Design of Engineering Experiments

Part 5 – The 2k Factorial Design
• Text reference, Chapter 6
• Special case of the general factorial design; k factors,
all at two levels
• The two levels are usually called low and high (they
could be either quantitative or qualitative)
• Very widely used in industrial experimentation
• Form a basic “building block” for other very useful
experimental designs (DNA)

DOE 6E Montgomery

Relationship between the tools
and the practical problems
Practical Questions Statistical Answer
 We have varied all the controllable factors at the Orthogonality Concept
same time how can we estimate the effect of each
factor separately?
 Do all the controllable factors have the same effect Effect estimation
on the examined performance measure?
 How can we quantify the effect controllable factors
of rank them based on their importance?
 Do some of the factors depend on others in Interaction Estimation
affecting the performance measure?
 If we repeat the experiment, another time do we get ANOVA but its assumptions
the same conclusions? must be met.
 To what extent can we generalize the conclusions of
our experiment?

DOE 6E Montgomery

Orthogonallity & Orthogonal Arrays
(OA)
• Orthogonality comes from the Greek orthos, meaning "straight", and
gonia, meaning "angle". It has somewhat different meanings depending
on the context, but most involve the idea of perpendicular, non-
overlapping, varying independently, or uncorrelated.
• In mathematics, two lines or curves are orthogonal if they are perpendicular at
their point of intersection. Two vectors are orthogonal if and only if their dot
product is zero.
• Typically in Cartesian coordinates, one considers primarily bound vectors. A
bound vector is determined by the coordinates of the terminal point, its initial
point always having the coordinates of the origin O = (0,0). Thus the bound
vector represented by (1,0) is a vector of unit length pointing from the origin
up the positive x-axis.

DOE 6E Montgomery

Orthogonallity & Orthogonal Arrays
A B
• Using the 22=4 Orthogonal array, -1 -1
vector A (-1,1,-1,1) and B (-1,-1,1,1) are
1 -1
orthogonal because
– Their inner product is zero -1 1
– All the possible pair of levels appear 1 1
the same number of time

• It is essential to estimate the effect of
each of the studied factors
independently.

DOE 6E Montgomery

Example Tea Experiment Data
F M
Tea
  Type Sugar Cup Response Avg
1 L 1.5 P 3 5 4
2 L 1.5 M 8 4 6
3 L 2 P 5 6 5.5
4 L 2 M 4 5 4.5
5 R 1.5 P 6 5 5.5
6 R 1.5 M 5 4 4.5
7 R 2 P 4 5 4.5
8 R 2 M 6 7 6.5

DOE 6E Montgomery

another time do we get ANOVA but its assumptions the same conclusions? must be met. Relationship between the tools and the practical problems Practical Questions Statistical Answer  We have varied all the controllable factors at the Orthogonality Concept same time how can we estimate the effect of each factor separately?  Do all the controllable factors have the same effect Effect estimation on the examined performance measure?  How can we quantify the effect controllable factors of rank them based on their importance?  Do some of the factors depend on others in Interaction Estimation affecting the performance measure?  If we repeat the experiment.  To what extent can we generalize the conclusions of our experiment? DOE 6E Montgomery .

5 6 R(+1) 1.Determining the most influential factors— calculation of the main effect F M Tea   Type Sugar Cup Response Avg 1 L (-1) 1.375 .875 Avg (+1) 5.5 (-1) M (+1) 5 4 4.5 (-1) P (-1) 6 5 5.5 (-1) M (+1) 8 4 6 3 L (-1) 2 (+1) P (-1) 5 6 5.5 (-1) P (-1) 3 5 4 2 L (-1) 1.5 5 R(+1) 1.25 5.5 Avg (- 1) 5 5 4.5 8 R(+1) 2 (+1) M (+1) 6 7 6.5 7 R(+1) 2 (+1) P (-1) 4 5 4.25 5.5 4 L (-1) 2 (+1) M (+1) 4 5 4.

then there is no interdependence DOE 6E Montgomery . Evaluating the Interdependence • What is interaction? Does tea depend on sugar in impacting the taste? • Fix tea at its (-1) level (Lepton) and vary the sugar-compute the effect of sugar • Fix tea at its (+1) level Rabee and vary the sugar-compute the effect of sugar • Compare the effects calculated in the above two steps-if the are equal.

Evaluating the Interdependence DOE 6E Montgomery .

Evaluating the Interdependence DOE 6E Montgomery .

