eollor.t tnlno applied linear regression models
TRANSCRIPT
TnlnoEollor.t Applied Linear
Regression Models
'F
John NeterUniversitl, of Georgia
Michael H. KutnerThe Cleveland Clinic Foundation
Christopher J. NachtsheimUnit ersity of Minneso ta
William WassermanSyracuse Llniversitt'
IBWINChicago . Bogatd .Lc.ndon . Madrid .
al ,nW|t concerned about Our Environment
D- -l| In recognition of the {act that our company is a large end-user of ftagile yet
Lt replenis-hable resources, we at IRWIN can assure you that every etfort is made to
meetorexceedEnvironmentalProtect ionAgency(EPA)recommendat ionsandrequire.ments for a'greener" workplace.
To preserve these natural assets, a number ol environmental policies' both companywroe
and departmenlspeciiic, have been rmplemented. From the use of 50o/o recycled paper in
ourtextbookstothepr int ingolpromot iona|mater ia|swithrecyc|edstockandsoyinkstoouroffice paper recycllng program, we are committed to reducing waste and replacing environ'
mentally unsafe products with safer alternatives'
@ The McGraw-Hil l Companies, Inc., 1983. 1989, and 1996
All rights reserved. No part of this publication may be reproduced, stored in a letrieval system, oI trans-
mitted, in any form or by any means, electronic, mechanical, photocopying, recording' or otherwise'
without the prior written permission of the publisher'
Trademark Acknowledgments:IBM and IBM pc are registefed trademarks oflnternational Business Machines corporation.
Disclaimer
Richard D. Irwin, Inc. makes no waranties, either expressed or implied, regarding the enclosed com-
puter software package, its merchantability or its fitness for any particular purpose. The exclusion of im-
plied *urrantiei is not permitted by some states. The exclusion may not apply to you' This warranty pro-
uid., you with specific legal rights. There may be other rights that you may have which may vary from
state to state.
Irwin Book Team
Publisher: Tom CassonSenior sponsoring editor: Richard T. Hercher, Jr'
Senior developmental editor: Gail Korosa
Senior marketing manager: Brian Kibby
Project editor: LYnne BaslerProduction supervisor: Laurie KerschDesigner: Matthew BaldwinCover designer: Stuart PatersonAssistant Manager, Graphics: Charlene R. Perez
Compositor/Art Studio: Graphic Composition, Inc'
Typeface: 10/12 Times RomanPrinter: R. R. Donnelley & Sons Company
Library of Congress Cataloging-in-Publication Data
Applied linear regression models / John Neter ' ' ' [et al']' - 3rd ed'
p.cm
lncludes index.
ISBN 0-256-08601-X
1. Regression analysis. I. Neter, John
QA278.2.A65 1996
519.5'36-dc20
Prfitted in the United States of America
34567890DO210987
95-37447
5.1 Matrices
Detinition of MatrixA matrix is a rectangular array of elements arranged in rows and columns. An example d
a matrix is: i:
D,
Matrix Approach toCnlpren
5 Simple LinearRegression AnalYsis
Matrix algebra is widely used for mathematical and statistical analysis' The matrix approach
is practicllly a necessity in multiple regression analysis, since it permits extensive systemf
ofequationi ancl large anays ofdata to be denoted compactly and operated upon efficiently'
In this chapter, we tirstiake up a brief introduction to matrix algebra. (A fuller treatmenl
of matrix algebra may be found in specialized texts such as Reference 5'1') Then wc
apply matrix-methods to the simple linear regression model discussed in previous chapten'
ettnougn matrix algebra is not really required for simple linear regression, the application
of matrix methocls to this case will provide a useful transition to multiple regression' which
will be taken uP in Parts II and III.
Readers lamiliar with matrix algebra may wish to scan the introductory parts of thit
chapter and l.cus upon the later parti dealing with the use of matrix methods in regressioR
analysis.
Column Column1)
Row I l- 16,000 nfRow 2 | 33,000 47 |Row 3 L 21,000 35 J
i
176
12 l6 ' le 8l
These two matrices have dimensions of 2 x 2 and 2 x 4, respectively. Norc thut in givingthe dimension of a matrix, we always specify the number of rows first urrtl thcrr lhc numberof columns.
As in ordinary algebra, we may use symbols to identify the elenrcnls ol :r rrurlrix:
" / : l j :2 j :3
i : l l a, atz on1i :2 | a2, azz ozz l
Note that the first subscript identifies the row number and the second thc r.rlurrrrr nurnber.We shall use the general notation aii for the element in the i th row and thc i rlr t.olrrrln. Inour above example, i : 1,2 and 7 - 1,2,3.
A matrix may be denoted by a symbol such as A, X, or Z.The symbol is irr lroll l ircc toidentify that it refers to a matrix. Thus, we might define for the above marrrr:
A:l at t otz ' ' ' l
Lazt az2 a:z )
Reference to the matrix A then implies reference to the 2 x 3 array just givcrr.Another notation for the matrix A just given is:
A: [a i j l i :1,2: j :1,2,3
This notation avoids the need for writ ing out all elements of the matrix by stl l irr:: ,rrlv thcgeneral element. It can only be used, of course, when the elements of a matrix rrr.r, svrrrlrols.
To summarize, a matrix with r rows and c columns wil l be represented cit lrer.irr l ir l l :
Other examples of matrices are:
A_
or in abbreviated form;
A:[a;1) i :1, . . . , r ; j -1, . .
or simply by a boldface symbol, such as A.
[ r o l l+ jl_s rol l : rs
Secriotr 5.1 Matrices 177
tacl ,
em:.rtl),le i r i
lers..t ior,hich
rhi:siol
leoatt ar t c l , ;
a- t t ar at ;
a; r o; t a; ;
A,t A, t Q.;
(5.1)
an(i
umr000)
's b1'I the
.C
178 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
Comments
L Do not think of a matrix as a number. It is a set of elements arranged in an array. Onlywhen the matrix has dimension I x I is there a single number in a matrix, in which case onecan think of it interchangeably as either a matrix or a number.
2. The following is not a matrix:
I v Ils ll " ll l0 ls lI s 16_]
since the numbers are not arranged in columns and rows. r
Square Matrix
A matrix is said to beexamples are:
square if the num
l+ 7lL3 el
ber of rows equals the number of columns. Two
l -1
I on arz ort
II ol azz or, ILo:, az2 ax J
Vector
Transpose
A matrix containing only one column is called acolumnvector or simply avector.Twoexamples arc;
The vector A is a 3 x I matrix, and the vector C is a 5 x I matrix.A nratrix containing only one row is called a row vector. Two examples are:
B' : [15 25 50] F' : [/i fr]
We use the primc syrnbol for row vectors for reasons to be seen shortly. Note that the rowvectorB/ isa I x 3 nratr ixandtherowvectorF/ isa 1x 2matr ix.
A single subscript suffices to identify the elements of a vector.
The transpose of a matrix A is another matrix, denoted by A', that is obtained by inter-changing corresponding columns and rows of the matrix A.
