Introduction to Experimental Design

Alex Sanchez-Pla
Departamento de Genética, Microbiología y Estadística (UB)

Unidad de Estadística y Bioinformática (VHIR)

Versión 2021-11-17
Outline
1) Experiments: What, why, how

2) Principles of experimental design

3) Basic types of experimental designs

4) Some resources

2 / 48
Experiments: What, Why, How

3 / 48
Types of studies in medical research

source: Dtsch Arztebl Int 2009; 106(15): 262-8.

DOI: https://doi.org/10.3238/arztebl.2009.0262
4 / 48
Experimental Studies
Investigate changes in response variable,

by manipulating the levels of one or more treatment or independent variables.

Investigators play active role by (randomly) assigning each individual to each group.

5 / 48
Observational Studies
Investigate changes in response variable,

due to different levels of an independent variable,

Assignment of subjects into each other group is outside the control of the investigator
who can only observe the response variable in each group.

6 / 48
Why experiment?
A very common type of research study is experimental studies.

The purposes of the experimental studies are diverse. For example, we could propose
experiments to:

1. Compare responses to different treatments.

2. Determine the cause (s) of the response variation.
3. Find the conditions in which the optimal response is reached,
4. Develop a model to predict responses.

A characteristic of experimental versus observational studies is that, under the right

conditions, the former are the only ones that allow establishing causal relationships.

7 / 48
An example and basic ideas

Goal: investigate the effect of a drug to

prevent the development of diabetes

in genetically modified mice.

Effect will be tested and compared

with that of administering a placebo.
Effectiveness of the treatment may be
affected by the age of the animal, so it
is applied to mice "young" or "old"
Mice randomly selected within each
age group will receive either the drug
or a placebo (administered in the
same way as the drug).
To determine the effect of each
drug/age combination, weight, and
Source: EMBO Mol Med (2018) 10:e8791
decrease in age-related insulin
resistance will be determined. https://doi.org/10.15252/emmm.201708791

8 / 48
Some definitions
An experiment is any investigation in which a particular set of conditions is applied to
and the results of said experiment are observed and evaluated. app. For example, study
of the drug in pre-diabetic mice

Each group of experimental conditions is a treatment or factor. For example "Drug" and
"Age" are two factors.

Each particular condition within a factor is a level of that factor. For example
Drug/Placebo or Old/Young are two factors with two levels each.

The results that we observe after applying a treatment are the responses.

9 / 48
Some important definitions:
Experimental Unit (UE) The physical entity or subject exposed to the treatment
independently of other units.

Each mouse is an experimental unit

Unit of observation (UO) The unit in which they are carried out.
observations, that is,
measurements.

It can be a sample of the EU or be identical to the EU.

Experimental error: random variation observed between different repetitions carried

out, or not, under the same experimental conditions.

Observation error: The variation between multiple observations of the same

experimental unit.

10 / 48
Types of variability

Random variability
Differences expected to be observed
when different subjects from the same
sample are measured.
It is usually always present to a greater
or lesser degree.

It is usually reduced by increasing the

sample size

Systematic variability

Differences between subjects or

observations attributable to the
measurement process or to a non-
random selection of all the individuals
in the sample.
Usually can be corrected
11 / 48
What characterizes an experiment?

1. The treatments to be used

2. The experimental units to be used

3. The way treatment levels are assigned to the experimental units, that is, the
experimental design

4. Responses that are measured

12 / 48
What characterizes a good experimental
design?
It avoids biases or systematic errors

It allows a precise estimation of the response, which implies that the random error is
as low as possible.

It allows an adequate estimation of the error.

It has wide validity: the experimental units are a sample of the population in question,
so it is possible to extrapolate the conclusions of the sample to the population.

13 / 48
How to get a good design

Try to apply some ideas, basic and somewhat redundant, but which, together, guarantee
a good result.

1. Rely on a checklist of the experimental design.

2. Apply the scientific method in the appropriate study.

3. Be based on the basic principles of Experimental Design

Randomization,
Replication,
Local control.

And also
Plan design and analysis at the same time,
Involve your favorite statistician from the beginning (or before) of the experiment.

14 / 48
Design checklist
1. Define the objectives of the experiment.
2. Identify all possible sources of variation.
3. Select an appropriate experimental design.
4. Specify the experimental process
5. Conduct a pilot study
6. Specify the hypothesized model
7. Describe the tests to be performed.
8. Estimate the required sample size using the results.
of the pilot study
9. Review your decisions in Steps 1 through 8 and make the
necessary revisions.

15 / 48
Design, experimentation and analysis

16 / 48
Principles of experimental design

17 / 48
Basic principles of experimental design
Good experimental designs share common traits.

