Upload
ankush-bhandari
View
253
Download
1
Embed Size (px)
Citation preview
8/6/2019 Chapter 17 Two-way Anova
1/43
Chapter 17: Two-way ANOVA
In the lecture on one-way ANOVA, we discussed the following experiment:
Observe an emergency alone
Observe an emergency w/ 1 other
Observe an emergency w/ 2 others
P + 2 people P + 1
person
P + 0 people
X 2= 9 X 1= 8 X 0= 3
In this experiment, we have ONE IV: the # of bystanders present
What if we wanted to look at other IVs too?
Chapter 17: Page 1
IV = Number of people present
DV = Time it takes (in secs.) to call for help
8/6/2019 Chapter 17 Two-way Anova
2/43
EX: Would the amount of time people wait to help be affected by
the gender of the person in need of help?
Chapter 17: Page 2
8/6/2019 Chapter 17 Two-way Anova
3/43
Lets do an experiment where we stage an emergency situation:
IV1: The victim of the emergency is either male OR female
IV2: The P witnesses the emergency alone, w/ 1 other, OR 2 others
These two IVs are crossed, meaning that each level of one IV ispaired with all levels of the other IV
Put another way, we have all possible combinations of the IVs
P + 2 people P + 1
person
P + 0 people
Male victim X = 9 X = 8 X = 3
Female victim X = 4 X = 4 X = 2
Chapter 17: Page 3
8/6/2019 Chapter 17 Two-way Anova
4/43
Chapter 17: Page 4
8/6/2019 Chapter 17 Two-way Anova
5/43
VOCABULARY
Factor = Independent variable
Two-factor ANOVA / Two-way ANOVA: an experiment with 2
independent variables
Levels: number of treatment conditions (groups) for a specific IVNOTATION
3 X 2 factorial = experiment w/ 2 IVs: one w/ 3 levels, one w/ 2
levels
2 X 2 factorial = experiment w/ 2 IVs: both w/ 2 levels
3 X 2 X 2 = ????
Chapter 17: Page 5
8/6/2019 Chapter 17 Two-way Anova
6/43
Why do a two-factor (two-way) ANOVA?
1. Greatergeneralizability of results
--EX: If experiment is only done with a male victim, we dont
know if the results are also true for female victims
2. Allows one to look forinteractions--The effect of one IV depends on the level of the other IV
--EX: Sample of patients who have an infection:
get antibiotics and are not allergic
dont get antibiotics and are not allergic get antibiotics and are allergic
dont get antibiotics and are allergic
--Measure how well the patients feel the next day
Chapter 17: Page 6
8/6/2019 Chapter 17 Two-way Anova
7/43
3
5
3
1
0
1
2
3
4
5
6
No Yes
Receive Anti-biotic
How
doyoufeelrightnow?
No
Yes
This illustrates an interaction
Chapter 17: Page 7
Allergic?
8/6/2019 Chapter 17 Two-way Anova
8/43
3
5
3
5
0
1
2
3
4
5
6
No Yes
Receive anti-biotic
How
doyoufeelrightnow?
No
Yes
If there were no interaction, then the graph would have looked
like this
Chapter 17: Page 8
Allergic?
8/6/2019 Chapter 17 Two-way Anova
9/43
Lets return to our bystander intervention experiment:
P + 2 people P + 1
person
P + 0 people
Male victim X = 9 X = 8 X = 3
Female victim X = 4 X = 4 X = 2
When we do a two-way ANOVA, we will obtain three different
statistical tests:
1. Main effect of IV1: Gender of Victim
2. Main effect of IV2: # of Bystanders Present
3. Interaction b/n the two IVs (gender & # of bystanders)
Chapter 17: Page 9
8/6/2019 Chapter 17 Two-way Anova
10/43
Each is a hypothesis test:
Gender Main Effect:
H0: all levels of gender have the same mean
H1: all levels of gender do not have the same mean
Bystander Main Effect:H0: all levels of bystander have the same mean
H1: all levels of bystander do not have the same mean
Interaction:
H0: there is no interaction between the factorsH1: there is an interaction between the factors
Chapter 17: Page 10
8/6/2019 Chapter 17 Two-way Anova
11/43
Main Effects
Defined: The effect of ONE IV on the DV averaged across the levels of theother IV
In our example:
--Main effect of gender: Is there a difference in response time for male
versus female victims, averaging over the number of bystanders present?
