Page 1 of 8

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 1: Introduction to Design of Experiments

Lesson 1: Introduction to Design of

Experiments
Introduction

In this course we will pretty much cover the textbook - all of the concepts and designs
included. I think we will have plenty of examples to look at and experience to draw from.

Please note: the main topics listed in the syllabus follow the chapters in the book.

A word of advice regarding the analyses. The prerequisite for this course is STAT 501 -
Regression and STAT 502 - Analysis of Variance. However, the focus of the course is on
the design and not on the analysis. Thus, one can successfully complete this course
without these prerequisites, with just STAT 500 - Applied Statistics for instance, but it will
require much more work, and for the analysis less appreciation of the subtleties involved.
You might say it is more conceptual than it is math oriented.

Learning objectives & outcomes

Upon completion of this lesson, you should be able to do the following:

• understand the issues and principles of Design of Experiments (DOE),

• understand experimentation is a process,
• list the guidelines for designing experiments, and
• recognize the key historical figures in DOE.

Let's start with a simple question...

What is the Scientific Method?

Do you remember learning about this back in high school or junior high even? What were
those steps again?

Decide what phenomenon you wish to investigate. Specify how you can manipulate the
factor and hold all other conditions fixed, to insure that these extraneous conditions aren't
influencing the response you plan to measure.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/5/ 4/18/2019
 Page 2 of 8

Then measure your chosen response variable at several (at least two) settings of the
factor under study. If changing the factor causes the phenomenon to change, then you
conclude that there is indeed a cause-and-effect relationship at work.

How many factors are involved when you do an experiment? Some say two - perhaps this
is a comparative experiment? Perhaps there is a treatment group and a control group? If
you have a treatment group and a control group then in this case you probably only have
one factor with two levels.

How many of you have baked a cake? What are the factors involved to ensure a
successful cake? Factors might include preheating the oven, baking time, ingredients,
amount of moisture, baking temperature, etc.-- what else? You probably follow a recipe so
there are many additional factors that control the ingredients - i.e., a mixture. In other
words, someone did the experiment in advance! What parts of the recipe did they vary to
make the recipe a success? Probably many factors, temperature and moisture, various
ratios of ingredients, and presence or absence of many additives. Now, should one keep
all the factors involved in the experiment at a constant level and just vary one to see what
would happen? This is a strategy that works but is not very efficient. This is one of the
concepts that we will address in this course.

Tasks for this Lesson

• Read Chapter 1 of the textbook

• Read the Online supplement [1] for Chapter 1
• Complete the homework for this lesson.

1.1 - A Quick History of the Design of

Experiments (DOE)
The textbook we are using brings an engineering perspective to the design of
experiments. We will bring in other contexts and examples from other fields of study
including agriculture (where much of the early research was done) education and nutrition.
Surprisingly the service industry has begun using design of experiments as well.

"All experiments are designed experiments, it is just that some are poorly designed and
some are well-designed."

Engineering Experiments

If we had infinite time and resource budgets there probably wouldn't be a big fuss made
over designing experiments. In production and quality control we want to control the error
and learn as much as we can about the process or the underlying theory with the
resources at hand. From an engineering perspective we're trying to use experimentation
for the following purposes:

• reduce time to design/develop new products & processes

• improve performance of existing processes

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/5/ 4/18/2019
 Page 3 of 8

• improve reliability and performance of products

• achieve product & process robustness
• perform evaluation of materials, design alternatives, setting component & system
tolerances, etc.

We always want to fine tune or improve the process. In today's global world this drive for
competitiveness affects all of us both as consumers and producers.

Robustness is a concept that enters into statistics at several points. At the analysis stage
robustness refers to a technique that isn't overly influenced by bad data. Even if there is
an outlier or bad data you still want to get the right answer. Regardless of who or what is
involved in the process - it is still going to work. We will come back to this notion of
robustness later in the course (Lesson 12).

Every experiment design has inputs. Back to the cake baking example: we have our
ingredients such as flour, sugar, milk, eggs, etc. Regardless of the quality of these
ingredients we still want our cake to come out successfully. In every experiment there are
inputs and in addition there are factors (such as time of baking, temperature, geometry of
the cake pan, etc.), some of which you can control and others that you can't control. The
experimenter must think about factors that affect the outcome. We also talk about the
output and the yield or the response to your experiment. For the cake, the output might be
measured as texture, flavor, height, size, or flavor.

Four Eras in the History of DOE

Here's a quick timeline:

• The agricultural origins, 1918 – 1940s

◦ R. A. Fisher & his co-workers
◦ Profound impact on agricultural science
◦ Factorial designs, ANOVA
• The first industrial era, 1951 – late 1970s
◦ Box & Wilson, response surfaces

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/5/ 4/18/2019
 Page 4 of 8

◦ Applications in the chemical & process industries

• The second industrial era, late 1970s – 1990
◦ Quality improvement initiatives in many companies
◦ CQI and TQM were important ideas and became management goals
◦ Taguchi and robust parameter design, process robustness
• The modern era, beginning circa 1990, when economic competitiveness and
globalization is driving all sectors of the economy to be more competitive.

Notes:

A lot of what we are going to learn in this course goes back to what Sir Ronald Fisher
developed in the UK in the first half of the 20th century. He really laid the foundation for
statistics and for design of experiments. He and his colleague Frank Yates developed
many of the concepts and procedures that we use today. Basic concepts such as
orthogonal designs and Latin squares began there in the 20's through the 40's. World War
II also had an impact on statistics, inspiring sequential analysis, which arose from World
War II as a method to improve the accuracy of long-range artillery guns.

Immediately following World War II the first industrial era marked another resurgence in
the use of DOE. It was at this time that Box and Wilson (1951) wrote the key paper in
response surface designs thinking of the output as a response function and trying to find
the optimum conditions for this function. George Box died early in 2013. And, an
interesting fact here - he married Fisher's daughter! He worked in the chemical industry in
England in his early career and then came to America and worked at the University of
Wisconsin for most of his career.

The Second Industrial Era - or the Quality Revolution

The importance of statistical quality control was

taken to Japan in the 1950's by W Edward
Deming. This started what Montgomery calls a
second Industrial Era, and sometimes the
quality revolution. After the second world war
Japanese products were of terrible quality.
They were cheaply made and not very good. In
the 1960s their quality started improving. The
Japanese car industry adopted statistical quality
control procedures and conducted experiments
which started this new era. Total Quality
Management (TQM), Continuous Quality
Improvement (CQI) are management
W. Edwards Deming
techniques that have come out of this statistical
quality revolution - statistical quality control and
design of experiments.

Taguchi, a Japanese engineer, discovered and published a lot of the techniques that were
later brought to the West, using an independent development of what he referred to as
orthogonal arrays. In the West these were referred to as fractional factorial designs. These
are both very similar and we will discuss both of these in this course. He came up with the
concept of robust parameter design and process robustness.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/5/ 4/18/2019
 Page 5 of 8

The Modern Era

Around 1990 Six Sigma, a new way of representing CQI, became popular. Now it is a
company and they employ a technique which has been adopted by many of the large
manufacturing companies. This is a technique that uses statistics to make decisions
based on quality and feedback loops. It incorporates a lot of the previous statistical and
management techniques.

Clinical Trials

Montgomery omits in this brief history a major part of design of experimentation that
evolved - clinical trials. This evolved in the 1960's when medical advances were previously
based on anecdotal data; a doctor would examine six patients and from this wrote a paper
and published it. The incredible biases resulting from these kinds of anecdotal studies
became known. The outcome was a move toward making the randomized double-blind
clinical trial the gold standard for approval of any new product, medical device, or
procedure. The scientific application of the statistical procedures became very important.

1.2 - The Basic Principles of DOE

The first three here are perhaps the most important...

Randomization - this is an essential component of any experiment that is going to have

validity. If you are doing a comparative experiment where you have two treatments, a
treatment and a control for instance, you need to include in your experimental process the
assignment of those treatments by some random process. An experiment includes
experimental units. You need to have a deliberate process to eliminate potential biases
from the conclusions, and random assignment is a critical step.

Replication - is some in sense the heart of all of statistics. To make this point...
Remember what the standard error of the mean is? It is the square root of the estimate of
−− 2
the variance of the sample mean, i.e., √ sn . The width of the confidence interval is
determined by this statistic. Our estimates of the mean become less variable as the
sample size increases.

Replication is the basic issue behind every method we will use in order to get a handle on
how precise our estimates are at the end. We always want to estimate or control the
uncertainty in our results. We achieve this estimate through replication. Another way we
can achieve short confidence intervals is by reducing the error variance itself. However,
when that isn't possible, we can reduce the error in our estimate of the mean by increasing
n.

Another way is to reduce the size or the length of the confidence interval is to reduce the
error variance - which brings us to blocking.

Blocking - is a technique to include other factors in our experiment which contribute to

undesirable variation. Much of the focus in this class will be to creatively use various

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/5/ 4/18/2019
 Page 6 of 8

blocking techniques to control sources of variation that will reduce error variance. For
example, in human studies, the gender of the subjects is often important factor. Age is
another factor affecting the response. Age and gender are often considered nuisance
factors which contribute to variability and make it difficult to assess systematic effects of a
treatment. By using these as blocking factors, you can avoid biases that might occur due
to differences between the allocation of subjects to the treatments, and as a way of
accounting for some noise in the experiment. We want the unknown error variance at the
end of the experiment to be as small as possible. Our goal is usually to find out something
about a treatment factor (or a factor of primary interest), but in addition to this we want to
include any blocking factors that will explain variation.

Multi-factor Designs - we will spend at least half of this course talking about multi-factor
experimental designs: 2k designs, 3k designs, response surface designs, etc. The point to
all of these multi-factor designs is contrary to the scientific method where everything is
held constant except one factor which is varied. The one factor at a time method is a very
inefficient way of making scientific advances. It is much better to design an experiment
that simultaneously includes combinations of multiple factors that may affect the outcome.
Then you learn not only about the primary factors of interest but also about these other
factors. These may be blocking factors which deal with nuisance parameters or they may
just help you understand the interactions or the relationships between the factors that
influence the response.

Confounding - is something that is usually considered bad! Here is an example. Let's say
we are doing a medical study with drugs A and B. We put 10 subjects on drug A and 10
on drug B. If we categorize our subjects by gender, how should we allocate our drugs to
our subjects? Let's make it easy and say that there are 10 male and 10 female subjects. A
balanced way of doing this study would be to put five males on drug A and five males on
drug B, five females on drug A and five females on drug B. This is a perfectly balanced
experiment such that if there is a difference between male and female at least it will
equally influence the results from drug A and the results from drug B.

An alternative scenario might occur if patients were randomly assigned treatments as they
came in the door. At the end of the study they might realize that drug A had only been
given to the male subjects and drug B was only given to the female subjects. We would
call this design totally confounded. This refers to the fact that if you analyze the difference
between the average response of the subjects on A and the average response of the
subjects on B, this is exactly the same as the average response on males and the
average response on females. You would not have any reliable conclusion from this study
at all. The difference between the two drugs A and B, might just as well be due to the
gender of the subjects, since the two factors are totally confounded.

Confounding is something we typically want to avoid but when we are building complex
experiments we sometimes can use confounding to our advantage. We will confound
things we are not interested in order to have more efficient experiments for the things we
are interested in. This will come up in multiple factor experiments later on. We may be
interested in main effects but not interactions so we will confound the interactions in this
way in order to reduce the sample size, and thus the cost of the experiment, but still have
good information on the main effects.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/5/ 4/18/2019
 Page 7 of 8

1.3 - Steps for Planning, Conducting and

Analyzing an Experiment
The practical steps needed for planning and conducting an experiment include:
recognizing the goal of the experiment, choice of factors, choice of response, choice of the
design, analysis and then drawing conclusions. This pretty much covers the steps
involved in the scientific method.

1. Recognition and statement of the problem

2. Choice of factors, levels, and ranges
3. Selection of the response variable(s)
4. Choice of design
5. Conducting the experiment
6. Statistical analysis
7. Drawing conclusions, and making recommendations

What this course will deal with primarily is the choice of the design. This focus includes all
the related issues about how we handle these factors in conducting our experiments.

Factors

We usually talk about "treatment" factors, which are the factors of primary interest to you.
In addition to treatment factors, there are nuisance factors which are not your primary
focus, but you have to deal with them. Sometimes these are called blocking factors,
mainly because we will try to block on these factors to prevent them from influencing the
results.

There are other ways that we can categorize factors:

Experimental vs. Classification Factors

Experimental Factors - these are factors that you can specify (and set the levels)
and then assign at random as the treatment to the experimental units. Examples
would be temperature, level of an additive fertilizer amount per acre, etc.

Classification Factors - can't be changed or assigned, these come as labels on the

experimental units. The age and sex of the participants are classification factors which
can't be changed or randomly assigned. But you can select individuals from these
groups randomly.

Quantitative vs. Qualitative Factors

Quantitative Factors - you can assign any specified level of a quantitative factor.
Examples: percent or pH level of a chemical.

Qualitative Factors - have categories which are different types. Examples might be
species of a plant or animal, a brand in the marketing field, gender, - these are not
ordered or continuous but are arranged perhaps in sets.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/5/ 4/18/2019
 Page 8 of 8

Think About It:

Think about your own field of study and jot down several of the factors that are
pertinent in your own research area? Into what categories do these fall?

Get statistical thinking involved early when you are preparing to design an experiment!
Getting well into an experiment before you have considered these implications can be
disastrous. Think and experiment sequentially. Experimentation is a process where what
you know informs the design of the next experiment, and what you learn from it becomes
the knowledge base to design the next.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/5

Links:
[1] https://bcs.wiley.com/he-bcs/Books?
action=chapter&bcsId=7219&itemId=1118146921&chapterId=79009

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/5/ 4/18/2019
 Page 1 of 9

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 2: Simple Comparative Experiments

Lesson 2: Simple Comparative

Experiments
Lesson 2: Introduction

This chapter should be a review for most students who have the required prerequisites.
We included it to focus the course and confirm the basics of understanding the
assumptions and underpinnings of estimation and hypothesis testing.

Learning objectives & outcomes

Goals for this lesson include the following:

• to review basic statistical concepts

• to review sample size calculation for two sample problems based on the t-test
• to review the difference between two independent samples and paired comparison
design
• to review the assumptions underlying the t-test and how to test for these
assumptions

2.1 - Simple Comparative Experiments

Simple comparative experiments are not only preliminary to this course but this takes you
back probably into your first course in statistics. We will look at both hypothesis testing
and estimation and from these perspectives we will look at sample size determination.

Two Sample Experiment

Here is an example from the text where there are two formulations for making cement
mortar. It is hard to get a sense of the data when looking only at a table of numbers. You
get a much better understanding of what it is about when looking at a graphical view of the
data.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 2 of 9

Dot plots work well to get a sense of the distribution. These work especially well for very
small sets of data.

Another graphical tool is the boxplot, useful for small or larger data sets. If you look at the
box plot you get a quick snapshot of the distribution of the data.

Remember that the box spans the middle 50% of the data (from the 25th to the 75th
percentile) and the whiskers extend as far out as the minimum and maximum of the data,
to a maximum of 1.5 times the width of the box, or 1.5 times the Interquartile range. So if
the data are normal you would expect to see just the box and whisker with no dots
outside. Potential outliers will be displayed as single dots beyond the whiskers.

This example is a case where the two groups are different in terms of the median, which
is the horizontal line in the box. One cannot be sure simply by visualizing the data if there

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 3 of 9

is a significant difference between the means of these two groups. However, both the box
plots and the dot plot hint at differences.

Testing: The two sample t-test

For the two sample t-test both samples are assumed to come from Normal populations
with (possibly different) means μi and variances σ2. When the variances are not equal we
will generally try to overcome this by transforming the data. Using a metric where the
variation is equal we can use complex ANOVA models, which also assume equal
variances. (There is a version of the two sample t-test which can handle different
variances, but unfortunately this does not extend to more complex ANOVA models.) We
want to test the hypothesis that the means μi are equal.

Our first look at the data above shows that the means are somewhat different but the
variances look to be about the same. We estimate the mean and the sample variance
using formulas:
n n
∑ yi ∑ (yi − ȳ)2
i=1 i=1
ȳ = and s2 =
n n−1
We divide by n - 1 so we can get an unbiased estimate of σ2. These are the summary
statistics for the two sample problem. If you know the sample size, n, the sample mean,
and the sample standard deviation (or the variance), these three quantities for each of the
two groups will be sufficient for performing statistical inference. However, it is dangerous
to not look at the data and only look at the summary statistics because these summary
statistics do not tell you anything about the shape or distribution of the data or about
potential outliers, both things you'd want to know about to determine if the assumptions
are satisfied.

The two sample t-test is basically looking at the difference between the sample means
relative to the standard deviation of the difference of the sample means. Engineers would
express this as a signal to noise ratio for the difference between the two groups.

If the underlying distributions are normal then the z-statistic is the difference between the
sample means divided by the true population variance of the sample means. Of course if
we do not know the true variances -- we have to estimate them. We therefore use the
t-distribution and substitute sample quantities for population quantities, which is something
we do frequently in statistics. This ratio is an approximate z-statistic -- Gosset published
the exact distribution under the psuedonym "Student" and the test is often called the
"Student t" test. If we can assume that the variances are equal, an assumption we will
make whenever possible, then we can pool or combine the two sample variances to get
the pooled standard deviation shown below.

Our pooled statistic is the pooled standard deviation sp times the square root of the sum of
the inverses of the two sample sizes. The t-statistic is a signal-to-noise ratio, a measure of
how far apart the means are for determining if they are really different.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 4 of 9

Does the data provide evidence that the true means differ? Let's test H0: μ1 = μ2

We will now calculate the test statistic, which is

ȳ 1 − ȳ 2
t= −−−−−−
Sp √ n11 + n12

This is always a relative question. Are they different relative to the variation within the
groups? Perhaps, they look a bit different. Our t-statistic turns out to be -2.19. If you know
the t-distribution, you should then know that this is a borderline value and therefore
requires that we examine carefully whether these two samples are really far apart.

We compare the sample t to the distribution with the appropriate d.f.. We typically will
calculate just the p-value which is the probability of finding the value at least as extreme
as the one in our sample. This is under the assumption of the null hypothesis that our
means are equal. The p-value in our example is essentially 0.043 as shown in the Minitab
output below.

Normal probability plots look reasonable.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 5 of 9

Confidence intervals involve finding an interval, in this case the interval is about the difference in
means. We want to find upper and lower limits that include the true difference in the means with a
specified level of confidence, typically we will use 95%.

In the cases where we have a two-sided hypothesis test which rejects the null hypothesis,
then the confidence interval will not contain 0. In our example above we can see in the
Minitab output that the 95% confidence interval does not include the value 0, the
hypothesized value for the difference, when the null hypothesis assumes the two means
are equal.

2.2 - Sample Size Determination

The estimation approach to determining sample size addresses the question: "How
accurate do you want your estimate to be?" In this case we are estimating the difference
in means. This approach requires us to specify how large a difference we are interested in
detecting, say B for the Bound on the margin of error, and then to specify how certain we
want to be that we can detect a difference that large. Recall that when we assume equal
sample sizes of n, a confidence interval for μ1- μ2 is given by:
−−
¯ ¯ 2
{Y 1 − Y 2 ± t(1 − α/2; df) ⋅ s ⋅ √ }
n
Where n is the sample size for each group, and df = n + n - 2 = 2(n - 1) and s is the pooled
standard deviation. Therefore, we first specify B and then solve this equation:
−−
2
B = t(1 − α/2; df) ⋅ s ⋅ √
n
for n. Therefore,
– 2
√2 t2 (1 − α/2; df) ⋅ s2 ⋅ 2
n = [t(1 − α/2; df) ⋅ s ⋅ ] =[ ]
B B2
Since in practice, we don't know what s will be, prior to collecting the data, we will need a
guesstimate of σ to substitute into this equation. To do this by hand and we use z rather
than t since we don't know the df if we don't know the sample size n - the computer will
iteratively update the d.f. as it computes the sample size, giving a slightly larger sample
size when n is small.

So we need to have an estimate of σ2, a desired margin of error bound B, that we want to
detect, and a confidence level 1-α. With this we can determine sample size in this
comparative type of experiment. We may or may not have direct control over σ2, but by
using different experimental designs we do have some control over this and we will
address this later in this course. In most cases an estimate of σ2 is needed in order to
determine the sample size.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 6 of 9

One special extension of this method is when we have a binomial situation. In

this case where we are estimating proportions rather than some quantitative
mean level, we know that the worst-case variance, p(1-p), is where p (the true
proportion) is equal to 0.5 and then we would have an approximate sample
size formula that is simpler, namely n = 2/B2 for α = 0.05.

Another Two-Sample Example – Paired Samples

In the paired sample situation, we have a group of subjects where each subject has two
measurements taken. For example, blood pressure was measured before and after a
treatment was administered for five subjects. These are not independent samples, since
for each subject, two measurements are taken, which are typically correlated – hence we
call this paired data. If we perform a two sample independent t-test, ignoring the pairing for
the moment we lose the benefit of the pairing, and the variability among subjects is part of
the error. By using a paired t-test, the analysis is based on the differences (after – before)
and thus any variation among subjects is eliminated.

In our Minitab output we show the example with Blood Pressure on five subjects.

By viewing the output, we see that the different patients' blood pressures seem to vary a
lot (standard deviation about 12) but the treatment seems to make a small but consistent
difference with each subject. Clearly we have a nuisance factor involved - the subject -
which is causing much of this variation. This is a stereotypical situation where because the
observations are correlated and paired and we should do a paired t-test.

These results show that by using a paired design and taking into account the pairing of the
data we have reduced the variance. Hence our test gives a more powerful conclusion
regarding the significance of the difference in means.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 7 of 9

The paired t-test is our first example of a blocking design. In this context the subject is
used as a block, and the results from the paired t-test are identical to what we will find
when we analyze this as a Randomize Complete Block Design from lesson 4.

2.3 - Determining Power

We begin this part by defining the power of a hypothesis test. This also provides another
way of determining the sample size. The power is the probability of achieving the desired
outcome. What is the desired outcome of a hypothesis test? Usually rejecting the null
hypothesis. Therefore, power is the probability of rejecting the null hypothesis when in fact
the alternative hypothesis is true.

Decision HO HA
Reject Null
Type I Error - α OK
Hypothesis
Accept Null
OK Type II Error - β
Hypothesis

Note:

P(Reject H0 |H0 is true) = α: P(Type I Error)

P(Accept H0 | HA is true) = β: P(Type II Error)

Therefore the power of the test is P(Reject H0 | HA is true) = 1-β.

Before any experiment is conducted you typically want to know how many observations
you will need to run. If you are performing a study to test a hypothesis, for instance in the
blood pressure example where we are measuring the efficacy of the blood pressure
medication, if the drug is effective there should be a difference in the blood pressure
before and after the medication. Therefore we want to reject our null hypothesis, and thus
we want the power (i.e. the probability of rejecting the HO when it is false) to be as high as
possible.

We will describe an approach to determine the power, based on a set of operating

characteristic curves traditionally used in determining power for the t-test. Power depends
on the level of the test, α , the actual true difference in means, and n (the sample size).
Figure 2.13 (2.12 in 7th ed) in the text gives the operating characteristic curves where β is
calculated for n* = 2n - 1 for an α = 0.05 level test. When you design a study you usually
plan for equal sample size, since this gives the highest power in your results. We will look
at special cases where you might deviate from this but generally this is the case.

To use the Figure in the text, we need to first calculate the difference the difference in
means measured in numbers of standard deviation, i.e. |μ1 - μ2| / σ. You can think of
this as a signal to noise ratio, i.e. how large or strong is the signal, |μ1 - μ2| , in relation to
the variation in the measurements, σ. We are not using the symbols in the text, because

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 8 of 9

the 2 editions define d and δ differently. Different software packages or operating

characteristic curves may require either |μ1 - μ2| / σ or |μ1 - μ2| / 2σ to compute sample
sizes or estimate power, so you need to be careful in reading the documentation. Minitab
avoids this by asking for |μ1 - μ2| and σ separately, which seems like a very sensible
solution.

Again,

Example calculations: Let's consider an example in the two sample situation.

We will let α = .05, |μ1 - μ2| = 8 (the difference between the two means), and
the sigma (assumed true standard deviation) would equal 12, and finally, let
the number of observations in each group n = 5.

In this case, |μ1 - μ2|/σ = 8/12 = .66, and n* = 2n - 1 = 9.

If you look at the Figure you get approximately a β of about 0.9. Therefore,
power - or the chance of rejecting the null hypothesis prior to doing the
experiment is 1 - β or 1 - 0.9 = 0.1 or about ten percent of the time. With such
low power we should not even do the experiment!

If we were willing to do a study that would only detect a true difference of, let's
say, |μ1 - μ2| = 18 then and n* would still equal 9, then figure 2-12 the Figure
shows that β looks to be about .5 and the power or chance of detecting a
difference of 18 is also 5. This is still not very satisfactory since we only have a
50/50 chance of detecting a true difference of 18 even if it exists.

Finally, we calculate the power to detect this difference of 18 if we were to use

n = 10 observations per group, which gives us n* = 19. For this case β = 0.1
and thus power = 1- β = 0.9 or 90%, which is quite satisfactory.

These calculations can also be done in Minitab as shown below. Under the
Menu: Stat > Power and Sample Size > 2-sample t, simply input sample sizes,
n = 10, differences δ = 18, and standard deviation σ = 12.

Another way to improve power is to use a more efficient procedure - for example if we
have paired observations we could use a paired t-test. For instance, if we used the paired
t-test, then we would expect to have a much smaller sigma – perhaps somewhere around
2 rather than 12. So, our signal to noise ratio would be larger because the noise
component is smaller. We do pay a small price in doing this because our t-test would now
have degrees of freedom n - 1, instead of 2n - 2.

The take-home message here is:

If you can reduce variance or noise, then you can achieve an incredible savings in the
number of observations you have to collect. Therefore the benefit of a good design is to
get a lot more power for the same cost or much decreased cost for the same power.

We now show another approach to calculating power, namely using software tools rather
than the graph in Figure 2.12. Let's take a look at how Minitab handles this below.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 9 of 9

You can use these dialog boxes to plug in the values that you have assumed and have
Minitab calculate the sample size for a specified power, or the power that would result, for
a given sample size.

Exercise: Use the assumptions above, and confirm the calculations of power for these
values.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/8

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/8/ 4/18/2019
 Page 1 of 20

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 3: Experiments with a Single Factor - the Oneway ANOVA - in the Completely Randomized Design (CRD)

Lesson 3: Experiments with a Single Factor -

the Oneway ANOVA - in the Completely
Randomized Design (CRD)
Lesson 3: Introduction

By the end of this chapter we will understand how to proceed when the ANOVA tells us that the mean responses
differ, (i.e., the levels are significantly different), among our treatment levels. We will also briefly discuss the
situation that the levels are a random sample from a larger set of possible levels, such as a sample of brands for a
product. (Note that this material is in Chapter 3.9 of the 8th edition and Chapter 13.1 of the 7th edition.) We will
briefly discuss multiple comparison procedures for qualitative factors, and regression approaches for quantitative
factors. These are covered in more detail in the STAT 502 course, and discussed only briefly here.

Learning objectives & outcomes

We focus more on the design and planning aspects of these situations:

• How many observations do we need?

◦ to achieve a desired precision when the goal is estimating a parameter, and
◦ to achieve a desired level of power when hypothesis testing.
• Which multiple comparison procedure is appropriate for your situation?
• How should we allocate our observations among the k treatment groups? Usually equally, but
the Dunnett Test situation has a different optimum allocation.
• The last section describes the F-test as an example of the General Linear Test.

3.1 - Experiments with One Factor and Multiple

Levels
Lesson 3 is the beginning of the one-way analysis of variance part of the course, which extends the two
sample situation to k samples. In addition to these notes, read Chapter 3 of the text and the on-line
supplement. (If you have the 7th edition, also read 13.1.)

We review the issues related to a single factor experiment, which we see in the context of a Completely
Randomized Design (CRD). In a single factor experiment with a CRD the levels of the factor are
randomly assigned to the experimental units. Alternatively, we can think of randomly assigning the
experimental units to the treatments or in some cases, randomly selecting experimental units from
each level of the factor.

Example - Cotton Tensile Strength

Let's take a look at an Example, taken from Problem 3.10 of

Montgomery
Typesetting math:(3.8
100%in the 7th edition).

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 2 of 20

This is an investigation into the formulation of synthetic fibers

that are used to make cloth. The response is tensile strength, the
strength of the fiber. The experimenter wants to determine the
best level of the cotton in terms of percent, to achieve the
highest tensile strength of the fiber. Therefore, we have a single
quantitative factor, the percent of cotton combined with synthetic
fabric fibers.

The five treatment levels of percent cotton are evenly spaced from 15% to 35%. We have five
replicates, five runs on each of the five cotton weight percentages.

The box plot of the results shows an indication that there is an increase in strength as you increase the
cotton and then it seems to drop off rather dramatically after 30%.

Makes you wonder about all of those 50% cotton shirts that you buy?!

The null hypothesis asks: does the cotton percent make a difference? Now, it seems that it doesn't take
statistics to answer this question. All we have to do is look at the side by side box plots of the data and
there appears to be a difference – however this difference is not so obvious by looking at the table of
raw data. A second question, frequently asked when the factor is quantitative: what is the optimal level
of cotton if you only want to consider strength?

There is a point that I probably should emphasize now and repeatedly throughout this course. There is
often more than one response measurement that is of interest. You need to think about
multiple responses in any given experiment. In this experiment, for some reason, we are interested in
only one response, tensile strength, whereas in practice the manufacturer would also consider
comfort, ductility, cost, etc.

This single factor experiment can be described as a completely randomized design (CRD). The
completely randomized design means there is no structure among the experimental units. There are 25
runs which differ only in the percent cotton, and these will be done in random order.If there were
different machines
Typesetting math: 100% or operators, or other factors such as the order or batches of material, this would

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 3 of 20

need to be taken into account. We will talk about these kinds of designs later. This is an example of a
completely randomized design where there are no other factors that we are interested in other than the
treatment factor percentage of cotton.

Analysis of Variance

The Analysis of Variance (ANOVA) is a somewhat misleading name for this procedure. But we call it
the analysis of variance because we are partitioning the total variation in the response measurements.

The Model Statement

Each measured response can be written as the overall mean plus the treatment effect plus a random
error.

Yij = μ + τi + ϵij
i = 1, ... , a, j = 1, ... , ni

Generally we will define our treatment effects so that they sum to 0, a constraint on our definition of our
parameters, ∑ τi = 0. This is not the only constraint we could choose, one treatment level could be a
reference such as the zero level for cotton and then everything else would be a deviation from that.
However, generally we will let the effects sum to 0. The experimental error terms are assumed to be
normally distributed, with zero mean and if the experiment has constant variance then there is a single
variance parameter σ2. All of these assumptions need to be checked. This is called the effects model.

An alternative way to write the model, besides the effects model, where the expected value of our
observation, E(Yij) = μ + τi or an overall mean plus the treatment effect. This is called the means model
and is written as:

Yij = μ + ϵij
i = 1, ... , a, and j = 1, ... , ni.

In looking ahead there is also the regression model. Regression models can also be employed but for
now we consider the traditional analysis of variance model and focus on the effects of the treatment.

Analysis of variance formulas that you should be familiar with by now are provided in the textbook,
(Section 3.3).

The total variation is the sum of the observations minus the overall mean squared, summed over all a ×
n observations.

The analysis of variance simply takes this total variation and partitions it into the treatment component
and the error component. The treatment component is the difference between the treatment mean and
the overall mean. The error component is the difference between the observations and the treatment
mean, i.e. the variation not explained by the treatments.

Notice when you square the deviations there are also cross product terms, (see equation 3-5), but
these sum to zero when you sum over the set of observations. The analysis of variance is the partition
of the total variation into treatment and error components. We want to test the hypothesis that the
means are equal versus at least one is different, i.e.

Ho: μ1 = … = μa versus Ha: μi ≠ μi’ for some i, i’

Corresponding to the sum of squares (SS) are the degrees of freedom associated with the treatments,
a - 1, and the degrees of freedom associated with the error, a × (n - 1), and finally one degree of
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 4 of 20

freedom is due to the overall mean parameter. These add up to the total N = a × n, when the ni are all
equal to n, or N = ∑ ni otherwise.

The mean square treatment (MST) is the sum of squares due to treatment divided by its degrees of
freedom.

The mean square error (MSE) is the sum of squares due to error divided by its degrees of freedom.

If the true treatment means are equal to each other, i.e. the μi are all equal, then these two quantities
should have the same expectation. If they are different then the treatment component, MST will be
larger. This is the basis for the F-test.

The basic test statistic for testing the hypothesis that the means are all equal is the F ratio, MST/MSE,
with degrees of freedom, a-1 and a×(n-1) or a-1 and N-a.

We reject H0 if this quantity is greater than 1-α percentile of the F distribution.

Back to the example - Cotton Weight Percent

Here is the Analysis of Variance table from the Minitab output:

Note a very large F statistic that is, 14.76. The p-value for this F-statistic is < .0005 which is taken from
an F distribution pictured below with 4 and 20 degrees of freedom.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 5 of 20

We can see that most of the distribution lies between zero and about four. Our statistic, 14.76, is far out
in the tail, obvious confirmation about what the data show, that indeed the means are not the same.
Hence, we reject the null hypothesis.

Model Assumption Checking

We should check if the data are normal - they should be approximately normal - they should certainly
have constant variance among the groups. Independence is harder to check but plotting the residuals
in the order in which the operations are done can sometimes detect if there is lack of independence.
The question in general is how do we fit the right model to represent the data observed. In this case
there's not too much that can go wrong since we only have one factor and it is a completely
randomized design. It is hard to argue with this model.

Let's examine the residuals, which are just the observations minus the predicted values, in this case
treatment means. Hence, eij = yij − ȳ i .

These plots don't look exactly normal but at least they don't seem to have any wild outliers. The normal
scores plot looks reasonable. The residuals versus the order of the data plot are a plot of the error
residuals data in the order in which the observations were taken. This looks a little suspect in that the
first six data points all have small negative residuals which are not reflected in the following data points.
This looks like it might be a start up problem? These are the kinds of clues that you look for... if you are
conducting this experiment you would certainly want to find out what was happening in the beginning.

Post-ANOVA Comparison of Means

So, we found the means are significantly different. Now what? In general, if we had a qualitative factor
rather than a quantitative factor we would want to know which means differ from which other ones. We
would probably want to do t-tests or Tukey maximum range comparisons, or some set of contrasts to
examine the differences in means. There are many multiple comparison procedures.

Two methods in particular are Fisher's Least Significant Difference (LSD), and the Bonferroni Method.
Both of these are based on the t-test. Fisher's LSD says do an F-test first and if you reject the null
hypothesis, then just do ordinary t-tests between all pairs of means. The Bonferroni method is similar,
but only requires that you decide in advance how many pairs of means you wish to compare, say g,
and then perform the g t-tests with a type I level of α / g. This provides protection for the entire family of
g tests that the type I error is no more than α. For this setting, with a treatments, g = a(a-1)/2 when
comparing all pairs of treatments.

All of these multiple comparison procedures are simply aimed at interpreting or understanding the
overall F-test
Typesetting math: --which
100% means are different? They apply to many situations especially when the factor is

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 6 of 20

qualitative. However, in this case, since cotton percent is a quantitative factor, doing a test between two
arbitrary levels e.g. 15% and 20% level, isn't really what you want to know. What you should focus on is
the whole response function as you increase the level of the quantitative factor, cotton percent.

Whenever you have a quantitative factor you should be thinking about modeling that relationship with a
regression function.

Review the video that demonstrates the use of polynomial regression to help explain what is going on.

Here is the Minitab output where regression was applied:

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 7 of 20

Here is a link to the Cotton Weight % dataset (cotton_weight.MTW [1]). Open this in Minitab so that you
can try this yourself.

You can see that the linear term in the regression model is not significant but the quadratic is highly
significant. Even the cubic term is significant with p-value = 0.015. In Minitab we can plot this
relationship in the fitted line plot as seen below:

This shows the actual fitted equation. Why wasn't the linear term significant? If you just fit a straight line
to this data it would be almost flat, not quite but almost. As a result the linear term by itself is not
significant. We should still leave it in the polynomial regression model however, because we like to
have a hierarchical model when fitting polynomials. What we can learn from this model is that tensile
strength of cotton is probably best between the 25 and 30 weight.

This is a more focused conclusion than we get from simply comparing the means of the actual levels in
the experiment because the polynomial model reflects the quantitative relationship between the
treatment and the response.

We should also check whether the observations have constant variance σ2, for all treatments. If they
are all equal we can say that they are equal to σ2. This is an assumption of the analysis and we need to
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 8 of 20

confirm this assumption. We can either test it with the Bartlett's test, the Levene's test, or simply use
the 'eye ball' technique of plotting the residuals versus the fitted values and see if they are roughly
equal. The eyeball approach is almost as good as using these tests, since by testing we cannot ‘prove’
the null hypothesis.

Bartlett's test is very susceptible to non-normality because it is based on the sample variances, which
are not robust to outliers. (See Section 3.4 in the text.) We must assume that the data are normally
distributed and thus not very long-tailed. When one of the residuals is large and you square it, you get
a very large value which explains why the sample variance is not very robust. One or two outliers can
cause any particular variance to be very large. Thus simply looking at the data in a box plot is as good
as these formal tests. If there is an outlier you can see it. If the distribution has a strange shape you
can also see this in a histogram or a box plot. The graphical view is very useful in this regard.

Levene's test is preferred to Bartlett’s in my view, because it is more robust. To calculate the Levene's
test you take the observations and obtain (not the squared deviations from the mean but) the absolute
deviations from the median. Then, you simply do the usual one way ANOVA F-test on these absolute
deviations from the medians. This is a very clever and simple test that has been around for a long time,
created by Levene back in the 1950's. (See 3.4 in the text.) It is much more robust to outliers and non-
normality than Bartlett's test.

3.2 - Sample Size Determination

An important aspect of designing an experiment is to know how many observations are needed to
make conclusions of sufficient accuracy and with sufficient confidence. We review what we mean by
this statement. The sample size needed depends on lots of things; including what type of experiment is
being contemplated, how it will be conducted, resources, and desired sensitivity and confidence.

Sensitivity refers to the difference in means that the experimenter wishes to detect, i.e., sensitive
enough to detect important differences in the means.

Generally, increasing the number of replications increases the sensitivity and makes it easier to
detect small differences in the means. Both power and the margin of error are a function of n and a
function of the error variance. Most of this course is about finding techniques to reduce this
unexplained residual error variance, and thereby improving the power of hypothesis tests, and reducing
the margin of error in estimation.

Hypothesis Testing Approach to Determining Sample Size

Our usual goal is to test the hypothesis that the means are equal, versus the alternative that the means
are not equal.

The null hypothesis that the means are all equal implies that the τi's are all equal to 0. Under this
framework we want to calculate the power of the F-test in the fixed effects case.

Example - Blood Pressure

Consider the situation where we have four treatment groups that will be using four different blood
pressure drugs, a = 4. We want to be able to detect differences between the mean blood pressure for
the subjects after using these drugs.