For example, if:
^: f i l . : [ r i ]L'ol l " rLcsl
,i,:li ' i]
Sulirn 5.1 Matices l7gthen the transpose A/ is:
A' : l ' - 7 3l2x3 L5 l0 4)
Note that the first column of A is the-first row_ of Ar, and similarly the sec.ncl c.lunrn of Ais the-secoJrdiow of A'..corr.espo"ai"gly,it"irst row of A has become rhe rirsr corumnof A/' and to onior. that theiime"r;;; ;f ;, indicated under the symbor A. bcconresreversed for the dimension of A,.As another example, consider:
Thus, the transpose of a column vector is a row vector, and vice versa. This is the reason;:1.ffi::3:?:'Jft'*i":,:lf:to iaeniriv a row vector, since ir may be thought oras
In general, we have:
= t l i t ) i :1, . . . ,c: i=1, . . . , r,v\
Row Columnindex index
ffiLf;"i:ment in the i th row and the 7th column in A is found in the 7th row and irh
Equality of Matrices
Two matrices A and B are said to be equal if they have the same dimension and if all:[Tii,","*T:;i;f;::;n5:i';:"*rserv,irtwomatricesareequar,rheircorresponding
,t, : f ,J] ,g, :ro 7 lol
T
I o ' ot , f
'X": l^ ' ; l : 1 ' i l i :1," . ' r ; j : r , . . . ,clu ' l or , | 'z
" \
J Row Column
(5.3) index index
,r,:li,l ,x, : fi]fo, , onl f n 21
.! . : lo, orr l n: f t+ t lLo, ot, _J 3x2
L t: 9l
A':cXr
thenA:Bimpl ies:
Similarly, if:
at:4 a2=7 a3:3
l - rI
or t or t I
l : : lt lLor, or ,
)
180 Chapter 5 Matix Approach to Simple Linear Regression Analysis
thenA:BimPl ies:
at t :17 at2:2
azt: 14 azz: 5
ay: 13 azz:9
Regression Examples
ln regression analysis, one basic matrix is the vector Y, consisting of the n observations onlhe response variable:
Note that thc transpose Y' is the row vector:
Yl : [Ir Yz Yn]lXn
Another basic matrix in regression analysis is the X matrix, which is defined as followsfor simple linear regression analysis:
5.2 Matrix Addition and SubtractionAdding or subtracting lwo matrices requires that they have the same dimension. The sum,or difference, of two matrices is another matrix whose elements each consist of the sum. orJiff'erence, of the corresponding elements of the two matrices. Suppose:
I r , IY:11' l
nxl | : IlY^ )
(s.4)
[l ::1(5.6.1 I" : l : : lnx2 l : ILt x ' )
The nratrix X consists of a column of ls and a column containing the n observations on thepredictor variatrlc X. Note that the transpose of X is:
(5 7, x, =l;, ;, . ": ]
A : l 1 l B :1 113x2 [ : 6J 3x2 l : 4 )
M ut ri.r At lJ i t i r trt und Subt raction 181
then:
A+B3x2
Similarly:
A_B3x2
In general, if:
A : [a i i lrx.c
: lbt j l l :1, ; - 1
then:
(s.8)1I"u
: lail * bill and ^,;j
: [a;i - b;i)
Formula (5.8) generalizes in an obvious way to addition and subtraction of more than twomatrices. Note also that A + B : B * A, as in ordinary algebra.
lz 6 l:14 8 |
L6 r0l
[o 21: lo 2l
L0 2)
Section 5.2
I t+r 4+21: lz+2 5+3 |
L3+3 6+4)
I t - t 4-21:12_2 5_31
L3-3 6-41ns on
rllows
on thc
e sunl,um, ()l
BrXc
flegression Example
The regression model:
f;*'Y'i""n comPactrY
Yi:E{Yi I*ei i :1, . . . ,n
in matrix notation. First, let us define the vector of the mean
/5 ql
and the vector of the error teffns:
E{Y}nXl
Iurr , t lI twrl I: l,,r,l
Y in (5.4), we can write the regression
r82 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
because:
5.3 MatrixMultiPlication
Multiplication of a Matrix by a Scalar
suppose the matrix A is given bY:
Similarly,l.A equals:
ln general, if A
(5.rr)
L-scalqt_i_s an ordinary number or a symbol reple!941!1ng a number. In multiplication of n '.matrix by a sialar, every element of the matrix is multiplied by the scalar. For example. I
o:13 i ]Then 4A, where 4 is the scalar, equals:
tll - f aall - f:: l : | ;ulrllL'l
: l',,,]
. L"l : L'''.i *'.1
Thus, the observations vector Y equals the sum of two vectors, a vector containing the
expected values and another containing the error terms'
T5 21l i i l r lsl i s l : tL3L, i . ) .1
= loil and .1. is a scalar, we have:
trA:Al- lLai j
oo: ol? i] : [,:2sl12)
where I denotes a scalar.If every element of a matrix has a common factor, this factor can be taken outside the '
matrix and treated as a scalar. For example:
Sirni lar ly:
^o:^[3 ] ] : [3]
I s 271 " f : e lLtr rs. l : ' ls e. l
7r l3r l
'11; l
Section 5.3 Matrix Multiplication 1g3
tllultiplication of a Matrix by a MatrixMultiplication of a matrix by a matrix may appear somewhat complicated at flrst, but a littlepractice will make it a routine operation.
Consider the two matrices:
,1,:11 i ] ,y, :u f ]The product AB will be a 2 x 2 matrix whose elements are obtained by finding the crossproducts of rows of A with columns of B and summing the cross p.oauar. For instance, tofind the element in the first row and the first column of the product AB, we work with thefirst row of A and the first column of B, as follows:
Al
*"* ' Ifzll l-f-4-lno*zf
^ t ) Lb" j
2(4)+5(5):33
The number 33 is the element in the first row and first column of the matrix AB.To find the element in the first row and second column of AB, we work with the firstrow of A and the second column of B:
Col. I
We take the cross products and sum:
A
*o*' [D-3-l l-+l+ l lRow2[4 t) ls
Col. 1
The sum of the cross products is:
2(6)+5(8):52
Continuing this process, we find the product AB to be:
Let us consider another example:
AB
now r [33 IL]Col. I
6l8J
Col. 2
AB
Row r [33 s2lLJ
Col. 1 Col .2
B
i6-llLEllCol. 2
f:,:l '^ i][l :] : i;i #l
,x, : l ; 3 i ] ,* , : f ; ]r i : f ; 3 3l [; ] : [r]
Chapter 5 Matrix Approach to Simple Linear Regression Analysis
When obtaining the product AB, we say that A is postmultiplied by B or B is premulti-plied by A. The reason for this precise terminology is that multiplication rules for ordinaryalgebra do not apply to matrix algebra. In ordinary algebra, xy : yx.In matrix algebra,AB + BA usually. In fact, even though the product AB may be defined, the product BA
may not be defined at all.In general, the product AB is only defined when the number of columns in A equals the
number of rows in I| so that there will be corresponding terms in the cross products. Thus,in our previous two cxamples, we had:
Equal
A,/ \BI x: 7x2\ ,vI ) inrcrrs ionol prot luct
Equal
A/\B : AB2x3 3xl 2xl
\ ,vDimensionof product
:AB2x2
Note thar r l re t l i r r rcnsiorro1' theproductABisgivenbythenumberof rowsinAandtherrurrrbcl ol coltrrrrns in l l . Note also that in the second case the product BA would not betlclirrctl sirrcr' t lrc rrrrrrrbcr of columns in B is not equal to the number of rows in A:
Unequal
z/\
l lcrc is rrrrollrel cxirrnple of matrix multiplication:
; t t t - [ ' r r r ar2 ot t l
L,1: I AZZ AZZ )
I r rgcrrcr ' : r l . i l Ah:rsdimensionrxcandBhasdimensioncxs,theproductABisamatrix ol t l irrrcrrsion r. x .r whose element in the i th row and lth column is:
aitbtrj
so lh i l l :
(5. 1 1) At l
'l'hus. in thc lirlcgoinu example, the element in the first row and second column of thegrlorluct Ali is:
3
,L.ortbm: atbt2 * arrbr, * arrby
as indeed we found by taking the cross products of the elements in the first row of A and
second column of B and summing.
A2x3
B3xl
f utt btrlIb^ bzzILbt, bt, )
* arrb, arrbr, * arrbr, + orrbrr l
* arrb, arrb,,2 * arrb22 * ozzbn )|
, r , ,b, , * arrb^-
f , , , ,b, , lazzbzt
c\-L
: [ i o,obo,l i -Lk=r
' l
Section 5.4 Special Types of Matrices 185
Additional Examples
t .
ti iiltnr:t1,i r,,i,lRegression Examples. A product frequently needed is Y'Y, where Y is tlrt' vectrlr ofobservations on the response variable as defined in (5.4):
I t , Ilv , l
t5. l3r y 'y: [Xr yz y, ,1 | . ' l : ly i +r l+. . .*1, ] l I ) ] ,1rxr -
| : Ir . . lL t,, -l
Note that Y'Y is a I x I matrix, or a scalar. We thus have a compact wly ()l l r ir irrr :r srrrrrof squared terms: Y/Y : rV?.
