building models for complex design of experiments
DESCRIPTION
This talk was presented live at JMP Discovery Summit 2012 in Cary, North Carolina, USA. More information about design of experiments is available at http://www.jmp.com/applications/doe/TRANSCRIPT
Copyright © 2010 SAS Institute Inc. All rights reserved.
Building Models for Complex DOEs Donald McCormack, JMP
2
Copyright © 2010, SAS Institute Inc. All rights reserved.
Intro
Basic Designs
Adding nuisance variables – Latin Squares
When blocks matter – Split Plots
Three random effects – Strip and Split-Split Plots
Crossover Designs
Other designs – Split Plot and Latin Square variations.
3
Copyright © 2010, SAS Institute Inc. All rights reserved.
Basic Designs
Typical DOE − Completely Randomized Design (CRD)
Temp: 25° Temp: 30°
pH: 6.0 pH: 7.0
Strain A Strain B
Factor 3 Factor 2 Factor 1
A, 6.0, 30° B, 7.0, 25° A, 6.0, 25° B, 6.0, 30°
B, 7.0, 30° A, 7.0, 30° A, 7.0, 25° B, 6.0, 25°
4
Copyright © 2010, SAS Institute Inc. All rights reserved.
Basic Designs
Typical DOE − Completely Randomized Block Design (CRBD)
Temp: 25° Temp: 30°
Factor 3 pH: 6.0 pH: 7.0
Factor 2 Strain A Strain B
Factor 1
A, 6.0, 30° B, 7.0, 25° A, 6.0, 25° B, 6.0, 30°
B, 7.0, 30° A, 7.0, 30° A, 7.0, 25° B, 6.0, 25° CRD1
Growth Media 1
B, 6.0, 25° A, 7.0, 25° A, 6.0, 30° A, 7.0, 30°
B, 7.0, 30° B, 7.0, 25° A, 6.0, 25° B, 6.0, 30° CRD2
Growth Media 2
Growth Media
Factor 4
5
Copyright © 2010, SAS Institute Inc. All rights reserved.
Latin Squares
Two blocking variables, rows and columns, used for nuisance variables. Two restrictions on randomization – there must be unique
combinations of treatments across rows and down columns. Number of levels must be identical for row, column, and
treatment variables.
Assumption: No two way or higher interaction between row, column, and treatment factors.
More than two nuisance variables? Graeco-Latin and Hyper-Graeco Latin designs.
JMPer Cable Spring 2002
6
Copyright © 2010, SAS Institute Inc. All rights reserved.
Latin Squares - Examples
Emissions Box, Hunter, & Hunter p. 157 Fuel additive is the treatment. Drivers and cars are blocking variables, 4 of each.
Emissions 2 Example 1 with two replicated LS Same Drivers and Cars?
1 2 3 4
1 A B D C
2 D C A B
3 B D C A
4 C A B D
Emissions Example
Car
Driver
7
Copyright © 2010, SAS Institute Inc. All rights reserved.
Latin Squares - Summary
Treat nuisance (blocking) variables as random effects
Unbound the variance components
No nesting or crossing unless there is replication If there are different sets of nuisance variables across replication,
nest the nuisance variable in the replication variable. For example, if the cars in Rep 1 were different than the cars in Rep two, next Car in Rep (Car [Rep]).
8
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots
Am I free to let any factors change at any run? Yes – CRD No, I have to restrict where, when, or how often one or more
factors is changed. » Test for statistical differences in at least one restricted factor?
» No – RCBD, Latin Square » Yes – Split Plot
What’s the difference? RCBD, Latin Square – I’m estimating (nuisance) variability so it
can be removed from experimental variability. Split Plot – I’m estimating both the signal and noise variability of
the affected factor and comparing the former to the later as my statistical test.
9
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots
Two columns are needed One for the block (noise variability) One for the factor (signal)
Two ways block column can be arranged: CR – Each time a factor level changes the block ID changes. RCB – Blocks correspond to groups of unrepeated factor levels.
The nature of the factor often dictates whether you’ll have CR or RCB blocks. Customer Designer uses CR.
You’ll need at least the number of factor levels plus one CR blocks or two RCBD blocks with the same level appearing at least once in both blocks. More is better.
Block arrangement affects how the model is built.
10
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots – Set Up: Example
Heat treatment in oven. Three factors: Temperature, Time, and Power. Oven can fit four units. Scenario 1 – Only one temp per oven run. Scenario 2 – Two temperature zones in an oven with two items
per zone.
11
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots – Set Up: Example Scenario 1
Only one temperature per whole plot (Oven Run). Set Temp to Nominal and nest Oven Run in Temp. JMP default –Leave Temp continuous and ignore the nesting
(keep Oven Run random). You’ll get the same results. In both cases, use REML and unbounded variance components.
Oven Run as CR Block JMP Default
Both give the same results
12
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots – Set Up: Example Scenario 2
Include Oven Run.
Cross Temp with Oven Zone.
Make both Random.
Oven Run*Temp&Random is used as the noise estimate to test for differences in Temp. It removes the run to run variability between ovens.