One possible scenario is that two of the drugs are effective and two are not. e.g. say two of them result
in blood pressure at 110 and two of them at 120. In this case the sum of the τi2 for this situation is 100,
i.e. τi = (-5, -5, 5, 5) and thus Σ τi2 = 100.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 9 of 20

Another scenario is the situation where we have one drug at 110, two of them at 115 and one at 120. In
this case the sum of the τi2 is 50, i.e. τi = (-5, 0, 0, 5) and thus Σ τi2 = 50.

Considering both of these scenarios, although there is no difference between the minimums and the
maximums, the quantities Σ τi2 are very different.

Of the two scenarios, the second is the least favorable configuration (LFC). It is the configuration of
means for which you get the least power. The first scenario would be much more favorable. But
generally you do not know which situation you are in. The usual approach is to not to try guess exactly
what all the values of the τi will be but simply to specify δ, which is the maximum difference between
the true means, or δ = max(τi) – min(τi).

Going back to our LFC scenario we can calculate this again using Σ τi2 = δ2/2, i.e. the maximum
difference squared over 2. This is true for the LFC for any number of treatments, since Σ τi2 = (δ/2)2 × 2
= δ2/2 since all but the extreme values of τi are zero under the LFC.

The Use of Operating Characteristic Curves

The OC curves for the fixed effects model are given in the Appendix V.

The usual way to use these charts is to define the difference in the means, δ = max (μi) - min (μi), that
you want to detect, specify the value of σ2, and then for the LFC use :

nδ 2
Φ2 =
2aσ 2
for various values of n. The Appendix V gives β, where 1 - β is the power for the test where ν1 = a - 1
and ν2 = a(n - 1). Thus after setting n, you must calculate ν1 and ν2 to use the table.

Example: We consider an α = 0.05 level test for a = 4 using δ = 10 and σ2 = 144 and we want to find
the sample size n to obtain a test with power = 0.9.

Let's guess at what our n is and see how this work. Say we let n be equal to 20, let δ = 10, and σ = 12
then we can calculate the power using Appendix V. Plugging in these values to find Φ we get Φ = 1.3.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 10 of 20

Now go to the chart where ν2 is 80 - 4 = 76 and Φ = 1.3. This gives us a Type II error of β = 0.45 and
power = 1 - β = 0.55.

It seems that we need a larger sample size.

Well, let's use a sample size of 30. In this case we get Φ2 = 2.604, so Φ = 1.6.

Now with ν2 a bit more at 116, we have β = 0.30 and power = 0.70.

So we need a bit more than n = 30 per group to achieve a test with power = 0.8.

Review the video below for a 'walk-through' this procedure using Appendix V in the back of the text.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 11 of 20

3.3 - Multiple Comparisons

Scheffé's method for investigating all possible contrasts of the means corresponds exactly to the
F-test in the following sense. If the F-test rejects the null hypothesis at level α, then there exists at least
one contrast which would be rejected using the Scheffé procedure at level α . Therefore, Scheffé
provides α level protection against rejecting the null hypothesis when it is true, regardless of how many
contrasts of the means are tested.

Fisher’s LSD -- which is the F test, followed by ordinary t-tests among all pairs of means, but only if
the F-test rejects the null hypothesis. The F-test provides the overall protection against rejecting Ho
when it is true. The t-tests are each performed at α level and thus likely will reject more than they
should, when the F-test rejects. A simple example may explain this statement: assume there are eight
treatment groups, and one treatment has a mean higher than the other seven, which all have the same
value, and the F-test will rejects Ho. However, when following up with the pairwise t-tests, the 7 × 6 / 2
= 21 pairwise t-tests among the seven means which are all equal, will by chance alone reject at least
one pairwise hypothesis, Ho: μi = μi' at α = 0.05. Despite this drawback Fisher's LSD remains a favorite
method since it has overall α level protection, and offers simplicity to understand and interpret.

Bonferroni method for g comparisons – use α / g instead of α for testing each of the g comparisons.

Comparing the Bonferroni Procedure with the Fishers LSD

Fishers’s LSD method is an alternative to other pairwise comparison methods (for post ANOVA
analysis). This method controls the α-level error rate for each pairwise comparison so it does not
control the family error rate. This procedure uses the t statistic for testing Ho : μi = μj for all i and j pairs.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 12 of 20

ȳ i − ȳ j
t= −− −−−−− −−− −− rate, by performing the pairwise
Alternatively, the Bonferroni method does control
√MSE( the family
1
+ error
1
)
ni nj
comparison tests using α/g level of significance, where g is the number of pairwise comparisons.
Hence, the Bonferroni confidence intervals for differences of the means are wider than that of Fisher’s
LSD. In addition, it can be easily shown that the p-value of each pairwise comparison calculated by
Bonferroni method is g times the p-value calculated by Fisher’s LSD method.

Tukey’s Studentized Range considers the differences among all pairs of means divided by the
estimated standard deviation of the mean, and compares them with the tabled critical values provided
in Appendix VII. Why is it called the studentized range? The denominator uses an estimated standard
deviation, hence, the statistic is studentized like the student t-test. The Tukey procedure assumes all ni
are equal say to n.

ȳ i − ȳ j
q= −−−−−−−−
√MSE( n1 )

Comparing the Tukey Procedure with the Bonferroni Procedure

The Bonferroni procedure is a good all around tool, but for all pairwise comparisons the Tukey
studentized range procedure is slightly better as we show here.

The studentized range is the distribution of the difference between the maximum and a minimum over
the standard error of the mean. When we calculate a t-test, or when we're using the Bonferroni
adjustment where g is the number of comparisons, we are not comparing apples and oranges. In one
case (Tukey) the statistic has a denominator with the standard error of a single mean and in the other
case (t-test) with the standard error of the difference between means as seen in the equation for t and
q above.

Example - Tukey vs. Bonferroni approaches

Here is an example we can work out. Let's say we have 5 means, so a = 5, we will let α = 0.05, and the
total number of observations N = 35, so each group has seven observations and df = 30.

If we look at the studentized range distribution for 5, 30 degrees of freedom, (the distribution can be
found in Appendix VII, p. 630.), we find a critical value of 4.11.

If we took a Bonferroni approach - we would use g = 5 × 4 / 2 = 10 pairwise comparisons since a = 5.

Thus, again for an α = 0.05 test all we need to look at is the t-distribution for α / 2g = 0.0025 and N - a
=30 df . Looking at the t-table (found in Appendix II, p. 614) we get the value 3.03. However, to
compare with the Tukey Studentized Range statistic, we need to multiply the tabled critical value by
–
√2 = 1.414 , therefore 3.03 x1.414 = 4.28, which is slightly larger than the 4.11 obtained for the
Tukey table.

The point that we want to make is that the Bonferroni procedure is slightly more conservative than the
Tukey result, since the Tukey procedure is exact in this situation whereas Bonferroni only approximate.

The Tukey's procedure is exact for equal samples sizes. However, there is an approximate procedure
called the Tukey-Kramer test for unequal ni .

If you are looking at all pairwise comparisons then Tukey's exact procedure is probably the best
procedure to use.
Typesetting math: 100%The Bonferroni, however, is a good general procedure.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 13 of 20

Contrasts of Means

A pairwise comparison is just one example of a contrast of the means. A general contrast can be
written as a set of coefficients of the means that sum to zero. This will often involve more than just a
pair of treatments. In general we can write a contrast to make any comparison we like. We will also
consider sets of orthogonal contrasts.

Example - Gas Mileage

We want to compare the gas mileage on a set of cars: Ford Escape (hybrid), Toyota Camry, Toyota
Prius (hybrid), Honda Accord, and the Honda Civic (hybrid). A consumer testing group wants to test
each of these cars for gas mileage under certain conditions. They take n prescribed test runs and
record the mileage for each vehicle.

Now they first need to define some contrasts among these means. Contrasts are the coefficients which
provide a comparison that is meaningful. Then they can test and estimate these contrasts. For the first
contrast, C1, they could compare the American brand to the foreign brands. We need each contrast to
sum to 0, and for convenience only use integers. How about comparing Toyota to Honda (that is C2),
or hybrid compared to non-hybrid (that is C3).

Ford Toyota Toyota Honda Honda

Escape Camry Prius Accord Civic
Y1. Y2. Y3. Y4. Y5.
C1 4 -1 -1 -1 -1
C2 0 -1 -1 1 1
C3 2 -3 2 -3 2
C4 0 -1 1 0 0
C5 0 0 0 -1 1

So the first three contrast coefficients would specify the comparisons described, and the C4 and C5 are
comparisons within the brands with two models.

After we develop a set of contrasts, we can then test these contrasts or we can estimate them. We can
also calculate a confidence intervals around the true contrast of the means by using the estimated
contrast ± the t-distribution times the estimated standard deviation of the contrast. See equation 3-30
in the text.

Concerning Sets of Multiple Contrasts

Scheffé’s Method provides α-level protection for all possible contrasts - especially useful when we don't
really know how many contrasts we will have in advance. This test is quite conservative, because this
test is valid for all possible contrasts of the means. Therefore the Scheffé procedure is equivalent to the
F-test, and if the F-test rejects, there will be some contrast that will not contain zero in its confidence
interval.

What is an orthogonal contrast?

Two contrasts are orthogonal if the sum of the product of the coefficients of the two contrasts sum to
zero. An orthogonal set of contrasts are also orthogonal to the overall mean, since the coefficients sum
to zero. See Section 3.5.4 and 3.5.5 of the text.

Look at the
Typesetting table
math: above
100% and locate which contrasts are orthogonal.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 14 of 20

There always exists a-1 orthogonal contrasts of a means. When the sample sizes are equal, the sum
of squares for these contrasts, when added up, total the sum of squares due to treatment. Any set of
orthogonal contrasts partition the variation such that the total variation corresponding to those a-1
contrasts equals the total sum of squares among treatments. When the sample sizes are not equal,
the definition of orthogonal contrasts involves the sample sizes - this is explained in Section 3.5.5

Dunnett's Procedure

Dunnett’s procedure is another multiple comparison procedure specifically designed to compare each
treatment to a control. If we have a groups, let the last one be a control group and the first a - 1 be
treatment groups. We want to compare each of these treatment groups to this one control. Therefore,
we will have a - 1 contrasts, or a - 1 pairwise comparisons. To perform multiple comparisons on these a
- 1 contrasts we use special tables for finding hypothesis test critical values, derived by Dunnett.
Section 3.5.8 in the text and compare the test statistics di for i - 1, … , a - 1.

Comparing Dunnett’s procedure to the Bonferroni procedure

We can compare the Bonferroni approach to the Dunnett procedure. The Dunnett procedure calculates
the difference of means for the control versus treatment one, control versus treatment two, etc. to a - 1.
Which provides a - 1 pairwise comparisons.

So, we now consider an example where we have six groups, a = 6, and t = 5 and n = 6 observations
per group. Then, Dunnett's procedure will give the critical point for comparing the difference of means.
From the table in the appendix, VIII, we get α=0.05 two-sided comparison d(a-1, f) = 2.66, where a - 1 =
5 and f = df = 30.

Using the Bonferroni approach, if we look at the t-distribution for g = 5 comparisons and a two-sided
test with 30 degrees of freedom for error we get 2.75.

Comparing the two, we can see that the Bonferroni approach is a bit more conservative. The Dunnett's
is an exact procedure for comparing a control to a-1 treatments. Bonferroni is a general tool but not
exact. However, there is not much of a difference in this example

Fisher's LSD has the practicality of always using the same measuring stick, the unadjusted t-test. Every
one knows that if you do a lot of these tests, that for every 20 tests you do, that one could be wrong by
chance. This is another way to handle this uncertainty. All of these methods are protecting you from
making too many Type I errors whether you are either doing hypothesis testing or confidence intervals.
In your lifetime how many tests are you going to do?

So in a sense you have to ask yourself the question what is the set of tests that I want to protect
against making a Type I error. So, in Fisher's LSD procedure each test is standing on its own and is not
really a multiple comparisons test. If you are looking for any type of difference and you don't know how
many you are going to end up doing, you should probably using Scheffé as to protect you against all of
them. But if you know it is all pairwise and that is it, then Tukey's would be best. If you're comparing a
bunch of treatments against a control then Dunnett's would be best.

There is a whole family of step-wise procedures which are now available, but we will not consider them
here. Each can be shown to be better in certain situations. Another approach to this problem is called
False Discovery Rate control. It is used when there are hundreds of hypotheses - a situation that
occurs for example in testing gene expression of all genes in an organism, or differences in pixel
intensities for pixels in a set of images. The multiple comparisons procedures discussed above all
guard against the probability of making one false significant call. But when there are hundreds of tests,
we might prefer to make a few false significant calls if it greatly increases our power to detect true
difference. False Discovery Rate methods attempt to control the expected percentage of false
significant calls among the tests declared significant.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 15 of 20

3.4 - The Optimum Allocation for the Dunnett

Test
The Dunnett test for comparing means is a multiple comparison procedure, but is precisely designed to
test t treatments against a control.

We compared the Dunnett test to the Bonferroni - and there was only a slight difference, reflecting the
fact that the Bonferroni procedure is an approximation. This is a situation where we have a = t + 1
groups; a control group and t treatments.

I like to think of an example where we have a standard therapy, (a control group), and we want to test t
new treatments to compare them against the existing acceptable therapy. This is a case where we are
not so much interested in comparing each of the treatments against each other, but instead we are
interested in finding out whether each of the new treatments are better than the original control
treatment.

We have Yij distributed with mean μi , and variance σ2, where i = 1, ... , t, and j = 1, ... , ni for the t
treatment groups and a control group with mean μ0 with variance σ2.

We are assuming equal variance among all treatment groups.

The text describes the Dunnett test in Section 3.5.8.

The question that I want to address here is the design question.

The Dunnett procedure is based on t comparisons for testing Ho that μi = μ0, for i = 1, ... , t. This is
really t different tests where t = a - 1.

The Ha is that the μi are not equal to μ0.

Or viewing this as an estimation problem, we want to estimate the t differences μi - μ0.

How Should We Allocate Our Observations?

This is the question we are trying to answer. We have a fixed set of resources and a budget that only
allows for only N observations. So, how should we allocate our resources?

Should we assign half to the control group and the rest spread out among the treatments? Or, should
we assign an equal number of observations among all treatments and the control? Or what?

We want to answer this question by seeing how we can maximize the power of these tests with the N
observations that we have available. We approach this using an estimation approach where we want to
estimate the t differences μi - μ0. Let's estimate the variance of these differences.

What we want to do is minimize the total variance. Remember that the variance of (ȳ i − ȳ 0 ) is σ2 / ni
+ σ2 / n0. The total variance is the sum of these t parts.

We need to find n0, and ni that will minimize this total variance. However, this is subject to a constraint,
the constraint being that N = n0 + (t × n), if the ni = n for all treatments, an assumption we can
reasonably make when all treatments are of equal importance.

Given N observations and a groups, where a = t + 1:

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 16 of 20

the model is: yij = μi + εij, where i = 0, 1, ... , t and j = 1, ... , ni

ni
sample mean: ȳ i. = n1 ∑ yij
i
j

2
and V ar(ȳ i. ) = σn
i

2 2
Furthermore, V ar(ȳ i. − ȳ 0 ) = σn + σn
i 0

2
^ = MSE and assume ni = n for i = 1, ... , t.
Use σ

t 2 2
ˆ(ȳ i. − ȳ 0. ) = t( σn + σn0 )
Then the Total Sample Variance (TSV) = (T SV ) = ∑ var
i=1

We want to minimize tσ 2 ( n1 + n1 ) where N = tn + n0

0

This is a LaGrange multiplier problem (calculus): min {TSV + λ(N - tn - n0}:

Solve:
∂(∗) −tσ 2
1)
δn
= n2
− λt = 0
∂(∗) −tσ 2
2)
∂n0
= n20
−λ=0

−tσ 2
From 2) λ = we can then substitute into 1) as follows:
n20

−tσ 2 −tσ 2 2
n20 n0
= λt = ⟹ n = ⟹ n = ⟹ n0 = n√t
n2 n20 t √t
N
Therefore, from N = tn + n0 = tn + √tn = n(t + √t) ⟹ n =
(t+√t)

When this is all worked out we have a nice simple rule to guide our decision about how to allocate our
observations:

n0 = n√t
Or, the number of observations in the control group should be the square root of the number of
treatments times the number of observations in the treatment groups.

If we want to get the exact n based on our resources, let n = N/(t + √t) and n0 = √t × n and
then round to the nearest integers.

Back to our example....

In our example we had N = 60 and t = 4. Plugging these values into the equation above gives us n = 10
and n0 = 20. We should allocate 20 observations in the control and 10 observations in each of the
treatments. The purpose is not to compare each of the new drugs to each other but rather to answer
whether or not the new drug is better than the control.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 17 of 20

These calculations demonstrate once again, that the design principles we use in this course are almost
always based on trying to minimize the variance and maximizing the power of the experiment. Here is a
case where equal allocation is not optimal because you are not interested equally in all comparisons.
You are interested in specific comparisons i.e. treatments versus the control, so the control takes on a
special importance. In this case we allocate additional observations to the control group for the purpose
of minimizing the total variance.

3.5 - One-way Random Effects Models

With quantitative factors, we may want to make inference to levels not measured in the experiment by
interpolation or extrapolation on the measurement scale. With categorical factors, we may only be able
to use a subset of all possible levels - e.g. brands of popcorn - but we would still like to be able to make
inference to other levels. Imagine that we randomly select a of the possible levels of the factor of
interest. In this case, we say that the factor is random. As before, the usual single factor ANOVA
applies which is

i = 1, 2, … , a
yij = μ + τi + εij {
j = 1, 2, … , n
However, here both the error term and treatment effects are random variables, that is

εij  is NID(0, σ 2 ) and τi is NID(0, στ2 )

Also, τi and εij are independent. The variances στ2 and σ2 are called variance components.

In the fixed effect models we test the equality of the treatment means. However, this is no longer
appropriate because treatments are randomly selected and we are interested in the population of
treatments rather than any individual one. The appropriate hypothesis test for a random effect is:

H0 : στ2 = 0
H1 : στ2 > 0
The standard ANOVA partition of the total sum of squares still works; and leads to the usual ANOVA
display. However, as before, the form of the appropriate test statistic depends on the Expected Mean
Squares. In this case, the appropriate test statistic would be

F0 = M STreatments /M SE
which follows an F distribution with a-1 and N-a degrees of freedom. Furthermore, we are also
interested in estimating the variance components σ2τ and σ2. To do so, we use the analysis of
variance method which consists of equating the expected mean squares to their observed values.

^2 = M SE  and σ
σ ^ 2 + nσ
^2τ = M STreatments
M STreatment − M SE
^2τ =
σ
n
^2 = M SE
σ
Potential problem that may arise here is that the estimated treatment variance component may be
negative. It such a case, it is proposed to either consider zero in case of a negative estimate or use
another method which always results in a positive estimate. A negative estimate for the treatment
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 18 of 20

variance component can also be viewed as a evidence that the model is not appropriate, which
suggests looking for a better one.

Example 3.11 (13.1 in the 7th ed) discusses a single random factor case about the differences among
looms in a textile weaving company. Four looms have been chosen randomly from a population of
looms within a weaving shed and four observations were made on each loom. Table 13.1 illustrates the
data obtained from the experiment. Here is the Minitab output for this example using Stat> ANOVA>
Balanced ANOVA command.

The interpretation made from the ANOVA table is as before. With the p-value equal to 0.000 it is
obvious that the looms in the plant are significantly different, or more accurately stated, the variance
component among the looms is significantly larger than zero. And confidence intervals can be found
for the variance components. The 100(1-α)% confidence interval for σ2 is

(N − a)MSE (N − a)MSE
≤ σ2 ≤
χ2α/2,N−a χ21−α/2,N−a

Confidence intervals for other variance components are provided in the textbook. It should be noted
that a closed form expression for the confidence interval on some parameters may not be obtained.

3.6 - The General Linear Test

This is just a general representation of an F-test based on a full and a reduced model. We will use this
frequently when we look at more complex models.

Let's illustrate the general linear test here for the single factor experiment:

First we write the full model, Yij = μ + τi + εij and then the reduced model, Yij = μ + εij where you don't
have a τi term, you just have an overall mean, μ. This is a pretty degenerate model that just says all the
observations are just coming from one group. But the reduced model is equivalent to what we are
hypothesizing when we say the μi would all be equal, i.e.:

H0 : μ1 = μ2 = ... = μa

This is equivalent to our null hypothesis where the τi's are all equal to 0.
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 19 of 20

The reduced model is just another way of stating our hypothesis. But in more complex situations this is
not the only reduced model that we can write, there are others we could look at.

The general linear test is stated as an F ratio:

(SSE(R) − SSE(F))/(dfR − dfF)

F=
SSE(F )/dfF
This is a very general test. You can apply any full and reduced model and test whether or not the
difference between the full and the reduced model is significant just by looking at the difference in the
SSE appropriately. This has an F distribution with (df R - df F), df F degrees of freedom, which
correspond to the numerator and the denominator degrees of freedom of this F ratio.

Let's take a look at this general linear test using Minitab...

Example - Cotton Weight

Remember this experiment had treatment levels 15, 20, 25, 30,
35 % cotton weight and the observations were the tensile
strength of the material.

The full model allows a different mean for each level of cotton
weight %.

We can demonstrate the General Linear Test by viewing the

ANOVA table from Minitab:

STAT > ANOVA > Balanced ANOVA

The SSE(R) = 636.96 with a dfR = 24, and SSE(F) = 161.20 with dfF = 20. Therefore:

(636.96 − 161.20)/(24 − 20)

F∗ =
161.20/20
This demonstrates the equivalence of this test to the F-test. We now use the General Linear Test (GLT)
to test for Lack of Fit when fitting a series of polynomial regression models to determine the appropriate
degree of polynomial.

We can demonstrate the General Linear Test by comparing the quadratic polynomial model (Reduced
model), with the full ANOVA model (Full model). Let Yij = μ + β1xij + β2xij2 + εij be the reduced model,
where xij is the cotton weight percent. Let Yij = μ + τi + εij be the full model.

[2]

The viewlet above shows the SSE(R) = 260.126 with dfR = 22 for the quadratic regression model. The
ANOVA shows the full model with SSE(F) = 161.20 with dfF = 20.

Therefore the GLT is:

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 20 of 20

(SSE(R) − SSE(F))/(dfR − dfF)

F∗ =
SSE(F )/dfF
(260.126 − 161.200)/(22 − 20)
=
161.20/20
We reject H0 : Quadratic Model =and98.926/2
claim there is Lack of Fit if F* > F1-α (2, 20) = 3.49.
8.06
Therefore, since 6.14 is > 3.49 we 49.46
reject the null hypothesis of no Lack of Fit from the quadratic
= From the viewlet above we noticed that the cubic term in the
equation and fit a cubic polynomial.
equation was indeed significant with8.06
p-value = 0.015.
= 6.14
We can apply the General Linear Test again, now testing whether the cubic equation is adequate. The
reduced model is:

Yij = μ + β1xij + β2xij2 + β3xij3 + εij

and the full model is the same as before, the full ANOVA model:

Yij = μ + τi + εij

The General Linear Test is now a test for Lack of Fit from the cubic model:

(SSE(R) − SSE(F))/(dfR − dfF)

F∗ =
SSE(F )/dfF
(195.146 − 161.200)/(21 − 20)
=
161.20/20
33.95/1
=
8.06
= 4.21
We reject if F* > F0.95 (1, 20) = 4.35.

Therefore we do not reject Ha: Lack of Fit and conclude the data are consistent with the cubic
regression model, and higher order terms are not necessary.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/12

Links:
[1]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson03/cotton_weight.MTW
[2] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson03/L03_cotton_weight_viewlet_swf.html',
'l03_cotton_weight', 704, 652 );

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/12/ 4/18/2019
 Page 1 of 33

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 4: Blocking

Lesson 4: Blocking
Introduction

Blocking factors and nuisance factors provide the mechanism for explaining and controlling variation
among the experimental units from sources that are not of interest to you and therefore are part of the
error or noise aspect of the analysis. Block designs help maintain internal validity, by reducing the
possibility that the observed effects are due to a confounding factor, while maintaining external validity by
allowing the investigator to use less stringent restrictions on the sampling population.

The single design we looked at so far is the completely randomized design (CRD) where we only have a
single factor. In the CRD setting we simply randomly assign the treatments to the available experimental
units in our experiment.

When we have a single blocking factor available for our experiment we will try to utilize a randomized
complete block design (RCBD). We also consider extensions when more than a single blocking factor
exists which takes us to Latin Squares and their generalizations. When we can utilize these ideal designs,
which have nice simple structure, the analysis is still very simple, and the designs are quite efficient in
terms of power and reducing the error variation.

Learning outcomes & objectives

By the end of this lesson, students are supposed to know

• Concept of Blocking in Design of Experiment

• Dealing with missing data cases in Randomized Complete Block Design
• Application of Latin Square Designs in presence of two nuisance factors
• Application of Graeco-Latin Square Design in presence of three blocking factor sources of variation
• Crossover Designs and their special clinical applications
• Balanced Incomplete Block Designs (BIBD)

References

In this lesson specific references to material in the textbook come from Chapter 4, including:

The Hardness Testing Example, Section 4.1

Vascular Graft Example, Example 4.1

The Latin Square Design, Section 4.2

4.1 - Blocking Scenarios

To compare the results from the RCBD, we take a look at the table below. What we did here was use the
one-way analysis of variance instead of the two-way to illustrate what might have occurred if we had not
blocked, if we had ignored the variation due to the different specimens. Blocking is a technique for
Loading
dealing[MathJax]/extensions/MathMenu.js
with nuisance factors.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 2 of 33

A nuisance factor is a factor that has some effect on the response, but is of no interest to the
experimenter; however, the variability it transmits to the response needs to be minimized or explained.
We will talk about treatment factors, which we are interested in, and blocking factors, which we are not
interested in. We will try to account for these nuisance factors in our model and analysis.

Typical nuisance factors include batches of raw material if you are in a production situation, different
operators, nurses or subjects in studies, the pieces of test equipment, when studying a process, and time
(shifts, days, etc.) where the time of the day or the shift can be a factor that influences the response.

Many industrial and human subjects experiments involve blocking, or when they do not, probably should
in order to reduce the unexplained variation.

Where does the term block come from? The original use of the term block for removing a source of
variation comes from agriculture. Given that you have a plot of land and you want to do an experiment on
crops, for instance perhaps testing different varieties or different levels of fertilizer, you would take a
section of land and divide it into plots and assigned your treatments at random to these plots. If the
section of land contains a large number of plots, they will tend to be very variable - heterogeneous.

A block is characterized by a set of homogeneous plots or a set of similar experimental units. In

agriculture a typical block is a set of contiguous plots of land under the assumption that fertility, moisture,
weather, will all be similar, and thus the plots are homogeneous.

Failure to block is a common flaw in designing an experiment. Can you think of the consequences?

If the nuisance variable is known and controllable, we use blocking and control it by including a
blocking factor in our experiment.

If you have a nuisance factor that is known but uncontrollable, sometimes we can use analysis of
covariance (see Chapter 15) to measure and remove the effect of the nuisance factor from the analysis.
In that case we adjust statistically to account for a covariate, whereas in blocking, we design the
experiment with a block factor as an essential component of the design. Which do you think is
preferable?

Many times there are nuisance factors that are unknown and uncontrollable (sometimes called a
“lurking” variable). We use randomization to balance out their impact. We always randomize so that
every experimental unit has an equal chance of being assigned to a given treatment. Randomization is
our insurance against a systematic bias due to a nuisance factor.

Sometimes several sources of variation are combined to define the block, so the block becomes an
aggregate variable. Consider a scenario where we want to test various subjects with different treatments.

Age classes and gender

In studies involving human subjects, we often use gender and age classes as the blocking factors. We
could simply divide our subjects into age classes, however this does not consider gender. Therefore we
partition our subjects by gender and from there into age classes. Thus we have a block of subjects that is
defined by the combination of factors, gender and age class.

Institution (size, location, type, etc)

Often in medical studies, the blocking factor used is the type of institution. This provides a very useful
blocking factor, hopefully removing institutionally related factors such as size of the institution, types of
populations served, hospitals versus clinics, etc., that would influence the overall results of the
experiment.

The Hardness Testing Example

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 3 of 33

In this example we wish to determine whether 4 different tips (the treatment factor) produce different
(mean) hardness readings on a Rockwell hardness tester. The treatment factor is the design of the tip for
the machine that determines the hardness of metal. The tip is one component of the testing machine.

To conduct this experiment we assign the tips to an experimental unit; that is, to a test specimen (called
a coupon), which is a piece of metal on which the tip is tested.

If the structure were a completely randomized experiment (CRD) that we discussed in lesson 3, we would
assign the tips to a random piece of metal for each test. In this case, the test specimens would be
considered a source of nuisance variability. If we conduct this as a blocked experiment, we would
assign all four tips to the same test specimen, randomly assigned to be tested on a different location on
the specimen. Since each treatment occurs once in each block, the number of test specimens is the
number of replicates.

Back to the hardness testing example, the experimenter may very well want to test the tips across
specimens of various hardness levels. This shows the importance of blocking. To conduct this experiment
as a RCBD, we assign all 4 tips to each specimen.

In this experiment, each specimen is called a “block”; thus, we have designed a more homogenous set of
experimental units on which to test the tips.

Variability between blocks can be large, since we will remove this source of variability, whereas variability
within a block should be relatively small. In general, a block is a specific level of the nuisance factor.

Another way to think about this is that a complete replicate of the basic experiment is conducted in each
block. In this case, a block represents an experimental-wide restriction on randomization. However,
experimental runs within a block are randomized.

Suppose that we use b = 4 blocks as shown in the table below:

Notice the two-way structure of the experiment. Here we have four blocks and within each of these
blocks is a random assignment of the tips within each block.

We are primarily interested in testing the equality of treatment means, but now we have the ability to
remove the variability associated with the nuisance factor (the blocks) through the grouping of the
experimental units prior to having assigned the treatments.

The ANOVA for Randomized Complete Block Design (RCBD)

In the RCBD we have one run of each treatment in each block. In some disciplines each block is called
an experiment (because a copy of the entire experiment is in the block) but in statistics we call the block
to be a replicate. This is a matter of scientific jargon, the design and analysis of the study is an RCBD in
both cases.

Suppose
Loading that there are a treatments
[MathJax]/extensions/MathMenu.js (factor levels) and b blocks.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 4 of 33

A statistical model (effects model) for the RCBD is:

i = 1, 2, … , a
Yij = μ + τi + βj + εij {
j = 1, 2, … , b
This is just an extension of the model we had in the one-way case. We have for each observation Yij an
additive model with an overall mean, plus an effect due to treatment, plus an effect due to block, plus
error.

The relevant (fixed effects) hypothesis for the treatment effect is:

b
H0 : μ1 = μ2 = ⋯ = μa where μi = (1/b) ∑(μ + τi + βj ) = μ + τi
j=1

b
if ∑ βj = 0
j=1

We make the assumption that the errors are independent and normally distributed with constant variance
σ2.

The ANOVA is just a partitioning of the variation:

a b a b
2
∑ ∑(yij − ȳ .. ) = ∑ ∑[(ȳ i. − ȳ .. ) + (ȳ .j − ȳ .. )
i=1 j=1 i=1 j=1

+(yij − ȳ i. − ȳ .j + ȳ .. )]2
= b ∑ a(ȳ i. − ȳ .. )2 + a ∑ b(ȳ .j − ȳ .. )2
i=1 j=1
a b
+ ∑ ∑(yij − ȳ i. − ȳ .j + ȳ .. )2
i=1 j=1

S ST = S STreatments + SSBlocks + S SE
The algebra of the sum of squares falls out in this way. We can partition the effects into three parts: sum
of squares due to treatments, sum of squares due to the blocks and the sum of squares due to error.

The degrees of freedom for the sums of squares in:

S ST = S STreatments + SSBlocks + S SE
are as follows for a treatments and b blocks:

ab − 1 = (a − 1) + (b − 1) + (a − 1)(b − 1)
The partitioning of the variation of the sum of squares and the corresponding partitioning of the degrees of
freedom provides the basis for our orthogonal analysis of variance.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 5 of 33

ANOVA Display for the RCBD

In Table 4.2 we have the sum of squares due to treatment, the sum of squares due to blocks, and the
sum of squares due to error. The degrees of freedom add up to a total of N-1, where N = ab. We obtain
the Mean Square values by dividing the sum of squares by the degrees of freedom.

Then, under the null hypothesis of no treatment effect the ratio of the mean square for treatments to the
error mean square is an F statistic that is used to test the hypothesis of equal treatment means.

The text provides manual computing formulas; however, we will use Minitab to analyze the RCBD.

Back to the Tip Hardness example:

Remember, the hardness of specimens (coupons) is tested with 4 different tips.

Note: tips are the treatment factor levels, and the coupons are the block levels, composed of
homogeneous specimens.

Here is the data for this experiment (tip_hardness.txt [1]):

Obs Tip Hardness Coupon

1 1 9.3 1
2 1 9.4 2
3 1 9.6 3
4 1 10.0 4
5 2 9.4 1
6 2 9.3 2
7 2 9.8 3
8 2 9.9 4
9 3 9.2 1
10 3 9.4 2
11 3 9.5 3
12 3 9.7 4
13 4 9.7 1
14 4 9.6 2
15 4 10.0 3
16 [MathJax]/extensions/MathMenu.js
Loading 4 10.2 4

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 6 of 33

Here is the output from Minitab. We can see four levels of the Tip and four levels for Coupon:

The Analysis of Variance table shows three degrees of freedom for Tip three for Coupon, and the error
degrees of freedom is nine. The ratio of mean squares of treatment over error gives us an F ratio that is
equal to 14.44 which is highly significant since it is greater than the .001 percentile of the F distribution
with three and nine degrees of freedom.

Our 2-way analysis also provides a test for the block factor, Coupon. The ANOVA shows that this factor is
also significant with an F-test = 30.94. So, there is a large amount of variation in hardness between the
pieces of metal. This is why we used specimen (or coupon) as our blocking factor. We expected in
advance that it would account for a large amount of variation. By including block in the model and in the
analysis, we removed this large portion of the variation, such that the residual error is quite small. By
including a block factor in the model, the error variance is reduced, and the test on treatments is more
powerful.

The test on the block factor is typically not of interest except to confirm that you used a good blocking
factor. The results are summarized by the table of means given below.

Here is the residual analysis from the two-way structure.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 7 of 33

Comparing the CRD to the RCBD

To compare the results from the RCBD, we take a look at the table below. What we did here was use the
one-way analysis of variance instead of the two-way to illustrate what might have occurred if we had not
blocked, if we had ignored the variation due to the different specimens.

This isn't quite fair because we did in fact block, but putting the data into one-way analysis we see the
same variation due to tip, which is 3.85. So we are explaining the same amount of variation due to the tip.
That has not changed. But now we have 12 degrees of freedom for error because we have not blocked
and the sum of squares for error is much larger than it was before, thus our F-test is 1.7. If we hadn't
blocked the experiment our error would be much larger and in fact we would not even show a significant
difference among these tips. This provides a good illustration of the benefit of blocking to reduce error.
Notice that the standard deviation, S = √MSE, would be about three times larger if we had not blocked.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 8 of 33

Other Aspects of the RCBD

See Text, Section 4.1.3.

The RCBD utilizes an additive model – one in which there is no interaction between treatments and
blocks. The error term in a randomized complete block model reflects how the treatment effect varies from
one block to another.

Both the treatments and blocks can be looked at as random effects rather than fixed effects, if the levels
were selected at random from a population of possible treatments or blocks. We consider this case later,
but it does not change the test for a treatment effect.

What are the consequences of not blocking if we should have? Generally the unexplained error in the
model will be larger, and therefore the test of the treatment effect less powerful.

How to determine the sample size in the RCBD? The OC curve approach can be used to determine
the number of blocks to run. See Section 4.1.3. In a RCBD, b, the number of blocks represents the
number of replications. The power calculations that we looked at before would be the same, except that
we use b rather than n, and we use the estimate of error, σ2, that reflects the improved precision based
on having used blocks in our experiment. So, the major benefit or power comes not from the number of
replications but from the error variance which is much smaller because you removed the effects due to
block.

4.2 - RCBD and RCBD's with Missing Data

Vascular Graft Example 4.1.

This example investigates a procedure to create artificial arteries using a resin. The resin is pressed or
extruded through an aperture that forms the resin into a tube.

To conduct this experiment as a RCBD, we need to assign all 4 pressures at random to each of the 6
batches of resin. Each batch of resin is called a “block”, since a batch is a more homogenous set of
experimental units on which to test the extrusion pressures. Below is a table which provides percentages
of those products that met the specifications.

Note: Since percent response data does not generally meet the assumption of constant variance, we
might consider a variance stabilizing transformation, i.e., the arcsine square root of the proportion.
However, since the range of the percent data is quite limited, it goes from the high 70s through the 90s,
this data seems fairly homogeneous.

Figure 4.2 in the text gives the output from the statistical software package Design Expert:

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 9 of 33

Notice that Design Expert does not perform the hypothesis test on the block factor. Should we test the
block factor?

Below is the Minitab output which treats both batch and treatment the same and tests the hypothesis of
no effect.

This example shows the output from the ANOVA command in Minitab (Menu >> Stat >> ANOVA >>
Balanced ANOVA). It does hypothesis tests for both batch and pressure, and they are both significant.
Otherwise, the results from both programs are very similar.

Again, should we test the block factor? Generally, the answer is no, but in some instances this might
be helpful. We use the RCBD design because we hope to remove from error the variation due to the
block. If the block factor is not significant, then the block variation, or mean square due to the block
treatments is no greater than the mean square due to the error. In other words, if the block F ratio is close
to 1 (or generally not greater than 2), you have wasted effort in doing the experiment as a block design,
and used in this case 5 degrees of freedom that could be part of error degrees of freedom, hence the
design could actually be less efficient!

Therefore, one can test the block simply to confirm that the block factor is effective and explains variation
that would otherwise be part of your experimental error. However, you generally cannot make any
stronger conclusions from the test on a block factor, because you may not have randomly selected the
blocks from any population, nor randomly assigned the levels.

Why did I first say no?

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 10 of 33

There are two cases we should consider separately, when blocks are: 1) a classification factor and 2) an
experimental factor. In the case where blocks are a batch, it is a classification factor, but it might also be
subjects or plots of land which are also classification factors. For a RCBD you can apply your experiment
to convenient subjects. In the general case of classification factors, you should sample from the
population in order to make inferences about that population. These observed batches are not necessarily
a sample from any population. If you want to make inferences about a factor then there should be an
appropriate randomization, i.e. random selection, so that you can make inferences about the population.
In the case of experimental factors, such as oven temperature for a process, all you want is a
representative set of temperatures such that the treatment is given under homogeneous conditions. The
point is that we set the temperature once in each block; we don't reset it for each observation. So, there is
no replication of the block factor. We do our randomization of treatments within a block. In this case there
is an asymmetry between treatment and block factors. In summary, you are only including the block factor
to reduce the error variation due to this nuisance factor, not to test the effect of this factor.

