Bivariate generalized linear mixed model structure

tinyVAST is a bivariate extension of a generalized linear mixed model (see Tables 1 and 2 for notation), which includes two linear predictors that each include four additive components:

1. Spatial interactions among variables: The user can specify interactions among variables at a given site (or for spatially correlated latent variables) using arrow notation derived from path analysis, based on the interface from R package sem;

2. Temporal interaction among variables: The user can specify simultaneous and lagged interactions among variables over time using an expanded arrow-and-lag notation that is derived from R package dsem, where these interactions the annual intercept for a given variable therefore apply uniformly for all locations.

3. Spatio-temporal interactions among variables: The user can specify simultaneous and lagged interactions among variables over time, where these interactions occur on a site-by-site basis.

4. Generalized additive model: The user specifies a formula for a generalized additive model (GAM) that is interpreted by R package mgcv. If other model components are missing, tinyVAST estimates parameters that are similar to mgcv, with small differences resulting from different methods for parameter estimation;

These four components are assembled into two linear predictors:

$$ \begin{aligned} p_{\mathrm 1,i} &= \underbrace{\mathbf X_{\mathrm 1,i} \mathbf\alpha_{\mathrm 1} + \mathbf Z_{\mathrm 1,i} \mathbf\gamma_{\mathrm 1}}_\text{GAM} + \underbrace{\mathbf A_{i} \mathbf\Omega_{\mathrm 1,c[i]}}_\text{space_term} + \underbrace{\mathbf D_{\mathrm 1,c[i],t[i]}}_\text{time_term} + \underbrace{\mathbf A_i \mathbf E_{\mathrm 1,c[i],t[i]}}_\text{spacetime_term} \\ p_{\mathrm 2,i} &= \underbrace{\mathbf X_{\mathrm 2,i} \mathbf\alpha_{\mathrm 2} + \mathbf Z_{\mathrm 2,i} \mathbf\gamma_{\mathrm 2}}_\text{GAM} + \underbrace{\mathbf A_{i} \mathbf\Omega_{\mathrm 2,c[i]}}_\text{space_term} + \underbrace{\mathbf D_{\mathrm 2,c[i],t[i]}}_\text{time_term} + \underbrace{\mathbf A_i \mathbf E_{\mathrm 2,c[i],t[i]}}_\text{spacetime_term} \end{aligned} $$

where

• p_1, i is the first linear predictor;
• X_1, iα₁ + Z_1, iγ₁ is the GAM component
• α₁ is the GAM fixed effects with associated design matrix X₁, where X_1, i is the row of the design matrix for sample i;
• γ₁ is the GAM random effects with associated design matrix Z, where Z_1, i is the row of the design matrix;
• A_iΩ_1, c[i] is the projection of the space-veriable interaction Ω to sample i, where A is the interpolation matrix with dimension I × S such that row A_i interpolates spatial vector Ω_1, c[i] for sample i and variable c[i];
• D_{1, c[i], t[i]} is the time-variable interaction D for sample i;
• A_iE_{1, c[i], t[i]} is the projection of the space-variable-time interaction E to sample i;

and terms are defined similarly for the second linear predictor p_2, i except using subscript-2. The linear predictors are then passed through a bivariate inverse-link function to specify the distribution for errors:

y_i ∼ f_e[i](g_e[i]⁻¹(p_1, i, p_2, i), θ_e[i])

where

• f_e[i] probability density or mass function for sample i;
• g_e[i]⁻¹(p_1, i, p_2, i) is the bivariate inverse-link function that transforms linear predictors to the central tendancy parameters of the distribution;
• θ_e[i] is the dispersion parameters for distribution e[i] for sample i;

In the simple case, the distribution only requires a single linear predictor such that p₂ = 0 by construction and drops out of the model. In this case, the model collapses to a generalized linear mixed model. For example we might have a log-link and Poisson distribution, such that it collapses to a log-linked Poisson GLMM, y_i ∼ Poisson(e^p_1, i).

