The Matérn class4.4 Variance model functions 46asreml uses an extended Matérn class which accomodates geometric anisotropy and achoice of metrics for random fields observed in two dimensions. This extension, describedin detail in Haskard [2006], is given byρ(h; φ) = ρ M (d(h; δ, α, λ); φ, ν)where h = (h x , h y ) T is the spatial separation vector, (δ, α) governs geometric anisotropy,(λ) specifies the choice of metric and (φ, ν) are the parameters of the Matérn correlationfunction. The function isρ M (d; φ, ν) = { 2 ν−1 Γ(ν) } ( ) ν (−1 d dK ν , (4.1)φ φ)estimationwhere φ > 0 is a range parameter, ν > 0 is a smoothness parameter, Γ(·) is the gammafunction, K ν (.) is the modified Bessel function of the third kind of order ν (Abramowitzand Stegun, 1965, section 9.6) and d is the distance defined in terms of X and Y axes:h x = x i − x j ; h y = y i − y j ; s x = cos(α)h x + sin(α)h y ; s y = cos(α)h x − sin(α)h y ;d = (δ|s x | λ + |s y | λ /δ) 1/λ .For a given ν, the range parameter φ affects the rate of decay of ρ(·) with increasing d.The parameter ν > 0 controls the analytic smoothness of the underlying process u s , theprocess being ⌈ν⌉ − 1 times mean-square differentiable, where ⌈ν⌉ is the smallest integergreater than or equal to ν (Stein, 1999, page 31). Larger ν correspond to smootherprocesses. asreml uses numerical derivatives for ν when its current value is outside theinterval [0.2,5].When ν = m + 1 2 with m a non-negative integer, ρ M (·) is the product of exp(−d/φ) anda polynomial of degree m in d. Thus ν = 1 2yields the exponential correlation function,ρ M (d; φ, 1 2) = exp(−d/φ), and ν = 1 yields Whittle’s elementary correlation function,ρ M (d; φ, 1) = (d/φ)K 1 (d/φ) (Webster and Oliver, 2001).When ν = 1.5 thenρ M (d; φ, 1.5) = exp(−d/φ)(1 + d/φ)which is the correlation function of a random field which is continuous and once differentiable.This has been used recently by Kammann and Wand [2003]. As ν → ∞ thenρ M (·) tends to the gaussian correlation function.The metric parameter λ is not estimated by asreml; it is usually set to 2 for Euclideandistance. Setting λ = 1 provides the cityblock metric, which together with ν = 0.5 modelsa separable AR1timesAR1 process. Cityblock metric may be appropriate when the dominantspatial processes are alighned with rows/columns as occurs in field experiments.Geometric anisotropy is discussed in most geostatistical books [Webster and Oliver, 2001,Diggle et al., 2003] but rarely are the anisotropy angle or ratio estimated from the data.Similarly the smoothness parameter ν is often set a-priori [Kammann and Wand, 2003,Diggle et al., 2003]. However Stein [1999] and Haskard et al. [2005] demonstrate that νcan be reliably estimated even for modest sized data-sets, subject to caveats regardingthe sampling design.EstimationThe order of the parameters in mtrn(), with their defaults, is (φ, ν = 0.5, δ = 1, α =0, λ = 2). Parameters are fixed or estimated depending on the data type (numeric orcharacter) of the argument to the respective parameter.• If an argument is numeric, it is treated as a starting value for estimation and giventhe constraint code P (positive).
4.4 Variance model functions 47• This behaviour can be altered by concatenating the numeric value followed by theconstraint code (P, U or F) into a character string.• If an argument is absent from the call, the corresponding parameter is held fixed atits default value.For example, to fit a Matérn model with only φ estimated and the other parameters set attheir defaults then we could use mtrn(phi = 0.1) where the starting value for estimationis given as 0.1.To fix ν some value other than the default and estimate φ, the fixed value and constraintcode are given as a single string to the nu argument. That is mtrn(phi = 0.1, nu = ”1.0F”)The parameters φ and ν are highly correlated so it may be better to <strong>manual</strong>ly cover agrid of ν values.We note that there is non-uniqueness in the anisotropy parameters of this metric d(·)since inverting δ and adding π 2to α gives the same distance. This non-uniqueness can beremoved by constraining 0 ≤ α < π 2and δ > 0, or by constraining 0 ≤ α < π and either0 < δ ≤ 1 or δ ≥ 1. With λ = 2, isotropy occurs when δ = 1, and then the rotation angleα is irrelevant: correlation contours are circles, compared with ellipses in general. Withλ = 1, correlation contours are diamonds.4.4.4 General structure modelscor(obj, init=NA)corb(obj, k=1, init=NA)corg(obj, init=NA)diag(obj, init=NA)us(obj, init=NA)chol(obj, k=1, init=NA)cholc(obj, k=1, init=NA)ante(obj, k=1, init=NA)fa(obj, k=1, init=NA)DescriptionThe class of general variance models includes the simple, banded and general correlationmodels (cor, corb, corg), the diagonal, unstructured, Cholesky and antedependencevariance models (diag, us, chol, cholc, ante) and the factor analytic structure (fa).Required argumentsobja factor in the data frame.Optional argumentsinita vector of initial parameter values. This vector can have an optional namesattribute to set the boundary constraint for each parameter. In this case, thename of each element may be one of ”P”, ”U” or ”F” for positive, unconstrainedor fixed, respectively.