ANOVA: Yield versus Batch, Pressure

The residual analysis for the Vascular Graft example is shown:

The pattern does not strike me as indicating an unequal variance.

Another way to look at these residuals is to plot the residuals against the two factors. Notice that pressure
is the treatment factor and batch is the block factor. Here we'll check for homogeneous variance. Against
treatment these look quite homogeneous.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 11 of 33

Plotted against block the sixth does raise ones eyebrow a bit. It seems to be very close to zero.

Basic residual plots indicate that normality, constant variance assumptions are satisfied. Therefore,
there seems to be no obvious problems with randomization. These plots provide more information about
the constant variance assumption, and can reveal possible outliers. The plot of residuals versus order
sometimes indicates a problem with the independence assumption.

Missing Data

In the example dataset above, what if the data point 94.7 (second treatment, fourth block) was missing?
What data point can I substitute for the missing point?

If this point is missing we can substitute x, calculate the sum of squares residuals, and solve for x which
minimizes the error and gives us a point based on all the other data and the two-way model. We
sometimes call this an imputed point, where you use a least squares approach to estimate this missing
data point.

After calculating x, you could substitute the estimated data point and repeat your analysis. Now you have
an artificial point with known residual zero. So you can analyze the resulting data, but now should reduce
your error degrees of freedom by one. In any event, these are all approximate methods, i.e., using the
best fitting or imputed point.

Before high-speed computing, data imputation was often done because the ANOVA computations are
more readily done using a balanced design. There are times where imputation is still helpful but in the
case of a two-way or multiway ANOVA we generally will use the General Linear Model (GLM) and use the
full and reduced model approach to do the appropriate test. This is often called the General Linear Test
(GLT). Note that text book has mentioned this test as the General Regression Significance Test in
Section 4.1.4.

Let's take a look at this in Minitab now...

[3]

The sum of squares you want to use to test your hypothesis will be based on the adjusted treatment sum
of squares, R(τi | μ, βj) using the notation in Section 4.1.4 of the text, for testing:

Ho : τi = 0

The numerator of the F-test, for the hypothesis you want to test should be based on the adjusted SS's
that is last in the sequence or is obtained from the adjusted sums of squares. That will be very close to
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 12 of 33

what you would get using the approximate method we mentioned earlier. The general linear test is the
most powerful test for this type of situation with unbalanced data.

The General Linear Test can be used to test for significance of multiple parameters of the model at the
same time. Generally, significance of all those parameters which are in the Full model but are not
included in the Reduced model are tested, simultaneously. The F test statistic is defined as

SSE(R) − SSE(F ) SSE(F)

F∗ = ÷
dfR − dfF dfF
Where F stands for “Full” and R stands for “Reduced.” Numerator and denominator degrees of freedom
for the F statistic are dfR - dfF and dfF , respectively.

Here are the results for the GLM with all the data intact. There are 23 degrees of freedom total here so
this is based on the full set of 24 observations.

When the data are complete this analysis from GLM is correct and equivalent to the results from the two-
way command in Minitab. When you have missing data, the raw marginal means are wrong. What if the
missing data point were from a very high measuring block? It would reduce the overall effect of that
treatment, and the estimated treatment mean would be biased.

Above you have the least squares means that correspond exactly to the simple means from the earlier
analysis.

We now illustrate the GLM analysis based on the missing data situation - one observation missing (Batch
4, pressure 2 data point removed). The least squares means as you can see (below) are slightly different,
for pressure 8700. What you also want to notice is the standard error of these means, i.e., the S.E., for
the second treatment is slightly larger. The fact that you are missing a point is reflected in the estimate of
error. You do not have as many data points on that particular treatment.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 13 of 33

The overall results are similar. We have only lost one point and our hypothesis test is still significant, with
a p-value of 0.003 rather than 0.002.

Here is a plot of the least squares means for Yield with all of the observations included.

Here is a plot of the least squares means for Yield with the missing data, not very different.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 14 of 33

Again, for any unbalanced data situation we will use the GLM. For most of our examples GLM will be a
useful tool for analyzing and getting the analysis of variance summary table. Even if you are unsure
whether your data are orthogonal, one way to check if you simply made a mistake in entering your data is
by checking whether the sequential sums of squares agree with the adjusted sums of squares.

4.3 - The Latin Square Design

Text reference, Section 4.2.

Latin Square Designs are probably not used as much as they should be - they are very efficient designs.
Latin square designs allow for two blocking factors. In other words, these designs are used to
simultaneously control (or eliminate) two sources of nuisance variability. For instance, if you had a plot
of land the fertility of this land might change in both directions, North -- South and East -- West due to soil
or moisture gradients. So, both rows and columns can be used as blocking factors. However, you can use
Latin squares in lots of other settings. As we shall see, Latin squares can be used as much as the RCBD
in industrial experimentation as well as other experiments.

Whenever, you have more than one blocking factor a Latin square design will allow you to remove the
variation for these two sources from the error variation. So, consider we had a plot of land, we might have
blocked it in columns and rows, i.e. each row is a level of the row factor, and each column is a level of the
column factor. We can remove the variation from our measured response in both directions if we consider
both rows and columns as factors in our design.

The Latin Square Design gets its name from the fact that we can write it as a square with Latin letters to
correspond to the treatments. The treatment factor levels are the Latin letters in the Latin square design.
The number of rows and columns has to correspond to the number of treatment levels. So, if we have
four treatments then we would need to have four rows and four columns in order to create a Latin square.
This gives us a design where we have each of the treatments and in each row and in each column.

This is just one of many 4×4 squares that you could create. In fact, you can make any size square you
want, for any number of treatments - it just needs to have the following property associated with it - that
each treatment occurs only once in each row and once in each column.

Consider another example in an industrial setting: the rows are the batch of raw material, the columns are
the operator of the equipment, and the treatments (A, B, C and D) are an industrial process or protocol for
producing a particular product.

What is the model? We let:

yijk = μ + ρi + βj + τk + eijk
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 15 of 33

i = 1, ... , t
j = 1, ... , t
[k = 1, ... , t] where - k = d(i, j) and the total number of observations

N = t2 (the number of rows times the number of columns) and t is the number of treatments.

Note that a Latin Square is an incomplete design, which means that it does not include observations for
all possible combinations of i, j and k. This is why we use notation k = d(i, j). Once we know the row and
column of the design, then the treatment is specified. In other words, if we know i and j, then k is specified
by the Latin Square design.

This property has an impact on how we calculate means and sums of squares, and for this reason we can
not use the balanced ANOVA command in Minitab even though it looks perfectly balanced. We will see
later that although it has the property of orthogonality, you still cannot use the balanced ANOVA
command in Minitab because it is not complete.

An assumption that we make when using a Latin square design is that the three factors (treatments, and
two nuisance factors) do not interact. If this assumption is violated, the Latin Square design error term
will be inflated.

The randomization procedure for assigning treatments that you would like to use when you actually apply
a Latin Square, is somewhat restricted to preserve the structure of the Latin Square. The ideal
randomization would be to select a square from the set of all possible Latin squares of the specified size.
However, a more practical randomization scheme would be to select a standardized Latin square at
random (these are tabulated) and then:

1. randomly permute the columns,

2. randomly permute the rows, and then
3. assign the treatments to the Latin letters in a random fashion.

Consider a factory setting where you are producing a product with 4 operators and 4 machines. We call
the columns the operators and the rows the machines. Then you can randomly assign the specific
operators to a row and the specific machines to a column. The treatment is one of four protocols for
producing the product and our interest is in the average time needed to produce each product. If both the
machine and the operator have an effect on the time to produce, then by using a Latin Square Design this
variation due to machine or operators will be effectively removed from the analysis.

The following table gives the degrees of freedom for the terms in the model.

AOV df df for the example

Rows t-1 3
Cols t-1 3
Treatments t-1 3
Error (t-1)(t-2) 6
Total (t2 - 1) 15

A Latin Square design is actually easy to analyze. Because of the restricted layout, one observation per
treatment in each row and column, the model is orthogonal.

If the row, ρi, and column, βj, effects are random with expectations zero, the expected value of Yijk is μ +
τk. In other words, the treatment effects and treatment means are orthogonal to the row and column
effects. We can also write the sums of squares, as seen in Table 4.10 in the text.

We can test for row and column effects, but our focus of interest in a Latin square design is on the
treatments.
Loading Just as in RCBD, the row and column factors are included to reduce the error variation but
[MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 16 of 33

are not typically of interest. And, depending on how we've conducted the experiment they often haven't
been randomized in a way that allows us to make any reliable inference from those tests.

Note: if you have missing data then you need to use the general linear model and test the effect of
treatment after fitting the model that would account for the row and column effects.

In general, the General Linear Model tests the hypothesis that:

Ho : τi = 0 vs. Ha : τi ≠ 0

To test this hypothesis we will look at the F-ratio which is written as:

MS(τk |μ, ρi , βj )
F= ∼ F((t − 1), (t − 1)(t − 2))
MSE(μ, ρi , βj , τk )
To get this in Minitab you would use GLM and fit the three terms: rows, columns and treatments. The F
statistic is based on the adjusted MS for treatment.

The Rocket Propellant Problem – A Latin Square Design (Table 4.9 in 8th ed and Table 4-8in 7th ed)

Table 4-13 (4-12 in 7th ed) shows some other Latin Squares from t = 3 to t = 7 and states the number of
different arrangements available.

Statistical Analysis of the Latin Square Design

The statistical (effects) model is:

⎪ i = 1, 2, … , p
⎧ 
Yijk = μ + ρi + βj + τk + εijk ⎨ j = 1, 2, … , p
⎩ 
⎪ 
k = 1, 2, … , p
but k = d(i, j) shows the dependence of k in the cell i, j on the design layout, and p = t the number of
treatment levels.

The statistical analysis (ANOVA) is much like the analysis for the RCBD.

See the ANOVA table, Table 4.10 (Table 4-9 in 7th ed)

The analysis for the rocket propellant example is presented in Example 4.3.

4.4 - Replicated Latin Squares

Latin Squares are very efficient by including two blocking factors, however the d.f. for error are often too
small. In these situations, we consider replicating a Latin Square. Let's go back to the factory scenario
again as an example and look at n = 3 repetitions of a 4 × 4 Latin square.
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 17 of 33

We labeled the row factor the machines, the column factor the operators and the Latin letters denoted the
protocol used by the operators which was the treatment factor. We will replicate this Latin Square
experiment n = 3 times. Now we have total observations equal to N = t2n.

You could use the same squares over again in each replicate, but we prefer to randomize these
separately for each replicate. It might look like this:

Ok, with this scenario in mind, let's consider three cases that are relevant and each case requires a
different model to analyze. The cases are determined by whether or not the blocking factors are the same
or different across the replicated squares. The treatments are going to be the same but the question is
whether the levels of the blocking factors remain the same.

Case 1

Here we will have the same row and column levels. For instance, we might do this experiment all in the
same factory using the same machines and the same operators for these machines. The first replicate
would occur during the first week, the second replicate would occur during the second week, etc. Week
one would be replication one, week two would be replication two and week three would be replication
three.

We would write the model for this case as:

Yhijk = μ + δh + ρi + βj + τk + ehijk
where:

h = 1, ... , n
i = 1, ... , t
j = 1, ... , t
k = dh (i, j) - the Latin letters
This is a simple extension of the basic model that we had looked at earlier. We have added one more
term to our model. The row and column and treatment all have the same parameters, the same effects
that we had in the single Latin square. In a Latin square the error is a combination of any interactions that
might exist and experimental error. Remember, we can't estimate interactions in a Latin square.

Let's take a look at the analysis of variance table.

df for Case
AOV df SS
1
See text p.
rep=week n-1 2
143
row=machine t-1 3
column=operator
Loading [MathJax]/extensions/MathMenu.js t-1 3

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 18 of 33

treatment=protocol t - 1 3
(t - 1) [n( t + 1) -
error 36
3]
Total nt2 - 1 47

Case 2

In this case, one of our blocking factors, either row or column, is going to be the same across replicates
whereas the other will take on new values in each replicate. Back to the factory example e.g., we would
have a situation where the machines are going to be different (you can say they are nested in each of the
repetitions) but the operators will stay the same (crossed with replicates). In this scenario, perhaps, this
factory has three locations and we want to include machines from each of these three different factories.
To keep the experiment standardized, we will move our operators with us as we go from one factory
location to the next. This might be laid out like this:

There is a subtle difference here between this experiment in a Case 2 and the experiment in Case 1 - but
it does affect how we analyze the data. Here the model is written as:

Yhijk = μ + δh + ρi(h) + βj + τk + ehijk

where:

h = 1, ... , n
i = 1, ... , t
j = 1, ... , t
k = dh (i, j) - the Latin letters
and the 12 machines are distinguished by nesting the i index within the h replicates.

This affects our ANOVA table. Compare this to the previous case:

df for Case
AOV df SS
2
See text p.
rep = factory n-1 2
144.
row (rep) = machine
n(t - 1) 9
(factory)
column = operator t-1 3
treatment = protocol t-1 3
error 30
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 19 of 33

(t - 1) (nt -
2)
Total nt2 - 1 47

Note that Case 2 may also be flipped where you might have the same machines, but different operators.

Case 3

In this case we have different levels of both the row and the column factors. Again, in our factory scenario
we would have different machines and different operators in the three replicates. In other words, both of
these factors would be nested within the replicates of the experiment.

We would write this model as:

Yhijk = μ + δh + ρi(h) + βj(h) + τk + ehijk

where:

h = 1, ... , n
i = 1, ... , t
j = 1, ... , t
k = dh (i, j) - the Latin letters
Here we have used nested terms for both of the block factors representing the fact that the levels of these
factors are not the same in each of the replicates.

The analysis of variance table would include:

df for Case
AOV df SS
3
See text p.
rep = factory n-1 2
144.
row (rep) = machine (factory) n(t - 1) 9
column (rep) = operator
n(t - 1) 9
(factory)
treatment protocol t-1 3
(t - 1) [n(t - 1) -
error 24
1]
Total nt2 - 1 47
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 20 of 33

Which case is best? There really isn't a best here... the choice of case depends on how you need to
conduct the experiment. If you are simply replicating the experiment with the same row and column
levels, you are in Case 1. If you are changing one or the other of the row or column factors, using different
machines or operators, then you are in Case 2. If both of the block factors have levels that differ across
the replicates, then you are in Case 3. The third case, where the replicates are different factories, can
also provide a comparison of the factories. The fact that you are replicating Latin Squares does allow you
to estimate some interactions that you can't estimate from a single Latin Square. If we added a treatment
by factory interaction term, for instance, this would be a meaningful term in the model, and would inform
the researcher whether the same protocol is best (or not) for all the factories.

The degrees of freedom for error grows very rapidly when you replicate Latin squares. But usually if you
are using a Latin Square then you are probably not worried too much about this error. The error is more
dependent on the specific conditions that exist for performing the experiment. For instance, if the protocol
is complicated and training the operators so they can conduct all four becomes an issue of resources then
this might be a reason why you would bring these operators to three different factories. It depends on the
conditions under which the experiment is going to be conducted.

Situations where you should use a Latin Square are where you have a single treatment factor and you
have two blocking or nuisance factors to consider, which can have the same number of levels as the
treatment factor.

4.5 - What do you do if you have more than 2

blocking factors?
When might this occur? Let's consider the factory example again. In this factory you have four machines
and four operators to conduct your experiment. You want to complete the experimental trials in a week.
Use the animation below to see how this example of a typical treatment schedule pans out...

As the treatments were assigned you should have noticed that the treatments have become confounded
with the days. Days of the week are not all the same, Monday is not always the best day of the week!
Just like any other factor not included in the design you hope it is not important or you would have
included it into the experiment in the first place.

What we now realize is that two blocking factors is not enough! We should also include the day of the
week in our experiment. It looks like day of the week could affect the treatments and introduce bias into
the treatment effects, since not all treatments occur on Monday. We want a design with 3 blocking factors;
machine, operator, and day of the week.

One way to do this would be to conduct the entire experiment on one day and replicate it four times. But
this would require 4 × 16 = 64 observations not just 16. Or, we could use what is called a Graeco-Latin
Square...

Graeco-Latin Squares

We write the Latin square first then each of the Greek letters occurs alongside each of the Latin letters. A
Graeco-Latin square is a set of two orthogonal Latin squares where each of the Greek and Latin letters is
a Latin square and the Latin square is orthogonal to the Greek square. Use the animation below to
explore a Graeco-Latin square:

The Greek letters each occur one time with each of the Latin letters. A Graeco-Latin square is orthogonal
between rows, columns, Latin letters and Greek letters. It is completely orthogonal.

How do we use this design?

We let the row be the machines, the column be the operator, (just as before) and the Greek letter the day,
(you could
Loading also think of this as the
order in which it was produced). Therefore the Greek letter could serve
[MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 21 of 33

the multiple purposes as the day effect or the order effect. The Latin letters are assigned to the treatments
as before.

We want to account for all three of the blocking factor sources of variation, and remove each of these
sources of error from the experiment. Therefore we must include them in the model.

Here is the model for this design:

Yijkl = μ + ρi + βj + τk + γl + eijkl
So, we have three blocking factors and one treatment factor.

and i, j, k and l all go from 1, ... , t , where i and j are the row and column indices, respectively, and k and l
are defined by the design, that is, k and l are specified by the Latin and Greek letters, respectively.

This is a highly efficient design with N = t2 observations.

You could go even farther and have more than two orthogonal latin squares together. These are referred
to a Hyper-Graeco-Latin squares!

Fisher, R.A. The Design of Experiments, 8th edition, 1966, p.82-84 [4], gives examples of hyper-Graeco-
Latin squares for t = 4, 5, 8 and 9.

(Note: it is impossible to have a 6 × 6 Graeco-Latin square! So in designing your experiment with a

Graeco-Latin Square - don't have 6 treatments! Add another, or drop one!)

4.6 - Crossover Designs

Crossover designs use the same experimental unit for multiple treatments. The common use of this
design is where you have subjects (human or animal) on which you want to test a set of drugs -- this is a
common situation in clinical trials for examining drugs.

The simplest case is where you only have 2 treatments and you want to give each subject both
treatments. Here as with all crossover designs we have to worry about carryover effects.

Here is a timeline of this type of design.

We give the treatment, then we later observe the effects of the treatment. This is followed by a period of
time, often called a washout period, to allow any effects to go away or dissipate. This is followed by a
second treatment, followed by an equal period of time, then the second observation.

If we only have two treatments, we will want to balance the experiment so that half the subjects get
treatment A first, and the other half get treatment B first. For example, if we had 10 subjects we might
have half of them get treatment A and the other half get treatment B in the first period. After we assign
the first treatment, A or B, and make our observation, we then assign our second treatment.

This situation can be represented as a set of 5, 2 × 2 Latin squares.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 22 of 33

We have not randomized these, although you would want to do that, and we do show the third square
different from the rest. The row effect is the order of treatment, whether A is done first or second or
whether B is done first or second. And the columns are the subjects. So, if we have 10 subjects we could
label all 10 of the subjects as we have above, or we could label the subjects 1 and 2 nested in a square.
This is similar to the situation where we have replicated Latin squares - in this case five reps of 2 × 2 Latin
squares, just as was shown previously in Case 2.

This crossover design has the following AOV table set up:

AOV df df for this example

rep = square n-1 4
column = subject(sq) n(t - 1) 5
row = order t-1 1
treatment = A vs. B t-1 1
error (t - 1) [n(t - 1) - 1] 8
2
Total nt - 1 19

We have five squares and within each square we have two subjects. So we have 4 degrees of freedom
among the five squares. We have 5 degrees of freedom representing the difference between the two
subjects in each square. If we combine these two, 4 + 5 = 9, which represents the degrees of freedom
among the 10 subjects. This representation of the variation is just the partitioning of this variation. The
same thing applies in the earlier cases we looked at.

With just two treatments there are only two ways that we can order them. Let's look at a crossover design
where t = 3. If t = 3 then there are more than two ways that we can represent the order. The basic building
block for the crossover design is the Latin Square.

Here is a 3 × 3 Latin Square. To achieve replicates, this design could be replicated several times.

In this Latin Square we have each treatment occurring in each period. Even though Latin Square
guarantees that treatment A occurs once in the first, second and third period, we don't have all sequences
represented. It is important to have all sequences represented when doing clinical trials with drugs.

Crossover Design Balanced for Carryover Effects.

The following crossover design, is based on two orthogonal Latin squares.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 23 of 33

Together, you can see that going down the columns every pairwise sequence occurs twice, AB, BC, CA,
AC, BA, CB going down the columns. The combination of these two Latin squares gives us this additional
level of balance in the design, than if we had simply taken the standard Latin square and duplicated it.

To do a crossover design, each subject receives each treatment at one time in some order. So, one of its
benefits is that you can use each subject as its own control, either as a paired experiment or as a
randomized block experiment, the subject serves as a block factor. For each subject we will have each of
the treatments applied. The number of periods is the same as the number of treatments. It is just a
question about what order you give the treatments. The smallest crossover design which allows you to
have each treatment occurring in each period would be a single Latin square.

A 3 × 3 Latin square would allow us to have each treatment occur in each time period. We can also think
about period as the order in which the drugs are administered. One sense of balance is simply to be sure
that each treatment occurs at least one time in each period. If we add subjects in sets of complete Latin
squares then we retain the orthogonality that we have with a single square.

In designs with two orthogonal Latin Squares we have all ordered pairs of treatments occurring twice and
only twice throughout the design. Take a look at the animation below to get a sense of how this occurs:

All ordered pairs occur an equal number of times in this design. It is balanced in terms of residual effects,
or carryover effects.

For an odd number of treatments, e.g. 3, 5, 7, etc., it requires two orthogonal Latin squares in order to
achieve this level of balance. For even number of treatments, 4, 6, etc., you can accomplish this with a
single square. This form of balance is denoted balanced for carryover (or residual) effects.

Here is an actual data example for a design balanced for carryover effects. In this example the subjects
are cows and the treatments are the diets provided for the cows. Using the two Latin squares we have
three diets A, B, and C that are given to 6 different cows during three different time periods of six weeks
each, after which the weight of the milk production was measured. In between the treatments a wash out
period was implemented.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 24 of 33

How do we analyze this? If we didn't have our concern for the residual effects then the model for this
experiment would be:

Yijk = μ + ρi + βj + τk + eijk
where:

ρi = period
βj = cows
τk = treatment

i = 1, ..., 3 (the number of treatments)

j = 1 , .... , 6 (the number of cows)
k = 1, ..., 3 (the number of treatments)

Let's take a look at how this is implemented in Minitab using GLM. Use the viewlet below to walk through
an initial analysis of the data (cow_diets.MTW [5]) for this experiment with cow diets. Reference: W.G.
Cochran and G.M. Cox, 1957, Experimental Designs, 2nd edition, p. 135.

These demonstrations [Inspect] are based on Minitab Version 16 or earlier. The GLM command menus
in Minitab Version 17 have changed.

[6]

Why do we use GLM? We do not have observations in all combinations of rows, columns and treatments,
since the design is based on the Latin square.

Here is a plot of the least square means for treatment and period. We can see in the table below that the
other blocking factor, cow, is also highly significant.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 25 of 33

So, let's go one step farther...

Is this an example of the Case 2 or the Case 3 of the multiple Latin Squares that we had looked at
earlier?

This is a Case 2 where the column factor, the cows are nested within the square, but the row factor,
period, is the same across squares.

Notice the sum of squares for cows is 5781.1. Let's change the model slightly using the general linear
model in Minitab again. Follow along with the animation.

[7]

Now I want to move from Case 2 to Case 3. Is the period effect in the first square the same as the period
effect in the second square? If it only means order and all the cows start lactating at the same time it
might mean the same. But if some of the cows are done in the spring and others are done in the fall or
summer, then the period effect has more meaning than simply the order. Although this represents order it
may also involve other effects you need to be aware of this. A Case 3 approach involves estimating
separate period effects within each square.

[8]

My guess is that they all started the experiment at the same time - in this case the first model would have
been appropriate.

How Do We Analyze Carryover Effect?

OK, we are looking at the main treatment effects. With our first cow, during the first period, we give it a
treatment or diet and we measure the yield. Obviously you don't have any carryover effects here because
it is the first period. However, what if the treatment they were first given was a really bad treatment? In
fact in this experiment the diet A consisted of only roughage, so, the cow's health might in fact deteriorate
as a result of this treatment. This could carry over into the next period. This carry over would hurt the
second treatment if the washout period isn't long enough. The measurement at this point is a direct
reflection of treatment B but may also have some influence from the previous treatment, treatment A.

If you look at how we have coded data here, we have another column called residual treatment. For the
first six observations we have just assigned this a value of 0 because there is no residual treatment. But
for the first observation in the second row, we have labeled this with a value of one indicating that this
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 26 of 33

was the treatment prior to the current treatment (treatment A). In this way the data is coded such that this
column indicates the treatment given in the prior period for that cow.

Now we have another factor that we can put in our model. Let's take a look at how this looks in Minitab:

[9]

We have learned everything we need to learn. We have the appropriate analysis of variance here. By
fitting in order, when residual treatment (i.e., ResTrt) was fit last we get:

SS(treatment | period, cow) = 2276.8

SS(ResTrt | period, cow, treatment) = 616.2

When we flip the order of our treatment and residual treatment, we get the sums of squares due to fitting
residual treatment after adjusting for period and cow:

SS(ResTrt | period, cow) = 38.4

SS(treatment | period, cow, ResTrt) = 2854.6

Which of these are we interested in? If we wanted to test for residual treatment effects how would we do
that? What would we use to test for treatment effects if we wanted to remove any carryover effects?

4.7 - Incomplete Block Designs

In using incomplete block designs we will use the notation t = # of treatments. We define the block size as
k. And, as you will see, in incomplete block designs k will be less than t. You cannot assign all of the
treatments in each block. In short,

t = # of treatments,
k = block size,
b = # of blocks,
ri = # of replicates for treatment i, in the entire design.

Remember that an equal number of replications is the best way to be sure that you have minimum
variance if you're looking at all possible pairwise comparisons. If ri = r for all treatments, the total number
of observations in the experiment is N where:

N = t(r) = b(k)

The incidence matrix which defines the design of the experiment, gives the number of observations say nij
for the ith treatment in the jth block. This is what it might look like here:

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 27 of 33

Here we have treatments 1, 2, up to t and the blocks 1, 2, up to b. For a complete block design we would
have each treatment occurring one time within each block, so all entries in this matrix would be 1's. For an
incomplete block design the incidence matrix would be 0's and 1's simply indicating whether or not that
treatment occurs in that block.

Example 1

The example that we will look at is Table 4.22 (4.21 in 7th ed). Here is the incidence matrix for this
example:

Here we have t = 4, b = 4, (four rows and four columns) and k = 3 ( so at each block we can only put three
of the four treatments leaving one treatment out of each block). So, in this case, the row sums (ri ) and the
columns sums, k, are all equal to 3.

In general, we are faced with a situation where the number of treatments is specified, and the block size,
or number of experimental units per block (k) is given. This is usually a constraint given from the
experimental situation. And then, the researcher must decide how many blocks are needed to run and
how many replicates that provides in order to achieve the precision or the power that you want for the
test.

Example 2

Here is another example of an incidence matrix for allocating treatments and replicates in an incomplete
block design. Let's take an example where k = 2, still t = 4, and b = 4. That gives us a case r = 2. In This
case we could design our incidence matrix so that it might look like this:

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 28 of 33

This example has two observations per block so k = 2 in each case and for all treatments r = 2.

Balanced Incomplete Block Design (BIBD)

A BIBD is an incomplete block design where all pairs of treatments occur together within a block an equal
number of times ( λ ). In general, we will specify λii' as the number of times treatment i occurs with i', in a
block.

Let's look at the previous cases. How many times does treatment one and two occur together in this first
example design?

It occurs together in block 2 and then again in block 4 (highlighted in light blue). So, λ12 = 2. If we look at
treatment one and three, this occurs together in block one and in block two therefore λ13 = 2. In this
design you can look at all possible pairs. Let's look at 1 and 4 - they occur together twice, 2 and 3 occur
together twice, 2 and 4 twice, and 3 and 4 occur together twice. For this design λii' = 2 for all ii' treatment
pairs defining the concept of balance in this incomplete block design.

If the number of times treatments occur together within a block is equal across the design for all pairs of
treatments then we call this a balanced incomplete block design (BIBD).

Now look at the incidence matrix for the second example.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 29 of 33

We can see that:

λ12 occurs together 0 times.

λ13 occurs together 2 times.
λ14 occurs together 0 times.
λ23 occurs together 0 times.
λ24 occurs together 2 times.
λ34 occurs together to 0 times.

Here we have two pairs occurring together 2 times and the other four pairs occurring together 0 times.
Therefore, this is not a balanced incomplete block design (BIBD).

What else is there about BIBD?

We can define λ in terms of our design parameters when we have equal block size k, and equal
replication ri = r. For a given set of t, k, and r we define λ as:

λ = r(k-1) / t-1

So, for the first example that we looked at earlier - let's plug in the values and calculate λ:

λ = 3 (3 - 1) / (4 -1) = 2

Here is the key: when λ is equal to an integer number it tells us that a balanced incomplete block design
exists. Let's look at the second example and use the formula and plug in the values for this second
example. So, for t = 4, k = 2, r = 2 and b = 4, we have:

λ = 2 (2 - 1) / (4 - 1) = 0.666

Since λ is not an integer there does not exist a balanced incomplete block design for this experiment. We
would either need more replicates or a larger block size. Seeing as how the block size in this case is
fixed, we can achieve a balanced complete block design by adding more replicates so that λ equals at
least 1. It needs to be a whole number in order for the design to be balanced.

We will talk about partially balanced designs later. But in thinking about this case we note that a balanced
design doesn't exist so what would be the best partially balanced design? That would be a question that
you would ask if you could only afford four blocks and the block size is two. Given this situation, is the
design in Example 2 the best design we can construct? The best partially balanced design is where λii'
should be the nearest integers to the λ that we calculated. In our case each λii' should be either 0 or 1,
the integers nearest 0.667. This example is not as close to balanced as it could be. In fact, it is not even a
connected design where you can go from any treatment to any other treatment within a block. More about
this later...

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 30 of 33

How do you Construct a BIBD?

In some situations, it is easy to construct the best IBD, however, for other cases it can be quite difficult
and we will look them up in a reference.

Let's say that we want six blocks, we still want 4 treatments and our block size is still 2. Calculate λ = r(k -
1) / (t - 1) = 1. We want to create all possible pairs of treatments because lambda is equal to one. We do
this by looking at all possible combinations of four treatments taking two at a time. We could set up the
incidence matrix for the design or we could represent it like this - entries in the table are treatment labels:
{1, 2, 3, 4}.

However, this method of constructing a BIBD using all possible combinations, does not always work as
we now demonstrate. If the number of combinations is too large then you need to find a subset - - not
always easy to do. However, sometimes you can use Latin Squares to construct a BIBD. As an example,
let's take any 3 columns from a 4 × 4 Latin Square design. This subset of columns from the whole Latin
Square creates a BIBD. However, not every subset of a Latin Square is a BIBD.

Let's look at an example. In this example we have t = 7, b = 7, and k = 3. This means that r = 3 = (bk) / t .
Here is the 7 × 7 Latin square :

We want to select (k = 3) three columns out of this design where each treatment occurs once with every
other treatment because λ = 3(3 - 1) / (7 - 1) = 1.

We could select the first three columns - let's see if this will work. Click the animation below to see
whether using the first three columns would give us combinations of treatments where treatment pairs are
not repeated.

Since the first three columns contain some pairs more than once, let's try columns 1, 2, and now we need
a third...how about the fourth column. If you look at all possible combinations in each row, each treatment
pair occurs only one time.

What if we could afford a block size of 4 instead of 3? Here t = 7, b = 7, k = 4, then r = 4. We calculate λ =

r(k - 1) / (t - 1) = 2 so a BIBD does exist. For this design with a block size of 4 we can select 4 columns (or
rows) from a Latin square. Let's look at columns again... can you select the correct 4?

Now consider the case with 8 treatments. The number of possible combinations of 8 treatments taking 4
at a time is 70. Thus with 70 sets of 4 from which you have to choose 14 blocks - - wow, this is a big job!
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 31 of 33

At this point, we should simply look at an appropriate reference. Here is a handout - a catalog that will
help you with this selection process [10] - taken from Cochran & Cox, Experimental Design, p. 469-482.

Analysis of BIBD's

When we have missing data, it affects the average of the remaining treatments in a row, i.e., when
complete data does not exist for each row - this affects the means. When we have complete data the
block effect and the column effects both drop out of the analysis since they are orthogonal. With missing
data or IBDs that are not orthogonal, even BIBD where orthogonality does not exist, the analysis requires
us to use GLM which codes the data like we did previously. The GLM fits first the block and then the
treatment.

The sequential sums of squares (Seq SS) for block is not the same as the Adj SS.

We have the following:

Seq SS
SS(β | μ) 55.0
SS(τ | μ, β) = 22.50

Adj SS
SS(β | μ, τ) = 66.08
SS(τ | μ, β) = 22.75

Switch them around...now first fit treatments and then the blocks.

Seq SS
SS(τ | μ) = 11.67
SS(β | μ, τ) = 66.08

Adj SS
SS(τ | μ, β) = 22.75
SS(β | μ, τi) = 66.08

The 'least squares means' come from the fitted model. Regardless of the pattern of missing data or the
design we can conceptually think of our design represented by the model:

Yij = μ + +βi + τj + eij

i = 1, ... , b, j = 1, ... , t

You can obtain the 'least squares means' from the estimated parameters from the least squares fit of the
model.

Optional Section

See the discussion in the text for Recovery of Interblock Information, p. 154. This refers to a procedure
which allows us to extract additional information from a BIBD when the blocks are a random effect.
Optionally you can read this section. We illustrate the analysis by the use of the software, PROC Mixed in
SAS (L03_sas_Ex_4_5.sas [11]):

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 32 of 33

SAS output file (L03_sas_Ex_4_5.lst [13] )

[12]

Note that the least squares means for treatments, when using PROC Mixed, correspond to the combined
intra- and inter-block estimates of the treatment effects.

Random Effect Factor

So far we have discussed experimental designs with fixed factors, that is, the levels of the factors are
fixed and constrained to some specific values. However, this is often not the case. In some cases, the
levels of the factors are selected at random from a larger population. In this case, the inference made on
the significance of the factor can be extended to the whole population but the factor effects are treated as
contributions to variance.

Random effect models are the topic of Chapter 13 in the text book and we will go through them in Lesson
12. Minitab’s General Linear Command handles random factors appropriately as long as you are careful
to select which factors are fixed and which are random.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/18

Links:
[1] https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/data/tip_hardness.txt
[2] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L04_block_tip_viewlet_swf.html', 'l04_block_tip', 718,
668 );
[3] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L04_missing_data_viewlet_swf.html', 'l04_missing_data',
718, 668 );
[4]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/graeco_latin_fisher.pdf
[5] https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/cow_diets.MTW
[6] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 33 of 33

( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L04_cow_GLM_step01_viewlet_swf.html',
'l04_cow_glm_step01', 718, 668 );
[7] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L04_cow_GLM_step02_viewlet_swf.html',
'l04_cow_glm_step02', 718, 668 );
[8] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L04_cow_GLM_step03_viewlet_swf.html',
'l04_cow_glm_step03', 718, 668 );
[9] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L04_cow_GLM_step04_viewlet_swf.html',
'l04_cow_glm_step04', 718, 668 );
[10]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/Cochran_Cox.pdf
[11]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L03_sas_Ex_4_5.sas
[12] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L04_interblock_viewlet_swf.html', 'l04_interblock', 821,
653 );
[13]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson04/L03_sas_Ex_4_5.lst

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/18/ 4/18/2019
 Page 1 of 11

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 5: Introduction to Factorial Designs

Lesson 5: Introduction to Factorial

Designs
Introduction

Factorial designs are the basis for another important principle besides blocking -
examining several factors simultaneously. We will start by looking at just two factors and
then generalize to more than two factors. Investigating multiple factors in the same design
automatically gives us replication for each of the factors.

Learning objectives & outcomes

Goals for this lesson includes the following

• Introductory understanding of Factorial Designs as among the most common

experimental designs
• Two factor Factorial Design and its extension to the General Factorial Designs
• Sample size determination in Factorial Designs

5.1 - Factorial Designs with Two

Treatment Factors
For now we will just consider two treatment factors of interest. It looks almost the same as
the randomized block design model only now we are including an interaction term:

Yijk = μ + αi + βj + (αβ)ij + eijk

where i = 1, ..., a, j = 1, ..., b, and k = 1, ..., n. Thus we have two factors in a factorial
structure with n observations per cell. As usual, we assume the eijk ∼ N(0, σ2), i.e.
independently and identically distributed with the normal distribution. Although it looks like
a multiplication, the interaction term need not imply multiplicative interaction.

The Effects Model vs. the Means Model

The cell means model is written:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 2 of 11

Here the cell means are: μ11, ... , μ1b, ... , μa1 ... μab. Therefore we have a × b cell means,
μij. We will define our marginal means as =
Yijk theμsimple average over our cell means as
ij + eijk
shown below:
1 1
μ̄i. = b
∑ μij , μ̄.j = a ∑ μij
j i

From the cell means structure we can talk about marginal means and row and column
means. But first we want to look at the effects model and define more carefully what the
interactions are We can write the cell means in terms of the full effects model:

μij = μ + αi + βj + (αβ)ij
It follows that the interaction terms (αβ)ij are defined as the difference between our cell
means and the additive portion of the model:

(αβ)ij = μij − (μ + αi + βj )
If the true model structure is additive then the interaction terms (αβ)ij are equal to zero.
Then we can say that the true cell means, μij = (μ + αi + βj), have additive structure.

Example 1

Let's illustrate this by considering the true means μij:

Note that both a and b are 2, thus our marginal row means are 8 and 12, and our marginal
column means are 7 and 13. Next, let's calculate the α and the β effects; since the overall
mean is 10, our α effects are -2 and 2 (which sum to 0), and our β effects are -3 and 3
(which also sum to 0). If you plot the cell means you get two lines that are parallel.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 3 of 11

The difference between the two means at the first β factor level is 9 - 5 = 4. The difference
between the means for the second β factor level is 15 - 11 = 4. We can say that the effect
of α at the first level of β is the same as the effect of α at the second level of β. Therefore
we say that there is no interaction and as we will see the interaction terms are equal to 0.

This example simply illustrates that the cell means in this case have additive structure. A
problem with data that we actually look at is that you do not know in advance whether the
effects are additive or not. Because of random error, the iteraction terms are seldom
exactly zero. You may be involved in a situation that is either additive or non-additive, and
the first task is to decide between them.

Now consider the non-additive case. We illustrate this with Example 2 which follows.

Example 2

This example was constructed so that the marginal means and the overall means are the
same as in Example 1. However, it does not have additive structure.

Using the definition of interaction:

(αβ)ij = μij - (μ + αi + βj)

which gives us (αβ)ij interaction terms that are -2, 2, 2, -2. Again, by the definition of our
interaction effects these (αβ)ij terms should sum to zero in both directions.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 4 of 11

We generally call the αi terms the treatment effects for treatment factor A and the βj terms
for treatment factor B, and the (αβ)ij terms the interaction effects.

