Loading...
Comparison of Transformations used in the Analysis of Data From Agricultural Experiments
A study was undertaken to empirically examine the suitability of the various commonly used transformation techniques on the analysis enumerative data relating to agricultural experiments or surveys. The possibility of evolving better transformations for the analysis of data pertaining to certain sp...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Ph.D Thesis |
| Language: | Undetermined |
| Published: |
Vellanikkara
Department of Agricultural Statistics, College of Horticulture
1997
|
| Summary: | A study was undertaken to empirically examine the suitability of the various commonly used transformation techniques on the analysis enumerative data relating to agricultural experiments or surveys. The possibility of evolving better transformations
for the analysis of data pertaining to certain specific environments was also explored.
Data for the study were gathered from the available records of the project on
pest surveillance survey on paddy, those on the project on early stage pest control on
paddy of Regional Agricultural Research Station, Pattambi and those of the post emergence herbicidal evaluation trial for the control of Pennisetum pedicel/atum of the
All India Co-ordinated Research Project on Weed Control, College of Horticulture,
Vellanikkara.
Comparisons among the various commonly used transformations were made
either on the basis of a single criterion viz., Bartlett's chi-square test, Tukey's test of non-additivity, Levene's residual F test or Taylor's power low or on the basis of multiple criteria viz., likelihood method of Box and Cox (1964) or the graphical method of Draper and Hunter (1969).
The results of the analysis of the data relating to pest surveillance study on
paddy showed that logarithmic transformation was the most desirable in the analysis of
data on the counts of all the major types of insects on rice (stem borer, jassid, gall fly,
leaf folder, BPH) the only exception being case worm for which a squareroot
transformation was indicated. Box-Cox approach undoubtedly emphasised the utility of
the logarithmic transformation in analysing data on counts of insects and weeds. The
graphical plot of the log likelihood function against the exponent of the power transform had a maximum value around zero for all sets of data indicating the superiority of the logarithmic transformation over the others. The graphical method of Draper and Hunter failed to suggest a unique transformation for all sets of data. However, in most cases, the choice lied between squareroot and logarithmic transformations with a slight superiority for the squareroot transformation.
As per the method suggested by Berry (1987) a suitable location parameter
'C' was estimated for the analysis of sets of data involving extreme observations including zero values. The estimated value of the additive constant was found to be approximately 2.8 for all the different sets of data. The analysis of transformated data after incorporating the estimated value of the additive constant to each observation showed slightly better results than the ordinary analysis after incorporating the additive constant 'one' to each datum.
An alternative estimate of the parametric constant in the inverse hyperbolic
sine squareroot transformation was developed and the resultant estimate produced better results than those by the estimate proposed by Beal (1942).
Assuming a non-linear relationship between mean (u) and standard deviation
(σ) a new transformation x' = log(x2+k) where x = original observation, k = a
parametric constant to be estimated from the data, was derived theoretically. The best
estimate (k^) of the parameter k was derived to be
^ ∑σ/μ - n
k = ----------- where n is the number of observations.
∑ (1/μ2)
This transformation is expected to be useful in the analysis of data when the mean- standard deviation relationship is approximately parabolic. In general, the new transformation was found to be slightly better than the inverse hyperbolic sine
squareroot transformation in the analysis of data with disproportionate amount of
variability.
Rank transformations were also found to be helpful in the analysis of data
when there are model violations and were in general helpful for increasing the sensitivity of the F test.
|
|---|