Chapter

8
Analysis of Variance
 Outline
Introduc)on
Logic of ANOVA
Independent Samples One-way ANOVA
Eect Size and Power: Eta-squared ( 2 ) Omega-squared (
2 ) Cohens f. Power computa)ons
Mul)ple Comparison Procedures
Repeated Measures One-way ANOVA: Basic Ideas, Structural Model; Computa)on; Contrasts; Post Hoc test,
Assump)on of Sphericity
Factorial Analysis of Variance: Two-way Completely Independent ANOVA: Basic Ideas; Main Eect and Interac)on
eect; The Structural Model and Par))oning Sum-of-squares; Computa)onal Formula
The Alterna)ve Factorial ANOVA Designs: 1. Mixed-factorial ANOVA; 2. Completely Repeated Measures Factorial
ANOVA
Mul)variate Analysis of Variances (MANOVA)
 Statistical IV Levels
DV Example
Technique (Factor) of IV

Independent One Two One Experimental and Control

Samples t-test group comparison

Dependent One Two One Pre-test post-test design

Samples t-test
Independent One More One Comparisons of three
Samples One- than treatments (BT, CBT, SPT)
way ANOVA Two for depression
Dependent One More One Comparison of same
Samples One- than participants under three
way ANOVA Two experimental conditions
Factorial Two or Two or One 3 ! 2 factorial design;
ANOVA More more for IV1: Treatment (CBT, BT,
each IV and SPT) and IV2: BPD
(Present and Absent) and
DV is depression
MANOVA One or Two or More Gender difference on four
More more for than measures of Neuroticism
each IV one
 Logic of ANOVA
For k groups, the comparisons are

( k k 1) 2 = ( 4 3) 2 = 12 2 = 6 !
The Type I error for those many (for example, 3)
comparisons are (if we use t-test)
( )
P = 1 (1 )r = 1 (1 .05)3 = 1 .95 3 = 1 .857 = .1426 !
The chance that we reject null when we should have
accepted it is 0.1426 for at least one comparison.
Alterna)ve is ANOVA
 Founda<ons of ANOVA
F-random variable
12 1
F= 2 !
2 2
The F distribu)on is
+ 2 1
1
1
(1 + 2 )
2 1 2 1
1 1 2
f (x) = x 2
1 + x !
1 2 2
2
2 2

S 2
2
S
The F is equal to Fn 1,n 1 = 1 1 = 1 !
2

1 1
S2 2 S2
2 2 2
we have independent random samples
from k populations of size n.
Logic of ANOVA
denote xij as i th value in j th population.
If xij is normally distributed with mean
and variance 2 , then we can state that
xij = j + eij
If xij ~ N ( , 2 ) then
xij = + j + eij
where j = 1,, k; i = 1, 2,, n, = grand mean,
j is a treatment effect of the j th group.
The treatment effect is stated as
k

j =0
j=1

H 0 : j = 0
H 0 : j 0
 Independent Samples One-way
ANOVA
Omnibus null hypothesis regarding equality of popula)on means
for all k-groups.
Par)cipants are independent (o\en randomly assigned).
Assump)ons
Normality: Popula)ons are normally distributed
Homogeneity of variances: Popula)on variances of all k-groups
are equal.
H 0 : 1 = 2 = ... = k !
Independence:
HA : j !
 Basic Equa<ons
k n k k n

(X ij X)2 = n (X j X)2 + (Xij X j )2

j=1 i=1 j=1 j=1 i=1

SST = SSbet + SSwithin

k n
SSwithin = (Xij X j )2
j=1 i=1
k n
SSbet = (Xij X j )2
j=1 i=1

dfwithin = ( n1 1) + ( n2 1) + ...+ ( nk 1) dfBetween = k 1

dfTotal = ( ntotal 1)
 Basic Equa<ons

MSBet = SSbetween dfBetween

MSW = SSWithin dfWithin

MSBet
F=
MSW
Solve Example 8.2 in the book using these computa)ons
formula.
Summary Table for Example 8.2

Source SS df MS F
Between 170.6 3 2 = 2 85.30 6.908

Within 333.4 9 + 9 + 9 =27 12.35

Total 504 30 1 = 29

Use R Code 8.4 and 8.5 for compu)ng ANOVA.

 Eect Size
Eta-squared SSbet
2 = !
SSTotal
MSW
Adjusted = 1
2
!
MST

Omega-squared SS (k 1) MSW
2 = bet !
SSTotal + MSW

Cohens f 2
f= !
1 2
 Power Calcula<ons for One-way
ANOVA

= f n! = /k !
= /k!
2
k=!
2

The is a non-centrality parameter.

