Y ]

(OE)

(OE)

Katanomc Grammikn Sunduasmn Tuqawn Metablhtn

Anexarthsa Mhdenik Sundiakmansh

KWNSTANTINOU, FOUNTAS

KWNSTANTINOU, FOUNTAS

PA. MAK.

23 / 91

'Estw na tuqao dinusma x diastsewc n to opoo katanmetai wc


x , x ), kai stw na n 1 mh-stoqastik dinusma . Tte:
x (

0 x , 0 x )
0x (

Apdeixh
(1) Gia th majhmatik elpda qoume

x ] = E (1 , 2 , , n )
E [

x1
x2
..
.
xn

= E [ x + x + + x ]
n n

1 1
2 2

= 1 E [x1 ] + 2 E [x2 ] + + n E [xn ] = 1 1 + 2 2 + + n n


= 0 x .
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

24 / 91

Katanomc Grammikn Sunduasmn Tuqawn Metablhtn

Katanomc Grammikn Sunduasmn Tuqawn Metablhtn


Pardeigma
'Estw na tuqao dinusma x = (x1 , x2 )0 me mso x = (1 , 2 )0 kai mtra
diakmanshc sundiakmanshc
" 2
#
1 12
x =
.
12 22

Jerhma
Apdeixh
(2) Gia th diakmansh qoume:
h

i
0x ] = E 0 x E [
0x ] 0 x E [
0x ] 0
Var [
h
i
x
x
= E 0 (x
xE [xx ]) (x
xE [xx ])0
h
i
= 0 E (xx ) (xx )0

'Estw epshc na mh-tuqao dinusma = (a, b)0 . Orzoume wc


y = ax1 + bx2 = 0x . Qrhsimopointac to parapnw jerhma qoume ti:

0x ] = aE [x1 ] + bE [x2 ]
E [y] = E [
"
#
!

 2 12
a
1
Var [y] = a b
b
12 22

= 0 x

= a 2 21 + b 2 22 + 2ab12 = a 2 Var [x1 ] + b 2 Var [x2 ] + 2abCov [x1 , x2 ] .


KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

25 / 91

Katanomc Grammikn Sunduasmn Tuqawn Metablhtn

Jerhma

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

26 / 91

Up-Sunjkh Ropc

Orismc

'Estw na tuqao dinusma x diastsewc n pou katanmetai wc


x , x ), kai stw mia m n mh-stoqastik mtra A , h opoa
x (
upojtoume ti enai plrouc bajmo. Tte

H up sunjkh majhmatik elpda thc y orzetai wc:


Z
E [ y| x] =
yf ( y| x) dy

A
x, A xA 0) .
Ax (A

Ry

h opoa enai epshc mia palindrmhsh thc y pnw sth x kai enai
sunrthsh tou x. Parathrste ti kje t.m. mpore na grafe wc:

'Eqoume ti:

Ax
x
E [Ax
Ax] = A E [xx ] = A
0
Ax
Var [Ax
Ax] = A Var [xx ] A = A x A 0

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

y = E [ y| x] + (y E [ y| x]) = E [ y| x] + u

PA. MAK.

27 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

28 / 91

Up-Sunjkh Ropc

Sqseic Up-Sunjkh kai Perijwriakn Ropn


1. :

E [y] = Ex (E [ y| x])

Orismc

Ex x.
.
2. :

H up-sunjkh diakmansh thc y enai

o Z
Var [ y| x] = E (y E [ y| x]) x =
n

(y E [ y| x])2 f ( y| x) dy
Ry

H up-sunjkh diakmansh kaletai skedastik sunrthsh kai enai


sunrthsh tou x pwc kai h palindrmhsh thc y sthn x. 'Otan h
Var [ y| x] den metablletai me th x , lme ti h up-sunjkh diakmansh
enai omoskedastik


E [ y| x] x. E [ y| x] = 0 + 1 x. :

E [y] = Ex (E [ y| x]) = Ex (0 + 1 x) = 0 + 1 E [x]


0 = E [y] 1 E [x] .
Cov [y, x] = Cov [E [ y| x] , x] = Cov [0 + 1 x, x] = 1 Var [x]
1 =

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

(x E [x]) E [ y| x] fx (x) dx.

Cov [x, y] = Cov [x, E [ y| x]] =


Rx

i
= E y 2 x (E [ y| x])2
h

PA. MAK.

29 / 91

Diaqwrismc thc Diakmanshc

KWNSTANTINOU, FOUNTAS

(OE)

Cov [y, x]
Var [x]

Basikc Statis tikc 'Ennoiec

PA. MAK.

30 / 91

Diaqwrismc thc Diakmanshc

Gia th diakmansh thc y qoume:

Var [y] = Varx [E [ y| x]] + Ex [Var [ y| x]]


pou Varx enai h diakmansh gia lec tic timc thc x . H parapnw
sqsh deqnei ti se dimetablhtc katanomc pijanottwn, h diakmansh
thc y mpore na analuje se:
a. Th diakmansh thc up-sunjkh majhmatikc elpdac.
b. (+) Thn anamenmenh diakmansh grw ap ton up-sunjkh mso.