We also wil l need X'X, which is a 2 x 2 matrix. where X is defined in tr.r ' r
I r x, l[ r I " ' r l l r x, l [ , , r \ ,(5.14t X'X: l l l - l : l
i *z Lx ' x2 " , ,11, : I Ltx, r \ ,
Lr xn)
and X/Y, which is a2 x l matrix:
(s. rs)" ," : [ '
I " ' t l2xt LX' X2 Xr)
5.4 Special Types of MatricesCertain special types of matrices arise regularly in reurcssion analysis. We consider themost imoortant of these.
2.
I q 211 a,1 | +o, + 2a.1Is s jL, ;J:Lr, i +8, ; )
fz l12 3 s l l1 l : t '+32+s21 : ; :s1
L) l
Here, the product is a 1 x I matrix, which is equivalent to a scalar'. l'lrLrs. lhc rnatrixproduct here equals the number 38.
-).
Iv, I| ' ' l : [ ' , , I| : I L, \ , ) ,1L Y,,J
186 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
Symmetric Matrix
If A : A', A is said to be symmetric. Thus, A below is symmetric:
A,3x3: [ l i \ ][ r 4 61
A :14 2 5l3x3 L6 5 3l
Diagonal Matrix
A symmetric matrix necessarily is square. Symmetric matrices arise typically in regressionanalysis when we premultiply a matrix, say, X, by its transpose, X'. The resulting matrix,X/X, is symmetric, as can readily be seen from (5.14).
A diasonal matrix is a square matrix whose off-diagonal elements are all zeros, such as:v t - - .__, . . - :
We will often not show all zeros for a diagonal matrix, presenting it in the tbrm:
Two important types of diagonal matrices are the identity matrix and the scalar matrix.
Identity Matrix. The identity matrix or unit matrix is denoted by I. It is a diagonal matrix
whose elemenls on the main diagonal are all ls. Premultiplying orpostmultiplying any r x r
matrix A by thc r x r identity matrix I leaves A unchanged. For example:
atz onf f o, dtz or: Ia22 or.. l : I ou azz or, Ia3z an J Loy a32 an )
Sirni lar ly. we have:
AI:
Note that the identity matrix I therefore corresponds to the number I in ordinary algebra,
s ince we have there that 1 . x : x . l : x .In general, we have for any r x r matrix A:
,i,: [i + l] .*.: [l , ,l l]-'l
At^l' | l |
a1 ln:0l+x+
azl-J
3x3
[ t o ol [a, ,I t t 'lA:10I 0l lo,L0 0 l l [43'
f o, , atz a, , l I t o o- l f o, , arz o, : Ilorr azz atr l l0 I 0 l : lot ' azz onlLo' a32 a33l [0 0 l l Lo: ' a3z ai .3J
l0
(5.16) AI: IA:A
Section 5.4 Special Types of Matices 187
Thus, the identity matrix can be inserted or dropped from a matrix expression whenever itis convenient to do so.
Scalar Matrix. A scalar matrix is a diagonal matrix whose main-diagonal elements arethe same. Two examples of scalar matrices are:
| \ 0 0ll? 9l l ; ^ 0rLo 2)
[o o i_]
A scalar matrix can be expressed as ,r"I, where ), is the scalar. For irrslrrrrcc:
lz o l_, I r o lfo , ) : ' lo iJ : zr
[ ; , o ol [ r o ollo r o l :^ lo I o l :11L0 0 i ._ j L0 0 r_l
Mult ip ly inganrxrmatr ixAbytherxrscalarmatr ix) , I isequivalcrrr r ' r r r r r l r i ; r ly in,u{ by the scalar ),.
Vectcir and Matrix with All Elements lJnity
A column vector with all elements 1 will be de
(5. 17) I -rXl
and a square matrix with all elements I will be
I r(5.18) f :1,
'xr I r
For instance, we have:
l ' l : [ llx l
noted by 1:
[ ' lt l lLIJdenoted by J:
l lr_l
l r l f r I r lr : l t l I : l r I r l
3xr Lr_j 3x. i [ r I t ]
I vector I we obtain:
| - ' l"L;JNote that for an n x
: [ r1 l : 11
| 8tt ('lruptcr 5 Matrix Approach to Simple Linear Regression Analysis
and:
l t l [ r=l, lu rr : l iL'J Li
r l
i l : JnXn
Zero Vector
5.5 Linear Dependence and Rank of Matrix
Linear Dependence
Consider the followins marrix:
A zero vector is a vector containing only zeros. The zero column vector will be denotedbv o:
(5.1e)
For example, we have:
03x1
'*,:f:]: [3]
L0_l
[ r 2 sd, : lz 2 to
L3 4 rs
[ , ; j : , I j ]L rs l L3l
:lr l
Let us think now of the columns of this matrix as vectors. Thus, we view A as being madeup of four column vectors. It happens here that the columns are interrelated in a specialmanner. Note that the third column vector is a multiple of the first column vector:
We say that the columns of A tre linearly dependent. They contain redundant information,so to speak, since one column can be obtained as a linear combination of the others.
we define the set of c column vectors cI,..., c" in an r x c matrix to be l inearlydependent if one vector can be expressed as a linear combination of the others. If no vectorin the set can be so expressed, we define the set of vectors to be linearly independent. Arnore general, though equivalent, definition is:
(5.20) When c scalars i,1, . . . , i.", not all zero, can be found such that:
i .1cr + L2c2+.. .+ lccc - 0
Section 5.6 Inverse of a Matrix 189
where 0 denotes the zero column vector, the c column vectors are lin-early dependent. lf the only set of scalars for which the equality holds islr : 0, ...,1. : 0, the set of c column vectors is l inearly independent.
Toi l lustrate forourexample, l t = 5, lz = 0. l : : -1,1+ :0leads to:
'[i].'[1] 'Ir] .'[i] : [l]Hence, the column vectors are linearly dependent. Note that some of the,l.; equal zero here.For linear dependence, it is only required that not all .1,, be zero.
Rank of Matrix
The rank of a matrix is defined to be the maximum number of linearly independent columnsin the matrix. We know that the rank of A in our earlier example cannot be 4, since the fourcolumns are linearly dependent. we can, however, find three columns (1, 2, and 4) whichare l inearly independent. There are no scalars .].r. 1,, 14 such that ),1cl + ),rcr+ ).oco : 0other than Lt : Lz - L4 : 0. Thus, the rank of A in our example is 3.
The rank of a matrix is unique and can equivalently be defined as the maximum numberof linearly independent rows. It follows that the rank of an r x c matrix cannot exceedmin(r, c), the minimum of the two values r and c.
When a matrix is the product of two matrices, its rank cannot exceed the smaller of thetwo ranks for the matrices being multiplied. Thus, if C : AB, the rank of C cannot exceedmin(rank A, rank B).
5.6 Inverse of a MatrixIn ordinary algebra, the inverse of a number is its reciprocal. Thus, the inverse of 6 is l . A
onumber multiplied by its inverse always equals l:
=-f , .X :x ._{ : I
In matr ix algebra, the inverse of a matr ix A is iurothcl r r r r t r ix . t lcrrotcd bv A-1, suchthat:
(s.21) r \A'- . I
where I is the identity nrutrix. ' l 'hus. rrrrrrirr. thc idcrrtity nratrix I plays the same role as thenumber I in ordinary algebra. An invelsc of a matrix is defined only for square matrices.Even so, many square matrices do not have an inverse. If a square matrix does have aninverse, the inverse is unique.