Apart from the fact that they are based on the logic of experimentation and the
scientific method,
they usually rely on some ideas, whose application guarantees good designs, or, in
any case, better designs than those studies in which they are not taken into
account explicitly.

Some of these principles are:

Randomization

Replication

Local control or blocking

18 / 48
1. Randomization
Since it is not possible to avoid random variations, we can randomly assign treatments
to units to try to compensate for the effect of such variation.

This can be done

Randomly assigning individuals to treatments and/or

Running the experiments in random order.

Randomizing does not mean doing everything at random.

Randomization is also important to ensure the validity of statistical procedures.

Randomization helps preventing a preferential assignment of certain to certain

experimental units.

For example, a doctor may be "tempted" to give the drug she thinks works best to
patients with the worst prognosis.

19 / 48
2. Replication
There is general agreement on the need to apply each treatment independently to
several experimental units.

This ...

Helps to establish the reproducibility of the results.

Protects against eventual abnormal or unusual results.
Provides a way to estimate the variance of the experimental error in the absence of
systematic differences between the experimental units.
Provides the ability to increase the precision of estimates of means of treatment.

Replication is importanty but, it does not necessarily guarantee valid estimates of

experimental error or, what is more:

Having the appropriate sample size does not guarantee the presence of an effect: The -
often heard sentence- "we didn't detect any effect but if we can collect enough samples
the effect will be seen" can be considered a Statistical Myth.

20 / 48
Replicates, power and precision
The number of repetitions r is directly related to the precision of the experiment
Variability is inversely related with the precision of the experiment.

¯
¯¯¯¯ 2
1/var(X ) = r/σ (∗)

While this is stated for estimating the sample mean, the rule can be easily extended to
other characterristics.

From (*) it follows that:

the greater the number of replicatess, r,

and the lower the variability, σ 2 ,
the greater the precision a design provides.

From this relation, it is straightforward to derive formulae for the sample size needed
for estimation.

21 / 48
How many replicates are needed?
If the goal of an experiment is, not only estimating one characteristic, but also
comparing groups, that is detecting the effect of a treatment, this can also be accounted
for.

Given the relation between:

(1) The variability

(2) The desired effect size,
(3) The level of significance ("alpha", type I error) of the test,

(4) The power (1-"beta", type II error) wished to attain

One can compute the sample size needed given the previous four values or,

One can fix any four and compute the other one (for instance the power given a
sample size, etc.)

Sample size online calculator example: [granmo]

(https://www.imim.es/ofertadeserveis/software-public/granmo/)

22 / 48
Technical and biological replicates

for
Technical replications allow quantifying
variability associated with the technique used.

for
Biological replications allow quantifying the
variability associated with the study population.

The total variability can be decomposed into various components of the variance.

2 2 2 2
σ(T OT AL) = σ(T EC) + σ(BI O) + σ(ERR)

In general :

2 2
σ(T EC) < σ(BI O)

source: https://www.licor.com/bio/blog/technical-and-biological-replicates

23 / 48
Replicates or pools?
Sometimes it may be decided to combine mRNA from different samples to form a
"pooled sample" or pool

This can be done because ...

Each separate sample does not provide enough mRNA

You want to compensate for excess variability by "averaging" similar samples.

This can be misleading, but correct if done apprpriately:

Combining several samples in each group but ...

Using several groups of different samples

What not to do:

Don't use groups when individual information is important (e.g. paired designs).
A sample with 3 grouped individuals is not the same as 3 individual samples!

24 / 48
3. Local control
In many situations it is common for not all samples to be homogeneous.

For example, in an experiment to compare two treatments using expression microarrays,

there may be different types of subjects:

Male or female animals

Animals from several litters (those from the same litter look more like each other)
Samples processed in different days due to the capacity of the equipment
and other sources of known but unavoidable variability.

If there are systematic differences between groups of samples ("blocks") the effects of
interest (for example the effect of a treatment) can be affected by differences between
samples of different blocks,

In other words, it may not be clear if the differences observed are attributable to the
effect of the treatment or other factors that we call confounding.

Local control or blocking, that is distributing each treatment evenly among the different
blocks is the way to minimize this undesired effect.

25 / 48
How to apply local control

This design does not apply good local This design applies good local control.
control.
The possible effect of sex or of the
Treatment effect can be confused with production batch is distributed among
the different levels of treatment, which
the effect of age or will allow them to be analyzed
that of the production batch. separately.

26 / 48
Batch effect and its adjustment

A Principal Component Analysis (PCA) can reveal the presence of undetected blocks in
the design.

If the different levels of the "batch" are distributed among the treatment levels, it is
possible to correct them.

