Tropical animal feeding: a manual for research workers 155

Chapter 8

Design and analysis

of experiments

This chapter was contributed by Andrew Speedy, University of Oxford,

UK. The objective is to assist researchers to compile and analyze data.
To this end, use is made of one of the simpler statistics programs
(MINITAB, Minitab Inc., Philadelphia, USA) as the model. More
powerful statistical packages may be required for studies in plant and
animal genetics and agricultural economics. But, in line with the
general philosophy of this manual, it is considered that simplicity and
ease of understanding are the principal attributes required of a
computer program and, in this respect, Minitab has much to commend
it and is therefore selected as the example. But obviously there are
various other - often more sophisticated - statistical software packages
available on the market.

THE OVERALL APPROACH

The objectives
The most important aspect of conducting good research is the definition
of the objective(s). No matter how good the design of the experiment,
how sophisticated the methods used or how clever is the statistical
treatment of results, the work is of little value if it does not answer a
question of scientific importance and practical relevance. Studying the
literature, thinking about the questions and discussing them with
colleagues, and especially the farmers who will ultimately apply the
technology, is the most important part of planning a research
programme. Research must be oriented to solving farmers' problems.

The methodology
Once the objectives are clear, the methodology can be considered. This
should be planned to provide the data to answer the questions raised and
156 Design and analysis of experiments

to satisfy the needs of the researcher and also others who may wish to
adopt the findings and apply them in other situations. It must also be
possible within the confines of the resources available (land, animals,
buildings, pens, laboratory equipment, etc.). Some of these problems
(such as numbers of replicates and land resources) may be overcome by
conducting the research 'on-farm', which also has important implications
for short-cutting the process of research application or technology
transfer.

Analysis of the data

When the data are finally collected, they must be analyzed in a way that
will provide meaningful conclusions.
Planning the analysis of the data is part of the initial process of
setting up the research programme. Knowing how the data can be
correctly analyzed and interpreted will affect how the data are collected
and the numbers of observations required. It is often valuable to produce
a 'dummy' set of data, calculated on the computer, to test the statistical
method.
The following section describes the rules and basic methods for
planning, analyzing and interpreting data relating to feed resources and
their use by animals.

PLANNING, ANALYZING AND INTERPRETING DATA

Statistical programs
There are many computer packages available for statistical analysis.
Throughout this chapter, examples will be given from data analyzed
using the package MINITAB which is available for IBM-compatible,
Apple Mackintosh and also mainframe computers. The necessary inputs
and outputs for this package will be shown. It is taken as an example of
a simple yet accurate system for the research worker as well as the
student.

Management of experimental data

Collection of data on a daily or weekly basis will yield results that must
be used to calculate the variables required for analysis: average daily
gain (kg) for each animal, average daily food intake, etc. Such initial
calculations (although they may be managed with MINITAB) are best
Tropical animal feeding: a manual for research workers 157

stored and manipulated with a spreadsheet package such as LOTUS

1-2-3. These data can be read into the MINITAB worksheet by, e.g.
RETRIEVE 'WEIGHTS.WK1';
LOTUS.

Types of data
Many of the measurements made in this type of work will be of the kind
that are called 'continuous variables': weight, food intake, blood levels,
etc. The pattern of variation of such variables conforms to the 'normal
distribution'. These can be analyzed by a range of tools called
parametric statistics, including regression analysis and analysis of
variance.
Certain variables of the type 'success/failure', 'germinated/not
germinated', 'conceived/not conceived' are 'discontinuous variables' and
the variation conforms to the binomial distribution. Also amongst this
type may be records of the type 'class 1, class 2 or class 3', where a
measurement 1.1 or 1.2 is not possible. These conform to the Poisson
distribution. These data cannot be analyzed by techniques like analysis
of variance but require 'non-parametric statistics'. However, in many
cases, such data can be 'transformed' using mathematical devices (e.g.
logarithm, square root, etc.) to make them conform to a normal
distribution. Percentage data should also be transformed.

Types of analysis
Although ANOVAR and regression are well-known techniques, both
may be analyzed using a newer method called the Generalized Linear
Model (GLM) which has a number of advantages. It fits a 'model' to the
data and predicts the means and variance from the model. Equivalent
examples of MINITAB instructions are:
REGRESS C1 1 C2 GLM C1=C2;
COVARIATE C2.
REGRESS C1 2 C2 C3 GLM C1=C2+C3;
COVARIATE C2 C3
ANOVA C1=C5 GLM C1=C5
ANOVA C1=C5+C6 GLM C1=C5+C6
No equivalent GLM C1=C2+C5;
COVARIATE C2.
No equivalent GLM C1=C2+C3+C5+C6;
COVARIATE C2 C3.
158 Design and analysis of experiments

Types of variables
From the above, it is clear that some variables are suited to 'regression
analysis' and some to ANOVA. The former are called continuous
variables and the latter are discrete variables. Levels of fertilizer, levels
of feed protein, etc. are continuous variables and can be analyzed in
GLM with the 'COVAR' subcommand. Discrete variables like variety
of crop, breed of livestock, etc. are analyzed with ANOVA or the
equivalent GLM command.

