Geometric Anisotropy in Spatial Effects

Why Direction Matters

An isotropic spatial effect has the same covariance between two locations whenever their separation distance is the same, regardless of direction. That can be restrictive. For example, a latent spatial pattern might vary slowly along a valley but change more rapidly across it. Two locations equally far from a third can then have different latent-effect correlations because they lie in different directions.

Geometric anisotropy represents this pattern by rotating and rescaling coordinates, then applying an isotropic covariance function to the transformed locations. The covariance model keeps a single correlation function, but its distance measure accounts for a preferred direction and a longer range along that direction. In a GLMM, this structure can be assigned to a Gaussian spatial latent effect; the resulting dependence is on the latent-effect scale, not automatically a statement about equal correlations between observed binary or count responses.

Definition: Let a location be $s=(x,y)$, with coordinates in a common planar coordinate system. Let $\theta$ be the angle, measured counterclockwise from the positive $x$-axis, of the direction with the longer spatial range. Let $\rho\geq 1$ be the ratio of the longer range to the shorter range. For a displacement $h=(h_x,h_y)$, define $$ h_{\parallel}=h_x\cos\theta+h_y\sin\theta,\qquad h_{\perp}=-h_x\sin\theta+h_y\cos\theta, $$ and the anisotropic effective distance $$ d_A(h)=\sqrt{\left(\frac{h_{\parallel}}{\rho}\right)^2+h_{\perp}^2}. $$ The longer-range direction is called the major direction; the perpendicular, shorter-range direction is the minor direction. When $\rho=1$, the model is isotropic.

From Transformed Coordinates to Covariance

The same definition can be written as a coordinate transformation. For each location $s_i=(x_i,y_i)$, set $$ z_i=T s_i,\qquad T= \begin{pmatrix} \cos\theta/\rho & \sin\theta/\rho\\ -\sin\theta & \cos\theta \end{pmatrix}. $$ Then $d_A(s_i-s_j)=\lVert z_i-z_j\rVert$, the ordinary Euclidean distance between transformed locations. This representation makes the geometry clear: first rotate the axes to align with the major and minor directions, then shrink the major-direction coordinate by $1/\rho$. A larger original separation along the major direction is therefore needed to produce the same effective distance as a smaller separation along the minor direction.

Let $u(s)$ be a mean-zero stationary Gaussian spatial effect with marginal variance $\sigma_u^2$. Applying an isotropic correlation function $r(d)$ to effective distance gives $$ \operatorname{Cov}\{u(s_i),u(s_j)\}=\sigma_u^2 r\{d_A(s_i-s_j)\}. $$ For example, the exponential correlation is $r(d)=\exp(-d/\phi)$, where $\phi>0$ is a distance-scale parameter. At effective distance $d=\phi$, correlation is $e^{-1}$. The Gaussian correlation is $r(d)=\exp\{-(d/\phi)^2\}$. Other valid isotropic correlation functions can also be used with transformed distance; the transformation does not require a particular one.

If independent site-specific noise is included, a covariance matrix can be written $$ \operatorname{Cov}(w_i,w_j)=\sigma_u^2 r\{d_A(s_i-s_j)\}+\tau^2 I(i=j), $$ where $w_i=u(s_i)+e_i$, the independent noise terms $e_i$ have variance $\tau^2$, and $I(i=j)$ is one when $i=j$ and zero otherwise. This diagonal contribution is distinct from directional range: a nugget changes the variance at the observation or site level, whereas anisotropy changes covariance with direction and separation.

Worked Example: Effective Distance at a Specified Direction

Suppose hypothetical locations are modeled with an exponential spatial correlation, $\theta=30^\circ$, and $\rho=2$. Consider two locations separated by $h=(2,0)$ coordinate units. Let $\phi=3$ coordinate units and $\sigma_u^2=1.5$. Calculate their effective distance and latent-effect covariance, and compare with the isotropic calculation using the same $\phi$.

