kriging

Hi All,
I had a look at the paper suggested by Edzer. Some of my data is of 
similar nature (% cover rather than direct counts; but many zeroes in 
the data set). It is difficult to assess the presence of trends with 
this variable due to 1) a limited number of covariates (i.e x,y, and 
depth) but also due to the nature of variation of the data (the 
distribution is quite patchy due to variations in substrate, exposure 
degree, etc). For example;

panel.hist <- function(x, ...)
 {
     usr <- par("usr"); on.exit(par(usr))
     par(usr = c(usr[1:2], 0, 1.5) )
     h <- hist(x, plot = FALSE)
     breaks <- h$breaks; nB <- length(breaks)
     y <- h$counts; y <- y/max(y)
     rect(breaks[-nB], 0, breaks[-1], y, col="cyan", ...)
 }
 pairs(fulmar, diag.panel=panel.hist)

shows that the variables of interest are highly left skewed for both 
coast and fulmar variables. I guess I'm confused as to why glm usage in 
this case was chosen as the trend predictor? Was it simply due to the 
limited covariates and the skewed distributions? If one looks as the 
residuals of the glm, they show some characteristics of non-homogeneity. 
Eg; plot(glm98)

Does the glm procedure use GLS to estimate the model as opposed to OLS? 
Please excuse my beginner level ignorance, but I was under the 
assumption that when there was a trend present, GLS should be used. I 
guess part of my confusion also stems from difficulty visualizing the 
"linear" model specified in the fulmar demo (i.e. fulmar ~ coast + 
depth)....there appears in the scatterplots to be only weak 
relationships between each of the variables, a case not unlike the data 
I am working with. I've tried variations of ordinary, and universal 
kriging and also "regression kriging" as discussed below...they all seem 
to give similar predictions of distribution and actual data values. I 
had avoided using the kriging variance due to my perhaps incorrect 
assumption that the variance computed on trend residuals was not optimal.

I guess it really depends on the way in which one models the data and 
any underlying trends...and that there does not appear to be a "gold 
standard" for data such as the fulmar data set?

kriging

Thread (6 messages)