Module 3: Experimental Design

1
Module 3 Experimental Design
2
Outline - Module 2

• definition and motivation
• limitations of routine operating data
• terminology
• considerations in planning an investigation
• limitations of one-variable-at-a-time strategies
• two-level factorial designs

3
Definition

• An experimental design is a disciplined plan for
  collecting data
• What should we observe, and how should we perturb
  the process?
• How can we maximize the information content of
  the data?

4
Motivation

• Experimental design is an integral component of
  quality improvement, and supports improvement in
• product design
• process design
• process operation

5
Process Investigations
The Course

Process Operation
Experimental Design
Statistically Designed Expts.
Data
Statistical Analysis
Regression Analysis
Information
Insight Through Experience
Knowledge
Quality Improvement
Improve Process/ Product Performance
6
The Iterative Nature of Process Investigations

Identify Objectives
Design Data Collection
Collect Data
Analyze Data
7
Outline - Module 2

• definition and motivation
• limitations of routine operating data
• terminology
• considerations in planning an investigation
• limitations of one-variable-at-a-time strategies
• two-level factorial designs

8
Why not use routine operating data?

• Routine operating data frequently does not
  contain sufficient information of interest, due
  to
• limited range of operating variables due to tight
  control
• values dont vary significantly, so the effects
  of the variables cant be seen
• systematic relationships between operating
  variables
• arising from process control and/or other
  operating policies
• coincidental (correlation) relationships that
  dont necessarily represent cause and effect

9
Active vs. Passive Data Collection

• Active Data Collection
• we actively intervene in the process, and cause
  changes
• Passive Data Collection
• we passively observe, without introducing any
  perturbations into the process
• The only way to ensure that our observations
  represent cause and effect is to introduce
  perturbations (causes) and observe the
  responses (effects).

10
Outline - Module 2

• definition and motivation
• limitations of routine operating data
• terminology
• considerations in planning an investigation
• limitations of one-variable-at-a-time strategies
• two-level factorial designs

11
Terminology

• Responses - measurable outcomes of interest
• frequently have more than one response
• e.g., melt grafting - degree of grafting,
  grafting efficiency
• Factors - controllable variable thought to have
  influence on response
• deliberately manipulated to determine effect on
  response
• Level - value or setting of a factor
• Test run - set of factor level combinations for
  one experimental run

12
Terminology

• Noise variables (covariates) - variables
  affecting process or product performance which
  cannot or are not controlled
• Extraneous variation - variation in measured
  response values in an experiment attributable to
  sources other than deliberate changes in the
  levels of the factors
• Design - selection of test run factor levels -
  set of experimental runs
• Effect - of factors on response, measured by
  change in average response under two or more
  factor level combinations

13
Example - Wave Solder Process

• Scenario - wave solder process producing too many
  defective items - investigate extent to which
  conveyor speed, solder pot temperature and flux
  density affect incidence of defects
• Response -
• Factors -
• Potential noise variables -

14
Outline - Module 2

• definition and motivation
• limitations of routine operating data
• terminology
• considerations in planning an investigation
• limitations of one-variable-at-a-time strategies
• two-level factorial designs

15
Considerations in Planning an Investigation

• What are the objectives of the investigation?
• What are the performance characteristics of
  interest?
• What responses will be used to assess these
  characteristics?
• What factors will be deliberately manipulated?
• What is the feasible operating region?
• What is the operating region of interest?
• What other variables may influence the results?
• Will future additional tests be possible?
• What sets of operating conditions are to be
  tested?
• In what order will the test runs be carried out?

16
Considerations in Planning an Investigation

• Factor effects -
• eliminate systematic bias by including as much as
  possible all factors suspected of having an
  effect - e.g., in a screening study, to
  identify which factors do have a significant
  effect
• what types of relationships do we think exist?
• quadratic, or linear?

