University of Queensland MERCI Statistical Training Series

Experimental design for agricultural field experiments

Alison Kelly Clayton Forknall Bethany Rognoni

June 26, 2023
 Introduction to experimental design
Background

I We will begin with a discussion of experimental design principles.

I Then we will install the ODApp on your computers and generate

some designs.

I Example 1 - Equally replicated row-column designs, eg NVT trials

I Example 2 - Unequally replicated designs, eg combined PVT and
NVT entries in one trial

I Example 3 - Partially replicated designs, eg ON trials with large

numbers and limited seed
 Key design principles

I Choice of treatments and measurement

I Choice of trial location - The target population of environments

(TPE)

I Denition of experimental unit and observational unit

I Randomisation and replication

I Control of variation through blocking structures

 The aim of the experiment
Choice of treatments

I Before beginning any experiment it is important to be clear about

the research aims.

I A clearly dened research question, or scientic hypothesis, should

be the basis of all comparative experiments.

I This then provides the framework for hypothesis testing in a

statistical analysis of the data arising from the experiment.

I The aims of the experiment clearly dene the type of treatments

that will be applied in the experiment.

In a plant breeding trial, the aim is usually to compare and select the
best variety.
 The aim of the experiment

Sometimes you may also wish to run experiments for other reasons.

I Selection for pathogen resistance, and does this relationship vary

between varieties?
I Selection for grain quality traits . For example,
I Soaking time in chickpea
I Cooking time in common bean
I Flour yield of grain in bread wheat
 Choice of treatments

The treatments are the things that will be manipulated in the

experiment. For example,

I A number of varieties are grown in trials across dierent years and

locations to expose varieties to dierent environments and measure
GxE.

I A number of varieties will be exposed to a range of plant spacing

treatments, and the resulting yield of plants will be measured.

I A new variety will be compared to a control or standard variety, and

the response will be level of disease infection.
 Choice of treatments

Each study should also consider the need for a control in the
experimental treatments.

I In eld trials measuring yield, control varieties or Checks are included

to demonstrate that new varieties are superior to existing varieties.

I In dedicated diseaese nurseries, susceptible and resistant Checks are

included to demonstrate that new varieties are superior to existing
varieties.
 The frame of reference for the study
Choice of experimenal material

A basic component of any experiment is the choice of experimental

material to be used in the study. In agricultural experiments this can be:

I An area of land which is typically divided into smaller units, or plots

of land.

I A set of pots to be grown in controlled glasshouse conditions,

comprising the characteristics of pot size, growing medium and pot
placement on benches in the glasshouse.
 The reference frame of the study
Choice of experimental material

The experimental material represents a sample from the population of all

possible experimental conditions. Welham et al. (2015) clearly articulates
that

it is important to recognise the frame of reference implied by the choice

of experimental units, so that appropriate conclusions can be drawn from
the results".

I The choice of trial location links in with the concept of the Target
Population of Environments.

I The choice and management of trial site should be representative of

farmers elds.
 The experimental unit

The experimental unit is the smallest unit to which a treatment can be

applied, such that any two units receive dierent treatments. Let's
consider the examples given above.

I One variety will be planted on one small plot of land. Hence variety
is the treatment and plot is the experimental unit.

I Three plants of each variety will be grown in one pot in the

glasshouse. Hence, variety is the treatment and pot is the
experimental unit.
 The observational unit

The observational unit is the smallest unit on which a measurement is

taken. It may be the same as the experimental unit, or on a ner scale
than the experimental unit.

I Variety is the treatment and plot is the experimental unit. If yield is

measured for the whole plot, then plot is both the experimental and
the observational unit.

I Variety is the treatment and pot is the experimental unit. If disease

measurements are taken on each plant, then plant is the
observational unit, but plot is the experimental unit.
 Experimental design process

Experimental design is simply the process of allocating treatments to

experimental units.

I Experimental material may be variable, and this causes problems for

fair experimental comparisons.

I Variability may arise from non-homogeneous experimental material.

I Variability may arise from experimental conduct, for example

planting and harvesting.