. l61
x.-
I__.h_l
6
190 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
Examples
l. The inverse of the matrix:
IS:
sr l lcc:
A_1A:
oI:
AA-I _
l . ' l ' l rc i r rvcrsr ' o l ' the matr ix:
f ' t 41
,!., : Ll il
A-, : i-2x2 L
I r 4112 4l1.3 - .2113 r l
l : i l 11 -11: [ ; ?]: '
. l .41I
: [ ; ?] : '
[ r o olA :10 4 0l
3x3 L0 0 2)
l i ; ; l[+ g ol i , oro i oro.3 ' l : i ; ?3- l : ,Lo o i_ l Lo o 2J [o o rJ
A-rJAJ
since:
A',A =
Nolc thlt tltt' ittvcrsc o1'a diagonal matrix is a diagonal matrix consisting simply of thercciprocirls ol t lrc clcntents on the diagonal.
Finding the lnverse
lJp to this Poirrt. rhc inverse of a matrix A has been given, and we have only checked tonrakc surc it is the invcrse by seeing whether or not A-1A _: I. But how does one find theinversc, ancl wlrcn tklcs i1 exist?
An inverse tll a square r x r matrix exists if the rank of the matrix is i.. Such a matrixis said to be ntttrsittl4ulur <>r of .full rank. An r x r matrix with rank less than r is said to be:;ingulctr or not ol'.litll ntnk, and does not have an inverse. The inverse of an r x r matrix oflull rank also has rank r'
Finding the inverse of a matrix can often require a large amount of computing. We shalltake the approach in this book that the inverse of a2 x 2 matrix and a 3 x 3 matrix can becalculated by hand. For any larger matrix, one ordinarily uses a computer or a programmable
Section 5.6 Inverse of a Matrix 191
calculator to find the inverse, unless the matrix is of a special form such as a diagonal matrix.It can be shown that the inverses for 2 x 2 and 3 x 3 matrices are as follows:
1. I f :
*, : lx 3 lthen:
(s.22)
where:
(5.22a) D:ad-bc
D is called the determinant of the matrix A. If A were singular, its determinant would equalzero and no inverse of A would exist.
2. rf:
F ' '1
- l
4-r : lo o, l2x2 Lc dJ
f d -bf
: I D D Il -c a ILD Dl
then:
r5 ?1r
where:
B-1 :3x3
B:l i2; l3x3
Ls h k)
fa b , f - ' le B clld n f | : lD E r lLs h k) LG H K)
7:(ek-fh) lz B:-(bk-ch) lz C:(bf-ce) lZ(5.23a) p:-(dk-f i l /Z 6:(ak-cC)lZ F:-(af-cd) lZ
6:(dh-edlZ g:-(ah-b7) lZ K-(ae-bd)/Z
and;
(s.23b) Z : a(ek - fh) - b(dk -fil * c(dh - eg)
Z is called the determinant of the matrix B.
Let us use (5.2D to find the inverse of:
We have:
f ' t r ,1A:l l , I
LJ TI
a:2 b:4
c:3 d:1
192 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
D : ad - bc :2(1) - 4(3) : -19
Hence:
- . r .41.3 - .2 )
as was given in an earlier example.When an inverse A-l has been obtained by hand calculations or from a computer
il#tti;tf s!:iff :"tri;:;ffi ilH,lf;1'ffi :"iil,Trilr;l,ffr"hffi ::
Regression Example
Ihe principal inverse matrix encountered in regression analysis is the inverse of the matrixX'X in (-5. l4 l :
Using rule (5.22), we have:
a:n b:LXi
, :LXt d:LX?
so that:
I t -4 IA-,=l =T f t l : f
l_3 2l LL -10 - l0 l
f n tX, l
I 'T :1"", txi-l
D:,, fx i - ( I " , ) ( f r , ) : " l t r f -
(-5.24 )
Since X X ; : t tX ancl X(X; - I ) ' :>X? - n*2,wecan simpl i fy (5.24):
r r v. i2 l\ut ' i ) l :n l ,1x, -*)2| 4. Inl
| >x? -EXi I, | ^>A:-ZP ,>d-W I
(X'X;- ' : ; I2x2 | -XXi n I
t,>,,x-W ;t1;|:VF )
fL- r ' - t I--. _ _, I r - t(X, rTP t(x=r I(X'X)- ' : ; I2x2 l -X1l
L ;A;;7 ;6;vP )
Hencc:
(5.24a)
Uses of lnverse Matrix
Section 5.6 Inverse of a Matrix 193
In ordinary algebra, we solve an equation of the type:
5Y :20
by multiplying both sides of the equation by the inverse of 5, namely:
lrsrr : lrrurf )
and we obtain the solution:
, : !p0,1 :4-5
In matrix algebra, if we have an equation:
AY: C
we corespondingly premultiply both sides by A-l , assuming A has an inverse:
A-IAY: A-rC
Since A-lAY : IY : Y, we obtain the solution:
Y: A- lC
To illustrate this use, suppose we have two simultaneous equations:
2y, I 4yr:29
31tfY' : lQ
which can be written as follows in matrix notation:
lz +l f i r I : I ro lL3 l l [ r : l L lo l
The solution of these equations then is:
[v, ' l - [z 4.1-r [2olL,, l :L: r l L 'o l
Earlier we found the required inverse, so we obtain:
[ r , l - l . t 4.1 [20' l : [ r . l
Hence, yt :2 and yr - 4 satisfy these two equations'
194 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
5.7 Some Basic Theorems for MatricesWe list here, without proof, some basic theorems for matrices which we will utilize in laterwork.
(5 ?5)
(s.26)
(s.27)
(s.28)
(s.2e)
(s.30)
(s.31)
(s.32)
(5 3?\
r5 34)
(5 ?5\
(5 ?6\
(s.37)
A+B:B*A
(A+B)fC:A+(B+C)
(AB)C: A(BC)
c(A+B):CA+Cg
).(A+B):) ,A+i"B
(A,), : A
(A + B)/ : A'+ B'
(AB)/ : B'A/
(ABC), :C,B,A,
(AB)- ' - B-rA-r
(ABC)- ' - C-tB-tA-r
(A-t ; - t - O
(A')- t : (A-r) '
5.8 Random Vectors and MatricesA random vector or a random matrix contains elements that are random variables. Thus,the observations vector Y in (5.4) is a random vector since the Y, elements are randomrtariables.
Expectation of Random Vector or Matrix
Suppose we have r : 3 observations in the observations vector Y:
:[l]The expected value of Y is a vector, denoted by E{Y}, that is defined as follows:
I t t t ' l IE{Y} : lElYzl l3x1
Luttr t - l
Y3xl
Section 5.8 RandomVectors and Matrices 195
Thus, the expected value of a random vector is a vector whose elements are the expectedvalues of the random variables that are the elements of the random vector. Similarly, theexpectation of a random matrix is a matrix whose elements are the expected values of thecorresponding random variables in the original matrix. We encountered a vector of expectedvalues earlier in (5.9).
In general, for a random vector Y the expectation is:
(s.38) E{Y} : lEIYiJlnX I
i : I , . . . ,n
and for a random matrix Y with dimension n x p, the expectation is:
(s.39) i :1, . . . ,n i j : I , . . . ,PE{Y} : lE{Yi i } lnxp
Regression Example. Suppose the number of cases in a regression application is n : 3The three error terms €t, €2, t3 each have expectation zero. For the error terms vector:
I3xl
we have:
E{e} : 03x1 3x1
since:
,lariance-Covariance Matrix of Random Vector
Consider again the random vector Y consisting of three observations Yr, Yz, L,. Thevariances of the three random variables. o:{Yi}, and the covariances between any twoof the random variables, o {Yi, Y1lt, are assembled in the yariance-covariance matrix of Y ,denoted by 62{Y}, in the following form:
(s.40) 62{Y}:
Note that the variances are on the main diagonal, and the covariance o {Y; , Y i} is foundin the i th row and 7th column of the matrix. Thus, o {I2, I, } is found in the second row, firstcolumn, and o {Y ,, 12 } is found in the flrst row, second column. Remember, of course, thato{Yr,Yr i ' - o{Yr, I2} . Since o{Y;,Y;} : o{Yi , Y;} for a l l i t ' j .o2{Y} is a symmerr icmatrix.