17
Outline - Module 2

• definition and motivation
• limitations of routine operating data
• terminology
• considerations in planning an investigation
• limitations of one-variable-at-a-time strategies
• two-level factorial designs

18
One-variable-at-a-time investigations

• Starting at a nominal point, conduct experiments
  in which the first factor is varied, then conduct
  experiments in which only the second factor is
  varied, and so forth
• Example - reactor yield vs. temperature,
  concentration

maximum yield
contours of constant reactor yield

temperature (C)
concentration ()
19
One-variable-at-a-time investigations

• If we conduct one sequence of a
  one-variable-at-a-time investigation (i.e.,
  conduct an experimental sequence in each factor
  only once, rather than repeat the sequence of
  experiments), we will not locate the true value
  of the point of maximum yield
• one factor at a time testing does NOT account for
  possible interactions between the effects of the
  variables - two-factor interactions
• the yield surface contours are rotated ellipses,
  which have cross-product terms --gt two-factor
  interactions

20
Outline - Module 2

• definition and motivation
• limitations of routine operating data
• terminology
• considerations in planning an investigation
• limitations of one-variable-at-a-time strategies
• two-level factorial designs

21
Two-Level Factorial Designs

• Suppose the we have k factors being
  investigated, and we have a region of interest
  defined by low and high limits for each factor

H
range for temperature
temperature
L
L
H
concentration
range for concentration
22
Two-Level Factorial Designs

• we conduct an experiment at every combination of
  high and low values for all factors

H
Runs Coded Values L,L -1,-1 L,H -1,1 H,L 1,-1 H
,H 1,1
temperature
L
L
H
concentration
23
Coding

• The standard coding is
• so that -
• -1 corresponds to the low limit of interest
• 1 corresponds to the upper limit of interest
• note that the average of the upper and low limit
  is essentially the midpoint of the interval of
  interest

24
Coding for Qualitative Factors

• Sometimes the factors under consideration are not
  numerical, but are instead qualitative
• catalyst types A and B
• catalyst preparation method I, II
• suppliers A and B
• machines I and II
• These factors can be coded as -1, 1
• e.g., -1 for catalyst type A, 1 for catalyst
  type B

25
Two-Level Factorial Designs

• If we have k factors under investigation, a
  two-level factorial design will consist of 2k
  runs
• the number of combinations of high and low values
  (two levels) for k factors
• These designs are also known as 2k designs, which
  identifies the number of levels (2), and the
  number of factors (k).

26
General Factorial Designs

• If we have k factors, each considered at mi
  levels, a general factorial design consists of
  experimental runs at all possible combinations of
  the levels for each factor, yielding
• experimental runs.
• Examples
• 2k - three-level factorial design
• 3k - three-level factorial design

27
Two-Level Factorial Designs

• Why place the runs at the limits of the region of
  interest?
• Think of the variance of the slope parameter
  estimate in a straight line model
• Placing the xis as far as possible from the
  average minimizes the variance of the parameter
  estimates
• improved precision of the parameter estimates

28
Two-Level Factorial Designs

• Parameter estimates contain information about the
  effects of the factors --gt precision in parameter
  estimates translates into precision of the
  knowledge of the effects of the factors
• Multiple linear regression case
• placing the points as far from the average point
  as possible maximizes the determinant of XTX
• covariance matrix is based on inverse of XTX -
  area of joint confidence region is proportional
  to 1/sqrt(det(XTX))
• maximizing determinant minimizes area of joint
  confidence region
• yields most precise parameter estimates

29
Randomization

• When implementing a designed experiment, the runs
  should be conducted in a completely randomized
  manner
• guard against sytematic trends caused by other
  variables which would lead to misinterpretation
  of the results -- biased results
• e.g., systematic noise component associated with
  increasingly higher temperatures, slow drift in
  one of the instruments, all low pressure runs
  conducted on a cool day,

30
Information Provided by a Designed Experiment

• Given M distinct sets of factor levels in the
  experimental design, we can estimate
• the overall average response
• M-1 pieces of information about the effects of
  the factors on the response
• This is viewed as providing M-1 independent
  pieces of information about the process (the
  overall average is not viewed as a piece of
  information about the factor effects).
• Link to regression - for M distinct sets of
  experimental runs, we can estimate
• intercept parameter
• M-1 other parameters

total of M parameters
31
Example - Reactor Yield

• What is the effect of concentration (C) and
  temperature (T) on chemical reactor yield?
• prepare a 22 factorial design in T, P -- 4 runs
• Experimental Design

H
Runs Coded Values L,L -1,-1 L,H -1,1 H,L 1,-1 H
,H 1,1
temperature
L
L
H
concentration
32
Example - Reactor Yield

• Information -
• main effects - effect of C, T on yield (2
  pieces of info)
• interaction effect - effect of CT on yield (1
  piece of info)
• total of 3 pieces of information from 4 runs
• remaining run helps provide overall average yield

33
Main Effects

• The main effect of a factor is the average
  influence of a change in level of the single
  factor on the response.
• In the 2-level factorial design,
• Main Effect Average of - Average of
• of a Factor Responses at Responses
• High Level of Factor at Low Level
• of Factor