Sunjwc, endiafermaste gia to poio tmma thc Var [y] enai megaltero.
Gia aut to skop mporome na qrhsimopoiome:
Suntelestc Prosdiorismo =

Varx [E [ y| x]]
.
Var [y]

Epomnwc h metablhtthta thc y mpore na apodoje se do phgc:


1

Th metablhtthta exaitac tou gegontoc ti h E [ y| x] metablletai


me to x , ra h diakmansh palindrmhshc = Varx [E [ y| x]]
Th metablhtthta pou prokptei epeid se kje up-sunjkh
katanom, h y metablletai grw ap ton up-sunjkh mso, ra
diakmansh sflmatoc = Ex [Var [ y| x]]

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

31 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

32 / 91

Tuqaa Degmata - Statistikc

Tuqaa Degmata - Statistikc


Orismc
'Estw (x1 , ..., xT ) mia sunrthsh, thc opoac to pedo orismo
perilambnei to deigmatik qro twn x1 , ..., xT . H sunrthsh aut
kaletai statistik
statistik. H tuqaa metablht y = (x1 , ..., xT ) akolouje mia
katanom, h opoa onomzetai katanom deigmatolhyac
deigmatolhyac.

Orismc
Oi t.m. x1 , x2 , ..., xT kalontai tuqao degma megjouc T ap ton
plhjusm f (x; ), en enai anexrthtec t.m. kai h oriak sunrthsh
puknthtac - pijanthtac enai h f (x; ), dhl. xt IID f (x; ).

f (x1 , x2 , ..., xT ; ) =

T
Y

PROSOQH
PROSOQH: H (x1 , ..., xT ) enai sunrthsh mno twn timn tou
degmatoc, kai den exarttai ap tic timc twn paramtrwn .
Suqn qrhsimopoiomenec statistikc:

f (xt ; )

t=1

Diaforetikc timc twn paramtrwn odhgon se diaforetikc idithtec


tou tuqaou degmatoc.

x =

T
1X
xt : deigmatikc msoc
T
t=1

1 X
(xt x )2 : deigmatik diakmansh
T 1
t=1

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

33 / 91

Tuqaa Degmata - Statistikc

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

34 / 91

Deigmatolhptikc Katanom Deigm. Msou - Diakmanshc

Ta aklouja jewrmata enai qrsima gia th sunqeia.


En ta x1 , ..., xT enai tuqao degma ap plhjusm me mso kai
diakmansh 2 , tte o x enai t.m. me mso kai diakmansh 2 /T .
Epiplon o msoc thc s 2 enai 2 .


'Estw x1 , ..., xT IID , 2

Jerhma
P
'Estw x1 , ..., xT kpoioi arijmo kai x = T1 Tt=1 xt . Tte isqoun oi
akloujec do sunjkec:
P
P
min Tt=1 (xt )2 = Tt=1 (xt x )2 , dhl. = x
P
P
(T 1) s 2 = Tt=1 (xt x )2 = Tt=1 xt2 T x 2
1

h P
i
T
1
1 PT
1 PT
t=1 E (xt ) = T
t=1 =
Th t=1 xt = iT
hP
i
2
2
T
1
Var [x ] = E (x E [x ]) = T 2 E
t=1 (xt )
nP
o
h
i P
2
T
T
= T12
t,s E [(xt ) (xs )]
t=1 E (xt ) +
P
P
= T12 Tt=1 Var [xt ] +
0
= T12 Tt=1 2 = 2 /T
|{z}
E [x ] = E

Jerhma
'Estw x1 , ..., xT na tuqao degma ap kpoion plhjusm kai stw g (x)
mia sunrthsh ttoia ste na uprqoun h E [g (x1 )] kai h Var [g (x1 )].
Tte:
hP
i
E Tt=1 g (xt ) = TE [g (x1 )]
hP
i
Var Tt=1 g (xt ) = T Var [g (x1 )]
1

anexarthsa
En ta xt prorqontai ap kanonik plhjusm, o x katanmetai
epshc kanonik wc grammik sunrthsh kanonikn metablhtn.

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

35 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

36 / 91

Mjodoi Ektmhshc - Mjodoc Elaqstwn Tetragnwn

Mjodoi Ektmhshc - Mjodoc Elaqstwn Tetragnwn

'Estw ti na tuqao degma y1 , ..., yT ap plhjusm me mso kai


diakmansh 2 .

To statistik updeigma pou ja qrhsimopoisoume enai:

Kje paratrhsh tou degmatoc mpore na grafe wc yt = + ut ,


dhl. to ut = yt mac deqnei ti oi parathrseic diafroun ap
ton mso.
Jloume na ektimsoume thn parmetro , h opoa den parathretai,
me th qrsh tou tuqaou degmatoc pou qoume. Isqei ti:

E [ut ] = E [yt ] = E [yt ] = = 0.


Epshc ti:

yt = + ut , t = 1, ..., T
E [ut ] = 0.

Var [ut ] = 2
Cov [ut , us ] = 0 gia t , s.


dhl. ut IID 0, 2 . H teleutaa idithta Cov [ut , us ] = 0 prokptei ap
thn upjesh ti ta yt enai anexrthta (tuqao degma) kai epomnwc
asusqtista.
3

Parathrste epshc ti gia kje ektmhsh tou lambnoume mia


t .
ektmhsh tou {sflmatoc} u

Var [ut ] = Var [yt ] = Var [yt ] = 2 .


10

BASIC STATISTICS

The least squares estimator of is obtained by minimizing the sum of squares errors, SSE, defined by
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

SSE =

n
X

e2i =

i=1

n
X

(yi )

PA. MAK.

37 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

38 / 91

(45)

i=1

The idea is to pick the value of to estimate which minimizes SSE. Pictorially we select the
value of which minimizes the sum of squares of the vertical deviations in figure 1.