That is: Ignoring the number of bystanders, does response time differ for
male versus female victims?
--Main effect of # of bystanders: Is there a difference in response time when
there is 1 versus 2 versus 3 bystanders present, averaging over the victimsgender?
That is: Ignoring the gender of the victim, does response time differ based
on the number of bystanders present?
Chapter 17: Page 11
8/6/2019 Chapter 17 Two-way Anova
12/43
One way to understand main effects is to examine something
called marginal means
P + 2 people P + 1 person P + 0 people Marginal
Means
Male victim
X = 9 X = 8 X = 3 X = 6.67Female victim
X = 4 X = 4 X = 2 X = 3.33
Marginal MeansX = 6.5 X = 6 X = 2.5
Chapter 17: Page 12
Response time
for 3bystanders,
averaging over
gender
(9 + 4) / 2 = 6.5
Response time
for 2bystanders,
averaging over
gender
(8 + 4) / 2 = 6
Response timefor 1 bystander,
averaging over
gender
(3 + 2) / 2 = 2.5
Response time for
male victims,
averaging over # of
bystanders
(9 + 8 + 3) / 3 =
6.67
Response time for
female victims,
averaging over # of
bystanders
(4 + 4 + 2) / 3 =
3.33
8/6/2019 Chapter 17 Two-way Anova
13/43
A main effect of gender asks:
Do the marginal means of 6.67 (male victims) and 3.33 (female
victims) differ?
A main effect of # of bystanders asks:
Do the marginal means of 6.5 (3 bystanders), 6 (2 bystanders), &
2.5 (1 bystander) differ?
Chapter 17: Page 13
8/6/2019 Chapter 17 Two-way Anova
14/43
Calculations
To perform all the calculations for Sums of Squares for a two-way
anova by hand is time and labor intensive
These are almost exclusively done using the aid of a statistical
package, like SPSS
Thus, Ill just explain the calculations conceptually
Chapter 17: Page 14
8/6/2019 Chapter 17 Two-way Anova
15/43
Calculations: Main Effects Sums of Squares
Calculations for main effects SS in a two-way ANOVA are very similar to thecalculations we used in one-way ANOVA
Conceptually, to calculate the SS for a main effect, one is comparing each
marginal mean to the overall (grand) mean
P + 2 people P + 1 person P + 0 people Marginal Means
Male victimX = 9 X = 8 X = 3 X = 6.67
Female victimX = 4 X = 4 X = 2 X = 3.33
Marginal MeansX = 6.5 X = 6 X = 2.5
Overall: 5
For a main effect of # of bystanders, one is taking the squared difference b/n:
The mean from 3 bystanders & the overall mean (6.5 5)The mean from 2 bystanders & the overall mean (6-5)
The mean from 1 bystander & the overall mean (2.5 5)
These squared differences are multiplied by the # of scores per mean
Chapter 17: Page 15
8/6/2019 Chapter 17 Two-way Anova
16/43
For a main effect of gender, one is taking the squared difference b/n:
The mean from male victims & the overall mean (6.67 5)The mean from female victims & the overall mean (3.33-5)
These squared differences are multiplied by the # of scores per mean
Calculations: Sums of SquaresTotal
The Total SS is calculated by summing the squared deviations
between each score and the overall grand mean
Chapter 17: Page 16
8/6/2019 Chapter 17 Two-way Anova
17/43
Calculations: Sums of SquaresCells
The SS Cells is calculated by summing the squared deviations
between each cell mean and the overall grand mean
Each deviation is weighted by the number of observations in that
cell
Why would we want this?