34
Main Effects - Example

• For temperature -
• average yield at high T is 70
• average yield at low T is 57
• main effect is 70 - 57 13

average 70
72
68
H
temperature
L
60
54
L
H
concentration
average 57
35
Interaction Effects

• Interaction - extent to which influence of one
  factor on response depends on level of another
  factor
• Visually - for reactor example

concentration high
72
Examine influence of Temperature at low conc.,
high conc.
60
68
yield
54
concentration low
L
H
temperature
36
Interaction Effects

• Reactor Yield example -
• the influence of temperature at high
  concentration is slightly larger than the
  influence of temperature at low concentration --
  mild interaction effect
• Interaction effect between temperature and
  concentration -
• 1/2 influence of T at high conc. - influence
  of T at low conc.
• 1/2 14 - 12 1

37
Interaction Effects - General Definition

• For two factors, x1 and x2, the interaction
  effect is
• 1/2 effect of factor 1 on response at high
  level of factor 2
• - effect of factor 1 on response at low level
  of factor 2
• Why divide by 2? - place assessment of
  interaction effect on same basis as that of main
  effects

38
Interaction and Main Effects

• We can return to the interaction plot, and
  visualize the main effects as well

main effect of X2
response
main effect of X1
L
H
X2
39
Using Regression to Estimate Effects

• For the 22 case for the reactor example, we can
  estimate the main effects and 2-factor
  interaction by fitting the following model to the
  data
• Main effect of factor 1 - from defn -
  difference between avg high, avg low

HH
HL
LH
LL
40
Using Regression to Estimate Effects

• General case - to obtain main and 2-factor
  interaction effects for 2k design, fit a
  first-order plus 2-factor interaction model
• In general
• Main effect of factor i
• Interaction effect between factors i, j

41
Example - Chemical Reactor Yield

• Form the X matrix The observation matrix
• The parameter estimate vector

int. x1 x2 (x1 x2 )
42
Effects - for Reactor Example, from Regression

• Using these coefficients, the effects are
• Main effect of x1 2(-2.5) -5
• Main effect of x2 2(6.5) 13
• Interaction effect x1 x2 2(0.5) 1

43
The Effects Representation - another approach!

• The Effects Representation is another approach to
  compute effects for 2-level factorial designs
• Steps
• 1) Form data table
• 2) Compute weighted sum of factor column values
  corresponding response column values
• e.g., for column 1 (the x1 column),
  -160154(-1)72168-10
• 3) Effect for the column is obtained by dividing
  the weighted sum by 2k-1 where k is the number
  of factors
• e.g., for column 1, main effect for factor 1 is
  (-5)/2 -5

44
The Effects Representation - Example

• Sumarizing for the reactor example
• which compares to the results from the other
  approaches.

45
Two-Level Factorial Designs - the 23 Case

• We have 3 factors, and we conduct a 23 design.
• What information can we obtain?
• we have 8-1 7 pieces of independent information
• main effects for 3 factors 3 pieces of info
• 2-factor interaction effects x1x2 , x1x3 , x2x3
  3 pieces of info.
• 3-factor interaction effect x1x2x3 1 piece of
  info
• total of 7 pieces

46
The 23 Design

• Picture

H
x3
H
L
L
x2
H
L
x1
47
Example - Reactor Yield
83
80

Information Available gtgt main effects (3) gtgt
two factor interactions (3) gtgt 3-factor
interaction (1) gtgt total of 7 pieces in 8 runs
72
68
180
45
52
Temperature
II
160
I
catalyst type
60
54
20
40
concentration
48
Example - Reactor Yield
83
80

• Main effect of catalyst type -

72
68
180
45
52
Temperature
II
160
I
catalyst type
60
54
20
40
concentration
49
Example - Reactor Yield

• Main effect for catalyst type

72
72
68
68
-
avg 65
avg 63.5
1.5
60
60
54
54
Catalyst Type II
Catalyst Type I
50
Example - Reactor Yield

• 2-factor interaction effect between catalyst type
  and temperature
• 1/2effect of cat type at high T - effect of
  cat type at low T

81.5
48.5
83
80
52
45
II
72
60
68
54
70
57
I
effect at high T 11.5
effect at low T -8.5
2-factor interaction effect for cat type and T is
1/211.5-(-8.5)10
51
Example - Reactor Yield

• The 2-factor interaction is essentially the
  difference of the averages on the following two
  planes

83
80
cat type T 1
avg69.25
72
68
180
45
52
II
160
I
catalyst type
60
54
cat type T -1
20
40
avg59.25
concentration
52
Example - Reactor Yield

• We can also summarize the 2-factor interaction
  information as we did before, but averaging over
  the values for different concentrations

81.5
70
yield
57
48.5
catalyst type
53
Constructing Factorial Designs

• Standard order -
• levels for first factor alternate (-1, 1)
• levels for second factor alternate with every
  pair of runs (-1,-1,1,1)
• levels for third factor alternate every four runs
  (-1,-1,-1,-1,1,1,1,1)
• and so forth...