However, tinyVAST can also handle a delta-model using either logit-log or Poisson-linked link functions:

g_e[i]⁻¹(p_1, i, p_2, i) = (μ_1, i, μ_2, i)

where μ_1, i and μ_2, i are the two linear predictors, such that 𝔼[y_i] = μ_1, iμ_2, i;

For the conventional logit-log bivariate link function we use a logit-link for encounter probabilities and a log-link for positive catch rates:

$$ \begin{aligned} \mu_{\mathrm 1,i} &= \frac{e^{p_{\mathrm 1,i}}}{1+e^{p_{\mathrm 1,i}}} \\ \mu_{\mathrm 2,i} &= e^{p_{\mathrm 2,i}} \end{aligned} $$

while for the Poisson-linked link function we specify a complemetary log-log link for encounter probabilities and define the second link function such that μ_1, iμ_2, i = e^p_1, ie^p_2, i:

$$ \begin{aligned} \mu_{\mathrm 1,i} &= 1 - e^{-e^{p_{\mathrm 1,i}}} \\ \mu_{\mathrm 2,i} &= \frac{e^{p_{\mathrm 1,i}}}{\mu_{\mathrm 1,i}} e^{p_{\mathrm 2,i}} \end{aligned} $$ where e^p_1, i is the density for an underlying point process, and e^p_2, i is biomass per point (i.e., animal or group).

In either the conventional or Poisson-linked delta model, μ_1, i = Pr(Y > 0) is the encounter probability (and is fitted with a Bernoulli distribution), while μ_2, i is the central tendancy for positive values, i.e., μ_2, i = 𝔼(Y|Y > 0).

Spatial domains

Linear predictors include spatially autocorrelated latent variables. These variables are treated as Gaussian Markov random fields (GMRFs), and evaluating the probability density of GMRFs involves calculating the precision matrix Q_domain = Σ⁻¹ as the inverse of the spatial covariance matrix. tinyVAST involves three options for specifying this spatial precision:

• Stochastic partial differentiaul equation (SPDE): The analyst can approximate spatial variation over a continuous two-dimensional surface by constructing finite element mesh (FEM), treating the value at vertices as a GMRF, and then using bilinear interpolation (i.e., a piecewise linear approximation) to interpolate from vertices to the spatial domain. In this case, the precision is constructred as:

Q_domain = τ²(κ⁴M₀ + 2κ²M₁ + M₂) where every row A_i of the interpolation matrix A is nonzero for only the three vertices of the triangle that contains sample i

• Simultaneous autoregressive (SAR): Alternatively, the analyst can specify an areal model representing the value within spatial strata:

Q_domain = τ²(I − κA^*)² where A^* is the adjacency matrix of the graph specified by the analyst, and each row A_i of interpolation matrix A is nonzero only for the single spatial stratum containing sample i (and noting that adjacency matrix A^* is different from interpolation matrix A, but we use the same notation for both due to a collision in notational standards).

• Stream networks: Finally, the analyst can specify that sites are partially correlated if they are adjacent along a stream network (while ignoring flow direction). This results in an Onstein-Uhlenbeck process along the stream network, or an exponential tail-down model.

For each spatial domain, the term τ is defined such that precision Q_domain has unit variance. This is done because the spatial domain always arises in combination with other parameters, which are used to define the variance of the associated spatial variable.

Structural equation models

tinyVAST also involves specifying a structural equation model (SEM). This SEM be viewed either:

1. Weak interpretation: as an expressive interface to parameterize the correlation among variables, using as many or few parameters as might be appropriate; or
2. Strong interpretation: as a structural causal model, allowing predictions about the consequence of counterfactual changes to the system.

To specify a SEM, the user uses arrow notation derived from package sem. For example, to specify a linear model this involves:

w1 -> w2, b_12
w1 <-> w1, sd_1
w2 <-> w2, sd_2

This then estimates a single slope parameter (represented with a one-headed arrow), as well as the variance of W₁ and W₂ (specified with two-headed arrows). In a more complicated case, W₁ might cause W₂, which in turn causes W₃. This is then represented as:

# Path, parameter_name, start value
w1 -> w2, b_12, 0
w2 -> w3, b_23, 0
w1 <-> w1, s_1, 1
w2 <-> w2, s_2, 1
w3 <-> w3, s_3, 1

SEM interactions can be as complicated or simple as desired, and can include:

1. Latent variables and loops (i.e., they are not restricted to directed acyclic graphs);
2. Values that are fixed a priori, where the parameter_name is provided as NA and the starting value that follows is the fixed value;
3. Values that are mirrored among path coefficients, where the same parameter_name is provided for multiple rows of the text file.