13
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots – Summary
The hard to change/batch factor needs two columns, one for the factor and one for the block
CR blocks Each time the factor changes so does the block ID Nest the block variable in the hard to change/batch factor. Make
it a random effect. You can also use the JMP default and ignore the nesting.
RCB blocks Group sets of the factor changes into blocks such that no level is
repeated in a given block. Cross the hard to change factor with the block factor and make it
random.
14
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split-Split and Strip Plots
Randomization restriction on two factors
A1B1
A2B1
A1B2
A2B2
B1 B2 A1
B1 B2 A2
Split Split-Split
Strip A1
A2
A1
A2
B1
B2
B1
B2
15
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots: Two Hard to Change Factors Change Simultaneously Just like a split plot: one additional source of error.
CR Block – ID changes if either factor changes.
RCB Block – Grouping based on unique combinations of both factors.
CR Blocks
RCB Blocks
JMP Default
16
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots: Two Hard to Change Factors Change Simultaneously How to ID the blocks
A1B1
A2B1
A1B2
A2B2
A1B1
A2B1
A1B2
A2B2
1
2
2
5
4
3
6
7
8
1
CR Blocks RCB Blocks
17
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots: Split-Split Plot
Two additional sources of error: whole plot and subplot Subplot is more frequently changing, but still restricted, block
inside of whole plots. Whole plots are very hard to change and subplot are hard to change.
Example: High throughput reactor (see Castillo, Quality Engineering 2010)
Reactor Module
Temperature Pressure
Catalyst Type Concentration
Reactor Block
Purge Type
18
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots: Split-Split Plot
Because both whole plot and subplot are arranged as CR blocks, both Fit Models produce the same results.
JMP Default CR Blocks
19
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots: Split-Split Plot
Runs 20 – 42
20
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots: Split-Split Plot
How to ID the blocks – Whole Plots
B1 B2 A1
B1 B2 A2
B1 B2 A1
B1 B2 A2
2
3
4
1 2
1
CR Blocks
RCB Blocks
21
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split Plots: Split-Split Plot
How to ID the blocks – Subplots
B1 B2 A1
B1 B2 A2
B1 B2 A1
B1 B2 A2
2
3
4
1
1
RCB Blocks
8 7
2
3
6
4
5
CR Blocks
22
Copyright © 2010, SAS Institute Inc. All rights reserved.
Strip Plots: Example
Two step semiconductor process: ion implant followed by a thermal anneal. Implant: Three factors – O+ Dose, Energy, Implant Temp Anneal: Three factors - O+ Conc, Anneal Temp, Time Both are batch processes.
The treatment combinations for each step come from a full factorial (32) plus center point. Nine unique combinations possible.
Nine wafers are processed at each step.
For each implant run (i.e., for a unique implant treatment combination) randomly assign each wafer to a unique anneal treatment combination.
Replicate the experiment for 162 wafers total.
23
Copyright © 2010, SAS Institute Inc. All rights reserved.
Strip Plots: Example
1, 1, 1
Implant
1, 1, -1
1, -1, 1
-1, 1, 1
-1, -1, -1
1, 1, 1
Anneal
1, 1, -1
1, -1, 1
-1, 1, 1
-1, -1, -1
9 wafers each step
1 wafer from each implant step randomly assigned
to anneal step
X 2
24
Copyright © 2010, SAS Institute Inc. All rights reserved.
Strip Plots
How to ID the blocks – CR blocks
A1
A2
A1
A2
B1
B2
B1
B2
1
2
3
4
1
2
3
4
WP1 WP2
B2 B1 B2 B1
A1
A1
A2
A2
1
2
3
4
1
2
3 4 WP2
WP1
25
Copyright © 2010, SAS Institute Inc. All rights reserved.
Strip Plots
How to ID the blocks – RCB blocks Count each set of treatment combinations
A1
A2
A1
A2
B1
B2
B1
B2
Rep - 1
B2 B1 B2 B1
A1
A1
A2
A2
Rep - 1
26
Copyright © 2010, SAS Institute Inc. All rights reserved.
Split-Split and Strip Plots
Split-Split
Strip
27
Copyright © 2010, SAS Institute Inc. All rights reserved.
Example – Split-Strip Plot
Fertilizer
S3 S2 S1
Soil Type
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
Ca0 Ca1
F0
F1
F2
F3
28
Copyright © 2010, SAS Institute Inc. All rights reserved.
Crossover Designs
Only one random effect – Subject[Sequence] Biggest challenge is setting up the dataset to estimate the
carryover effect. Example - Three periods, two treatments
JMPer Cable Fall 2006
29
Copyright © 2010, SAS Institute Inc. All rights reserved.
Additional Designs
30
Copyright © 2010, SAS Institute Inc. All rights reserved.
Other Designs: Latin Squares
Two factor full factorial in LS: Radar Detection Montgomery DOE 7th Ed, table 5.23
Hyper-Graeco-Latin Square: Wear testing Box, Hunter, & Hunter p. 163
Wear Testing Radar Detection
31
Copyright © 2010, SAS Institute Inc. All rights reserved.
Other Designs: Split Plots
Split-Split-Split
Strip with multiple treatments assigned to the strips.
Copyright © 2010 SAS Institute Inc. All rights reserved.