54
Design Decisions for 2-level Factorial Designs

• 1) High and low levels for each factor
• from process understanding, objectives,
  preliminary investigations
• 2) Number of runs at each factor level
• 3) Inclusion of centre point runs
• estimate inherent noise variance
• assess curvature over the experimental region
• 4) Balanced design
• preserve cancellation structure of runs

55
Number of Runs

• Conceptually -
• Perform enough runs so that the precision of the
  predicted effects is sufficient to allow
  detection of a certain effect size
• strengthen signal to noise
• Precision
• depends on inherent noise variance (noise) and
  runs
• as of runs increases, precision increases
  (increased information from data)

56
Number of Runs

• Assessment of Significance of Effects
• hypothesis test - effect is not significant
• two types of risk
• type I error - erroneous conclusion that effect
  IS significant (reject null hypothesis) -
  alpha-risk
• type II error - failure to detect a significant
  effect (accept null hypothesis) - beta-risk
• analogy - control charts

57
Number of Runs

• Given - size of effect to be detected
• - variance of inherent noise
• - type I error risk (false detection)
• - type II error risk (failure to detect)
• the number of runs n required at each factor
  level in 2-level factorial design in order to
  detect an effect of the stated size is

value of standard normal r.v. with upper tail
probability alpha/2
58
Number of Runs

• Note- if inherent noise variance is estimated,
  replace Z with Students t-distribution value
• degrees of freedom from noise variance estimate
• Example
• want to detect effect whose magnitude gt 2sigma
• alpha-risk 0.05, beta-risk0.1
• n 2 (1.961.28)2 (0.5)2 5.26 --gt 6 runs at
  each factor level
• Interpretation
• suppose we have a number of factors
• if we conduct a 2-level design with 16 runs or
  more, at least 8 runs will be conducted at each
  factor level --gt we need at least a 16-run design

59
Centre-Point Runs

• The 2-level factorial design is improved by
  adding several runs at the centre point of the
  design ( i.e., set factor levels to 0).
• Benefits -
• have replicates to enable estimation of inherent
  noise variation
• can assess curvature of response surface
• compare average of corner values to average at
  the center - if there is a significant
  difference, curvature is present

60
Centre-Point Runs

• Centre-point runs contribute no additional
  information about main and interaction effects
• think of effects representation - multiplication
  by zeros
• To assess curvature of the response surface
• calculate average of 2k runs
• calculate average of replicate runs at the centre
• use a t-test for differences in means - assuming
  both have the same variance (that estimated from
  the replicates at the centre)

61
Balanced Designs and Replicate Runs

• Replicate runs are
• independent
• repeat runs to estimate inherent noise variance
• Balanced design
• design in which each level of every individual
  factor appears the same number of times in
  combination with each of the levels of every
  other factor
• e.g., 24 design - low level of X1 appears 8 times
  with low level of X2, and .
• Main point - changing the balance of the design
  can dramatically alter the properties provided by
  the design

62
Upsetting the Balance

• Example - 22 design for two factors
• add an additional run at the LL combination
• XTX is no longer diagonal
• parameter estimates are no longer uncorrelated
• effects calculations are no longer uncorrelated
• potential for misleading conclusions
• imbalance because we no longer have an equal
  number of -1, 1 combinations

63
Properties of 2k Designs

• 1) Parameter estimates are uncorrelated - XTX is
  diagonal
• 2) Parameter estimates have uniform precision
• entries in XTX are identical (equal to 2k)
• addition of centre-points improves precision of
  intercept estimate (overall average), but not the
  single factor and interaction term parameters
• 3) Optimality - for any 2-level experimental
  design, 2-level factorial designs -
• provide the most precise parameter estimates
• provide the most precise predicted responses for
  any prediction at a point in the experimental
  region

64
Properties of 2k Designs

• 4) Terms - 2-level factorial designs
• allow estimation of main effects - linear in x
• allow estimation of interactions - products of
  xs
• do NOT allow estimation of quadratics - only two
  levels, minimum of three levels required

65
Precision of Predicted Effects

• Recall the definition of the main effect
• Each average involves half of the 2k points in
  the factorial design - what is the variance of an
  average?
• when m points are used in the average, and the
  variance of the measurements is sigma2.