Evaluating the Interdependence DOE 6E Montgomery .

of three-factor interaction in 2k design = – k!/(3!*(k-3)!) • No. of two-factor interaction in 2k design = – k!/(2!*(k-2)!) • No. of X-factor interaction in 2k design = – k!/(x!*(k-x)!) • Each interaction column is obtained by multiplying the elements of the factors that comprise (form) it DOE 6E Montgomery .Generating Interactions Columns in OAs • No.

of two-factor interaction in 2k design = – 3!/(2!*(3-2)!) = 3 • No. of X-factor interaction in 2k design = – k!/(x!*(k-x)!) • Each interaction column is obtained by multiplying the elements of the factors that comprise (form) it DOE 6E Montgomery . • No. of three-factor interaction in 2k design = – 3!/(3!*(3-3)!) = 1 • No. Interactions in Tea experiments • Three factors (k) were studied.

Estimate the main effects & interactions in the Tea experiments DOE 6E Montgomery .

Plot the main effects in the Tea experiments DOE 6E Montgomery .

Plot the Sugar-Tea type interaction DOE 6E Montgomery .

Plot the Cup-Tea type interaction DOE 6E Montgomery .

Relationship between the tools and the practical problems Practical Questions Statistical Answer  We have varied all the controllable factors at the Orthogonality Concept same time how can we estimate the effect of each factor separately?  Do all the controllable factors have the same effect Effect estimation on the examined performance measure?  How can we quantify the effect controllable factors of rank them based on their importance?  Do some of the factors depend on others in Interaction Estimation affecting the performance measure?  If we repeat the experiment.  To what extent can we generalize the conclusions of our experiment? DOE 6E Montgomery . another time do we get ANOVA but its assumptions the same conclusions? must be met.

53846153 ABC 9 1 9 8 0. Tea Experiment ANOVA   Sum of   Mean F.7051 0.25 4 0.25 4 0.4554 0.15384615 A 0.61538461 C 1 1 1 5 0.0000 5.25 1 0.25 1 0.25 1 0.15384615 B 0.7051 AC 0 1 0 0 1.15384615 AB 0.625     Total 23.25 4 0.   Source Squares DF Square Value P-value 0.0000 BC 0 1 0 0 1.0464 Pure Error 13 8 1.7051 0.75 15       .

5 5 R 1. Example Tea Experiment Data F M Tea   Type Sugar Cup Response Avg 1 L 1.5 M 5 4 4.5 M 8 4 6 3 L 2 P 5 6 5.5 P 6 5 5.5 P 3 5 4 2 L 1.5 7 R 2 P 4 5 4.5 8 R 2 M 6 7 6.5 4 L 2 M 4 5 4.5 6 R 1.5 DOE 6E Montgomery .

• Vertical vs. • Number of data points that provide information -the number of points that can freely vary. ANOVA • Tool for identifying the causes of variances and quantifying their contributions. . horizontal variance • Variance equation: • Every value in a random sample provide a piece of information about the population & to do so it must be unpredictable.

• In hypothesis testing. it is assumed that the effect is zero and the calculated F-value is evaluated on this basis . • If an effect is not significant. ANOVA • If two S2 are computed for two samples that come from two normal populations with a common variance then their ratio S12/S22 follows Fdf1. its variance should be close to the error variance.df2.

F-distribution F  Calculated Rejection Region F 0 F  table .

Two-Factor Analysis of Variance SSA Factor A SSB Factor B SST Interacti SSAB on Between A and B Inherent SSE Variation (Error) .

Two-Factor Analysis of Variance .

Tea Experiment ANOVA .

15384615 AB 0.   Source Squares DF Square Value P-value 0.7051 AC 0 1 0 0 1.7051 0.53846153 ABC 9 1 9 8 0.25 4 0.0000 5. Tea Experiment ANOVA   Sum of   Mean F.15384615 A 0.0000 BC 0 1 0 0 1.25 4 0.7051 0.4554 0.15384615 B 0.75 15       .25 4 0.625     Total 23.0464 Pure Error 13 8 1.61538461 C 1 1 1 5 0.25 1 0.25 1 0.25 1 0.

75 15 .5423 ABC 9 1 9 73 0.   Source s DF Square Value P-value 8.0535 Residual 14. Tea Experiment ANOVA Sum of   Square   Mean F.75 14 71     Total 23.0111 1.

75ABC • This should be used to predict the values of the responses and estimate the error . Tea Experiment Regression • Y = average + (Effect/2)X • Y =5.13+0.

375 0.125 6 5 4.125 3 5 5.375 0.875 11 6 5.125 9 5 4.625 16 7 5.375 5 6 5.875 0.625 13 5 5.Residuals Estimation Standa Predict   rd Actual ed Residu Order Value Value al 1 3 4.625 7 4 4.625 10 4 5.875 4 4 4.875 -1.375 -0.875 0.875 2.375 15 5 4.375 0.875 0.125 12 5 4.375 -0.875 1.375 -0.375 -1.375 0.875 14 4 4.375 8 6 5.375 2 8 5.125 .875 -0.875 -0.

