R EP EATED -M EA S U R ES

A N O VA

O ld and N ew
One-Way ANOVA looks at differences between

different samples exposed to different

manipulations (or different samples that may
come from different groups) within a single factor.
That is, 1 factor, k levels k separate groups are

compared.

NEW: Often, we can answer the same research

questions looking at just 1 sample that is exposed

to different manipulations within a factor.
That is, 1 factor, k levels 1 sample is compared

across k conditions.

M ultiple m easurem ents for

each sam ple.
That is, a single sample is measured on the

same dependent variable once for each

condition (level) within a factor.
Investigate development over time
(Quazi-Independent: time 1, time 2, time 3)
Chart learning (manipulate different levels of

practice)
(Independent: 1 hour, 4 hours, 7 hours)
Compare different priming effects in a LDT
(Independent: Forward, backward, no-prime, non-word)
Simply examine performance under different

conditions with the same individuals.

(Independent: suspect, equal, detective, audio format)

 Comparing two

Ext
ending t-tests
independent
samples?

Independent-samples t-

test!
T-test

Comparing the

dependent measures
of a single sample
under two different
conditions?
Related- (or dependent-

or paired-) sample ttest!

ANOVA cousin
Comparing more than

two independent samples

within a single factor?
One-way ANOVA!

Comparing the

dependent measures of a
single sample under
more than two different
conditions?
Repeated-Measures

ANOVA!

R-M AN O VA
Like the Related-samples t, repeated measures

ANOVA is more powerful because we eliminate

individual differences from the equation.
In a One-way ANOVA, the F-ratio is calculated

using the variance from three sources:

F = Treatment(group) Effect + Individual differences +
Experimenter error/Individual differences +
Experimenter error.
The denominator represents random error and we

do not know how much was from ID and EE.

This is error we expect from chance.

W hy R-M AN O VA is CO O L
With a Repeated-Measure ANOVA, we can

measure and eliminate the variability due

to individual differences!!!!
So, the F ratio is conceptually calculated
using the variance from two sources:
F = Treatment(group) Effect + Experimenter

error/ Experimenter error.

The denominator represents truly random error
that we cannot directly measurethe leftovers.
This is error we expect just from chance.
What does this mean for our F-value?

R-M vs.O ne-W ay:Pow er

All else being equal, Repeated-Measures

ANOVAs will be more powerful because

they have a smaller error term bigger Fs.
Let me demonstrate, if you will.
Assume treatment variance = 10, experimental-

error variance = 1, and individual difference

variance = 1000.
In a One-Way ANOVA, F conceptually = (10 + 1 +
1000)/(1 + 1000) = 1.01
In a Repeated-Measures ANOVA, F conceptually =
(10 + 1)/(1) = 11

W hat is a R-M AN O VA doing?

Again, SStotal = SSbetween + Sswithin
Difference: We break the SSwithin into two

parts:
SSwithin/error = SSsubjects/individual differences + SSerror
To get SSerror we subtract SSsubjects from SSwithin/error.
This time, we truly have SS error , or random

variability.

Other than breaking the SSwithin into two

components and subtracting out SSsubjects,

repeated measures ANOVA is similar to OneWay ANOVA.

Lets learn through exam ple.

A Priming Experiment!
10 participants engage in an LDT,

within which they are exposed to 4

different types of word pairs. RT to
click Word or Non-Word recorded and
averaged for each pair type.
Forward-Prime pairs (e.g., baby-stork)
Backward-Prime pairs (e.g., stork-baby)
Unrelated Pairs (e.g., glass-apple)
Non-Word Pairs (e.g., door-blug)

H ypotheses
For the overall RM ANOVA:
Ho: f = b = u= n
Ha: At least one treatment mean is

different from another.

Specifically:
Ho: f < b
Ho: f and b < u and n
Ho: u < n

The D ata
Par
t

Forwa
rd

Backwa
rd

Unrelat
ed

Nonword

Sum
Part.