Main Effect Average of - Average of of a
Factor Responses at Responses High Level of
Factor at Low Level of Factor
66
Precision of Predicted Effects

• Main effect calculation consists of difference
  between two averages, each with 2k/2 points
• remember that variances are ADDITIVE even if we
  subtract random variables
• Variance of calculated main effect is
• Standard devn of calculated main effect is

67
Precision of Predicted Effects

• Precision of interaction effects for 2k designs
• - precision can be derived by thinking of -
• interaction effect as difference between averages
  of 2k-1 points (the diagonal planes
• the formal defn
• the effects representation
• The precision of the interaction effects is the
  same as for the main effects

68
Precision of Predicted Effects

• For cases with replicate runs, think of the
  underlying principle -
• effects are differences of averages at different
  levels
• variances of averages are those of the noise
  divided by the number of points in the average
• variances of sums (and differences) of random
  variables are ADDITIVE

69
Precision of Predicted Effects

• Regression Perspective -
• can examine variance of associated parameter
  estimate
• To determine variance of the effect (as opposed
  to the parameter), consider

for a 2k design
70
Using Precision of Predicted Responses

• Is an effect statistically significant?
• Hypothesis Test
• null hypothesis H0 effect is zero
• alternate hypothesis Ha effect is non-zero
• test statistic

if noise variance is known
if noise variance is estimated
71
Testing Significance of Effects

• To test for significance, compare against
• Zalpha/2 if noise variance is known
• tdf,alpha/2 if noise variance is estimated - df
  degrees of freedom of variance estimate
• this is a two-tailed test - compare absolute
  value of test ratio against the entry from the
  table with upper tail area of alpha/2
• if test ratio exceeds the fence - significant
  effect
• if test ratio is inside the fence - effect is not
  statistically significant
• Confidence intervals can also be formed in a
  manner similar to those for parameter estimates.

72
Obtaining Estimates of the Noise Variance

• - three possible methods -
• replicate runs in the current experimental design
• replicate runs from a previous design
• from nonsignificant effects
• Replicate estimates of variance
• pool if more than one replicate set (e.g., centre
  points vs. vertices)
• test for constant variance
• if replicates from previous design are used,
  confirm that conditions for data collection were
  same as those for current experimental design

best
least
73
Variance Estimate from Nonsignificant Effects

• If certain effects are not statistically
  significant, we conclude that the values reflect
  the extraneous variation (inherent noise).
• These nonsignificant effects can be used directly
  to estimate the variance of the significant
  calculated effects -

This is an estimate of the variance of a
(significant) calculated effect - it is NOT an
estimate of the noise variance.
74
Variance Estimate from Nonsignificant Effects

• The Regression Perspective -
• identifying effects as nonsignificant is
  equivalent to identifying nonsignificant
  parameters in a model
• deleting terms from the model provides additional
  degrees of freedom for variance estimate - we
  now estimate model in p parameters (p-1 effects),
  and we have more data points than parameters
• we can estimate noise variance as MSE
• use variance estimate to form confidence
  intervals, perform hypothesis tests on parameters

75
Normal Probability Plots and Effects

• Normal probability plot - plot of cumulative
  probability vs. observation
• Premise - if values are from a normal
  distribution, normal probability plot will be
  LINEAR, centred at zero
• Procedure -
• order calculated effects from smallest to largest
  
• assign rank i to each effect, from 1 to n ( of
  calculated effects)
• calculate cumulative probability Pi for each
  effect
• plot Pi vs. effects on normal probability paper

76
Normal Probability Plots and Effects

• Nonsignificant effects -
• form line centred about zero
• behave as normal deviates --gt likely associated
  with noise
• Significant effects -
• will not lie on straight line
• e.g., kinks, steeper tail

77
Repeat vs. Replicate Runs

• Replicates -
• represent additional, independent trials at the
  same factor levels
• all sources of extraneous variation must be
  present
• Repeat measurements -
• represent repeated measurements for a given run
• dont have all sources of variation present
• provide indication of measurement noise, not
  entire extraneous variation
• Example - two separate experiments conducted at
  high P, high T vs. two measurements from an
  experiment at high P, high T