Transcript AnalysisofVariance

Analysis of Variance: A Difference of
Means Tests for Two or More Levels of
an IV
• An analysis of variance looks for the causal
impact of a nominal level independent variable
(factor) on an interval or better level
dependent variable
• The basic question you seek to answer with an
difference of means test is whether or not
there is a difference in scores on the
dependent variable attributable to membership
in one or the other category of the
independent variable
Types of Difference of Means Tests
•
Varieties of test for difference of means where there is a
single independent variable or factor
•
t-test: two levels of the independent variable
•
Analysis of Variance (ANOVA): Two or more levels or
conditions of the independent variable
•
What is the impact of gender (M, F) on annual salary?
•
What is the impact of ethnicity (Hispanic, African-American, AsianPacific Islander, Caucasian, etc) on annual salary?
ANOVA models can be fixed or random
Fixed model overwhelmingly used
Effects obtained in the fixed model only generalizable to other
identical levels of the factor studied (e.g., only to treatments A, B,
C such as online vs. classroom instruction)
Effects obtained in the random model generalizable to a wider
range of values of the IV than just the three levels
•
•
•
•
•
•
Time of day could be a random factor and you randomly decide to
compare classes taught at 8 am, noon, and 3 pm but these values
are “replaceable” by other randomly drawn values or you could add
more time periods
Subject matter or teacher could be another random factor
Repeated Measures and Analysis of
Covariance
• In a repeated measures ANOVA design,
the same Ss are tested across different
levels of the factor (example, time 1, time
2, time 3, …time n)
• In an analysis of covariance, we
statistically control for the effects of preexisting differences among subjects on the
DV of interest (e.g. controlling for the
effects of an individual’s computer
experience in evaluating impact of presence
or absence of narrative on enjoyment of
computer game play)
More on Tests of Difference of Means:
Analysis of Variance with Two
Independent Variables (Factors)
Two-way ANOVA: two or
more levels of two IVs
or factors
•
•
What is the impact of
diet type and
educational attainment
on pounds lost in six
months, and how do
they interact?
This data suggests two
significant factors that
behave the same way
regardless of the level
of the other factor
(Diet C is always
better, post grad
always better); don’t
interact
Diet A
Diet B
Diet C
High
School
6
8
10
College
10
12
16
Post
Graduate
12
16
20
Average pounds lost as a function
of educational attainment and diet
type
When Factors Interact
In this data set there
seems to be an
interaction between diet
type and educational
attainment, such that
Diet C is more effective
for people with lower
educational attainment,
Diet A works better for
people with high
attainment, and Diet B
works equally well
regardless of educational
attainment. Impact of
one factor depends on
the level of the second
factor
Diet A
Diet B
Diet C
High
School
8
10
12
College
10
10
10
Post
Graduate
12
10
8
Average pounds lost as a function
of educational attainment and diet
type
Single-factor ANOVA Example (one
Independent Variable)
•
Suppose you believed that
interviewer status (a
manipulated variable in which
you systematically varied the
dress of the same interviewer
across the conditions, high
medium, and low would have
an effect on interviewee selfdisclosure, such that the
amount of disclosure of
negative personal information
would vary across conditions.
