Three forms of spatial analysis were compared to the analysis of the normal Gauss-Markov model in genetical experiments, supposing progeny effects as random: moving averages on raw data (MM), moving averages on residual data (Papadakis - PPD), and spatial analysis modelling with residual covariances (AE). The local control information was initially ignored to test the effectiveness of spatial analysis. Thereafter, the different forms of spatial analysis were applied to the complete model, considering the local control of lattice design. The average values of proportions between estimates of variance components and of heritability were used as a discussion guide to determine the best form of analysis. Results showed that ignoring the experimental design, using spatial information was not effective, in general. In average, MM and PPD improved the original model justified by design, in contrast to AE. AE, although ineffective, did not change variance component estimates and heritability. This property guarantees that the combination of random effects for progenies and AE does not violate the assumptions (some of these justified by the design). This is specially useful in large experiments, with a huge number of progenies.
statistical analysis; moving average; Papadakis
