Design of Experiments

Prof. Romain Glèlè Kakaï

Dr. ir. Valère Salako

 Objectives

 Understand the terminologies commonly used in design of

experiments especially agricultural experiments

 Distinguish between the common experimental designs, their

principles, conditions of use and limitations

 Establish a design of experiment

 Conduct an experimentation
 CONTENTS
1. Basic concepts

2. Completely Randomised Designs (CRD)

3. Randomised Complete Blocks Designs (RCBD)

4. Split-Plot Designs (SPD)

5. Incomplete Blocks Designs (IBD)

4. Latin Square Designs (SLD)

 1. Basic Concepts
Experience
– «Test carried out to study a phenomenon »

– «in scientific language, a phenomenon caused or expected to test an

hypothesis, and thereby achieve a theoretical knowledge of how
things work»

– The concept of experience contrasts with that of surveys.

• Survey: we observe phenomena as they naturally occur without
influencing them
• Experience: we provoke a phenomenon of interest by
controlling underlying factors
Examples:
 Effect of the dose of compost (from waste) on tomato yield [Experience]
 Farmers' strategies to adapt to climate change [Survey]
4 4
 1. Basic Concepts
Factor :
group of objects of the same nature that could be
compared during experiments.
Example: group of maize varieties, provenances of teak; different
doses of a given mineral fertilizer, etc.
Qualitative Factor:
factor whose levels cannot be sorted a priori. Ex: varieties,
provenances of teak, diets, etc.
Example: comparison of the effect of 3 different treatments (A, B,
C) of organic fertilizer on yield of a given variety.

Quantitative Factor:
factor whose levels can be sorted before the experiment.
Example: different doses of a fertilizer, different temperature of
cooking, number of fermentation days of a product, etc.
 1. Basic Concepts
Levels or modalities of a factor:
Different objects of same nature associated with a factor.

In Ex1: A, B and C are levels of factor «fertilizer ».

Treatment or object :
Any level of a unique factor or any combination of modalities of 2
or more factors.
Ex1: Panicum C1 is a treatment in an experiment of comparison of
3 types of a grass.
Ex2: effect of dose of a fertlizer (30, 40) and varieties (v1, v2, v3)
on the yield of maize. One of the treatment is dose 30 kg/ha applied
to plants of variety v1.
 1. Basic Concepts
Experimental Unit:

basic object associated with each treatement of the experiment.

Examples: plot (agricultural experiment), plot (forestry), tree (production

parameters), animal (animal production), Petri dish (biotechnology).

Replication:
number of times a treatment is considered in an experiment.

Variable:

Basis of comparison of level of factor. Ex: yield, diameter of trees, basal

area, etc.
 1. Basic Concepts

Fixed Factor:

factor whose levels cannot change without changing the experiment or

without modifing the aim of the experiment. In other words, the levels
are controlled by the researcher and the results are used only for
conditions considered.
Example: 3 varieties of maize compared in an experiments. Factor
« variety » is fixed.

8 8
 1. Basic Concepts
Random Factor
Its levels can change without affecting the purpose of the experiment. In
other words, modalities of this factor are randomly selected from a large
number of modalities and the expected results from the experiments have to
be applied to all possible modalities of this factor.

Example: we would like to analyze the effect of 3 diets on the mean daily
weight gain of 4 differents groups of 5 rabbit. Two factors are the
considered: « Diets » and « group of 5 ». The First factor is fixed whereas
the second is random.

Nested Factor:
A factor is nested (to another) when its levels change according to the levels of
the main factor. In addition, a nested factor is always random.

9 9
 1. Basic Concepts
Control
A reference object in an experiment

Examples:
 In a comparison study of the agronomic performance of varieties in a region,
or a few widely used varieties (s) in the area concerned can be used as
control,

 In a comparison test the effects of different doses of fertilizer, parcels that

are not subject to any of the treatments studied (plots without fertilizer) can
be witnesses
 In a study evaluating the effect of pruning on reproductive performance, the
witness may be trees that have not been pruning

 In a different treatment effect evaluation study on seed germination rate, the

witness may be seeds having the subject of any treatment.
 1. Basic Concepts
Control

Avoid to include automatically a control in an experiment when it is not

essential to achieve the goal of the experiment, especially when it is
know a priori that differences with control are considerable.

Example:

In an experience for comparison of organic manures on very poor soils, it

may be omitted to consider plots without fertilizer. If necessary, we rather
introduce a reference fertilizer commonly used in the region, provided that such
fertilizer exists.

If the control is necessary because of extension or demonstration such plots

could be established in the margin or edge of the actual experience, without
necessarily devote the whole place due to each of the other treatments.
 1. Basic Concepts

Experimental design

How experimental units are assigned to treatments

in space or in time.