H Mjodoc Elaqstwn Tetragnwn

H Mjodoc Elaqstwn Tetragnwn

F IGURE 1. Least Squares Estimation

Jloume na elaqistopoisoume th sunrthsh

min SSE =

H arq twn elaqstwn tetragnwn baszetai sthn elaqistopohsh


thc apstashc twn problepmenwn timn, ed yt = ap tic
pragmatikc timc tou degmatoc yt . Wc mtro apstashc epilgoume
to tetrgwno thc Eukledeiac apstashc. To apl jroisma twn
sfalmtwn ja enai mhdenik.
Parathrste ti gia kje ektmhsh lambnoume kai mia ektmhsh
t . Stqoc mac enai h elaqistopohsh tou
tou sflmatoc u
ajrosmatoc twn tetragnwn twn ektimhmnwn sfalmtwn.

T
X
t=1

t2 =
u

T 
X

2
yt .

t=1

Lambnontac tic sunjkec prthc txhc qoume:

The solution is obtained by finding the value of that minimizes equation 45.
n
X
SSE

(yi )(1)
= 0
=2

i=1

n
1 X
=
yi = y
n i1

(46)

This method chooses values of the parameters of the underlying distribution, , such that the
distance between the elements of the random sample and predicted values are minimized.

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

39 / 91

T 
T
X

SSE
1X
= 2
yt = 0 =
yt = y .
T

t=1

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

t=1

PA. MAK.

40 / 91

Idithtec Deigmatolhyac se Peperasmna Degmata

Idithtec se Peperasmna Degmata (Amerolhya)

: .

h i
, . E = .
 
h i
bias E .
BASIC STATISTICS

'Estw x1 , ..., xT na tuqao degma ap plhjusm me mso .

17

 

4.2.1. Unbiasedness.
is said to be an
estimator
of if E = .


unbiased


In figure 2, is an unbiased estimator of , while is a biased estimator.

0 , .
F IGURE 2. Unbiased Estimator

fHL

Pardeigma (Amerlhpth) O x enai amerlhpth ektimtria tou :


E[x ] =
Pardeigma (Merolhptik): Jewrome to x 2 wc ektimtria tou 2
 P
h i
2 
h i
PP
P
E x 2 = E T1 Tt=1 xt
= T 2 Tt=1 E xt2 + T 2
E [xt xs ]
t,s

fHL

Qrhsimopointac
h i
2 = Var [xt ] = E xt2 2

fHL

E [xt xs ] = E [xt ] E [xs ] = 2 lgw anexarthsac

h i


E x 2 = T 1 2 + 2 + T 2 T (T 1) 2 = 2 +

2
T

, 2

(OE)
'Ennoiec
4.2.2. Minimum
variance.Basikc
is said to beStatis
a minimumtikc
variance estimator
of if

KWNSTANTINOU, FOUNTAS

 
 
V ar V ar

PA. MAK.

41 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

42 / 91

(70)

where is any other estimator of . This criterion has its disadvantages as can be seen by noting
that = constant has zero variance and yet completely ignores any sample information that we may

have. In figure 3, has a lower variance than .

Idithtec se Mikr Degmata (Apotelesmatikthta)

Idithtec se Mikr Degmata (Apotelesmatikthta)

4.2.3. Mean squared error efficient. is said to be a MSE efficient estimator of if


 

 

M SE M SE
(71)
H amerolhya enai mia epijumht
idithta all spnia
where is any other estimator of . This criterion takes into account both the variance and bias
of the estimator
under consideration.
4 shows three alternative
of .
qrhsimopoietai
mnh
thc wcFigure
kritrio
miacestimators
ektimtriac.
O lgoc enai
4.2.4. Best linear unbiased estimators.
is the best linear unbiased
(BLUE) of if
ti pollc amerlhptec
ektimtriec
denestimator
qrhsimopoion
apotelesmatik ta stoiqea tou degmatoc.
Metax do amerlhptwn ektimhtrin kai aut me th mikrterh
diakmansh
enai apotelesmatikterh.
enai
h i Sto hdigramma,
i
18
BASIC STATISTICS

apotelesmatikterh , diti Var < Var .

F IGURE 3. Estimators with the Same Mean but Different Variances

fHL

fHL

fHL

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

43 / 91

Axzei na shmeiwje pwc endqetai na uprqoun merolhptikc


ektimtriec pou qoun mikrterec diakumnseic ap amerlhptec
ektimtriec. 'Ena kritrio pou anagnwrzei thc {sqsh antallagc}
(trade-off) metax merolhyac kai diakmanshc, enai to
Mso Tetragwnik Sflma (Mean Squared Error).
Mso Tetragwnik Sflma
Sflma:

2 

MSE = E
n
h i  h i
o2 
= E E + E

 h i
h i2 
2 
h
h i  h i
i
= E E
+ E E + 2E E E

 h i
h i2 
2 
[]]=E []E []
E [E
=
E E
+ E E
h i
 2
= Var + bias
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

44 / 91

Idithtec se Mikr Degmata (Apotelesmatikthta)

Idithtec se Mikr Degmata (Apotelesmatikthta)


BASIC STATISTICS

19

F IGURE 4. Three Alternative Estimators

fHL

Apotelesmatikthta: 'Estw do ektimtriec 1 kai 2 thc


gnwsthc paramtrou . Tte:
H 1 enai apotelesmatikterh kat mso tetrgwno ap thn 2 en
kai mno en MSE(1 ) < MSE(2 )

fHL

En kai oi do ektimtriec enai amerlhptec, tte h 1 enai


apotelesmatikterh ap thn 1 en kai mno en Var[1 ] < Var[2 ]

fHL

fHL

large sampling properties (will return in more detail to this later)

 unbiased

In most cases, whether an estimator is exactly
or what its exact sampling variance
n 1

2 =

S2

, an(For
Gia pardeigma
h ektimtria

enai
merolhptik
ektimtria
thcbe able
is in samples
of a given size
will be unknown
the
these properties
we need to



n 1
n

n
paramtrou
,the
enai
apotelesmatikterh
ap
ticbeimpossible
kai . to calculate or
to compute
moments of
estimators.
These
moments
may


E
2

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

45 / 91

E S2

(74)

n 1
=
2
n
FOUNTAS (OE)
Basikc Statis tikc 'Ennoiec
Also from theorem 3 and equation 20, we have that

may not even exist, i.e., they may be innite.) In this case, the idea is to work with results

KWNSTANTINOU,

PA. MAK.

