Covariance Structures

Here we describe the covariance structures which are currently available in mmrm.

Introduction

We use some concepts throughout the different covariance structures and introduce these here.

Covariance and Correlation Matrices

The symmetric and positive definite covariance matrix $$ \Sigma = \begin{pmatrix} \sigma_1^2 & \sigma_{12} & \dots & \dots & \sigma_{1m} \\ \sigma_{21} & \sigma_2^2 & \sigma_{23} & \dots & \sigma_{2m}\\ \vdots & & \ddots & & \vdots \\ \vdots & & & \ddots & \vdots \\ \sigma_{m1} & \dots & \dots & \sigma_{m,m-1} & \sigma_m^2 \end{pmatrix} $$ is parametrized by a vector of variance parameters θ = (θ1, …, θk)⊤. The meaning and the number (k) of variance parameters is different for each covariance structure.

In many covariance structures we use the decomposition Σ = DPD where the diagonal standard deviation matrix is $$ D = \begin{pmatrix} \sigma_1 & 0 & \cdots & & 0 \\ 0 & \sigma_2 & 0 & \cdots & 0 \\ \vdots & & \ddots & \ddots & \vdots \\ & & & \sigma_{m-1} & 0 \\ 0 & \cdots & & 0 & \sigma_m \end{pmatrix} $$ with entries σi > 0, and the symmetric correlation matrix P is $$ P = \begin{pmatrix} 1 & \rho_{12} & \rho_{13} & \cdots & \cdots & \rho_{1,m-1} \\ \rho_{12} & 1 & \rho_{23} & \ddots & & \vdots \\ \rho_{13} & \rho_{23} & \ddots & \ddots & \ddots & \vdots \\ \vdots & \ddots & \ddots & \ddots & \rho_{m-2,m-1} & \rho_{m-2,m} \\ \vdots & & \ddots & \rho_{m-2,m-1} & 1 & \rho_{m-1,m} \\ \rho_{1, m-1} & \cdots & \cdots & \rho_{m-2,m} & \rho_{m-1,m} & 1 \end{pmatrix} $$ with entries ρij ∈ (−1, 1). Since these covariance structures assume different variances for each time point they are called “heterogeneous” covariance structures. Assuming a constant σ = σ1 = ⋯ = σm gives a “homogeneous” covariance structure instead.

Transformation to Variance Parameters

For standard deviation parameters σ we use the natural logarithm log (σ) to map them to ℝ. For correlation parameters ρ we consistently use the transformation $$ \theta = f(\rho) = \mathop{\mathrm{sign}}(\rho) \sqrt{\frac{\rho^2}{1 - \rho^2}} $$ which maps the correlation parameter to θ ∈ ℝ. It has the inverse $$ \rho = f^{-1}(\theta) = \frac{\theta}{\sqrt{1 + \theta^2}}. $$ This is important because the resulting variance parameters θ can be optimized without constraints over the whole of ℝ.

Covariance Structures

Unstructured (us)

Any covariance matrix can be represented by this saturated correlation structure. Here k = m(m + 1)/2. See the algorithm vignette for details.

Homogeneous (ad) and Heterogeneous Ante-dependence (adh)

The ante-dependence correlation structure (Gabriel 1962) is useful for balanced designs where the observations are not necessarily equally spaced in time. Here we use an order one ante-dependence model, where the correlation matrix P has elements $$ \rho_{ij} = \prod_{k=i}^{j-1} \rho_k. $$ So we have correlation parameters ρk, k = 1, …, m − 1.

We use a heterogeneous covariance structure to allow for different within subject variances. So here we can identify θ = (log (σ1), …, log (σm), f(ρ1), …, f(ρm − 1)) and we have in total 2m − 1 variance parameters. Assuming a constant variance yields a homogeneous ante-dependence covariance structure with total m variance parameters.

Note our naming convention for the homogeneous and heterogeneous covariance structures that a suffix h is used to denote the heterogeneous version, e.g, ad for homogeneous and adh for heterogeneous ante-dependence. This is different from the name used in SAS for ante-dependence covariance structure, where ANTE(1) refers to heterogeneous ante-dependence covariance structure and a homogeneous version is not provided in SAS.

