Isobel Clark wrote:

> Behrang

>

> What weighting do you use in the weighted least squares?

>

> Isobel

I have found choosing suitable weights always a frustrating

event. Cressie's weights, let's say N_h/[(gamma(h))^2], has

attractive properties, both intuitively and statistically. Here,

gamma(h) is the model value, not the sample variogram

value (because that might be zero; think of binary data). N_h

is the number of point pairs used to estimate semivariance

at lag (interval) h.

It's downside is that while fitting the variogram, gamma(h)

changes, and so the weights change. This has consequences:

while fitting, the criterion you try to minimize may actually

increase while the fit gets better. This is hard to deal with.

If you calculate e.g. a weighted R^2, and look at the trace,

it will go up, down, and then up, down, etc. The context changes.

If you fix gamma(h), say to it's starting values, then the final

fit may very much depend on which starting values you used.

Isobel, how do you deal with this?

As an alternative, (and the default value in gstat under R or

S-Plus), I now tend to use N_h/h^2 [*], which is equivalent to

Cressie's weights for a linear variogram with no nugget. It

works often, but will give rediculusly large weight to a

semivariance value with h very close to zero (think duplicate

measurements). Besides these two, gstat has the options

of weights N_h, and of no (=constant) weights.

--

Edzer

[*] If I'm correct, this was first suggested in a paper by

Zhang and ... in Computers & Geosciences, early nineties.

I strongly disliked it then, but consider it acquired taste.