Week 7 Flashcards
A 2 level full factorial design in f factors has how many treatments
t = 2^f
Unit treatment model for 2 level full 3 factorial design
Main effect for 2 level factor
Where first term is high level, second term is low level
2 factor interaction between 2 level factors
Where the first and second indices correspond to factors A and B respectively
Interaction between p 2-level factors
Generic2 level factorial effect estimator
Variance of generic factorial effect estimator
(Which can be calculated as such because they are unbiased)
Required assumptions for Inference
r > 1 (must be replication)
Error terms must be normally distributed
Why is replication required for inference
r = 1 would result in there being no residual DoF and hence no estimate for σ^2
r > 1 also ensure estimate of s^2 provided by MSE from full treatment model is independent of which factorial effects we choose to estimate
Comparisons made in inference
Is compared to appropriate t dist with N - 2^f DoF
how to determine r
(#of trials)/(#factors in full factorial)
X fractional factorial design confounds Y factorial effects with the mean
X = 2f-q
Y = 2q - 1
q of these effects can be chosen independently
The others are formed as the set of all elementwide products of the first q
General steps to choose a 2f-q design with f factors
1) choose q factorial effects to confound with mean, E1, …, Eq
2) DEFINING RELATION: formed from set of 2q - 1 effects consisting of E1, .., Eq and all of their products: I = E1 = … = Eq = E1E2 = … = E1…Eq
3) find aliasing scheme by multiplying defining relation by each combination of factors
4) find treatment combinations in design (in coded units) by finding treatments that satisfy q equations:
E1 = +/- 1 , E2 = +/- 1, … , Eq +/- 1
Resolution of a 2f-q design is
Length of shortest word in defining relation
Word length pattern
For a 2f-q design
Ai denotes number of factorial effects including i factors in the defining relation
Minimum aberration
For two 2f-q designs labelled δ1 and δ2
let l b the smallest value such that Ai(δ1) != Ai(δ2)
Then δ1 is said to have LESS aberration than δ2 if Ai(δ1) <= Ai(δ2)
If no design has less aberration than δ1 then δ1 has minimum aberration
Basically the design with higher resolution is preferred (according to ex sheet 6)
How to find full aliasing scheme given generators
Multiply generators together, any squared characters become I