(The null hypothesis would be
that the interviewees all
came from the same
population of interviewers)
Let’s say you conducted your
study and got the data on the
right, where higher scores
equal more self-disclosure
Interviewer Status
1. High Status
Sums
2. Medium Status
3. Low Status
X11
3
X12
3
X13
4
X21
2
X22
4
X23
5
X31
1
X32
2
X33
3
X41
2
X42
3
X43
4
8
12
16
Means
2
3
4
N
4
4
4
Self-disclosure scores for 12 subjects;
4 subjects in each of three interviewer
conditions
Some Typical Data for ANOVA
The sum over all rows
and columns, denoted
as
∑∑Xij, = 36
i j
Interviewer Status
High Status
(That’s 8 + 12 + 16)
The grand mean,
denoted Xij, is 3
(That’s 2 + 3 + 4 divided
by 3)
The overall N is 12
(That’s 4 subjects in each
of three conditions)
Sums
Medium Status
Low Status
X11
3
X12
3
X13
4
X21
2
X22
4
X23
5
X31
1
X32
2
X33
3
X41
2
X42
3
X43
4
8
12
16
Means
2
3
4
N
4
4
4
Partitioning the Variance for ANOVA: Within
and Between Estimates: how to obtain the
test statistic, F, for the difference of means
•
To obtain the F statistic, we are going to make two
estimates of the common population variance, σ2
•
•
•
The first is called the “within” estimate, which will be a
weighted average of the variances within each of the three
samples. This is an unbiased estimate of σ2 and is an estimate
of how much of the variance in self-disclosure scores is
attributable to more or less random individual differences
The second estimate of the common variance σ2 is called the
“between” (or “among”) estimate and it involves the
variance of the sample means about the grand mean. This is
an estimate of how much of the variation in self-disclosure
scores is attributable to the levels of the factor (interviewer
status). The “between” refers to between-levels variation
If our factor has a meaningful effect the “between estimate”
should be large relative to the “within estimate”; that is, there
should be more variation between the levels of interviewer
status than within them
Meaning of the F Statistic, the
Statistic used in ANOVA
•
•
•
•
•
•
The sampling distribution of the F ratio will be used to determine how
probable it is that our obtained value of F was due to sampling error
The null hypothesis would be that the population means for the three
treatment levels would not differ
If the null hypothesis is false, and the population means are not equal,
then the F ratio will be greater than unity (one). Whether or not the
means are significantly different will depend on how large this ratio is
There is a sampling distribution for F (see p. 479 in Kendrick) called
the “Distribution of the Critical Values of F”; note that there are
separate tables for the .05 and .01 confidence levels). (see also the
next slide)
The columns refer to n1, the DF of the between groups estimate (K-1,
where K is the number of conditions or treatments of the independent
variable) and the rows refer to n2, the DF of the within groups estimate
(N (total) – K)
For our example n1, the between DF, would be 2 and n2, the within DF,
would be 9
Critical values of F
Partitioning the Variation in ANOVA
•
•
The twelve self-disclosure scores we have obtained vary quite a bit
from the grand mean of all the scores, which was 3
The total variation is the sum of the squared deviations from the
grand (overall) mean. This quantity is also called the “total sum
of squares” or the total SS. Its DF is equal to N-1, where N is
the total over all the cases. The total variation has two
components
•
•
•
The within sum of squares: the sum of the squared deviations of the
individual scores from their own category (group) mean. We divide
this by the df (N-K) to obtain the within estimate. This represents
the variability among individuals within the sample
The between (among) sum of squares: this is based on the squared
deviations of the means of the IV levels from the grand mean, and is a
measure of the variability between the conditions. We want this
quantity to be big! We divide the betwenn SS by the df K-1 to get the
between estimate
The within and between estimates are also called the between
and within “mean squares”
A Hand Calculation of ANOVA: Obtaining the
Between and Within Estimates
• To get the between
estimate, the first
thing we calculate is
the between sum of
squares:
• We find the difference
between each group
mean and the grand
mean (3), square this
deviation, multiply by
the number of scores
in the group, and sum
these quantities
Interviewer Status
Sums
High Status
Medium Status
Low Status
X11
3
X12
3
X13
4
X21
2
X22
4
X23
5
X31
1
X32
2
X33
3
X41
2
X42
3
X43
4
8
12
16
Means
2
3
4
N
4
4
4
Between Estimate Calculations
•
•
•
•
•
•
•
So we have
High Status: 2-3 squared
X4=4
Medium Status: 3-3
squared X 4 = 0
Low Status: 4-3 squared X
4=4
So the between sum of
squares = 4+ 0 + 4 = 8
And the between estimate
is obtained by dividing the
between SS by the
between degrees of
freedom, K-1
Thus the between
estimate is 8/2 or 4
Interviewer Status
Sums∑
High Status
Medium Status
Low Status
X11
3
X12
3
X13
4
X21
2
X22
4
X23
5
X31
1
X32
2
X33
3
X41
2
X42
3
X43
4
8
12
16
Means
2
3
4
N
4
4
4
∑∑=36
Calculating the Total Sum of
Squares
•
The next thing we
calculate is the total
sum of squares. This
figure is obtained by
summing the squared
deviations of each of the
individual scores from
the grand mean of 3.
