Introduction to Experimental Design

Report
Testing Multiple Means
and the Analysis of Variance
(§8.1, 8.2, 8.6)
• Situations where comparing more than two
means is important.
• The approach to testing equality of more than
two means.
• Introduction to the analysis of variance table, its
construction and use.
Study Designs and Analysis
Approaches
1. Simple Random Sample
from a population with
known s - continuous
response.
2. Simple Random Sample
from a population with
unknown s - continuous
response.
3. Simple RandomSamples
from 2 popns with known s.
4. Simple Random Samples
from 2 popns with unknown
s.
One sample z-test.
One sample t-test.
Two sample z-test.
Two sample t-test.
Sampling Study with t>2
Populations
One sample is drawn independently and randomly from
each of t > 2 populations.
Objective: to compare the means of the t populations for
statistically significant differences in responses.
Initially we will assume all populations have common
variance, later, we will test to see if this is indeed true.
(Homogeneity of variance tests).
Sampling Study
Cholesterol
Levels
Meat &
Potato
Eaters
Health
Eaters
Vegetarians
Random
Sample
Random
Sample
Random
Sample
y 11
y 12

y 1n1
y 21
y 22

y 2n 2
y 31
y 32

y 3 n3
Experimental Study
with t>2 treatments
Experimental Units: samples of size n1, n2, …, nt, are
independently and randomly drawn from each of the t
populations.
Separate treatments are applied to each sample.
A treatment is something done to the experimental units
which would be expected to change the distribution
(usually only the mean) of the response(s).
Note: if all samples are drawn from the same population
before application of treatments, the homogeneity of
variances assumption might be plausible.
Experimental Study
Male College Undergraduate Students
Random Sampling
Health
Diet
Set of
Experimental
Units
Cholesterol
Levels @ 1
year.
Veg.
Diet
y 11
y 12

y 1n1
Set of
Experimental
Units
y 21
y 22

y 2n 2
Set of
Experimental
Units
y 31
y 32

y 3 n3
M&P
Diet
Responses
Hypothesis
Let i be the true mean of treatment group i (or population i ).
H0 :
1   2 
 t
Ha :
some of the means are different
Hence we are interested in whether all the groups
(populations) have exactly the same true means.
The alternative is that some of the groups (populations)
differ from the others in their means.
Let 0 be the hypothesized common mean under H0.
A Simple Model
Let yij be the response for experimental unit j in group i,
i=1,2, ..., t, j=1,2, ..., ni. The model is:
yij  i   ij
E(yij)= i: we expect the group mean to be i.
ij is the residual or deviation from the group mean.
Each population has normally distributed responses
around their own means, but the variances are the same
across all populations.
Assuming yij ~ N(i, s2), then ij ~ N(0, s2)
If H0 holds, yij = 0 + ij , that is, all groups have the same
mean and variance.
A Naïve Testing Approach
Test each possible pair of groups by
performing all pair-wise t-tests.
H0 : 1   2
H0 : 1  3
H0 :  2  3
• Assume each test is performed at the a=0.05 level.
• For each test, the probability of not rejecting Ho when Ho is true is 0.95 (1-a).
• The probability of not rejecting Ho when Ho is true for all three tests is (0.95)3
= 0.857.
• Thus the true significance level (type I error) for the overall test of no
difference in the means will be 1-0.857 = 0.143, NOT the a=0.05 level we
thought it would be.
Also, in each individual t-test, only part of the information available to
estimate the underlying variance is actually used. This is inefficient WE CAN DO MUCH BETTER!
Testing Approaches - Analysis of Variance
The term “analysis of variance” comes from the fact
that this approach compares the variability observed
between sample means to a (pooled) estimate of
the variability among observations within each
group (treatment).
-
y
y
y
Within groups
variance
4
- 3
- 2
- 1
01234
4
3
- 2
- 1
01234
- 4
- 3
- 2
- 1
012
3
4
x
y1
x
y2
x
y3
Between sample
means variance
y
y
y
Extreme Situations
-
4
- 3
- 2
- 1
01234
4
3
- 2
- 1
01234
- 4
- 3
- 2
- 1
012
3
4
x
x
x
yy
y
Within group variance is small compared to variability between means.
Clear separation of means.
-4
3
2
1
0
1
2
3
4
4
3
2
1
0
1
2
34
4
3
2
1
0
1
2
3
4
x
x
x
Within group variance is large compared to variability between means.
Unclear separation of means.
Pooled Variance
From two-sample t-test with assumed equal variance, s2, we produced a
pooled (within-group) sample variance estimate.
y1  y 2
t
1 1
sp

n1 n2
(n1  1)s12  (n2  1)s22
sp 
n1  n2  2
Extend the concept of a pooled variance to t groups as follows:
(n1  1)s12  (n2  1)s22    (nt  1)s2t SSW
s 