The components along the major and minor directions are $$ h_{\parallel}=2\cos(30^\circ)+0\sin(30^\circ)=1.732,\qquad h_{\perp}=-2\sin(30^\circ)+0\cos(30^\circ)=-1.000. $$ Thus $$ d_A=\sqrt{(1.732/2)^2+(-1.000)^2} =\sqrt{0.750+1.000} =1.323\text{ coordinate units}. $$ The exponential correlation and covariance are $$ r(d_A)=\exp(-1.323/3)=\exp(-0.441)=0.643,\qquad \operatorname{Cov}=1.5(0.643)=0.965. $$ Using ordinary isotropic distance instead would give $d=2$ and covariance $1.5\exp(-2/3)=0.770$. The larger anisotropic covariance is consistent with this displacement having a component in the longer-range direction. The coordinates must be in a suitable planar system for these distances and units to have this interpretation.

Why the Transformed Model Is Valid

A spatial covariance function must produce a positive semidefinite covariance matrix for every finite set of locations. Merely drawing elliptical contours does not establish that property. The coordinate-transformation construction does: if $r(\lVert z_i-z_j\rVert)$ is a valid isotropic correlation function in the transformed coordinates, then using it at $z_i=T s_i$ gives a valid covariance matrix in the original coordinates.

The justification is direct. For any finite set of original locations $s_1,\ldots,s_n$, transformation by $T$ produces a finite set $z_1,\ldots,z_n$. By validity of the isotropic covariance function, the matrix $R$ with entries $R_{ij}=r(\lVert z_i-z_j\rVert)$ is positive semidefinite. Multiplication by the nonnegative variance $\sigma_u^2$ preserves this property. Adding an independent nugget adds $\tau^2 I$, also positive semidefinite for $\tau^2\geq0$. Therefore the resulting covariance matrix is valid. The proof relies on starting with a valid isotropic correlation function; it does not make an arbitrary proposed directional formula valid.

Worked Example: Comparing Ranges in the Two Principal Directions

Consider a hypothetical exponential model with $\rho=3$, $\theta=45^\circ$, $\phi=5$ km, and $\sigma_u^2=4$. Compare two pairs of locations: one separated by 15 km in the major direction and another separated by 5 km in the minor direction.

For the 15 km major-direction displacement, $h_{\parallel}=15$ and $h_{\perp}=0$. Hence $$ d_A=\sqrt{(15/3)^2+0^2}=5\text{ km},\qquad \operatorname{Cov}=4\exp(-5/5)=4e^{-1}=1.472. $$ For the 5 km minor-direction displacement, $h_{\parallel}=0$ and $h_{\perp}=5$, so $$ d_A=\sqrt{0^2+5^2}=5\text{ km},\qquad \operatorname{Cov}=4\exp(-5/5)=1.472. $$ The two separations have equal effective distance and therefore equal covariance. In physical coordinates, the major-direction separation is three times as large, matching the specified range ratio. Since the exponential correlation at effective distance 5 km is $e^{-1}$, these calculations compare covariance at that model scale; they do not define a universal practical-range convention.

Worked Example: A Small Valid Covariance Matrix

Take a hypothetical exponential spatial effect with $\theta=0^\circ$, $\rho=2$, $\phi=1$ coordinate unit, and $\sigma_u^2=1$, with no nugget. Use locations $s_1=(0,0)$, $s_2=(2,0)$, and $s_3=(0,1)$.

At this orientation, $d_A(h)=\sqrt{(h_x/2)^2+h_y^2}$. The effective distances for pairs $(1,2)$, $(1,3)$, and $(2,3)$ are respectively $1$, $1$, and $\sqrt{2}$. Their correlations are $a=e^{-1}=0.3679$, $a=0.3679$, and $b=e^{-\sqrt{2}}=0.2431$. The covariance matrix is therefore $$ \Sigma= \begin{pmatrix} 1 & 0.3679 & 0.3679\\ 0.3679 & 1 & 0.2431\\ 0.3679 & 0.2431 & 1 \end{pmatrix}. $$ For a numerical check, its first two leading principal minors are $1$ and $1-a^2=0.8647$, and its determinant is $$ 1+2a^2b-2a^2-b^2 \approx 1+0.0658-0.2707-0.0591 =0.7360. $$ All are positive, so this particular matrix is positive definite. The transformation argument above, rather than checking a single matrix, is what guarantees validity for every finite set of locations.

