Sociology 5811:
Lecture 13: ANOVA
Copyright © 2005 by Evan Schofer
Do not copy or distribute without
permission
Announcements
• Midterm in one week
• Bring a Calculator
• Bring Pencil/Eraser
• Mid-term Review Sheet handed out today
• New topic today: ANOVA
5811 Midterm Exam
• Exam Topics:
• All class material and readings up through
ANOVA
• Emphasis: conceptual understanding, interpretation
• Memorization of complex formulas not required
• I will provide a “formula sheet”… But, formulas won’t be
labeled!
• Exam Format:
• Mix of short-answer and longer questions
• Mix of math problems and conceptual questions.
Review: Mean Difference Tests
• For any two means, the difference will also fall in
a certain range. Example:
• Group 1 means range from 6.0 to 8.0
• Group 2 means range from 1.0 to 2.0
• Difference in means will range from 4.0 to 7.0
• If it is improbable that the sampling distribution
overlaps with zero, then the population means
probably differ
• A corollary of the C.L.T provides formulas to
estimate standard error of difference in means
Z-tests and T-tests
• If N = large, we can do a Z-test:
Z(Y1 Y2 ) 
Y1  Y2
s N1  s N 2
2
1
2
2
• This Z-score for differences in means indicates:
How far the difference in means falls from zero
(measured in “standard errors”)
– If Z is large, we typically reject the null hypothesis…
group means probably differ.
Z-tests and T-tests
• If N = small, but samples are normal, with equal
variance, we can do a t-test:
t(N1  N 2 2 )
(Y1  Y2 )

1
1
s(Y1 Y2 )

N1 N 2
• Small N requires a different formula to determine
the standard error of difference in means
• Again: Large t = reject null hypothesis
T-Test for Mean Difference
• Question: What if you wanted to compare 3 or
more groups, instead of just two?
• Example: Test scores for students in different educational
tracks: honors, regular, remedial
• Can you use T-tests for 3+ groups?
• Answer: Sort of… You can do a T-test for every
combination of groups
• e.g., honors & reg, honors & remedial, reg & remedial
• But, the possibility of a Type I error
proliferates… 5% for each test.
• With 5 groups, chance of error reaches 50%.
ANOVA
• ANOVA = “ANalysis Of VAriance”
• “Oneway ANOVA” : The simplest form
• ANOVA lets us test to see if any group mean
differs from the mean of all groups combined
• Answers: “Are all groups equal or not?”
• H0: All groups have the same population mean
• m1 = m2 = m3 = m4
• H1: One or more groups differ
• But, doesn’t distinguish which specific group(s) differ
• Maybe only m2 differs, or maybe all differ.
ANOVA and T-Tests
• ANOVA and T-Test are similar
• Many sociological research problems can be addressed by
either of them
• But, they rely on very different mathematical approaches
• If you want to compare two groups, both work
• If there are many groups, people usually use
ANOVA
• Also, there are more advanced forms of ANOVA that are
very useful.
ANOVA: Example
• Suppose you suspect that a firm is engaging in
wage discrimination based on ethnicity
• Certain groups might be getting paid more or less…
• The company counters: “We pay entry-level
workers all about the same amount of money.
No group gets preferential treatment.”
• Given data on a sample of employees, ANOVA
lets you test this hypothesis.
• Are observed group differences just due to chance?
• Or do they reflect differences in the underlying population?
(i.e., the whole company)
ANOVA: Example
• The company has workers of three ethnic groups:
• Whites, African-Americans, Asian-Americans
• You observe:
• Y-barWhite = $8.78 / hour
• Y-barAfAm = $8.52 / hour
• Y-barAsianAm = $8.91 / hour
• Even if all groups had the same population mean
(mWhite = mAfAm = mAsianAm), samples differ randomly
• Question: Are the observed differences so large it
is unlikely that they are due to random error?