12 12
 1. Basic Concepts
Design of experiment: different steps of an experiment

• The definition of the goals and conditions of the experiment

• The definition of the factors considered in the experiment as well as
their levels/modalities,
• The definition of individuals or more generally, of the experimental
units,
• The definition of observations or variables to be measured
• How the different objects will be assigned to the different experimental
units, constituting the experimental design
• Methods for data analyses

13 13
 1. Basic Concepts
Design of experiment

Example

 Aim :
- compare 4 fungicides with a control (with no fungicide treatment), and
compare the fungicides on potato in small plots

 Objects :
– The four fungicides are labbelled 1, 2, 3 and 4 and the control is labelled
5.

 Experimental units
– Plot of 15m × 3m. The five objects are replicated four times, making a
total of 20 plots

 Observations / Variables to be measured

– Yield
– Growth (basal diameter, length)
 1. Basic Concepts
Design of experiment
Example

 Experimental design

Bloc 1 Bloc 2 Bloc 3 Bloc 4

3 2 1 4 5 1 4 5 3 2 2 3 5 1 4 5 4 3 1 2

Randomized complete blocks design

 Methods for data analyses

– Two-way anova (Fungicides, Blocks)
– Newman and Keuls test (to compare fungicides) + Dunnett test (to
compare fungicides to the control)
2. Completely randomised designs
Principles:
Random assignment of experimental units to different treatments
Example. Study of the effect of scarification and provenance of baobab
seeds on their germination rate.
2 levels of scarification: scarified seeds (S2) and natural seeds (S1); 2
provenances: sudanian (P1), sudano-guinean (P2). Three replications,
then 12 expérimental units (pots).

Random assignment of the 12 treatments of experimental units:

P1S1 P1S2 P1S2 P2S1 P2S2 P2S1 P2S2 P2S2 P1S1 P1S2 P2S1 P1S1
2. Completely randomised designs

Data analysis: ANOVA with 2 fixed factors

Provenance (P) p-1
Treatment (T) q-1
PxT (p-1)(q-1)
Error n-pq
Total n-1

Advantages: very simple design (no difficulty to establish)

disadvantages: Difficulty to control heterogeneity; often use in
laboratory essays.
2. Completely randomised designs
Case study: CRD to test growth of young plants of Parkia biglobosa
 3. Randomised complete
blocks designs

Principles: control of heterogeneity in one direction.

Experimental units here are distributed in homogeneous blocks.
The objective is to create groups of experimental units as similar as
possible to each other, so as to reduce, within blocks, the
heterogeneity of experimental conditions.

Forestry: comparison of effect of type of wind and humidity on the

yield of carbonisation: consideration of three villages as blocks.

Biotechnology: blocks can be here the plots on which the same tests
were performed.
 3. Randomised complete
blocks designs

Example: Effect of provenance and scarification of baobab seeds on

aerial biomass of two-months plants: consideration of a factor “block”
with factor 3 levels.

P2S1 P1S1 P2S2 P1S2 P1S1 P1S2 P2S2 P2S1 P1S2 P1S1 P2S2 P2S1

Block 1 Block 2 Block 3

Data analysis
Crossed mixte ANOVA with 3 factors.
 3. Randomised complete
blocks designs

Model of ANOVA with 3 or more factors.

SSV DF SSE MS F P
Provenance p-1
Scarification k-1
Provenance*Scarification (q-1)(k-1)
Blocs q-1
Provenance*Block (p-1)(q-1)
Scarification*Block (k-1)(q-1)
Error (q-1)(p-1)(k-1)
Total n-1
 3. Randomised complete
blocks designs

Advantages: Simple randomization of experimental units in blocks;

effectiveness of the design compare to the completely randomized design.

Disadvantages: difficulty of randomization of some factors levels: water,

fertilizer, seeding or planting, etc.
 3. Randomised complete
blocks designs
Case study 1: RCBD to analyze growth of young plants of T. indica
 3. Randomised complete
blocks designs
Case study 2: RCBD to analyze growth of seedlings of Artemisia annua
 4. Split-plot design
Principles
Let’s consider two factors with p and q levels and r blocks. We assign
the p levels of the first factor in the r blocks, having then pr plots. It then
distributes the q terms of the 2nd factor inside pr parcels constituting
pqr and subplots. Let’s assign now the q modalities of the 2nd factor
inside the pr plots, having then pqr sub-plots.

- 2 types of scarification: S1
and S2
A2S1 A2 A2S2 A1S2 A1 A1S1 Block1
- 2 levels of watering: water
(A1), stress (A2)
A2 A2S2 Block2
- 3 blocks. A1S2 A1 A1S1 A2S1

A2S1 A2 A2S2 A1S1 A1 A1S2 Block3

 4. Split-plot design
Data analysis
Source of variation Degre of freedom

Watering (A) p –1 = 1
Blocks r –1 = 2
Error1 (p -1)(r -1) = 2
Scarification (S) q -1 = 1
AxS (p –1)(q -1)=1
Error2 p(q –1)(r –1)=4
Total pqr -1 = 11

Advantages: Favor a factor relative to another;

Disadvantages: High loss of effectiveness of the second factor
related to the « whole plots ».
 4. Split-plot design
Case study 1: Split plot for analyzing effect of weed on rice plants
 4. Split-plot design
Case study 2: Split plot design for testing drought resistance in rice
 4. Split-plot design
Case study 3: Split plot design for analyzing pre-treatment effect on
baobab seed germination
 5. Incomplete Block
designs
Overview
In some situations, the randomized complete block design is not
suitable for the study of large number of objects (eg more than 20).
Problem of homogeneity of blocks.