The model we have written gives us a way to represent in a mathematical form a two
factor design, whether we use the means model or the effects model, i.e.,

Yijk = μij + eijk

or

Yijk = μ + αi + βj + (αβ)ij + eijk

There is really no benefit to the effects model when there is interaction, except that it gives
us a mechanism for partitioning the variation due to the two treatments and their
interactions. Both models have the same number of distinct parameters. However, when
there is no interaction then we can remove the interaction terms from the model and use
the reduced additive model.

Now, we'll take a look at the strategy for deciding whether our model fits, whether the
assumptions are satisfied and then decide whether we can go forward with an interaction
model or an additive model. This is the first decision. When you can eliminate the
interactions because they are not significantly different from zero, then you can use the
simpler additive model. This should be the goal whenever possible because then you
have fewer parameters to estimate, and a simpler structure to represent the underlying
scientific process.

Before we get to the analysis, however, we want to introduce another definition of effects -
rather than defining the αi effects as deviation from the mean, we can look at the
difference between the high and the low levels of factor A. These are two different
definitions of effects that will be introduced and discussed in this chapter and the next, the
αi effects and the difference between the high and low levels, which we will generally
denote as the A effect.

Factorial Designs with 2 Treatment Factors, cont'd

For a completely randomized design, which is what we discussed for the one-way
ANOVA, we need to have n × a × b = N total experimental units available. We randomly
assign n of those experimental units to each of the a × b treatment combinations. For the
moment we will only consider the model with fixed effects and constant experimental
random error.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 5 of 11

The model is:

Yijk = μ + αi + βj + (αβ)ij + eijk

i = 1, ... , a
j = 1, ... , b
k = 1, ... , n

Read the text section 5.3.2 for the definitions of the means and the sum of squares.

Testing Hypotheses

We can test the hypotheses that the marginal means are all equal, or in terms of the
definition of our effects that the αi's are all equal to zero, and the hypothesis that the βj's
are all equal to zero. And, we can test the hypothesis that the interaction effects are all
equal to zero. The alternative hypotheses are that at least one of those effects is not equal
to zero.

How we do this, in what order, and how do we interpret these tests?

One of the purposes of a factorial design is to be efficient about estimating and testing
factors A and B in a single experiment. Often we are primarily interested in the main
effects. Sometimes, we are also interested in knowing whether the factors interact. In
either case, the first test we should do is the test on the interaction effects.

The Test of H0: (αβ)ij = 0

If there is interaction and it is significant, i.e. the p-value is less than your chosen cut off,
then what do we do? If the interaction term is significant that tells us that the effect of A is
different at each level of B. Or you can say it the other way, the effect of B differs at each
level of A. Therefore, when we have significant interaction, it is not very sensible to even
be talking about the main effect of A and B, because these change depending on the level
of the other factor. If the interaction is significant then we want to estimate and focus our
attention on the cell means. If the interaction is not significant, then we can test the main
effects and focus on the main effect means.

The estimates of the interaction and main effects are given in the text in section 5.3.4.

Note that the estimates of the marginal means for A are the marginal means:

1 σ2
ȳ i.. = bn
∑ ∑ yijk , with var(ȳ i.. ) = bn
j k

A similar formula holds for factor B, with

σ2
var(ȳ .j. ) = an

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 6 of 11

Just the form of these variances tells us something about the efficiency of the two factor
design. A benefit of a two factor design is that the marginal means have n × b number of
replicates for factor A and n × a for factor B. The factorial structure, when you do not have
interactions, gives us the efficiency benefit of having additional replication, the number of
observations per cell times the number of levels of the other factor. This benefit arises
from factorial experiments rather than single factor experiments with n observations per
cell. An alternative design choice could have been to do two one-way experiments, one
with a treatments and the other with b treatments, with n observations per cell. However,
these two experiments would not have provided the same level of precision, nor the ability
to test for interactions.

Another practical question: If the interaction test is not significant what should we do?

Do we get remove the interaction term in the model? You might consider dropping that
term from the model. If n is very small and your df for error are small, then this may be a
critical issue. There is a 'rule of thumb' that I sometimes use in these cases. If the p-value
for the interaction test is greater than 0.25 then you can drop the interaction term. This is
not an exact cut off but a general rule. Remember, if you drop the interaction term, then a
variation accounted for by SSab would become part of the error and increasing the SSE,
however your error df would also become larger in some cases enough to increase the
power of the tests for the main effects. Statistical theory shows that in general dropping
the interaction term increases your false rejection rate for subsequent tests. Hence we
usually do not drop nonsignificant terms when there are adequate sample sizes.
However, if we are doing an independent experiment with the same factors we might not
include interaction in the model for that experiment.

What if n = 1, and we have only 1 observation per cell? If n = 1 then we have 0 df for
SSerror and we cannot estimate the error variance with MSE. What should we do in order
to test our hypothesis? We obviously cannot perform the test for interaction because we
have no error term.

If you are willing to assume, and if it is true that there is no interaction, then you can use
the interaction as your F-test denominator for testing the main effects. It is a fairly safe and
conservative thing to do. If it is not true then the MSab will tend to be larger than it should
be, so the F-test is conservative. You're not likely to reject a main effect if it is not true.
You won't make a Type I error, however you could more likely make a Type II error.

Extension to a 3 Factor Model

The factorial model with three factors can be written as:

Yijk = μ + αi + βj + γk + (αβ)ij + (αγ)ik + (βγ)jk + (αβγ)ijk + eijkl

where i = 1, ... , a, j = 1 , ... , b, k = 1 , ... , c, l = 1 , ... , n.

We extend the model in the same way. Our analysis of variance has three main effects,
three two-way interactions, a three-way interaction and error. If this were conducted as a

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 7 of 11

Completely Randomized Design experiment, each of the a × b × c treatment combinations

would be randomly assigned to n of the experimental units.

Sample Size Determination [Section 5.3.5]

We first consider the two factor case where N = a × b × n, (n = the number of replicates
per cell). The non-centrality parameter for calculating sample size for the A factor is:

Φ2 = ( nb × D2 ) / ( 2a × σ2 )

where D is the difference between the maximum of μ¯i. and the minimum of μ¯i. , and
where b is the number of observations in each level of factor A.

Actually at the beginning of our design process we should decide how many observations
we should take, if we want to find a difference of D, between the maximum and the
minimum of the true means for the factor A. There is a similar equation for factor B.

Φ2 = ( na × D2 ) / ( 2b × σ2 )

where na is the number of observations in each level of factor B.

In the two factor case this is just an extension of what we did in the one factor case. But
now we have the marginal means benefiting from a number of observations per cell and
the number of levels of the other factor. In this case we have n observations per cell, and
we have b cells. So, we have nb observations.

5.2 - Another Factorial Design Example -

Cloth Dyes
Consider the cloth dyes data from Problem 5.19 in the text:

For each combination of time, temperature and operator, there are three observations.
Now we have a case where there are three factors and three observations per cell. Let's
run this model in Minitab.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 8 of 11

The ANOVA table shows us that the main effects due to cycle time, operator, and
temperature are all significant. The two-way interactions for cycle time by operator and
cycle time by temperature are significant. But the operator by temperature is not
significant but the dreaded three-way interaction is significant. What does it mean when a
three-way interaction is significant?

Let's take a look at the factor plots:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 9 of 11

These interaction plots show us the three sets of two-way cell means, each of the three
are plotted in two different ways. This is a useful plot to try to understand what is going on.
These are all the two-way plots.

Typically a three-way interaction would be plotted as two panels... showing how the two-
way interactions differ across the levels of the third factor. Minitab does not do that for you
automatically.

Let's think about how this experiment was done. There are three observations for each
combination of factors. Are they actually separate experimental units or are they simply
three measurements on the same experimental unit? If they are simply three
measurements on the same piece of cloth that was all done in the same batch, for
instance, then they are not really independent. If this is the case, then another way to look
at this data would be to average those replications. In this case there is only 1
observation for each treatment, so that there would be no d.f. for error. However, the way
the problem is presented in the text, they appear to have been treated independently and
thus are true replicates, leading to 36 d.f. for error.

You could also think about the operator not as a factor that you're interested in but more
as a block factor, i.e. a source of variation that we want to remove from the study. What
we're really interested in is the effect of temperature and time on the process of dyeing the
cloth. In this case we could think about using the operator as a block effect. Running the
analysis again, now we get the same plot but look at the ANOVA table: now the
interactions related to operator have been pooled as a part of the error. So the residual
error term now has 2 + 4 + 4 + 36 = 46 df for error. Note also that if you do use the
operator as a treatment factor, it probably should be considered random. In this case, you
would probably want to consider the 2 and 3-way interactions involving operator to be
random effects. Experiments in which some factors are fixed and others are random are
called mixed effects experiments. The analysis of mixed effects experiments is discussed
in Chapter 13.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 10 of 11

What this points out is the importance of distinguishing what is a block factor, and which
are the treatment factors when you have a multifactor experimental design. This should be
apparent from how the experiment was conducted, but if the data are already collected
when you are introduced to the problem, you need to inquire carefully to understand how
the experiment was actually conducted to know what model to use in the analysis.

Let's take a look two examples using this same dataset using Minitab v.16. First we will
analyze the quantitative factors involved, Cycle Time and Temperature and as though
they were qualitative - simply nominal factors.

[1]

Next, using Operator as a block we will now use Minitab v.16 to treat the quantitative
factors as qualitative factors and apply these in a regression analysis.

[2]

Source URL: https://onlinecourses.science.psu.edu/stat503/node/26

Links:
[1] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson05/L05_clothdyes_viewlet_swf.html',
'l05_clothdyes', 724, 708 );
[2] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson05/L05_clothdyes_02_viewlet_swf.html',
'l05_clothdyes_02', 724, 708 );

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 11 of 11

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/26/ 4/18/2019
 Page 1 of 27

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 6: The 2^k Factorial Design

Lesson 6: The 2^k Factorial Design

Introduction

The 2k designs are a major set of building blocks for many experimental designs. These designs
are usually referred to as screening designs. The 2k refers to designs with k factors where each
factor has just two levels. These designs are created to explore a large number of factors, with
each factor having the minimal number of levels, just two. By screening we are referring to the
process of screening a large number of factors that might be important in your experiment, with
the goal of selecting those important for the response that you're measuring. We will see that k
can get quite large. So far we have been looking at experiments that have one, two or three
factors, maybe a blocking factor and one or two treatment factors, but when using screening
designs k can be as large as 8, 10 or 12. For those of you familiar with chemical or laboratory
processes, it would not be hard to come up with a long list of factors that would affect your
experiment. In this context we need to decide which factors are important.

In these designs we will refer to the levels as high and low, +1 and -1, to denote the high and the
low level of each factor. In most cases the levels are quantitative, although they don't have to
be. Sometimes they are qualitative, such as gender, or two types of variety, brand or process. In
these cases the +1 and -1 are simply used as labels.

Learning objectives & outcomes

Upon completion of this lesson, you should be able to do the following:

• The idea of 2-level Factorial Designs as one of the most important screening designs
• Defining a “contrast” which is an important concept and how to derive Effects and Sum of Squares
using the Contrasts
• Process of analyzing Unreplicated or Single replicated factorial designs, and
• How to use Transformation as a tool in dealing with inadequacy of either variance homogeneity or
normality of the data as major hypothetical assumptions.

6.1 - The Simplest Case

The simplest case is 2k where k = 2. We will define a new notation which is known as Yates
notation. We will refer to our factors using the letters A, B, C, D, etc. as arbitrary labels of the
factors. In the chemical process case A is the concentration of the reactant and B is the amount
of catalyst, both of which are quantitative. The yield of the process is our response variable.

Since there are two levels of each of two factors, 2k equals four. Therefore, there are four
treatment combinations and the data are given below:
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 2 of 27

You can see that we have 3 observations at each of 4 = 2k combinations for k = 2. So we have n
= 3 replicates.

A B Yates Notation
- - (1)
+ - a
- + b
+ + ab

The table above gives the data with the factors coded for each of the four combinations and
below is a plot of the region of experimentation in two dimensions for this case.

The Yates notation used for denoting the factor combinations is as follows:

We use "(1)" to denote that both factors are at the low level, "a" for when A is at its
high level and B is at its low level, "b" for when B is at its high level and A is at its low
level, and "ab" when both A and B factors are at their high level.

The use of this Yates notation indicates the high level of any factor simply by using the small
letter of that level factor. This notation actually is used for two purposes. One is to denote the
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 3 of 27

total sum of the observations at that level. In the case below b = 60 is the sum of the three
observations at the level b.

This shortcut notation, using the small letters, shows which level for each of our k factors we are
at just by its presence or absence.

We will also connect this to our previous notation for the two factor treatment design:

Yijk = μ + αi + βj + (α β)ij + eijk

What is the primary goal of these screening experiments?

The goal is to decide which of these factors is important. After determining which factors are
important, then we will typically plan for a secondary experiment where the goal is to decide
what level of the factors gives us the optimal response. Thus the screening 2k experiment is the
first stage, generally, of an experimental sequence. In the second stage one is looking for a
response surface or an experiment to find the optimal level of the important factors.

Estimation of Factors Effects (in the Yates tradition)

The definition of an effect in the 2k context is the difference in the means between the high and
the low level of a factor. From this notation A is the difference between the averages of the
observations at the high level of A minus the average of the observations at the low level of A.

Therefore, A = ȳ A+ − ȳ A− , in the example above:

A = 190/6 - 140/6 = 50/6 = 8.33

Similarly, B = ȳ B+ − ȳ B− , is the similar only looking in the other direction. In our

example:

B = 150/6 - 180/6 = 25 - 30 = -5
ab+(1) a+b
and finally, AB = 2n
− 2n

AB = [(90 + 80)/6 - (100 + 60)/6] = 10/6 = 1.67

Therefore in the Yates notation, we define an effect as the difference in the means between the
high and
Loading the low levels of a factor whereas in previous models we defined an effect as the
[MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 4 of 27

coefficients of the model, which are the differences between the marginal mean and the overall
mean. To restate this, in terms of A, the A effect is the difference between the means at the high
levels of A versus the low levels of A, whereas the coefficient, αi , in the model is the difference
between the marginal mean and the overall mean. So the Yates "effect" is twice the size of the
estimated coefficient αi in the model, which is also usually called the effect of the factor A.

The confusion is all in the notation used in the definition.

Let's look at another example in order to reinforce your understanding of the notation for these
types of designs. Here is an example in three dimensions, with factors A, B and C. Below is a
figure of the factors and levels as well as the table representing this experimental space.

In table 6.4 you can see the eight points coded by the factor levels +1 and -1. This example has
two replicates so n = 2. Notice that the Yates notation is included as the total of the two
replicates.

One nice feature of the Yates notation is that every column has an equal number of pluses and
minuses so these columns are contrasts of the observations. For instance, take a look at the A
column. This column has four pluses and four minuses, therefore, the A effect is a contrast
defined on page 216.

This is the principle that gives us all sorts of useful characterizations in these 2k designs.

In the example above the A, B and C each are defined by a contrast of the data observation
totals. Therefore you can define the contrast AB as the product of the A and B contrasts, the
contrast AC by the product of the A and C contrasts, and so forth.

Therefore all the two-way and three-way interaction effects are defined by these contrasts. The
product of any two gives you the other contrast in that matrix. (See Table 6.3 in the text.)

From these contrasts we can define the effect of A, B, and C, using these coefficients. The
general form of an effect for k factors is:

Effect = (1/2(k-1)n) [contrast of the totals]

The sum of the products of the contrast coefficients times the totals will give us the estimate of
the effects. See equations (6-11), (6-12), and (6-13).
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 5 of 27

We can also write the variance of the effect using the general form used previously. This would
be:

V ariance(Effect) = [1/(2(k−1) n)2 ]V (contrast), or

= [1/(2(k−1) n)2 ]2k nσ 2
= σ 2 /2(k−2) n
Also, we can write the sum of squares for the effects which looks like:

SS(effect) = (contrast)2 / 2kn

To summarize what we have learned in this lesson thus far, we can write a contrast of the totals
which defines an effect, we can estimate the variance for this effect and we can write the sum of
squares for an effect. We can do this very simply using Yates notation which historically has
been the value of using this notation.

6.2 - Estimated Effects and the Sum of

Squares from the Contrasts
How can we apply what we learned in the preceding section?

In general for 2k factorials the effect of each factor and interaction is:

Effect = (1/2(k-1)n) [contrast of the totals]

We also defined the variance as follows:

Variance(Effect) = σ2 / 2(k-2)n

The true but unknown residual variance σ2, which is also called the within cell variance, can be
estimated by the MSE.

If we want to test an effect, for instance, say A = 0, then we can construct a t-test which is the
effect over the square root of the estimated variance of the effect as follows:

Effect
t∗ = − −−− ∼ t(2k (n − 1))
√ MSE
n2k−2

where ~ means that it has a t distribution with 2k(n-1) degrees of freedom.

Finally, here is the equation for the sum of squares due to an effect to complete the story here:

SS(Effect) = (contrast of totals)2 / 2kn

Where does all of this come from? Each effect in a 2k model has one degree of freedom. In the
simplest case we have two main effects and an interaction. They each have 1 degree of
freedom. So the t statistic is the ratio of the effect over its estimated standard error (standard
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 6 of 27

deviation of the effect). You will recall that if you have a t statistic with ν degrees of freedom and
square it, you get an F distribution with one and ν degrees of freedom.

t2 (v) = F (1, v)
We can use this fact to confirm the formulas just developed. We see that the

∗ 2 (Effect)2
(t (v)) =
MSE/n2k−2
and from the definition of an F-test, when the numerator has 1 degree of freedom:

SS(Effect)/1 (contrast)2
F (1, v) = = k
MSE 2 n(MSE)

But from the definition of an Effect, we can write (Effect)2 = (contrast)2 / (n2k-1)2 and thus F(1, ν)
= (t*(ν))2 which you can show by some algebra or by calculating an example.

Hint: Multiply F(1, ν) by (2(k-1)n)2 / (2(k-1)n)2 and simplify.

Once you have these contrasts, you can easily calculate the effect, you can calculate the
estimated variance of the effect and the sum of squares due to the effect as well.

Creating a Factorial Design in Minitab

Let's use Minitab to help us create a factorial design and then add data so that we can analyze
it. Click on the 'Inspect' button to walk through this process using Minitab v.16. The data come
from Figure 6.1.

[1]

In Minitab we use the software under Stat > Design of Experiments to create our full factorial
design. We will come back to this command another time to look at fractional factorial and other
types of factorial designs.

In the example that was shown above we did not randomize the runs but kept them in standard
order for the purpose of the seeing more clearly the order of the runs. In practice you would
want to randomize the order of run when you are designing the experiment.

Once we have created a factorial design within the Minitab worksheet we then need to add the
response data so that the design can be analyzed. These response data, Yield, are the
individual observations not the totals. So, we again go to the Stat >> DOE >> Factorial menu
where we will analyze the data set from the factorial design.

We began with the full model with all the terms included, both the main effects and all of the
interactions. From here we were able to determine which effects were significant and should
remain in the model and which effects were not significant and can be removed to form a
simpler reduced model.
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 7 of 27

A Second Example - The Plasma Etch Experiment

Similar to the previous example, in this second industrial process example we have three
factors, A equals Gap, B = Flow, C = Power and our response y = Etch Rate. (The data are from
Table 6-4 in the text.) Once again in Minitab we will create a similar layout for a full factorial
design for three factors with two replicates which gives us 16 observations. Next, we add the
response data, Etch Rate, to this worksheet and analyze this data set. These are the results we
get:

The analysis of variance shows the individual effects and the coefficients, (which are half of the
effects), along with the corresponding t-tests. Now we can see from these results that the A
effect and C effect are highly significant. The B effect is not significant. In looking at the
interactions, AB, is not significant, BC is not significant, and the ABC are not significant.
However the other interaction, AC is significant.

This is a nice example to illustrate the purpose of a screening design. You want to test a number
of factors to see which ones are important. So what have we learned here? Two of these factors
are clearly important, A and C. But B appears not to be important either as a main effect or
within any interaction. It simply looks like random noise. B was the rate of gas flow across the
edging process and it does not seem to be an important factor in this process, at least for the
levels of the factor used in the experiment.

The analysis of variance summary table results show us that the main effects overall are
significant. That is because two of them, A and C, are highly significant. The two-way
interactions overall are significant. That is because one of them is significant. So, just looking at
this summary information wouldn't tell us what to do except that we could drop the 3-way
interaction.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 8 of 27

Now we can go back to Minitab and use the Analyze command under Design of Experiments
and we can remove all the effects that were seemingly not important such as any term having to
do with B in the model. In running this new reduced model we get:

For this model, all three terms are significant.

6.3 - Unreplicated 2^k Factorial Designs

These are 2k factorial designs with one observation at each corner of the "cube". An
unreplicated 2k factorial design is also sometimes called a "single replicate" of the 2k
experiment.

You would find these types of designs used where k is very large or the process for instance is
very expensive or takes a long time to run. In these cases, for the purpose of saving time or
money, we want to run a screening experiment with as few observations as possible. When we
introduced this topic we wouldn't have dreamed of running an experiment with only one
observation. As a matter of fact, the general rule of thumb is that you would have at least two
replicates. This would be a minimum in order to get an estimate of variation - but when we are in
a tight situation, we might not be able to afford this due to time or expense. We will look at an
example with one observation per cell, no replications, and what we can do in this case.

Where are we going with this? We have first discussed factorial designs with replications, then
factorial designs with one replication, now factorial designs with one observation per cell and no
replications, which will lead us eventually to fractional factorial designs. This is where we are
headed, a steady progression to designs with more and more factors, but fewer observations
and less direct replication.

Unreplicated 2k Factorial Designs

Let's look at the situation where we have one observation per cell. We need to think about
where[MathJax]/extensions/MathZoom.js
Loading the variation occurs within this design. These designs are very widely used. However,

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 9 of 27

there are risks…if there is only one observation at each corner, there is a high chance of an
unusual response observation spoiling the results. What about an outlier? There would be no
way to check if this was the case and thus it could distort the results fairly significantly. You have
to remind yourself that these are not the definitive experiments but simply just screening
experiments to determine which factors are important.

In these experiments one really cannot model the "noise" or variability very well. These
experiments cannot really test whether or not the assumptions are being met - again this is
another shortcoming, or the price of the efficiency of these experiment designs.

Spacing of Factor Levels in the Unreplicated 2k Factorial Designs

When choosing the levels of your factors, we only have two options - low and high. You can pick
your two levels low and high close together or you can pick them far apart. As most of you know
from regression the further apart your two points are the less variance there is in the estimate of
the slope. The variance of the slope of a regression line is inversely related the distance
between the extreme points. You can reduce this variance by choosing your high and low levels
far apart.

However, consider the case where the true underlying relationship is curved, i.e., more like this:

... and you picked your low and high level as illustrated above, then you would have missed
capturing the true relationship. Your conclusion would probably be that there is no effect of that
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 10 of 27

factor. You need to have some understanding of what your factor is to make a good judgment
about where the levels should be. In the end, you want to make sure that you choose levels in
the region of that factor where you are actually interested and are somewhat aware of a
functional relationship between the factor and the response. This is a matter of knowing
something about the context for your experiment.

How do we analyze our experiment when we have this type of situation? We must realize that
the lack of replication causes potential problems in statistical testing:

• Replication provides an estimate of "pure error" (a better phrase is an internal estimate of

error), and
• With no replication, fitting the full model results in zero degrees of freedom for error.

Potential solutions to this problem might be:

• Pooling high-order interactions to estimate error, (something we have done already in

randomized block design),
• Normal probability plotting of effects (Cuthbert and Daniels, 1959), and/or
• Dropping entire factors from the model and other methods.

Example of an Unreplicated 2k Design

The following 24 factorial (Example 6-2 in the text) was used to investigate the effects of four
factors on the filtration rate of a resin for a chemical process plant. The factors are A =
temperature, B = pressure, C = mole ratio (concentration of chemical formaldehyde), D = stirring
rate. This experiment was performed in a pilot plant.

Here is the dataset for this Resin Plant experiment. You will notice that all of these factors are
quantitative.

Notice also the use of the Yates notation here that labels the treatment combinations where the
high level for each factor is involved. If only A is high then that combination is labeled with the
Loading [MathJax]/extensions/MathZoom.js
small letter a. In total, there are 16 combinations represented.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 11 of 27

Here is a visual representation of this - it would be impossible to show this in a 4 dimensional

cube but here are two cubes which attempt to do the same thing.

...

Sequential Procedure for Strategically Finding a Model

Let's use the dataset (Ex6-2.MTW [2]) and work at finding a model for this data with Minitab...

[3]

Even with just one observation per cell, by carefully looking at the results we can come to some
understanding as to which factors are important. We do have to take into account that these
actual p-values are not something that you would consider very reliable because you are fitting
this sequence of models, i.e., fishing for the best model. We have optimized with several
decisions that invalidates the actual p-value of the true probability that this could have occurred
by chance.

This is one approach to assume that some interactions are not important and use this to test
lower order terms of the model and finally come up with a model that is more focused. Based on
this for this example that we have just looked at, we can conclude that following factors are
important, A, C, D, (of the main effects) and AC and AD of the two-way interactions.

Now I suggest you try this procedure and then go back and check to see what the final model
looks like. Here is what we get when we drop factor B and all the interactions that we decided
were not important:

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 12 of 27

The important factors didn't change much here. However, we have slightly higher degrees of
freedom for error. But now what the design looks like, by having dropped B totally, is that we
now have a 23 design with 2 replicates per cell. We have moved from a four factor with one
observation per cell, to a three factor with two observations per cell.

So, we have looked at two strategies here. The first is to take a higher order interactions out of
the model and use them as the estimate of error. Next, what we did at the end of the process is
drop that factor entirely. If a particular factor in the screening experiment turns out to be not
important either as a main effect or as part of any interaction we can remove it. This is the
second strategy, and for instance in this example we took out factor B completely from the
analysis.

Graphical Approaches to Finding a Model

Let's look at some more procedures - this time graphical approaches for us to look at our data in
order to find the best model. This technique is really cool. Get a cup of coffee and click:

[4]

Normal Probability Plot for the Effects

Having included all the terms back into a full model we have shown how to produce a normal
plot. Remember that all of these effects are 1 degree of freedom contrasts of the original data,
each one of these is a linear combination of the original observations, which are normally
distributed with constant variance. Then these15 linear combinations or contrasts are also
normally distributed with some variance. If we assume that none of these effects are significant,
the null hypothesis for all of the terms in the model, then we simply have 15 normal random
variables, and we will do a normal random variable plot for these. That is what we will ask
Minitab
Loading to plot for us. We get a normal probability plot, not of the residuals, not of the original
[MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 13 of 27

observations but of the effects. We have plotted these effects against what we would expect if
they were normally distributed.

In the middle - the points in black, they are pretty much in a straight line - they are following a
normal distribution. In other words, their expectation or percentile is proportionate to the size of
the effect. The ones in red are like outliers and stand away from the ones in the middle and
indicate that they are not just random noise but there must be an actual affect. Without making
any assumptions about any of these terms this plot is an overall test of the hypothesis based on
simply assuming all of the effects are normal. This is a very helpful - a good quick and dirty first
screen - or assessment of what is going on in the data, and this corresponds exactly with what
we found in our earlier screening procedures.

The Pareto Plot

Let's look at another plot - the Pareto plot. This is simply a plot that can quickly show you what is
important. It looks at the size of the effects and plots the effect size on a horizontal axis ranked
from largest to smallest effect.

Having dropped some of the terms out of the model, for instance the three and four way
interactions, Minitab plots the remaining effects, but now it is the standardized effect. Basically it
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 14 of 27

is plotting the t-value, the effect over its standard deviation and then plotting it in ranked order. It
also displays the t critical point as a red line at alpha = 0.05.

Effects and Interaction Plots

Another Minitab command that we can take a look at is the subcommand called Factorial Plots.
Here we can create plots for main effects telling Minitab which factors you want to plot. As well
you can plot two-way interactions. Here is a plot of the interactions (which are more interesting
to interpret), for the example we've been looking at:

You can see that the C and D interaction plot the lines are almost parallel and therefore do not
indicate interaction effects that are significant. However the other two combinations, A and C
and A and D, indicate that significant interaction exists. If you just looked at the main effects plot
you would likely miss the interactions that are obvious here.

Checking Residuals Using Minitab's Four in One Plot

We have reduced the model to include only those terms that we found were important. Now we
want to check the residuals in order to make sure that our assumptions are not out of line with
any conclusions that we are making. We can ask Minitab to produce a Four in One residuals plot
which, for this example, looks like this:

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 15 of 27

In visually checking the residuals we can see that we have nothing to complain about. There
does not seem to be any great deviation in the normal probability plot of the residuals. There's
nothing here that is very alarming and it seems acceptable. In looking at the residuals versus the
fitted values plot in the upper right of this four in one plot - except for the lower values on the left
where there are smaller residuals and you might be somewhat concerned here, the rest do not
set off any alarms - but we will come back to this later.

Contour and Surface Plots

We may also want contour plots of all pairs of our numeric factors. These can be very helpful to
understand and present the relationship between several factors on the response. The contour
plots below for our example show the color coded average response over the region of interest.
The effect of these changes in colors is to show the twist in the plane.

In the D*C plot area you can see that there is no curvature in the colored areas, hence no
evidence of interaction. However, if you look at C*A display you can see that if C is low you get a
dramatic change. If C is high it makes very little difference. In other words, the response due to
A depends on the level of C. This is what the interaction means and it shows up nicely in this
contour plot.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 16 of 27

Finally, we can also ask Minitab to give us a surface plot. We will set this up the same way in
Minitab and this time Minitab will show the plot in three dimensions, two variables at a time.

The surface plot shows us the same interaction effect in three dimensions in the twisted plane.
This might be a bit easier to interpret. In addition you can ask Minitab to provide you with 3-D
graphical tools that will allow you to grab these boxes and twist them around so that you can
look at these boxes in space from different perspectives. Pretty cool! Give it a try. These
procedures are all 'illustrated in the "Inspect" Flash movie at the beginning of this section.

Another Example - The Drilling Example 6.3

This is another fairly similar example to the one we just looked at. This drilling example
(Example 6-3) is a 24 design - again, the same design that we looked at before. It is originally
from C. Daniel, 1976. It has four factors, A = Drill load, B = Flow of a lubricant, C = Speed of
drill, D = Type of mud, Y is the Response - the advance rate of the drill, (how fast can you drill
an oil or gas well?).

We've used Minitab to create the factorial design and added the data from the experiment into
the Minitab worksheet. First, we will produce a normal probability plot of the effects for this data
with all terms included in a full model.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 17 of 27

Here's what it looks like. It shows a strange pattern! No negative and all positive effects. All of
the black dots are in fairly straight order except for perhaps the top two. If we look at these
closer we can see that these are the BD and the BC terms, in addition to B, C, and D as our
most important terms. Let's go back to Minitab and take out of our model the higher order
interactions, (i.e. the 3-way and 4-way interactions), and produce this plot again (see below) just
to see what we learn.

The normal probability plot of residuals looks okay. There is a gap in the histogram of other
residuals but it doesn't seem to be a big problem.

When we look at the normal probability plot below, created after removing 3-way and 4-way
interactions, we can see that now BD and BC are significant.

We can also see this in the statistical output of this model as shown below:

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 18 of 27

The combined main effects are significant as seen in the combined summary table. And the
individual terms, B, C, D, BC and BD, are all significant, just as shown on the normal probability
plot above.

Now let's go one step farther and look at the completely reduced model. We'll go back into
Minitab and get rid of everything except for the significant terms. Here is what you get:

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 19 of 27

What do you think?

Residuals versus the fitted values plot in the upper right-hand corner has now a very distinct
pattern. It seems to be a classic as the response gets larger the residuals get more spread
apart.

What does this suggest is needed? For those of you who have studied heteroscedastic variance
patterns in regression models you should be thinking about possible transformations.

A transformation - the large values are more variable than smaller values. But why does this
only show up now? Well, when we fit a full model it only has one observation per cell and there's
no pure way to test for residuals. But when we fit a reduced model, now there is inherent
replication and this pattern becomes apparent.

Take a look at the data set and you will find the square root and the log already added in order
to analyze the same model using this transformed data. What do you find happens?

6.4 - Transformations
When you look at the graph of the residuals as shown below you can see that the variance is
small at the low end and the variance is quite large on the right side producing a fanning effect.
Consider the family of transformations that can be applied to the response yij.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 20 of 27

Transformations towards the bottom of the list are stronger in how they shrink large values more
than they shrink small values that are represented on the plot. This pattern of the residuals is
one clue to get you to be thinking about the type of transformations you would select.

The other consideration and thinking about transformations of the response yij is what it does to
the relationship itself. Some of you will recall from other classes the Tukey one-degree-of-
freedom test for interaction. This is a test for interaction where you have one observation per cell
such as with a randomized complete block design. But with one observation per cell and two
treatments our model would be :

Yijk = μ + αi + βj + (αβ)ij + ϵijk

where,

i = 1 ... a,
j = 1 ... b, with
k = 1 ... 1, (only have one observation per cell)

There is no estimate of pure error so we cannot fit the old model. The model proposed by
Tukey's has one new parameter (γ) gamma :

Yij = μ + αi + βj + γαi βj + ϵij

This single parameter, gamma, is the 1 degree of freedom term and so our error, εij, has (a-1)
(b-1) -1 degrees of freedom. This model allows for just a single additional parameter which is
based on a multiplicative effect on the two factors.

Now, when is this applicable?

Let's go back to the drill rate example (Ex6-3.MTW [5]) where we saw the fanning effect in the
plot of the residuals. In this example B, C and D were the three main effects and there were two
interactions BD and BC. From Minitab we can reproduce the normal probability plot for the full
model.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 21 of 27

But let's first take a look at the residuals versus our main effects B, C and D.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 22 of 27

All three of these residuals versus the main effects show same pattern, the large predicted
values tend to have larger variation.

Next, what we really want to look at is the factorial plots for these three factors, B, C and D and
the interactions among these, BD and BC.

What you see in the interaction plot above is a pattern that is non-parallel showing there is
interaction present. But, from what you see in the residual graph what would you expect to see
on this factor plot?

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 23 of 27

The tell-tale pattern that is useful here is an interaction that does not have crossing lines - a
fanning effect - and it is exactly the same pattern that allows the Tukey model to fit. In both
cases, it is a pattern of interaction that you can remove by transformation. If we select a
transformation that will shrink the large values more than it does the small values and the overall
result would be that we would see less of this fan effect in the residuals.

We can look at either the square root or log transformation. It turns out that the log
transformation is the one that seems to fit the best. On a log scale it looks somewhat better - it
might not be perfect but it is certainly better than what we had before.

Let's also look at the analysis of variance.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 24 of 27

The overall main effects are still significant. But the two 2-way interactions effects combined are
no longer significant, and individually, the interactions are not significant here either. So, the log
transformation which improved the unequal variances pulled the higher responses down more
than the lower values and therefore resulted in more of a parallel shape. What's good for
variance is good for a simple model. Now we are in a position where we can drop the
interactions and reduce this model to a main effects only model.

Now our residual plots are nearly homoscedastic for B, C and D. See below...

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 25 of 27

Serendipity - good things come in packages! When you pick the correct transformation, you
sometimes achieve constant variance and a simpler model.

Many times you can find a transformation that will work for your data - giving you a simpler
analysis but it doesn't always work.

Transformations are typically performed to:

• Stabilize variance - to achieve equal variance

• Improve normality - this is often violated because it is easy to have an outlier when
variance is large which can be 'reined in' with a transformation
• Simplify the model

Sometimes transformations will solve a couple of these problems.

Is there always a transformation that can be applied to equalize variance? Not really ... there are
two approaches to solving this question. First, we could use some non-parametric method.
Although non-parametric methods have fewer assumptions about the distribution, you still have
to worry about how you are measuring the center of the distribution. When you have a non-
parametric situation you may have a different shaped distribution in different parts of the
experiment. You have to be careful about using the mean in one case, and the media in
another ... but that is one approach.

The other approach is a weighted analysis, where you weight the observations according to the
inverse of their variance. There are situations where you have unequal variation for maybe a
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 26 of 27

known reason or unknown reason, but if you have repeated observations and you can get
weights, then you can do a weighted analysis.

It is this course author's experience many times you can find a transformation when you have
this kind of pattern. Also, sometimes when you have unequal variance you just have a couple of
bad outliers, especially when you only have one or a few observations per cell. In this case it is
difficult to the distinguish whether you have a couple of outliers or the data is heteroscedastic - it
is not always clear.

Empirical Selection of Lambda

Prior (theoretical) knowledge or experience can often suggest the form of a transformation.
However, another method for the analytical selection of lambda for the exponent used in the
transformation is the Box-Cox (1964). This method simultaneously estimates the model
parameters and the transformation parameter lambda.

Box-Cox method is implemented in some statistical software applications.

Example 6.4

This example is a four factor design in a manufacturing situation where injection molding is the
focus. Injection molding is a very common application in industry; a 2k design where you have
many factors influencing the quality which is measured by how many defects are created by the
process. Almost anything that you can think of which have been made out of plastic was created
through the injection molding process.

See the example in (Ex6-4.MTW [6])

In this example we have four factors again: A = temperature of the material, B = clamp time for
drying, C = resin flow, and D = closing time of the press. What we are measuring as the
response is number of defects. This is recorded as an index of quality in terms of percent. As
you look through the data in Figure 6.29 (7th edition) you can see percent of defects as high as
15.5% or as low as 0.5%. Let's analyze the full model in Minitab.

The normal probability plot of the effects shows us that two of the factors A and C are both
significant and none of the two-way interactions are significant.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 27 of 27

What we want to do next is look at the residuals vs. variables A, B, C, D in a reduced model with
just the main effects as none of the interactions seemed important.

For each factor you see that the residuals are more dispersed (higher variance) to the right than
to the left. Overall, however, the residuals do not look too bad and the normal plot also does not
look too bad. When we look at the p-values we find that A and C are significant but B and D are
not.

But there is something else that can be learned here. The point of this example is that although
the B factor is not significant as it relates to the response, percentage of product defects -
however, if you are looking for a recommended setting for B you should use the low level for B.
A and C, are significant and will reduce the number of defects. However, by choosing B at the
low level you will produce a more homogeneous product, products with less variability. What is
important in product manufacturing is not only reducing the number of defects but also
producing products that are uniform. This is a secondary consideration that should be taken into
account after the primary considerations related to the percent of product defects.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/34

Links:
[1] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson06/L06_factorial_design_viewlet_swf.html',
'l06_factorial_design', 718, 668 );
[2]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson06/Ex6-2.MTW
[3] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson06/L06_unreplicated_viewlet_swf.html',
'l06_unreplicated', 718, 668 );
[4] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson06/L06_graphical_viewlet_swf.html',
'l06_graphical', 718, 668 );
[5]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson06/Ex6-3.MTW
[6] https://onlinecourses.science.psu.edu/stat503/node/38/graphics/EX6-4.MTW

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/34/ 4/18/2019
 Page 1 of 16

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 7: Confounding and Blocking in 2^k Factorial Designs

Lesson 7: Confounding and Blocking in

2^k Factorial Designs
Introduction

In Lesson 4 we discussed blocking as a method for removing extraneous sources of

variation. In this lesson we consider blocking in the context of 2k designs. We will then
make a connection to confounding, and show a surprising application of confounding where
it is beneficial rather than a liability.