.2

.1

.4

.7

1.4

.5

.3

.8

.9

2.5

.4

.3

.6

.8

2.1

F2= 1.36

.4

.2

.8

.9

2.3

B2= .58

.6

.4

.8

.8

2.6

U2= 3.82

.3

.3

.5

.8

1.9

N2= 6.63

.1

.1

.5

.7

1.4

P2= 385

.2

.1

.6

.9

1.8

Xt2=
12.39

.3

.2

.4

.7

1.6

10

.4

.2

.6

.9

2.1

3.4

2.2

8.1

19.7

All = 19.7

Form ula for SS total

2
(
X
TOT
)

X 2TOT

t
NTOT

Ians SS =

Remember, this is the sum of the squared

deviations from the grand mean.

So, SStotal = 12.39 (19.72/40)

= 12.39 9.70225
= 2.68775
Importantly, SStotal = SSwithin/error +

SSbetween

The D ata
Par
t

Forwa
rd

Backwa
rd

Unrelat
ed

Nonword

Sum
Part.

.2

.1

.4

.7

1.4

.5

.3

.8

.9

2.5

.4

.3

.6

.8

2.1

F2= 1.36

.4

.2

.8

.9

2.3

B2= .58

.6

.4

.8

.8

2.6

U2= 3.82

.3

.3

.5

.8

1.9

N2= 6.63

.1

.1

.5

.7

1.4

P2= 385

.2

.1

.6

.9

1.8

Xt2=
12.39

.3

.2

.4

.7

1.6

10

.4

.2

.6

.9

2.1

3.4

2.2

8.1

19.7

All = 19.7

W ithin-G roups Sum of Squares:

SS w ithin/error
SSwithin/error = the sum of each SS with each

group/condition.
Measures variability within each condition, then
adds them together.
( X 1
2

( X 1 ) 2
n1

) ( X 2
2

( X 2 ) 2
n2

) ... ( X k
2

( X k ) 2
nk

So, SSwithin/error =
(1.36 ([3.4]2/10)) + (.58 ([2.2]2/10)) + (3.82

([6]2/10) + (6.63 ([8.1]2/10)

= (1.36 1.156) + (.58 .484) + (3.82 3.6) + (6.63

6.561)
= .204+ .096+ .22+ .069= .589

The D ata
Par
t

Forwa
rd

Backwa
rd

Unrelat
ed

Nonword

Sum
Part.

.2

.1

.4

.7

1.4

.5

.3

.8

.9

2.5

.4

.3

.6

.8

2.1

F2= 1.36

.4

.2

.8

.9

2.3

B2= .58

.6

.4

.8

.8

2.6

U2= 3.82

.3

.3

.5

.8

1.9

N2= 6.63

.1

.1

.5

.7

1.4

P2= 385

.2

.1

.6

.9

1.8

Xt2=
12.39

.3

.2

.4

.7

1.6

10

.4

.2

.6

.9

2.1

3.4

2.2

8.1

19.7

All = 19.7

Breaking up SS w ithin/error
We must find SSSUBJECTS and subtract that

from total within variance to get SSERROR

( p )
( p )
SSSUBJECTS =

TOT

TOT
K is generic for the number
of conditions, as

usual.
SSSUBJECTS = (1.42/4 +2.52/4 +2.12/4 +2.32/4
+2.62/4 +1.92/4 +1.42/4 +1.82/4 +1.62/4 +2.12/4
+) 19.72/40 = .
49+1.5625+1.1025+1.3225+1.69+.9025+.49+.
81+.64+1.1025) -9.70225 = 10.1125-9.70225
=.41025

N ow for SS error
SSerror = SSwithin/error SSsubjects
SSerror = .589 - .41025 = .17875 or .

179
Weeeeeeeeee! We have pure
randomness!

The D ata
Par
t

Forwa
rd

Backwa
rd

Unrelat
ed

Nonword

Sum
Part.

