Comparative analysis of the effect of transformation techniques on some robust regression measures
Loading...
Date
October 22, 2016
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
When the distribution of the errors in a linear model is normal, the OLSE pro-
vides reliable parameter estimates for the linear model. When errors deviate from
normality, the OLS provides unreliable estimates. This study however evaluated
the e ect of transformation techniques on some robust regression measures by
comparing four robust regression methods namely M-estimation, MM-estimation,
S-estimation and LTS estimation to the OLS estimation using the Mean Absolute
Residuals, R2, and Relative e ciencies. Data was simulated and analyzed from
a linear model under various situations where the error term has normal distri-
bution, lognormal distribution, truncated normal distribution and contaminated
normal distribution. The log and Box-Cox transformation were applied to the
simulated data before estimating and comparing parameters. The results showed
that on the average the robust methods were 2.4% more e cient than the OLSE
for the Normal case when no transformation was applied with the exception of
the SE which was 10% less e cient. For the lognormal case, the robust methods
were 26.9% more e cient under no transformation, 24.5% more e cient under log
transformation and 22.8% more e cient under Box-Cox transformation. In the
contaminated normal case, the robust methods were 6.3% more e cient under
no transformation, 5.8% more e cient under log transformation and 3.8% more
e cient under Box-Cox transformation. In the case of the truncated normal, 10%
tail truncation resulted in the robust methods being 0.6% more e cient under no
transformation and 7.2% more e cient under log and Box-Cox transformation.
Increasing the tail truncation to 20% made the robust methods 3.3% more e -
cient under no transformation and 10.8% more e cient under log and Box-Cox
transformation. Based on the study, it is recommended that, the robust methods
should be used when the basic assumptions fail. Also, the Box-Cox transfor-
mation technique should be used to linearize data if non-linear relationship is
established before using the robust methods.
Description
A thesis submitted to the Department of Mathematics,
Kwame Nkrumah University of Science and Technology in
partial fufillment of the requirement for the degree
of Master of Philosophy in Mathematical Statistics,