If there is confusion between the two (for example, each treatment has been done
in a distinct batch), the effects cannot be separated. 27 / 48
Basic types of experimental designs

28 / 48
Experimental designs
A key point in any experiment is the way in which
the experimental units are assigned to
the treatments.

This assignment must be done in such a way that:

it is possible to estimate the effects that interest the researcher

the random variability is as small as possible ("maximum precision") with the

available resources.

the best possible local control is achieved, given the circumstances of the
experiment.

To achieve the best possible design, we will take into account the components that
define each design.

29 / 48
Design components
When considering the choice of a design for an experiment we must take into account:

The design of the treatments.

Which and how many treatments are included in the study?

Are they considered separately or in combination?
What are the levels of each treatment?

The error control design.

How are treatments assigned to experimental units ?.

This depends on the resources, the available units, the desired precision, the
heterogeneity between UEs.

The Observational Design

At what level are the observations made?

Is each EU an OU or are there several OUs per EU (subsampling)?

30 / 48
From components to design
Treatments Design Error control D. Observational Design

Completely 1 factor, (k niv.), r

Assign trat. 1 ... k UE 1 UU = 1 UO
Randomized replicates/level

Completely
1 factor (k level), 1 block (l Assign trats. 1 ... k to EU, in
randomized block 1 EU = 1 OU
level), k l EU each block
design

Two-factor design 2 factors (k, l levels.), r Assign each combination 1

1 EU = 1 OU
with interaction replicates of each combination ... k l of treatments to EU

1 Trat. (k niv), l repeated

Repeated Assign trats. 1 ... k to For each UE there are 'l' OU
measures, r
Measures Design individuals at time 1 (temporal measurements)
replicates/combination

31 / 48
Designs, Models and Analysis
The design of the treatments and the observational design help us to choose the
appropriate design for an experiment.

Each design can be represented by a linear model. This model:

Represents the relationships between responses, treatments and experimental and

observational units.
Is the basis for the analysis of the data once it has been collected.

Error control design defines how the randomization is carried out, that is, the
assignment of individuals to the treatments.

This should be done when planning the investigation.

32 / 48
Experimental design and ANOVA
Sometimes the design of the experiment is confused with its analysis, which is carried
out using Analysis of the Variance techniques.

This is understandable, because when one defines the experimental design the way
it will be analyzed is set. That is they are related, but they are not the same.
It is a common problem, in some books or statistics courses, which do not pay
attention to how treatments were allocated between individuals and provide the
data already collected.
This makes it difficult for students to realize that experimental design had been
carried out before the data were collected.

Summarizing: Although the treatment design suggests a certain analysis model the
experimental design should not be confused with the analysis of the data collected in
the experiment!

33 / 48
Experimental design and ANOVA
Treatments Design Error control D. Analysis

Completely 1 factor, (k niv.), r

Assign trat. 1 ... k UE 1-way ANOVA
Randomized replicates/level

Completely 2-way ANOVA

1 factor (k level), 1 block (l Assign trats. 1 ... k to EU,
randomized without
level), k * l EU in each block
block design interaction

Two-factor 2 factors (k, l levels.), r Assign each combination 3-way ANOVA

design with replicates of each 1 ... k * l of treatments to with
interaction combination EU interaction

1 Trat. (k niv), l repeated ANOVA of

Repeated Assign trats. 1 ... k to
measures, r repeated
Measures Design individuals at time 1
replicates/combination measures

34 / 48
Completely randomized design
Gene therapy experiment: compare four techniques to correct faulty genes

A: Normal gene inserted in a nonspecific location.

B: Abnormal gene exchanged for a normal gene.

C: Abnormal gene repaired by selective reversion mutation.

D: Regulation of a particular altered gene.

20 genetically identical and modified mice, affected by the disease to be treated, are
selected.

Treatments are randomized between the mice.

The response variable is gene expression.

35 / 48
Completely randomized design
The simplest design, suitable for comparing several treatments on a homogeneous
sample.

Randomization is performed by randomly assigning each of the 1 .... k treatments to

individuals out of a total of N = k * r

The basic linear model for one-factor experiments is as follows:

Yij = μi + eij = μ + τi + eij , i = 1 … k, j = 1 … r.

The analysis will usually be carried out by means of a one-way analysis of variance
(ANOVA).

36 / 48
Randomization in a DCA
There are many libraries that allow randomization, but it can also be done easily with a
small script.

Randomization is carried out before the experiment and it only indicates which
treatment will receive each experimental unit

Once the experiment is carried out, it is usual to present the data ordered by the
treatments received, which eliminates the evidence that the assignment has been made
randomly.

37 / 48
Random block design
After exposure to a poison, cells can be treated by different substances that accelerate
regeneration.

