Dynamic Spatio-Temporal Ordinal Model

• The model is a Bayesian framework that analyzes ordered categorical responses using sequential conditional probabilities.
• It recasts the ordinal likelihood as a series of binary components, enabling dynamic generalized linear modeling for sparse spatio-temporal data.
• The approach integrates latent Gaussian random effects to capture spatial and temporal dependencies without strict threshold constraints.

The dynamic spatio-temporal sequential ordinal model is a Bayesian framework for spatio-temporal ordinal data in which the response is observed on an ordered categorical scale and dependence is represented through both covariates and latent spatio-temporal random effects. In the formulation developed for invasive weed abundance, the model uses a sequential ordinal likelihood, recasts that likelihood as a product of binary likelihoods, and embeds the resulting observation model within a dynamic generalized linear model structure. This construction is intended for settings with sparse observations in space and time, aggregated ordinal categories, and high-dimensional latent dependence, while avoiding the threshold-ordering constraints required by cumulative ordinal models (Hosack et al., 2 Sep 2025).

1. Definition and modeling scope

In this model class, the observed response is an ordinal variable $z_j(t) \in \{1, \ldots, C\}$, recorded for observation $j$ at time $t$, with location $\bm{s}_j$. The model allows multiple observations per location and time. Its central feature is the use of a sequential ordinal model, in which category probabilities are specified through conditional probabilities of being in category $c$ given membership in category $c$ or higher (Hosack et al., 2 Sep 2025).

The defining conditional probability is

$$P(z_j(t) = c \mid z_j(t) \geq c, \eta_{c \mid j}(t)) = \delta_{c \mid j}(t) = g^{-1}(\eta_{c \mid j}(t)),$$

with link function $g$, commonly complementary log-log in the weed application. The marginal probability of category $c$ is then

$$P(z_j(t) = c) = \delta_{c \mid j}(t) \prod_{\tilde{c}=1}^{c-1} (1 - \delta_{\tilde{c} \mid j}(t)).$$

This construction differs from cumulative ordinal formulations in a specific and consequential way: the threshold parameters do not need to satisfy strict ordering constraints. In the reported application, that simplification is presented as especially useful for high-dimensional space-time problems under Bayesian inference. The model is also described as multivariate in the invasive-weed study, where four species are analyzed and where ordinal abundance is measured with a modified Braun-Blanquet score (Hosack et al., 2 Sep 2025).

A common misunderstanding is to treat “sequential” here as referring only to temporal succession. In this model, “sequential” also refers to the ordinal likelihood itself: the observation is generated through a sequence of conditional transitions across ordered categories. Temporal evolution enters separately through the latent spatio-temporal process.

2. Sequential ordinal likelihood and binary reformulation

The linear predictor for category $j$0 and observation $j$1 at time $j$2 is

$j$3

where $j$4 is a threshold parameter for level $j$5, $j$6 are coefficients for covariates, $j$7 are covariates shared across categories for observation $j$8, and $j$9 is the spatio-temporal random-effect vector projected to the observation. The covariates and random effects are therefore modeled as global effects, meaning they are constant across categories for a given observation, whereas only the threshold parameters vary with category (Hosack et al., 2 Sep 2025).

The most important computational property of the model is that the sequential ordinal likelihood can be rewritten as a product of binary likelihoods. For each observation and threshold up to the observed category, a binary variable is introduced: it is coded as 1 if the observed category is $t$0, and 0 otherwise, for all thresholds up to the observed category. The full likelihood becomes

$t$1

This binary equivalence is the mechanism that connects the model to the dynamic generalized linear model framework. Conditional on the regression coefficients and spatio-temporal random effects, estimation and prediction can proceed using machinery developed for binary generalized linear models rather than requiring bespoke multivariate ordinal algorithms. In the reported formulation, this is not merely a computational convenience; it is a structural identity of the sequential likelihood (Hosack et al., 2 Sep 2025).

A plausible implication is that the model occupies an intermediate position between conventional ordinal regression and latent-Gaussian spatio-temporal GLMs: it retains an explicitly ordinal observation model while exploiting binary latent-Gaussian computational infrastructure.