Number of treatments
The number of treatments which can be applied may depend on what is
available and the amount of experimental resources. However, with
continuous variables, it is always better to have more levels of a factor
where possible. For example, with 30 experimental units, more
information on the type of response to a factor will be obtained with 5
levels and 6 replicates than with 3 levels and 10 replicates. Very little is
lost in precision whereas much is gained in knowledge about the shape
of the response (linear, quadratic or cubic). Thus we can find the
maximum or optimum response level of a factor.

Numbers of replicates
An experiment uses a sample of a population as the experimental unit.
In general, the more replicates that are used, the greater the difference
that can be detected. However, experimental facilities are always limited
and therefore it is important to be economical with the use of resources.
There overriding rule is never to have less than 3 replicates per
(sub-)treatment. A more precise estimate of the numbers required
to detect the desired percentage difference with a t-test is given by the
formula:

%r
Expected difference = t x CV

where t = Student's t (at given treatment and error degrees of freedom);

CV = coefficient of variation; and r = number of replicates.

The appropriate CV can be found by finding comparable experiments in

published articles in the literature. It may vary between 3-25% in this
type of work.
Tropical animal feeding: a manual for research workers 159

The actual size of the experiment will vary with both the number of
replicates and the number of treatments. Fewer replicates are needed in
factorial experiments where the overall total is greater. Again, as a
general rule, ensure that the design has at least 15 degrees of freedom for
error (residual degrees of freedom).

Blocks
Blocking is a way to deal with known sources of variation which may be
sites on a gradient of fertility down a slope, different litters of pigs,
different farms, etc. Each block contains all treatments with replicates.
The analysis enables the variable 'block' to be measured and removed
from the error variation, eg:

GLM output = block + treatment

It is good to block experiments wherever a known source of variation

occurs. There is little point in including an interaction between treatment
and block because this will be difficult to interpret even if it is
significant.
Covariates
Inclusion of covariates in an analysis is another way of taking out known
variation. Covariates are continuous variables such as initial weight,
initial milk yield, etc. Their use is vital in experiments involving dairy
production where it is normal for animals of different ages, stages of
lactation and potential yield to be used. The command is:
GLM yield = initial + treatment;
covariate initial.

ANALYSIS OF CONTINUOUS DATA

Experiments with two treatments
The simplest experiment compares the results of two treatments. We
may wish to compare two or more populations (breeds of animal or
varieties of plant) and take samples from each. Our samples must be
taken at random and must represent the populations and their variation.
Experiments often involve applying some action or actions to a
sample of the population to measure its effect. The sample of the
160 Design and analysis of experiments

population is divided and the treatment(s) applied to part(s) of the

sample. If we want to know how the treated sample differs from the
untreated one, we need to keep the untreated ones as a 'control'.
Treatments must be applied at random.
When the data have been collected, we want to analyze the results to
compare the two (or more) samples or the treated groups with the
control. This is done by calculating the variance and partitioning it
between that due to treatment and the natural ('residual' or 'error')
variance. The process is called an 'analysis of variance'.
An example involves two treatments (or a treated group and control)
with 10 replicates of each. The means of the two treatments are 10 and
11.

MTB > PRINT C1-C2

ROW C1 C2
1 11.5 9.7
2 11.6 11.6
3 10.5 10.9
4 10.1 10.8
5 10.2 9.6
6 9.0 10.7
7 9.1 12.5
8 8.5 11.2
9 10.2 12.6
10 9.3 11.5

MTB > TWOSAMPLE C2 C1;

SUBC> POOLED.
TWOSAMPLE T FOR C2 VS C1
N MEAN STDEV SE MEAN
C2 10 11.11 1.01 0.32
C1 10 10.00 1.04 0.33
95 PCT CI FOR MU C2 - MU C1: (0.15, 2.07)
TTEST MU C2 = MU C1 (VS NE): T= 2.43 P=0.026 DF= 18
POOLED STDEV = 1.02

Explanation:
The data consist of two sets of values (two treatments) stored in C1 and
C2. These are listed with the MINITAB command 'PRINT'. Then the
data are compared using a 't-test' with the command 'TWOSAMPLE'.
The printout shows the means, standard deviations and standard error of
the means and calculated t value. The probability value of 0.026 is less
Tropical animal feeding: a manual for research workers 161

than 0.05 and therefore the null hypothesis that C2 is NOT different to
C1 is rejected, i.e. C2 is significantly greater than C1 (P<0.05).