Power is computed using NCP.
Sample size can also be determined by using the eect size
f= n Compute power by using R Code
2
! 8.6 and 8.7 in the book.
2 f = n
Mul<ple Comparison Procedures

Error Rate
Per Comparison Type I Error Rate ()
Family-wise Type I Error Rate (fw)
Types of Comparisons
Planned comparisons or a priory comparisons
Post hoc or unplanned comparisons or posteriori
comparisons
Do we need Signicance of F?
 Planned Comparisons

Contrasts
k

a X j j = a1 X1 + a2 X2 + + ak X k = !
j=1

k
Linear Contrasts c j =0!

j=1
k
= cj X j !
j=1
 Planned Comparisons
Mul<ple t-test
Xi X j Xi X j
t= = !
MSW MSW 2MSW
+
n n n

The BonferroniDunn test

Bonferroni Inequality
( )
P Ei P ( Ei )
i =1
i =1
Planned Comparisons
fw c = pc !
fw c pc = c fw c ! ( )
2
F= !
a j MSW
2

n
Evaluate F at pc

Dunnidk Correc<on
 Planned Comparisons
Holms and Larzelere and Mulaik Test
Mul)stage Test
(i) Compute the t-values as is discussed in Bonferroni
Dunn.
(ii) Disregarding the sign, arrange all t-values as per the
ascending order, c is the number of contrasts tested or
comparisons carried out.
(iii) The t is tested for signicance using Dunns cri)cal
c

value at c contrasts. Then t is tested for signicance

c 1

using Dunns cri)cal value at c 1 contrasts, and so on

 Post Hoc Tests

Least Signicant Dierence (LSD)

1
(
LSD = t /2,dfw MSW 1 n + 1 n + + 1 n !
2 k
)
DunneXs Procedure
2MSW
CDD = t D !
n
 Post Hoc Tests
Studen<zed Range Sta<s<cs (q)

X L XS
qr = !
MSW n

Tukeys HSD or WSD Test

MSW
HSD = q !
n
 Post Hoc Tests
TukeyKramer Procedure: HSD for Unequal
Sample MSe MSe
+
n1 n2
!
2
GamesHowell Procedure: Unequal n and
Unequal Variances 2 2
S1 S2
+
n1 n2
CDGH = q.05(r, df ) !
2
 Post Hoc Tests

Sche Test

( k 1)( Fdfb, dfw )

2MSW
CDS = !
n

Newman Keuls Procedure

MSW
CDNK = q(spets,dfw) !
n
 Post Hoc Tests
RyanEinotGabrielWelsch Q (REGWQ) Procedure:
Hochberg GT2
Hochberg H2, Tamhane T2, and Dunnei T3
BenjaminiHochberg Test
WallerDuncan Bayesian Test
Use R Code 8.8 , 8.9, 8.10, 8.11, 8.12, and 8.13 from the
book for mul)ple comparison procedures.
 Repeated Measures One-way
ANOVA
Same par)cipant in dierent condi)ons
Par)cipants are matched across condi)ons
Prac)ce Eect
Counterbalancing
La)n Squares Design
Randomiza)on of Orders
 Structural Model
where Xij = + i + j + eij !
Xij is a score of i th participants in j th treatment
is a grand mean
i is a constant of i th person
j is a constant associated with j th treatment
eij is a error associate with i th participants in j th treatment
i represents the extent to which each persons/subjects score departs
from the mean of person/subject. It is assumed to distributed with mean zero.
j is representative of the difference between treatment mean and
average treatment mean.
eij is distributed around mean zero.
eij and i are iid
 Computa<on
( )!
ntotal
SST = Xij X
2
dfT = ntotal 1 !
i=1

( )
k
SStreat = n X j X
2
! dftreat = k 1 !
j=1

dfsub = nsub 1 !
( )
nsub
SSsub = k X sub X
2
!
i=1

SSerror = SST ( SStreat + SSsub ) ! dferror = dftreat dfsub !

MStreat = SStreat dftreat ! MSerror = SSerror dferror !

MStreat Solve with example data in Table 8.9

F= ! from the book.
MSerror
 Summary Table

Source' SS' df# MS' F# P#

SSsub ' 34.83' dfsub = nsub 1 = 10 1 = 9 ' 3.87' MStreat <.01%
F=
SStreat ' 35.00' dftreat = k 1 = 3 1 = 2 ' 17.50' MSerror '
'
SSerror ' 23.67' dferror = dftreat dfsub = 2 9 = 18 ' 1.315' 17.5 '
= = 13.31
SST ' 93.50' dfT = ntotal 1 = 30 1 = 29 ' ' 1.315 '
%

