If you think there is any spatial autocorrelation it is better to check that in a fixed level of predictor on your response variable. To do this, you can take samples with different distances in a line (the simplest method) and check the correlation between the plot distance and change in your response variable. If you find significant correlation then you should detrend your data for your predictor effect.

But, keep this in mind that the spatial autocorrelation decreases with increasing distance between plots logarithmically and you should find the inflection point to remove spatial autocorrelation practically rather than mathematically.

Cheers

thanks a lot for your recommendations! especially concerning the logarythmic scale.

however, if the predictor value does not change with distance, this would mean lack of spatial structure determined by the given predictor. Then, wouldn't be better to change predictor, than to perform complicated analysis to remove autocorrelation?

If you can change it you'd better do this way. First select your sites, weather near or far, then think of any predictor that you can use. After that, choose the most effective, non-correlated predictors to spatial changes and test their spatial autocorrelation and again, test their within group variance.

In last step, choose the one with lower spatial autocorrelation and lower within group variance (in regression called lack of the fit in some way).

I think what your are getting at is that you can have spatial autocorrelation for some variables and not for others, depending on the spatial scale in question. So, for example, imagine your response was tree height and one of your predictors was the slope of the terrain where the trees were growing. If you sampled in an area that is relatively flat with regularly spaced ridges, you might see a pattern of spatial autocorrelation between the slope predictor at the scale of the distance between ridges. Similarly, you may see spatial autocorrelation between the response, tree height, in the flat areas at a smaller spatial scale. You could resample your data to remove the spatial autocorrelation based on a semivariogram. Semivariiograms can be easily calculated in R using the "variog" commend in the "geoR" package.

Ok, Peter, I understand. If the scale is to small I shall miss the presence of regular ridges, if it is too big I shall miss flat areas. So, checking for spatial autocorrelation is in fact answering the question "how good is my sampling scheme", but in case if I use stratified random sampling this question probably wouldn't be so topical...