Relationships between variables

In some types of data, the objective is to test the relationship between
two variables and to produce an equation which describes this
relationship. This is frequently done by regression analysis. In the
example here, the alternative 'GLM' command (Generalized Linear
Model) is used to perform the regression analysis. The example is to test
the relationship between OIL and ENERGY in feed samples:
MTB > brief 3
MTB > glm Energy=Oil;
SUBC> covariate Oil.

Analysis of Variance for Energy

Source DF Seq SS Adj SS Adj MS F P
Oil 1 0.26167 0.26167 0.26167 7.94 0.011
Error 18 0.59334 0.59334 0.03296
Total 19 0.85501

Term Coeff Sdev t-value P

Constant 10.9324 0.9046 12.09 0.000
Oil 0.5116 0.1816 2.82 0.011

Unusual Observations for Energy

Obs. Energy Fit Sdev.Fit Residual Sd.Resid
5 13.9461 13.4981 0.0412 0.4480 2.53R
17 13.6340 13.7359 0.1000 -0.1020 -0.67 X
R denotes an obs. with a large Sd.Resid.
X denotes an obs. whose X value gives it large influence.

Explanation:
The two variables are stored in columns C1-C2 and labelled Energy and
Oil. The GLM model to test is C2=C1 and the subcommand
COVARIATE C1 (abbreviated to 'cova C1') tells MINITAB to treat C!
as a continuous variable and not a discrete series of treatments. The
probability value (P=0.011) tells us that there IS a significant
relationship between Energy and Oil (P<0.05) and the equation is given
below. Badly fitting data are also indicated. The constant and
coefficient of the regression equation are given and the equation can be
derived as:
162 Design and analysis of experiments

Energy = 10.9324 (± 0.9046) + 0.5116 (± 0.1815) Oil

When more than two variables are involved, these may be included in the
model to give a multiple regression analysis. Only significant factors
should be included in the equation. The COEFFICIENT OF
DETERMINATION (r²) is found by dividing the SSx by the SStotal.
In this case:

r² = 0.26167/0.85501 = 0.306 (30.6%)

(Equations with r² less than 70% should not be used for prediction).

Experiments with more than two treatments.

When more than two factors are involved in an experiment, the technique
of ANALYSIS OF VARIANCE can be used. This can also be carried
out using GLM with the appropriate model. The following is a simple
experiment with three treatments and their effect on live weight gain
(LWG). The model is: LWG = Treat
MTB > table 'Treat';
SUBC> stats 'LWG'.
ROWS: Treat
LWG LWG LWG
N MEAN STD DEV
1 10 507.38 53.45
2 10 576.98 48.31
3 10 656.45 71.04
ALL 30 580.27 83.75

MTB > glm LWG=Treat;

SUBC> means Treat.
Factor Levels Values
Treat 3 1 2 3
Analysis of Variance for LWG
Source DF Seq SS Adj SS Adj MS F P
Treat 2 111286 111286 55643 16.30 0.00
Error 27 92145 92145 3413
Total 29 203431

Unusual Observations for LWG

Obs. LWG Fit Stdev.Fit Residual St.Resid
25 775.515 656.455 18.474 119.060 2.15R
R denotes an obs. with a large st. resid.
Tropical animal feeding: a manual for research workers 163

Means for LWG

Treat Mean Stdev
1 507.4 18.47
2 577.0 18.47
3 656.5 18.47

Explanation:
The TABLE command is used to give the means and standard deviations
of the treatments. These are the figures that should be presented in a
published paper. Then the GLM test is used as shown to produce the
analysis of variance table.
From the results shown it can be seen that the effect of treatment is
highly significant (P<0.001). A significant F test must be obtained
before it is valid to compare treatments by a t-test.
The final table lists the means and pooled standard deviation of the
mean. This is used to test for differences. The least significant
difference is t x SE of the difference. In cases where there are a
reasonable number of replicates, t will be approximately 2. Therefore
differences between means greater than 2 x SE(difference) are
significant. In this example, there are significant differences between all
treatments.

Experiments with blocks

As was explained earlier, where there is a known source of variation
(such as site, farm, litter of pigs, etc.) the treatments should be applied
equally to each block and the block taken account of in the analysis. The
following example gives the MINITAB printout for an experiment with
three blocks and three treatments:
MTB > glm LWG=block treat;
SUBC> means treat.

Factor Levels Values

block 3 1 2 3
treat 3 1 2 3

Analysis of Variance for LWG

Source DF Seq SS Adj SS Adj MS F P
block 2 19211 19211 9605 3.47 0.044
treat 2 28331 28331 14165 5.11 0.012
Error 31 85916 85916 2771
Total 35 133457
164 Design and analysis of experiments

Unusual Observations for LWG

Obs. LWG Fit Stdev.Fit Residual Sd.Resid
27 627.328 517.695 19.620 109.633 2.24R
R denotes an obs. with a large sd. resid.