Learning Objectives & Outcomes

By the end of this lesson, readers are supposed to understand:

• Concept of Confounding
• Blocking of replicated 2k factorial designs
• Confounding high order interaction effects of the 2k factorial design in 2p blocks
• How to choose the effects to be confounded with blocks
• That a 2k design with a confounded main effect is actually a Split Plot design
• The concept of Partial Confounding and its importance for retrieving information on
every interaction effect

Blocking in Replicated Designs

In 2k replicated designs where we have n replications per cell and perform a completely
randomized design we randomly assign all 2k times n experimental units to the 2k treatment
combinations. Alternatively, when we have n replicates we can use these n replicates as
blocks, and assign the 2k treatments to the experimental units within each of the n blocks. If
we are going to replicate the experiment anyway, at almost no additional cost, you can
block the experiment, doing one replicate first, then the second replicate, etc. rather than
completely randomize the n times 2k treatment combinations to all the runs.

There is almost always an advantage to blocking when we replicate the treatments. This is
true even if we only block using time due to the order of the replicates. However, there are
often many other factors that we have available as potential sources of variation that we
can include as a block factor, such as batches of material, technician, day of the week, or
time of day, or other environmental factors. Thus if we can afford to replicate the design
then it is almost always useful to block.
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 2 of 16

To give a simple example, if we have four factors, the 2k design has 16 treatment
combinations, so say we plan to do just two replicates of the design. Without blocking, the
ANOVA has 24 = 16 treatments, but with n = 2 replicates, the MSE would have 16 degrees
of freedom. If we included a block factor, with two levels, the ANOVA would use one of
these 16 degrees of freedom for the block, leaving 15 degrees of freedom for MSE. Hence
the statistical cost of blocking is really the loss of one degree of freedom for error, and the
potential gain if the block explains significant variation would be to reduce the size of the
MSE and thereby increase the power of the tests.

The more interesting case that we will consider next is when we have an unreplicated
design. If we are only planning to do one replicate, can we still benefit from the advantage
ascribed to blocking our experiment?

7.1 - Blocking in an Unreplicated Design

We begin with a very simple replicated example of blocking. Here we have 22 treatments
and we have n = 3 blocks. In the graphic below the treatments are labeled using the
standard Yates notation. Here the 22 treatments are the full set of treatment combinations
so we can simply put each replicate within a block and assign them in this way.

We can use the Minitab software to construct this design as seen in the video below.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 3 of 16

Now let’s consider the case when we don't have any replicates, hence when we only have
one set of treatment combinations. We go back to the definition of effects that we defined
before. We did this using following table, where {(1), a, b, ab} is the set of treatment
combinations, and A, B, and AB are the effect contrasts:

trt A B AB
(1) -1 -1 1
a 1 -1 -1
b -1 1 -1
ab 1 1 1

The question is: what if we want to block this experiment? Or, more to the point, when it is
necessary to use blocks, how would we block this experiment?

If our block size is less than four we are only going to consider, in this context of 2k
treatments, block sizes in the same family, i.e. 2p number of blocks. So in the case of this
example let's use blocks of size 2, which is 21. If we have blocks of size two then we must
put two treatments in each block. One example would be twin studies where you have two
sheep from each ewe. The twins would have homogeneous genetics and the block size
would be two for the two animals. Another example might be two-color micro-arrays where
you have only two colors in each micro-array.

So now the question: How do we assign our four treatments to our blocks of size two?

In our example each block will be composed of two treatments. The usual rule is to pick an
effect you are least interested in, and this is usually the highest order interaction, as a
means of specifying how to do blocking. In this case it is the AB effect that we will use to
determine our blocks. As you can see in the table below we have used the high level of AB
to denote Block 1, and the low-level of AB to denote Block 2. This determines our design.

trt A B AB Block
(1) -1 -1 1 1
a 1 -1 -1 2
b -1 1 -1 2
ab 1 1 1 1

Now, using this design we can assign treatments to blocks. In this case treatment (1) and
treatment ab will be in the first block, and treatment a and treatment b will be in the second
block.

Blocks of size 2

Block 1 2
Typesetting math: 100% AB + -

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 4 of 16

(1) a
ab b

This design confounds blocks with the AB interaction. You can see this by these contrasts -
the comparison between block 1 and Block 2 is the same comparison as the AB contrast.
Note that the A effect and the B effect are orthogonal to the AB effect. This design gives
you complete information on the A and the B main effects, but it totally confounds the AB
interaction effect with the block effect.

Although our block size is fixed at size = 2 we still might want to replicate this experiment in
addition. What we have above is two blocks which is one unit of the experiment. We could
replicate this design additionally let's say r times and each replicate of the design would be
2 blocks of the design laid out in this way.

We show how to construct this with four replicates. Review the movie below to see how this
occurs in Minitab.

7.2 - The 2^3 Design

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 5 of 16

Let's look now at the 23 design. Here we have 8 treatments and we could create designs
with blocks of size 2p - which could either be blocks of size 4 or 2. As before, we can write
this out in a table as:

trt I A B C AB AC BC ABC
(1) + - - - + + + -
a + + - - - - + +
b + - + - - + - +
ab + + + - + - - -
c + - - + + - - +
ac + + - + - + - -
bc + - + + - - + -
abc + + + + + + + +

In the table above we have defined our seven effects: three main effects {A, B, C}, three
2-way interaction effects {AB, AC, BC}, and one 3-way interaction effect {ABC}. We need to
define our blocks next by selecting an effect that we are willing to give up by confounding it
within the blocks. Let's first look at an example where we let the block size = 4.

Now we need to ask ourselves, what is typically the least interesting effect? The highest
order interaction. Do we will use the contrast of the highest order interaction, the three-
way, as the effect to guide the layout of our blocks.

trt I A B C AB AC BC ABC Block

(1) + - - - + + + - 1
a + + - - - - + + 2
b + - + - - + - + 2
ab + + + - + - - - 1
c + - - + + - - + 2
ac + + - + - + - - 1
bc + - + + - - + - 1
abc + + + + + + + + 2

Under the ABC column, the - values will be placed in Block 1, and the + values will be
placed in Block 2. Thus we can layout the design by defining the two blocks of four
observations like this:

Block 1 2
ABC - +
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 6 of 16

(1) a
ab b
ac c
bc abc

Let's take a look at how Minitab would run this process ...

What if we have 23 treatments but we want the block size to be 2?

Now for each replicate we need four blocks with only two treatments per block.

Thought Questions: How should we assign our treatments? How many and which effects
must you select to confound with the four blocks?

To define the design for four blocks we need to select two effects to confound, and then we
will get four combinations of those two effects.

Note: Is there a contradiction here? If we pick two effects to confound that is only two
degrees of freedom. But how many degrees of freedom are there among the four blocks?
Three! So, if we confound two effects then we have automatically also confounded the
interaction between those two effects. That is simply a result of the structure used here.
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 7 of 16

What if we first select ABC as one of the effects? Then, it would seem logical to pick one of
the 2-way interactions as the other confounding factor. Let's say we use AB. If we do this,
remember, we also confound the interaction between these two effects. What is the
interaction between ABC and AB. It is C. We can see this by multiplying the elements in the
columns for ABC and AB. Try it and you get the same coefficients as you have in the
column for C. This is called the generalized interaction. Although it intuitively seemed as
though ABC and AB would be a good choice, it is not because it also confounds the main
effect C.

Another choice would be to pick two of the 2-way interactions such as AB and AC. The
interaction of these is BC. In this case you have not confounded a main effect, but instead
have confounded the three two-way interactions. The four combinations of the AB and AC
interactions define the four blocks as seen in this color coded table.

trt I A B C AB AC BC ABC Block

(1) + - - - + + + - 4
a + + - - - - + + 1
b + - + - - + - + 3
ab + + + - + - - - 2
c + - - + + - - + 2
ac + + - + - + - - 3
bc + - + + - - + - 1
abc + + + + + + + + 4

Look under the AB and the AC columns. Where there are - values for both AB and AC
these treatments will be placed in Block 1. Where there is a + value for AB and a - value for
AC these treatments will be placed in Block 2. Where there is a - value for AB and a +
value for AC these treatments will be placed in Block 3. And finally, where there are +
values for both AB and AC these treatments will be placed in Block 4. From here we can
layout the design separating the four blocks of two observations like this:

Block 1 2 3 4
AB, AC -, - +, - -, + +, +
a ab b (1)
bc c ac abc

Let's take a look at how Minitab would run this process ...

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 8 of 16

For the 23 design the only two possibilities are either block sizes of two or four. When we
look at more than eight treatments or 23, then we have more combinations possible. We
typically want to confound the highest order of interaction possible remembering that all
generalized interactions are also confounded. This is a property of the geometry of
designs.

In the next lesson we will look at how we can analyze the data if we take replications of
these basic designs, considering one replicate as just the basic building block. This is
typically determined by the fact that the block size is usually imposed by some cost or size
restrictions on the experiment. However, given adequate resources you can replicate that
whole experiment multiple times. So then the question becomes how to analyze these
designs and how do we pull out the treatment information.

7.3 - Blocking in Replicated Designs

In the previous section, we saw a 22 treatment design with 4 runs constructed in two blocks
confounded with the AB contrast. We also saw a 23 design constructed in two blocks, with
ABC confounded with blocks. We say this is a 23 design in 21 blocks of size 22 per
replicate. And we also saw a 23 design in 22 = 4 blocks of size 21 = 2 per replicate with
effects AB, AC, and therefore AB × AC = A2BC = BC confounded with blocks.
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 9 of 16

Now let’s consider this last situation when we have n = 3 replicates of this basic design with
b = 4 blocks. We can write a model:

Yijklm = μ + ri + bj(i) + αk + βl + γm +. . .
where “i” is the index for replicates and “j” is the index for blocks within the replicates. “k”,
“l” and “m” are indices for the different treatment factors.

AOV df
Rep n-1 =3-1=2
Blk(Rep) n(b - 1) = 3(4 - 1) = 9
A 2-1 1
B 2-1 1
C 2-1 1
ABC 2-1 1
Error (n - 1)*4 8
Total n*23 - 1 23

Now we consider another example: in figure 7.3 of the text we see four replicates with ABC
confounded in each of the four replicates. The ANOVA for this design is seen in table 7.5
which shows that the Block effect (Block 1 vs. Block 2) is equivalent to the ABC effect and
since there are four replicates of this basic design, we can extract some information about
the ABC effect, and indeed test the hypothesis of no ABC effect, by using the Rep × ABC
interaction as error.

See the analysis of this design using Minitab:

Stat >> ANOVA >> General Linear Model >>

and fitting the following model:

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 10 of 16

If Reps is specified as a random effects factor in the model, as above, GLM will produce
the correct F-tests based on the Expected Means Squares. The reason is analogous to the
RCBD with random blocks (Reps) and a fixed treatment (ABC). The topic of random factors
is completely covered in chapter 13 of the text book

For Minitab Stat >> ANOVA >> GLM to analyze this data, you need to first construct a
pseudo-factor called "ABC" which is constructed by multiplying the levels of A, B, and C
using 'Calculator' under the 'Data' menu in Minitab. Click on the 'Inspect' button below
which will walk you through this process using Minitab v.16.

[1]

In addition you can open this Minitab project file 2-k-confound-ABC.MPJ [2] and review the
steps leading to the output. The response variable Y is random data simply to illustrate the
analysis.

Here is an alternative way to analyze this design using the analysis portion of the fractional
factorial software in Minitab v.16.

[3]

A similar exercise can be done to illustrate the confounded situation where the main effect,
say A, is confounded with blocks. Again, since this is a bit nonstandard, we will need to
generate a design in Minitab using the default settings and then edit the worksheet to
create the confounding we desire and analyze it in GLM.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 11 of 16

7.4 - Split-Plot Example – Confounding a

Main Effect with blocks
Let us consider three replicates, and at each replicate we have two large fields. A diagram
of this would look like this:

We will randomly assign the low (-1) or high (+1) level of factor A to each of the two fields.
In our example A = +1 is irrigation, and A = -1 is no irrigation. We will randomly assign the
levels of A to the two fields in each replicate. Then the experiment layout would look like
this for one replicate:

Similar to the previous example, we would then assigned the treatment combinations of
factors B and C to the four experimental units in each block. We could call these
experimental units plots -- or using the language of split plot designs -- the blocks are
whole plots and the subplots are split plots.

The analysis of variance is similar to what we saw in the example above except we now
have A rather than ABC confounded with blocks.

See the Minitab project file 2-K-Split-Plota.MPJ [4] as an example. In addition, here is a
viewlet that will walk you through this example using Minitab v.16.

[5]

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 12 of 16

7.5 - Blocking in 2^k Factorial Designs

Now we will generalize what we have shown by example. We will look at 2k designs in 2p
blocks of size 2k-p. We do this by choosing k and if we want to confound the design in 2p
blocks then we need to choose p effects to confound. Then, due to the intereactions among
these effects, we get 2p-1 effects confounded with blocks.

To illustrate this, if p = 2 then we have 2p = four blocks, and thus 2p-1 = 3 effects
confounded, i.e., the 2 effects we chose plus the interaction between these two. In general,
we choose p effects and in addition to the p effects we choose, 2p-p-1 other effects are
automatically confounded. We will call these "generalized interactions" which are also
confounded.

Earlier we looked at a couple of examples - for instance when k = 3 and p = 2. We chose

ABC and AB. Then the ABC × AB = A2B2C = C which was also confounded. This shows
that the generalized interaction can be a main effect, i.e. the generalized interaction affect
can be a lower order term. This is not a good outcome. A better outcome that we settled on
was to pick two 2-way interactions, AB and AC, which gave us AB × AC = A2BC which =
BC, another 2-way interaction. In this case we have all three 2-way interactions
confounded, but all the main effects were estimable.

7.6 - Example 1
Let's take another example where k = 4, and p = 2. This is one step up in the number of
treatment factors. And now we have block size = 24 - 2 or 4. Again, we have to choose two
effects to confound. We will show three cases to illustrate this.

a. Let's try ABCD and a 3-way, ABC. This implies ABCD × ABC = A2B2C2D = D
is also confounded. We usually do not want to confound a main effect. It seems
that if you reach too far then you fall short. So, the question is: what is the right
compromise?

b. We could try ABCD and just AB. In this case we get ABCD × AB = A2B2CD =
CD. Here we have the 4-way interaction and just two of the 2-way interactions
confounded. Can we do better than this? Do you know that one or more of your
2-way interaction effects are not important? This is something you probably
don't know, but you might. In this case you could pick this interaction and very
carefully assign treatments based on this knowledge.

c. One more try. How about confounding two 3-way interactions? What if we
use ABC and BCD. This would give us the interactions of those, ABC × BCD =
AB2C2D which = AD.

Which of these three attempts is better? The first try (a) is definitely not good because it
confounds the main effect. So, which of the second or third do you prefer? The third (c) is
probably the best because it has the fewest lower order interactions confounded. Generally
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 13 of 16

it is assumed that the higher order interactions are less important, so this makes the (c)
case the best choice. Both cases (b) and (c) confound 2-way interactions but the (b) case
confounds two of them and the (c) case only one.

If we look at Minitab the program defaults are always set to choose the best of these
options. Use this short viewlet to see how Minitab v.17 selects these:

7.7 - Example 2
Let's try an example where k = 5, and p = 2.

a. If we choose to confound two 4-way interactions ABCD and BCDE, this would give us
ABCD × BCDE = AB2C2D2E = AE, confounded as well, which is a 2-way interaction. Not so
good.

If we choose ABC and CDE, this would give us ABC × CDE = ABC2DE = ABDE. So, with
this choice we are confounding the higher level 4-way interaction and two 3-way
interactions instead of the 2-way interaction as above.

Let's see what Minitab chooses...

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 14 of 16

If you were planning to replicate one of these designs, you would not need to use the same
three factors for blocking in each replicate of the design, but instead could choose a
different set of effects to use for each replicate of the experiment. More on that later.

7.8 - Alternative Method for Assigning

Treatments to Blocks
We began this section by looking at the +'s and -'s that were assigned by looking at
whether the treatment level was high or low. And in our simplest example we looked at our
contrast as +1's and -1's and used these to determine which treatments were assigned to
which blocks.

An alternative to using the -'s and +'s is to use 0 and 1. In this case, the low level is 0 and
the high level is 1. You can think of this method as just another finite math procedure that
can be used to determine which treatments go in which block. We introduce this here
because as we will see later, this alternative method generalizes to designs with more than
two levels.

Here is a 23 design using this notation:

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 15 of 16

X3 X2 X1
LABC LBC LAC LAB C B A
0 0 0 0 0
1 1 0 0 1
0 1 0 1 0
1 0 0 1 1
1 0 1 0 0
0 1 1 0 1
1 1 1 1 0
0 0 1 1 1

Defining Contrasts

LAB = X1 + X2 (mod 2)
LAC = X1 + X3 (mod 2)
LBC = X2 + X3 (mod 2)
LABC = X1 + X2 + X3 (mod 2)

Note: (mod 2) refers to modular arithmetic where you divide a number by 2 and keep the
remainder, e.g., ( 5 (mod 2) = 1)

If you look at LAB all we are doing here is just summing the 0 and 1 combinations,
therefore, LAB = the sum of the row of 0's and 1's for AB (in blue for the first row only). What
we are doing is defining the linear combinations using modular 2 arithmetic in this way.

If we want to construct a design for k = 3, p = 2 by choosing AB and AC as our defining

contrasts then we would construct our design in the following manner:

4 3 2 1 Block
1, 1 0, 1 1, 0 0, 0 LAB, LAC
a ab b (1)
bc c ac abc

We are using LAB and LAC to define our blocks, so, what we need to do is exactly what we
did before, but this time we are using the 0's and 1's to determine the layout for the design.
We are simply using a different coding mechanism here for determining the design layout.

Why is this important?

For two level designs both methods work the same. You can either use the +'s and -'s as
the two levels of the factor to divide the treatment combinations into blocks, or you can use
zero and one, which is simply a different way to do this and gives us a chance to define the
contrastsmath:
Typesetting where:
100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 16 of 16

L = a1X1 + a2X2 + a3X3 (mod 2)

where ai is the exponent of the ith factor in the effect to be confounded (either a 0 or a 1 in
each case) and Xi is the level of the ith factor appearing in a particular treatment
combination.

Both approaches will give us the same set of treatment combinations in blocks. These
functions translate the levels of A and B to the levels of the AB interaction.

When we get to designs with more than two levels using +'s and -'s doesn't work.
Therefore, we need another method and using this 1's and 0's approach generalizes. We
will come back to this method when we look at 3 level designs - but we will get to that later
in Lesson 9.

Partial Confounding

In the above designs, we had to select one or more effects that we were willing to confound
with blocks, and therefore not be able to estimate. Generally, we should have some prior
knowledge about which effects to neglect or which effects are zero. Even if we do replicate
a blocked factorial design, we would not be able to obtain good intra-block estimates the
effect(s) which are confounded with blocks. To avoid this issue, there is a method of
confounding called partial confounding” which is widely used.

In partial confounding, the experimenter uses a different interaction effect to be confounded

with blocks throughout different replicates. In this way, information regarding each
interaction effect which is confounded in one of the replicates can be retrieved from the
remaining replicates. Figure 7.7 in the text book shows a partial confounding of 23 design
where ABC, AB, BC and AC are confounded with blocks in the first through fourth
replicates, respectively. Since each interaction is unconfounded in three-quarters of
replicates, ¾ is the relative information for the confounded effects. The analysis is shown
in Table 7.10. Example 7.3 in the text book illustrates a 23 design with partial confounding.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/39

Links:
[1] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson07/L07_2k_confnd_ABC_viewlet_swf.html',
'l07_2k_confnd_abc', 704, 652 );
[2]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson07/2-k-
confound-ABC.MPJ
[3] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson07/L07_2k_confnd_default_viewlet_swf.html',
'l07_2k_confnd_default', 704, 652 );
[4]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson07/2-K-
Split-Plota.MPJ
[5] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson07/L07_split_plot_viewlet_swf.html',
'l07_split_plot', 704, 652 );

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/39/ 4/18/2019
 Page 1 of 28

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 8: 2-level Fractional Factorial Designs

Lesson 8: 2-level Fractional Factorial Designs

Learning Objectives & Outcomes

The learning objectives for this lesson include:

• Understanding the application of Fractional Factorial designs, one of the most important
designs for screening
• Becoming familiar with the terms “design generator”, “alias structure” and “design resolution”
• Knowing how to analyze fractional factorial designs in which there aren’t normally enough
degrees of freedom for error
• Becoming familiar with the concept of “foldover” either on all factors or on a single factor and
application of each case
• Being introduced to “Plackett-Burman Designs” as another class of major screening designs

Introduction to Fractional Factorial Designs

What we did in the last chapter is consider just one replicate of a full factorial design and run it in
blocks. The treatment combinations in each block of a full factorial can be thought of as a fraction of
the full factorial.

In setting up the blocks within the experiment we have been picking the effects we know would be
confounded and then using these to determine the layout of the blocks.

We begin with a simple example.

In an example where we have k = 3 treatments factors with 23 = 8 runs, we select 2p = 2 blocks, and
use the 3-way interaction ABC to confound with blocks and to generate the following design.

trt A B C AB AC BC ABC I
(1) - - - -
a + - - +
b - + - +
ab + + - -
c - - + +
ac + - + -
bc - + + -
abc + + + +

Here are the two blocks that result using the ABC as the generator:
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 2 of 28

Block 1 2
ABC - +
(1) a
ab b
ac c
bc abc

A fractional factorial design is useful when we can't afford even one full replicate of the full factorial
design. In a typical situation our total number of runs is N = 2k-p, which is a fraction of the total
number of treatments.

Using our example above, where k = 3, p = 1, therefore, N = 22 = 4

So, in this case, either one of these blocks above is a one half fraction of a 23 design. Just as in the
block designs where we had AB confounded with blocks - where we were not able to say anything
about AB. Now, where ABC is confounded in the fractional factorial we can not say anything about
the ABC interaction.

Let's take a look at the first block which is a half fraction of the full design. ABC is the generator of
the 1/2 fraction of the 23 design. Now, take just the fraction of the full design where ABC = -1 and we
place it within its own table:

trt A B C AB AC BC ABC I
(1) - - - + + + - +
ab + + - + - - - +
ac + - + - + - - +
bc - + + - - + - +

Notice the contrast defining the main effects (similar colors) - there is an aliasing of these effects.
Notice that columns with the same color are just -1 times one another.

In this half fraction of the design we have 4 observations, therefore we have 3 degrees of freedom to
estimate. The degrees of freedom estimate the following effects: A - BC, B - AC, and C - AB. Thus,
this design is only useful if the 2-way interactions are not important, since the effects we can
estimate are the combined effect of main effects and 2-way interactions.

This is referred to as a Resolution III Design. It is called a Resolution III Design because the
generator ABC has three letters, but the properties of this design and all Resolution III designs are
such that the main effects are confounded with 2-way interactions.

Let's take a look at how this is handled in Minitab:

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 3 of 28

This design is only useful if you can be assured that the 2-way interactions are not important. If this
is the case then you will find Resolution III designs to be very useful and efficient. When runs are
expensive and factors are plentiful these are popular designs.

8.1 - More Fractional Factorial Designs

We started our discussion with a single replicate of a factorial design. Then we squeezed it into
blocks, whether it was replicated or not. Now we are going to construct even more sparse designs.
There will be a large number of factors, k, but the total number of observations will be N = 2k-p, so we
keep the total number of observations relatively small as k gets large.

The goal is to create designs that allow us to screen a large number of factors but without having a
very large experiment. In the context where we are screening a large number of factors, we are
operating under the assumption that only a few are very important. This is called sparsity of effects.
We want an efficient way to screen the large number of factors knowing in advance that there will
likely be only two or three factors that will be the most important ones. Hopefully we can detect those
factors even with a relatively small experiment.

We started this chapter by looking at the 23-1 fractional factorial design. This only has four
observations. This is totally unrealistic but served its purpose in illustrating how this design works.
ABC was the generator, which is equal to the Identity, (I = ABC or I = -ABC). This defines the
generator of the design and from this we can determine which effects are confounded or aliased with
which other effects

Let's use the concept of the generator and construct a design for the 24-1 fractional factorial. This
gives us
Loading a one half fraction of the 24 design. Again, we want to pick a high order interaction. Let's
[MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 4 of 28

select ABCD as the generator (I = ABCD) and by hand we can construct the design. I = ABCD
implies that D = ABC. First of all, 24-1 = 23 = 8. So, we will have eight observations in our design.
Here is a basic 23 design in standard Yates notation defined by the levels of A, B, and C:

trt A B C D=ABC
(1) - - - -
a + - - +
b - + - +
ab + + - -
c - - + +
ac + - + -
bc - + + -
abc + + + +

We can then construct the levels of D by using the relationship where D = ABC. Therefore, in the first
row where all the treatments are minus, D = -1*-1*-1 = -1. In the second row, +1, and so forth. As
before we write - and + as a shorthand for -1 and +1.

This is a one half fraction of the 24 design. A full 24 design would have 16 factors.

This 24-1 design is a Resolution IV design. The resolution of the design is based on the number of
the letters in the generator. If the generator is a four letter word, the design is Resolution IV. The
number of letters in the generator determines the confounding or aliasing properties in the resulting
design.

We can see this best by looking at the expression I = ABCD. We obtain the alias structure by
multiplying A × I = A × ABCD = A2BCD which implies A = BCD. If we look at the aliasing that occurs
we would see that A is aliased with BCD, and similarly all of the main effects are aliased with a three-
way interaction:

B = ACD
C = ABD
D = ABC

Main effects are aliased with three-way interactions. Using the same process, we see that two-way
interactions are aliased with other two-way interactions:

AB = CD
AC = BD
AD = BC

In total, we have seven effects, the number of degrees of freedom in this design. The only effects
that are estimable from this design are the four main effects assuming the 3-way interactions are
zero and the three 2-way interactions that are confounded with other 2-way interactions. All 16
effects are accounted for with these seven contrasts plus the overall mean.

Let's take a look at how this type of design is generated in Minitab...

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 5 of 28

Resolution IV Designs

What you need to know about Resolution IV designs:

• the main effects are aliased with the 3-way interactions. This is just the result of the fact that
this is a four letter effect that we are using as the generator.
• the 2-way interactions are aliased with each other. Therefore, we can not determine from this
type of design which of the 2-way interactions are important because they are confounded or
aliased with each other.

Resolution IV designs are preferred over Resolution III designs. Resolution III designs do not have
as good properties because main effects are aliased with two-way interactions. Again, we work from
the assumption that the higher order interactions are not as important. We want to keep our main
effects clear of other important effects.

The 5 Factor Design

Here we let k = 5 and p = 1, again, so that we have a one half fraction of a 25 design. Now we have
five factors, A, B, C, D and E, each at two levels. What would we use as our generator? Since we
are only picking one generator, we should choose the highest order interaction as possible. So we
will choose I = ABCDE, the five-way interaction.

Let's use Minitab to set this up. Minitab gives us a choice of a one half or one fourth fraction.We will
select the one half fraction. It says it is a Resolution V design because it has a five letter generator I
= ABCDE or (E = ABCD).
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 6 of 28

We get a 25-1, so there are 16 observations. Taking a look at the design:

E = ABCD gives us the basis for the resolution of the design.

Let's look at the properties of a Resolution V design. We can see that:

• the main effects are 'clear' of 2-way and 3-way interactions.

• the main effects are only confounded with 4-way interactions or higher, so this gives us really
good information, and
• the 2-way interactions are 'clear' of each other but are aliased with 3-way interactions.

The Resolution V designs are everybody's favorite because you can estimate main effects and two-
way interactions if you are willing to assume that three-way interactions and higher are not important.
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 7 of 28

You can go higher, with Resolution VI, VII etc. designs, however, Resolution III is more or less the
minimum, and Resolution IV and V are increasing in good properties in terms of being able to
estimate the effects.

A One-Fourth Fractional Design, or a 1/22 Fraction of a 2k design

Let's try to construct a 1/4 fractional design using the previous example where k = 4 factors. In this
case p = 2, therefore we will have to pick 2 generators in order to construct this type of design.

As in the previous example k = 4, but now p = 2, so this would give us 24 -2 = 4 observations. A

problem that we can foresee here is that we only have a total of 3 degrees of freedom to estimate.
But we have four main effects, so a main effect is going to have to be confounded or aliased with
another main effect. Hence, this design is not even a Resolution III. Let's go ahead anyway.

Let's pick ABCD, as we did before, as one generator and ABC as the other. So we would have
ABCD × ABC = D as our third generator.

This is not good ... now we have a main effect as a generator which means the main effect would be
confounded with the mean .... we can do better than that.

Let's pick ABCD and then AB as a second generator, this would give us ABCD × AB = CD as our
third generator. We pick two but we must also include a generalized interaction.

Now the smallest word in our generator set is a two letter word - so this means that this is a
Resolution II design. But we found out that a Resolution II designs tell us that the main effects are
aliased with each other, ... hence not a good design if we want to learn which main effects are
important.

Let's try another example...

Let's say we have k = 5 and p = 2. We have five factors, so again we need to pick two generators.
We want to pick the generators so that the generators and their interactions are each as large a word
as possible. This is very similar to what we were doing when we were confounding in blocks.

Let's pick the 4-way interaction ABCD, and CDE. Then the generalized interaction is ABCD × CDE =
ABE. In this case, in the way we picked them the smallest number of letters is 3 so this is a
Resolution III design.

We can construct this design in the same way we had previously. We begin with 25-2 = 23 = 8
observations which are constructed from all combinations of A, B, and C, then we'll use our
generators to define D and E. Note that I = ABCD tells us that D = ABC, and the other generator I =
CDE tells us that E = CD. Now we can define the new columns D = ABC and E = CD. Although D
and E weren't a part of the original design, we were able to construct them from the two generators
as shown below:

trt A B C D=ABC E=CD

(1) - - - - +
a + - - + -
b - + - + -
ab + + - - +
c
Loading [MathJax]/extensions/MathMenu.js
- - + + +

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 8 of 28

ac + - + - -
bc - + + - -
abc + + + + +

Now we have a design with eight observations, 23, with five factors. Our generator set is: I = ABCD =
CDE = ABE. This is a Resolution III design because the smallest word in the generator set has only
three letters. Let's look at this in Minitab ...

A Resolution IV Design Example

Let's take k = 6 and p = 2, now we again have to choose two generators with the highest order
possible, such that the generalized interaction is also as high as possible. We have factors A, B, C,
D, E and F to choose from. What should we choose as generators?

Let's try ABCD and CDEF. The generalized interaction of these two = ABEF. We have strategically
chosen two four letter generators whose generalized interaction is also four letters. This is the best
that we can do. This results in a 26-2 design, which is sometimes written like this, 26-2IV, because it is
a Resolution IV design.

In Minitab we can see the available designs for six factors in the table below:

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 9 of 28

... with six factors, a 26-2 = 24 design, which has 16 observations, is located in the six factor column,
the 16 observation row. This tells us that this design is a Resolution IV, (in yellow). We know from
this table that this type of design exists, so in Minitab we can specify this design.

... which results in the following output.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 10 of 28

In Minitab by default ABCE and BCDF were chosen as the design generators. The design was
constructed by starting with the full factorial of factors A, B, C, and D. Minitab then generated E by
using the first three columns, A, B and C. Then it could choose F = BCD.

Because the generator set, I = ABCE = ADEF = BCDF, contains only four letter words, this is
classified as a Resolution IV design. All the main effects are confounded with 3-way interactions and
a 5-way interaction. The 2-way interactions are aliased with each other. Again, this describes the
property of the Resolution IV design.

8.2 - Analyzing a Fractional Factorial Design

We discussed designing experiments, but now let's discuss how we would analyze these
experiments. We take an example we saw before. The response Y is filtration rate in a chemical pilot
plant and the four factors are: A = temperature, B = pressure, C = concentration and D = stirring rate.
(Example 2 from Chapter 6, Ex6-2.MTW [1])

This experimental design has 16 observations, a 24 with one complete replicate. This is the example
we looked at with one observation per cell when we introduced a normal scores plot.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 11 of 28

Our final model ended up with three factors, A, C and D, and two of their interactions, AC and AD.
This was based on one complete replicate of this design. What might we have learned if we had
done an experiment half this size, N = 8? If we look at the fractional factorial - one half of this design
- where we have D = ABC or I = ABCD as the generator - this creates a design with 8 observations.

The alias structure is a four letter word, therefore this is a Resolution IV design, A, B, C and D are
each aliased with a 3-way interaction, (so we can't estimate them any longer), and the two way
interactions are aliased with each other.

If we look at the analysis of this 1/2 fractional factorial design and we put all of the terms in the
model, (of course some of these are aliased with each other), and we will look at the normal scores
plot. What do we get? (The data are in Ex6_2Half.MTW)

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 12 of 28

We only get seven effects plotted, since there were eight observations. The overall mean does not
show up here. These points are labeled but because there are only seven of them there is no
estimate of error. Let's look at another plot that we haven't used that much yet - the Pareto plot. This
type of plot looks at the effects and orders them from largest to smallest showing you the relative
sizes of the effects. Although we do not know what is significant and what is not significant, this still
might be a helpful plot to look at to better understand the data.

This Pareto plot shows us that the three main effects A, C, and D that were most significant in the full
design are still important as well as the two interactions, AD and AC. However, B and AB are clearly
not as large. (You can do this using the Stat >> DOE >> Factorial >> Analyze and click on graph.)

What can we learn from this? Let's try to fit a reduced model from the information that we gleaned
from this first step. We will include all the main effects and the AC and AD interactions.

In the analysis, we have four main effects ...

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 13 of 28

... overall they are almost significant, (.052), and the overall two-way interactions, (.038) but we only
have one degree of freedom of error - so this makes this a very low-power test. However, this is the
price that you would pay with a fractional factorial. If we look above at the individual effects, B as we
saw on the plot appears to be not important, we have further evidence that we should drop this from
the analysis.

Back to Minitab and let's drop the B term because it doesn't show up as a significant main effect nor
as part of any of the interactions.

Now the overall main effects and 2-way interactions are significant. Residual error still only has 2
degrees of freedom, but this gives us an estimate at least and we can also look at the individual
effects.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 14 of 28

So, fractional factorials are useful when you hope or expect that not all of the factors are going to be
significant. You are screening for factors to drop out of the study. In this example, we started with a
24 - 1 design but when we dropped B we ended up with a 23 design with 1 observation per cell.

This is a typical scenario, you begin by screening a large number of factors and end up with a
smaller set. We still don't know much about the factors and this is still a pretty thin or weak design
but it gives you the information that you need to take the next step. You can now do a more complete
experiment on fewer factors.

8.3 - Foldover Designs

Foldover designs are useful when you are involved in sequentially testing a set of factors. You begin
with very small experiments and proceed in stages. We consider this type of design through two
examples.

1/8th fractional factorial of a 26 design

First, we will look at an example with 6 factors and we select a 26-3 design, or a 1/8th fractional
factorial of a 26 design.

In order to select a 1/8 fraction of the full factorial, we will need to choose 3 generators and make
sure that the generalized interactions among these three generators are of sufficient size to achieve
the higher resolution. In this case it will be a Resolution III as Minitab shows us above.

Let's remind ourselves how we do this. We can choose I = ABD = ACE = BCF as the generators.

Since N = 26-3 = 23 observations, we start with a basic 23 design which would be set up using the
following framework. First write down the complete factorial for factors A, B, and C. From that we can
generate additional factors based on the available interactions, i.e. we will make D = AB, E = AC,
and F = BC. Complete the table below ...

trt A B C D=AB E=AC F=BC ABC

(1) - - -
a + - -
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 15 of 28

b - + -
ab + + -
c - - +
ac + - +
bc - + +
abc + + +

Our generators are ABD, ACE and BCF. So, our alias structure is created by this equivalence:

I = ABD = ACE = BCF

If these are our generators then all of the generalized interactions among these terms are also part
of the generator set. Let's take a look at them:

The 2-way interactions:

• ABD × ACE = BCDE,

• ABD × BCF = ACDF,
• ACE × BCF = ABEF,

and the 3-way interaction:

• ABD × ACE × BCF = DEF

We still have a Resolution III design because the generator set is composed of words, the smallest
of which has 3 letters. So you could fill in the framework above for these factors just by multiplying
from the basic design, the pluses and minuses.

Minitab does this for you. And the worksheet will look like this:

We can estimate all of the main effects and one of the aliased two-way interactions. What this also
suggests is that there is one more factor that we could include in a design of this size, N = 8.

1/16th fraction of a 27 design

Now we consider a 1/16th fraction of a 27 design, or a 27-4 design. Again, we will have only N = 23 =
8 observations but now we have seven factors. Thus k = 7 and p = 4.

Let's look at this in Minitab - for seven factors here are the design options ...

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 16 of 28

The 1/16 fraction design is a Resolution III design and it is the smallest possible one. Here is what
the design looks like:

The generators are listed on top. The same first three are as before and then G = ABC, the only one
left. The alias structure gets quite convoluted. The reason being that if we were taking a complete
replicate of this design, 27, we could put it into 16 blocks. In this case,we are only looking at one of
the 16 blocks in the complete design. In these 16 blocks there are 15 degrees of freedom among
these blocks. So, you see I + the 15 effects.

Sometimes people are not interested in seeing all of these higher order interactions, after all five way
interactions are not all that interesting. You can clean up this output a bit by using this option found in
the 'Results...' dialog box in Minitab:

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 17 of 28

Notice now that the only thing you find in the table are main effects. No 2-way interactions are
available. This is a unique design called a Saturated Design. This is the smallest possible design
that you could use for 7 factors. Another way to look at this is that for a design with eight
observations the maximum number of factors you can include in that design is seven. We are using
every degree of freedom to estimate the main effects.

If we moved to the next smallest design where N = 16, then what would the saturated design be? 15
factors. You would have a 215-11, which would give us a 24 basic design. Then we could estimate up
to 15 main effects.

So, you can see with fairly small designs, only 16 observations, we can test for a lot of factors if we
are only interested in main effects, using a Resolution III design. Let's see what the options are in
Minitab.

Notice that the largest design shown has 128 runs which is already a very large experiment for 15
factors. You probably wouldn't want more than that.

Folding a Design

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 18 of 28

We will come back to Saturated designs - but first let's consider the 27-4 design, which is saturated
and is a Resolution III design and let's fold it over.

Let's assume we ran this design, we found some interesting effects but we have no degrees of
freedom for error. So we want to look at another replicate of this design. Rather than repeating this
exact same design we can fold it over.

We can fold it over on all factors, or specify a single factor for folding.

What folding means is to take the design and reverse the sign on all the factors. This would be a fold
on all factors.

