Incomplete Block Design (IBD)
If in a block the number of experimental units is smaller than the number of
treatments, then the block is said to be incomplete and a design constituted of such
blocks is called an incomplete block design. In order to ensure equal or nearly equal
precision of the comparison of different pairs of treatments, the treatments are so
allotted to the different blocks that each pair of treatment has the same or nearly the
same number of replications and each treatment has an equal number of replications.
Balanced IBD
When the number of replications of all pairs of treatments in a design is the same,
then the design is known as BIBD. The design ensures equal precisions of the estimates
of all pairs of treatment effect.
Definition:- An IBD is said to be BIBD if it satisfies the following conditions
1. The experimental material is divided in to b blocks of k units each, different
treatments being applied to the units in the same block
2. There are t treatments each of which occurs in r blocks
3. Any two treatments occur together in exactly blocks.
The quantities t,b,r,k and are called parameters of BIBD. The necessary relationships
between the parameters of BIBD are
(i) rt=kb
(ii) (t-1)=r(k-1)
(iii) b t ; r>
Lattice design
A balance two dimensional (design) with k 2 treatments having one restriction is called
simple lattice. Also in this design, the treatments should be assigned to each block in
such a manner that is same for all pairs of treatments. For a design to be balanced,
minimum required number of blocks is k(k+1). Thus, at least k+1 replications are
needed for a k2 simple lattice. This property of having separate replications for BIBD
holds only when t is a multiple of k and especially for lattice square design. Most of the
BIB design do not hold this property. In general, in an m-dimensional balanced lattice
design, the number of treatments is km, where k is a prime number or prime power.
Response surface methodology
The response surface methodology seeks to relate an average response to the
value of input factors which are quantitative in nature. If the input factors are amount of
fertilizer, soil moisture, time and the like, the yield of a crop or response may be
expressed as a function of the levels of these factors. For example, an experimenter
might want to find out how the yield of a crop depends upon the amount of N, P, K. He
is interested in the presumed functional relationship Y=f(N,P.K) that expresses the
yield Y as function of the variables N, P, K. The response actually observed in a
particular experiment differs from Y because of experimental error, which is denoted
by e . Thus,
Y=f(N, P, K)+e
is the required model relating the observed response to the levels of input factors. In
general, if we have r input factors (variables) X1, X2,…..Xr, the model can be written as
Y=f(X1, X2,…..Xr)+e
The function, f is called the response surface.
There are two uses of response surface. First, it has been applied to describe how the
response is affected by a number of quantitative variables over some already chosen
levels of interest. The second use is to locate the neighborhood of maximal or minimal
response. In agriculture experiments, it is of interest to determine optimum level of a
factor or combination of factors that will maximize the yield. When the response is cost
of production per unit of output, the objective may be to minimize the response.
Cross- over Designs
Cross over Designs are used in situation in which treatment are applied in sequence
over several periods to a group of individual items and the number of experimental
units may be less than number of observation. The design has been used for comparing
two or four treatments in dairy husbandry and other biological studies. The cross over
design has two restrictions imposed on randomization of the treatments to the
experimental units. The treatments are all included in each replicate or group. The
experimental units are rated with regards to time of application in each replicate or
group. The second restriction is that each treatment must be applied an equal number of
times in each period or time in the replicates.
For example, suppose that we have to compare the effect of two feeding rations, A & B
on the amount and quality milk produced by the cow. Since cows vary greatly in their
milk production, each ration is tested on every cow by feeding it either the first or the
second half of the period of lactation, so that each cow gives a separate replicate. The
rations are allotted to the periods at random with the restriction that half of the cows
receive first ration and the other half receives 2 nd ration B in the period 1 and cows
receiving A receive B in period 2.
The experimental design for the six replicates (six cows) is of the following
Cows or Replication
Rows 1 2 3 4 5 6
Period I B B A A B A
Period II A A B B A B
If the above design were applied to an experimental situation which require a separate
experimental unit for each replicate, the analysis would be the same as given alone. For
example, suppose that two treatments A and B are applied to dairy cows, that treatment
period is used, that twelve cows are paired in to 6 pairs with each member of pair being
rated as superior or inferior and that one half of the superior and one half of the inferior
cows receive treatment B. The experimental design might be of the following form
Cows or Replication
Rows 1 2 3 4 5 6
superior B B A A B A
Inferior A A B B A B
The cross over design may be used for any number of treatments with the condition that
the number of replicates must be a multiple of the number of treatments.
TRANSFORMATIONS
ANOVA has three assumptions.
1. The effects are additive.
2. The experimental errors are independent.
3. The errors are distributed normally with mean zero and common variance 2.
When the above assumptions of ANOVA are violated we have to transform the data.
Whenever the standard deviations of samples are roughly
proportional to the means, an effective transformation may be a log
transformation. Frequency distributions skewed to the right are often
made more symmetrical by transformation to a logarithmic scale. While
logarithms to any base can be used, common logarithms (base 10) or
natural logarithms (base e) are generally the most convenient. The
presence of multiplicative effects and a rough proportionality between
standard deviations and means suggest that a logarithmic transformation
may be appropriate. For example, a log transformation is often appropriate
when the dependent variable is a concentration. This cannot be less than zero, and may
have several moderately high observations, but may have a small number of very high
values. Taking logs (one can be added to each observation, if some are zero) often
normalizes the data.
Whenever the response variable is a count of relatively rare events
(e.g. insect counts on a leaf, blood cells within a gridded region of a
hematocytometer, etc.) the data tend to follow a special distribution
called a Poisson distribution . In such situations square root
transformation is used. It is better to use (y+0.5) instead of (y). If
there is negative values in the data, use appropriate constant to make it
positive by adding it through out. For example, counts such as the numbers
of cells in a haemocytometer square, can sometimes produce data which can be
analysed by the ANOVA. If the mean count is low, say less than about five, then the
data may have a Poisson distribution. This can be transformed by taking the square root
of the observations.
Another kind of data that may require transformation is that based on counts
expressed as percentages or proportions of the total sample. Such data generally
exhibit what is called a binomial distribution rather than a normal distribution. One of
the characteristics of such a distribution is that the variances are related to the means.
In such situation we go for arc sin transformation to the square root of proportion or
percentages. i.e., sin-1(p). It is used to stabilize the data when observed proportions are
in the range of 0 to 30% or 70 to 100%. When the data contains 0 or 1, transformation
is improved by replacing 0 by (1/4n) and 1 by (1-(1/4n)) before taking angular values,
where n is the number of observations based on which p is estimated for each group.
A logit transformation {loge(p/(l-p))} where p is the proportion, will often correct
percentages or proportions in which there are many observations less than 0.2 or
greater than 0.8 (assuming the proportions cannot be < 0 or > 1)
When the treatment S.D. are proportional to square root of means, the appropriate
transformation is x to 1/x. It is mostly used when time is the independent variable.
In general transformations are used to reduce the heteroscedasticity of the data or
to make the data more closely resemble the normal distribution.
When a transformation has been made, the analysis is carried out with the
transformed data. The conclusions are drawn from such analysis. However, while
presenting the results, the mean and standard errors are transformed back in to original
units. While transforming back to original units, some corrections have to be made. In
case of logtransformed data, if the mean value is , the mean value of original units
will be , where is variance of the mean .
If the square root transformation has been used, then .
If no suitable transformation can be found, a nonparametric test can often be used