46 / 91

that come about as n ! 1 and treat the n !


1
results as approximations. Usually


V ar X

(75)

As n ! 1,se
we Megla
are typicallyDegmata
able to learn the true moments of the distribution that
Idithtec

Idithtec se Megla Degmata

and S2 where X1, X2, . . . Xn are a


Now consider the mean square error of the two estimators X
random sample from a normal population with a mean of and a variance of 2 .

is generating our observations

Stic perissterec periptseic en mia ektimtria enai akribc


amerlhpth poia enai akribc h deigmatolhptik thc diakmansh
se degma dedomnou megjouc enai gnwsto (qreizetai na
upologsoume tic ropc tic katanomc/ apaitetai parxh twn
ropn).
Se aut thn perptwsh, h ida enai na qrhsimopoisoume
apotelsmata ta opoa prokptoun kajc T kai na
jewrsoume ta apotelsmata gia T wc proseggseic.
Sunjwc:

Sunpeia: Mia ektimtria thc paramtrou enai sunepc, en


kajc to mgejoc tou degmatoc auxnei, h {proseggzei} thn .

As n ! 1, moments that were innite may become nite

As n !:
1, certain things
start
looking
normally
distributed,
even if they



were not
so when
we started
from n small. ,

Denition:
Consistency.
An estimator b of is consistent if, when the sample size
closer
(

:
increases, b gets
to h: (Example:
if our
random
on UK
food
i )
h
isample
  expenditures

= the

limT
Var to
be
= 0arbitrarily
= p limclose
=to,
we
limwould
T Eexpect
is innitely large,
sample
average
the mean
P

UK food expenditure)
.

T ,
.
T ,
.
T ,
(. ),
T .
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

47 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

48 / 91

Idithtec se Megla Degmata (Sunpeia)

Deigmatolhptikc Katanomc - Kanonik Katanom




Tuqaa metablht: x N , 2

O deigmatikc msoc x enai sunepc ektimtria tou msou tou


plhjusmo.

f (x) =

E [x ] =
Var [x ] = 2 /T limT Var [x ] = 2 /T = 0.



1
exp 2 (x )2
2
22

2 2
, ,

Jerhma
Jerhma Slutsky . Gia mia suneq sunrthsh g(xT ) pou den enai
sunrthsh tou T , isqei ti: plim g (xT ) = g (plim xT ).

f ( x)

H ektimtria x 2 enai sunepc ektimtria tou 2 .


Slutsky, x
.
, x 2 2 .
.
-3

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

49 / 91

KWNSTANTINOU, FOUNTAS

(OE)

0-2 -

+2

+3

Basikc Statis tikc 'Ennoiec

PA. MAK.

Here is the distribution of b2 for the general case. Again, for the time being we are
assuming that we know its standard deviation (sd).

50 / 91

Deigmatolhptikc Katanomc - Kanonik Katanom

Deigmatolhptikc Katanomc - Kanonik Katanom

x , x ), pou x, x diansmata T 1,
Tuqao dinusma: En x N (
kai x h T T mtra diakmanshc - sundiakmanshc:
T2

f (xx ) = (2)

T2

= (2)
pou Rij =

ij
i j , ui


1
x | exp (xx x )0 1
x
(x
)

|
x
x
2


1
1
1
R x | 2 exp u 0x R 1
u
|R
x
x
(1 2 n )
2
12

xi i
i .

En x akolouje polumetablht kanonik katanom,


tte

 kje xt
2
x , x ) xt N t , t t = 1, ..., T .
enai epshc kanonik : x N (

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

51 / 91



En x N 0,
0,2I T tte
f (xx ) =

22

T /2

T
1 X

xt2
exp 2
2
t=1



= f (x1 ) ...f (xT ) pou xt IID N 0, 2
Kje grammikc metasqhmatismc tou x enai epshc na kanonik
dinusma:
x , x )
x N (
y = a + Bx
Bx, a B
,
a + B x , B x B 0 )
y N (a

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

52 / 91

Deigmatolhptikc Katanomc - Kanonik Katanom

Kanonik Katanom - Palindrmhsh

En oi x kai y enai ap-koino kanonikc kai enai asusqtistec,


tte oi x kai y enai anexrthtec.
H Pr (X x) den qei apl morf.


x
x N , 2 z = N (0, 1)

 x 

x
Pr (X x) = Pr Z =

Jerhma
En

w=
tte

 x  SOME SPECIFIC PROBABILITY DISTRIBUTIONS


3
H enai h CDF thc tupopoihmnhc kanonikc katanomc, gia
tic opoac tic timc uprqoun pnakec. Isqei ti (z) = 1 (z).

Figure 3. Normal pdf and cdf

Probability Density Function

Cumulative Distribution Function

"

x
y

"
N

P


W s N( ; )
x , xx ) , y N y , yy
Perijwriakc Katanomc: x N (
V = a + BW 
linear transformation
y
x

a constant matrix
Palindrmhsh thc y sthna xis: a constant
| N vector
y.x and
yy.xB ,ispou
then V s N (a + B ; B

0.35

F(X)

f(X)

0.2

0.1

B0)

Y are independent.