Now instead of eight observations we have 16. And if you look at the first eight and compare these
with the second set of eight you will see that the signs have simply been reversed

Look at row 1 and row 9 and you will see that they have the exact opposite signs. Thus you double
the basic design with all factors exchanged. Or, you can think of this somewhat as taking one
replicate and putting it in blocks, we've now taken two of the blocks to create our design.

These designs are used to learn how to proceed from a basic design, where you might have learned
something about one of the factors that looks promising, and you want to give more attention to that
factor. This would suggest folding, not on all factors, but folding on that one particular factor. Let's
examine why you might want to do this.

In our first example above we started with a Resolution III design, and by folding it over on all factors,
we have increased the resolution by one number, in this case it goes from resolution III to IV. So,
instead of the main effects being confounded with two-way interactions, which they were before, now
they are all clear of the two-way interactions. We still have the two-way interactions confounded with
each other however.

Now, let's look at the situation where after the first run we were mostly intrigued by factor B.

Now, rather than fold on all factors we want to fold on just factor B.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 19 of 28

Notice now that in the column for B, the folded part is exactly the opposite. None of the other
columns change, just the column for factor B. All of the other columns stayed the same.

Now look at the alias structure for this design...

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 20 of 28

This is still a Resolution III, (we haven't folded on all factors so we don't jump a resolution number).
But look at the B factor which we folded. The main effect, B, is aliased with only the four way
interactions and higher. Also, notice that all of the 2-way interactions with B are clear of other two-
way interactions, so they become estimable. So by only folding on one factor, you get very good
information on that factor and its interactions. However, it is still a resolution three design.

There are two purposes for folding; one is taking on another replicate for the purpose of moving to a
higher resolution number. The other reason would be to isolate the information on a particular factor.
Both of these would be done in the context of doing a sequential experiment, doing an analysis of
that and then doing a second stage experiment. If you do this two stage experiment, performing a
second stage based on the first experiment, you should also use stage as the block factor in the
analysis.

All of these designs, even though they are fractions of an experiment, should be blocked, if they are
done in stages.

One more example ...

Let's go to 8 factors. The minimal design now can not be eight observations but must be 16. This is a
Resolution IV design.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 21 of 28

This design has 4 generators BCDE, ACDF, ABCG and ABDH. It is a Resolution IV design and it is a
design with16 observations. OK, now we are going to assume that we can only run these
experiments eight at a time so we have to block. We will use two blocks, and we will still have the
same fractional design, eight factors in 16 runs but now we want to have two blocks.

We let Minitab pick the blocks:

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 22 of 28

In this design, we have eight factors, 16 runs, and the same generators but now we need an
additional generator, the block generator. Minitab is using AB as the block generator. Notice in the
alias structure that the blocks are confounded with the AB term.

Notice also that the AB term does not show up as an estimable effect below. It would have been an
effect we could have estimated but it is now confounded with blocks. So, one additional degree of
freedom is used in this confounding with blocks.

The only choice the program had was to select one of these effects that were previously estimable
and confound them with blocks. The program picked one of those 2-way interactions and this means
blocks are now confounded with a 2-way interaction.

We can still block these fractional designs and it is useful to do this if you can only perform a certain
number at a time. However, if you are doing sequential experimentation you should block just
because you are doing it in stages.

In summary, when you fold over a Resolution III design on all factors, then you get a Resolution IV
design. Look at the table of all possible designs in Minitab below:

If you fold any of the red Resolution III designs you go to the next level, it has twice as many
observations and becomes a Resolution IV design. If you fold many of the Resolution IV designs,
even though you double the number of observations by folding, you are still at the Resolution IV.

Whereas, if you fold a Resolution III or IV design on one factor, you get better information on that
factor and all its 2-way interactions would be clear of other 2-way interactions. Therefore, it serves
that purpose well for Resolution III or IV designs.

8.4 - Plackett-Burman Designs

We looked at 2k-p designs, which give us designs that have 8, 16, 32, 64, 128, etc. number of runs.
We noted that all of these numbers are some fraction of 1/2p of a 2k design.
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 23 of 28

However, when you look at these numbers there is a pretty big gap between 16 and 32, 32 to 64,
etc. We sometimes need other alternative designs besides these with a different number of
observations.

A class of designs that allows us to create experiments with some number between these fractional
factorial designs are the Plackett-Burman designs. Plackett-Burman designs exist for

N = 12, [16], 20, 24, 28, [32], 36, 40, 44, 48, ...

... any number which is divisible by four. These designs are similar to Resolution III designs,
meaning you can estimate main effects clear of other main effects. Main effects are clear of each
other but they are confounded with other higher interactions.

Look at the table of available designs in Minitab. The Plackett-Burman designs are listed below:

So, if you have 2 to 7 factors you can create a Plackett-Burman design with 12, 20, 24, ... up to 48
observations. Of course, if you have 7 factors with eight runs then you have a saturated design.

In the textbook there is a brief shortcut way of creating these designs, but in Minitab we simply select
the Plackett-Burman option.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 24 of 28

You specify how many runs and how many factors are in your experiment. If we specified eight
factors and 12 runs, we get a design that looks like this:

This look very much like the designs we had before. In this case we have eight factors, A through H,
each with two levels. And each factor is defined by a 12 run design, 6 pluses and 6 minuses. Again,
these are contrasts. Half of the observations at a high level and half at the low-level, and if you take
any two columns they are orthogonal to each other. So, these are an orthogonal set of columns just
as we had for the 2k-p design. If you take the product of any two of these and add them up, the sum
of the products you get is zero.

Because these are orthogonal contrasts we get clean information on all main effects. The main
effects are not confounded as required by the orthogonality of those columns.

Here is a quick way to manually create this type of design. First of all, one would fill out the first
column of the design table, this would be column A. Then you can create the B column by taking the
last element for permuting and then slide everything down. This process can be repeated for each
column of factors needed in the design. Click the 'Create Design' button below to see how this
works:

You can generate these designs by just knowing the first 11 elements, permuting these into the next
column and adding an additional row of minuses across the bottom. It has this cyclical pattern and it
works for most of these types of designs, (12, 20, 24, 36, but not for 28!). Here is what it looks like for
20 runs with 16 factors:

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 25 of 28

The cyclical pattern is a result of number theory properties that generate these orthogonal arrays.
There is a lot of mathematical research behind these designs to achieve a matrix with orthogonal
columns which is what we need.

We point out that these designs are a little different than the 2k-p designs. When you have a 2k-p
design you have an alias structure that confounds some factors with other factors. Let's look at two
examples to illustrate this.

Example: FF2LevelCorr.MPJ [2]

The first is a fractional factorial, 4 factor design, Resolution IV with one generator ABCD or D = ABC.
From this design we get an alias structure that we are familiar with. Main effects are aliased with
3-way interactions which means that they are completely confounded with those factors. Two-way
interactions are confounded with each other.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 26 of 28

Let's look at the correlation among these factors, A, B, C and D, and then a couple of interaction
columns.

This is just a simple Pearson correlation. What Minitab gives us is the coefficient and the p-value.
We can ignore the p-values because we are not really interested in testing, however a correlation
between A and B = 0, A and C = 0, A and D = 0, etc. The correlation between all these factors = 0
because of the orthogonality.

Look back up at the alias structure and you will see the D is confounded with ABC. As we look back
at the correlation table the correlation between D and ABC = 1. The correlation between two factors
that are confounded = 1. This is appropriate because they are completely correlated with each other.
Therefore, in these 2k-p designs we can see through correlation that factors are either orthogonal
(correlation = 0) or they are completely confounded (correlation = 1).

Next, let's look at the Plackett-Burman designs and see how this differs. Below, we have created a
design for 9 factors, 12 runs and we are looking at the correlation among the main effects, A, B, C,
D, and E.

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 27 of 28

These are the main factors themselves already set in orthogonal columns so these correlations = 0.
If we look at the next design, however, in this case we have the 12 runs and then we have created
new 2-way interactions through multiplication of the factors already determined. Again, ignoring the
p-values we produced a correlation matrix, (partially displayed below).

A is orthogonal to every other factor, the correlation value = 0. B is not correlated with all the other
main effects, where correlation = 0, but with some of these two-way interactions the correlation =
0.333. This shows partial confounding with the two-way interaction. Likewise, C has partial
confounding with AB and AD. D is partially confounded with AB and AC. F is partially confounded
with AB and AC and AD, ... and so forth.
Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 28 of 28

Plackett-Burman designs have partial confounding, not complete confounding, with the 2-way and
3-way and higher interactions. Although they have this property that some effects are orthogonal
they do not have the same structure allowing complete or orthogonal correlation with the other two
way and higher order interactions.

Like other Resolution II designs, these designs are also good for screening for important factors. But
remember, in a Resolution II design a main effect might look important, because some combination
of interactions is important and the main effect itself might not be the important effect.

If you assume that your interactions = 0 or are not important these are great designs. If your
assumption is wrong and there are interactions, then it could show up as influencing one or the other
main effects. These designs are very efficient with small numbers of observations and useful, but
remember the caveat, you are assuming that the main effects are going to show up as larger effects
than interactions so that they will dominate the interaction effects.

Using Minitab we can ask for up to 47 factors. In doing so you want to select a sufficient number of
runs over the number of factors so that you have a reasonable number of degrees of freedom for
error. At this stage a statistical test really isn't that important, you are just screening for a place to
start.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/48

Links:
[1] https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson08/Ex6-2.MTW
[2]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson08/FF2levelCorr.MPJ

Loading [MathJax]/extensions/MathMenu.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/48/ 4/18/2019
 Page 1 of 22

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 9: 3-level and Mixed-level Factorials and Fractional Factorials

Lesson 9: 3-level and Mixed-level

Factorials and Fractional Factorials
Learning Objectives and Outcomes

By the end of this lesson, students are assumed to know

• Application of 3k factorial designs, the interaction components and relative degrees of

freedom
• How to perform blocking of 3k designs in 3p number of blocks and how to choose the
effect(s) which should be confounded with blocks
• Concept of “Partial Confounding” in replicated blocked designs and its advantages
• How to generate reasonable 3k-p fractional factorial designs and understand the alias
structure
• The fact that Latin square and Graeco-Latin square designs are special cases of 3k
fractional factorial design
• Mixed level factorial designs and their applications

Introduction

Basic material

k
These designs are a generalization of the 2 designs. We will continue to talk about coded
variables so we can describe designs in general terms, but in this case we will be assuming
k k
in the 3 designs that the factors are all quantitative. With 2 designs we weren't as strict
k
about this because we could have either qualitative or quantitative factors. Most 3 designs
k
are only useful where the factors are quantitative. With 3 designs we are moving from
screening factors to analyzing them to understand what their actual response function looks
like.

With 2 level designs, we had just two levels of each factor. This is fine for fitting a linear,
straight line relationship. With three level of each factor we now have points at the middle so
we will are able to fit curved response functions, i.e. quadratic response functions. In two
dimensions with a square design space, using a 2k design we simply had corner points,
which defined a square that looked like this:
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 2 of 22

In three dimensions the design region becomes a cube and with four or more factors it is a
hypercube which we can't draw.

We can label the design points, similar to what we did before – see the columns on the left.
However for these design we prefer the other way of coding, using {0,1,2}which is a
generalization of the {0,1} coding that we used in the 2k designs. This is shown in the
columns on the right in the table below:

A B A B
- - 0 0
0 - 1 0
+ - 2 0
- 0 0 1
0 0 1 1
+ 0 2 1
- + 0 2
0 + 1 2
+ + 2 2

For either method of coding, the treatment combinations represent the actual values of X1
and X2, where there is some high level, a middle level and some low level of each factor.
Visually our region of experimentation or region of interest is highlighted in the figure below
when k = 2:

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 3 of 22

If we look at the analysis of variance for a k = 2 experiment with n replicates, where we have
three levels of both factors we would have the following:

AOV df
A 2
B 2
AxB 4
Error 9(n-1)
Total 9n-1

Important idea used for confounding and taking fractions

How we consider three level designs will parallel what we did in two level designs, therefore
we may confound the experiment in incomplete blocks or simply utilize a fraction of the
design. In two-level designs, the interactions each have 1 d.f. and consist only of +/-
components, so it is simple to see how to do the confounding. Things are more complicated
in 3 level designs, since a p-way interaction has 2^p d.f. If we want to confound a main
effect (2 d.f.) with a 2-way interaction (4 d.f.) we need to partition the interaction into 2
orthogonal pieces with 2 d.f. each. Then we will confound the main effect with one of the 2
pieces. There will be 2 choices. Similarly, if we want to confound a main effect with a 3-way
interaction, we need to break the interaction into 4 pieces with 2 d.f. each. Each piece of the
interaction is represented by a psuedo-factor with 3 levels. The method given using the
Latin squares is quite simple . There is some clever modulus arithmetic in this section, but
the details are not important. The important idea is that just as with the 2k designs, we can
purposefully confound to achieve designs that are efficient either because they do not use
the entire set of 3k runs or because they can be run in blocks which do not disturb our ability
to estimate the effects of most interest.

Following the text, for the A*B interaction, we define the pseudo factors, which are called the
AB component and the AB2 component. These components could be called pseudo-
interaction effects. The two components will be defined as a linear combination as follows,
where X1 is the level of factor A and X2 is the level of factor B using the {0,1,2} coding
system.
Loading Let the AB component be defined as
[MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 4 of 22

LAB = X1 + X2 (mod3)

and the AB2 component will be defined as:

LAB2 = X1 + 2X2 (mod3)

Using these definitions we can create the pseudo-interaction components. Below you see
that the AB levels are defined by LAB and the AB2 levels are defined by LAB2.

A B AB AB2
0 0 0 0
1 0 1 1
2 0 2 2
0 1 1 2
1 1 2 0
2 1 0 1
0 2 2 1
1 2 0 2
2 2 1 0

This table has entries {0, 1, 2} which allow us to confound a main effect or either component
of the interaction A*B. Each of these main effects or pseudo interaction components have
three levels and therefore 2 degrees of freedom.

This section will also discuss partitioning the interaction SS's into 1 d.f. sums of squares
associated with a polynomial, however this is just polynomial regression. This method does
not seem to be readily applicable to creating interpretable confounding patterns.

9.1 - 3^k Designs in 3^p Blocks

Let's begin by taking the 3k designs and we will describe partitioning where you take one
replicate of the design and put it into blocks. We will then take that structure and look at 3k-p
factorials. These designs are not used for screening as the 2k designs were; rather with
three levels we begin to think about response surface models. Also, 3k designs become
very large as k gets large. With just four factors a complete factorial is already 81
observations, i.e. N = 34. In general we won't consider these designs for very large k, but we
will point out some very interesting connections that these designs reveal.

Reiterating what was said in the introduction, consider the two factor design 32 with factors
A and B, each at 3 levels. We denote the levels 0, 1, and 2. The A×B interaction, with 4
degrees of freedom, can be split into two orthogonal components. One way to define the
components is that AB component will be defined as a linear combination as follows:
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 5 of 22

LAB = X1 + X2 (mod3)

and the AB2 component will be defined as:

LAB2 = X1 + 2X2 (mod3)

A B AB AB2
0 0 0 0
1 0 1 1
2 0 2 2
0 1 1 2
1 1 2 0
2 1 0 1
0 2 2 1
1 2 0 2
2 2 1 0

In the table above for the AB and the AB2 components we have 3 0's, 3 1's and 3 2's, so this
modular arithmetic gives us a balanced set of treatments for each component. Note that we
could also find the A2B and A2B2 components but when you do the computation you
discover that AB2=A2B and AB=A2B2.

We will use this to construct the design as shown below.

We will take one replicate of this design and partition it into 3 blocks. Before we do, let’s
consider the analysis of variance table for this single replicate of the design.

AOV df
A 3-1=2
B 3-1=2
A×B 2*2=4
AB 3-1=2
AB2 3-1=2

We have partitioned the A×B interaction into AB and AB2, the two components of the
interaction, each with 2 degrees of freedom. So, by using modular arithmetic, we have
partitioned the 4 degrees of freedom into two sets, and these are orthogonal to each other.
If you create two dummy variables for each of these factors, A, B, AB and AB2 you would
see that each of these sets of dummy variables are orthogonal to the other.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 6 of 22

These pseudo components can also be manipulated using a symbolic notation. This is
included here for completeness, but it is not something you need to know to use or
understand confounding. Consider the interaction between AB and AB2. Thus AB × AB2
which gives us A2B3 which using modular (3) arithmetic gives us A2B0 = A2 = (A2)2 = A.
Therefore, the interaction between these two terms gives us the main effect. If we wanted
to look at a term such as A2B or A2B2, we would reduce it by squaring it which would give
us: (A2B)2 = AB2 and likewise (A2B2)2 = AB. We never include a component that has an
exponent on the first letter because by squaring it we obtain an equivalent component. This
is just a way of partitioning the treatment combinations and these labels are just an arbitrary
identification of them.

Let's now look at the one replicate where we will confound the levels of the AB component
with our blocks. We will label these 0, 1, and 2 and we will put our treatment pairs in blocks
from the following table.

A B AB AB2
0 0 0 0
1 0 1 1
2 0 2 2
0 1 1 2
1 1 2 0
2 1 0 1
0 2 2 1
1 2 0 2
2 2 1 0

Now we assign the treatment combinations to the blocks, where the pairs represent the
levels of factors A and B.

LAB
0 1 2
0, 0 1, 0 2, 0
2, 1 0, 1 1, 1
1, 2 2, 2 0, 2

This is how we get these three blocks confounded with the levels of the LAB component of
interaction.

Now, let's assume that we have four reps of this experiment - all the same - with AB
confounding with blocks using the LAB. (each replicate is assigned to 3 blocks with AB
confounded with blocks). We have defined one rep by confounding the AB component, and
then we
Loading will do the same with 3 more reps.
[MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 7 of 22

Let's take a look at the AOV resulting from this experiment:

AOV df
Rep 4-1=3
Blk = AB 3-1=2
Rep × AB 3*2=6
Inter-block Total 11
A 3-1=2
B 3-1=2
A × B = AB2 3-1=2
(2+2+2)*
Error
(4-1)=18
Total 3*3*4-1=35

Note that Rep as an overall block has 3 df. Within reps we have variation among the 3
blocks, which are the AB levels - this has 2 df. Then we have Rep by blk or Rep by AB
which has 6 df. This is the inter-block part of the analysis. These 11 degrees of freedom
represents the variation among the 12 blocks (3*4).

Next we consider the intra-block part: A with 2 df, B with 2 df and the A × B or AB2
component that also has 2 df. Finally we have error, which we can get by subtraction, (36
observations = 35 total df, 35 - 17 = 18 df). Another way to think about the Error is the
interaction between the treatments and reps which is 6 × 3 = 18, which is the same logic as
in a randomized block design, where the SSE is (a-1)(b-1). A possible confusion here is
using the terminology of blocks at two levels, the reps are at an overall level, and then within
each rep we have the smaller blocks which are confounded with the AB component.

We now examine another experiment, this time confounding the AB2 factor. We can
construct another design using this component as our generator to confound with blocks.

A B AB AB2
0 0 0 0
1 0 1 1
2 0 2 2
0 1 1 2
1 1 2 0
2 1 0 1
0 2 2 1
1 2
Loading [MathJax]/extensions/MathZoom.js 0 2

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 8 of 22

2 2 1 0

Using the AB2 then gives us the following treatment pairs (A,B) assigned to 3 blocks:

LAB2
0 1 2
0, 0 1, 0 2, 0
1, 1 2, 1 0, 1
2, 2 0, 2 1, 2

This partitions all nine of the treatment combinations into the three blocks.

Partial Confounding (optional section)

We now consider a combination of these experiments, in which we have 2 reps confounding

AB and 2 reps confounding AB2. We again will have 4 reps but our AOV will look a little
different:

AOV df
Rep 4-1=3
Blk = AB 3-1=2
Blk = AB2 3-1=2
Rep × AB (2-1)*2=2
Inter-block Error 4
Rep × AB2 (2-1)*2=2
Inter-block Total 11
A 3-1=2
B 3-1=2
A×B 2*2=4
AB 3-1=2
4
AB2 3-1=2
2*(4-1)+2*
(4-1)+2*
Error
(2-1)+2*
(2-1)=16
Total 3*3*4-1=35

There are only two reps with AB confounded, so Rep × AB = (2-1) * (3-1) = 2 df . The same
is true for the AB2 component. This gives us the same 11 df among the 12 blocks. In the
intra-block
Loading section, we can estimate A and B, so they will have 2 df. A × B will have 4 df
[MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 9 of 22

now, and if we look at what this is in terms of the AB and the AB2 component each accounts
for 2 df. Then we have Error with 16 df and the total stays the same. The 16 df comes from
the unconfounded effects - (A: 2 x 3 = 6 and B: 2 x 3 = 6) - that's 12 of these df, plus each of
the AB and the AB2 components which are confounded in two reps, and unconfounded in
the other two reps - (2 * (2-1) = 2 for AB and 2 * (2-1) = 2 for AB2) - which accounts for the
remaining 4 of the total 16 df for error.

We could determine the Error df simply by subtracting from the Total df, but, if it is helpful to
think about randomized block designs where you have blocks and treatments and the error
is the interaction between them. Note that here we use the term replicates instead of blocks,
so actually we consider replicates as sort of super-blocks. In this case error would be the
interaction between replicates and unconfounded treatments. This RCBD framework is a
foundational structure that we use again and again in experimental design.

This is a good example of the benefit of partial confounding because the interaction of the
pseudo factors are confounded in only half of the design, so we can estimate the interaction
A*B from the other half. You get overall exactly half the information on the interaction from
this partially confounded design.

Confounding a main effect (an important idea)

Now let’s think further outside of the box. What if we confound the main effect A? What
would this do to our design? What kind of experimental design would this be?

Now we define or construct our blocks by using levels of A from the table above. A single
replicate of the design would look like this.

A
0 1 2
0, 0 1, 0 2, 0
0, 1 1, 1 2, 1
0, 2 1, 2 2, 2

Then we could replicate this design four times. Let's consider an agricultural application and
say that A = irrigation method, B = crop variety, and the Blocks = whole plots of land to
which we apply the irrigation type. By confounding a main effect we're going to get a split-
plot design in which the analysis will look like this:

AOV df
Reps 3
A 2
Rep × A 6
Inter-block Total 11
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 10 of 22

B 2
A×B 4
Error 18
Total 35

In this design there are four reps (3 df), and the blocks within reps are actually the levels of
A which has 2 df, Rep × A has 6 df. The interblock part of the analysis here is just a
randomized complete block analysis of four reps, three treatments and their interactions.
The intra-block part contains B which has 2 df, and the A x B interaction which has 4 df.
Therefore this is another way to understand a split plot design, where you confound one of
the main effects .

More examples of confounding.

Let's look at the k = 3 case - an increase in the number of treatments by one. Here we will
look at a 33 design confounded in 31 blocks, or we could look at a 33 design confounded in
32 blocks. In a 33 design confounded in three blocks, each block would have nine
observations now instead of three.

To create the design shown in Figure 9-7 below, follow the following
commands:

Stat > DOE > Factorial > Create Factorial Design

• click on General full factorial design,

• set Number of factors to 3
• set Number of levels of each factor to 3
• under options, unclick the randomize button
• Then use Calc menu and subtract 1 from each of column A, B,
and C (We could have initially made levels 0, 1 and 2).

Now the levels of the three factors are coded with (0, 1, 2). We are
ready to calculate the pseudo factor, AB2C2, which we will abbreviate as
AB2C2.

Label the next blank column, AB2C2. Again, using the Calc menu, let
AB2C2 = Mod(A + 2*B + 2*C, 3), which creates the levels of the pseudo
factor LAB2C2 described on the page 371.

Here is a link to a Minitab project file that implements this: Figure-

9-7.MPJ [1]

Let's look at the k = 3 case - a 33 design confounded in 31 blocks. In a 33 design confounded

in three blocks, each block would have nine observations now.

Loading [MathJax]/extensions/MathZoom.js
A B C

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 11 of 22

0 0 0
1 0 0
2 0 0
0 1 0
1 1 0
2 1 0
0 2 0
1 2 0
2 2 0
0 0 1
1 0 1
2 0 1
0 1 1
1 1 1
2 1 1
0 2 1
1 2 1
2 2 1
0 0 2
1 0 2
2 0 2
0 1 2
1 1 2
2 1 2
0 2 2
1 2 2
2 2 2

With 27 possible combinations, without even replicating, we have 26 df. These can be
broken down in the following manner:

AOV df
A 2
B 2
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 12 of 22

C 2
A×B 4
A×C 4
B×C 4
A ×B × C 8
Total 26

The main effects all have 2 df, the three two-way interactions all have 4 df, and the three-
way interaction has 8 df. If we think about what we might confound with blocks to construct
a design we typically want to pick a higher order interaction.

The three way interaction A × B × C can be partitioned into four orthogonal components
labeled, ABC, AB2C, ABC2 and AB2C2. These are the only possibilities where the first letter
has exponent = 1. When the first letter has an exponent higher than one, for instance A2BC,
to reduce it we can first square it, A4B2C2, and then using mod 3 arithmetic on the exponent
get AB2C2, i.e. a component we already have in our set. These four components partition
the 8 degrees of freedom and we can define them just as we have before. For instance:

LABC = X1 + X2 + X3 (mod 3)

This column has been filled out in the table below in two steps, the first column carries out
the arithmetic (sum) and the next column applies the mod 3 arithmetic:

A B C A+B+C LABC
0 0 0 0 0
1 0 0 1 1
2 0 0 2 2
0 1 0 1 1
1 1 0 2 2
2 1 0 3 0
0 2 0 2 2
1 2 0 3 0
2 2 0 4 1
0 0 1 1 1
1 0 1 2 2
2 0 1 3 0
0 1 1 2 2
1 1 1 3 0
Loading [MathJax]/extensions/MathZoom.js
2 1 1 4 1

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 13 of 22

0 2 1 3 0
1 2 1 4 1
2 2 1 5 2
0 0 2 2 2
1 0 2 3 0
2 0 2 4 1
0 1 2 3 0
1 1 2 4 1
2 1 2 5 2
0 2 2 4 1
1 2 2 5 2
2 2 2 6 0

Using the LABC component to assign treatments to blocks we could write out the following
treatment combinations for one of the reps:

LABC
0 1 2
0, 0, 0 1, 0, 0 2, 0, 0
2, 1, 0 0, 1, 0 1, 1, 0
1, 2, 0 2, 2, 0 0, 2, 0
2, 0, 1 0, 0, 1 1, 0, 1
1, 1, 1 2, 1, 1 0, 1, 1
0, 2, 1 1, 2, 1 2, 2, 1
1, 0, 2 2, 0, 2 0, 0, 2
0, 1, 2 1, 1, 2 2, 1, 2
2, 2, 2 0, 2, 2 1, 2, 2

This partitions the 27 treatment combinations into three blocks. The ABC component of the
three-way interaction is confounded with blocks.

If we performed one block of this design perhaps because we could not complete 27 runs in
one day - we might be able to accommodate nine runs per day. So perhaps on day one we
use the first column of treatment combinations, on day two we used the second column of
treatment combinations and on day three we use the third column of treatment
combinations. This would conclude one complete replicate of the experiment. We can then
continue a similar approach in the next three days to complete the second replicate. So, in
twelve[MathJax]/extensions/MathZoom.js
Loading days four reps would have been performed.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 14 of 22

How would we analyze this? We would use the same structure.

AOV df
Rep 4-1=3
ABC = Blk 2
Rep × ABC 6
A 2
B 2
C 2
A×B 4
A×C 4
B×C 4
A×B×C 6
AB2C 2

ABC2 2

AB2C2 2
Error 72
Total 108-1=107

We have (4 - 1) or 3 df for Rep, ABC is confounded with blocks so the ABC component of
blocks has 2 df, the Rep by ABC (3*2) has 6 df. In summary to this point we have twelve of
these blocks in our 4 reps so there are 11 df in our inter-block section of the analysis.
Everything else follows below. The main effects have 2 df, the two-way interactions have 4
df, and the A × B × C would have 8 df, but it only has 6 df because the ABC component is
gone, leaving the other three components with 2 df each.

Error will be the unconfounded terms times the number of reps -1, or 24 × (4 - 1) = 72.

Likewise, LAB2C = X1 + 2X2 + X3 (mod 3) can also be defined as another pseudo component
in a similar fashion.

9.2 - 3^k Designs in 3^p Blocks cont'd.

We again start out with a 33 design which has 27 treatment combinations and assign them
to 3 blocks. What we want to do in this lesson, going beyond the 32 design, is to describe
the AOV for this 33 design. Then we also want to look at the connection between
confounding in blocks and 3k-p fractional factorials. This story will be very similar to what we
did in the 2k-p designs previously. There is a direct analogue here that you will see.
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 15 of 22

From the previous section we had the following design, 33 treatments in 3 blocks with the
ABC pseudo factor confounded with blocks, i.e.,

LABC
0 1 2
0, 0, 0 1, 0, 0 2, 0, 0
2, 1, 0 0, 1, 0 1, 1, 0
1, 2, 0 2, 2, 0 0, 2, 0
2, 0, 1 0, 0, 1 1, 0, 1
1, 1, 1 2, 1, 1 0, 1, 1
0, 2, 1 1, 2, 1 2, 2, 1
1, 0, 2 2, 0, 2 0, 0, 2
0, 1, 2 1, 1, 2 2, 1, 2
2, 2, 2 0, 2, 2 1, 2, 2

The three (color coded) blocks are determined by the levels of the ABC component of the
three-way interaction which is confounded with blocks. If we only had one replicate of this
design we would have 26 degrees of freedom. So, let's pretend that this design is Rep 1 and
we will add Reps 2, 3, 4, just as we did with the two factor case. This would result in a total
of 12 blocks.

If we did this as our basic design and replicate it three more times our AOV would look like
the following:

AOV df
Reps 3
Blocks(Rep) 4 × (3-1) = 8
ABC 2
Rep × ABC 6
A 2
B 2
C 2
A×B 4
A×C 4
B×C 4
A×B×C 6
Error 72
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 16 of 22

Total 107

We would have Reps with 3 df, blocks nested in Reps with 2 df × 4 Reps = 8 df, then we
would have all of the unconfounded effects as shown above. The A × B × C would only have
6 df because one component (ABC) is confounded with blocks. Error is 24 × 3 = 72 df and
our total is (33 × 4) - 1 = 107 df .

Now, we have written Blocks(Rep) with 8 df equivalently (in the blue font above) as ABC
with 2 df, and Rep × ABC with 6 df, but now we are considering the 0, 1, and 2 as levels of
the ABC factor. In this case ABC is one component of the interaction and still has meaning
in terms of the levels of ABC, just not very interesting since it is part of the three way
interaction. Had we confounded a main effect with blocks, we certainly would have wanted
to analyze it, as seen above where a main effect was confounded with blocks. Then it had
an important meaning and you certainly would want to pull this out and be able to test it.

Now we have a total of 3 × 4 = 12 blocks and the 11 df among them are the interblock part
of the analysis. If we averaged the nine observations in each block and got a single number,
we could analyze those 12 numbers and this would be the inter-block part of this analysis.

How do we accomplished this in Minitab? If you have a set of data labeled by rep, blocks,
and A, B, and C, then you would have everything you need and you can fit a general linear
model:

Y = Rep Blocks(Rep) A | B | C

This would generate the analysis since A | B | C expands to all main effects and all
interactions in GLM of Minitab.

An Alternate Design - Partial Confounding

In thinking about how this design should be implemented a good idea would be to followed
this first Rep with a second Rep that confounds LAB2C, confound LABC2 in Rep three, and
finally confound LAB2C2 in fourth Rep. Now we could estimate all four components of the
three-way interactions because in three of the Reps they would be unconfounded. There is
no information available in the way we had approached it previously. There is lots of
information available using this partial confounding strategy of the three-way interactions.

3k-p designs - Fractional Factorial 3-level Designs

The whole point of looking at this structure is because sometimes we want to only conduct a
fractional factorial. We sometimes can't afford 27 runs, certainly not 108 runs. Often we can
only afford a fraction of the design. So, let's construct a 33-1 design which is a 1/3 fraction of
a 33 design. In this case, N = 33-1= 32 = 9, the total number of runs. This is a small, compact
design. For the case where we use the LABC pseudo factor to create the design, we would
use just one block of the design above, and below here is the alias structure:

I = ABC
A = A × ABC = (A2BC) = AB2C2
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 17 of 22

A = A × (ABC)2 = A 3B2C2 = (B2C2)2 = BC

B = B × ABC = AB2C
B = B × (ABC)2 = A2B3C2 = AC
C = C × ABC = ABC2
C = C × (ABC)2 = A 2B2C3 = (A2B2)2 = AB

Here A is confounded with part of the 3-way and part of the 2-way interaction, likewise for B
and for C. This design only has 9 observations. It has A, B and C main effects estimable and
if we look at the AOV we only have nine observations so we can only include the main
effects:

AOV df
A 2
B 2
C 2
Error 2
Total 8

Below is the 33 design where we partitioned the treatment combinations for one Rep of the
experiment using the levels of LABC. It is of interest to notice that a 33-1 fractional factorial
design is also a design we previously discussed. Can you guess what it is?

If we look at the first light blue column, we can call A the row effect, B the column effect and
C the Latin letters, or in this case 0, 1, 2. We would use this procedure to assign the
treatments to the square. This is how we get a 3 × 3 Latin square. So, a one third fraction of
a 33 design is the same as a 3×3 Latin square design that we saw earlier in this course.
Click on the 'Start' button above to see how this works.

It is important to see the connection here. We have three factors, A, B, C, and before when
we talked about Latin squares, two of these were blocking factors and the third was the
treatment factor. We could estimate all three main effects and we could not estimate any of
the interactions. And now you should be able to see why. The interactions are all aliased
with the main affects.

Let's look at another component LAB2C of the three factor interaction: A×B×C:

LAB2C = X1 + 2X2 + X3 (mod 3)

We can now fill out the table by first plugging in the levels of X1, X2 and X3 from the levels of
A, B and C to generate the column LAB2C. When you assign treatments to the level of LAB2C
= 0 you get an arrangement that follows (only the principle block filled in):

LAB2C
0 1 2
0, 0,
Loading [MathJax]/extensions/MathZoom.js 0

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 18 of 22

1, 1, 0
2, 2, 0
2, 0, 1
0, 1, 1
1, 2, 1
1, 0, 2
2, 1, 2
0, 2, 2

Then it also generates its own Latin square using the same process that we used above.
You should be able to follow how this Latin square was assigned to the nine treatment
combinations from the table above.

B
C 0 1 2
0 0 1 2
A 1 2 0 1
2 1 2 0

The benefit of doing this is to see that this one third fraction is also a Latin square. This is a
Resolution III design, (it has a three letter word generator), and so it has the same
properties that we saw at the two level designs, i.e. the main effects are clear of each other
and estimable and aliased with higher order interactions including two-way. In fact, since the
ABC and the AB2C are orthogonal to each other - they partition the A×B×C interaction - the
two Latin squares we constructed are orthogonal Latin Squares.

The Next Level Example - Four Factors

Now let's take a look at the 34-2 design. How do we create this design? In this case we
would have to pick 2 generators. We have four factors, A, B, C and D. So, let's say we will
begin (trial and error) by selecting I = ABC = BCD as our generators then we will also have
the generalized interactions between those generators which are also included. Thus we will
also confound:

ABC × BCD = AB2C2D, and

ABC × (BCD)2 = AD2

This is a Resolution II design - there are only two letters in the second component and we
should be able to do better.

Let's try again, how about

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 19 of 22

I = ABC = BC2D as our generators. This confounds:

ABC × BC2D = AB2D

ABC × (BC2D)2 = AC2D2

This is much better because there is nothing smaller than a three letter word in the
generator set so this is a Resolution III design. Now, how do we generate the design? It is a
design with four factors but how many observations are there? Nine. It is still a design with
only nine observations, or a 1/9th fraction of a 34 design or 81 observations. If we can write
out the basic design with nine observations, which we can do with A and B, it gives us the
basic design, and then we use our generators to give us C and D. We can use ABC such
that:

LABC = 0 this principle fraction implies that X3 = 2X1 + 2X2 (mod 3).

LBC2D = 0 this implies that X4 = 2X2 + X3 (mod 3)

If we were confounding this in blocks we will want a principal block where these two defining
relationships are both zero. You will see that by defining X3 and X4 in this way results in
ABC being equal to zero. Take a look and make sure that you understand how column C
was generated by the function X3 = 2X1 + 2X2 (mod 3) yet still preserves the principle
implied where LABC = 0. Also, by the same process column D was generated using the
function X4 = 2X2 + X3 (mod 3) in such a way that it preserves the principle implied where in
LBC2D = 0.

And so, the 34-2 design is equivalent to the Graeco-Latin square. There are two Latin
squares, one for each component, C and D, superimposed as shown below:

So we can see that the Graeco-Latin Square with three treatments is simply a fractional
factorial of this 34 design!

9.3 - Mixed Factorials

We have been talking about 2-level designs and 3-level designs. 2 level designs for
screening factors and 3 level designs analogous to the 2 level designs, but the beginning of
our discussion of response surface designs.

Since a 2 level design only has two levels of each factor, we can only detect linear
effects.We have been mostly thinking about quantitative factors but especially when
screening two level designs the factors can be presence/absence, or two types and you can
still throw it into that framework and decide whether that's an important factor. If we go to
three level designs we are almost always thinking about quantitative factors. But, again, it
doesn't always have to be, it could be three types of something. However, in the general
application we are talking about quantitative factors.

If we take a 2 level design that has center points.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 20 of 22

Then, if you project into the A axis or the B axis, you have three distinct values, -1, 0, and
+1.

In the main effect sense, a two level design with center points gives you three levels. This
was our starting point towards moving to a three level design. Three-level designs require a
whole lot more observations. With just two factors, i.e., k = 2, you have 3k = 9 observations,
but as soon as we get to k = 4, now you already have 34 = 81 observations, and with k = 5
becomes out of reach - 35 = 243 observations. These designs grow very fast so obviously
we are going to look for more efficient designs.

Mixed Level Designs

When we think of next level designs we think of factors with 4 or 5 levels, or designs with
combinations of 2, 3, 4, or 5 levels of factors. In an Analysis of Variance course, which most
of you have probably taken, it didn't distinguish between these factors. Instead, you looked
at general machinery for factors with any numbers of level. What is new here is thinking
about writing efficient designs. Let's say you have a 23 × 32 - this would be a mixed level
design with 8 × 9 = 72 observations in a single replicate. So this is growing pretty rapidly! As
this gets even bigger we could trim the size of this by looking at fractions for instance, 23-1, a
fractional factorial of the first part. And, as these numbers of observations get larger you
could look at crossing fractions of factorial designs.

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 21 of 22

A Note about Factors with 4 levels - 22

This design is 22, so in some sense there is nothing new here. By using the machinery of
the 2k designs you can always take a factor with four levels and call it the four combinations
of 22.

A Note about Factors with 5 levels

Design with factors with 5 levels... Think quantitative - if it is quantitative then you have five
levels, and we should then be thinking about fitting a polynomial regression function.

This leads us to a whole new class of designs that we will look at next - Response Surface
Designs.