=
[" ]
I t { ' ' } I Io lLf[;]l: L3l
f " t { r r } o{Y1,Y2l
" ty, , y:} l
I o{Yr,Y,} o2{yr l otYr, \ } |
lo{Yr,Y,} o{ \ ,Y2} o ' {Y) J
196 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
It follows readily that:
(5.41) o2{y} : E{ty - E{y}lty - E{y}l '}
For our illustration, we have:
or{Y} :
Multiplying the two matrices and then taking expectations, we obtain:
Lclcation in Product Expected Value
o2lYr lo lY 1, Y2lro {Yr, Yr}o {Yr, Yr]r
etc.
This. ol'course , lcads to the variance-covariance matrix in (5.40). Remember the definitionsol'variarrcc and covariance in (A.15) and (A.21), respectively, when taking expectations.
To ge rrctlrliz.c. the variance-covariance matrix for an n x I random vector y is:
'{[l -ritil ,'-E{Yt} Y2- E[Y2]
"
-U,rr" ]
(5 4) l
Row l, column 1llow l, column 2l{ow l, column 3ll.ow 2, column I
etc.
( '2{Y}nX,l
(Yt _ E{YJ)z(yt- E{Yt})(Y2- E{Y2})(Yt - ElYl ))(r: - E{yz]j)(Y2- EIy\DVt - EIYJ)
etc.
olYr,Yr)oz IYr|
| " ' t 'I o lY, .: l 'I
Lo{Y" 'Yt\ o lYn, Y2)
Note again thar o21Y] is a symmetric matrix.
Regression llxarnple. Let us return to the example based on n :3 cases. Suppose thatthe three error lertt ls have constant variance, o2le,] : oz,andare uncorrelated so that o {ei,e; ) : 0 l-or i I .i. "l'he variance-covariance matrix for the random vector e of the previousexamplc is t lrcrclirrc as follows:
62{€}3x3
Notc: thrrt rtll r,:trianccs are o2 and all covariances are zero. Note also that this variance-ctlvariancc ttr:tlrix is a scalar matrix, with the common variance o2 the scalar. Hence, wec:rn cxprcss thc variunce-covariance matrix in the following simple fashion:
o2 {El - o2l3x3 3x3
o {Yt, Yr\
o {Y2, Ynl'r ]Yrl
fo ' o ol:10 o2 0 |
Io o "r ]
[ t o ol fo, o ol" ' lo I o l : l o o2 o I
L0 0 lJ L0 0 o ' )o2l =
Section 5.8 Randont Vettors and Matrices 197
Some Basic Theorems
Frequently, we shall encounter a random vector W which is obtained hy prcrrrultiplying therandom vector Y by a constant matrix A (a matrix whose elements arc Iixctl):
(5.43) W: AY
Some basic theorems for this case are:
where o2{Y} is the variance-covariance matrix of Y.
Example. As a simple illustration of the use of these theorems, consider:
I
(s.44)
{ \ l l ' \ l
(5.46)
We then have by (5.45):
E{w)2x1
and by (5.46):
o21W1 :2x2
Thus:
E{A} : A
E{W} :E{AY} :AE{Y}
62{w} : o2{AY} : Ao2{Y}A'
w2x1
A2x2
I Iu,r , ] l : IE{rr}- r { r , } lI Lr t r , t l Lr ty, ]+ EtY) l
I r: I r
Y2xl
-1
I
- o ' lYt - Y2J : o2{yr l + o21vry - 2olYr,Yr)
- o ' lY, * Y) : o2{Yr l + 02lY2} * 2olYr, Y2l
- o lyt * yz,y, * y2| : o ' lyr j - o2lyr j
I r - ' l I o?1Y,1 o{Y,.r , } l [ | t lI t r J [ "1rr . r ,1 "r1vry ) l - t r I
f " ' {v,} + 02lyzJ - 2olyr ,yr l o2{y) - o2{yr l
L o21Yrl - o ' lyr \ o2lyr l + o21rry *2o{Y,, , , , ]
o21w,l
o21wrl
o lW1, W2)
198 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
5 e si mp'e'';::J"::il"::::l *::;:::ff:T: ::J.T:, Remember again,ha,
we will not present any new results, but shall only state in matrix terms the results obtained
earlier. We beein with the normal error regression model (2.1):
(5.47a)
Yt: f ro * B,X, * e,Yz:f l0*BtXrte,
Yn:frolBtXr+e,
We defined earlier the observations vector Y in (5.4), the X matrix in (5.6), and the I vector
in (5.10). Let us repeat these definitions and also define the p vector of the regression
coefficients:
(s.48)
(s.41)
This implies:
Yi: f lo l BrX,+ e, i :1. . . . .n
(s.4e)
since:
[ r x, I f ' l ll l 1 ' l [ ! :1.1: ,Lt xn) L ' ,_J
IBo* F,x,1 [ ' ' l Ino* B,x,+e,)
lao+0," ' l * | ' : l : | , ' * P'x,+ '=
|Lr,*'u,r,) L"l Lr,* u,'*,*,,)
Note that Xp is the vector of the expected values of the I; observations since E{Y;}
Bo+ Brx, ;hence:
E{Y} : XpnXl nXl
Y:nXl
[ r , I [ r x, I [ ' , I
I t l .L: l l " ' l ,g, : [ f l ] , i , : l ' ' ILr,_l [ t x, , ] L ' ,1
(5.47a) in matrix terms compactly as follows:Now we can write
YnXl
=x p+€nX2 2\ I nXl
t l ll . . l :Lr, l
t
f.e
(s.s0)
Section 5.9 Simple Linear Regression Moclel in Matrix Terms lgg
where E{Y} is defined in (5.9).The column of ls in the X matrix may be viewed as consisting of the dummy variable
Xo : I in the alternative regression model (1.5):
Yi : FsXo -t BtX; * e; where Xn : I
Thus, the X matrix may be considered to contain a column vector of the dummy variableXo and another column vector consisting of the predictor variable observations X;.
with respect to the error terms, regression model (2.1) assumes that E{€;} : 0,o21e;\ : o2, and that the e; are independent normal random variables. The conditionE{r;} : 0 in matr ix terms is:
(s.s r) E{€}nXl
since (5.51) states:
Iat . , t l I o- ll t { . , } l_ l o Il : l l : l
fur ' ,r j L r lThe condition that the error terms have constant variance o2 and that all covariances
6l€i ,€ j l for i I j arezero (s incethee, areindependent) isexpressedinmatr ixtermsthrough the variance-covariance matrix of the error terms:
:0nXl
/5 52 r
Since this is a scalar matrix, we know from the earlier example that it can be expressed inthe following simple fashion:
f",^ l0
o'{e} : Inxn' |
:
L0
olol. l: l
" t )
00o2o
::00
(5.52a)
Thus, the normal error regression
(s.s3)
where:
e is a vector ofoz{e} : 621
62{E} - o2rt rxn nxn
model (2.1,) in matr ix terms is:
Y:XF+e
independent normal random variables with E{e} : 0 and
Chapter 5 Matrix Approach to Simple Linear Regression Analysis
5.10 Least Squares Estimation of Regression Parameters
Normal Equations
The normal equations (1.9):
(5.54)
in nratrix ternrs are:
(.5..s-5 ) X'X b : X/Y2x.2 zxl 2xl
wlrcrc b is thc vcctor of the least squares regression coefficients:
(5.55a) b : [? ' l2x1 Lbr )
To see lhis. rccirl l that we obrained x/X in (5.14) and x'Y in (5.15). Equation (5.55) thus
["i, ;;i][i:1:t;';]I nbr+btrx i l_[ t t , - l
Iao:x, +bt,x? ) Lr" , r , lThesc irrc prcciscly the normal equations in (5.54).