• Thus, it is unlikely that: mWhite = mAfAm = mAsianAm
ANOVA: Concepts & Definitions
• The grand mean is the mean of all groups
• ex: mean of all entry-level workers = $8.75/hour
• The group mean is the mean of a particular subgroup of the population
• As usual, we hope to make inferences about
population grand and group means, even though
we only have samples and observed grand and
group means
• We know Y-bar, Y-barWhite, Y-barAfAm,Y-barAsianAm
• We want to infer about: m, mWhite, mAfAm , mAsianAm
ANOVA: Concepts & Definitions
• Hourly wage is the dependent variable
• We are looking to see if wage “depends” upon the particular
group a person is in
• The effect of a group is the difference between
that group’s mean from the grand mean
• Effect is denoted by alpha (a)
• If m  $8.75, mWhite = $8.90, then aWhite= $0.15
• Effect of being in group j is:
αj  μ j μ
• Calculated for samples as: α  Y
j
j
• It is like a deviation, but for a group
Y
ANOVA: Concepts & Definitions
• ANOVA is based on partitioning deviation
• We initially calculated deviation as the distance
of a point from the grand mean: d i  Yi  Y
• But, you can also think of deviation from a group
mean (called “e”): ei ,White  Yi ,White  YWhite
• Or, for any case i in group j:
eij  Yij  Y j
• Thus, the deviation (from group mean) of the 27th
person in group 4 is: e
 Y Y
27, 4
27, 4
4
ANOVA: Concepts & Definitions
• The location of any case is determined by:
• The Grand Mean, m, common to all cases
• The group “effect” a, common to group members
• The distance between a group and the grand mean
• The within-group deviation (e): called “error”
• The distance from group mean to an case’s value
The ANOVA Model
• This is the basis for a formal model:
• For any population with mean m
• Comprised of J subgroups, Nj in each group
• Each with a group effect a
• The location of any individual can be expressed
as follows:
Yij  μ  α j  eij
• Yij refers to the value of case i in group j
• eij refers to the “error” (i.e., deviation from group
mean) for case i in group j
Sum of Squared Deviation
• We are most interested in two parts of the model:
• The group effects:
aj
• Deviation of the group from the grand mean
• Individual case error:
eij
• Deviation of the individual from the group mean
• Each are deviations that can be “summed up”
• Remember, we square deviations when summing
• Otherwise, they add up to zero
• Remember variance is just squared deviation.
Sum of Squared Deviation
• The total deviation can partitioned into aj and eij
components:
• That is, aj + eij = total deviation:
α j  Yj  Y
eij  Yij  Yj
eij  α j  (Yj  Y)  (Yij  Yj )  Yij  Y
Sum of Squared Deviation
• The total deviation can partitioned into aj and eij
components:
• The total variance (SStotal) is made up of:
–
–
–
aj : between group variance (SSbetween)
eij : within group variance (SSwithin)
SStotal = SSbetween + SSwithin
Sum of Squared Deviation
• Given a sample with J sub-groups:
Grand Mean  Y
Group Means  Y1 , Y2 , Y3 ,..., Y j
• Formula for the squared deviation can be rewritten as follows:
N
J
nj
d   (Yi  Y )   (Yij  Y )
2
2
i 1
2
j 1 i 1
• This is called the “Total Sum of Squares” (SStotal)
Sum of Squared Deviation
• The between group (a) variance is the distance
from the grand mean to each group mean
(summed for all cases):
J
SSBetween   n j (Y j  Y )
2
j 1
• The within group variance (e) is the distance from
each case to its group mean (summed):
J
nj
SSWithin   (Yij  Y j )
j 1 i 1
2
Sum of Squared Variance
• The sum of squares grows as N gets larger.
• To derive a more comparable measure, we “average” it, just
as with the variance: i.e, (divide) by N-1
• It is desirable, for similar reasons, to “average”
the Sum of Squares between/within
• Result the “Mean Square” variance
– MSbetween and MSwithin
Sum of Squared Variance
• Choosing relevant denominators we get:
J
MSBetween 
 n (Y
j 1
j
Y )
J 1
J
MSWithin 
j
2
nj
 (Y
j 1 i 1
ij
 Yj )
NJ
2
Mean Squares and Group Differences
• Question: Which suggests that group means are
quite different?