What we have plotted here is a 22 design, which are the four corners of a 22. We have
center points. And then to achieve what we will refer to as a central composite design we
will add what are called star points (axial points). These are points that are outside the range
of -1 and 1 in each dimension. If you think in terms of projecting, we now have 5 levels of
each of these 2 factors obtained in some automatic way. Instead of having 25 points which
is what a 5 x 5 requires, we only have 9 points. It is a more efficient design but still in a
projection we have five levels in each direction. What we want is enough points to estimate
a response surface but at the same time keep the design as simple and with as few
observations as possible.

The primary reason that we looked at the 3k designs is to understand the confounding that
occurs. When we have quantitative variables we will generally not use a 3 level designs. We
use this more for understanding of what is going on. In some sense 3 level designs are not
as practical as CCD designs. We will next consider response surface designs to address to
goals of fitting a response surface model.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/53

Links:
[1]
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 22 of 22

https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson09/Figure-
9-7.MPJ

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/53/ 4/18/2019
 Page 1 of 1

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 10: Simple Linear Regression

Lesson 10: Simple Linear Regression

This lesson corresponds to Chapter 10: Fitting Regression Models.

A course on fitting regression models is a prerequisite for this course. Chapter 10 covers
standard topics in regression. Please read the Chapter if you feel you need to review and
then proceed to Chapter 11.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/91

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/91/ 4/18/2019
 Page 1 of 29

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 11: Response Surface Methods and Designs

Lesson 11: Response Surface Methods and

Designs
Overview of Response Surface Methods

We are now going to shift from screening designs where the primary focus of previous lessons
was factor screening – (two-level factorials, fractional factorials being widely used), to trying to
optimize an underlying process and look for the factor level combinations that give us the
maximum yield and minimum costs. In many applications, this is our goal. However in some
cases we are trying to hit a target or aim to match some given specifications - but this brings up
other issues which we will get to later.

Here the objective of Response Surface Methods (RSM) is optimization, finding the best set of
factor levels to achieve some goal. This lesson aims to cover the following goals:

Learning Objectives & Outcomes

• Response Surface Methodology and its sequential nature for optimizing a process
• First order and second order response surface models and how to find the direction of
steepest ascent (or descent) to maximize (or minimize) the response
• How to deal with several responses simultaneously (Multiple Response Optimization)
• Central Composite Designs (CCD) and Box-Behnken Designs as two of the major
Response Surface Designs and how two generate them using Minitab
• Design and Analysis of Mixture Designs for cases where the sum of the factor levels equals
a constant, i.e. 100% or the totality of the components
• Introductory understanding of designs for computer models

RSM dates from the 1950's. Early applications were found in the chemical industry. We have
already talked about Box. Box and Draper have some wonderful references about building RSMs
and analyzing them which are very useful.

RSM as a Sequential Process

The text has a graphic depicting a response surface method in three dimensions, though actually
it is four dimensional space that is being represented since the three factors are in 3-dimensional
space the the response is the 4th dimension.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 2 of 29

Instead, let's look at 2 dimensions - this is easier to think about and visualize. There is a response
surface and we will imagine the ideal case where there is actually a 'hill' which has a nice
centered peak. (If only reality were so nice, but it usually isn't!). Consider the geologic ridges that
exist here in central Pennsylvania, the optimum or highest part of the 'hill' might be anywhere
along this ridge. There's no clearly defined centered high point or peak that stands out. In this
case there would be a whole range of values of X1 and X2 that would all describe the same 'peak'
-- actually the points lying along the top of the ridge. This type of situation is quite realistic where
there does not exist a predominate optimum.

But for our purposes let's think of this ideal 'hill' and the problem is that you don't know where this
is and you want to find factor level values where the response is at its peak. This is your quest, to
find the values X1optimum and X2optimum, where the response is at its peak. You might have a
hunch that the optimum exists in certain location. This would be good area to start - some set of
conditions, perhaps the way that the factory has always been doing things - and then perform an
experiment at this starting point.

The actual variables in their natural units of measurement are used in the experiment. However,
when we design our experiment we will use our coded variables, X1 and X2 which will be
centered on 0, and extend +1 and -1 from the center of the region of experimentation. Therefore,
we will take our natural units and then center and rescale them to the range from -1 to +1.

Our goal is to start somewhere using our best prior or current knowledge and search for the
optimum spot where the response is either maximized or minimized.

Here are the models that we will use.

Screening Response Model

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 3 of 29

y = β0 + β1x1 + β2x2 + β12x1x2 + ε (1)

The screening model that we used for the first order situation involves linear effects and a single
cross product factor, which represents the linear x linear interaction component.

Steepest Ascent Model

If we ignore cross products which gives an indication of the curvature of the response surface
that we are fitting and just look at the first order model this is called the steepest ascent model:

y = β0 + β1x1 + β2x2 + ε (2)

Optimization Model

Then, when we think that we are somewhere near the 'top of the hill' we will fit a second order
model. This includes in addition the two second-order quadratic terms.

y = β0 + β1x1 + β2x2 + β12x1x2 + β11x12 + β22x22 + ε (3)

Steepest Ascent - The First Order Model

Let's look at the first order situation - the method of steepest ascent. Now, remember, in the first
place we don't know if the 'hill' even exists so we will start somewhere where we think the
optimum exists. We start somewhere in terms of the natural units and use the coded units to do
our experiment. Consider the example 11.1 in the textbook. We want to start in the region where
x1 = reaction time (30 - 40 seconds) and x2 = temperature (150 - 160 degrees), and we want to
look at the yield of the process as a function of these factors. In a sense, for the purpose of
illustrating this concept, we can superimpose this region of experimentation on to our plot of our
unknown 'hill'. We obviously conduct the experiment in its natural units but the designs will be
specified in the coded units so we can apply them to any situation.

Specifically, here we use a design with four corner points, a 22 design and five center points. We
now fit this first-order model and investigate it.

We put in the actual data for A and B and the response measurements Y.

We fit a full model first: See: Ex-11-1-output.doc [1]

We fit the surface. The model has two main effects, one cross product term and then one
additional parameter as the mean for the center point. The residuals in this case have four df
which come from replication of the center points. Since there are five center points,
i.e., four df among the five center points. This is a measure of pure error.

We start by testing for curvature. The question is whether the mean of the center points is
different from the values at (x1,x2) = (0,0) predicted from the screening response model (main
effects plus interaction). We are testing whether the mean of the points at the center are on the
plane fit by the four corner points. If the p-value had been small, this would have told you that a
mean of the center points is above or below the plane indicating curvature in the response
surface. The fact that, in this case, it is not significant indicates there is no curvature. Indeed the
center points fall exactly on the plane that fits the quarter points.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 4 of 29

There is just one degree of freedom for this test because the design only has one additional
location in terms of the x's.

Next we check for significant effects of the factors. We see from the ANOVA that there is no
interaction. So, let's refit this model without the interaction term, leaving just the A and B terms.
We still have the average of the center points and our AOV now shows 5 df for residual error.
One of these is lack of fit of the additive model and there are 4 df of pure error as before. We
have 1 df for curvature, and lack of fit in this case is just the interactions from the model.

What do we do with this? See the Minitab analysis and redo these results in EX11-1.MPJ [2]

Our estimated model is: y^ = 40.43 + 0.775x1 + 0.325x2

So, for any X1 and X2 we can predict y. This fits a flat surface and it tells us that the predicted y is
a function of X1 and X2 and the coefficients are the gradient of this function. We are working in
coded variables at this time so these coefficients are unitless.

If we move 0.775 in the direction of X1 and then 0.325 in the direction of X2 this is the direction of
steepest ascent. All we know is that this flat surface is one side of the 'hill'.

The method of steepest ascent tells us to do a first order experiment and find the direction that
the 'hill' goes up and start marching up the hill taking additional measurements at each (x1, x2)
until the response starts to decrease. If we start at 0, in coded units, then we can do a series of
single experiments on this path up the 'hill' of the steepest descent. If we do this at a step size of
x1 = 1, then:

1 / 0.775 = x2 / 0.325 → x2 = 0.325 / 0.775 = 0.42

and thus our step size of x1 = 1 determines that x2 = 0.42, in order to move in the direction
determined to be the steepest ascent. If we take steps of 1 in coded units, this would be five
minutes in terms of the time units. And for each step along that path we would go up 0.42 coded
units in x2 or approximately 2° on the temperature scale.

Here is the series of steps in additional meanures of five minutes and 2° temperature. The
response is plotted and shows an increase that drops off towards the end.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 5 of 29

This is a pretty smooth curve and in reality you probably should go a little bit more beyond the
peak to make sure you are at the peak. But all you are trying to do is to find out approximately
where the top of the 'hill' is. If your first experiment is not exactly right you might have gone off in
a wrong direction!

So you might want to do another first-order experiment just to be sure. Or, you might wish to do a
second order experiment, assuming you are near the top. This is what we will discuss in the next
section. The second order experiment will help find a more exact location of the peak.

The point is, this is a fairly cheap way to 'scout around the mountain' to try to find where the
optimum conditions are. Remember, this example is being shown in two dimensions but you may
be working in three or four dimensional space! You can use the same method, fitting a first-order
model and then moving up the response surface in k dimensional space until you think you are
close to where the optimal conditions are.

If you are in more than 2 dimensions, you will not be able to get a nice plot. But that is OK. The
method of steepest ascent tells you where to take new measurements, and you will know the
response at those points. You might move a few steps and you may see that the response
continued to move up or perhaps not - then you might do another first order experiment and
redirect your efforts. The point is, when we do the experiment for the second order model, we
hope that the optimum will be in the range of the experiment - if it is not, we are extrapolating to
find the optimum. In this case, the safest thing to do is to do another experiment around this
estimated optimum. Since the experiment for the second order model requires more runs than
experiments for the first order model, we want to move into the right region before we start fitting
second order models.

Steepest Ascent - The Second Order Model

y = β0 + β1x1 + β2x2 + β12x1x2 + β11x12 + β22x22 + ε

This second order model includes linear terms, cross product terms and a second order term for
each of the x's. If we generalize this to k x's, we have k first order terms , k second order terms
and then we have all possible pairwise first-order interactions. The linear terms just have one
subscript. The quadratic terms have two subscripts. There are k*(k-1)/2 interaction terms. To fit

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 6 of 29

this model, we are going to need a response surface design that has more runs than the first
order designs used to move close to the optimum.

This second order model is the basis for response surface designs under the assumption that
although the hill is not a perfect quadratic polynomial in k dimensions, it provides a good
approximation to the surface near the maximum or a minimum.

Assuming that we have 'marched up this hill' and if we re-specified the region of interest in our
example, we are now between 80-90 in terms of time and 170-180 in terms of temperature. We
would now translate these natural units into our coded units and if we fit the first order model
again, hopefully we can detect that the middle is higher than the corner points so we would have
curvature in our model, and could now fit a quadratic polynomial.

After using the Steepest Ascent method to find the optimum location in terms of our factors, we
can now go directly to the second order response surface design. A favorite design that we
consider is sometimes referred to as a central composite design. The central compositive design
is shown on Figure 11.3 above and in more detail in the text in Figure 11.10. The idea is simple -
take the 2k corner points, add a center point, and then create a star by drawing a line through the
center point orthogonal to each face of the hypercube. Pick a radius along this line and place a
new point at that radius. The effect is that each factor is now measured at 5 levels - center, 2
corners and the 2 star points. This gives us plenty of unique treatments to fit the 2nd order model
with treatment degrees of freedom left over to test the goodness of fit. Replication is still usually
done only at the center point.

11.1 - Multiple Responses

In many experiments more than one response is of interest for the experimenter. Furthermore,
we sometimes want to find a solution for controllable factors which result in the best possible
value for each response. This is the context of multiple response optimization, where we seek a
compromise between the responses; however, it is not always possible to find a solution for
controllable factors which optimize all of the responses simultaneously. Multiple response
optimization has an extensive literature in the context of multiple objective optimization which is
beyond the scope of this course. Here, we will discuss the basic steps in this area.

As expected, multiple response analysis starts with building a regression model for each
response separately. For instance, in Example 11.2 we can fit three different regression models
for each of the responses which are Yield, Viscosity and Molecular Weight based on two
controllable factors: Time and Temperature.

One of the traditional methods way to analyze and find the desired operating condition one is
overlaid contour plots. This method is mainly useful when we have two or maybe three
controllable factors but in higher dimensions it loses its efficiency. This method simply consists of
overlaying contour plot for each of the responses one over another in the controllable factors
space and finding the area which makes the best possible value for each of the responses.
Figure 11.16 (Montgomery, 7th Edition) shows the overlaid contour plots for example 11.2 in
Time and Temperature space.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 7 of 29

Figure 11.16 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th Edition)

The unshaded area is where yield > 78.5, 62 < viscosity < 68, and molecular weight < 3400. This
area might be of special interest for the experimenter because they satisfy given conditions on
the responses.

Another dominant approach for dealing with multiple response optimization is to form a
constrained optimization problem. In this approach we treat one of the responses as the
objective of a constrained optimization problem and other responses as the constraints where the
constraint’s boundary is to be determined by the decision maker (DM). The Design-Expert
software package solves this approach using a direct search method.

Another important procedure that we will discuss here, also implemented in Minitab, is the
desirability function approach. In this approach the value of each response for a given
combination of controllable factors is first translated to a number between zero and one known as
individual desirability. Individual desirability functions are different for different objective types
which might be Maximization, Minimization or Target. If the objective type is maximum value, the
desirability function is defined as

⎧ 
⎪  0 y<L
⎪ 
y−L r
d = ⎨ ( T−L ) L≤y≤T
⎪ 
⎩ 
⎪ 
1 y>T
When the objective type is a minimum value the, the individual desirability is defines as

⎧ 
⎪  1 y<T
⎪ 
U−y r
d = ⎨ ( U−T ) T ≤y≤U
⎪ 
⎩ 
⎪ 
0 y>U

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 8 of 29

Finally the two-sided desirability function with target-the-best objective type is defined as

⎧ 
⎪  0 y<L
⎪ 
⎪ 
⎪ 
⎪ 
⎪  y−L r1
⎪ ( T−L ) L≤y≤T
d = ⎨ 
⎪  U−y r2
⎪ 
⎪ 
⎪ ( U−T
) T ≤y≤U
⎪ 
⎪ 
⎩ 
⎪ 
0 y>U
Where the r1 , r2 and r define the shape of the individual desirability function (Figure 11.17 in the
text shows the shape of individual desirability for different values of shape parameter). Individual
desirability is then used to calculate the overall desirability using the following formula:

D = (d1 d2 … dm )1/m
where m is the number of responses. Now, the design variables should be chosen so that the
overall desirability will be maximized. Minitab’s Stat > DOE > Response Surface > Response
Optimizer routine uses the desirability approach to optimize several responses, simultaneously.

11.2 - Response Surface Designs

After using the Steepest Ascent method to find the optimum location in terms of our factors, we
can now go directly to the second order response surface design. A favorite design that we consider
for a second order model is referred to as a central composite design.

We give here an example in two dimensions, Example11.2 in the text. We have 2k corner points
and we have some number of center points which generally would be somewhere between 4 and
7, (five here). In two dimensions there are 4 star points, but in general there are 2k star points in
k dimensions. The value of these points is something greater than 1. Why is it something greater
than 1? If you think about the region of experimentation, we have up to now always defined a
box, but if you think of a circle the star points are somewhere on the circumference of that circle,
or in three dimensions on the ball enclosing the box. All of these are design points around the
region where you expect the optimum outcome to be located. Typically the only replication, in
order to get some measure of pure error, is done at the center of the design.

The data set for the Example 11.2 is found in the Minitab worksheet, Ex11-2.MTW [3]. The
analysis using the Response Surface Design analysis module is shown in the Ex11-2.MPJ [4].

11.2.1 - Central Composite Designs

In the last section we looked at the example 11.2 which was in coded variables and was a central
composite design.

In this section we examine a more general central composite design. For k = 2 we had a 22
design with center points, which was required for our first order model; then we added 2*k star
points. The star or axial points are, in general, at some value α and -α on each axis.

There are various choices of α. If α = 1, the star points would be right on the boundary, and we
would just have a 32 design. Thus α = 1 is a special case, a case that we considered in the 3k

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 9 of 29

designs. A more common choice of α is which gives us a spherical design as shown

below.
−
α = √k

Our 22 design gives us the box, and adding the axial points (in green) outside of the box gives us
−
a spherical design where α = √k . The corner points and the axial points at α, are all points on
the surface of a ball in three dimensions, as we see below.

This design in k = 3 dimensions can also be referred to as a central composite design, chosen so
that the design is spherical. This is a common design. Much of this detail is given in Table 11.11
of the text.
1
An alternative choice where α = (nF ) 4 , or the fourth root of the number of points in the
factorial part of the design, gives us a rotatable design.

If we have k factors, then we have, 2k factorial points, 2*k axial points and nc center points. Below
is a table that summarizes these designs and compares them to 3k designs:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 10 of 29

k=2 k=3 k=4 k=5

Factorial points 2k 4 8 16 32
Central Composite
Designs Star points 2k 4 6 8 10
Center points nc (varies) 5 5 6 6
Total 13 19 30 48
3k Designs 9 27 81 243
−
Spherical design ( α = √k ) 1.4 1.73 2 2.24
Choice of α 1
Rotatable design ( α = (nF ) 4 ) 1.4 1.68 2 2.38

Compare the total number of observations required in the central composite designs versus the
3k designs. As the number of factors increases you can see the efficiencies that are brought to
bear.

The spherical designs are rotatable in the sense that the points are all equidistant from the
center. Rotatable refers to the variance of the response function. A rotatable design exists when
there is an equal prediction variance for all points a fixed distance from the center, 0. This is a
nice property. If you pick the center of your design space and run your experiments, all points that
are equal distance from the center in any direction, have equal variance of prediction.

You can see in the table above that the difference in the variation between the spherical and
rotatable designs are slight, and don't seem to make much difference. But both ideas provide
justification for selecting how far away the star points should be from the center.

Why do we take about five or six center points in the design? The reason is also related to the
variance of a predicted value. When fitting a response surface you want to estimate the
response function in this design region where we are trying to find the optimum. We want the
prediction to be reliable throughout the region, and especially near the center since we hope the
optimum is in the central region. By picking five to six center points, the variance in the middle is
approximately the same as the variance at the edge. If you only had one or two center points,
then you would have less precision in the middle than you would have at the edge. As you go
farther out beyond a distance of 1 in coded units, you get more variance and less precision. What
we are trying to do is to balance the precision at the edge of the design relative to the middle.

How do you select the region where you want to run the experiment? Remember, for each factor
X we said we need to choose the lower level is and the upper level for the region of
experimentation. We usually picked the -1 and 1 as the boundary. If the lower natural unit is really
the lowest number that you can test, because the experiment won't work lower than this, or the
lower level is zero and you can't put in a negative amount of something, then, the star point is not
possible because it is outside the range of experimentation.

If this is the case, one choice that you could make would be to use the -α as the lowest point.
Generally, if you are not up against a boundary then this is not an issue and the star points are a
way to reach beyond the region that you think the experiment should be run in. The issue isn't
selecting the coding of the design relative to the natural units. You might lose some of these
exact properties, but as long as you have the points nicely spread out in space you can fit a
regression function. The penalty for not specifying the points exactly, would be seen in the
variance, and it would be actually very slight.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 11 of 29

Generating These Designs in Minitab

Minitab will show you the available designs and how to generate these designs.

We can create central composite designs using a full factorial, central composite designs with
fractional factorials, half fraction and a quarter fraction, and they can be arranged in blocks. Later,
we will look at the Box-Behnken designs.

As an example, we look at the k = 3 design, set up in Minitab using a full factorial, completely
randomized, in two blocks, or three blocks with six center points and the default α = 1.633 (or α =
1.682 for a rotatable design).

If you do not want the default α you can specify your own in the lower left. A Face Centered
design is obtained by putting the α at +1 and - 1 on the cube. Here is the design that results:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 12 of 29

The first block is a one half fraction of the 23 plus 2 center points. Block 2 is the second half
fraction of the factorial part with two center points. The third block consists of 6 star points, plus to
center points. Each of the three blocks contains 2 center points and the first two blocks have half
of the corner points each. The third block contains the star points and is of size 8.

Rollover the words 'Block 1', 'Block 2', and 'Block 3' in the graphic above. Do you see how they
use center points strategically to tie the blocks together? They are represented in each block and
they keep the design connected.

The corner points all have +1 or - 1 for every dimension, because they're at the corners. They are
either up or down, in or out, right or left. The axial points have + α, or -α (+1.6330 or -1.6330) for
A, but are 0 for factors B and C. The center points have zero on all three axes, truly the center of
this region. We have designed this to cover the space in just the right way so that we can
estimate a quadratic equation.Using a Central Composite Design, we can't estimate cubic terms,
and we can't estimate higher order interactions. If we had utilized a 3k design, one that quickly
becomes unreasonably large, then we would have been able to estimate all of the higher order
interactions.

However, we would have wasted a lot of resources to do it. The CCD allows us to estimate just
linear and quadratic terms and first order interactions.

An Example - Polymers and Elasticity

This example is from the Box and Draper (1987) book and
the data from Tables 9.2 and 9.4 are in Minitab
(BD9-1.MTW [5]).

This example has three variables and they are creating a

polymer, a kind of plastic that has a quality of elasticity.
The measurement in this experiment is the level of
elasticity. We created the design in Minitab for this
experiment, however the data only has two center points:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 13 of 29

Variables A and B are the concentration of two ingredients that make up the polymer, and C is
the temperature, and the response is elasticity. There are 8 corner points, a complete factorial, 6
star points and 2 center points.

Let's go right to the analysis stage now using Minitab ...

A video demonstration is given here: https://screencast.com/t/yfyiPXdPq [6]

or you can view the flash animation...

[7]

Before we move on I would like to go back and take a look again at the plot of the residuals. Wait
a minute! Is there something wrong with this residual plot?

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 14 of 29

Residual plot for the polynomial fit

Look at the plot in the lower right. The first eight points tend to be low, and then the next eight
points are at a higher level. This is a clue, that something is influencing the response that is not
being fit by the model. This looks suspicious. What happened? My guess is that the experiment
was run in two phases. They first ran the 2k part - (block 1). And then they noticed the response
and added the star points to make a responsive surface design in the second part. This is often
how these experiments are conducted. You first perform a first-order experiment, and then you
add center points and star points and then fit the quadratic.

Add a block term and rerun the experiment to see if this makes a difference.

Two Types of Central Composite Designs

The central composite design has 2*k star points on the axial lines outside of the box defined by
the corner points. There are two major types of central composite designs: the spherical central
composite design where the star points are the same distance from the center as the corner
points, and the rotatable central composite design where the star points are shifted or placed
such that the variances of the predicted values of the responses are all equal, for x’s which are
an equal distance from the center.

When you are choosing, in the natural units, the values corresponding to the low and high, i.e.
corresponding to -1 and 1 in coded units, keep in mind that the design will have to include points
further from the center in all directions. You are trying to fit the design in the middle of your
region of interest, the region where you expect the experiment to give the optimal response.

11.2.2 - Box-Behnken Designs

Box-Behnken Designs

Another class of response surface designs are called Box-Behnken designs. They are very useful
in the same setting as the central composite designs. Their primary advantage is in addressing
the issue of where the experimental boundaries should be, and in particular to avoid treatment
combinations that are extreme. By extreme, we are thinking of the corner points and the star
points, which are extreme points in terms of region in which we are doing our experiment. The
Box-Behnken design avoids all the corner points, and the star points.

One way to think about this is that in the central composite design we have a ball where all of the
corner points lie on the surface of the ball. In the Box-Behnken design the ball is now located
inside the box defined by a 'wire frame' that is composed of the edges of the box. If you blew up a
balloon inside this wire frame box so that it just barely extends beyond the sides of the box, it
might look like this, in three dimensions. Notice where the balloon first touches the wire frame;
this is where the points are selected to create the design.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 15 of 29

Therefore the points are still on the surface of a ball, but the points are never further out than the
low and high in any direction. In addition, there would be multiple center points as before. In this
type of design you do not need as many center points because points on the outside are closer to
the middle. The number of center points are again chosen so that the variance of is about the
same in the middle of the design as it is on the outside of the design.

In Minitab we can see the different designs that are available. Listed at the bottom are the Box-
Behnken Designs.

A Box-Behnken (BB) design with two factors does not exist. With three factors the BB design by
default will have three center points and is given in the Minitab output shown above. The last
three observations are the center points. The other points, you will notice, all include one 0 for
one of the factors and then a plus or minus combination for the other two factors.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 16 of 29

If you consider the BB design with four factors, you get the same pattern where we have two of
the factors at + or - 1 and the other two factors are 0. Again, this design has three center points,
and a total of 27 observations.

Comparing the central composite design with 4 factors, which has 31 observations, a Box-
Behnken design only includes 27 observations. For 5 factors, the Box-Behnken would have 46
observations, and a central composite would have 52 observations if you used a complete
factorial, but this is where the central composite also allows you to use a fractional factorial as a
means of making this experiment more efficient. Likewise for six factors, the Box-Behnken
requires 54 observations, and this is the minimum of the central composite design.

Both the CCD and the BB design can work, but they have different structures, so if your
experimental region is such that extreme points are a problem then there are some advantages
to the Box-Behnken. Otherwise, they both work well.

The central composite design is one that I favor because even though you are interested in the
middle of a region, if you put all your points in the middle you do not have as much leverage
about where the model fits. So when you can move your points out you get better information
about the function within your region of experimentation.

However, by moving your points too far out, you get into boundaries or could get into extreme
conditions, and then enter the practical issues which might outweigh the statistical issues. The
central composite design is used more often but the Box-Behnken is a good design in the sense
that you can fit the quadratic model. It would be interesting to look at the variance of the predicted
values for both of these designs. (This would be an interesting research question for somebody!)
The question would be which of the two designs gives you smaller average variance over the
region of experimentation.

The usual justification for going to the Box-Behnken is to avoid the situation where the corner
points in the central composite design are very extreme, i.e. they are at the highest level of
several factors. So, because they are very extreme, the researchers may say these points are not
very typical. In this case the Box Behnken may look a lot more desirable since there are more
points in the middle of the range and they are not as extreme. The Box-Behnken might feel a little
'safer' since the points are not as extreme as all of the factors.

The Variance of the Predicted Values

Let's look at this a little bit. We can write out the model:

y^x = b0 + b1x1 + b2x2 + b11x12 + b22x22 + b12x1x2

Where the b0, b1, etc are the estimated parameters. This is a quadratic model with two x's. The
question we want to answer is how many center points should there be so that the variance of the
predicted value, var(y^x ) when x is at the center is the same as when x is at the outside of the
region?

See handout Chapter 11: Supplemental Text Material [8]. This shows the impact on the variance
of a predicted value in the situation with k = 2, full factorial and the design has only 2 centerpoints
rather than the 5 or 6 that the central composite design would recommend.

What you see (S11-3) is that in the middle of the region the variance is much higher than further
out. So, by putting more points in the center of the design, collecting more information there,
(replicating a design in the middle), you see that the standard error is lower in the middle and

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 17 of 29

roughly the same as farther out. It gets larger again in the corners and continues growing as you
go out from the center. By putting in enough center points you balance the variance in the middle
of the region relative to further out.

Another example (S11-4) is a central composite design where the star points are on the face. It is
not rotatable design and the variance changes depending on which direction you're moving out
from center of the design.

It also shows another example (S11-4), also a face-centered design with zero center points,
which shows a slight hump in the middle on the variance function.

Notice that we only need two center points for the face centered design. Rather than having our
star points farther out, if we move them closer into the face we do not need as many center points
because we already have points closer to the center. A lot of factors affect the efficiencies of
these designs.

Rotatability

Rotatability is determined by our choice of alpha. A design is rotatable if the prediction variance
depends only on the distance of the design point from the center of the design. This is what we
were observing previously. Here in the supplemental material (S11-5) is an example with a
rotatable design, but the variance contours are based on a reduced model. It only has one
quadratic term rather than two. As a result we get a slightly different shape, the point being that
rotatability and equal variance contours depend both on the design and on the model that we are
fitting. We are usually thinking about the full quadratic model when we make that claim.

11.3 - Mixture Experiments

This is another class of response surface designs where the components are not just levels of
factors but a special set where the x1, x2, ... are coded and are the components of the mixture
such that the sum of the xi = 1. So, these make up the proportions of the mixture.

Examples

If you are making any kind of product it usually involves mixtures of ingredients. A classic
example is gasoline which is a mixture of various petrochemicals. In polymer production,
polymers are actually mixtures of components as well. My favorite classroom example is baking a
cake. A cake is a mixture of flour, sugar, eggs, and other ingredients depending on the type of
cake. It is a mixture where the levels of x are the proportions of the ingredients.

This constraint that the sum of the x's sum to 1, i.e.,

0 ≤ xi ≤ 1

has an impact on how we analyze these experiments.

Here we will write out our usual linear model:

Yi = β0 + β1 xi1 + β2 xi2 + … + βk xik + εi

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 18 of 29

where,
k
1 = ∑ xthis
If you want to incorporate ij constraint then we can write:
j=1
Yi = β1 xi1 + β2 xi2 + … + βk xik + εi
in other words, if we drop the β0, this reduces the parameter space by 1 and then we can fit a
reduced model even though the x's are each constrained.

In the quadratic model:

k
Yh = ∑ βi xhi + ∑ ∑ βij xhi xhj + εh
i=1 i<j

This is probably the model we are most interested in and will use the most. Then we can
generalize this into a cubic model which has one additional term.

A Cubic Model

k k k k
Yh = ∑ βi xhi + ∑ ∑ βij xhi xhj + ∑ ∑ ∑ βijl xhi xhj xhl + εh
i=1 i<j i<j<l

These models are used to fit response surfaces.

Let's look at the parameter of space. Let's say that k = 2. The mixture is entirely made up of two
ingredients, x1 and x2. The sum of both ingredients is a line plotted in the parameter space below:
An experiment made up of two components is either all of x1or all of x2 or something in between,
a proportion of the two. Use your mouse to click and drag the intersection point along the line that
serves as a boundary to this region of experimentation.

Let's take a Look at the parameter space in three dimensions. Here we have three components:
x1, x2 and x3. As we satisfy our constraint that the sum of all the components equal 1 and then
our parameter space is the plane that cuts the three-dimensional surface, intersecting these three
points in the graph below scratch that in the plot below.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 19 of 29

Plot showing the plane where the sum: x1+x2+x3 = 1

The triangle represents the full extent of the region of experimentation in this case with the points
sometimes referred to as the Barycentric coordinates. The design question we want to address is
where do we do our experiment? We are not interested in any one of the corners of the triangle
where only one ingredient is represented, we are interested in some way on the middle where
there is a proportion of all three of the ingredients included. We will restrict it to a feasible region
of experimentation somewhere in the middle area.

Let's look at an example, for instance, producing cattle feed. The ingredients might include the
following: corn, oats, hay, soybean, grass, ... all sorts of things.

In some situations it might work where you might have 100% of one component, but many
instances of mixtures we try to partition off a part of the space in the middle where we think the
combination is optimal.

In k = 4 the region of experimentation can be represented by the shape of a tetrahedron where

each of the four sides of the tetrahedron is an equalateral triangle and has its own set of
Barycentric coordinates. Each face of the tetrahedron corresponds to design region where one
coordinate is zero, and the remaining three must sum to 1.

11.3.1 - Two Major Types of Mixture Designs

Simplex Lattice Design

A {p,m} simplex lattice design for p factors (components) is defined as all possible combination of
factor levels defined as
1 2
xi = 0, m, m,⋯,1 i = 1, 2, … , p
As an example, the simplex lattice design factor levels for the case of {3,2} will be

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 20 of 29

Which results in the following design points:

xi = 0, 12 , 1 i = 1, 2, 3
(x1 , x2 , x3 ) = {(1, 0, 0), (0, 1, 0), (0, 0, 1), ( 12 , 12 , 0), ( 12 , 0, 12 ), (0, 12 , 12 )}

Simplex Centroid Design

This design which has 2p-1 design points consist of p permutations of (1,0,0,…,0), permutations
p p
of (1, 0, 0, … , 0), ( ) , permutations of ( 12 , 12 , 0, … , 0), ( ) , and the overall centroid
2 3
1 1 1
( , , ⋯ , ) . Some simplex centroid designs for the case of p = 3 and p = 4 can be find in
p p p
Figure 11.41.

Minitab handles mixture experiments which can be accessed through Stat > DOE > Mixture. It
allows for building and analysis of Simplex Lattice and Simplex Centroid designs. Furthermore, it
covers a third design which is named, Extreme Vertex Design. Application of Extreme Vertex
designs are for cases where we have upper and lower constraints on some or all of the
components making the design space smaller than the original region.

11.3.2 - Mixture Designs in Minitab

How does Minitab handle these types of experiments?

Mixture designs are a special case of response surface designs. Under the stat menu in many
tab, select design of experiments, then mixture, create mixture design. Minitab then presents you
with the following dialog box:

Simplex lattice option will look at the points that are extremes. Simplex lattice creates a design for
p components of degree m. In this case, we want points that are made up of 0, 1/m, 2/m, ... up to 1.
Classifying the points in this way tells us how we will space the points. For instance, if m = 2, then
the only points we would have would be 0, 1/2, and 1 to play with in all key dimensions. You can

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 21 of 29

create this design in Minitab, for 3 factors, using tab Stat > DOE > Mixture > Create Mixture
Design and select Simplex Centroid. See the image here:

If we are in a design with the m = 3, then we would have 0, 1/3, 2/3, and 1. In this case we would
have points a third of the way along each dimension. Any point on the boundary can be
constructed in this way.

All of these points are on the boundary which means that they are made up of mixtures that omit
one of the components. (This is not always desirable but in some settings it is fine.)

The centroid is the point in the middle. Axial points are points that lie along the lines that intersect
the region of experimentation, points that are located interior and therefore include part of all of
the components.

You can create this design in Minitab, for 3 factors, using tab Stat > DOE > Mixture > Create
Mixture Design and select Simplex Centroid. See the image here:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 22 of 29

This should give you the range of points that you think of when designing in a mixture. again, you
want points in the middle but like regression in an unconstrained space you typically want to have
your points farther out so you have good leverage. From this perspective, the points on the
outside make a lot of sense. From an actual experimentation situation, you would have to be in a
scientific setting also where those points make sense. If not, we would constrain this region to
begin with. We will get in to this later.

How Rich of a Design?

Let's look at the set of possible designs that Minitab gives us.

Where it is labeled on the left Lattice 1, Lattice 2, etc., here minitab is referring to degree 1, 2, etc.
So, if you want a lattice of degree 1, this is not very interesting. This means that you just have a 0
and 1. If you go to a lattice of degree 2 then you need six points in three dimensions. This is
pretty much what we looked at previously... (roll over the red mixture points, below).

Here is a design table for a lattice with degree 3:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 23 of 29

Now let's go into Minitab and augment this design by including axial points. Here is what results:

This gives us three more points. Each of these points is 2/3, 1/6, 1/6. These are interior points.

These are good designs if you can run your experiment in the whole region.

Let's take a look at four dimensions and see what the program will do here. Here is a design with
four components, four dimensions, and a lattice of degree three. We have also selected to
augment this design with axial and center points.

This gives us 25 points in the design and the plot shows us the four faces of the tetrahedron. It
doesn't look like it is showing us a plot of the interior points.

11.3.3 - The Analysis of Mixture Designs

Example - Elongation of Yarn

Ex11.5.MTW [9] from the text.

This example has to do with the elogation of yarn based on

its component fabrics. There are three components in this
mixture and each component is a synthetic material. The
mixture design was one that we had looked at previously.
It is a simple lattice design of degree 2. This means that it
has mixtures of 0, 1/2, 100%. The components of this
design are made up of these three possibilities.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 24 of 29

In the Minitab program, the first 6 runs show you the pure components, and in addition you have
the 5 mixed components. All of this was replicated 3 times so that we have 15 runs. There were
three that had missing data.

You can also specify in more detail which type of points that you want to include in the mixture
design using the dialog boxes in Minitab if your experiment requires this.

Analysis

In the analysis we fit the quadratic model ( the linear + the interaction terms). Remember we only
have 6 points in this design, the vertex, the half-lengths, so we are fitting a response surface to
these 6 points. Let's take a look at the analysis:

Here we get 2 df linear, 3 df quaratic, these are the five regression parameters. If you look at the
individual coefficients, six of them because they are is no intercept, three linear and three cross
product terms... The 9 df for error are from the triple replicates and the double replicates. This is
pure error and there is no additional df for lack of fit in this full model.

If we look at the coutour service plot we get:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 25 of 29

We have the optimum somewhere between a mixture of A and C, with B essentially not
contributing very much at all. So, roughly 2/3rds C and 1/3 A is what we would like in our mixture.
Let's look at th optimizer to find the optimum values.

It looks like A = about .3 and B = about .7, with B not contributing nothing to the mixture.

Unless I see the plot how can I use the analysis output? How else can I determine the
appropriate levels?

Example - Gasoline Production

Pr11-31.MTW [10] from text

This example focuses on the production of an efficient

gasoline mixture. The response variable is miles per gallon
(mpg) as a function of the 3 components in the mixture.
The data set contains these 14 points - which has

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 26 of 29

duplicates at the centroid, labeled (1/3, 1/3, 1/3), and the

three vertices, labeled (1,0,0), (0,1,0), and (0,0,1).

This is a degree 2 design that has points at the vertices, middle of the edges, the center and axial
points, which are interior points, (2/3, 1/6, 1/6), (1/6, 2/3, 1/6) and (1/6, 1/6, 2/3). Also the design
includes replication at the vertices and the centroid.

If you analyze this dataset without having first generated the design in Minitab, you need to tell
Minitab some things about the data since you're importing it.

The model shows a linear term significant, the quadratic terms not significant, and the lack of fit,
( a total of 10 points and we are fitting a model sex parameters - 4 df), it shows that there is no
lack of fit from the model. It is not likely that it would make any difference.

If we look at the contour plot for this data:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 27 of 29

We can see that the optimum looks to be about 1/3, 2/3 between components A and B.
Component C does not play hardly any role at all. Next, let's look at the optimizer for this data
where we want to maximize a target of about 24.9.

And, again, we can see that component A at the optimal level is about 2/3rds and component B is
at about 1/3rd. Component C plays no part, as a matter of fact if we were to add it to the gasoline
mixture it would probably lower our miles per gallon average.

Let's go back to the model and take out the factors related to component C and see what
happens. When this occurs we get the following contour plot...

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 28 of 29

... and the following analysis:

Our linear terms are still significant, our lack of fit is still not significant. the analysis is saying that
linear is adequate for this situation and this set of data.

One says 1 ingredient and the other says a blend - which one should we use?

I would like look at the variance. ...

24.9 is the predicted value.

By having a smaller, more parimonious model you decrease the variance. This is what you would
expect with a model with fewer parameters. The standard error of the fit is a function of the
design, and for this reason, the fewer the parameters the smaller the variance. But is also a
function of residual error which gets smaller as we throw out terms that were not significant.