A study wants to compare six of these growth factors (5 are treatments and 1 is a
control).

A problem has caused that there is not enough culture medium to grow all treatments
with replicates. Instead, there are 4 culture media available.

Since a complete randomization is not possible, it is decided to block by type of culture

medium.

We prepare 4 groups of 6 plates, each group of a type of culture

Within each group a different treatment is randomly assigned to each of the six
plates.

38 / 48
Random block design
The completely randomized design loses utility if the experimental material is not
homogeneous.

In these cases, we can apply local control (blocking) and divide the experimental
material in homogeneous subgroups, which we will call blocks.

Once the samples have been distributed among the blocks, the treatments are applied
to the experimental units randomly and independently of the other blocks.

This design is called Random Block Design (RBD).

The linear model that describes the experiment is the following:

Yij = μ + ρi + τj + eij , i = 1 … k, j = 1 … l.

The analysis will usually be carried out by means of an analysis of variance (ANOVA) of
two factors without interaction.
Obviously, if it is not possible to distribute the samples evenly between the blocks, the
situation becomes complicated and we are faced with unbalanced designs

39 / 48
Block or randomize?
Block what you can and randomize what you cannot - Box, Hunter & Hunter (1978)

Randomization provides a rough balance between variables that have not been taken
into account.

Local control eliminates the effect of differences between blocks, thereby ensuring that
differences between treatments cannot be due to differences between blocks.

40 / 48
Factorial design
A study was conducted to study the effect of a drug and a diet on systolic blood
pressure.

20 people with high blood pressure were randomized to one of four treatment
conditions.

Control group (neither diet nor drug modification)

Diet modification only
Drug only
Modification of both drugs and diet

At the end of the treatment period, systolic blood pressure was assessed.

It is a factorial design in which each of the two treatments (drug, diet) can be randomly
assigned to each individual.

By having 20 individuals, there can be replicates of each treatment combination.

41 / 48
Factorial design
This design is useful to study the effects of several factors simultaneously.

The "treatments" are all combinations of the different factors under study.

Randomization is similar to the completely randomized design, that is, each

combination of treatments is randomly assigned to independent r EUs.

The fact that each combination is replicated makes it possible to study, not only the
effects of each factor separately, but also the interaction between them.

The linear model that describes a two-factor design with interaction with t and s levels
and r replicates respectively is the following:

Yijk = μ + ρi + τj + τ ρij + eijk , i = 1 … t, j = 1 … l, k = 1 … r.

The analysis will usually be carried out by means of an analysis of variance (ANOVA) of
two factors with interaction.

42 / 48
Repeated measures design
A study wanted to measure the concentration of certain metabolites in plasma after two
dietary interventions consisting of adding an amount of olive oil or an equivalent
amount of walnuts to the standard diet.

21 mice subjected to the same diet were taken and an intervention (water, olive oil or
nuts) was randomly assigned.

The concentration of the metabolite in blood was measured after before the
intervention and at 24h, 48h and one week.

43 / 48
Repeated measures design
When we take more than one measurement in each experimental unit, we have a
within-subjects design.

In this case, the data have different characteristics from the previous ones.

The measurements taken on the same individual are correlated.

There is a new source of variation that must be taken into account: variability
within subjects.

Apart from this, they offer the same possibilities as with other designs, but with an
additional source of variability, "time".

The analysis of repeated measures data is a whole world. Although the ANOVA of
repeated measures is traditionally used, the current trend is to perform the analyzes
using linear mixed models which are much more flexible.

44 / 48
Summarizing ...
A good experimental design is essential to carry out good experiments.

Experimental design means planning in advance, that is, before and not after the
experiment.

The experimental design must consider all steps: from sampling to data analysis.

Applying grounded principles such as randomization, replication and local control is key
to obtain good experimental designs.

The analysis of designed experiments is carried out with the Analysis of the Variance
(ANOVA). While each design can be asociated with an ANOVA model they should not be
confused.

Whenever possible we should have statistical support from the beginning of the study

45 / 48
And, as the master said ...

46 / 48
References and resources

47 / 48
References and resources
3rs-reduction.co.uk

A short and interactive introductory course on the design of experiments focused

on the benefits derived from an adequate design for the reduction of suffering in
experimental animals.

The "Statistical Analysis" section takes a brief tour of some experimental design
models and their analysis.

[A First Course in Design and Analysis of Experiments]

(http://users.stat.umn.edu/~gary/Book.html)

A book for an introductory course to design of experiments that, after being sold
out in bookstores, the author decided to provide freely on the internet.

It takes a "traditional" approach to the subject and contains aspects that today
would be approached differently, but it continues to be very interesting.

The author updated the examples by implementing them in R.

48 / 48