Dynamic Spatio-Temporal Sequential Ordinal Models: Application to Invasive Weeds

arXiv:2509.01976 · stat.AP · Submitted 2025-09-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Dynamic Spatio-Temporal Sequential Ordinal Models".

Title and authors: Jane: I think the core takeaway for someone who doesn't read papers is that this model isn't just a complicated math trick; it’s a framework that lets us handle the uncertainty of location and time simultaneously when we have data in categories, not just numbers >

Tom: Exactly. They were looking at four invasive weeds in the ACT from two thousand nineteen to two thousand twenty-three and their main finding was that these species respond differently based on where they are and how much control effort is applied >

Lu: The summary emphasizes that this model avoids the typical MCMC bottleneck associated with fitting these complex likelihoods, which is a major methodological win because it makes estimation feasible in practice >

Meng: Feasible is one thing, but I still need to know the cost. If we can run this without bespoke procedures, does that mean a small team can deploy this on standard cloud infrastructure without needing massive computational resources just to get a result?

Lalam: The paper summarizes that the DSTSOM allows estimation and prediction in INLA exactly as it does for univariate spatio-temporal models, which is a big piece of context for anyone familiar with those systems >

Jane: So, if I’m driving, and I see an area where weeds are really dense, this model gives me a way to predict not just how much there will be later in time or in that specific spot >

Tom: That’s the application. It moves us from just looking at static maps to understanding the dynamic spread of these species based on control actions and environmental factors >

Lu: The paper specifically points out how the structure helps interpret results even when you have aggregated ordinal categories, which is a feature that makes it more transparent for biological interpretation >

Meng: Transparency is important for stakeholders. If we can show them *why* the model predicts something, even if it's in a category, that’s how we get buy-in for control strategies >

Lalam: The paper summarizes the application context by mentioning the Australian Capital Territory's Biosecurity Strategy two thousand sixteen through twenty twenty-six which aims to prevent new weed infestations and minimize impact from weeds > <ref:2509.01976#pg2>

The paper's summary: Jane: What are these future ideas? Are they suggesting we can model something completely different than just weed cover now?

Tom: They suggested extending it to model biological population growth using a Gompertz growth model, but with time-varying intrinsic rates of growth >

Lu: That’s interesting because it links the ordinal categories directly to the actual population size in a way that respects how populations naturally grow over time >

Meng: If we can link the category level to an actual population estimate, that changes what our prediction means. It moves it from just "high cover" to "this many individuals" which is more actionable for management teams >

Lalam: They mentioned a linear predictor formulation where the linear predictor of Eq. two for category c of observation j with the sign reversal convention of Eq. ten is then eta cj(t) equals beta c minus w transpose alpha j log xi t >

Jane: That sounds incredibly dense. For someone listening, what does that equation actually do in plain language?

Tom: It essentially links the ordinal categories to population size using those growth dynamics, showing how the category level relates to the actual population size over time >

Lu: The paper also showed a special case where DSTSOM is equivalent to a spatio-temporal cumulative ordinal proportional hazards model, which has that useful property of collapsibility that maintains the interpretability of regression coefficients across aggregated ordinal categories >

Meng: That collapsibility thing sounds promising. It means we don't lose the ability to interpret the underlying effects just because we group things into categories for modeling purposes >

Jane: So, if I’m looking at a map, and I see a certain pattern of weed density, this model helps me predict not just the density but also how fast that area will change over time based on growth rates >

Tom: And they also highlighted that their inclusion of spatio-temporal dependence led to a ninety-five percent confidence interval that was greater than zero for duration, which is a key finding when you compare model structures like M1, M2 and M3 >

Lu: That comparison shows the value of adding that spatial dependence. It proves that incorporating how things spread over space and time actually changes the inference about how long something lasts >

Meng: So, if we run this comparison on our own data, we can see if adding that spatial term actually makes a statistically significant difference in our duration estimates >

Lalam: They also touched on using the binary likelihood representation to apply standard univariate spatio-temporal modeling procedures for latent random effects without needing approximation in an ordinal GLM >

The paper's improvements: Jane: So, if I had to sum up the implications for my day—the listener—what does this actually change for me when I’m out there?