11.4 - Experiments with Computer Models

In many cases, performing actual experiments can be much too costly and cumbersome. Instead,
there might be a computer simulation of the system available which could be used as a means to
generate the response values at each design point -- as an proxy for the real system output.
Generally, there are two types of simulation models: Deterministic and Stochastic. Deterministic
simulation models are usually complex mathematical models which provide deterministic outputs
and not a random variable. The output from a stochastic simulation model is a random variable.
Normally, and for optimization purposes, a program of the simulation model is built (which is

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 29 of 29

called Metamodel) and based on the assumption that the simulation model is a true
representation of reality, the achieved optimum condition should be in compliance with the real
system. Research into optimal designs for complex models and optimal interpolation of the model
output have become hot areas of research in recent years. However, in this course we will not
cover any details about “experiments with computer models." More information can be found in
the text and of course the relative references.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/57

Links:
[1] https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson11/Ex-11-
1-output.doc
[2] https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/Ex11-1.mpj
[3]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson11/Ex11-2.MTW
[4]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson11/EX11-2.MPJ
[5]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson11/BD9-1.MTW
[6] https://screencast.com/t/yfyiPXdPq
[7] https://onlinecourses.science.psu.edu/stat503/javascript:popup_window
( '/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson11/L11_response_viewlet_swf.html',
'l11_response', 718, 668 );
[8]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson11/ch11.pdf
[9]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson11/Ex11-3.MTW
[10] https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson11/Pr11-
29.MTW

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/57/ 4/18/2019
 Page 1 of 3

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 12: Robust Parameter Designs

Lesson 12: Robust Parameter Designs

Introduction

In what we have discussed so far in the context of optimization only the average location
of the response variable has been taken into account. However, from another perspective
the variation of the response variable could be of major importance as well. This variation
could be due to either usual noise of the process or randomness in the nature of one or
more controllable factors of the process.

The Robust Parameter Design (RPD) approach initially proposed by Japanese engineer,
Genichi Taguchi, seeks a combination of controllable factors such that two main
objectives are achieved:

• The mean or average location of the response is at the desired level, and
• The variation or dispersion of the response is as small as possible.

Taguchi proposed that only some of the variables cause the variability of the process,
which he named noise variables or uncontrollable variables. Please note that noise
variables may be controllable in the laboratory, while in general they are a noise factor,
and uncontrollable. An important contribution of RPD efforts is to identify both the
controllable variables and the noise variables and find settings for the controllable variable
such that the variation of response due to noise factors is minimized.

The general ideas of Taguchi widely spread throughout the world; however, his philosophy
and methodology to handle RPD problems caused lots of controversy among statisticians.
With the emergence of Response Surface Methodology (RSM), many efficient
approaches were proposed which could nicely handle RPD problems. In what follows,
RSM approaches for Robust Parameter Design will be discussed.

Learning Objectives and Outcomes

• Understanding the general idea of Robust Parameter Design approaches

• Getting familiar with Taguchi’s crossed array design and its relative weaknesses
• Understanding combined array design and response model approach to RPD

12.1 - Crossed Array Design

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/74/ 4/18/2019
 Page 2 of 3

Crossed array design was originally propose by Taguchi. These designs consist of an
inner array and an outer array. The inner array consists of the controllable factors while
the outer array consists of the noise factors. The main feature of this design is that these
two arrays are “crossed”; that is, every treatment combination in the inner array is run in
combination with every treatment combination in the outer array. Table 12.2 is an example
of crossed array design, where the inner array consists of four controllable factors and
outer array consists of three noise factors. Note the typo in the levels of the 6th column of
data. It should be {+,-,+}.

Crossed array designs provide sufficient information about the interaction between
controllable factors and noise factors existing in the model which is an integral part of RPD
problems. However, it can be seen that crossed array design may result in a large number
of runs even for a fairly small number of controllable and noise factors. An alternative for
these designs are combined array designs which is discussed in the next section.

The dominant method used to analyze crossed array designs is to model the mean and
variance of the response variable separately, where the sample mean and variance can
be calculated for each treatment combination in the inner array across all combinations of
outer array factors. Consequently, these two new response variables can be considered
as a dual response problem where the response variance needs to be minimized while
response mean could be maximized, minimized or set close to a specified target. The text
book has an example about the leaf spring experiment in which the resulting dual
response problem has been solved by the overlaid contour plots method (See Figure
12.6) for multiple response problems, discussed in section 11.3.4.

12.2 - Combined Array Design

The combined array design approach treats all the variables the same, no matter they are
controllable or noise. These models are capable of modeling main effects of controllable
and noise factors and also their interactions. To illustrate, consider a case with two
controllable and one noise factor. Equation 12.1 in the textbook gives a first-order model
as:

y = β0 + β1 x1 + β2 x2 + β12 x1 x2 + γ1 z1 + δ11 x1 z1 + δ21 x2 z1 + ε

where the βi are coefficients of controllable factors, β12 is the coefficient of interaction of
controllable factors, γ1 is the coefficient of the noise factor and δij are the coefficients of
interaction between controllable and noise factors. As can be seen, the response model
approach puts all of the variables, no matter they are controllable or noise, in a single
experimental design. There exist some assumptions which are mentioned as follows:

• ε is a random variable with mean zero and variance σ2

• Noise factors are random variables (although controllable in experiment) with mean
zero and variance σz2
• If there exist several noise factors their covariance is zero

Under these general assumptions, we will find the mean and variance for the given
example, as following:
Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/74/ 4/18/2019
 Page 3 of 3

and
E(y) = β0 + β1 x1 + β2 x2 + β12 x1 x2
V ar(y) = σz2 (γ1 + δ11 x1 + δ21 x2 )2 + σ 2
Notice that although the variance model involves only controllable variables but it also
considers the interaction regression coefficients between the controllable and noise
factors.

Finally, as before, we perform the optimization using any dual response approach like
overlaid contours, desirability functions or etc. (Example 12.1 from the text book is a good
example of overlaid contour plots approach).

From the design point of view, using any resolution V (or higher) design for the two level
factor designs is efficient. Because these designs allow any main effect or two factor
interaction to be estimated separately, assuming that three and higher factor interactions
are negligible.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/74

Loading [MathJax]/extensions/MathZoom.js

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/74/ 4/18/2019
 Page 1 of 9

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 13: Experiments with Random Factors

Lesson 13: Experiments with Random Factors

Introduction
Throughout most parts of this course we have discussed experiments with fixed factors. That is, the levels used
for the factors are those of interest by the experimenter and the inference made is confined to those specific levels.
However, when factor levels are chosen at random from a larger population of potential levels, the factor is called a
random factor. In this case, the statistical inference applies to the whole population of levels. Random factor
models have many industrial applications including measurement system studies.

Learning Objectives & Outcomes

• Understanding the concept of random effect
• Getting familiar with random effect models and components of variance in each model
• Learning how to deal with models containing two random factors
• Getting familiar with how to analyze experiments where one of the factors is fixed and the other one is random
• Finding the expected mean squares using a simple algorithm

13.1 - Random Effects Models

Imagine that we randomly select a of the possible levels of the factor of interest. In this case, we say that the factor
is random. Typically random factors are categorical. While continuous covariates may be measured at random
levels, we usually think of the effects as being systematic (such as linear, quadratic or even exponential) effects.
Random effects are not systematic. The model helps make this clear.

As before, the usual single factor ANOVA applies which is

i = 1, 2, … , a
yij = μ + τi + εij {
j = 1, 2, … , n
However, herein, both the error term and treatment effects are random variables, that is

εij  is NID(0, σ 2 ) and τi is NID(0, στ2 )

Also, τi and εij are independent. The variances σ2τ and σ2 are called variance components.

There might be some confusion about the differences between noise factors and random factors. Noise factors
may be fixed or random. In Robust Parameter Designs we treat them as random because, although we control
them in our experiment, they are not controlled under the conditions under which our system will normally be run.
Factors are random when we think of them as a random sample from a larger population and their effect is not
systematic.

It is not always clear when the factor is random. For example, if a company is interested in the effects of
implementing a management policy at its stores and the experiment includes all 5 of its existing stores, it might
consider "store" to be a fixed factor, because the levels are not a random sample. But if the company has 100
stores and picks 5 for the experiment, or if the company is considering a rapid expansion and is planning to
implement the selected policy at the new locations as well, then "store" would be considered a random factor. We
seldom consider random factors in 2k or 3k designs because 2 or 3 levels are not sufficient for estimating
variances.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 2 of 9

In the fixed effect models we test the equality of the treatment means. However, this is no longer appropriate
because treatments are randomly selected and we are interested in the population of treatments rather than any
individual one. The appropriate hypothesis test for a random effect is:

H0 : στ2 = 0
H1 : στ2 > 0
The standard ANOVA partition of the total sum of squares still works; and leads to the usual ANOVA display.
However, as before, the form of the appropriate test statistic depends on the Expected Mean Squares. In this case,
the appropriate test statistic would be

F0 = M STreatments /M SE

which follows an F distribution with a-1 and N-a degrees of freedom. Furthermore, we are also interested in
estimating the variance components σ2τ and σ2. To do so, we use the analysis of variance method which
consists of equating the expected mean squares to their observed values.

^2 = M SE  and σ
σ ^2 + nσ
^2τ = M STreatments
M STreatment − M SE
^2τ =
σ
n
^2 = M SE
σ
Potential problem that may arise here is that the estimated treatment variance component may be negative. It such
a case, it is proposed to either consider zero in case of a negative estimate or use another method which always
results in a positive estimate. A negative estimate for the treatment variance component can also be viewed as a
evidence that the linear model in not appropriate, which suggests looking for a better one.

Example 3.11 from the text discusses a single random factor case about the difference of looms in a textile weaving
company. Four looms have been chosen randomly from a population of looms within a weaving shed and four
observations of fabric strength were made on each loom.The data obtained from the experiment are below.

Loom Obs 1 Obs 2 Obs 3 Obs 4 row sum

1 98 97 99 96 390
2 91 90 93 92 366
3 96 95 97 95 383
4 95 96 99 98 388

Here is the Minitab output for this example using Stat > ANOVA > Balanced ANOVA command.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 3 of 9

The interpretation made from the ANOVA table is as before. With the p-value equal to 0.000 it is obvious that the
looms in the plant are significantly different, or more accurately stated, the variance component among the looms is
significantly larger than zero. And confidence intervals can be found for the variance components. The 100(1-α)%
confidence interval for σ2 is

(N − a)MSE (N − a)MSE
≤ σ2 ≤
χ2α/2,N−a χ21−α/2,N−a

Confidence intervals for other variance components are provided in the textbook. It should be noted that a closed
form expression for the confidence interval on some parameters may not be obtained.

13.2 - Two Factor Factorial with Random Factors

Imagine that we have two factors, say A and B, that both have a large number of levels which are of interest. We
will choose a random levels of factor A and b random levels for factor B and n observations are made at each
treatment combination. The corresponding linear model for this case and the respective variance components
would be

⎧ i = 1, 2, … , a
⎪ 
yijk = μ + τi + βj + (τβ)ij + εijk ⎨  j = 1, 2, … , b
⎩ 
⎪ 
k = 1, 2, … , n
2 , V (ε ) = σ 2
V (τi ) = στ2 , V (βj ) = σβ2 , V [(τβ)ij ] = στβ ijk

2 + σ2
V (yijk ) = στ2 + σβ2 + στβ

Where τi , βj , (τβ)ij and εijk are all NID random variables with mean zero and variance as shown above. The
relevant hypotheses that we are interested in testing are:

H0 : στ2 = 0 H0 : σβ2 = 0 2 =0
H0 : στβ
H1 : στ2 > 0 H1 : σβ2 > 0 2 >0
H1 : στβ

The numerical calculations for the analysis of variance are exactly like in the fixed effect case. However, we state
once again, that to form the test statistics, the expected mean squares should be taken into account. We state the
expected mean squares (EMS) here and assuming the hypothesis is true, we form the F test statistics, so that
under that assumption, both the numerator and denominator of the F statistic have the same expectation. Note that
the test for the main effects are no longer what they were in the fixed factor situation.
MSA
E(M SA ) = σ 2 + nστβ
2 + bnσ 2 ⟹ F =
τ 0 MSAB

MSB
E(M SB ) = σ 2 + nστβ
2 + anσ 2 ⟹ F =
β 0 MSAB

MSAB
E(M SAB ) = σ 2 + nστβ
2 ⟹ F0 = MSE

E(M SE ) = σ 2
Furthermore, variance components can again be estimated using the analysis of variance method by equating the
expected mean squares to their observed values.
M SA −M SAB
^2τ =
σ bn

M SB −M SAB
^2β =
σ an

M SAB −M SE
^2τβ =
σ n
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 4 of 9

Example 13.2 in the textbook discusses a two-factor factorial with random effects on a measurement system
capability study. These studies are often called gauge
^ capability
σ = M SE studies or gauge repeatability and reproducibility
(R&R) studies. In this example three randomly selected operators are selected to measure twenty randomly
selected parts, each part twice. Data obtained from the experiment is shown in Table 13.3. The variance
components are 2

2 + σ2
σy2 = στ2 + σβ2 + στβ

Typically, σ2 is called gauge repeatability because it shows the variation of the same part measured by the same
operator and σ2β + σ2τβ

which reflects the variation resulting from operators is called gauge reproducibility. Table 13.4 shows the analysis
using Minitab’s Balanced ANOVA command.

Table 13.4 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th Edition)

As it can be seen, the only significant effect is part. Estimates for components of variance and expected mean
square for each term are given at the lower part of the table. Notice that the estimated variance for interaction term
part*Operator is negative. The fact that the p-value for the interaction term is large along with the negative estimate
of its variance is a good sign that the interaction term is actually zero. Therefore, we can proceed and fit a reduced
model without part*operator term. The analysis of variance for the reduced model can be found in Table 13.5.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 5 of 9

Table 13.5 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th Edition)

Since the interaction term is zero, both of the effects is tested against the error term. Estimates of the variance
component are given below at lower part of the table. Furthermore, as mentioned before, estimate of the variance
of the gauge can be achieved as

^2gauge = σ
σ ^2 + σ
^2β = 0.88 + 0.01 = 0.89

which is relatively small comparing to the variability of the product.

13.3 - The Two Factor Mixed Models

Next, consider the case that one of the factors is fixed, say A, and the other one (B) is a random factor. This case is
called the two-factor mixed model and the linear statistical model and respective components of variance is

⎧ i = 1, 2, … , a
⎪ 
yijk = μ + τi + βj + (τβ)ij + εijk ⎨  j = 1, 2, … , b
⎩ 
⎪ 
k = 1, 2, … , n
2 , V (ε ) = σ 2
V (βj ) = σβ2 , V [(τβ)ij ] = [(a − 1)/a]στβ ijk

a
∑ τi = 0
i=1

Here τi is a fixed effect but βj and (τβ)ij are assumed to be random effects and εijk is a random error. Furthermore, βj
and εijk are NID. The interaction effect is also normal but not independent. There often is a restriction imposed on
the interaction which is
a
∑(τβ)ij = (τβ).j = 0 j = 1, 2, … , b
i=1

Because of the sum of interaction effects over the levels of the fixed factor equals zero, this version of the mixed
model is called the restricted model. There exists another model which does not include such a restriction and is
discussed later. Neither of these models is "correct" or "wrong" - they are both theoretical models for how the data
behave. They have different implications for the meanings of the variance components. The restricted model is
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 6 of 9

often used in the ANOVA setting. The unrestricted model is often used for more general designs that include
continuous covariates and repeated or spatially correlated measurements.

Once again the tests of hypotheses for the mixed-model are:

H0 : τi = 0 H0 : σβ2 = 0 2 =0
H0 : στβ
H1 : τi ≠ 0 H1 : σβ2 > 0 2 >0
H1 : στβ

Furthermore, test statistics which are based on the expected mean squares are summarized as follows
a
bn ∑ τi2
MSA
E(M SA ) = σ 2 + nστβ
2 + i=1
a−1
⟹ F0 = MSAB

MSB
E(M SB ) = σ 2 + anσβ2 ⟹ F0 = MSE

MSAB
E(M SAB ) = σ 2 + nστβ
2 ⟹ F0 = MSE

E(M SE ) = σ 2
In the mixed model, it is possible to estimate the fixed factor effects as before which are shown here:

^ = ȳ ..
μ
τ^i = ȳ i.. − ȳ ...
The variance components can be estimated using the analysis of variance method by equating the expected mean
squares to their observed values:
M SB −M SE
^2β =
σ an

M SAB −M SE
^2τβ =
σ n

^2 = M SE
σ
Example 13.3 is the measurement system capability experiment where here we assume the operator has become a
fixed factor while part is left as a random factor. Assuming the restricted version of the mixed effect model,
Minitab’s balanced ANOVA routine output is given as follows.

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 7 of 9

Table 13.6 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th Edition)

Like before, there exists a large effect of parts, small operator effect and no part*operator interaction. Notice that
again the variance component estimate for the part*operator interaction is negative, which considering its
insignificant effect, leads us to assume it is zero and to delete this term from the model.

As mentioned before, there exist alternative analyses for the mixed effect models which are called the unrestricted
mixed models. The linear statistical model and components of variance for the unrestricted mixed model are given
as:

⎪ i = 1, 2, … , a
⎧ 
yijk = μ + αi + γj + (αγ)ij + εijk ⎨  j = 1, 2, … , b
⎩ 
⎪ 
k = 1, 2, … , n
2 , V (ε ) = σ 2
V (γj ) = σβ2 , V [(αγ)ij ] = σαγ ijk

a
∑ αi = 0
i=1

In the unrestricted mixed model, all of the random terms are assumed to be Normally and independently distributed
(NID) and there is not a restriction on the interaction term which was previously imposed. As before, the relevant
tests of hypotheses are given by:

H0 : αi = 0 H0 : σγ2 = 0 2 =0
H0 : σαγ
H1 : αi ≠ 0 H1 : σγ2 > 0 2 >0
H1 : σαγ
And the expected mean squares which determine the test statistics are
a
bn ∑ α2i
MSA
E(M SA ) = σ 2 + nσαγ
2 + i=1
a−1
⟹ F0 = MSAB
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 8 of 9

MSAB
E(M
E(MSSAB ) = σ 2 + nσ2 αγ
2
2 ⟹FF0 =
= MSB
B ) = σ + nσαγ + anσγ ⟹ 0 MSE
MSAB
E(M SE ) = σ 2
2
Again, to estimate the variance components, the analysis of variance method is used and the expected mean
squares are equated to their observed values which result in:
M SB −M SAB
^2γ =
σ an

M SAB −M SE
^2αγ =
σ n

^2 = M SE
σ
Example 13.4 uses the unrestricted mixed model to analyze the measurement systems capability experiment. The
Minitab solution for this unrestricted model is given here:

Table 13.7 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th Edition)

It is difficult to provide guidelines for when the restricted or unrestricted mixed model should be used, because
statisticians do not fully agree on this. Fortunately, the inference for the fixed effects does not differ for the 2 factor
mixed model which is most often seen, and is usually the same in more complicated models as well.

13.4 - Finding Expected Mean Squares

As we have demonstrated, determining the appropriate test statistics in the analysis of variance method depends
on finding the expected mean squares. In complex design situations and in the presence of random or mixed
models it is tedious to apply the expectation operator. Therefore, it would be helpful to have a formal procedure by
which we could derive the expected mean squares for the different terms in the model. Page 523 has listed a set of
rules which works for any set of balanced models to derive the expected mean squares. These rules are consistent
with the restricted mixed model and can be modified to incorporate the unrestricted model assumptions, as well.
Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 9 of 9

It is worth mentioning that the test statistic is a ratio of two mean squares where the expected value of the
numerator mean square differs from the expected value of the denominator mean square by the variance
component or the fixed factor in which we are interested. Therefore, under the assumption of the null hypothesis,
both the numerator and the denominator of the F ratio have the same EMS.

13.5 - Approximate F Tests

Sometimes in factorial experiments with three or more factors involving a random or mixed model, we determine
that there is no exact test statistic for certain effects in the model. Satterthwaite (1946) proposed a test procedure
which uses the linear combinations of the original mean squares to form the F-ratio. These linear combinations of
the original mean squares are sometimes called “synthetic” mean squares. Details on how to build the test statistic
and adjustments made to degrees of freedom based on Satterthwaite procedure can be found in Section 13.6.

Minitab will analyze these experiments and derive “synthetic” mean squares, although their “synthetic” mean
squares are not always the best choice. Approximate tests based on large samples (which use modified versions
of the Central Limit Theorem) are also available. Unfortunately, this is another case in which it is not clear that
there is a best method.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/82

Typesetting math: 100%

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/82/ 4/18/2019
 Page 1 of 9

Published on STAT 503 (https://onlinecourses.science.psu.edu/stat503)

Home > Lesson 14: Nested and Split Plot Designs

Lesson 14: Nested and Split Plot Designs

Introduction

Nested and Split Plot experiments are multifactor experiments that have some important industrial applications
although historically these come out of agricultural contexts. "Split plot" designs -- here we are originally talking about
fields which are divided into whole and split plots, and then individual plots get assigned different treatments. For
instance, one whole plot might have different irrigation techniques or fertilization strategies applied, or the soil might
be prepared in a different way. The whole plot serves as the experimental unit for this particular treatment. Then we
could divide each whole plot into sub plots, and each subplot is the experimental unit for another treatment factor.

Whenever we talk about split plot designs we focus on the experimental unit for a particular treatment factor.

Nested and split-plot designs frequently involve one or more random factors, so the methodology of Chapter 13 of
our text (expected mean squares, variance components) is important.

There are many variations of these designs – here we will only consider some of the more basic situations.

Learning Objectives & Outcomes

• Understanding the concept of nesting factors inside another factor.

• Getting familiar with the two-stage nested designs where either or both of the factors could be fixed or random.
• Getting familiar with split-plot designs and their applications where changing the level of some of the factors is
hard, relative to other factors.
• Understanding the two main approaches to analyze the split- plot designs and their derivatives and the basis for
each approach.
• Getting familiar with split-split-plot designs as an extension of split-plot designs.
• Getting familiar with strip- plot designs (or split-block designs) and their difference from the split-plot designs.

14.1 - The Two-Stage Nested Design

When factor B is nested in levels of factor A, the levels of the nested factor don't have exactly the same meaning
under each level of the main factor, in this case factor A. In a nested design, the levels of factor (B) are not identical
to each other at different levels of factor (A), although they might have the same labels. For example, if A is school
and B is teacher, teacher 1 will differ between the schools. This has to be kept in mind when trying to determine if
the design is crossed or nested. To be crossed, the same teacher needs to teach at all the schools.

As another example, consider a company that purchases material from three suppliers and the material comes in
batches. In this case, we might have 4 batches from each supplier, but the batches don't have the same
characteristics of quality when purchased from different suppliers. Therefore, the batches would be nested. When we
have a nested factor and you want to represent this in the model the identity of the batch always requires an index of
the factor in which it is nested. The linear statistical model for the two-stage nested design is:

⎧  i = 1, 2, … , a
⎪ 
yijk = μ + τi + βj(i) + εk(ij) ⎨  j = 1, 2, … , b
⎩ 
⎪ 
k = 1, 2, … , n
The subscript j(i) indicates that jth level of factor B is nested under the ith level of factor A. Furthermore, it is useful to
think of replicates as being nested under the treatment combinations; thus, k(ij) is used for the error term. Because
not every level of B appears with every level of A, there is no interaction between A and B. (In most of our designs,
the error is nested in the treatments, but we only use this notation for error when there are other nested factors in the
design).

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019
 Page 2 of 9

When B is a random factor nested in A, we think of it as the replicates for A. So whether factor A is a fixed or random
factor the error term for testing the hypothesis about A is based on the mean squares due to B(A) which is read "B
nested in A". Table 14.1 displays the expected mean squares in the two-stage nested design for different
combinations of factor A and B being fixed or random.

Table 14.1 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th and 8th Edition)

The analysis of variance table is shown in table 14.2.

Table 14.2 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th and 8th Edition)

Another way to think about this is to note that batch is the experimental unit for the factor 'supplier'. Does it matter
how many measurements you make on each batch? (Yes, this will improve your measurement precision on the
batch.) However, the variability among the batches from the supplier is the appropriate measure of the variability of
factor A, the suppliers.

Essentially the question that we want to answer is, "Is the purity of the material the same across suppliers?"

In this example the model assumes that the batches are random samples from each supplier, i.e. suppliers are fixed,
the batches are random, and the observations are random.

Experimental design: Select four batches at random from each of three suppliers. Make three purity determinations
from each batch. See the schematic representation of this design in Fig. 14-1.

Figure 14.1 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th and 8th Edition)

It is the average of the batches and the variability across the batches that are most important. When analyzing these
data, we want to decide which supplier should they use? This will depends on both the supplier mean and the
variability among batches?

Here is the design question: How many batches should you take and how many measurements should you make on
each batch? This will depend on the cost of performing a measurement versus the cost of getting another batch. If
measurements are expensive one could get many batches and just take a few measurements on each batch, or if it
is costly to get a new batch then you may want to spend more money taking many multiple measurements per batch.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019
 Page 3 of 9

At a minimum you need at least two measurements (n = 2) so that you can estimate the variability among your
measurements, σ2, and at least two batches per supplier (b = 2) so you can estimate the variability among batches,
σ2β. Some would say that you need at least three in order to be sure!

To repeat the design question: how large should b and n be, or, how many batches versus how many samples per
batch? This will be a function of the cost of taking a measurement and the cost of getting another batch. In order to
answer these questions you need to know these cost functions. It will also depend on the variance among batches
versus the variance of the measurements within batches.

Minitab can provides the estimates of these variance components.

Minitab General Linear Model (unlike SAS GLM), bases its F tests on what the expected mean squares determine is
the appropriate error. The program will tell us that when we test the hypothesis of no supplier effect, we should use
the variation among batches (since Batch is random) as the error for the test.

Run the example given in Minitab Example14-1.MPJ [1] to see the test statistic, which is distributed as
an F-distribution with 2 and 9 degrees of freedom.

Practical Interpretation – Example 14.1

There is no significant difference (p-value = 0.416) in purity among suppliers, but significant variation exists (p-value
= 0.017) in purity among batches (within suppliers)

What are the practical implications of this conclusion?

Examine the residual plots. The plot of residuals versus supplier is very important (why?)

An assumption in the Analysis of Variance is that the variances are all equal. The measurement error should not
depend on the batch means, i.e., the variation in measurement error is probably the same for a high-quality batch as
it is for low-quality batch. We also assume the variability among batches, σ2B, is the same for all suppliers. This is an
assumption that you will want to check! Because the whole reason one supplier might be better than another is
because they have lower variation among their batches. We always need to know what assumptions we are making
and whether they are true or not. It is often the most important thing to learn - when you learn there is a failed
assumption!

What if we had incorrectly analyzed this experiment as a crossed factorial rather than a nested design? The analysis
would be:

The inappropriate Analysis of variance for crossed effects is shown in Table 14.5.

Table 14.5 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th and 8th Edition)

This analysis indicates that batches differ significantly and that there is significant interaction between batch and
supplier. However, neither the main effect of Batch nor the interaction is meaningful, since batches are not the same
across suppliers. Note that the sum of the Batch and the S × B Sum of Squares and Degree of Freedom is the Batch
(Supplier) line in the correct Table.

For the model with the A factor also a random effect, analysis of variance method can be used to estimate all three
components of variance.

^2 = M SE
σ
M SB(A) − M SE
^2β =
σ
n

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019
 Page 4 of 9

And

M SA − M SB(A)
^2τ =
σ
bn

14.2 - The General m-Stage Nested Design

The results from the previous section can easily be generalized to the case of m completely nested factors. The text
book gives an example of a 3-stage nested design in which the effect of two formulations on the alloy harness is of
interest. To perform the experiment, three heats of each alloy formulation are prepared, two ingots are selected at
random from each heat, and two harness measurements are made on each ingot. Figure 14.5 shows the situation.

Figure 14.5 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th and 8th Edition)

The linear statistical model for the 3-stage nested design would be

⎧ 
⎪ i = 1, 2, … , a
⎪ 
⎪ j = 1, 2, … , b
yijk = μ + τi + βj(i) + γk(ij) + εl(ijk) ⎨ 
⎪ 
⎪ k = 1, 2, … , c
⎩ 
⎪ 
l = 1, 2, … , n
Where τi is the effect of the ith alloy formulation, βj(i) is the effect of the jth heat within the ith alloy, and γk(ij) is the effect
of the kth ingot within the jth heat and ith alloy and εl(ijk) is the usual NID error term. The calculation of the sum of
squares for the analysis of variance is shown in Table 14.8.

Table 14.8 (Design and Analysis of Experiments, Douglas C. Montgomery, 8th Edition) (Please note: the
Sum of Squares formulas for B(A) and C(B) have an error - they should have the A means and B means
subtracted, respectively, not the overall mean.)

To test the hypotheses and to form the test statistics once again we use the expected mean squares. Table 14.9
illustrates the calculated expected mean squares for a three-stage nested design with A and B fixed and C random.

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019
 Page 5 of 9

Table 14.9 (Design and Analysis of Experiments, Douglas C. Montgomery, 8th Edition)

14.3 - The Split-Plot Designs

Note: It is worth mentioning the fact that the notation used in this section (especially the use of Greek rather than
Latin letters for the random error terms in the linear models) is not our preference. But, we decided to keep the text
book’s notation to avoid any possible confusion.

There exist some situations in multifactor factorial experiments where the experimenter may not be able to randomize
the runs completely. Three good examples of split-plot designs can be found in the article: "How to Recognize a Split
Plot Experiment" [2] by Scott M. Kowalski and Kevin J. Potcner, Quality Progress, November 2003.

Another good example of such a case is in the text book in Section 14.4. The example is about a paper
manufacturer who wants to analyze the effect of three pulp preparation methods and four cooking temperatures on
the tensile strength of the paper. The experimenter wants to perform three replicates of this experiment on three
different days each consisting of 12 runs (3 × 4). The important issue here is the fact that making the pulp by any of
the methods is cumbersome. Thus method is a “hard to change” factor. It would be economical to randomly select
any of the preparation methods, make the blend and divide it into four samples and cook each of them with one of
the four cooking temperatures. Then the second method is used to prepare the pulp and so on. As we can see, in
order to achieve this economy in the process, there is a restriction on the randomization of the experimental runs.

In this example, each replicate or block is divided into three parts called whole plots (Each preparation method is
assigned to a whole plot). Next, each whole plot is divided into four samples which are split-plots and one
temperature level is assigned to each of these split-plots. It is important to note that since the whole-plot treatment in
the split-plot design is confounded with whole plots and the split-plot treatment is not confounded, if possible, it is
better to assign the factor we are most interested in to split plots.

Analysis of Split-Plot designs

In the statistical analysis of split-plot designs, we must take into account the presence of two different sizes of
experimental units used to test the effect of whole plot treatment and split-plot treatment. Factor A effects are
estimated using the whole plots and factor B and the A*B interaction effects are estimated using the split plots. Since
the size of whole plot and split plots are different, they have different precisions. Generally, there exist two main
approaches to analyze the split- plot designs and their derivatives.

1. First approach uses the Expected Mean Squares of the terms in the model to build the test statistics and is the
one discussed by the book. The major disadvantage to this approach is the fact that it does not consider the
randomization restrictions which may exist in any experiment.
2. Second approach which might be of more interest to statisticians and the one which considers any restriction in
randomization of the runs is considered as the tradition approach to the analysis of split-plot designs.

Both of the approaches will be discussed but there will be more emphasis on the second approach, as it is more
widely accepted for analysis of split-plot designs. It should be noted that the results from the two approaches may not
be much different.

The linear statistical model given in the text for the split-plot design is:

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019
 Page 6 of 9

i = 1, 2, … , r
yijk = μ + τi + βj + (τβ)ij + γk + (τγ)ik + (βγ)jk + (τβγ)ijk + εijk ⎨ j = 1, 2, … , a
⎩ 
Where, τi , βj and (τβ)ij represent the whole plot and γk, (τγ)ik, (βγ)jk and (τβγ)ijk represent⎪ 
the split-plot. Here τi , βj and
γk are block effect, factor A effect and factor B effect, respectively. The sums of squares⎧  ⎪  k the
for = 1, 2, …are
factors , b computed
as in the three-way analysis of variance without replication.

To analyze the treatment effects we first follow the approach discussed in the book. Table 14.17 shows the expected
mean squares used to construct test statistics for the case where replicates or blocks are random and whole plot
treatments and split-plot treatments are fixed factors.

Table 14.17 (Design and Analysis of Experiments, Douglas C. Montgomery, 8th Edition)

The analysis of variance for the tensile strength is shown in the Table 14.16.

Table 14.18 (Design and Analysis of Experiments, Douglas C. Montgomery, 8th Edition)

As mentioned earlier analysis of split-plot designs using the second approach is based mainly on the randomization
restrictions. Here, the whole plot section of the analysis of variance could be considered as a Randomized Complete
Block Design or RCBD with Method as our single factor (If we didn’t have the blocks, it could be considered as a
Complete Randomize Design or CRD). Remember how we dealt with these designs (Step back to Chapter 4). The
error term which we used to construct our test statistic (The sum of square of which was achieved by subtraction) is
just the interaction between our single factor and the Blocks. (If you recall, we mentioned that any interaction
between the Blocks and the treatment factor is considered part of the experimental error). Similarly, in the split-plot
section of the analysis of variance, all the interactions which include the Block term are pooled to form the error term
of the split-plot section. If we ignore method, we would have an RCBD where the blocks are the individual
preparations. However, there is a systematic effect due to method, which is taken out of the Block effect. Similarly,
the block by temperature has a systematic effect due to method*temperature, so a SS for this effect is removed from
the block*temperature interaction. SO, one way to think of the SP Error is that it is Block*Temp+Block*Method*Temp
with 2*3+2*2*3=18 d.f.

The Mean Square error terms derived in this fashion are then be used to build the F test statistics of each section of
ANOVA table, repectively. Below, we have implemented this second approach for data. To do so, we have first

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019
 Page 7 of 9

produced the ANOVA table using the GLM command in Minitab, assuming a full factorial design. Next, we have
pooled the sum of squares and their respective degrees of freedom to create the SP Error term as described.

As you can see, there is a little difference between the output of analysis of variance performed in this manner and
the one using the Expected Mean Squares because we have pooled Block*Temp and Blocks*Method*Temp to form
the subplot error.

Advantages and Disadvantages of Split-Plot Experiments

In summary, when one of the treatment factors needs more replication or experimental units (material) than another
or when it is hard to change the level of one of the factors, these design become important. The primary
disadvantage of these designs is the loss in precision in the whole plot treatment comparison and the statistical
complexity.

14.4 - The Split-Split-Plot Design

The restriction on randomization mentioned in the split-plot designs can be extended to more than one factor. For the
case where the restriction is on two factors the resulting design is called a split-split-plot design. These designs
usually have three different sizes or types of experimental units.

Example 14.4 of the text book (Design and Analysis of Experiments, Douglas C. Montgomery, 7th and 8th Edition)
discusses an experiment in which a researcher is interested in studying the effect of technicians, dosage strength
and wall thickness of the capsule on absorption time of a particular type of antibiotic. There are three technicians,
three dosage strengths and four capsule wall thicknesses resulting in 36 observations per replicate and the
experimenter wants to perform four replicates on different days. To do so, first, technicians are randomly assigned to
units of antibiotics which are the whole plots. Next, the three dosage strengths are randomly assigned to split-plots.
Finally, for each dosage strength, the capsules are created with different wall thicknesses, which is the split-split
factor and then tested in random order.

First notice the restrictions that exist on randomization. Here, we can not simply randomize the 36 runs in a single
block (or replicate) because we have our first hard to change factor, named Technician. Furthermore, even after
selecting a level for this hard to change factor (say technician 2) we can not randomize the 12 runs under this
technician because we have another hard to change factor, named dosage strength. After we select a random level
for this second factor, say dosage strength of level 3, we can then randomize the four runs under these two
combinations of two factors and randomly run the experiments for different wall thicknesses as our third factor.

The linear statistical model for the split-split-plot design would be:

yijkh = μ + τi + βj + (τβ)ij + γk + (τγ)ik + (βγ)jk + (τβγ)ijk + δh

⎧ 
⎪  i = 1, 2, … , r
⎪ 
⎪ j = 1, 2, … , a
+(τδ)ih + (βδ)jh + (τβδ)ijh + (γδ)kh + (τγδ)ikh + (βγδ)jkh + (τβγδ)ijkh + εijkh ⎨ 
⎪ 
⎪ k = 1, 2, … , b
⎩ 
⎪ 
h = 1, 2, … , c
Using the Expected Mean Square approach mentioned earlier for split-plot designs, we can proceed and analyze the
split-split-plot designs, as well. Based on Expected Mean Squares given in Table 14.25 to build test statistics
(assuming the block factor to be random and the other factors to be fixed), , and are whole plot, split-plot and split-
split-plot errors, respectively. Minitab handles this model exactly in this way by GLM. (This was Table 14.22 in the
7th edition. The 8th edition has only the factors and EMS without the list of subscripts.)

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019
 Page 8 of 9

Table 14.25 (Design and Analysis of Experiments, Douglas C. Montgomery, 8th Edition)

However, we can use the traditional split-plot approach and extend it to the case of split-split-plot designs as well.
Keep in mind, as mentioned earlier, we should pool all the interaction terms with the block factor into the error term
used to test for significance of the effects, in each section of the design, separately.

14.5 - The Strip-Plot Designs

These designs are also called Split-Block Designs. In the case where there are only two factors, Factor A is applied
to whole plots like the usual split-plot designs but factor B is also applied to strips which are actually a new set of
whole plots orthogonal to the original plots used for factor A. Figure 14.11from the 7th edition of the text is an
example of strip-plot design where both of the factors have three levels.

Figure 14.11 (Design and Analysis of Experiments, Douglas C. Montgomery, 7th Edition)

The linear statistical model for this two factor design is:

⎪ i = 1, 2, … , r
⎧ 
yijk = μ + τi + βj + (τβ)ij + γk + (τγ)ik + (βγ)jk + εijk ⎨ j = 1, 2, … , a
⎩ 
⎪ 
k = 1, 2, … , b

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019
 Page 9 of 9

Where, (τβ)ij , (τγ)ik and εijk are the errors used to test Factor A, Factor B and interaction AB, respectively.
Furthermore, Table 14.26 shows the analysis of variance assuming A and B to be fixed and blocks or replicates to be
random.

Table 14.26 (Design and Analysis of Experiments, Douglas C. Montgomery, 8th Edition)

It is important to note that the split-block design has three sizes of experimental units where the units for effects of
factor A and B are equal to whole plot of each factor and the experimental unit for interaction AB is a subplot which is
the intersection of the two whole plots. This results into three different experimental errors which we discussed
earlier.

Source URL: https://onlinecourses.science.psu.edu/stat503/node/68

Links:
[1] https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson14/Example14-1.MPJ
[2]
https://onlinecourses.science.psu.edu/stat503/sites/onlinecourses.science.psu.edu.stat503/files/lesson14/recognize_split_plot_experiment.pdf

https://newonlinecourses.science.psu.edu/stat503/print/book/export/html/68/ 4/18/2019