3. Dynamic spatio-temporal dependence structure

Spatio-temporal dependence is introduced through latent Gaussian random effects evolving over time:

$t$2

Here, $t$3 is the random-effect vector over all potential locations at time $t$4, $t$5 is the state-transition matrix, and $t$6 is the innovation covariance (Hosack et al., 2 Sep 2025).

For the invasive-weed application, a separable space-time structure is adopted:

$t$7

In this parameterization, $t$8 is a shared first-order temporal autocorrelation parameter, and $t$9 is the spatial covariance matrix, often Matérn. The spatial covariance is written as

$\bm{s}_j$0

with principal parameters described as spatial marginal variance $\bm{s}_j$1, range $\bm{s}_j$2, and smoothness $\bm{s}_j$3 (Hosack et al., 2 Sep 2025).

The separable formulation is motivated in the source description by two requirements: computational simplicity and maintenance of parameter interpretability in the presence of aggregated ordinal categories. This suggests a deliberate trade-off. Richer nonseparable dependence structures are not excluded in principle, but the reported model emphasizes a decomposition in which temporal persistence and spatial correlation remain directly interpretable.

The model also uses a sign reversal convention in the predictor so that increases in covariates or random effects increase the probability of higher categories. This convention matters for interpretation because it aligns coefficient signs with increased abundance or severity on the ordinal scale (Hosack et al., 2 Sep 2025).

4. Bayesian inference, INLA, and prior structure

The estimation strategy is Bayesian and uses integrated nested Laplace approximation (INLA). In the reported implementation, the binary reformulation of the sequential ordinal likelihood allows inference to be carried out within INLA’s standard univariate binary GLM framework, rather than requiring specialized multivariate ordinal machinery. The expanded binary responses, expanded design matrices for covariates and random effects, and the spatio-temporal latent field represented through meshes and projection matrices together make the model compatible with univariate spatio-temporal methods already implemented in INLA (Hosack et al., 2 Sep 2025).

This is presented as a direct advantage over cumulative ordinal models. In the cumulative case, one must impose and preserve ordered thresholds, such as $\bm{s}_j$4, and the source description notes that such models often require multivariate MCMC-based inference or approximate procedures, including Dirichlet approximations in INLA. By contrast, the sequential formulation removes the strict ordering constraint on thresholds and simplifies prior specification and computation (Hosack et al., 2 Sep 2025).

The prior structure reported for the weed application includes:

Parameter class	Prior specification
Matérn variance $\bm{s}_j$5	Penalized so $\bm{s}_j$6
Range $\bm{s}_j$7	Penalized so $\bm{s}_j$8
Temporal autocorrelation $\bm{s}_j$9	Prior favoring high autocorrelation
Regression coefficients $c$0	Diffuse normal

The prior on $c$1 is motivated by the expectation of persistent weed populations. The model is also described as naturally accommodating missing or sparse data, which is relevant because the invasive-weed observations are sparse in both space and time and concentrated in lower cover categories (Hosack et al., 2 Sep 2025).

A frequent misconception is that INLA is restricted to standard Gaussian or binary outcomes in a way that excludes ordinal space-time modeling. In this framework, ordinal inference is achieved precisely by exploiting the binary decomposition of the sequential likelihood; the ordinal structure is preserved at the model level even though the computational engine is binary.

5. Parameter interpretation, aggregation, and applied ecological use

The model distinguishes several parameter classes with different inferential roles. Threshold parameters $c$2 delimit transitions between categories and affect category probabilities non-monotonically. The covariate coefficients $c$3 are global effects that shift the entire probability profile and, under the adopted sign convention, influence the likelihood of higher categories. The dependence parameters $c$4 govern the magnitude of latent variation, the spatial range of correlation, and temporal persistence. The random effects $c$5 capture residual local space-time variation and modify prediction for specific contexts (Hosack et al., 2 Sep 2025).