(n1  1)  (n2  1)    (nt  1)
nT  t
2
w
nT   ni
If all the ni are equal to n then this reduces to an average variance.
1 t 2
s   si
t i1
2
w
t
i1
Variance Between Group Means
t
Consider the variance between the group
means computed as:
y 
y
i 1
i
t
t
s2 
 (y  y
i
i 1

)2
t 1
If we assume each group is of the same size, say n, then under H0, s2 is an
estimate of s2/n (the variance of the sampling distribution of the sample mean).
Hence, n times s2 is an estimate of s2. When the sample sizes are unequal,
the estimate is given by:
where
t
1
SSB
2
sB2 
n
(
y

y
)

 i i 
t  1 i 1
t 1
1
yi  
ni
ni
y
j 1
ij
1
y 
nT
t
ni
 y
i 1 j 1
ij
F-test
Now we have two estimates of s2, within and between means. An F-test
can be used to determine if the two statistics are equal. Note that if the
groups truly have different means, sB2 will be greater than sw2. Hence the
F-statistic is written as:
sB2
F  2 ~ F( t 1), (nT t )
sW
If H0 holds, the computed F-statistics should be close to 1.
If HA holds, the computed F-statistic should be much greater than 1.
We use the appropriate critical value from the F - table to
help make this decision.
Hence, the F-test is really a test of equality of means under the
assumption of normal populations and homogeneous variances.
Partition of Sums of Squares
  y
t
2
ni
i 1 j1
 y
ij
    y
t
i 1 j1
=
TSS
Total
Sums of
Squares
ni
=
ij
SSW
Sums of
Squares
Within
Groups
ni
t
 y i
  n y  y 
t
2
i
i 1

i
SSB
+
Sums of
Squares
Between
Means
+
t
2
ni
TSS   ( yij  y ) 2  (nT  1)sT2   yij2  nT y2
i 1 j 1
i 1 j 1
t
t
i 1
i 1
SSB   ni ( yi  y ) 2   ni yi2  nT y2
SSW  TSS  SSB
TSS measures
variability about
the overall mean
The AOV (Analysis of Variance) Table
The computations needed to perform the F-test for
equality of variances are organized into a table.
Partition of
the sums of
squares
Partition of
the degrees
of freedom
Variance
Estimates
Test
Statistic
Example-Excel
=average(b6:b10)
=var(b6:b10)
=sqrt(b13)
=count(b6:b10)
=(B15-1)*B13
=(sum(B15:D15)-1)*var(B6:D10)
=sum(b16:d16)
=b18-b19
Excel Analysis Tool
Pac
Example SAS
proc anova;
class popn;
model resp = popn ;
title 'Table 13.1 in Ott Analysis of Variance';
run;
Table 13.1 in Ott - Analysis of Variance
31
Analysis of Variance Procedure
Dependent Variable: RESP
Source
DF
Sum of
Squares
Mean
Square
F Value
Pr > F
Model
2
2.03333333
1.01666667
5545.45
0.0001
Error
12
0.00220000
0.00018333
Corrected Total
14
2.03553333
R-Square
C.V.
Root MSE
RESP Mean
0.998919
0.247684
0.013540
5.466667
DF
Anova SS
Mean Square
F Value
Pr > F
2
2.03333333
1.01666667
5545.45
0.0001
Source
POPN
GLM in SAS
General Linear Models Procedure
Dependent Variable: RESP
DF
Sum of
Squares
Mean
Square
F Value
Pr > F
Model
2
2.03333333
1.01666667
5545.45
0.0001
Error
12
0.00220000
0.00018333
Corrected Total
14
2.03553333
R-Square
C.V.
Root MSE
RESP Mean
0.998919
0.247684
0.013540
5.466667
DF
Type I SS
Mean Square
F Value
Pr > F
2
2.03333333
1.01666667
5545.45
0.0001
DF
Type III SS
Mean Square
F Value
Pr > F
2
2.03333333
1.01666667
5545.45
0.0001
Source
proc glm;
class popn;
model resp =
popn / solution;
title 'Table
13.1 in Ott’;
run;
Source
POPN
Source
POPN
Parameter
INTERCEPT
POPN
1
2
3
Estimate
5.000000000
0.900000000
0.500000000
0.000000000
B
B
B
B
T for H0:
Parameter=0
Pr > |T|
Std Error of
Estimate
825.72
105.10
58.39
.
0.0001
0.0001
0.0001
.
0.00605530
0.00856349
0.00856349
.
NOTE: The X'X matrix has been found to be singular and a generalized inverse
was used to solve the normal equations.
Estimates followed by the
letter 'B' are biased, and are not unique estimators of the parameters.
Minitab
Example
STAT > ANOVA > OneWay (Unstacked)
One-way Analysis of Variance
Analysis of Variance
Source
DF
SS
Factor
2 2.033333
Error
12 0.002200
Total
14 2.035533
Level
EG1
EG2
EG3
MS
1.016667
0.000183
N
Mean
StDev
5
5
5
5.9000
5.5000
5.0000
0.0158
0.0071
0.0158
Pooled StDev =
0.0135
F
5545.45
P
0.000
Individual 95% CIs For Mean
Based on Pooled StDev
----+---------+---------+---------+-
(*
*)
(*
----+---------+---------+---------+5.10
5.40
5.70
6.00
R Example
Function: lm( )
> hwages <- c(5.90,5.92,5.91,5.89,5.88,5.51,5.50,5.50,5.49,5.50,5.01,
5.00,4.99,4.98,5.02)
> egroup <- factor(c(1,1,1,1,1,2,2,2,2,2,3,3,3,3,3))
> wages.fit <- lm(hwages~egroup)
> anova(wages.fit)
Df
Sum Sq
Mean Sq
F value Pr(>F)
factor(egroup) 2
2.03333
1.01667
5545.5 2.2e-16 ***
Residuals
12
0.00220
0.00018
--Signif. codes: 0 `***' 0.001 `**' 0.01 `*' 0.05 `.' 0.1 ` ' 1
y1
> summary(wages.fit)
R Example
Call: lm(formula = hwages ~ egroup)
Residuals:
Min
1Q
Median
3Q
-2.000e-02 -1.000e-02 -1.941e-18 1.000e-02
Max
2.000e-02
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 5.900000
0.006055 974.35 < 2e-16 ***
egroup2
-0.400000
0.008563 -46.71 6.06e-15 ***
egroup3
-0.900000
0.008563 -105.10 < 2e-16 ***
--Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 0.01354 on 12 degrees of freedom
Multiple R-Squared:
0.9989,
Adjusted R-squared: 0.9987
y
F-statistic: 5545 on 2 and 12 DF, p-value: < 2.2e-16
1
(Intercept)
5.9
y1
egroup2
-0.4
y2  y1
egroup3
-0.9
y3  y1
A Nonparametric Alternative to
the AOV Test: The Kruskal Wallis Test (§8.5)
What can we do if the normality assumption is rejected in
the one-way AOV test?
.
We can use the standard nonparametric alternative: the
Kruskal-Wallis Test. This is an extension of the
Wilcoxon Rank Sum Test to more than two samples.
Kruskal - Wallis Test
Extension of the rank-sum test for t=2 to the t>2 case.
H0: The center of the t groups are identical.
Ha: Not all centers are the same.
2
t
Ti
12
Test Statistic: H 
 3(nT  1)

