Outline - stat.berkeley.eduwfithian/courses/stat210a/lectures/lecture16.pdfGaussian adjacent...
Transcript of Outline - stat.berkeley.eduwfithian/courses/stat210a/lectures/lecture16.pdfGaussian adjacent...
OutlineDX E and F distributions
2 Canonical linear model
3 General linear model
ReviewMultiparameter exp fans
OTCx X'UG ACOX e hCx
For Ho OEO vs H O Oo l side
or Ho 0 00 us H OF Oo 2 sided
The UMPu test 0 4 conditions on UCD
rejects for conditionally largelextreme TEX
Gaussian adjacent distributions
If Z Zd dN 0,1 thenshape scale
EZE XI Gamma 42,2
EV d Vad 2nd
As d as
I 51 PCIE H e 0 too
a Nld 2d
If Zn Nco D and V X'd ZIV then
retd Nco D as d 70
If V XI and K XIs V Vz then
V IdFd dz d XI as de a
Note if T td then TZ F.ae
Recall 2 nNd u E A c IRK'd.beRkAZ to NL Auto AEA
EI 1 sayle t test
X X El Ncu o7 nerd o o
We showed UMPU test of Ho M O us H M 0
rejects for large GI wheregz
I Lexi Ncm E5 1
E x ITn l
02L HX II RN XIrII IT E t
r is n 1
Why Orthogonal Projection 4 2
Xs N X
µII 11 11 41 711 ok
11km111 orarok Br X In Proj X
p
mm n1nn
7 X
s
I n I
Let Q g n Qr
where q In InEz En complete orthonormal basis
e.g via Gram Schmidt
X n Nn n In o In
New basis 1
Q'x n a IIIHQ'rXH HQxp Heixll
11 112 Nye Q'Q Inn 1 5
Q'X n N o Inh o
Zr Qi X N O o In
5 HZ.IT E X'mand 5215 we already knew from Basu
I
CanonicallinearModelf
a.me
fI.ynn l ozIn
dr n do d
Mo E IRI M E IRd
o 0
Test Ho u 0 us Hi M 0
or possibly one sided if d L
F xp Fam
plz eZ't E Zo Io 112112
1
Cord on Zo reject for largelsmalllextreme Z
Zo HZ test stat is Z Nlm o
Ei Io N o l Z testo
id reject for large 112,11117,114oz XI test
OZ unknown d I
Cord on Zo 112112 112,112 1112 112 1 241
Reject for largefsmalllextreme Z
Reject for large 71 11211
ZReject for largeqzI
E ta
Ct te st
id Reject for conditionally large HZ.lt
Reject for large 117114dL He f
FYd d dr
Here n XIfunctioning as estimator of o
IEE E Var 2
20ydCompare Z ZYo t Zyg
XZ 112,11 F 1172
LinealMany problems can be put into canonical
linear model after change of basis
Basic setup i
Observe Y NIO TIN or o
known or unknownTest O E
vs OE Iwhere
o E are subspaces of IRdimioto do dim D dotd
idea rotate into canonical formdo d n d
Q n Q Q Qrorthonormal o b for ab forbasis for n ot IR n
2 Q'Y Nn 8 8 o In
Ho Qi 0 0
Do Z X t or F test as appropriate
EI Linear Regression Xi C IRD fixed
Yi X B t ee e Nco E
Y Nd XP TIN X c IRD
Gin
Assume X has full column rank
O XB E Span Xi Xd
Ho 13 Bd _0 1 Ed Ed
O C Span Xd Xd
or 03 if d D
117112 HY Proj Y 112
HY xpno.si 2Bo.s argminHY XpH2
ECYi xipjzx'x x'y
Residual sunofsquares Rss
17,1121117112 HY Rojo.CH2
RSSo null Rss
F statistic is Zido Rss Rsscd do
HZrlf Cn d RSS kn d
n d called residualdegreesof freedom
d L Let Xo X XD E IRdo
Let X X Pro's Xi
X X x xD xo X
B Xie 4 11 11112 s e B 041 11
why O xp O Xup t Xp same p
E X'thx tl Q Y
t statistic E gB µRSSkn.FI Ie B
EI Two sample t test equal variance
Y Y id Ncn E y y dNCu ogI m until I n 1M
Model O EY no OE Sean CIN 1
Ho M _u c Span Imm
do L D 2 d ntm 2
orthogonalize HI
Reject for large
t.E.fi I.EYi I Ja
Eth aE FI