Use R Code 8.14 from the book for solving the repeated
measures ANOVA.
 Assump<on of Sphericity
Constant variance and constant covariance of
the popula)on variance covariance matrix.
Its is es)mated by sample matrix.
Mauchly's sphericity test.
Use R Code 8.15 from the book to carry out
Mauchly's sphericity test.
 Factorial Analysis of Variance:
Two-way Completely Independent
ANOVA
Factorial Designs: More than one IV
Each IV is at Two or more levels
Three basic types
Completely randomized (All cells independents)
Completely Repeated
Mixed (One IV is Randomized other Repeated)
Main Eect and Interac<on Eect
The main eect in factorial ANOVA is referred to
the impact of individual factor or IV on the
dependent variable.
Interac)on eect is the joint eect of
independent variables.
Impact of one of the factor on DV changes as
levels of the other factor changes.
 Example
IV1: Media Condi)on (2 levels: Audiotape or Videotape)
IV2: Openness to Experience (2 levels: Low-openness and
High-openness)
! ! IV#1:#Media#Condition# !
! ! L1:!Audiotape! L2:!Videotape! !
# L1:!Low1 Low1O!&! Low1O! Low1O!!
IV2:# Openness! Audiotape!! Videotape!! n!=!60!
Openness# n!=!30! n!=!30!
L2:!High1 High1O!&! High1O!&! High1O!!
Openness! Audiotape!! Videotape! n!=!60!
n!=!30! n!=!30!
! ! n!Audiotape!=! n!Videotape!=! Total!n!=!120!
60! 60!
!
Null and Alterna<ve Hypothesis
H 0 : mAudiotape = mVideotape
H A : mAudiotape mVideotape
H 0 : LowO = HighO !
H A : LowO HighO !

H0 : Absence of Interaction Effect

 Structural Model
Xijk = + j + k + ( ) jk + eijk !
Xijk is#any#ith#observation#for#combination#of#(jk)th#cell.##
#is#constant#population#mean##
j is#effect#of#first#factor#(A)#which#is# j #
k is#effect#of#second#factor#(B)#which#is# k #
( ) jk is#interaction#effect#between#first#and#second#factor#(A*B)#
eijk is#error#associated#with#observation# Xijk .#The#error#follows#normal#with#mean#
zero#and#variance# 2 .##
#
 Computa<onal Formula
( )
n
dftotal = ntotal 1
2
SST = Xijk X !
i=1

( )
n n
SSerror = SSWC = Xijk X jk = ( Xi Xcell ) !
2 2
dfWC = (n jk 1)
i=1 i=1

( )!
2
SSIV 1 = n j X j X dfIV 1 = J 1

= n (X X) !
2
SSIV 2 k k dfIV 2 = k 1
SSIV 1IV 2 = SST ( SSIV 1 + SSIV 2 + SSerror ) ! df
IV 1IV 2 = ( j 1) ( k 1)

( )!
2
SScell = n jk X jk X
 Summary Table
Sum!Sq! Df! Mean!Sq! F!value! Pr(>F)!
!
Media! 8.1$ 1$ 8.1$ 10.196$ 0.00292$
Opn! 44.1$ 1$ 44.1$ 55.51$ <$0.0001$
Opn*Media! 3.6$ 1$ 3.6$ 4.531$ 0.04019$
Residuals! 28.6$ 36$ 0.79$
Total! 84.4$ $ $
! $ $ $ $

Use example data in Table 8.14 from the

book.
Use R Code 8.16 from the book.
 Alterna<ve Factorial ANOVA
Designs

Mixed-factorial ANOVA
SSTotal = SSIV 1 + SSwithin + SSIV 2 + SSIV 1IV 2 + SSIV 2within !

Use R Code 8.17 from book.

Completely Repeated Measures, Factorial ANOVA
Use R Code 8.18 from book.
 Mul<variate Analysis of
Variances (MANOVA)
More dependent variables and DVs are
intercorrelated

For example, IV = Gender (Levels: malefemale)

DV: Anxiety (A), Impulsiveness (Im),

Vulnerability (Vu) and Hos)lity (Ho) subscales of
Assump<ons and Other Important
Aspects of MANOVA
Mul)variate Normal Distribu)on (MVN) and Outliers
Homogeneity of VarianceCovariance Matrices
Mul)collinearity and Singularity
Linearity
Homogeneity of Regressions
Reliability of Covariates
Sample Size
Null and Alterna<ve Hypothesis for
MANOVA

Null: All groups have the same vector of popula)on

means.
Male Female Male Female
A A A A

lm lm lm lm !
H0 : = HA :
Vu Vu Vu Vu

Ho Ho Ho Ho

 Test
Sta)s)cal test is Wilks Lambda
S error
= !
S effect + S error
where,&&
S&is&as&Sum&of&Squares&and&Cross&Products&(SSCP)&matrix;&
Serror is#a#SSCP#of#within##
Seffect is#SSCP#of#effect#(or#between).##
Serror is#a#determinant#of# Serror #
Seffect + Serror is#a#determinant#of#addition#of# Seffect + Serror matrices.##
The# will#be#close#to#zero#if# Serror is#small##
The# will#be#close#to#one#if# Serror is#large.##
#
Distribu<on and Signicance
Signicance of the Wilks lambda
1 y df2
Approximate F(df1 ,df2 ) = !
y df1
y= ! ( )
1/s
s = min p,dfeffect !
( )
p dfeffect + 1 p dfeffect 2
(
df1 = p dfeffect ! )
df2 = ( dferror )
2
2
!

2 = ( N 1) 0.5 ( p + k ) ln
df = p(k 1)
Use 8.8.4 small sample worked out example from book.
Usr R Code 8.19 and 8.21.