In this preceding example, path coefficients for one-headed arrows then define path matrix P:

$$ \mathbf P = \begin{pmatrix} 0 & 0 & 0 \\ b_{12} & 0 & 0 \\ 0 & b_{23} & 0 \end{pmatrix} $$

and coefficents for two-headed arrows define the Cholesky G of the exnogenous covariance matrix G^TG:

$$ \mathbf G = \begin{pmatrix} s_{1} & 0 & 0 \\ 0 & s_{2} & 0 \\ 0 & 0 & s_{3} \end{pmatrix} $$

These matrices are define a simultaneous equation model:

$$ \mathbf{ w = P w + \epsilon} \\ \mathbf\epsilon \sim \mathrm{MVN}( \mathbf 0, \mathbf G^T \mathbf G ) $$ where the variance Var(w) = (I − P)⁻¹G²(I − P^T)⁻¹. This then results in a sparse precision matrix:

Q = (I − P^T)G⁻¹G^−T(I − P)

where this precision matrix Q is then used as a modular component in the larger tinyVAST model.

Dynamic structural equation models

Similarly, tinyVAST involves specifying dynamic structural equation models (DSEM). To specify a DSEM, the user uses arrow-and-lag notation. For example, to specify a univariate first-order autoregressive process:

w1 -> w1, 1, rho, 0.8

If there were four time-intervals (T = 4) this would then result in the path matrix:

$$ \mathbf P = \begin{pmatrix} 0 & 0 & 0 & 0 \\ \rho & 0 & 0 & 0 \\ 0 & \rho & 0 & 0 \\ 0 & 0 & \rho & 0 \end{pmatrix} $$

and when the DSEM involves multiple times and variables, the sparse precision is formed by summing across the Kronecker product of time-lag and interaction matrices. This DSEM defines a GMRF over a nonseparable interaction of time and variables, represented by a matrix Q_{time_term} with dimension CT × CT. The user can specify a separate arrow-and-lag notation to define the precision matrix for the time-variable interaction Q_{time_term} and for the space-time-variable interaction Q_{spacetime_term}.

The precision matrix Q_{time_term} for the time term D is used to define the time-varying intercept for each variable:

vec(D) ∼ MVN(0, Q_{time_term}) Meanwhile, the space-time term Q_{spacetime_term} is combined with the spatial precision Q_{space_term} as we explain next.

Spatial interactions for SEM and DSEM

tinyVAST uses the SEM and DSEM notation to construct the joint precision for the space-variable interaction Ω with dimension S × C, and the space-time-variable interaction E with dimension S × C × T. To do so, it constructs the separable precision for each process:

vec(E) ∼ MVN(0, Q_{spacetime_term} ⊗ Q_domain)

where the precision matrix Q_{spacetime_term} ⊗ Q_domain has dimension STC × STC to match the length of vec(E), and vec(Ω) ∼ MVN(0, Q_{space_term} ⊗ Q_domain)

where the precision matrix Q_{space_term} ⊗ Q_domain has dimension SC × SC, to match the length of vec(Ω).

Generalized additive model

Finally, the analyst can specify a generalized additive model using syntax from package mgcv. For example this might involve:

count ~ year + offset(log_area) + s(depth) + s(species, bs="re")

If year and species are factors and depth and log_area are continuous, then this would specify a fixed effect for each level year, a spline smoother for depth, using log_area as an offset, and estimating a random intercept for each level of species. This formula is parsed internally to assemble fixed effects in design matrix X and the basis functions for spline smoothers and random effects in design matrix Z. The coefficients γ associated with smoothers and random effects are then specified to follow a GMRF:

γ ∼ GMRF(0, Q_gam) where Q_gam is a blockwise diagonal matrix, assembled from estimated variance parameters and matrices constructed by mgcv.