So the total sum of
squares is 3-3 squared
plus 2-3 squared plus 13 squared plus 2-3
squared plus 3-3
squared plus 4-3
squared…. plus 4-3
squared = 14
Interviewer Status
Sums∑
High Status
Medium Status
Low Status
X11
3
X12
3
X13
4
X21
2
X22
4
X23
5
X31
1
X32
2
X33
3
X41
2
X42
3
X43
4
8
12
16
Means
2
3
4
N
4
4
4
Calculating the Within Estimate
•
•
•
Finally, we calculate the within sum of squares. We obtain
that by subtracting the between SS (8) from the total SS
(14). So the within SS = 6. And the within estimate is
obtained by dividing the within SS by its DF, so the within
estimate or within mean square is 6/(N-k) or 6/9 or .667
Recall that for the null hypothesis, that the population
means for the three conditions are equal, to be true, the
between estimate should equal the within estimate, yet our
between estimate is very large in relation to the within
estimate. This is good; it means that the variance
“explained” by the status manipulation is much greater than
what individual differences alone can explain
See the table on the next page which shows the estimates
for the different sources of variation
Basic Output of an ANOVA
Source of
Variation
Sums of
Squares
(SS)
DF
Estimates
Total
14
N-1(11)
Between
8
K-1(2)
4
Within
6
N – K (9)
.667
F
6
Called
“mean
squares”
The between and within estimates are obtained by dividing the between and
within SS by their respective DFs. The F statistic is obtained by dividing
the between estimate by the within estimate (4/.667 = 6)
The obtained value of F tells us that the variation between the conditions is
much greater than the variation within each condition. We look up the F
statistic in the table with 2 DF (conditions minus 1) in the numerator and 9 DF
(total N minus number of conditions) in the denominator and we find that we
need a F of 4.26 to reject the null hypothesis at p < .05. (see next slide)
Looking up the F Value in the Table
of Critical Values of F
With our obtained F of 6 we can reject the null hypothesis
ANOVA in SPSS
•
•
•
•
•
Now let’s try that in SPSS. Go here to download the
data file disclosure.sav and open it in SPSS
In Data Editor go to Analyze/Compare Means/OneWay Anova
Move the Interviewer Status variable into the Factor
window and move the Self-Disclosure variable into the
Dependent List window
Under Options select Descriptive, then press Continue
and then OK
Compare the results in your Output Window to the
hand calculations and to the next slide
SPSS Output, One-Way ANOVA
Descriptives
Self-disclosure
N
Low Status
Medium Status
High Status
Total
4
4
4
12
Mean
4.0000
3.0000
2.0000
3.0000
Std. Deviation
.81650
.81650
.81650
1.12815
Std. Error
.40825
.40825
.40825
.32567
95% Confidence Interval for
Mean
Lower Bound
Upper Bound
2.7008
5.2992
1.7008
4.2992
.7008
3.2992
2.2832
3.7168
Minimum
3.00
2.00
1.00
1.00
Maximum
5.00
4.00
3.00
5.00
ANOVA
Self-disclosure
Between Groups
Within Groups
Total
Sum of
Squares
8.000
6.000
14.000
df
2
9
11
Mean Square
4.000
.667
F
6.000
Sig .
.022
The results of this analysis suggest that interviewer status
has a significant impact on interviewee self-disclosure, F
(2,9) = 6, p < .05 ( or p = .022)
Planned Comparisons vs. Post-hoc
Comparison of Means
•
•
•
Even if we have obtained a significant value of F and the overall
difference of means is significant, the F statistic isn’t telling us
anything about how the mean scores varied among the levels of
the IV. Fortunately, we know that this will be the case in advance,
and so we can plan some comparisons between the pairwise group
means that we will specify in advance. These are called planned
comparisons.
