Design and Analysis of

Single-Factor Experiments:
The Analysis of Variance
 Engineering Experiments
• Experiments are a natural part of the engineering
decision-making process
• Designed to improve the performance of a subset
of processes
• Processes can be described in terms of controllable
variables
• Determine which subset has the greatest influence
• Such analysis can lead to
– Improved process yield
– Reduced variability in the process and closer
conformance to nominal or target requirements
– Reduced cost of operation
 Steps In Experimental Design
• Usually designed sequentially
• Determine which variables are most important
• Used to refine the information to determine the
critical variables for improving the process
• Determine which variables result in the best
process performance
 Single Factor Experiment
• Assume a parameter of interest
• Consist of making up several specimens in two
samples
• Analyzed them using the statistical hypotheses
methods
• Can say an experiment with single factor
– Has two levels of investigations
– Levels are called treatments
– Treatment has n observations or replicates
Designing Engineering Experiments
• More than two levels of the factor
• This chapter shows
– ANalysis Of VAriance (ANOVA)
– Discuss randomization of the experimental
runs
• Design and analyze experiments with
several factors
 Linear Statistical Model
• Following linear model • Each treatment defines
Yij= µ+i+ij a population
• Mean µi consisting of
• i=1, 2,…,a, and j=1,
the overall mean µ
2,…,n
• Plus an effect i
• Yij is (ij)th observation
• µ called the overall
mean
• i called the ith
treatment effect
• ij is a random error
component with mean Pg. 471 Fig 13-1b
zero and variance 2
Completely Randomized Design
• Table shows the underlying model
– Following observations are taken in random order
– Treatments are used as uniform as possible
• Called completely randomized design
 Fixed-effects and Random
Models
• Chosen in two different ways
– Experimenter chooses the a treatments
• Called the fixed-effect model
– Experimenter chooses the treatments from a
larger population
• Called random-effect model
 Development of ANOVA
• Total of the observations and the average of the
observations under the ith treatment

• Grand total of all observations and the grand mean

– N=an is the total number of observations

– “dot” subscript notation implies summation
 Hypothesis Testing
• Interested in testing the equality of the
following a treatment means
1,2 ….. a
• Equivalent
H0: 1=2=…=a=0
H1: a#0 for at least one i
• If the null hypothesis is true, changing the
levels of the factor has no effect on the
mean response
 Components of Total Variability
• Total variability in data is described by the total
sum of squares
• Partitions this total variability into two parts
a n a a n

 ij ..
( y − y
i =1 j =1
) =2
n  i. ..  ij i.
( y −
i =1
y ) +2
( y − y ) 2

i =1 j =1

• Measure the differences between treatments

• Measure the random error effect
 Computational Formulas
• Efficient formulas • Mean square for
• Total sum of squares treatments
2
MSTreatments=SSTreatments /(a-1)
a n
y..
SST =  yij −
2

i =1 j =1 N
• Error mean square
MSE=SSE /[a(n-1)]
• Treatment sum of
squares 2 2 a
yi y
SSTreatments =  − ..
i =1 n N
• Error sum of squares
SSE=SST - SSTreatment
ANOVA TABLE
 Using Computer Software
• Packages have the capability to analyze data
from designed experiments
• Presents the output from the Minitab one-way
analysis of variance routine
 Example
• The tensile strength of a synthetic fiber is of interest to the
manufacturer. It is suspected that strength is related to the
percentage of cotton in the fiber. Five levels of cotton
percentage are used, and five replicates are run in
random order, resulting in the data below. Use α=0.05.

a) Does cotton percentage affect breaking strength?

 Solution
• Use the general steps in hypothesis testing
1. Parameter of interest is the cotton percentage
2. H0: 1=2= 3=4=5=0
3. H1: i #0 for at least one I
4. α = 0.05
5. Test statistic
Fo = MSTR /MSE
6. Reject Ho if fo > fα,(a-1)n(a-1)
7. Computations
 Initial calculations
• Compute the last two columns

Conc 1 2 3 4 5 yi. yi.

15 7 7 15 11 9 49 9.8
20 12 17 12 18 18 77 15.4
25 14 18 18 19 19 88 17.6
30 19 25 22 19 23 108 21.6
35 7 10 11 15 11 54 10.8
y.. =376
 Solution - Cont.
7. Compute SST, SSTR, SSE , MSTR, and MSE
a n 2
y
SST =  yij − .. =(7)2 +(7)2 + ….+(376)2/25= 636.96
2

i =1 j =1 N
a 2 2
y y
SSTreatments =  i − .. =((49)2 +(77)2 +..+(54)2)/5 -376/5
i =1 n N
=475.7
SSE = 636.96-475.75 = 161.20
MSTR= SSTR/a-1 = 475.76/4 = 118.9
MSE = SSE/a(n-1)=161.20/5(5-1) = 8.0
• Hence, the test statistic
Fo = MSTR /MSE = 118.96/8.06 = 14.75
8. Since fo=14.75> f0.05,4,20= 2.87, reject Ho
 Solution
• ANOVA results
Source DF SS MS F P
COTTON 4 475.76 118.94 14.76 0.000
Error 20 161.20 8.06
Total 24 636.96
• Reject H0 and conclude that cotton percentage
affects breaking strength
Multiple Comparisons Following
the ANOVA
• When H0:1=2=…=a=0 is rejected
• Know that some of the treatment are different
• Doesn’t identify which means are different
• Called multiple comparisons methods
• Called Fisher’s least significant difference (LSD)
method
 Fisher’s Least Significant
Difference (LSD) Method
• Compares all pairs of means with the H0:  i =  j
for all i # j
• Test statistic to = yi − y j . .