Assumptions, Diagnostics, and Alternatives

The model makes substantive and computational assumptions. Locations are represented in a common, meaningful planar coordinate system, and the same global orientation and range ratio apply throughout the modeled region. The spatial effect is stationary in transformed coordinates: its covariance depends on the displacement through $d_A$, not on the absolute position. Its transformed-coordinate correlation function must be valid, and the covariance parameters must be in their admissible ranges, including $\rho\geq1$ and $\phi>0$. In a GLMM, the Gaussian latent-effect assumption and the conditional response model are additional modeling choices; geometric anisotropy specifies the latent spatial covariance rather than the response distribution.

The practical alternative is an isotropic covariance model, which sets $\rho=1$. It is simpler and can be preferable when the data provide little evidence about orientation or unequal ranges. Geometric anisotropy uses a global ellipse: it preserves the form of the chosen isotropic correlation function after coordinate transformation. A more flexible model could allow directional behavior that cannot be represented by one orientation and one range ratio, but such added flexibility requires enough spatial information to estimate it. An adjacency-based spatial effect represents spatial relationships through a neighborhood structure rather than continuous transformed distances; it answers a different modeling question and is not interchangeable with this construction.

Useful diagnostics include plotting locations and fitted covariance contours, comparing empirical residual dependence across directional bands, and checking whether residual spatial structure remains after fitting. Directional comparisons should account for distance as well as angle: apparent directional differences can arise simply because one direction has more pairs at short separations. Inspect parameter uncertainty and sensitivity to plausible coordinate systems, starting values, and covariance choices. In a GLMM, residuals are not direct observations of the latent effect, so diagnostic summaries must respect the fitted response model rather than treat estimated latent effects as error-free data.

Anisotropy can be weakly identified when locations are sparse, concentrated along a narrow band, or distributed in only a few directions. The orientation can be poorly determined when $\rho$ is near one, because the major direction then has little practical meaning. Estimation of $\theta$, $\rho$, $\phi$, and the marginal variance can also be confounded, particularly when there are few short- and long-distance pairs. A fitted ellipse should not be interpreted as evidence of a physical mechanism or as a causal effect; it summarizes modeled dependence in the observed spatial domain.

Common Pitfalls

Using the angle convention inconsistently. If $\theta$ describes the minor rather than the major direction, or uses a different rotation convention, the transformed distances change. State which axis the angle identifies and use the same convention throughout.
Applying the range ratio in the wrong direction. Dividing the major component by $\rho$ makes equal effective distance require a longer physical separation along the major direction. Multiplying it by $\rho$ would reverse the intended geometry.
Confusing a nugget with anisotropy. A nugget changes diagonal variance; anisotropy changes off-diagonal covariance according to direction and separation.
Assuming an ellipse proves validity. Validity follows from a valid isotropic covariance applied to linearly transformed coordinates, not from the visual shape of a contour alone.
Ignoring coordinate units or projection. Distances in longitude and latitude are not uniform physical distances. An unsuitable coordinate system can distort both estimated ranges and orientation.

Key takeaway: Geometric anisotropy is isotropic covariance evaluated after a linear transformation of spatial coordinates. The angle sets the major direction, the ratio $\rho$ sets how much longer the range is in that direction, and the transformed-distance construction preserves covariance validity when the starting correlation function is valid.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

If $\rho=1$, what happens to effective distance, and what does that imply about the model?
A model has $\theta=0^\circ$, $\rho=4$, and a displacement of $(8,0)$. Calculate its effective distance.
For the same model, calculate effective distance for displacement $(0,2)$. Which displacement has the larger effective distance?
Why does transforming coordinates preserve covariance validity when the original isotropic correlation function is valid?
Give one reason the major-axis orientation may be difficult to estimate from data.

Solutions. (1) With $\rho=1$, $d_A=\sqrt{h_{\parallel}^2+h_{\perp}^2}=\sqrt{h_x^2+h_y^2}$, so the distance is ordinary Euclidean distance and the model is isotropic. (2) The major direction is the $x$-axis, so $d_A=\sqrt{(8/4)^2+0^2}=2$ coordinate units. (3) The displacement lies along the minor direction, giving $d_A=\sqrt{0^2+2^2}=2$ units. The two displacements have equal effective distance, as expected for a four-to-one range ratio. (4) Any finite set of original locations maps to a finite set of transformed locations; the valid isotropic covariance produces a positive semidefinite matrix for those transformed locations, which is the same covariance matrix indexed by the original locations. (5) For example, if $\rho$ is close to one, the model is nearly isotropic, so many orientations imply nearly the same covariance and the angle is weakly identified.