Normally distributed with constant variance . • It should be normally distributed  any linear function of normally distributed variable is normal (Central limit theory) • It should be independent of the run order • Considerably large values should be examined • Its variance with the predicted values should be the same (homogeneous variance) • So errors should be independent. Residuals (error) assessment • Subtracting the actual values from the predicted ones (Y A-Yp) e • Errors average should be zero (the positive impacts cancel the negative ones).

Residuals Plot (Normality) .

Residuals Plot (Normality) • Sort the calculated errors from the smallest to the largest • Assign a rank value (i) to each of them • Calculate (i-0. of residuals) .5)/(No.

625 0.21875 5 -0.40625 8 0.375 0.65625 12 0.125 0.625 0.53125 10 0.375 0.15625 4 -0.03125 2 -1.625 0.875 0.Residuals Plot (Normality) Rank Reseduals (i-0.125 0.875 0.78125 14 0.28125 6 -0.5)/16 1 -1.34375 7 -0.125 0.125 0.59375 11 0.84375 15 1.96875 .625 0.90625 16 2.46875 9 0.875 0.71875 13 0.125 0.09375 3 -0.375 0.375 0.

Residuals Plot (Normality) .

Residuals Plot (Independence) .

Residuals Plot: Variance Constance .

5 8 R 2 M 6 7 6.5 7 R 2 P 4 5 4.5 6 R 1.5 P 3 5 4 2 L 1.5 M 5 4 4.5 .5 P 6 5 5.Use the table of the data and locate the maximum (or the minimum) F M Tea   Type Sugar Cup Response Avg 1 L 1.5 5 R 1.5 M 8 4 6 3 L 2 P 5 6 5.5 4 L 2 M 4 5 4.

If the main effects are the only significant effects use their plot to identify the best settings. .

the impact of tea type on the taste depends on both the sugar quantity and the cup type. • The best performance was attained with Rabee tea when used with two sugar cubes and a Mug. . • Because of the significance of the interaction ABC. Conclusions From the Experiments • None of the studied factors has an independent effect on the tea taste as all the main effects were statistically not significant. In fact only the variance constancy assumption is suspicious. • The residuals analysis revealed no serious violation of the ANOVA assumption.

ABC interaction plot C=P C=M .

5 P 6 6 R 1.5 M 8 3 L 2 P 5 4 L 2 M 4 5 R 1.Single Replicate Experiments   Tea Type Sugar Cup Res 1 L 1.5 M 5 7 R 2 P 4 8 R 2 M 6 .5 P 3 2 L 1.

Single Replicate Experiments • Estimate the effects and divide them into two groups-the small verses the large ones. • Use half-normal probability plot • Use ANOVA –values of the sums of squares should be divided into two groups. .

75 C 1.25 .25 AB 0.75 ABC 2. Estimate the effects and divide them into two groups-the small verses the large ones Term Effect A 0.25 AC -0.25 B -0.75 BC -0.

• Calculate (i-0. of Effects) .5)/(No. Half-Normal Probability Plot • Obtain the absolute values of the effects • Sort the calculated absolute Effects from the smallest to the largest. • Assign a rank value (i) to each Effect.

642857 6 C 1.214286 3 B 0.357143 4 AC 0.5 5 BC 0.25 0.75 0.5)/7 1 A 0.Half-Normal Probability Plot Rank (i) Term Effect (i-0.928571 .75 0.75 0.25 0.25 0.785714 7 ABC 2.071429 2 AB 0.25 0.

Half-Normal Probability Plot .

310345 0.Half-Normal Probability Plot Sum of Mean F Source Squares DF Square Value Prob > F C 3.125 1 10.625 5 0.125 13.725 Total 16.0135 Error 3.125 4.96552 0.0925 ABC 10.125 1 3.875 7 .

Half-Normal Probability Plot   Sum of   Mean F   Source Squares DF Square Value Prob > F ABC 10.125 1 10.125     Total 16.875 7       .0240 Residual 6.125 9 0.75 6 1.

square root.What to do if the Residuals plots reveal unusual pattern • Consider one of the conventional transformations of the response such as: log. inverse square root • Use Power transformation –Box-Cox plot • The above procedures are normally implemented using a software package. ln. • If these transformations are not effective. . the experiments must be repeated.

• If there is one or more significant two or higher factor interactions. use their plots along with the main effects that are not involved in them to identify the best settings. . Selecting Best Settings • Use the table of the data and locate the maximum (or the minimum) • If the main effects are the only significant effects use their plot to identify the best settings.