Means for LWG

treat Mean Stdev
1 546.7 15.20
2 604.6 15.20
3 607.7 15.20

In this example, both block and treatment are significant (P<0.05).

There are significant differences between treatment 1 and both the other
two treatments but not between T2 and T3.

Latin square design

A Latin Square is a special sort of block design with symmetrical
arrangement of treatments in two directions. It is particularly useful in
experiments where numbers are restricted by facilities. Take an animal
experiment to measure protein degradability by the nylon bag technique,
using 4 fistulated animals. Four feeds (A, B, C, D) are studied and each
feed is incubated in the rumen of each animal in turn. The design looks
as follows:

Animal
Period 1 2 3 4
1 B A D C
2 A D C B
3 C B A D
4 D C B A

The analysis would appear as follows:

Tropical animal feeding: a manual for research workers 165

MTB> table c1 c2;

SUBC> means c4.
ROWS: Row COLUMNS: Column
1 2 3 4 ALL
1 42.800 42.900 69.100 49.400 51.050
2 47.400 53.500 47.100 56.800 51.200
3 52.300 61.200 40.500 51.800 51.400
4 61.300 51.200 54.000 39.700 51.550
ALL 50.950 52.200 52.625 49.425 51.300
CELL CONTENTS -
C4:MEAN

MTB> table c3;

SUBC> stats c4.
ROWS: Feed
Dg Dg Dg
N MEAN STD DEV
1 4 42.575 3.504
2 4 48.525 4.176
3 4 52.300 2.061
4 4 62.100 5.117
ALL 16 51.300 8.138

MTB> glm dg = row column feed

Factor Levels Values
Row 4 1 2 3 4
Column 4 1 2 3 4
Feed 4 1 2 3 4

Analysis of Variance for Dg

Source DF Seq SS Adj SS Adj MS F P
Row 3 0.58 0.58 0.19 0.01 0.999
Column 3 24.81 24.81 8.27 0.32 0.811
Feed 3 812.88 812.88 270.96 10.49 0.008
Error 6 155.04 155.04 25.84
Total 15 993.32

Explanation:
The analysis shows a significant effect of feed (P<0.01); the table of
means is given at the top of this page, together with their standard
deviations.
In general, a 4x4 (or better, a 6x6) latin square is suitable for this
type of experiment. The design can be chosen at random from lists of
latin square designs in statistical textbooks.
166 Design and analysis of experiments

Experiments with interactions

When there are two factors in an experiment, we require to know not
only whether there is an effect of each factor alone but also whether there
is an INTERACTION between them (one factor affects the response of
the animal to the other). This can be analyzed using GLM by specifying
the terms:
MTB> GLM Y = A B A*B

Alternatively, the above expression can be abbreviated to:

MTB> GLM Y = A ! B

The following example refers to an experiment with three energy

treatments and three protein treatments.

MTB > table 'energy' 'protein';

SUBC> stats 'LWG'.

ROWS: Energy COLUMNS: Protein

1 2 3 ALL
1 3 3 3 9
513.11 636.95 650.69 600.25
50.23 63.41 35.10 79.06

2 3 3 3 9
640.75 650.09 711.54 667.46
65.69 18.29 22.36 48.96

3 3 3 3 9
649.58 737.08 627.28 671.32
40.65 59.81 54.66 67.68
ALL 9 9 27
601.15 674.71 663.17 646.34
80.60 64.84 50.98 71.94

CELL CONTENTS --
LWG:N
MEAN
STD DEV

MTB > glm LWG = energy ! protein;

SUBC> means energy ! protein.

Factor Levels Values

Energy 3 1 2 3
Protein 3 1 2 3
Tropical animal feeding: a manual for research workers 167

Analysis of Variance for LWG

Source DF Seq SS Adj SS Adj MS F P
Energy 2 28746 28746 14373 6.12 0.009
Protein 2 28172 28172 14086 6.00 0.010
Energy*
Protein 4 35364 35364 8841 3.76 0.021
Error 18 42287 42287 2349
Total 26 134569

Means for LWG

Energy Mean Stdev
1 600.3 16.16
2 667.5 16.16
3 671.3 16.16
Protein
1 601.1 16.16
2 674.7 16.16
3 663.2 16.16
Energy*Protein
1 1 513.1 27.98
1 2 636.9 27.98
1 3 650.7 27.98
2 1 640.7 27.98
2 2 650.1 27.98
2 3 711.5 27.98
3 1 649.6 27.98
3 2 737.1 27.98
3 3 627.3 27.98

Explanation:
Both energy and protein have significant effects. In addition there is a
significant interaction between energy and protein, that is, the effect of
one is mediated by the effect of the other.