E sti mated Reg ress i on C oeff i c i ents
Ttr ohtuilr lhc eslirnated regression coefficients from the normal equations (5.55) by matrixrncthotls. rvc Ple rrrult iply both sides by the inverse of X/X (we assume this exists):
(x'x)-rx'xb - (X'x;-t*'t
We then l ind, s incc (X'X;- t ;1 ' " : f and Ib : b:
tS 56r 5 : (X/X)-tX,y2x1 2x2 2xl
The estimators bo and b, tn b are the same as those given earlier in ( 1. l0a) and ( l. l0b). Weshall demonstrate this by an example.
nbo+brlXi : f Y,
boL xi * brL x! : 2 x,Y,
' | '
I
Sectbn 5. I0 Least Squares Estimation of Regression parameters 201
Example. We shall use matrix methods to obtain the estimated regression coefficients forthe Toluca Company example. The data on the I and X variables were given in Table l.l.Using these data, we def,ne the Y obserrrations vector and the X matrix as follows:
15 57\
[ :eel I r 8ol(a)
" : l '? ' | (b) * : l l ' : l1,,, J Li ̂ j
We now require the following matrix products:
I z.soz IL 617,180 I
I r 80lt ' l | 1 ,0 |Tol l ' : I
Lr 70)
[:eo Ir l | 12r I
7o. l | : l :1323J
l - r I . . .(5.58) X' ,X: | ̂ .180 30
_ t 2s 1.7s0'l- L l.7so t42.3ooJ
. l - r r(5.59) X'Y: l^ .180 30
Using (5.22), we find the inverse of X'X:
(5.60)
In subsequent matrix calculations utilizing this inverse matrix and other matrix results. weshall actually utilize more digits for the matrix elements than are shown.
Finally, we employ (5.56) to obtain:
t5.6lr b: | -?ol : rX'Xr- ,X'y: I l t '172 - 99] :1: I t 7.807. l- lb, l - ""^' l-.oo:s:s .0000s0s r .1 l e l7. ts0 J_l ez.n I
| 3.s102l
or bo: 62.37 andbr :3.5102. These results agree with the ones in Chapter l. Anydifferences would have been due to roundins effects.
Comments
I . To derive the normal equations by the method of least squares, we minimize the quantity:
Q:DIYi _ $o+ Fl))2
(x,x)-, : [_ 33]1li - 33il:;,, ]
In matrix notation:
(s.62)
Expanding, we obtain:
Q: (Y - Xp),g _ Xp)
e :Y'Y - F'x'Y - Y'xF + p'x'xp
2ll2 Clnpter 5 Matrix Approach to Simple Linear Regression Analysis
since (Xp)/ : p'X'by (5.32)- Note now that Y'Xp is I x 1, hence is equal ro its transpose,which according to (5.33) is p'X'Y. Thus, we find:
(5.63) Q:Y'Y -2F'x'Y + p'x'xp
To find the value of p that minimizes Q, we differentiate with respect to po and Br.Let'.
(5.64)
Then it follows that:
(s.6s)
f ao1| - |a '^ . I ado Iap(s '= | og I
lah )
eaF(Q:-2x'Y+2x'xp
Equating to the zero vector, dividing by 2, and substituting b for p gives the matrix form of theleast squares normal equations in (5.55).
2. A comparison of the normal equations and X'X shows that whenever the columns ofX'X are linearly dependent, the normal equations will be linearly dependent also. No uniquesolutions can then be obtained for bo and b,. Fortunately, in most regression applications, thecolumns of X'X are linearly independent, teading to unique solutions for bn and 0,.
5.11 Fitted Values and Residuals
Fitted Values
(s.66)
In matrix notation, we then have:
(s.67)
noted by Y:
-^-lv, l
l t , l: l ' ll^ ' lLY' l
Let the vector of the fitted values Y, be de
f a^+ u,x, lI u i+a' ,x" It - lIr, * r,", -]Ltl[l:]
tn\1
Y:nXl
xbnx2 2x1
f ? ' l :Lot J
;
Section 5.II FittedValues and Residuals 203
Example. For the Toluca Company example, we obtain the vector of fitted values usinsthe matrices in (5.57b) and (5.61):
(5.68)
The fitted values are the same, of course, as in Table 1.2.
Hat Matrix. We can express the matrix result for t in (S.OZ) as follows by usins theexpression for b in (5.56):
t : X(X'X)- 'X'y
or, equivalently:
I r 80 l [:+z.rs Iv:xb: I l , I f ,3 j l . , l : |
16e47 |
Li to j ' ' [ r228_]
(s.6e)
where:
(5.69a)
Y:HYnXl nXn nXl
: X(X'x)-rx'HnXtr
Residuals
We see from (5.69) that the fitted values Y, canbe expressed as linear combinations ofthe response variable observations f,, with the coefficients being elements of the matrixH. The H matrix involves only the observations on the predictor variable X, as is evidentfrom (5.69a).
The square n xn matrix H is calledthe /za tmatrix.ltplays animportantrole in diagnosticsfor regression analysis, as we shall see in Chapter 9 when we consider whether resressionresults are unduly influenced by one or a few observations. The matrix H is symmetric andhas the special property (called idempotency):
(s.70) HH: H
ln general, a matrix M is said to be idempolenl if MM : M.
Let the vector of the residuals e, :
(s.1r)nXl
Y.-Y.he'1 denoted bv e:
[ ' ' Il " , l1", l
2(l.l ('lnt1t11'v 5 Matrix Approach to Simple Linear Regression Analysis
In matrix notation, we then have:
(s.72) e=YnXl nXl
Y: Y -XbnX 1 nXl nXl
example, we obtain the vector of the residualsExample. For the Toluca Companyusing the results in (5.57a) and (5.68):
(s.73)[:ee I [:+z.es I
": | '1 ' l - | ' ' , i ' l -lzzz ) 1r,r.rrl
The residuals are the same as in Table 1.2.
Variance-Covariance Matrix of Residuals. The residuals e,, like the fitted values i,can be expressed as^linear combinations of the response variable observations I;, using theresult in (5.69) for Y:
e=Y-?:Y-HY:f l -mY
We thus have the important result:
(s.74) e:( IH)YnX I nxn nxn nXl
by
I sr .ozl
l-or;or I| '0 "-l
where H is the hat matrix defined in (5.69a). The matrix I - H, like the matrixsymmetric and idempotent.
The variance-covariance matrix of the vector of residuals e involves the matrix I
H, is
_H:
(s.7s)
and is estimated by:
(5.76)
Note
The variance-covariance matrixe : (I - H)Y, we obtain:
o2{e} =o2(I-H)nXn
s21e1 :MsE(I-H)nXn
a':
of e in (5.75) can be derived by means of (5.46). Since
or1e1 : ( I - H)62{Y}( I - H) '
Now oz{Y} : o2{s} : o2lfor the normal error model according to (5.52a). Also, (I - H)/I - H because of the symmetry of the matrix. Hence:
o21e1: o2(I-H)I( I -H)
: o2(t - H)( I - H)
i ,Y.:i--
5.12 Analysis of Variance Results
Sums of SquaresTo see how the sums of squares are expressed in matrix notation, we begin u'ith the total sum
of squares SSZO, definedin (2.43).It will be convenient to use an algebraically equivalent
expression:
Section 5.12 Anahsis ofVariance Results 205
lnviewofthefact thatthematr ix l -His idempotent,weknowthat( I -H)( I -H): I -H
and we obtain formula(5.15).
(s.11)
We know from (5.13) that:
Y,Y:L,Y?