– MSbetween > MSwithin or MSbetween < MSwithin
Mean Squares and Group Differences
MSbetween > MSwithin:
MSbetween < MSwithin:
Mean Squares and Group Differences
• Question: Which suggests that group means are
quite different:
– MSbetween > MSwithin or MSbetween < MSwithin
• Answer: If between group variance is greater
than within, the groups are quite distinct
• It is unlikely that they came from a population with the
same mean
• But, if within is greater than between, the groups
aren’t very different – they overlap a lot
• It is plausible that m1 = m2 = m3 = m4
The F Ratio
• The ratio of MSbetween to MSwithin is referred
to as the F ratio:
FJ 1, N  J
MS Between

MSW ithin
• If MSbetween > MSwithin then F > 1
• If MSbetween < MSwithin then F < 1
• Higher F indicates that groups are more separate
The F Ratio
• The F ratio has a sampling distribution
• That is, estimates of F vary depending on exactly which
sample you draw
• Again, this sampling distribution has known
properties that can be looked up in a table
• The “F-distribution”
– Different from z & t!
• Statisticians have determined how much area falls under the
curve for a given value of F…
• So, we can test hypotheses.
The F Ratio
• Assumptions required for hypothesis testing
using an F-statistic
• 1. J groups are drawn from a normally
distributed population
• 2. Population variances of groups are equal
• If these assumptions hold, the F statistic can be
looked up in an F-distribution table
• Much like T distributions
– But, there are 2 degrees of freedom: J-1 and N-J
• One for number of groups, one for N
The F Ratio
• Example: Looking for wage discrimination
within a firm
• The company has workers of three ethnic groups:
• Whites, African-Americans, Asian-Americans
• You observe in a sample of 200 employees:
• Y-barWhite = $8.78 / hour
• Y-barAfAm = $8.52 / hour
• Y-barAsianAm = $8.91 / hour
The F Ratio
• Suppose you calculate the following from your
sample:
• F = 6.24
• Recall that N = 200, J = 3
• Degrees of Freedom: J-1 = 2, N-J = 197
• If a = .05, the critical F value for 2, 197 is 3.00
• See Knoke, p. 514
• The observed F easily exceeds the critical value
• Thus, we can reject H0; we can conclude that the groups do
not all have the same population mean
Comparison with T-Test
• T-test strategy: Determine the sampling
distribution of the mean…
• Use that info to assess probability that groups have same
mean (difference in means = 0)
• ANOVA strategy
• Compute F-ratio, which indicates what kind of deviation is
larger: “between” vs. “within” group
• High F-value indicates groups are separate
• Note: For two groups, ANOVA and T-test
produce identical results.
Bivariate Analyses
• Up until now, we have focused on a single
variable: Y
• Even in T-test for difference in means & ANOVA, we just
talked about Y – but for multiple groups…
• Alternately, we can think of these as simple
bivariate analyses
• Where group type is a “variable”
– Ex: Seeing if girls differ from boys on a test
• … is equivalent to examining whether gender (a first
variable) affects test score (a second variable).
2 Groups = Bivariate Analysis
Group 1: Boys
Case
Score
1
57
2
64
3
48
Group 2: Girls
Case
1
2
3
Score
53
87
73
Case
1
2
3
4
5
6
Gender
0
0
0
1
1
1
Score
57
64
48
53
87
73
2 Groups = Bivariate
analysis of Gender
and Test Score
T-test, ANOVA, and Regression
• Both T-test and ANOVA illustrate fundamental
concepts needed to understand “Regression”
• Relevant ANOVA concepts
• The idea of a “model”
• Partitioning variance
• A dependent variable
• Relevant T-test concepts
• Using the t-distribution for hypothesis tests
• Note: For many applications, regression will
supersede T-test, ANOVA
• But in some cases, they are still useful…