nT (nT  1) i 1 ni
t
nT   ni
i 1
Ti denotes the sum of the ranks for the measurements in sample i
after the combined sample measurements have been ranked.
Reject if H > c2(t-1),a
With large numbers of ties in the ranks of the sample measurements use:
H'
H


3
1   (t 3j  t j ) /(nT  nT )
 j

where tj is the number of observations
in the jth group of tied ranks.
options ls=78 ps=49 nodate;
data OneWay;
input popn resp @@ ;
cards;
1 5.90 1 5.92 1 5.91 1 5.89 1 5.88
2 5.51 2 5.50 2 5.50 2 5.49 2 5.50
3 5.01 3 5.00 3 4.99 3 4.98 3 5.02
;
run;
proc print;
run;
Proc npar1way;
class popn;
var resp;
run;
OBS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
POPN RESP
1 5.90
1 5.92
1 5.91
1 5.89
1 5.88
2 5.51
2 5.50
2 5.50
2 5.49
2 5.50
3 5.01
3 5.00
3 4.99
3 4.98
3 5.02
Analysis of Variance for Variable resp
Classified by Variable popn
popn
N
Mean
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒ
1
5
5.90
2
5
5.50
3
5
5.00
Source
DF
Sum of Squares
Mean Square
F Value
Pr > F
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒ
Among
2
2.033333
1.016667
5545.455
<.0001
Within
12
0.002200
0.000183
Average scores were used for ties.
The NPAR1WAY Procedure
Wilcoxon Scores (Rank Sums) for Variable resp
Classified by Variable popn
Sum of
Expected
Std Dev
Mean
popn
N
Scores
Under H0
Under H0
Score
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒ
1
5
65.0
40.0
8.135753
13.0
2
5
40.0
40.0
8.135753
8.0
3
5
15.0
40.0
8.135753
3.0
Average scores were used for ties.
Kruskal-Wallis Test
Chi-Square
12.5899
DF
2
Pr > Chi-Square
0.0018
MINITAB
Stat > Nonparametrics > Kruskal-Wallis
Kruskal-Wallis Test: RESP versus POPN
Kruskal-Wallis Test on RESP
POPN
N
Median
Ave Rank
Z
1
5
5.900
13.0
3.06
2
5
5.500
8.0
0.00
3
5
5.000
3.0
-3.06
Overall
15
8.0
H = 12.50
DF = 2
P = 0.002
H = 12.59
DF = 2
P = 0.002 (adjusted for ties)
R
kruskal.test( )
> kruskal.test(hwages,factor(egroup))
Kruskal-Wallis rank sum test
data: hwages and factor(egroup)
Kruskal-Wallis chi-squared = 12.5899, df = 2, p-value =
0.001846
SPSS

similar documents