Homogeneous (toep) and Heterogeneous Toeplitz (toeph)

Toeplitz matrices (Toeplitz 1911) are diagonal-constant matrices. Here we can model the correlation matrix as a Toeplitz matrix: $$ P = \begin{pmatrix} 1 & \rho_1 & \rho_2 & \cdots & \cdots & \rho_{m-1} \\ \rho_1 & 1 & \rho_1 & \ddots & & \vdots \\ \rho_2 & \rho_1 & \ddots & \ddots & \ddots & \vdots \\ \vdots & \ddots & \ddots & \ddots & \rho_1 & \rho_2 \\ \vdots & & \ddots & \rho_1 & 1 & \rho_1 \\ \rho_{m-1} & \cdots & \cdots & \rho_2 & \rho_1 & 1 \end{pmatrix} $$ This means that the correlation between two time points only depends on the distance between them, i.e. ρij = ρ|i − j| and we have correlation parameters ρk, k = 1, …, m − 1.

We use a heterogeneous covariance structure to allow for different within subject variances. So here we can identify θ = (log (σ1), …, log (σm), f(ρ1), …, f(ρm − 1)) and we have in total 2m − 1 variance parameters. This is similar to the heterogeneous ante-dependence structure, but the correlation parameters are used differently in the construction of P. Assuming a constant variance yields a homogeneous Toeplitz covariance structure with total m variance parameters.

Homogeneous (ar1) and Heterogeneous (ar1h) Autoregressive

The autoregressive covariance structure can be motivated by the corresponding state-space equation yit = φyi, t − 1 + εt where the white noise εt has a normal distribution with mean zero and a constant variance. It can be shown that this gives correlations ρij = ρ|i − j|. where ρ is related to φ and the variance and is the single correlation parameter here.

Assuming a constant variance in the state-space equation yields a homogeneous autoregressive covariance structure with total only k = 2 variance parameters, otherwise we obtain the heterogeneous autoregressive covariance structure with k = m + 1 variance parameters.

Homogeneous (cs) and Heterogeneous (csh) Compound Symmetry

The compound symmetry covariance structures assume a constant correlation between different time points: ρij = ρ where ρ is the single correlation parameter here.

Assuming a constant variance in the state-space equation yields a homogeneous compound symmetry covariance structure with total only k = 2 variance parameters, otherwise we obtain the heterogeneous compound symmetry covariance structure with k = m + 1 variance parameters.

Spatial Covariance Structure

Spatial covariance structures, unlike other covariance structures, does not require that the timepoints are consistent between subjects. Instead, as long as the distance between visits can be quantified in terms of time and/or other coordinates, the spatial covariance structure can be applied. Euclidean distance is the most common case. For each subject, the covariance structure can be different. Only homogeneous structures are allowed (i.e. a common variance is used).

Please note that while printing the summary of an mmrm fit, the covariance displayed is a 2 * 2 square matrix. As the distance will be used to derive the corresponding element in that matrix, unit distance is used here. The distance matrix is

$$ \begin{pmatrix} 0 & 1\\ 1 & 0\\ \end{pmatrix} $$

Spatial exponential (sp_exp)

For spatial exponential, the covariance structure is defined as follows:

ρij = ρdij

where dij is the distance between time point i and time point j,

A total number of parameters k = 2 is needed:

The parameterization for θ is a little different from previous examples. In previous examples, ρ can take values from -1 to 1, but here we need to restrict ρ to (0, 1). Hence we have the following parametrization form.

θ = (log (σ), logit (ρ))

References

Gabriel KR (1962). “Ante-dependence Analysis of an Ordered Set of Variables.” The Annals of Mathematical Statistics, 33(1), 201–212. https://doi.org/10.1214/aoms/1177704724.
Toeplitz O (1911). “Zur Theorie Der Quadratischen Und Bilinearen Formen von Unendlichvielen Veränderlichen.” Mathematische Annalen, 70(3), 351–376. https://doi.org/10.1007/BF01564502.