It is therefore necessary to use incomplete block designs in which all the

objects are not present in each block.

Principles:
- each block has the same number k of treatments or experimental
units; this number is less than the number p of varieties (k<p);
- the p objects are replicated the same number r of times in the whole
design.

The number b of incomplete blocks that are considered is so that:

pr = bk
 5. Incomplete Block designs

Example: Incomplete blocks design for 30 varieties.

21 22 1 3 4 6 7 12 9 26 6 27 8 19 30 11 22 13 25 15 11 2 13 24 5 16 27 18 31 20

14 15 23 30 19 28 17 8 25 10 4 5 13 20 9 18 7 28 15 30 1 22 3 14 25 6 17 8 19 10

16 17 26 28 29 1 2 7 4 21 9 10 18 25 14 23 12 3 20 5 11 12 21 23 24 26 27 2 29 16

Number of incomplete blocks (b) = 9 ;

Number of treatments in each incomplete block (k)= 10;
Number of replications in each treatment (r) = 3;
Number of treatments (p) = 30.
 5. Incomplete Block designs

Example: Incomplete blocks designs for 30 varieties

Three incomplete blocks of 10 varieties each are constitued to obtain a
main complete block. Three main blocks named «replications» are
considered.

R1 21 22 1 3 4 6 7 12 9 26 14 15 23 30 19 28 17 8 25 10 11 2 13 24 5 16 27 18 31 20
10 10

R2 11 12 21 23 24 26 27 2 29 16 4 5 13 20 9 18 7 28 15 30 1 22 3 14 25 6 17 8 19 10
10 10

R3 16 17 26 28 29 1 2 7 4 21 9 10 18 25 14 23 12 3 20 5 6 27 8 19 30 11 22 13 25 15
10 10

Incomplete block1 Incomplete block2 Incomplete block3

 5. Incomplete Block designs
Data analysis
Mixte model ANOVA with two factors of classification.

SSV DF SSE MS F P
Treatments p-1=29
Blocks b-1=8
Error pr-p-b+1=52
Total pr-1=89
Particular case of incomplete blocks
Nested model of ANOVA with 3 factors.
SSV DF SSE MS F P
Treatments p-1=29
Replications r-1=2
IBlocks(Replications) b-r=6
Error pr-p-b+1=52
Total pr-1=89
 5. Incomplete Block designs

Avantages: - often used to overcome limitations due to space and

homogeneity
- increase the precision of the results when the number of
treatment is 20 or more.

Disadvantages: Non orthogonal ANOVA

An experiment is a question which science poses to

Nature, and a measurement is the recording of
Nature’s answer. MAX PLANCK, 1858-1947
 6. Latin square design
Principles

A design which consists of a number of plots (experimental unit) equal to a perfect

square (4, 9, 16, 25,...), which comprises as many rows of plot as columns of plots,
and within which each object is present one and only one time in each row and
column.
The rows and columns of a Latin square constitute, like a double network of
complete blocks perpendicular to each other, leading to a double control of the
heterogeneity of the experimental material. Col 1 Col 2 Col 3 Col 4

Example: row 1 V3 V1 V2 V4
Consider an experiment which aims
to compare agronomic performances
row 2 V1 V4 V3 V2
of four varieties of maize. (V1, V2, V3
and V4)

row 3 V4 V2 V1 V3

row 4 V2 V3 V4 V1
 6. Latin square design
Table: ANOVA table of an experiment of p×p

Source of variation DF p=4

Objects p-1 3
Rows p-1 3
Columns p-1 3
Residual variance (p-1)(p-2) 6
Total p2-1 15

Avantages: - often used to control bi-directional gradient of fertility

- More efficient than CRD and RCBD

Constraint: - Number of treatments (objects) should be equal to the

number of replications
 Home work
Group 1. 2k factorial designs

Group 2. Response Surface Design

Group 3. Mixture designs

Group 4. Augmented designs

Group 5. Balanced Incomplete blocks designs and

Lattice designs
 Home work
Your report should contain at least

1. Principles
2. Variants of the design
3. Methods for statistical analysis
4. Advantages and limits
5. Alternatives
6. Application in R
 Practice: field experiment
Species
1. Afzelia africana
2. Pterocarpus erinaceus
3. Bombax costatum
4. Diospyros mespiliformis

Experiment
1. Germination of four species in response to seed scarification
2. Germination of four species in response to soaking in hot water
3. Weed competitiveness of four tree species
4. Effect of substrates on seedlings growth and biomass allocation

Candidate designs
1. Completely Randomized Design
2. Randomized Complete Block Design
3. Split-plot Design
4. Latin Square Design