Cell means could differ b/c:
1. They come from different levels of gender
2. They come from different levels of # of bystanders
3. There is an interaction b/n gender & # of bystanders
Chapter 17: Page 17
8/6/2019 Chapter 17 Two-way Anova
18/43
SS cells is made up of 3 parts:
SS gender
SS # of bystanders
SS interaction
We can calculate SS gender & SS # of bystanders directly
Then, to find SS interaction:
SS interaction = SS cells SS gender SS # of bystanders
Chapter 17: Page 18
8/6/2019 Chapter 17 Two-way Anova
19/43
Calculations: Sums of SquaresError
The SS error is whats left over
Of the total SS, we know what is due to gender (SS gender), what
is due to the # of bystanders (SS bystanders) and what is due
to the interaction (SS interaction). Thus:
SSerror= SStotal (SSgender + SSbystanders + SSinteraction)
Chapter 17: Page 19
8/6/2019 Chapter 17 Two-way Anova
20/43
ANOVA Table
Source df SS MS FGender
# of Bystanders
Interaction (G*B)
ErrorTotal
Chapter 17: Page 20
8/6/2019 Chapter 17 Two-way Anova
21/43
Calculations: Degrees of Freedom
dftotal = N 1
(where N is the total sample size of the entire experiment)
dfgender = k 1 (where k is the # of gender levels)
dfbystanders = k 1 (where k is the # of bystander levels)
dfinteraction = dfgender * dfbystanders
(Product of the df for the two IVs)
dferror= dftotal - dfgender - dfbystanders - dfinteraction
Chapter 17: Page 21
8/6/2019 Chapter 17 Two-way Anova
22/43
Calculations: Mean Squares
To find any MS, take the SS & divide by its corresponding df
MSGender = GenderGender
df
SS
MSBystanders = dersBysdersBys
df
SS
tan
tan
MSInteraction = nInteractionInteractio
df
SS
MSError =Error
Error
df
SS
Chapter 17: Page 22
8/6/2019 Chapter 17 Two-way Anova
23/43
Calculations: F statistics
To find the 3 F statistics for our tests, take the MS for the 2 IVs &the MS for the interaction & divide them by the MSError
FGender =Error
Gender
MS
MS
FBystanders =Error
dersBys
MS
MStan
FInteraction= ErrornInteractio
MS
MS
Chapter 17: Page 23
8/6/2019 Chapter 17 Two-way Anova
24/43
Calculations: Critical Values
To determine if a particular F statistic is statistically significant,
you obtain the appropriate critical value from Table E.3 or E.4
For each F, you have two df: one for the corresponding factor and
the dfErrorAs usual, if the obtained F value equals or exceeds the critical
value, then we reject the null hypothesis and conclude that we
have a statistically significant effect
Chapter 17: Page 24
8/6/2019 Chapter 17 Two-way Anova
25/43
Tests of Between-Subjects Effects
Dependent Variable: How long do P wait to help?
200.000a 5 40.000 17.778 .000
750.000 1 750.000 333.333 .000
83.333 1 83.333 37.037 .000
95.000 2 47.500 21.111 .000
21.667 2 10.833 4.815 .017
54.000 24 2.250
1004.000 30254.000 29
SourceCorrected Model
Intercept
GENDER
BYSTAND
GENDER * BYSTAND
Error
TotalCorrected Total
Type III Sumof Squares df Mean Square F Sig.
R Squared = .787 (Adjusted R Squared = .743)a.
Interpreting SPSS Output
Chapter 17: Page 25
Between-Subjects Factors
15
15
10
10
10
f
m
Gender of
Victim
1.00
2.00
3.00
# of bystanders
present
N
Main
effect of
gender
Main effect
of
bystanders
Interaction
between
gender &
bystanders
8/6/2019 Chapter 17 Two-way Anova
26/43
Estimated Marginal Means
. Gender of Victim1
Dependent Variable: How long do P wait to help?
.3333 .387 .2534 .4133
.6667 .387 .5867 .7466
Gender of Victim
f
m
Mean Std. Error Lower Bound Upper Bound
% Confidence Interval95
. # of bystanders present2
Dependent Variable: How long do P wait to help?
.2500 .474 .1521 .3479
.6000 .474 .5021 .6979
.6500 .474 .5521 .7479
# of bystanders present
.100
.200
.300
Mean Std. Error Lower Bound Upper Bound
% Confidence Interval95
Chapter 17: Page 26
There was a
main effect
of gender, sothese means
differ
There was a
main effect
of bystander,
so there is a
difference
among these
means
8/6/2019 Chapter 17 Two-way Anova
27/43
. Gender of Victim * # of bystanders present3
Dependent Variable: How long do P wait to help?
.2000 .671 .615 .3385
.4000 .671 .2615 .5385
.4000 .671 .2615 .5385
.3000 .671 .1615 .4385
.8000 .671 .6615 .9385
.9000 .671 .7615 .10385
# of bystanders present
.100
.200
.300
.100
.200
.300
Gender of Victim
f
m
Mean Std. Error Lower Bound Upper Bound
% Confidence Interval95
Chapter 17: Page 27
There was a significant
interaction, so the
effect of bystander
depends on the level of
gender
8/6/2019 Chapter 17 Two-way Anova
28/43
Multiple-comparison procedures: Post-hoc tests VS simple effects
Understanding Main Effects
If there is a significant main effect, & that factor has only two levels, you knowthat those two marginal means differ significantly from each other
If there is a significant main effect, & that factor has 3 or more levels, youknow that at least two of those marginal means differ. Need multiple-comparison procedures (post-hocs) to determine which ones.
