Appropriate analysis of plant breeding trials is critical for the accurate assessment of test lines and selection decisions. The objectives of this study were two-fold: firstly, to examine the performance of two-dimensional spatial models based on the first order separable autoregressive process in comparison with randomised complete block (RCB) and randomisation based (RB) models in analysis of cotton breeding trials; secondly, to understand the presence and forms of spatial variations and their association with field layout. The different models were first used to analyse a lint yield dataset from the CSIRO cotton breeding program, which consisted of 96 trials under furrow-irrigated conditions from 1995 to 2002 and Residual Maximum Likelihood ratio test and the Akaike Information Criterion were used to identify adequate model (i.e. dataset-preferred model) for individual datasets. The spatial models fitted 62 trials adequately and outperformed the RB model (31) with the worse being RCB model (3). Spatial variations in various forms were commonly present in trials in which spatial models were adequate, and was dominant in planting row direction. Layouts with more plots in dimensional directions tended to have a higher level of spatial variation. Spatial models offered about 176 % mean relative efficiency over RCB, which was comparable with that achieved by the dataset-preferred models but about 20 % higher than the RB model. Therefore, a routine use of spatial analysis in conjunction with efficient trial designs would mitigate the impact of spatial variations on the yield estimate of cotton breeding trials and improve the accuracy of selection. © 2014 Springer Science+Business Media Dordrecht.