0.6



y yx 1

+ yx 1
x
xx
xx x
| {z
x) does |
not have{z
a simple}formula,
so }one must use tables to calcu
=

(d) Pr(X

0.4

0.15

1 and Y are uncorrelated, then X


(c) IfEX[y
yand
x x )
(jointly)
y.x = normal
y +and
| x ]Y are
yx ifxxX(x

0.8
0.3
0.25

# "
#!
xx xy
,
yx yy

(b) All linear transformations of W are normal

0.45
0.4

x
y

it.

0.2

0.05

0
10

KWNSTANTINOU, FOUNTAS

0
X
(OE)

0
10

10

0
X

Basikc Statis tikc 'Ennoiec

X=
sN
Z=
Var [yy | x ] Ify.x
(yy ;) yxthen
1
xx xy

10
PA. MAK.

53 / 91

KWNSTANTINOU, FOUNTAS

1.4. Evaluating probability statements with a normal random variable. If x N(,2 )


then,

Katanom 2

Z =

X


E (Z) = E

V ar (Z) = V ar

N (0, 1)

X
= 1 (E(X) ) =


X
= 12 V ar(X )

Therefore Pr(X

s N (0; 1)
x

x) = Pr(Z

Basikc Statis tikc 'Ennoiec

) = (x

PA. MAK.

54 / 91

is the CDF of the standard normal which is tabulated (see table D.1).

KatanomImportant
2
distributions derived from the normal distribution are:

0
(3)

En zt , t = 1, ..., p enai= anexrthtec


t.m. kai zt N(0, 1), tte

= 1
Pp

w = t=1 zt2 2p
2

(2) Chi-squared Distribution

(PDF) 2p ,
, p, .
.. 2p .
w 2p , E [w] = p, Var [w] = 2p p 2,
The

p diagram
2. shows the pdf and cdf for the chi-square distribution with parameters
following
=10.

, PDF CDF 210 .


6

(OE)

SOME SPECIFIC PROBABILITY DISTRIBUTIONS

En y1 , y2 , ..., yT enai tuqaec metablhtc ap nan plhjusm pou


If Z , i = 1; ::p; are independent N (0; 1) random variables,
katanmetai ikanonik me mso kai diakmansh 2 , tte

(T
The

2
p

=
1)sy2then W
2
T 1
2

p
X

Zi2

2
p

i=1

density function has a single parameter, p, the number of degrees of f

Katanomc
2p Ame
dom.

2
ligterouc
bajmoc eleujerac...
p random variable only takes positive values and is skewed to the ri

Figure 5. Chi-square pdf and cdf


Cumulative Distribution Function

0.1

0.08

0.8

0.06

0.6

F(X)

f(X)

Probability Density Function

0.04

0.4

0.02

0.2

10

20

30

df(

10

2.2. Properties(OE)
of the chi-square
random
variable.
Basikc
Statis
tikc

KWNSTANTINOU, FOUNTAS

20

)red >df(

)blue

30

'Ennoiec

If Z
PA. MAK.

55 / 91

KWNSTANTINOU, FOUNTAS

2
p
(OE)

E(Z) = p

Basikc Statis tikc 'Ennoiec

PA. MAK.

56 / 91

Figure 6. Students t distribution pdf and cdf


Probability Density Function

Cumulative Distribution Function

0.4
1

0.35
0.3

F(X)

Katanom t

f(X)

Katanom t

0.8

0.25
0.2
0.15

0.6

0.4

0.1
0.2
0.05

'Estw ti w tp , tte:
0
10

'Estw do tuqaec metablhtc z N(0, 1) kai y 2p . En oi z kai


y enai anexrthtec
anexrthtec, tte:

0
X

10

0
10

0
X

10

E[w] = 0 p > 1 ( )
Var[w]
= p/(p 2) p > 2
The following diagram shows the cdf for the Students t-distribution with parameters = 10 and
= 3.

, : w N (0, 1)
p

w= p

z
y/p

Figure 7. Students t-distribution with alternative parameter levels

tp

fHxL
0.3

v=3

PDF tp p, p
.
, ,
(.
).

0.2

0.1

-4

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

57 / 91

KWNSTANTINOU, FOUNTAS

-2

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

58 / 91

Efarmog: H katanom t Student

Katanom F
'Estw do t.m. x kai y h opoec enai anexrthtec kai katanmontai
wc 2 me m kai k bajmoc eleujerac antstoiqa. Tte

x/m
Fm,k
y/k

w=

10

Figure 9. Probability of Intervals

H1 = 12, 2 = 50L

T (
y )
N (0, 1)

2 (T 1) sy2 2T 1 ,

T (
y )
2 (T 1) sy2

0.6
H1 = 12, 2 = 10L

fHxL 0.4

H1 = 6, 2 = 30L

0.2

-1

Basikc Statis tikc


'Ennoiec
2

E(F ) =

..

4.3. moments of the F distribution.

(OE)

T (y )
tT 1
sy

Parathrste ti:

T (y )
T (y ) /
= q
sy
2 (T 1) sy2 / (T 1)

SOME SPECIFIC PROBABILITY DISTRIBUTIONS

0.8

'Estw y1 , y2 , ..., yT na tuqao degma megjouc T ap plhjusm pou


katanmetai kanonik me mso kai diakmansh 2 , tte:

x tm : y x 2 F1,m .
PDF Fm,k 2 m k (
). .
w Fm,k , E [w] = k k2 k > 2 ( )
w1 Fm,k ,

 w2 = 1/w
 1 Fk ,m .

