Dear all,
I posted a message this afternoon concerned with assessing
variance with data that was skewed. First of all thankyou to all
those that have replied. I have another question: the problem that i
have is this. I am attempting to use the full ICP-AES spectrum
which consists of 5500 data points. I am using the multivariate
routine PLS1. It has been found that there is a need to reduce the
number of variables prior to PLS modelling. What I have been
attempting to do is to rank these variables according to their
importance. PLS1 does not actually model the original data, but
forms latent variables which convey to the model data variability. it
is hoped that this variability is due to the concentration changes in
the calibration set!!
The most success to date has been obtained by ranking the
variables according to their std. However, as i have just found, a lot
of the variables display a skewed distribution, the use of std may
not therefore be appropriate. My main problem is that within the
ICP-AES spectrum there is a huge range in the emission
intensities. Thus we may have a huge Ce peak with very little
variation, but with a large std due solely to its large value. On the
other hand there are much smaller peaks, with potentially more
information for calibration purposes, which exhibit a much smaller
std because of thier low intensity. It is not intensity which dictates
potential use in calibration. by ranking according to the std I am
effectively saying that the more intense peaks are more informative,
this is not what i need. I need some sort of ranking criteria capable
of giving a relation between a variable and an analyte (chemical of
interest) which is not influenced by the intensity of the peak, but by
its variation due to the concentration changes only. I have
investigated the use of the covariance value, but am I right in
thinking that ths value is highly influenced by the stdev value again?
Any help would be greatly apprecieted
Thankyou
Mike
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|