Can be obtained in SPSS
Chapter 17: Page 28
8/6/2019 Chapter 17 Two-way Anova
29/43
Post Hoc Tests: # of bystanders presentMultiple Comparisons
Dependent Variable: How long do P wait to help?
LSD
- .35000* .6708 .000 - .48845 - .21155
- .40000* .6708 .000 - .53845 - .26155
.35000* .6708 .000 .21155 .48845
-.5000 .6708 .463 - .18845 .8845.40000* .6708 .000 .26155 .53845
.5000 .6708 .463 -.8845 .18845
(J) # of bystanders
present
.200
.300
.100
.300
.100
.200
(I) # of bystanders
present
.100
.200
.300
Mean
Difference
(I-J) Std. Error Sig. Lower Bound Upper Bound
% Confidence Interval95
Based on observed means.
The mean difference is significant at the . level.05*.
Chapter 17: Page 29
1 bystander
differs from2
1 bystander
differs from
3
2 bystanders
DO NOT
differ from 3
8/6/2019 Chapter 17 Two-way Anova
30/43
Multiple-comparison procedures: Post-hoc tests VS simple effects
Understanding Interactions
An interaction occurs when the effect of one IV depends on the level of the
other IV
If you obtain a significant interaction, you may want to examine it closely to seewhat is causing it
A useful first step is to graph the means to see the pattern of the interaction
Can be obtained in SPSS or done by hand
Chapter 17: Page 30
8/6/2019 Chapter 17 Two-way Anova
31/43
You could graph the means two different ways:
Estimated Marginal Means of How long do P wait to help?
Gender of Victim
mf
EstimatedMargin
alMeans
10
8
6
4
2
0
# of bystanders pres
.100
.200
.300
Graphed this way, we see that the bystander effect seems smaller for female
victims compared to male victims
Chapter 17: Page 31
8/6/2019 Chapter 17 Two-way Anova
32/43
Estimated Marginal Means of How long do P wait to hel
# of bystanders present
.300.200.100
EstimatedMarginalMeans
10
8
6
4
2
0
Gender of Victim
f
m
Graphed this way, we see that male and female victims get helped about equally
quickly with one bystander present, but when multiple bystanders arepresent, a gender difference emerges
Chapter 17: Page 32
8/6/2019 Chapter 17 Two-way Anova
33/43
Understanding Interactions
Another thing to do to understand an interaction is to calculate simple effects
The effect of one IV at one level of another IV
Can be done in SPSS using Syntax
Chapter 17: Page 33
8/6/2019 Chapter 17 Two-way Anova
34/43
Simple effects of gender at each level of bystander:
Univariate Tests
Dependent Variable: How long do P wait to help?
.2500 1 .2500 .1111 .302
.54000 24 .2250
.40000 1 .40000 .17778 .000
.54000 24 .2250
.62500 1 .62500 .27778 .000
.54000 24 .2250
Contrast
Error
Contrast
Error
Contrast
Error
# of bystanders present
.100
.200
.300
Sum of
Squares df Mean Square F Sig.
Each F tests the simple effects of Gender of Victim within each level combination of the other effects
shown. These tests are based on the linearly independent pairwise comparisons among the estimated
marginal means.
Chapter 17: Page 34
There is a difference
between male and femalevictims in response time
when 2 or 3 bystanders
are present
There is a difference
between male and femalevictims in response time
when 2 or 3 bystanders
are present
There is no difference in
response time for male and
female victims when only
1 bystander is present
8/6/2019 Chapter 17 Two-way Anova
35/43
Simple effects of bystander at each level of gender:
Univariate Tests
Dependent Variable: How long do P wait to help?
.13333 2 .6667 .2963 .071
.54000 24 .2250
.103333 2 .51667 .22963 .000
.54000 24 .2250
Contrast
Error
Contrast
Error
Gender of Victim
f
m
Sum of
Squares df Mean Square F Sig.
Each F tests the simple effects of # of bystanders present within each level combination of theother effects shown. These tests are based on the linearly independent pairwise comparisons
among the estimated marginal means.
