It is not as simple as "which geostatistical model is best". There is a spatial analysis add-on for ARCGIS10 and it includes some tutorial material. Also see the gstat package in R. I think you will have to learn something about geostatistics first. There are several good books, see the one by P. Goovaerts. Most geostatistical software will assume that the coordinates are given in terms of linear units, i.,e., feet, meters, etc using latitude and longitude, i,e. degrees minutes and seconds will not work, you need UTM coordinates. ARCGIS can handle this for you

Addendum I suspect that you are using the term "resolution" in two different senses. Rain gages generate "point data", whereas the climate model generates non-point data. In geostatistics these correspond to the "supports" of the data. Your problem is somewhat analogous to using rain gage data to calibrate doppler radar rain data, most authors have used cokriging for this. This is further evidence that it is not just a question of the "best geostatistical model". The other responses have not dealt with this aspect of the problem. Do a search on google for cokriging, non-point support data. In the case of the climate model as the resolution increases the spatial correlation decreases.

In my opinion the correct way to predict area averaged values (grid values) from point measurements with geostatistics is using "block kriging", not co-kriging as proposed by Donald. Co-kriging is useful when you are trying to predict a primary variable, for which you don't have a lot of spatial measurements (ex: rain gauges), using the information given by a more densely sampled secondary variable (ex: terrain height) that is correlated with the primary variable. Such information is contained in the variogram of the secondary variable and the cross-variogram between the primary and secondary variable. However, if terrain height is measured on a regular grid (ex: a digital elevation model), the kriging with external drift would be more adapted.

The observed re-analysis data that you are using is already interpolated by some sort of geostatistical method. If you cannot access the original meteorological stations data, you probably have to upscale/regrid the 0.3 degrees resolution re-analysis and the 0.44 degrees resolution RCM to a common grid. I don't know how to obtain grid average estimates from grid area "measurements" (instead of point measurements) using geostatistics. Depending on the required level of accuracy for your analyses, a more pragmatic (and time saving) re-gridding of the re-analysis to the RCM grid (or the contrary) may be sufficient.

The most pedagogic book I know to start with geostatistics is Isaaks, Srivastava: Applied Geostatistics.

It is not a question of "block kriging" vs "cokriging" since there is block cokriging. The question is whether the data has point support or non-point support, or more generally if the data is mixed, i.e. the data for one or more attributes is point support and the data for one or more other attributes is non-point support.

Thanks for reminding me about the block co-kriging. I guess the main problem is that both the re-analysis of point meteorological stations and the RCM model output are areal estimates (non-point support) at quite similar resolution. Therefore, I am trying to understand whether a geostatistical approach is motivated in terms of efforts or if a more pragmatic approach involving the re-gridding of data would be sufficient. This depends on the scientific goals and by which application the comparison between the two datasets is needed. If it is only a graphical exploration to depict global rainfall biases, a re-gridding would be faster and easier.