Transcript Statistical Data Analysis: Primer
Stat 470-13
•
Today:
Finish Chapter 3 •
Assignment 3
: 3.14 a, b (do normal qq-plots only),c, 3.16, 3.17
•
Additional questions
: 3.14 b (also use the IER version of Lenth’s method and compare to the qq-plot conclusions), 3.19
•
Important Sections in Chapter 3
: – – – 3.1-3.7 (Read these) 3.8 (don’t bother reading), 3.11-3.13 (Read these)
Analysis of Unreplicated 2
k
Factorial Designs
• For cost reasons, 2 k factorial experiments are frequently unreplicated • Can assess significance of the factorial effects using a normal or half normal probability plot • May prefer a formal significance test procedure • Cannot use an F-test or t-test because there are no degrees of freedom for error
Lenth’s Method
• Situation: – have performed an unreplicated 2 k factorial experiment – have 2 k -1 factorial effects – want to see which effects are significantly different from 0 • If none of the effects is important, the factorial effects are an iid sample of size n= 2 k -1 from a N( , ) • Can use this fact to develop an estimator of the effect variance based on the median of the absolute effects
Lenth’s Method
• s 0 = • PSE= • PSE=“pseudo standard error” • t PSE,i =
Lenth’s Method
• Both s 0 and the PSE are estimates of the variance of an effect • Use s 0 to screen out important effects from the calculation of the PSE • Once you have an estimate of the standard error, can construct a t-like statistic • Appendix H of text gives critical values for the t-stats • Will use the IER version of Lenth’s method
Example
: Original Growth Layer Experiment • Effect Estimates and QQ-Plot: -
Effect
A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD
Estimate
0.055
0.142
-0.109
0.836
-0.032
-0.074
-0.025
0.047
0.010
-0.037
0.060
0.067
-0.056
0.098
0.036
-1 0 1 Qu a n t i l e s o f S ta n d a rd No rm a l
• s 0 = • PSE= • t PSE,i = • Cut-off:
Lenth’s Method
Blocking in 2
k
Experiments
• The factorial experiment is an example of a completely randomized design • Often wish to block such experiments • As an example, you may wish to use the same paper helicopter for more than one trial • But which treatments should appear together in a block?
Blocking in 2
k
Experiments
• Consider a 2 3 factorial experiment in 2 blocks
A
-1 -1 -1 -1 +1 +1 +1 +1
B
-1 -1 +1 +1 -1 -1 +1 +1
C
-1 +1 -1 +1 -1 +1 -1 +1
AB
+1 +1 -1 -1 -1 -1 +1 +1
AC
+1 -1 +1 -1 -1 +1 -1 +1
BC
+1 -1 -1 +1 +1 -1 -1 +1
ABC
-1 +1 +1 -1 +1 -1 -1 +1
BLOCK
I II II I II I I II
Blocking in 2
k
Experiments
• What is the relationship between ABC interaction and block?
• What if estimate block effect?
Blocking in 2
k
Experiments
• Can use an interaction to determine which trials are performed in which blocks • Drawback:
Blocking in 2
k
Experiments
• What do the columns in the table mean for a regression model?
• If there was a column for the mean, what would it look like?
Blocking in 2
k
Experiments
• What would the interaction column between the block and ABC interaction look like?
• Can write as:
Blocking in 2
k
Experiments
• Presumably, blocks are important.
• Effect hierarchy suggest we sacrifice higher order interactions • Which is better: – b=ABC – b=AB • Can write as:
Blocking in 2
k
Experiments
• Suppose wish to run the experiment in 4 blocks • b 1 =AB and b 2 =BC • These imply a third relation • Group is called the defining contrast sub-group
Blocking in 2
k
Experiments
• Identifies which effects are confounded with blocks • Cannot tell difference between these effects and the blocking effects
Which design is better?
• Suppose wish to run the 2 3 experiment in 4 blocks • b 1 =AB and b 2 =BC • I=ABb 1 =BCb 2 =ACb 1 b 2 • Suppose wish to run the 2 3 experiment in 4 blocks • b 1 =ABC and b 2 =BC • I=ABCb 1 =BCb 2 =Ab 1 b 2
Ranking the Designs
• Let
D
denote a blocking design •
g i (D)
is the number of (
i=1,2,…k
)
i
-factor interactions confounded in blocks • For any 2 blocking schemes (
D 1 g r
(
D
2 ) and
D 2
) , let
r
be the smallest
i
such
Ranking the Designs
• Effect hierarchy suggests that the design that confounds the fewest lower order terms is best • So, if then D 1 has less aberration • A design has minimum aberration (MA) if no design has less aberration