Alternatively, we can compare the means of the groups on a pairwise basis after the fact
Doing comparison-of-means tests after the fact, when we have
had time to check out the means and see what direction they’re
tending (for example, we can look and see that there was more
disclosure to the low-status interviewer than to the high-status
interviewer), it’s not really the done thing to allow a low
confidence level like .10 when we know the direction of the results.
We should use a more conservative alpha region in order to reduce
the risk of Type I error (rejecting a true null hypothesis)
Post-hoc Tests in SPSS
•
•
•
•
•
•
•
In SPSS data editor, make sure you have the
disclosure.sav data file open
Go to Analyze/Compare Means/One-Way Anova
Move Interviewer Status into the Factor box (this is
where the IVs go)
Move Self-disclosure into the Dependent List box
Under Options, select Descriptive, Homogenity of
Variance test, and Means Plot, and click Continue
Under Post Hoc, click Sheffé and Tukey and set the
confidence interval to .05, then click Continue and OK
Compare your output to the next slide
Output for Post-Hoc Comparisons
Test of Homogeneity of Variances
Descriptives
Self-disclosure
Self-disclosure
N
Low Status
Medium Status
High Status
Total
4
4
4
12
Mean
4.0000
3.0000
2.0000
3.0000
Std. Deviation
.81650
.81650
.81650
1.12815
Std. Error
.40825
.40825
.40825
.32567
95% Confidence Interval for
Mean
Lower Bound
Upper Bound
2.7008
5.2992
1.7008
4.2992
.7008
3.2992
2.2832
3.7168
Minimum
3.00
2.00
1.00
1.00
Maximum
5.00
4.00
3.00
5.00
Levene
Statistic
.000
df1
df2
2
9
Sig .
1.000
Variances are equal
Important!
Both Tukey and Sheffé tests show significant differences between high and low
status condition but not between medium status and other two conditions. Tukey
can only be used with groups of equal size. Sheffé critical value (test statistic that
must be exceeded) = k-1 times the critical value of F needed for the one-way anova
at a particular alpha level. If variances are unequal by Levene, use the Tamhane’s
T2 test for post-hoc comparisons
Writing up Your Result
•
To test the hypothesis that interviewer status would have a significant effect
on interviewee self-disclosure, a one-way analysis of variance was
performed. Levene’s test for the equality of variances indicated that the
variances did not differ significantly across levels of the independent variable
(Levene statistic = 000, df = 2, 9, p=1.00). Interviewer status had a
significant main effect on interviewee self-disclosure (F (2,9) = 6, p = .022).
Sheffe’ post-hoc tests indicated that there were significant differences
between mean levels of disclosure for subjects in the high status (M = 2) and
low status (M = 4) conditions (p =.022), suggesting an inverse relationship
between interviewer status and interviewee disclosure. Subjects disclosed
more to the low-status interviewer.
Mean Interviewee Self-Disclosure as a Function of Level of Interviewee Status
High Status
2b*
Medium Status
3ab
Low Status
4a
*Higher scores indicate greater disclosure; means with common subscripts are not significantly different from one
another at p = .022
More SPSS ANOVA
•
•
•
•
•
•
•
Using the general social survey data, let’s test the
hypothesis that one’s father’s highest earned degree has a
significant impact on one’s current socio-economic status
Download the socialsurvey.sav file and open it in Data
Editor
Go to Analyze/Compare Means/One-Way Anova
Move Father’s Highest Degree into the Factor box and move
Respondent Socioeconomic Index into the Dependent List
box
Under Options, select Descriptive and Homogeneity of
Variance test and click Continue
Under Post Hoc select Sheffé and set the significance level
to .05, select Continue and then OK
Compare your output to the next slides
What Will Your Results Section
Say?
Test of Homogeneity of Variances
Respondent Socioeconomic Index
Levene
Statistic
1.784
df1
df2
1148
4
Sig .
.130
ANOVA
Respondent Socioeconomic Index
Between Groups
Within Groups
Total
Sum of
Squares
29791.484
382860.1
412651.5
df
4
1148
1152
Mean Square
7447.871
333.502
F
22.332
Sig .
