Independence of observations is one of the basic assumptions of the analysis of variance. Performed randomizations prevent results from being biased in cases when independence is violated. The objective of the present paper is to find out the predominant shape of spatial relationships in Polish wheat variety testing trials. One of the possibilities is to apply some geo-statistical method (Cressie Noel 1993, Grondona andCressie 1991). Such an approach is used in this paper. Using the results of nearly 200 trials on wheat varieties, conducted either in generalized lattice (GL) designs (Patterson and Hunter 1983) or in incomplete split-block designs, the empirical values of semivariance have been calculated. Residuals were computed from a model with fixed effects for varieties and replicates, ignoring incomplete blocks, then 10 different geo-statistical models have been fitted to empirical semivariances. With spatial models convergence problems occurred in some cases. The linear and bounded-linear models were the ones most often successfully fitted. Inclusion of revealed relationships into classical models for incomplete blocks did not improve substantially the effectiveness of analyses.