[R-sig-Geo] What's the difference between (universal) kriging and spatial autoregressive models?

[Guido Schulz]

Dear list,

as part of a course on missing observations in social/survey statistics
I am trying to explore existing methods of predicting either point
pattern or polygon data.

As far as I have seen, there is no particular package that includes
missing spatial data and imputation for spatial data methods, so I am
relying on prediction and interpolation.

I got quite confused by all the different terms and formalisations used
in geostatistics (more on the kriging side?) on the one hand and spatial
statistics/econometrics (more on the regression side?) on the other.

So far, I understood that both spatial autoregression and kriging

- seek spatial prediction,

- rely on a variance-covariance (or variogram) Matrix - in other words,
they rely on spatial autocorrelation,

- and, as far as universal kriging is concerned, assume an underlying
trend plus random residuals.

However, kriging seems to be used primarily for point patterns with
continuous regions while spatial autoregression involves aggregations of
phenomena into polygons.

So here are my questions:

- Is the underlying spatial structure the only difference between the two?

- Are there equivalent mathematical model formulations for universal
kriging and spatial autoregression?

- Would area-to-area universal kriging be exactly the same thing as
prediction models using spatial error autoregression?

- Is the "constrainedKriging" package the only one that supports
spatialpolygondataframe-to-spatialpolygondataframe kriging?


Best,

Guido