In fact, many models are possible and they vary in how they relate the correlation between residuals to the distance between observations: the relationship may be linear or non-linear, one- or two-dimensional, involve a moving mean or not, and so forth. ASReml provides a wide array of possible models that can be fit to the data
(Table 7.6 of Gilmour et al., 2014). In practice, we most often just fit autoregressive first-order (AR1) correlation structures to field data, as this model has proven robust across many different data sets. This is not to say that errors are really correlated in an AR1 fashion, but it means that the AR1 model is a useful approximation to a wide variety of error effect distributions.
In fact, many models are possible and they vary in how they relate the correlation between residuals to the distance between observations: the relationship may be linear or non-linear, one- or two-dimensional, involve a moving mean or not, and so forth. ASReml provides a wide array of possible models that can be fit to the data
(Table 7.6 of Gilmour et al., 2014). In practice, we most often just fit autoregressive first-order (AR1) correlation structures to field data, as this model has proven robust across many different data sets. This is not to say that errors are really correlated in an AR1 fashion, but it means that the AR1 model is a useful approximation to a wide variety of error effect distributions.
正在翻譯中..