A method for assessing the goodness of computer simulation of soil processes

Any satisfactory computer simulation model of a soil process must match actual behaviour in the laboratory or field; a model can be evaluated by how well it does so. This paper describes a method for assessing models using anion diffusion and nitrate leaching as examples. The method partitions the sum of squares of the differences between measurement and simulation into two components, one calculated from the differences between the simulation and the mean of replicate measurements (the 'lack of fit'), and the other calculated from the variance within each set of replicate measurements (the 'pure error'). If the former is not significantly larger than the latter than the data present no grounds for rejecting the model. Where a model simulates the change in a process with time the method can also take account of how experimental error in the initial measurements affects the goodness of fit of the simulation of subsequent measurements. The method is particularly valuable where it is difficult or costly to take many replicate measurements, such as often happens in soil science or agriculture; nonetheless, some replicates must be taken.