 : 
1
1
Pr (w1 < a) = Pr 1/w1 > a = Pr w2 > a = 1 Pr w2 < a1

KWNSTANTINOU, FOUNTAS

v = 10

(25)

PA. MAK.

59 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

60 / 91

Efarmog: H Katanom F

Eidik Apotelsmata

En sx2 kai sy2 enai oi deigmatikc diakumnseic ap do anexrthta


degmata megjouc m kai n antstoiqa, ta opoa elfjhsan ap
kanonikoc plhjusmoc, pou h diakmanshc tou x ston plhjusm
enai 2x kai h diakmansh tou y ston plhjusm 2y . Tte:

2y sx2
2x sy2

2
2
x (m 1) sx / (m 1)

2
y (n

1) sy2 / (n

(OE)

1)

Basikc Statis tikc 'Ennoiec

,
A ) = n, tte:
'Estw x Nn (,
). En A mia n n mtra me r (A
w = x 0 Ax 2n

Jerhma
,
0,
'Estw x Nn (,
). Mporome na orsoume z x N (0,
). Tte:

2
2
2
x (m 1) sx m1
2
2
2
y (n 1) sy n1 ,
2 (m
2
2
x
1) sx 2
y (n 1) sy

KWNSTANTINOU, FOUNTAS

Jerhma

PA. MAK.

0, I n )
q = 1/2 (xx
) Nn (0,
kai

w = z 0 1z (xx
)0 1 (xx
) 2n

61 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

Eidik Apotelsmata

Eidik Apotelsmata

Apdeixh

Jerhma

Mporome na gryoume = 1/2 1/2 . Tte:

,
'Estw x Nn (,
). En A n n mia tautodnamh mtra me
r (A ) = rA < n (rA = # {i = 1}), tte:


0
w = z 0 1 z =
== z 0 1/2 1/2z = z 0 1/2 1/2z

0
= 1/2z 1/2z = q 0q

w = x 0 Ax 2rA

Jerhma

pou q = 1/2z .
q ] = 1/2 E [zz ] = 0,
'Eqoume E [q
kai



0
q ] = 1/2 Var [zz ] 1/2 = 1/2 1/2 = I n .
Var [q

,
'Estw x Nn (,
). En A kai B enai do n n tautodnamec mtrec
A ) = rA < n (rA = # {i = 1}) kai r (B
B ) = rB < n (rB = # {i = 1})
me r (A
0
tte, oi tetragwnikc morfc w1 = x Ax kai w2 = x 0Bx enai
anexrthtec en kai mno en AB = 0 . Se aut thn perptwsh:

0, I n ) dhl. qi N (0, 1) .
Epomnwc, q = 1/2 (xx
) Nn (0,
P
Epshc, w = q 0 q = ni=1 qi2 2n .
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

62 / 91

PA. MAK.

x 0Ax
Ax/rA
FrA ,rB
x 0Bx
Bx/rB
63 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

64 / 91

Eidik Apotelsmata

Ektmhsh kai Elegqoc Upojsewn


Mia statistik elgqou enai nac kannac pou prosdiorzei en mia
sugkekrimnh tim 0 brsketai se sumfwna me ta
apotelsmata tou degmatoc.

Jerhma
0,
'Estw x Nn (0,
). En enai mh-stoqastik n 1 dinusma kai A mia
n n tautodnamh mtra, me r (A ) = rA < n (rA = # {i = 1}), tte oi
do t.m. w1 = 0x kai w2 = x 0Ax enai anexrthtec en kai mno en
0A = 0 . Se aut thn perptwsh
w=

KWNSTANTINOU, FOUNTAS

(OE)

0x
(xx 0Ax
Ax/rA )1/2

trA

Basikc Statis tikc 'Ennoiec

PA. MAK.

65 / 91

Diastmata Empistosnhc



,

, , .

mia ektimtria thc paramtrou . 'Ena disthma


'Estw
L ,
U ] ttoio ste:
100(1 )% gia thn enai na disthma [


L
U = 1 .
Pr

(OE)

Basikc Statis tikc 'Ennoiec

mia ektimtria thc paramtrou


'Estw

,
()
.


.

.

.
,
,
.

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

66 / 91

Ektmhsh Diastmatoc Empistosnhc (Pardeigma)

Anexrthta ap tic idithtec thc ektimtriac, h ektmhsh pou


lambnoume ap na degma, ja metablletai ap degma se degma,
kai uprqei kpoia pijanthta na enai esfalmnh.

KWNSTANTINOU, FOUNTAS

(
).

PA. MAK.

y1 , y2 , ..., yT N(, 2 ), 2
.
P
, = T1 Tt=1 yt .
h

i
95% 1.96/ T , + 1.96/ T
 2
N , T = y

)
T (
N (0, 1).



