SUMMARYThe present paper compares standard and novel methods for analysing aggregated patterns of plant death in designed field experiments; these methods include binomial (BN), beta-binomial (BBN), logistic-normal-binomial (LNB), BN models with random blocks, BN models with smooth-scale spatial components and principal coordinates of neighbour matrices (PCNM). PCNM is a relatively new technique used in ecology to determine how much observed variability can be explained by spatial and environmental variables, and has not yet been applied to agricultural studies. The survival data of two pasture species, collected from a designed field experiment that was replicated at multiple locations, were used. First, the occurrence of overdispersion was tested using the BN and BBN distributions. Goodness-of-fit tests proved that the BBN model provided a better description (better fit) of the observed data in some cases than did the BN distribution, indicating overdispersion was present. When overdispersion was not present, the BN distribution was adequate to describe the data, and the use of the BBN distribution was superfluous. It is then shown that the PCNM approach, the BN model with smooth-scale spatial components and the LNB model were able to account for some of the variation as spatial variability, thus reducing the species effect compared with that explained under the standard BN model. The amount of variation among species according to the BN model and the BN model with random blocks was similar. Therefore, it is argued that the novel PCNM approach warrants further testing when exploring the spatial variability in designed experiments in agriculture and using LNB, PCNM and BN with smooth-scale spatial components may provide better predictions of species effects than do other, more conventional, approaches.
Goodness-of-Fit Tests for Censored Survival Data