Introducing covariates
Another known source of variation may be a continuous variable such as
previous milk yield, starting weight, previous performance, etc. This is
very often the case in milking experiments with cows or goats when the
experimental animals will almost certainly have different yields and be
at different stages of lactation. The following example is an experiment
with three treatments to measure the effect on the milk yield of cow.
Initial yield is stored in the data table as the variable 'init' and the
analysis is as follows:
168 Design and analysis of experiments

MTB > glm yield=treat;

SUBC> cova init;
SUBC> means treat.

Factor Levels Values

treat 3 1 2 3

Analysis of Variance for yield

Source DF Seq SS Adj SS Adj MS F P
init 1 274.56 250.30 250.30 55.75 0.000
treat 2 44.08 44.08 22.04 4.91 0.014
Error 32 143.66 143.66 4.49
Total 35 462.31

Term Coeff Stdev t-value P

Constant 2.3086 0.9055 2.55 0.016
init 0.9670 0.1295 7.47 0.000

Means for Covariates

Covariate Mean Stdev
init 6.438 2.789

Adjusted Means for yield

treat Mean Stdev
1 7.100 0.6121
2 8.695 0.6130
3 9.808 0.6150

Explanation:
If the analysis had been performed without including initial milk yield as
a covariate, no significant differences between treatments would have
been found. However, with the inclusion of the term 'init' as a covariate,
there is a significant effect of treatment (P<0.05).
The final table of means shown are the values for each treatment
adjusted for initial milk yield. Treatment 1 again differs significantly
from the other 2.

Better experiments (more levels of the treatments)

In many feeding experiments we need to test the effects of the level of
feed inclusion or of the level of a nutrient such as energy or protein. All
the experiments described above have 3 treatments and the treatments are
treated as DISCRETE variables. However, we often want to know the
SHAPE of the response to a treatment, whether there is a maximum or
optimum level. With 2 or 3 treatments we can only see if there is a
Tropical animal feeding: a manual for research workers 169

response. By including more levels of the treatment we can test the

linear, quadratic and cubic effects. That is, we can see if the response
is curved. We can also find the equation which describes the curve. To
do this, we treat the factors as CONTINUOUS variables. As a general
rule, it is better to include more levels of treatments in this type of
experiment as we obtain more information about the response. We make
better use of the available experimental material and, provided we have
a reasonable number, we lose very little in precision (only 1 degree of
freedom for each level). The following is an experiment with 5 levels of
energy and 5 levels of protein. We can test for the response to both and
also for the interaction between energy and protein.

MTB > table 'Energy' 'Protein';

SUBC> stats 'LWG'.
ROWS: Energy COLUMNS: Protein
1 2 3 4 5 ALL
1 3 3 3 3 3 15
522.31 522.62 565.98 535.65 596.00 548.51
20.46 32.52 18.15 29.77 48.59 39.96
2 3 3 3 3 3 15
555.63 564.93 631.32 638.35 643.72 606.79
28.50 23.54 19.19 29.58 18.51 44.64
3 3 3 3 3 3 15
558.41 620.30 638.36 646.82 641.30 621.04
5.65 25.63 41.76 8.58 27.69 40.03
4 3 3 3 3 3 15
621.73 664.42 673.26 666.18 667.57 658.63
30.48 34.30 36.86 26.19 36.15 33.97
5 3 3 3 3 3 15
616.58 667.53 690.55 698.05 691.20 672.78
17.21 4.94 14.17 9.19 36.06 35.10
ALL 15 15 15 15 15 75
574.93 607.96 639.89 637.01 647.96 621.55
43.91 62.68 50.47 59.79 44.08 58.05
CELL CONTENTS --
LWG:N
MEAN
STD DEV

MTB > glm LWG=Energy Protein Energy*Energy Protein*Protein

Energy*Protein;
SUBC> cova Energy Protein;
SUBC> test Energy Protein Energy*Energy Protein*Protein
Energy*Protein /error.
170 Design and analysis of experiments

Analysis of Variance for LWG

Source DF Seq SS Adj SS Adj MS F P
Energy 1 135343 18105 18105 22.13 0.000
Protein 1 45991 14473 14473 17.69 0.000
Energy*
Energy 1 4514 4514 4514 5.52 0.022
Protein*
Protein 1 6682 6682 6682 8.17 0.006
Energy*
Protein 1 414 414 414 0.51 0.479
Error 69 56438 56438 818
Total 74 249382

Term Coeff Stdev t-value P

Constant 396.39 26.68 14.86 0.000
Energy 61.38 13.05 4.70 0.000
Protein 54.88 13.05 4.21 0.000
Energy*Energy 4.636 1.974 2.35 0.022
Protein*Protein 5.641 1.974 2.86 0.006
Energy*Protein 1.175 1.651 0.71 0.479