The source description emphasizes an additional property: in the complementary log-log global-effects case, collapsibility is maintained under category aggregation, matching the proportional hazards model. This is significant for ecological applications in which high-abundance categories may be sparsely observed and combined for analysis. In the motivating dataset, the multivariate ordinal observations for the four managed invasive weed species are sparsely distributed in space and time, and few observations occur in high foliage-cover categories. The separable model is therefore presented as one that preserves interpretability in the presence of aggregated ordinal categories (Hosack et al., 2 Sep 2025).

The reported application concerns foliage cover abundance of four actively managed invasive alien species observed in the Australian Capital Territory from 2019 to 2023 using Braun-Blanquet-based ordinal scores from 1, denoting absence, to 5, denoting high cover. Species-specific models include habitat, distance to road, year, and weed-control timing as covariates, together with spatio-temporal random effects. Bayesian estimation and prediction are reported to show that the four weed species differentially respond to habitat type, control effort, and accessibility, while sharing similar magnitudes of dependence characterized by short effective spatial ranges and strong temporal autocorrelations (Hosack et al., 2 Sep 2025).

The source also notes that the model can be interpreted in terms of a population process, such as Gompertz growth, when the link is complementary log-log and the random effects are connected to latent abundance. This does not redefine the model as a mechanistic population-dynamics model, but it provides a biologically interpretable lens for the ordinal regression parameters.

6. Relation to adjacent spatio-temporal sequence models

The dynamic spatio-temporal sequential ordinal model belongs to a broader family of models that combine space, time, and sequence structure, but it addresses a distinct problem. Several neighboring literatures use “sequential” or “ordinal” in materially different senses.

In video understanding, “Human Activity Recognition based on Dynamic Spatio-Temporal Relations” models successive actions through qualitative spatio-temporal graphs and a discrete Hidden Markov Model, with long videos partitioned into action segments and represented as sequences of “visual words” (Liu et al., 2020). “VideoMamba: Spatio-Temporal Selective State Space Model” uses selective state-space dynamics and bidirectional scans to capture temporal order and spatial context in video tokens, with linear complexity rather than quadratic self-attention (Park et al., 2024). These models are sequential in the sense of temporal evolution and order sensitivity, but they do not define an ordinal response likelihood over ordered categories.

In forecasting, “STG2Seq: Spatial-temporal Graph to Sequence Model for Multi-step Passenger Demand Forecasting” treats prediction as a sequence-to-sequence problem over future time points, using hierarchical graph convolutional modules and attention for dynamic temporal and channel-wise information (Bai et al., 2019). Here, “ordinal” refers to the ordered forecast horizon rather than to ordinal observations. Likewise, “Learning Spatiotemporal Dynamical Systems from Point Process Observations” jointly models latent continuous dynamics and a spatiotemporal point process for observation times and locations; its observation model uses a fixed-variance Gaussian rather than an ordinal likelihood (Iakovlev et al., 2024).

Within Bayesian latent-variable modeling, “Deep Markov Spatio-Temporal Factorization” provides a deep generative framework for nonlinear multimodal temporal dynamics and spatial factorization, but the source description explicitly states that its current formulation assumes Gaussian continuous distributions and makes no mention of explicit ordinal or count-data extensions (Farnoosh et al., 2020). This marks a clear methodological boundary: dynamic spatio-temporal latent structure alone does not supply ordinal-data semantics.

For large ordinal space-time datasets, a different Bayesian strategy appears in “Two-stage MCMC for Fast Bayesian Inference of Large Spatio-temporal Ordinal Data, with Application to US Drought,” which uses a latent continuous process with ordinal cutpoints, site-wise temporally dependent stage-one models, and a second-stage Metropolis-Hastings correction that restores spatial dependence while preserving the exact full posterior (Hepler et al., 30 May 2025). That work addresses ordinal spatio-temporal scaling, but through a cutpoint-based latent-variable model rather than a sequential ordinal likelihood.

These comparisons clarify the scope of the dynamic spatio-temporal sequential ordinal model. Its distinguishing combination is not merely that it is dynamic, spatio-temporal, or sequential in isolation, but that it integrates all three with an ordinal likelihood whose binary equivalence makes high-dimensional Bayesian inference tractable within established latent-Gaussian computational frameworks (Hosack et al., 2 Sep 2025).