)
T (
= 0.95
Pr 1.96

1.96



Pr 1.96 + 1.96
T

100(1 )% (), .
,
100(1 )%
.
..
.
67 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

68 / 91

Ektmhsh Diastmatoc Empistosnhc (Pardeigma)

Elegqoc Upojsewn
Oi Upojseic Mhdn : = 0
Oi Enallaktikc Upojseic
Monokatlhktec
Dikatlhktec

'Estw y1 , y2 , ..., yT tuqao degma T = 21 parathrsewn ap


plhjusm N(, 2 ), me gnwsth 2 . Edame ti h ektimtria
elaqstwn tetragnwn tou msou enai o deigmatikc msoc,
= 1 PT yt . 'Ena disthma empistosnhc 95% gia to enai

t=1
T
h

i
2.086sx / T , + 2.086sx / T

 2
N , T



)
T (
Pr 2.086

2.086
= 0.95
sx


Pr 2.086 sx + 2.086 sx
T

KWNSTANTINOU, FOUNTAS

(OE)

)
T (
sx

HA : > 0
HA : , 0

HA : < 0

H diadikasa elgqou enai nac kannac, se rouc twn stoiqewn,


pou upagoreei en h upjesh mhdn ja prpei na aporrifje qi.
H klassik mejodologa (Neyman-Pearson), enqei to diaqwrism
tou deigmatiko qrou se do perioqc.

tT 1

(., )

. -,
.

Basikc Statis tikc 'Ennoiec

PA. MAK.

69 / 91

Elegqoc Upojsewn - Dikatlhktoc Elegqoc

t = 2 0 tT 1

sx /T

H0 .
- / :


H0
0
0
0


2 > c/2 2 < c/2 , 2 > c/2 .
sx /T

(OE)

( )

t=

0
s

Basikc Statis tikc 'Ennoiec

tT 1

-
0 2.086s 0 + 2.086s
0
0

/2=2.5%

sx /T

0-ca/2s

(, c/2 ) (c/2 , +),


- (c/2 , c/2 ).
c/2 Pr (|t| > c/2 ) = :
Pr (|t| > c/2 ) = Pr (t > c/2 ) + Pr (t < c/2 )
2 Pr (t > c/2 ) = (1 2 Pr (t < c/2 )) .
= 0.05 (5%) T = 21, c/2 = 2.086.

KWNSTANTINOU, FOUNTAS

PA. MAK.

70 / 91

Elegqoc Upojsewn - Dikatlhktoc Elegqoc

'Estw y1 , y2 , ..., yT tuqao degma T = 21 parathrsewn ap


plhjusm N(, 2 ), me gnwsth diakmansh 2 . Edame ti h
ektimtria elaqstwn tetragnwn tou msou enai o deigmatikc
P
msoc, = T1 Tt=1 yt . Jloume na exetsoume thn H0 : = 0
nanti thc enallaktikc HA : , 0 .

sx /T

H0 : = 0

0+ca/2s

0
s

tT 1

/2=2.5%

t=

( )

f tH0

2.086 t 2.086

/2=2.5%

-ca/2

/2=2.5%

ca/2

33

tH0

33

summetra

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

71 / 91

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

72 / 91

Elegqoc Upojsewn - Monokatlhktoc Elegqoc

Elegqoc Upojsewn - Monokatlhktoc Elegqoc

'Estw y1 , y2 , ..., yT tuqao degma ap plhjusm N(, 2 ), me gnwst


diakmansh 2 . Edame ti h ektimtria elaqstwn tetragnwn tou
P
msou enai o deigmatikc msoc, = T1 Tt=1 yt . Jloume na
exetsoume thn upjesh mhdn H0 : = 0 nanti thc enallaktikc
HA : > 0 .
0

z =

2x /T

z=
f (z)

0
/T
2

N ( 0,1)

z=

- 0

z 1.645

z > 1.645

- / : H0

2 0 > c . (c , +),

N ( 0,1)

( ) -

> 1.645

0
1.645

N (0, 1) H0 .

2 /T

=5%

x /T

- (, c ).
c Pr (z > c ) = 1 (c ) =
= 0.05 (5%) , c = 1.645

=5%

zH0

c/1/2

36

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

73 / 91

Elegqoc Upojsewn ` Sflmata

:
H0
H0 ( HA ).

:
- H0
HA ( H0 ).

KWNSTANTINOU, FOUNTAS

Den Aporrptoume H0
:p =1
Sflma tpou II
II: p =
(OE)

Basikc Statis tikc 'Ennoiec

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

74 / 91

Elegqoc Upojsewn ` Eppedo Shmantikthtac kai


Dnamh

O kannac aprriyhc den enai tleioc, kai ta statistik kritria


pou qrhsimopoiome enai tuqaa. H dia diadikasa elgqou mpore
na odhgsei se diaforetik sumpersmata me th qrsh diaforetikn
deigmtwn. Epomnwc, endqetai na upopsoume se 2 tpouc
sfalmtwn:

H0 alhjc
HA alhjc

KWNSTANTINOU, FOUNTAS

36

Aporrptoume H0
Sflma tpou II: p =
:p =1
PA. MAK.

75 / 91

H pijanthta sflmatoc tpou I enai to eppedo statistikc


shmantikthtac tou elgqo. Aut sunjwc sumbolzetai me .
H dnamh tou elgqou enai h pijanthta aprriyhc thc upjeshc
mhdn tan enai yeudc.



Pr H0 H0 =



Pr H0 HA =



Pr H0 H0 =



1 Pr H0 HA = 1

Gia dedomno eppedo shmantikthtac ja jlame to na enai so to


dunatn mikrterh.
,
(
) !
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

76 / 91

Elegqoc Upojsewn ` Sflmata

Elegqoc Upojsewn ` Sflmata



H 0 : = 0


H 0 : = 0

=1%
=5%

=5%

2.5%

2.5%

0-c/2s

0.5%

0+c/2s


.
,
- .
-
5%. ,
, 5%.
1

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

77 / 91

Elegqoc Upojsewn ` Sflmata

1%,
1%.

.