F-test with denominator: Error

Denominator MS = 817.94 with 69 degrees of freedom

Numerator DF Seq MS F P
Energy 1 135343 165.47 0.000
Protein 1 45991 56.23 0.000
Energy*Energy 1 4514 5.52 0.022
Protein*Protein 1 6682 8.17 0.006
Energy*Protein 1 414 0.51 0.479

Explanation:
The first TABLE gives the means for each sub-treatment with standard
deviations. The mean for each main treatment is shown at the right hand
side and bottom of the table. Then the analysis of variance is performed.
Notice that both Energy and Protein are set as continuous variables with
the subcommand COVA Energy Protein. Notice also an additional
subcommand TEST. This requires some explanation.
TEST is used as a sub-command to GLM to force MINITAB to use
the sequential sums-of-squares and consequent mean squares in the test
of significance, rather than the adjusted sums-of-squares and mean
squares, which is the default action. The difference between them is that
the adjusted sum-of- squares refers to each factor when all the others
have been accounted for; the sequential sum-of-squares is calculated
sequentially from the top so that each factor is taken out in turn.
Tropical animal feeding: a manual for research workers 171

The TEST sub-command should always be used when the factors are
NOT independent, as is inevitably the case with linear, quadratic and
cubic effects (X, X*X, X*X*X). In other experiments where the
sequential sums-of-squares and adjusted sums-of-squares are very
different, non-independence is implied and the TEST sub-command
should be used to force the use of the sequential sums- of-squares. The
factors tested by the above commands are:

Energy: linear effect of energy

Protein: linear effect of protein
Energy*Energy: quadratic effect of energy
Protein*Protein: quadratic effect of protein
Energy*Protein: Energy x Protein interaction.

In assessing significance, the LAST table should be used (F test with

denominator: Error). In the example, the linear and quadratic effects for
both Energy and Protein are significant but there is no interaction (NS).
This shows that the effects of Energy and Protein are curvilinear
(diminishing response in this case as the quadratic coefficients are
negative).
There is little reason for a farmer to increase either energy or protein
above the third level in both cases. An accurate equation can be obtained
by rerunning the analysis with the interaction removed (because it was
not significant) and using the constant and coefficients to construct the
equation.
Note that in experiments with two treatments where we wish to test
the interaction, the model can be abbreviated to:
MTB> GLM LWG = FEED ! SYSTEM

This will test the main effects and the interaction (FEED, SYSTEM and
FEED*SYSTEM). This could not be used in the above example because
we excluded some of the more complex interactions.

Dealing with unbalanced designs

Particularly in on-farm research, we may not be able to apply all of the
treatments, all of the time. With ANOVA, this presented serious
problems and necessitated calculating 'missing plots'. However, GLM
172 Design and analysis of experiments

is a powerful tool for dealing with unbalanced designs and has less
limitations. A fuller explanation of the use of GLM for unbalanced
designs is given below.

Some Restrictions on Models in GLM

(Minitab Reference Manual 1991, 8-28)
Although models can be unbalanced in GLM, they must be "full rank."
Thus, there must be enough data to estimate all the terms in your model.
For example, suppose you have a two factor crossed model with one
empty cell. Then you can fit the model GLM Y = A B, but not GLM Y
= A B A*B. Don't worry about figuring out whether or not your model
is of full rank. Minitab will tell you if it is not. In most cases, eliminating
some of the high order interactions in your model (assuming, of course,
they are not important) will solve your problem.
There is another restriction: nesting must be balanced. Suppose A
has 3 levels, and B is nested within A. If B has 4 levels within the first
level of A, it must have 4 levels within the second and third levels of A
also. Minitab will tell you if you have unbalanced nesting.
In addition, the subscripts used to indicate the 4 levels of B within each
level of A must be the same. Thus, you cannot use ( 1 2 3 4 ) for the
levels of B within level 1 of A, and ( 5 6 7 8 ) for the levels of B within
level 2 of A."

ANALYZING EXPERIMENTS WITH DISCONTINUOUS VARIABLES

Chi-squared analysis
The use of chi-squared analysis enables the analysis of experiments
involving data of the yes/no type or when the results are counts. Note
that in the latter case the absolute data should be used and the results
should NOT be converted into percentages.

The data are arranged in 'contingency tables' of the type:

Germinated Not germinated

Control 149 51
Treated seed 182 18
Tropical animal feeding: a manual for research workers 173

The chi-squared statistic is rather like the SS in that it is the square of the
difference between the observed result and the expected result (if the
results were averaged between the two treatments). We compute a value
for each cell, then sum the values for all the cells and compare the value
with the value in tables.

chi ² =3 | observed - expected |²

expected

If the total chi-squared value is GREATER than the tabulated value, then
there is a significant difference between the rows or treatments.
The data should be entered into MINITAB in two columns and the
MINITAB command CHISQUARE used as follows:
MTB > chis c1 c2

Expected counts are printed below observed counts

C1 C2 Total
1 149 51 200
165.50 34.50
2 102 10 200
165.50 34.50
Total 331 69

ChiSq = 1.645 + 7.891 + 1.645 + 7.891 = 19.073

df = 1

Note that for a 2x2 table there is one degree of freedom (only one
comparison possible). Look up the tables on the line for 1 d.f.