.2

.1

.4

.7

1.4

.5

.3

.8

.9

2.5

.4

.3

.6

.8

2.1

F2= 1.36

.4

.2

.8

.9

2.3

B2= .58

.6

.4

.8

.8

2.6

U2= 3.82

.3

.3

.5

.8

1.9

N2= 6.63

.1

.1

.5

.7

1.4

P2= 385

.2

.1

.6

.9

1.8

Xt2=
12.39

.3

.2

.4

.7

1.6

10

.4

.2

.6

.9

2.1

3.4

2.2

8.1

19.7

All = 19.7

SS betw een-group :
The (same) Formula:
( X )
( X )
( X

...
2

n1

n2

nk

)2

( X TOT ) 2

NTOT

So = SSbetween =
[((3.4)2/10) + ((2.2)2/10) + ((6)2/10) +

((8.1)2/10)] 19.72/40
= (1.156+ .484+ 3.6+ 6.561) 9.70225
= 11.801 9.70225 = 2.09875 or
2.099

G etting Variance from SS

Need? DEGREES OF FREEDOM!
K=4
Ntotal = total number of scores = 40 (4x10)
DfTOTAL = Ntotal 1 = 39
DfBETWEEN/GROUP = k 1 = 3
Dfwithin/error= N K = 40 4 = 36
Dfsubjects = s-1 = 10-1 (where s is the # of
subjects) = 9
Dferror = (k-1)(s-1) = (3)(9) = 27 OR Dfwithin/error Dfsubjects = 36 9 = 27

M ean Squared

(deviations from the m ean)

We want to find the average squared

deviations from the mean for each type of

variability.
To get an average, you divide by n in some
form (or k which is n of groups) and do a little
correction with -1.
That is, you use df.

MSbetween/group =
MSwithin/error =

SS between
df between
=
2.099/3
SS ERROR
=
.179/27
df ERROR

= .7

= .007

H ow do w e interpret these M S
MS error is an estimate of population

variance.
Or, variability due to ___________?

F?
MS
F =MS
= .7/.007 = 105.671
BET

ERROR

(looks like it should be 100, but

there were rounding issues due

to very small numbers.
OK, what is Fcrit? Do we reject the

Null??

Pros and Cons

Advantages
Each participant serves as own control.
Do not need as many participants as one-way. Why? More

power, smaller error term.

Great for investigating trends and changes.

Disadvantages
Practice effects (learning)
Carry-over effects (bias)
Demand characteristics (more exposure, more time to think

about meaning of the experiment).

Control
Counterbalancing
Time (greater spacingbut still have implicit memory).
Cover Stories

Sphericity
Levels of our IV are not independent
same participants are in each level (condition).

Our conditions are dependent, or related.

We want to make sure all conditions are equally

related to one another, or equally dependent.

We look at the variance of the differences

between every pair of conditions, and assume

these variances are the same.
If these variances are equal, we have Sphericity

M ore Sphericity
Testing for Sphericity
Mauchlys test
Significant, no sphericity, NS Sphericity!
If no sphericity, we must engage in a correction of the Fratio. Actually, we alter the degrees of freedom associated
with the F-ratio.

Four types of correction (see book)

Estimate sphericity from 0 (no sphericity) to 1

(sphericity)
Greenhouse-Geiser (1/k-1)
Huynh-Feldt
MANOVA (assumes measures are independent)
non-parametric, rank-based Friedman test (one-factor only)

Sym m etry

Eff
ect Sizes
The issue is not entirely settled. Still

some debate and uncertainty on how

to best measure effect sizes given
the different possible error terms.
2 = See book for equation.

Specifi
c tests
Can use Tukey post-hoc for

exploration
Can use planned comparisons if you
have a priori predictions.
Sphericity not an issue

Contrast Form ula

Same as one way, except error term is different

Contrasts
Some in SPSS:
Difference: Each level of a factor is compared to the mean

of the previous level

Helmert: Each level of a factor is compared to the mean of
the next level
Polynomial: orthogonal polynomial contrasts
Simple: Each level of a factor is compared to the last level

Specific:
GLM forward backward unrelate nonword
/WSFACTOR = prime 4 special (1 1 1 1
-1 -1 1 1
-1 1 0 0
0 0 -1 1)

2+ W ithin Factors
Set up.
Have participants run on tread mill

for 30min.
Within-subject factors:
Factor A
Measure Fatigue every 10min, 3 time
points.
Factor B
Do this once after drinking water, and
again (different day) after drinking new
sports drink.