I Variability may arise from the measurement process.

 Barley uniformity trial - Hermitage Research Station, Qld
A motivating example

Consider the small eld plots on which most agricultural experiments are
conducted in the eld.

I One barley variety grown in a 50m by 50m area

I harvested in small rectangular plots of 5m by 1.75m,

I rectangular layout of 9 columns × 26 rows

 Barley uniformity trial - Hermitage Research Station, QLD
Yield (t/ha) of one variety

The phenotyping curse is that of plot-to-plot variation in eld trials

26
25
24
23
22
21
20
19
18 yield (t/ha)
17
16 2.0
15
Row

14 2.5
13
12
11 3.0
10
9 3.5
8
7
6
5
4
3
2
1
1 2 3 4 5 6 7 8 9
Column
 Barley uniformity trial - Hermitage Research Station, Qld
A motivating example

Suppose that we want to design a trial testing 3 Replicates of 15

Genotypes.

I Dene the row and column dimensions for this trial

I Plots = 3 × 15 = 45

I We could choose a trial layout with 3 columns × 15 rows

I We could choose a trial layout with 5 columns × 9 rows

I The options for row and column numbers depend on the factors of
the total number of plots
 Barley uniformity trial - Hermitage Research Station, QLD
Yield (t/ha) of one variety

26 26
25 25
24 24
23 23
22 22
21 21
20 20
19 19
18 18
17 yield (t/ha) 17 yield (t/ha)
16 16
15 2.0 15 2.0
Row

Row
14 2.5 14 2.5
13 13
12 3.0 12 3.0
11 11
10 3.5 10 3.5
9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
Column Column
 Barley uniformity trial - Hermitage Research Station, Qld
A motivating example

We can choose to place the trial at dierent locations in the eld.

I The trial results will vary according to where the trial is grown
I Consider 24 plots from the centre bottom of the eld.
I Layout 1 below has a trial mean yield of 3.12 t/ha
I Consider 24 plots from the bottom right hand corner of the eld.
I Layout 2 below has a trial mean yield of 2.64 t/ha
 Barley uniformity trial - Hermitage Research Station, QLD
Yield (t/ha) of one variety

26 26
25 25
24 24
23 23
22 22
21 21
20 20
19 19
18 18
17 yield (t/ha) 17 yield (t/ha)
16 16
15 2.0 15 2.0
Row

Row
14 2.5 14 2.5
13 13
12 3.0 12 3.0
11 11
10 3.5 10 3.5
9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
Column Column
 The experimental process

How do we ensure that comparisons are fair?

 The design process
Randomisation

Why Random?

I Each experiment is only one realisation of all possible sets of

occasions and conditions under which the experiment could be
undertaken.

I It is this non-repeatablity of the experiment that gives the plot

eects their randomness, (Bailey, 2008).

I It is the randomness of plots, and that fact that we are obtaining

only a small sample from all possible combinations that leads to the
need for replication and randomisation in the experimental design.

I In our example, we would obtain dierent treatment means,

depending on the plots where the treatments were grown.
 Experimental design
The role of randomisation

Why?
I Randomisation ensures that we avoid biassed or unrepresentative
results from the experiment.
I It is the insurance policy against any unknown systematic variation
in the experimental units. Bailey (2008) lists four types of potential
biasses that may occur due to lack of randomisation.
I systematic bias: for example measuring all plots of variety A at one
end of the paddock, and all plots of variety B at the other end.
I selection bias: choosing the unhealthiest plants to sample from the
high level of disease inoculation plot.
I accidental bias: allocating treatments based on the order in which
cattle may run through a race for sorting.
I deliberate bias (cheating): this may be common in medical research
where an experimenter may be tempted to place the sickest patients
into the treatment allocation which gives them the best chance of
recovery.
 Experimental design
The role of randomisation

How?
I The process of randomisation must be objective, so that there is no
possibility of intentional or accidental bias occurring.

I Computer generation through design packages is one of the most

common options available.

I Alternately, simple processes can also be used and can be as

eective as random number generators used in computer software.

I BUT, remember it must be random!!!