Chi-squared analysis may equally be used for simple experiments

with more than two treatments. A 3 x 2 table has 2 d.f. The degrees of
freedom is calculated as: (rows-1) x (columns-1)
It is also possible to have 3 x 3, 3 x 4, 4 x 4..... etc. tables.

The chi-squared statistic behaves like normal variance in that it may be

partitioned between several factors and the interaction may also be
calculated. Take the following example:
174 Design and analysis of experiments

Four treatments are applied to 100 cows each and the results measured
as 'conceived' or 'failed' to conceive:

Treatment Conceived Failed

High energy - high protein 81 19
High energy - low protein 88 12
Low energy - high protein 75 25
Low energy - low protein 43 57

First, compute the chi-squared value for the whole table (3 d.f.):
Total treatment effect (3 df) chi ² = 58.549 > 11.3 significant
(P<0.01)

Now combine rows 1+2 and 3+4 into a 2x2 table and calculate
chi-squared (1 df) to calculate the energy effect and combine rows 1+3
and 2+4 into another 2x2 table and calculate the chi-squared to test the
protein effect:
Energy effect (1 df) chi ² = 32.080 > 6.63 significant (P<0.01)
Protein effect (1 df) chi² = 7.709 > 6.63 significant (P<0.01)

Subtract the energy and protein chi-squared values from the total
chi-squared to get the remaining effect which is due to the interaction.
Energy x protein (1 df) chi ² = 6.760 > 6.63 significant (P<0.01)

There is a significant effect of energy and protein, and there is also an

interaction between energy and protein. Note how the chi-squared values
are additive and we partition the original 3 df into 1 for each main effect
and 1 for the interaction.

Numbers required for chi-squared analysis

(eg: animal reproductive performance)
The numbers required to obtain significant differences in this type of
analysis are usually greater than with measurements such as growth or
yield. Consider the results of chi-squared analysis where there is a
difference of 10% in fertility of cows:
Tropical animal feeding: a manual for research workers 175

25 cows per treatment

conceived failed
20 5 chi ² = 0.439
18 7

50 cows per treatment

conceived failed
40 10 chi² = 0.877
36 14

100 cows per treatment

conceived failed
80 20 chi² = 1.754
72 28

150 cows per treatment

conceived failed
120 30 chi² = 2.632
18 7

225 cows per treatment

conceived failed
180 45 chi² = 3.947
162 63

It is only when we have 225 cows per treatment that we can detect the
10% difference in fertility (P<0.05), which is an important practical
difference.

Limitations of chi-squared analysis

Certain rules must be considered when applying chi-squared analysis.
One of these is that all cells should contain values greater than 5
(Snedecor). Otherwise, chi-squared is unreliable particularly with only
1 df.
As an improvement, Yates (1939) proposed an adjustment known as
'Yate's Correction Factor'. This is simply an adjustment of the formula
as follows:
176 Design and analysis of experiments

chi² adjusted = (| observed - expected | - 0.5)²

expected

Exact probabilities
Occasionally it is possible to obtain only limited amounts of data, for
example, if to obtain data would destroy experimental units. When the
numbers in a 2 x 2 table are very small, it may be best to compute exact
probabilities rather than to rely on the chi-squared approximation.

Example:

Have Have not Total

Standard 5 2 7
Treatment 3 3 6
Total 8 5 13

We compute the probability of obtaining the observed distribution or a

more extreme one, the more extreme ones being:

6 1 7 7 0 7

2 4 6 1 5 6
and
8 5 13 8 5 13

We require the sum of the probabilities associated with the three

distributions. Marginal totals are the same for all three tables. The sum
of the probabilities will be used in judging significance. The probability
associated with the distribution:

n11 n12 n1.

n21 n22 n2.
n.1 n.2 n..
is
Tropical animal feeding: a manual for research workers 177

P = n1.!n2.!n.1!n.2!
n11!n12!n21!n22!n..!

where nij! is defined as:

n! = n(n-1)...1 and 0! = 1

Read n! as 'n factorial'.

The probabilities for the three tables above are

P= 7! 6! 8! 5! = 0.3263
5! 2! 3! 3! 13!

P= 7! 6! 8! 5! = 0.0816
6! 1! 2! 4! 13!

P = 7! 6! 8! 5! = 0.0047
7! 0! 1! 5! 13!

