Академический Документы
Профессиональный Документы
Культура Документы
Additive Models
STAT 512
Spring 2011
Background Reading
KNNL: 19.1, Chapter 20
28-1
Topic Overview
Additive Two-way ANOVA Model
Two-way ANOVA with n = 1
28-2
Additive Models
In an additive model, interaction terms are
assumed to be zero:
Yijk = + i + j + ijk
) are independent
and i = i = 0
SAS uses as constraints: a = 0, b = 0 , so
compares factor level means to last level of
A and B.
28-3
DF
2
3
6
11
28-4
SS
1544
12
24
62
1642
MS F Value
772
74.71
12
1.16
12
1.16
10.3
Pr > F
<.0001
0.3226
0.3747
28-6
DFpooled = 6 + 2 = 8
MSE pooled = 86 / 8 = 10.75
4 levels of A
4 levels of B
2 observations per cell
ANOVA table with interaction
Source
A
B
A*B
Error
Total
DF
3
3
9
16
31
SS
900
300
90
600
1890
MS
300
100
10
37.5
F Value
8.0
2.7
0.3
28-8
DF
3
3
25
31
SS
900
300
690
1890
MS
300
100
27.6
F Value
10.9
3.6
Pooling Summary
Pooling affects both the significance level
and power of the tests for the main effects and
is not fully understood.
Suggested rules for when to pool:
o Degrees of freedom associated with MSE
are small
o The test statistic MSAB/MSE very small
compared to the critical value (i.e., the pvalue for the interaction effect fairly large)
28-10
ANOVA for n = 1
Weve seen that interaction cannot be tested
for normally as there would be no degrees
of freedom remaining for error.
So for n = 1 we must assume additivity.
How would we check this assumption?
The rest of the analysis is virtually the same
as before except that we are only looking at
main effects.
28-11
Example
Problem 20.2 (terminals.sas)
University conducted experiment to see the
usage levels of coin-operated computer
graphics terminals
Two testing periods (during midterm week
and during finals week) at four different
locations.
Only one observation per cell, but no reason
to expect an interaction here.
28-12
Data
Loc #1
Loc. #2
Loc. #3
Loc. #4
Midterm
16.5
11.8
12.3
16.6
Final
21.4
17.3
16.9
21.0
28-13
Interaction Plot
28-14
Additive Model
Based on interaction plot, additive model
seems appropriate. Should of course also
check other assumptions (normality,
constant variance, etc.)
Source
loc
time
Error
Total
DF
SS
3 37.005
1 47.045
3
0.345
7 84.395
MS
F Value
12.335 107.26
47.045 409.09
0.115
Pr > F
0.0015
0.0003
28-15
28-16
LSMeans Output
location
1
2
3
4
i
1
1
1
2
2
3
j
2
3
4
3
4
4
usage LSMEAN
18.9500000
14.5500000
14.6000000
18.8000000
DIFF
4.40
4.35
0.15
-0.05
-4.25
-4.20
LSMean #
1
2
3
4
95% CI DIFF
2.763536
6.036464
2.713536
5.986464
-1.486464
1.786464
-1.686464
1.586464
-5.886464
-2.613536
-5.836464
-2.563536
28-17
LSMeans Output
Confidence intervals for (1,4) and (2,3)
contain zero.
Terminals 1 and 4 are used more often than
terminals 2 and 3.
time
Final
Midterm
usage LSMEAN
19.1500000
14.3000000
Pr > |t|
0.0003
28-20