M uch is the sam e,m uch

diff
erent
We have 2 factors (A and B) and an interaction

between A and B.
These are within-subjects factors
All participants go through all the levels of each factor.

Again, we will want to find SS for the factors and

interaction, and eventually the respective MS as

well.
Again, this will be very similar to a one-way
ANOVA.
Like a 1-factor RM ANOVA, we will also compute
SSsubject so we can find SSerror.

W hat is very diff

erent?
Again, we can parse up SS

w/e

into

SSsubject and SSerror.

NEW: We will do this for each F
we calculate.
For each F, we will calculate:
SS

Effect

; SS

Subject (within that effect)

; and SS

What IS SSError for each effect?

(We will follow the logic of factorial

ANOVAS)

Error

W hat are the SS Errors now ?

Lets start with main effects.
Looking at Factor A, we have
Variability due to Factor A: SSFactor A
Variability due to individual differences.
How do we measure that?
By looking at the variability due to a main effect
of Participant (i.e., Subject): SSSubject (within Factor A)
Variability just due to error.

How do we calculate that!?!?!?

Think about the idea that we actually have 2
factors here, Factor A and Subject.

The Error is in the

IN TERACTIO N S w ith Subject.
For FFactor A (Time)
SSAS is the overall variability looking at

Factor A and Subjects in Factor A (collapsing

across Drink).
To find SSerror for the FFactor A (Time) Calculate:
SSAS; SSFactor A ; SSSubject (within Factor A)
SSerror is: SSAS - (SSFactor A + SSSubject (within Factor A))
That is, SSerror for FFactor A (Time)is SS
A*subj

!!!!!!

Which measures variability within factor A due

to the different participants (i.e., error)

The Sam e for Factor B.

For FFactor B (Drink)
SSAS is the overall variability looking at Factor B

and Subjects in Factor B (collapsing across

Time).
To find SSerror for the FFactor B (Drink) Calculate:
SSBS; SSFactor B ; SSSubject (within Factor B)
SSerror is: SSBS - (SSFactor B + SSSubject (within Factor B))
That is, SSerror for FFactor B (Drink)is SS

!!!!!!
SS B*subj: Which measures variability within factor
B due to the different participants (i.e., error)
B*subj

Sim ilar for AxB Interaction

For FAxB
SSBetween groups is the overall variability due to

Factor A, Factor B, and Subjects.

To find SSerror for the FAxB Calculate:
SSBetween; SSFactor A ; SSFactor B ; SSSubject
SSerror is: SSBetween - (SSFactor A + SSFactor B +
SSSubject)
That is, SSerror for FAxB is SS
SS

A*B*subj

!!!!!!

: measures variability within the AxB

interaction due to the different participants (i.e.,
error)
A*B*subj

W e are fi
nely chopping SS W /E

SS

W/E

SSSub
(A)

SSSub
(B)

SSSub
(AxB)

G etting to F
Factor A (Time)
SSA = 91.2; dfA = (kA 1) = 3 1 = 2
SSError(A*S) = 16.467; dfError(A*S) = (kA 1)(s

1) = (3-1)(10-1) = 18
So, MSA = 91.2/ 2 = 45.6
So, MSError(A*S) = 16.467/ 18 = .915
FA = 45.6/.915 = 49.846

Snapshot ofother Fs
Factor B (Drink)
dfB = (kB 1) = 2 1 = 1
dfError(B*S) = (kB 1)(s 1) = (2-1)(10-1) =

AxB (Time x Drink)

dfAxB = (kA 1)(kB 1) = 2 x 1= 2
dfError(A*B*S) = (kA 1)(kB 1)(s 1) = 2 x 1

x 9 = 18

Use SSs to calculate the respective

MSeffect and MSerror for the other main