I AND, it must be a dierent randomisation for each experiment of
the same treatments (genotypes)!!!
 Experimental design
The role of replication

Why replicate?

I Replication of the experimental units is an essential element of

experimental design, as it ensures two important outcomes
surrounding the experimental results.

I Firstly, it increases the accuracy of the estimate of treatment eects,

as each treatment eect is averaged across more than one
experimental unit.

I Secondly, the variation between experimental units containing the

same treatment gives a measure of background variability,

I This can be used to assess the condence we place in the treatment

dierences measured in the experiment.
 Control of systematic variation
The role of blocking

Knowledge of variation inherent in the experimental units is essential for

the conduct of robust research. Variation can be classied as

I Systematic due to application of treatments to the experimental

units.

I Random due to uctuations in the experimental units.

I Systematic due to dierences between the experimental units or

induced by the experimental process. This type of variation can be
disastrous for the experimental process, if it is not accounted for
(Cobb, 2014)
 Control of systematic variation
The role of blocking

I Often the experimental units can be grouped into sets of units that
are more alike, and this principle is known as blocking.

I Blocking is commonly used to design experiments with either known

heterogeneity in the set of experimental units, or expected
heterogeneity imposed through operational procedures on the
experimental units.

I For example, blocking may be dened by batches or days within a

laboratory process, by groups of experimental units assessed by
dierent operators, or by plots in the eld with dierent underlying
fertility.
 Control of systematic variaiton
The role of blocking

I If blocking is implemented so that it groups like plots together, then

it generally increases the precision of the experiment, as it lessens
the unexplained (random) variation due to plots.

I Furthermore, it generally increases the accuracy of the experiment,

as it can ensure that systematic variation may be removed from
treatment comparisons.

I Blocking may or may not be eective in practice depending on the

homogeneity of plots within the blocks, and the heterogeneity
between/across blocks.

I Often blocking is adopted as an insurance policy in eld

experimentation, as the need to manage subsets of plots in the
experiment may arise.
 The barley uniformity trial - Warwick, QLD
Yield (t/ha) of one variety

The trial can be blocked into three replicate groups in dierent ways

24 24
23 23
22 22
21 21 Replicate 3
20 20
19 19
18 18
17 17
16 16
15 yield 15 yield
14 3.5 14 3.5
13 Replicate 1 Replicate 2 Replicate 3 13 Replicate 2
Row

Row
3.0 3.0
12 12
2.5 2.5
11 11
10 2.0 10 2.0
9 9
8 8
7 7
6 6
5 5 Replicate 1
4 4
3 3
2 2
1 1

1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
Column Column
 The barley uniformity trial - Warwick, QLD
Yield (t/ha) of one variety
3.5

3.5
3.0

3.0
●
2.5

2.5
2.0

2.0
1 2 3 1 2 3

Figure: Yield in tonnes/ha of a barley uniformity trial of a single variety

conducted at the Hermitage Research Station, Warwick. The trial can be
blocked into three replicate groups in dierent ways.
 The barley uniformity trial - Warwick, QLD
Yield (t/ha) of one variety

Field trend - Replicate mean Rep1 = 2.96, Rep2 = 3.13, Rep3 = 2.79

26
25
24
23
22
21
20
19
18 yield (t/ha)
17
16 2.0
15
Row

14 2.5
13
12
11 3.0
10
9 3.5
8
7
6
5
4
3
2
1
1 2 3 4 5 6 7 8 9
Column
 The barley uniformity trial - Warwick, QLD
Yield (t/ha) of one variety

Residual error around the Replicate Block Mean in Column 7

3.0
Yield (t/ha)

2.5

2.0

1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526
Row
 The barley uniformity trial - Warwick, QLD
Yield (t/ha) of one variety

Field trend - modelled as smooth trend across spatial dimensions of the

trial

26
25
24
23
22
21
20
19
18 yield (t/ha)
17
16 2.0
15
Row

14 2.5
13
12
11 3.0
10
9 3.5
8
7
6
5
4
3
2
1
1 2 3 4 5 6 7 8 9
Column