,
. (
.)
5

KWNSTANTINOU, FOUNTAS

Basikc Statis tikc 'Ennoiec

PA. MAK.

78 / 91

=1%

=1%


H A : = 1

=5%

0.5%

0.5%

0.5%

0.5%

PA. MAK.

. ,
,
5% 1%.
8


. ,
H0 ,
.
Basikc Statis tikc 'Ennoiec


H A : = 1

=5%

HA : = 1
.

(OE)

(OE)


H0 : = 0

KWNSTANTINOU, FOUNTAS

Elegqoc Upojsewn ` Sflmata


H0 : = 0

0.5%

79 / 91

.
, ,
.
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

80 / 91

Elegqoc Upojsewn ` Sflmata

Elegqoc Upojsewn ` Sflmata



H0 : = 0


H0 : = 0

-
=1%

-
=1%
=5%

0.5%

0.5%

0.5%

Tloc, stw mia ektmhsh . Ja lboume th swst apfash


(aprriyh) en qrhsimopoisoume eppedo shmantikthtac 5%, all
ja upopsoume se sflma tpou II en qrhsimopoisoume eppedo
shmantikthtac 1% .
8

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

81 / 91

Elegqoc Upojsewn ` Sflmata

()
1%
1% -, .
.
HA ,
- HA
.
14

KWNSTANTINOU, FOUNTAS

=1%

HA : = 1

=5%

0.5%

PA. MAK.

82 / 91

0.5%

0.5%

15

Basikc Statis tikc 'Ennoiec

HA : = 1

=5%

5%,
HA ,
HA
- 5%.
.
, = 0.05 = 0.01
.
(OE)

Basikc Statis tikc 'Ennoiec

=1%

KWNSTANTINOU, FOUNTAS

(OE)


H0 : = 0

Elegqoc Upojsewn ` Sflmata


H0 : = 0

0.5%

0.5%

HA : = 1

=5%


H A : = 1

PA. MAK.

83 / 91

, , H0
. ,
;
15

H0 , 1%
5%,
( ).
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

84 / 91

Elegqoc Upojsewn ` Tim p (pvalue)

Elegqoc Upojsewn ` Tim p (pvalue)

Orismc
H tim p (pvalue ) tim pijanthtac (probability value) h opoa
sqetzetai me thn upologizmenh tim ap to degma enc statistiko
krithrou orzetai wc to kattero eppedo statistikc shmantikthtac,
sto opoo h upjesh mhdn H0 mpore na aporrifje, dedomnhc thc
timc pou lambnei to statistik kritrio (statistik) sto degma.

Strathgik Elgqou
Elgqou: Prosdiorzoume en to p-value gia thn
upologismnh tim ap to degma tou statistiko krithrou , enai
mikrtero megaltero tou epilegmnou epipdou statistikc
shmantikthtac .
pvalue
, ,
H0 .
pvalues
H0 .
Mikr pvalue (kont sto mhdn) sunist isqur erhma antjeto
proc thn H0 .
Meglo pvalue (kont sth monda) suniston adnama eurmata
antjeta proc thn H0 .

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

85 / 91

Elegqoc Upojsewn ` Sflmata

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

86 / 91

Elegqoc Upojsewn ` Sflmata


Grafik:

Dnamh tou elgqou


elgqou: 'Estw y1 , y2 , ..., yT N(, 1). Jloume na
exetsoume thn upjesh:

H0 : = 0
HA

: , 0

O
aprriyhc enai na aporryoume thn upjesh en
kannac

0
> c/2 . Epomnwc, h dnamh tou elgqou enai
1/T
!



0


Pr 2 > c/2 , 0
x /T

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

87 / 91


= 0 .

0 .

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

88 / 91

Elegqoc Upojsewn ` Statistik 2

Elegqoc Upojsewn ` Statistik t

'Estw y1 , ..., yT tuqao degma ap plhjusm N(, 2 ), me gnwsth


diakmansh 2 . Jloume na elgxoume:
H0 : = 0
HA : , 0




= y , sy2 , 2 .
t =

sy / T

tT 1 H0 .


- [c/2 , c/2 ] , Pr (|t| > c/2 ) = .
() (t),
H0 t < [c/2 , c/2 ].

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

89 / 91

Elegqoc Upojsewn ` Statistik F


'Estw x1 , ..., xm kai y1 , ..., yn anexrthta tuqaa degmata ap
kanonikoc plhjusmoc, megjouc m kai n antstoiqa. Jloume na
elgxoume:
H0 : 2x = 2y
HA : 2x , 2y




sx2 , sy2 2x , 2y .
F =

sx2
sy2

Fm1,n1 H0 .


- [c1,/2 , c2,/2 ] , Pr (F < c1,/2 ) = /2
Pr (F > c2,/2 ) = /2.
(F),
H0 F < [c1,/2 , c2,/2 ].
c1,/2
Pr (F < c1,/2 ) = Pr (1/F > 1/c1,/2 ) 1/F
Fn1,m1 (
).
KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

91 / 91

'Estw y1 , ..., yT tuqao degma ap plhjusm N(, 2 ). Jloume na


elgxoume:
H0 : 2 = 20
HA : 2 , 20
sy2 2
C = (T 1) sy2 /20 2T 1
H0 .

- [c1,/2 , c2,/2 ] , Pr (C < c1,/2 ) = /2
Pr (C > c2,/2 ) = /2.
(C),
H0 C < [c1,/2 , c2,/2 ] .

KWNSTANTINOU, FOUNTAS

(OE)

Basikc Statis tikc 'Ennoiec

PA. MAK.

90 / 91

Вам также может понравиться