2 MS E
n
• Pair of means i and j would be different
yi. − y j.  LSD
• Least significant difference, LSD, is
2MS E
LSD = t / 2,a ( n −1)
n
 Example
• Use Fisher’s LSD method with α = 0.05 test to
analyze the means of five different levels of
cotton percentage content in the previous
example
• Recall H0 was rejected and concluded that cotton
percentage affects the breaking strength
• Apply the Fisher’s LSD method to determine
which treatment means are different
 Solution
• Summarize
– a =5 means, n=5, MSE = 8.06, and
t0.025,20=2.086
– Treatment means are
y1. 9.8
y2. 15.4
y3. 17.6
y4. 21.6
y5. 10.8
 Solution –Cont.
• Value of LSD
2MS E 2(8.06)
LSD = t / 2,a ( n −1) LSD = 2.086 = 3.7455
n 5
• Comparisons
5 Vs. 1=I10.8–9.8I=1
5 Vs. 2=I10.8-15.4I=4.6>3.74
5 Vs. 3=I10.8-17.6I=6.8>3.74
5 Vs. 4=I10.8-21.6I=10.8>3.74
4 Vs. 1=I21.6-9.8I=11.8>3.74
4 Vs. 2=I21.6-15.4I=6.2 > 3.74
4 Vs. 3=I21.6 –17.6I=4>3.74
3 Vs. 1=I17.6-9.8I=7.8>3.74
3 Vs. 2=I17.6 -15.4I=2.2
2 Vs. 1=I15.4-9.8I=5.6>3.74
• From this analysis, we see that there are significant differences
between all pairs of means except 5 vs. 1 and 3 vs. 2
 C.I. on Treatment Means
• Confidence interval on the mean of the ith
treatment µi
MSE MSE
yi . − t / 2,a ( n −1)  i  yi . + t / 2,a ( n −1)
n n

• Confidence interval on the difference in two

treatment means

2 MSE 2 MSE
yi. − y j. − t / 2,a ( n −1)  i −  j  yi. − y j. + t / 2,a ( n −1)
n n
 Determining Sample Size
• Choice of the sample size to use is important
• OC curves provide guidance in making this
selection
• Power of the ANOVA test is
1-β=P( Reject H0 | H0 is false)
=P(F0> fα, a-1, a(n-1) | H0 is false)
• Plot β against a parameter 
a
n i
2

2 = i =1

a 2
Sample OC Curves
 Example
• Suppose that four normal populations
have common variance 2=25 and means
µ1 =50, µ 2=60, µ3=50, and µ4=60. How
many observations should be taken on
each population so that the probability of
rejecting the hypothesis of equality of
means is at least 0.90? Use α=0.05
 Solution
• Average mean  = 55
• 1 = -5,a 2 = 5, 3 = -5, 4 = 5
n i
2
n(100)
 =
2 i =1  =
2
= n, a − 1 = 3 where a(n − 1) = 4(n − 1)
a 2 4(25)

• Various choices:

n 2  a(n-1) Power=1-
4 4 2 12 0.80
5 5 2.24 16 0.95

• Therefore, n = 5 is needed
 The Random-effects Model
• A large number of possible levels
• Experimenter randomly selects a of these
levels from the population of factor levels
– Called random-effect model
• Valid for the entire population of factor
levels
 Linear Statistical Model
• Following linear model
Yij= µ+i+ij
•  =1,2,….a, j=1,2,…n
• Yij is the (ij)th observation
• i and ij are independent random
variables
• Identical in structure to the fixed-effects
case
• Parameters have a different interpretation
• ij are with mean 0 and variance 2
• i are with mean zero and variance 2
 Testing the Hypothesis
• Testing the hypothesis that the individual treatment
effects are zero is meaningless
• Appropriate to test a hypothesis about the variance
of the treatment effect
H0: 2 =0 vs. H1: 2 >0
• 2 =0, all treatments are identical
• There is variability between them
• Total variability
SST=SSTreatments +SSE
• Expected values of the MS Eq 13 -21,22
E(MSTreatments)= 2 + n2 and E(MSE)= 2
• Computational procedure and construction of the
ANOVA table are identical to the fixed-effects case
Randomized Complete Block Design
• Desired to design an experiment so that the
variability arising from a nuisance factor can be
controlled
• Recall about the paired t-test
• When all experimental runs cannot be made
under homogeneous conditions
• See the paired t-test as a method for reducing
the noise in the experiment by blocking out a
nuisance factor effect
• Randomized block design can be viewed as an
extension of the paired t-test
• Factor of interest has more than two levels
• More than two treatments must be compared
 Randomized Complete Block Design
• General procedure for a randomized complete
block design consists of selecting b blocks
• Data that result from running a randomized
complete block design for investigating a single
factor with a levels and b blocks
 Linear Statistical Model
• Following linear model

• Yij is the (ij)th observation 

• j is the effect of the jth block

• µ called the overall mean
• i called the ith treatment effect
• ij is a random error component with mean zero
and variance 2
 Hypothesis Testing
• Interested in testing the equality of the a
treatment means
1,2 ….. a
• Equivalent
H0: 1=2=…=a=0
H1: a#0 for at least one i
• If the null hypothesis is true, changing the
levels of the factor has no effect on the
mean response
Displaying Data
 Components of Total Variability
• Total variability in data
• Partitions this total variability into three parts

• Or symbolically,
SST=SStreatments+SSblocks+SSerrors
 Computational Formulas
• Computing formulas for the sums of squares

• Error sum of squares

SSerrors=SST-SStreatments-SSblocks
• Computer software package will be used to
perform the analysis of variance
Analysis of Variance