.000
Using the General Linear Model in
SPSS
•
•
•
•
•
•
•
Now we are going to redo the same analysis but with a few more
bells and whistles. This time, for example, we are going to get
measures of the effect size (impact of the IV, father’s highest
degree) on the DV, respondent’s SES, and we will also get a power
estimate
In the Data Editor, make sure your socialsurvey.sav file is open
Go to Analyze/General Linear Model/Univariate (in the case of
ANOVA, univariate means you only analyze one DV at a time)
Put Father’s Highest Degree into the Fixed Factor box and
Respondent’s SES into the Dependent Variable box
Under Post Hoc, move padeg (shorthand for Father’s Highest
Degree) into the Post Hoc Tests for box and under Equal Variances
assumed select Sheffé (we can do this because we already know
that the variances are not significantly different from our previous
analysis) and click Continue
Click on Options and move padeg into the Display Means for box
Under Display, click on Descriptive Statistics, Estimates of Effect
Size, and Observed Power, and set the significance level to .05.
Click continue and then OK. Compare your result to the next slide
SPSS GLM Output, Univariate
Analysis
= Independent variable
corrected means
that the variance
accounted for by the
intercept has been
removed
Note that we have all the
power required to detect an
effect
Note partial eta squared which is the ratio of the
between-groups SS to the sum of the between groups
SS and the error SS. It describes the amount of
variation in the dependent variable explained by the
independent variable (Father’s highest degree). In this
case the amount of variation accounted for, about 7%,
is not very impressive despite a significant result
SPSS GLM Output, Univariate
Analysis, cont’d
Multiple Comparisons
Father's Highest Degree
Dependent Variable: Respondent Socioeconomic Index
Scheffe
Dependent Variable: Respondent Socioeconomic Index
Father's Highest Degree
LT High School
High School
Junior College
Bachelor
Graduate
Mean
43.143
50.338
52.960
54.818
59.393
Std. Error
.789
.904
3.652
1.703
2.198
95% Confidence Interval
Lower Bound
Upper Bound
41.595
44.690
48.564
52.112
45.794
60.126
51.477
58.159
55.079
63.706
(I) Father' s
Highest Degree
LT High School
High School
Note confidence intervals around the mean
difference estimates. These intervals
should not contain zero (recall that the null
hypothesis is of no differences on the
dependent variable between levels of the
IV) Note also above that some of the
confidence levels around the category
means themselves contain the mean of the
other category. So this sort of data should
be studied as well as significance tests
Junior College
Bachelor
Graduate
(J) Father's
Highest Degree
High School
Junior College
Bachelor
Graduate
LT High School
Junior College
Bachelor
Graduate
LT High School
High School
Bachelor
Graduate
LT High School
High School
Junior College
Graduate
LT High School
High School
Junior College
Bachelor
Mean
Difference
(I-J)
-7.195*
-9.817
-11.676*
-16.250*
7.195*
-2.622
-4.480
-9.055*
9.817
2.622
-1.858
-6.433
11.676*
4.480
1.858
-4.574
16.250*
9.055*
6.433
4.574
Based on observed means.
*. The mean difference is significant at the .05 level.
Std. Error
1.1998
3.7366
1.8768
2.3357
1.1998
3.7626
1.9281
2.3771
3.7366
3.7626
4.0299
4.2630
1.8768
1.9281
4.0299
2.7809
2.3357
2.3771
4.2630
2.7809
Sig .
.000
.142
.000
.000
.000
.975
.249
.006
.142
.975
.995
.685
.000
.249
.995
.608
.000
.006
.685
.608
95% Confidence Interval
Lower Bound
Upper Bound
-10.897
-3.493
-21.346
1.711
-17.466
-5.885
-23.456
-9.044
3.493
10.897
-14.231
8.987
-10.429
1.468
-16.389
-1.721
-1.711
21.346
-8.987
14.231
-14.291
10.575
-19.585
6.720
5.885
17.466
-1.468
10.429
-10.575
14.291
-13.154
4.005
9.044
23.456
1.721
16.389
-6.720
19.585
-4.005
13.154