The sum of the probabilities is 0.4126 (not significant, P>0.05). It is

clear that the computation of the first and second probability alone was
sufficient to answer the question of significance. In practice, one uses
this approach by computing the largest individual probability first, and
so on.

Other non-parametric tests

Chi-squared analysis can also be used to test whether a distribution
conforms to a particular type such as a binomial or Poisson. The
calculated distribution is tested against the observed one. Other
non-parametric tests which may be required are:

The sign test - for comparing medians.

Wilcoxon's signed rank test - an improvement on the above.
Friedman's test for randomized complete block design.
Wilcoxon's test for completely random design, two populations.
Mann-Whitney test for the same but with unequal samples.
178 Design and analysis of experiments

Kruskal-Wallis test for completely random design, any number of

populations.
Spearman's coefficient of rank correlation.
Tukey's test of association.

All the above can be used as quick tests without having to make
assumptions about the nature of the population, its type of distribution
and variance. However, where it is possible to make the necessary
assumptions for the use of anovar, etc., more information (on means,
variance, etc.) will be obtained.

Transforming non-normal data for analysis

To use the analysis of variance, we have to confirm the assumptions that:
1. Treatment and environmental effects are additive.
2. Experimental errors are random, independently and normally
distributed about zero mean and with a common variance (i.e. the data
are of the 'height' or 'weight' type).
Violation of these assumptions may result in unreliable statistical
tests and the unacceptability of the conclusions (particularly for
publication). Data which consist of counts and percentages, in particular,
do not conform to these requirements. We can use the non-parametric
tests (such as chi-squared) but these give us less information on
treatment effects. A solution is often to 'transform' the data to conform
to a normal probability distribution. For this, we take the original data,
apply a formula and carry out anovar on the transformed data. We do
NOT convert the data back to present the statistics but state that the data
were transformed before analysis. The following techniques apply:

Square root transformation

When data consist of small whole numbers, e.g. number of plants or
insects of a stated species in a given area, they often conform to the
'Poisson distribution', for which the mean and variance are equal. The
analysis of such numbers is often best done by first taking the square
root of each observation (%x) before carrying out the anovar.
Tropical animal feeding: a manual for research workers 179

Percentage data based on counts and a common denominator, where

be analyzed using %x. Percentages between 80-100 should be subtracted

the range of percentages is 0-20% or 80-100% (but not both), may also

from 100 before the transformation is made.

It can be seen that when there are mostly low counts with a few very
high ones, the probability will be skewed and taking the square root will
pull in the high 'tail'. Notice also that this type of data will have a fixed
end, 0 (or 100% in the case of high percentages) which prevents it from

When very small values are involved, %x tends to overcorrect and

showing a two sided normal distribution shape.

%(x+0.5) should be used when some of the values are <10 and especially
when zeros are present.

The logarithmic transformation

The logarithmic transformation (log10 x) is used with positive integers
which cover a wide range. This will again pull in a high 'tail' particularly
when the high values are 100's or 1000's. When values are low (and
obviously with 0), log(x+1) should be used. (The log transformation is
also appropriate in experiments in which the variable is the variance.)

The angular transformation(arcsin %x or sin-1 %x) is applicable to

The angular transformation

binomial data expressed as a decimal fraction or percentages when the

percentages cover a wide range. (%x was recommended for percentages
0-20 and 80-100. For percentages 30-70 it is doubtful if any
transformation is required.) Data may require to be divided by the
numerator or 100 in the case of percentages to produce the decimal
fractions required.
Classical binomial data are the 'success or failure' type variables -
conception rate, germination rate, etc. When given as a proportion, the
angular transformation is appropriate. The square root may be applied
when they are given as percentages (80-100%).
It is not always obvious which type of transformation is required. It
may be helpful to plot the data and data transformed by various methods
to check the effect on the shape of the curve.
180 Design and analysis of experiments

SIMULATING EXPERIMENTAL DATA

The data referred to here are not real. They were produced by
simulation, using MINITAB to produce sets of random data conforming
to normally distributed probabilities and with appropriate variance.
This can be a very useful technique, used to run the experiment in a
theoretical way (using appropriate means and SE's obtained from
previous experience or the literature). We can then try the statistical
analysis before the experiment starts and identify and limitations in the
design.
The appropriate MINITAB commands to create a set of 20 normally
distributed data, with mean 10 and SD±1 in column 1, are:
MTB> RANDOM 20 C1;
MTB> NORMAL 10 1.

We might do the same in C2, with mean 12 and SD±1 and perform an
ANOVAR on the two columns.
The technique can be used to simulate factorial experiments,
randomized block designs, latin squares, etc., using appropriate columns
for different effects and variances. These can be summed to produce the
simulated values for the data column and the appropriate analysis
performed.
It is a good method to 'practise' statistics, while gaining an
appreciation of the effects of numbers, different levels of variation and
different methods of analysis.