The subtraction terrn (rY)2 ln in matrix form uses J, the matrix of ls defined in t5.18), as
follows:
(s.7s) (r v' )r : (1) v'rv
n \n/
For instance, if n :2, we have:
SSIO : LtYi - Y l ' : L,rz - . lYi \2
n
(i) ,, ,,r Il l] [l;] :Hence, it follows that:
(Yt+y2)(Yr+Y,)
(s.19\
Just as X ff is represented by Y'Y in matrix terms, so 55p : L el
be reoresented as follows:
(s.80) SSE: e'e : (Y - Xb)'(Y - Xb)
which can be shown to equal:
(5.80a) SSE: Y'Y - b 'X'Y
Finally, it can be shown that:
sszo: Y/Y - (1)v'r
ssR: b'x'y - (1)v'n
: t ( r i - f ; t r can
(s.81)
Chapter 5 Matrix Approach to Simple Linear Regression Analysis
Example. Let us find SSE for the Toluca Company example by matrix methods, us-ing (5.80a). Using (5.57a), we obtain:
Y/Y: [399 121
and using (5.61) and (5.59), we find:
[:se Il r2r l
32311 | : 2,74s,173l : lL323 )
b,yly : t6237 3.s102t Ir , ] :?33]
:2.6s0,348
sSE : y'y - tr'xiy : 2,745,1',73 - 2,690,34g : 54,g25
which is the same resull as that obtained in Chapter l. Any difference would have been dueto rounding effects.
Note
To illustrate the derivation of the sums of squares expressions in matrix notation, consider SSE:
ssE: e/e : (y _ xb)/(y _ xb) : y,y _ 2b,x,y + b,x,xb
In substituting fur the rightmost b we obtain by (5.56):
ssE : y'y - 2btx,y + b,x,x(x/x)-,x,y
: y'y - 2b'x'y + b/Ix'y
tn dropping I and subtracting, we obtain the result in (5.80a).
Sums of Squares as Quadratic Forms
The ANOVA sums of squares can be showntobe quadraticforms. Anexample of aquadraticform of rhe observations Y, when n :2 is:
(s.82) 5Y? +6YtYz+4Y:
Note that this expression is a second-degree polynomial containing terms involving thesquares of the observations and the cross product. We can express (5.82) in matrix terms asfollows
(5.82a)Is
lY' Yr l l :_-t1L.-
llli:l : "'o"where A is a symmetric matrix.
Section 5.12 Analysis ofVariance Results
In general, a quadratic form is defined as:
(5.83)
(5.85a)
(s.8sb)
(5.85c)
(5.86a)
(s.86b)
(5.86c)
Y'AY: i io, ,y,y.rx l Et i !1 ' t ' t
where aii : aii
A is a symmetric n x n maftix and is called the matrix of the quadratic form.The ANovA sums of squares .ssro, ,s,tE, and ssR are ali quadratic forms, as can be
seen by reexpressing b/X/. From (5.67), we know, using (5.32), ihat:
b'X' : (Xb)': t '
We now use the result in (5.69) to obtain:
b'X' : (HY) '
Since H is a symmetric matrix so that H, : H, we finally obtain, using (5.32):
(5.84) b,X, _ Y,H
This result enables us to express the ANovA sums of squares as follows:
sszo: Y'ir - (;) r]'
SSE: Y,( I _ H)Y
ssR: " ' [ "
- ( ; ) ' ] "
Each of these sums of squares can now be seen to be of the form y/Ay, where the three Amatrices are:
' - (1)r\n/
I -H
"- 11)r\n/
Since each of these A matrices is symmetric, sszo, ssE, and ssR are quadratic forms,with the matrices of the quadratic forms given in (5.86). Quadratic forms play an importantrole in statistics because all sums of squares in the analysis of variance for linear statisticalmodels can be expressed as quadratic foms.
208 Chapter 5 MatrixApproachto Simple Linear Regression Ana[ysis
5.13 lnferences in Regression AnalysisAs we saw in earlier chapters, all interval estimates are of the following form: point estimator
plus and minus a certain number of estimated standard deviations of the point estimator'
Similarly, all tests require the point estimator and the estimated standard deviation of the
point esiimator or, in the case of analysis of variance tests, various sums of squares' Matrix
algebra is of principal help in inference making when obtaining the estimated standard
deviations and sums of squares. We have already given the matrix equivalents of the sums
of squares for the analysis of variance. We focus here chiefly on the matrix expressions for
the estimated variances of point estimators of interest'
Regression CoefficientsIhe variance-covariance matrix of b:
02{b}2x2
I o ' luol o{bg,btJ l:
Io{b1. be} o2lb] )
o2{b} : ozlx 'X)- l2x2
-Xoz
,(xi - x)'
oZ----.-----=-E(Xi - X) '
When MSE is substituted for o2 in (5.88a), we obtain the estimated variance-covariance
matrix o1'b. t ienoted bY s2 {b}:
(s.88)
or, using (5.24a):
(5.88a) 62{b}2x2
s21b1 : MSE(x'x)
f usr _r MSE(X:) - -xusE ^
s:1b1 :MsE(x'x)- ' I : l ' ?) ,1: :*" t (x i -x) '
rvr | -Xusn MSE
L tt",;f ,6=W
fozt-_ln- t
IL
I oas.:+: [ -a.+za
, o2x2* ;1x, -Vy-Xo2
-.---.:;t(xi - x)'
(s.8e)
In (5.88a), you will recognize the variances of bo in (2.22b) and of, b', in (2.3b) and the
covariance tr ao ana a, iln 1+.s1. Likewise, the esiimated variances in (5'89) are familiar
from earlier chaPters.
Example. We wish to find s2{bo} and s2{b, } for the Toluca Company example by matrix
methods. Using the results in Figure 2'2 and in (5'60)' we obtain:
-r . ,o, | - .287415 -.003535 I' : 2'56+ f-.oo:srs .oooo5o5l I
-8.428 I)2o4o )
(5.90)
hlean Response
Section 5.13 Inferences in Regression Analysis 209
Thus, s2{b0} : 685.34 and s2{b,} : .12040. These are the same as the results obtained inChapter 2.
Note
To derive the variance-covariance matrix of b, recall that:
b: (x/x)- ix,Y: AY
where A is a constant matrix:A: (X/X)-rX/
Hence, by (5.46) we have:o2{b} : Ao2{Y}A'
Now o2{Y} : d2L Fufther, it follows from (5.32) and the fact that (X'X)-l is symmerric that:
We find therefore:
A': X(X'X)-r
62 {b} : 1x'xy-tx' o2lx(x'x)- I
: o2(x'X)-rx'x1x'x;-r
: o2(x'x)-1I
: o2(x'x)-1
To estimate the mean response at Xp,let us define the vector:
(s.91)l - r I
Xh: lo I or X'h: i l Xnlzxl L"hJ lx2
The fitted value in matrix notation then is:
(s.e2)
since:
?lr :x 'nb
l- i" Ix;b : Ir xnl l lo | : [bo * b1X1) : [ lr ] : 11,
L"I J
Note that X[b is a I x I matrix; hence, we can write the final result as a scalar
The variance of ?1r, givenearlier in (2.29b), in matrix notation is:
(s.e3) o2 {? nl - o'x'h(x'x) - I xft
The variance of Y7, in (5.93 ) can be expressed as a function of o2 1b1. ttre variance-covariancematrix of the estimated regression coefficients, by making use of the result in (5.88):
(5.93a) o2 It'nl : x'ncz {b}xn
210 Chapter 5 Matrix Approach to Simple Linear Regression Analysis
The estimated variance of Il, given earlier in (2.30), in matrix notation is:
(s.e4) t2 1? o1 : MSE (x/h(x'x)- txr, )
Example. We wish to find s21ir) for the Toluca Company example when X1 : 65. Wedefine:
x; : t l 6sland use the result in (5.90) to obtain:
s21?^1: x i ,s21n1x,
. . . lass:q -8.428 l f r l=rl o ' l l -s.+zs . r2o4o) lo! l : rs ' :z
This is the same result as that obtained in Chaoter 2.
Note
The result in (5.93a) can be derived directly by using (5.46), since in : X,r,b:
Hence:
o21i '01 = 1t
o21?1,| : x ioz{b}Xr,
, ulJ*',ii',, " !;,'Ji'] [ ;, ]of:
(5.e5) o'{ i 'n\ : oz{bo} *2X1,olbo,b) + x|ozlbJ
Using the results from (5.88a), we obtain:
) , : . . o2 o2*2 2X t \ -X)o2 X?.o2- r 'ht
-n >(Xi - X) ' ' >1x, - x ' t2 ,(x i - x)2
which reduces to the familiar expression:
or l r ) : " r l ! * txn-x) 'z]' ln t(xt - x) ' /1
(5.95a)
Thus, we see explicitly that the variance expression in (5.95a) contains contributions fromo21Ul, o21b.,1, and o {bo,b1 }, which it musr according to theorem (A.30b) since ?1, : bo+brX 1,is a linear combination of bo and, br.