Chapter 17: Page 35
There is a difference amongthe levels of bystander for
male victims. There is no
such difference for female
victims
There is a difference amongthe levels of bystander for
male victims. There is no
such difference for female
victims
8/6/2019 Chapter 17 Two-way Anova
36/43
Interpretation
Reporting the results of a two-way anova is complex because there are so manytests conducted. Below is one way you might report the results of the
above analyses:
Chapter 17: Page 36
8/6/2019 Chapter 17 Two-way Anova
37/43
An experiment was conducted to determine if the number of bystanders present
in an emergency situation and the gender of the victim in an emergency
situation affect the time it takes a person to help. A two-way ANOVAfound a main effect of gender,F(1,24)=37.037,p .05, indicating thatfemale victims are helped sooner than male victims. There was also a
main effect of the number of bystanders present,F(2,24)=21.111,p .05.This effect showed that a lone bystander helped much sooner than when
there were 2 or 3 bystanders present. Finally, there was an interaction
between gender and the number of bystanders present,F(2,24)=4.815,p .05. Simple effects tests showed that male and female victims receive
help equally quickly when only 1 bystander is present,F(1,24)=1.11,p> .05. However, when 2 or 3 bystanders are present, female victims were
helped more quickly than male victims,F(1,24)=17.778,p .05 for 2bystanders andF(1,24)=27.778,p .05 for 3 bystanders.
Chapter 17: Page 37
8/6/2019 Chapter 17 Two-way Anova
38/43
Filling in an ANOVA Table
You should be able to complete a partially filled in ANOVA table
Lets suppose we did an experiment where we investigate the effectiveness of
advertisements. We manipulate:
IV1: Whether the ad has a celebrity spokesperson or a normal person
IV2: Whether the ad is in black/white or color
DV: How persuasive do Ps find the ad on a scale where 1 = Not at all persuaded;
7 = Extremely persuaded
There are 20 participants per cell
Chapter 17: Page 38
8/6/2019 Chapter 17 Two-way Anova
39/43
ANOVA Table
Source df SS MS FSpokesperson 200
Color 150
Interaction (S*C) 50
ErrorTotal 850
We should be able to complete everything else!
Calculate df
Spokesperson has 2 levels df = (k 1) = (2 1) = 1Color has 2 levels df = (k 1) = (2 1) = 1
Interaction df is the product of the df for the 2 IVs = 1*1 = 1
Total df = (N 1) = (80-1) = 79
Error df = whats left over (79 1 1 1) = 76
Chapter 17: Page 39
8/6/2019 Chapter 17 Two-way Anova
40/43
Source df SS MS FSpokesperson 1 200
Color 1 150
Interaction (S*C) 1 50
Error 76
Total 79 850
The SS for all factors and interactions should add up to the totalSS. Thus:
SSerror= SStotal SSspokesperson SScolor- SSinteraction
SSerror= 850 200 150 50 = 450
Chapter 17: Page 40
8/6/2019 Chapter 17 Two-way Anova
41/43
Source df SS MS FSpokesperson 1 200
Color 1 150
Interaction (S*C) 1 50
Error 76 450
Total 79 850
MSSpokesperson = 2001200
==
onspokespers
onspokespers
df
SSMScolor= 1501
150==
color
color
dfSS
MSInteraction = 50150
==
nInteractio
nInteractio
df
SS
MSError = 92.576450
==
Error
Error
df
SS
Chapter 17: Page 41
8/6/2019 Chapter 17 Two-way Anova
42/43
Source df SS MS FSpokesperson 1 200 200
Color 1 150 150
Interaction (S*C) 1 50 50
Error 76 450 5.92
Total 79 850
FSpokesperson = 7.3392.5200
==
Error
onSpokespers
MS
MSFColor = 34.2592.5
150==Error
Color
MSMS
FInteraction = 45.892.550
==
Error
nInteractio
MS
MS
Chapter 17: Page 42
8/6/2019 Chapter 17 Two-way Anova
43/43
Source df SS MS FSpokesperson 1 200 200 33.78
Color 1 150 150 25.34
Interaction (S*C) 1 50 50 8.45
Error 76 450 5.92
Total 79 850
Finally, wed compare each F to the appropriate critical F value
All tests in this example have 1,76 df. Assuming = .05, from
Table E.3, well use the df that are closest to but smaller than
the actual df, b/c they are not listed
Thus, we will use the df for (1,60) = 4.00
All Fs exceed the critical value
Chapter 17: Page 43