Prediction of New Observation
The estimated variance s2{pred}, given earlier in (2.38), in matrix notation is:
L__
(s.96) s2{pred} : MSE(I + X;(X'X)-rXft )
Problems 2ll
Cited Reference
Graybill, F. A. Matices with Applications in statistics. 2nd ed. Belmont, calif.:Wadsworth, 1982.
Problems
For the matrices below, obrain (1) A + B, (2) A _ B, (3) AC, (4) AB,, (5) B,A.
[ r + lA:12 6 |
L3 8J
lz r lo: l ; ; I
14 8_l ":fi]
T"c: l l
L)
[ : s l. : I I 9 |
l ) r I12 4l
I r : ln: l r + l
lz sJr t . lo0l
State the dimension of each resulting matrix.
5.2. For the matrices below, obtain (1) A + C, (2) A - C, (3) B/A, (4) AC,, (5) C,A.
State the dimension of each resulting matrix.
5.3. Show how the following expressions are written in terms of matrices: (1) yi _ ii : ei,(2) t X,ei : 0. Assume i : 1, . . . , 4.
t/ s'+. tr'lavor deterioration. The results shown below were obtained in a small-scale experimenrto study the relation between o F of storage temperature (X) and number of weeks beforeflavor deterioration of a food product begins to occur (y).
i :123
Xi:84Yi: 7.8 9.0
0-4-810.2 I 1.0 l r .1
Assume that first-order regression model (2.1) is applicable. Using matrix methods. find(1) Y'Y, (2) X'X, (3) X/Y.
5'5' Consumer finance. The data below show, for a consumer linance company operatingin six cities, the number of competing loan companies operating in the city (X) and thenumber per thousand of the company's roans made in that city that are currentry derinquent(Y):
r23456
Xi:412334yi : t6 5 10 15 13 22
Assume that flrst-order regression model (2.1) is applicable. using matrix methods, find( l ) Y'Y. (2) X' ,X, (3) X' ,Y.
('lurltrcr 5 Mutrix Approach to Simple Linear Regression Analysis
\i 5.6. Refer to Airfreight breakage problem 1.21. Using matrix merhods, find (l) y/y, (2)x'x, (3) x/Y.
Refer to Plastic hardness Problem 1.22. Using marrix methods, find (1) y,y, (2) x/x,(3) X/Y.
Let B be defined as follows:
a. Are the column vectors of B linearly dependent?
b. What is the rank of B?
c. What must be the determinant of B?
5.9. Let A be defined as follows:[n
-1r - l8 l
A:10 3 l l1055J
a. Are the column vectors of A linearly dependent?
b. Restate definition (5.20) in terms of row vectors. Are the row vectors of A linearlydependent?
c. What is the rank of A?
d. Calculate the determinant of A.
5.10. Find the inverse of each of the followins matrices:
" : [ i ;31Lr 0 5J
^: [3 i ]l+
B: | 6
Lr0
3215 10 Ir6J
5.12.
5.1 3.
5.r4.
Check in each case that the resulting matrix is indeed the inverse.
Find the inverse of the followins matrix:
[s r 3 lA:14 0 5l
Lr e 6)
Check that the resulting matrix is indeed the inverse.
Refer to Flavor deterioration Problem 5.4. Find (XrX;-t.
Refer to Consumer finance Problem 5.5. Find (X,y;-t.
Consider the simultaneous equalions:
i7yr: )5
*3yr:12
4y,
2yr
a. Write these equations in matrix notation.
Problems 213
b. Using matrix methods, find the solutions for y, and yr.
5.15. Consider the simultaneous equations:
5yt i_Zyt:8
23y1 4- 7 lz : 28
a. Write these equations in matrix notation.b. Using matrix methods, find the solutions for y, and yr.
5.16. Consider the estimated linear regression function in the form of (1.15). Write expressionsin this form for the fitted values i in matrix terms for I : 1, . . ., 5.
5.17. Consider the following functions of the random variables yr, yr, and, y.:
Wt:Yr+Y2+Y3
Wz: Yt - Yz
Wz:Yr-Yz-Yz
a. State the above in matrix notation.b. Find the expectation of the random vector W.c. Find the variance-covariance matrix of W.
5.18' consider the following functions of the random variables yr, yz, y3, and r_r:
1Wr:;(Yt+Y2+Y3+Y4l
l lWr:
i (Y, - r Yr) -
, (y3 + y4l
a. State the above in matrix notation.
b. Find the expectation of the random vector W.
c. Find the variance-covariance matrix of W.
5.19. Find the matrix A of the quadratic form:
3Y? +1oyy2+17y15.20. Find the matrix A of the quadratic form:
lYl -8YrY2+8Y:
5.21. For the matrix:[s ) f
A:1" - l
L2 1_l
find the quadratic form of the observations y, and yr.
5.22. For the matrix:10410 3 ol4 0 eJ
; Yr, Yr, and Y,find the quadratic form of the observations
t*_>---*
214 Chapter 5 Matix Approach to Simple Linear Regression Anlalysis
'l S.ZZ. Refer to Flavor deterioration Problems 5.4 and 5.12.
a. Using matrix methods, obtain the following: (1) vector of estimated regression coef-ficients, (2) vector of residuals, (3) s.tR, (4) ssE, (5) estimated variance-covariancematrix of b, (6) point estimate of E lY o] when X1 = -6, (7) estimated variance of iowhen X7, : -6.
b. What simplifications arose from the spacing of the X levels in the experiment?
Find the hat matrix H.
Find s2{ei.
5.24. Refer to Consumer finance Problems 5.5 and 5.13.
using matrix methods, obtain the following: (1) vector of estimated regression coef-ficients, (2) vector of residuals, (3) ssR, (4) ssE, (5) estimated variance-covariancematrix of b, (6) point estimate of E {Y 1,} when X1 : 4, (j) s2 {pred} when X o : 4.
From your estimated variance-covariance matrix in part (a5), obtain the following:(1) s{bn, b,} ; (2) s21br} ; (3) s{b,} .
Find the hat matrix H.
Find s2{e}.
a,
b.
c.
d.
1 5.25. Refer to Airfreight breakage Problems 1.21 and 5.6.
a. Using matrix merhods, obtain the following: (l) (X/D-r, (2) b, (3) e, (4) H, (5) ,SSE,(6) s2{b}, (7) i, when Xn = 2,18; s21io1 when X1, - 2.
b. From parl (a6), obtain the following: (l) 12{br}; (2) s{Do, b,i; (3) s{bo}.
c. Find thc ntarrix ofthe quadratic form for SSR.
5.26. Re fer to Plastic hardness Problems 1.22 and 5.7.
a. Llsing nrarrix nrerhods, obtain the following: (1) (X,X)-1, (2) b, (3) i, <+l H, (5) SSE,(6) s?{b}, (7) s2{pred} when X7, - 30.
Irrorn pi lrr (a6). obrain the fol lowing: ( l) s2{bo}; Q) s{bo, D, }; (3) s{b1}.
Obtain tlrc nr:rtrix ol'the quadratic form for SSE.
b.
Exercises
5.27. Refer to regression-through-the-origin model (4.10). Set up the expectation vector for €.Assumethat i :1, . . . ,4.
5.28. Consider model (4.10) for regression through the origin and the estimaror b, givenin (4.14). Obtain (4.14) by utilizing (5.56) with X suitably defined.
5.29. Consider the least squares estimator b given in (5.56). Using matrix methods, show thatb is an unbiased estimator.
5.30. Show that i, in (5.92) canbe expressed in matrix terms as b/X;,.
5.31. Obtain an expression for the variance-covariance matrix of the fined values f,, I :1. .... n. in terms of the hat matrix.
b-