Tom: It means we can get much more robust and accessible predictions about weed abundance and spread that help us target control efforts more effectively because they account for spatial and temporal dependencies >

Lu: And the authors were very clear that the DSTSOM is particularly apt choice for biological populations, which suggests it’s a good starting point for modeling things like dynamic growth models >

Meng: From a production standpoint, if we can use this without requiring those heavy MCMC simulations, it means we can deploy better tools faster and get actionable results sooner >

Lalam: The paper concludes that they successfully applied the DSTSOM to the invasive weeds application in Section seven, where they incorporated spatio-temporal dependence assists interpretation and prediction of the dynamic biological populations >

Jane: It seems like a solid piece of work, connecting complex statistical theory to real ecological problems in a way that’s actually useful for managers and decision-makers >

Tom: The full title is "Dynamic Spatio-Temporal Sequential Ordinal Models: Application to Invasive Weeds," and it gives us a concrete method for analyzing these types of spatio-temporal ordinal datasets >

Lu: To wrap up, the paper sets up clear directions for future work, like extending the model into non-separable latent space-time process models by adding more complex structure to that state transition matrix >

Meng: We'll keep an eye on those extensions, but for now, it’s a solid method to get started with our current weed monitoring data >

Conclusion: Jane: So we’ve been looking at how to use sequential ordinal models to fit spatio-temporal data, and this paper takes that idea and applies it directly to tracking four different invasive weed species in Australia >

Tom: Yeah, it uses a model called the Dynamic Spatio-Temporal Sequential Ordinal Model, or DSTSOM, and the authors show how you can frame that likelihood as a binary regression model conditional on unknown coefficients and spatio-temporal random effects >

Lu: The core innovation they are pushing is that this approach avoids those ordering constraints you usually have to deal with when fitting cumulative ordinal models because it considers the transition probabilities between levels instead of just forcing them into a strict sequence >

Meng: From a deployment standpoint, the paper shows how this lets us use standard dynamic generalized linear model frameworks without needing those computationally demanding Markov chain Monte Carlo simulations, which is pretty appealing for running things at scale >

Lalam: And they did show that by using the binary likelihood representation of the DSTSOM, you can apply standard univariate spatio-temporal modeling procedures to latent random effects in an effectively univariate estimation framework without needing approximation in an ordinal GLM >

Jane: So, what are the actual findings? What did they find when they looked at these invasive weeds observed from two thousand nineteen to two thousand twenty-three >

Tom: They found that the four different weed species respond differently based on habitat type and control effort, but they all share similar patterns of dependence with short effective spatial ranges and strong temporal autocorrelations >

Lu: Specifically, they looked at model structures like M1, M2, and M3 to assess the influence of spatio-temporal dependence on those control covariate coefficients >

Tom: The comparison showed that adding the spatial dependence led to a ninety-five percent confidence interval that was greater than zero for duration for some species >

Jane: That’s a big deal because it means incorporating how things spread over space and time actually changes our statistical inference about how long something lasts, which is what we care about when managing these weeds >

Meng: It’s good to see the authors show that the inclusion of spatio-temporal dependence made a difference in the duration estimates for some species >

Lalam: They also highlighted that the high temporal autocorrelations across all species indicate that their growth model is moving toward an exponential growth model over time >

Lu: And they did suggest ways to extend this, like modeling biological population growth using a Gompertz growth model with time-varying intrinsic rates of growth by formulating the linear predictor in a specific way >

Tom: So, if you’re looking at the whole picture, the DSTSOM gives us a much cleaner way to handle these complex biological responses than traditional methods might allow >

Jane: It’s a lot of stuff to take in about this Dynamic Spatio-Temporal Sequential Ordinal Model application to Invasive Weeds, but it really shows how we can get more predictive power when dealing with data that lives both in space and time >

Lu: The paper sets up some interesting next steps too, like extending the model to non-separable latent space-time process models by adding more complex structure to the state transition matrix >

Meng: For us in production, I’m watching how they handle those extensions because that’s where things get messy when you try to ship it live >

Lalam: I'm fairly confident in this direction; the results on the duration CI being greater than zero for some species is quite robust given the data structure they used >

CSIRO Data61 Hobart Tasmania Australia · CSIRO Data61 Dutton Park Queensland Australia · CSIRO Health and Biosecurity Townsville Queensland Australia · Office of Nature Conservation Australian Capital Territory Government Canberra Australian Capital Territory Australia · Parks and Conservation Service Australian Capital Territory Government Canberra Australian Capital Territory Australia · CSIRO Health and Biosecurity Canberra Australian Capital Territory Australia · CSIRO Health and Biosecurity Dutton Park Queensland Australia

stat.AP

Submitted: 2025-09-02

Updated: 2025-09-02

DOI: 10.1007/s00477-026-03389-7

Code: https://github.com/csiro/DSTSOM

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 74/100

The gist: The multivariate sequential ordinal model is investigated for use in the Bayesian analysis of spatio-temporal ordinal data and is applied to estimate foliage cover abundance of four actively managed…

Key concepts

Dynamic Spatio-Temporal Sequential Ordinal Model (DSTSOM)
This model is a method used to fit ordinal data that changes over both space and time. It works by treating the sequential ordinal likelihood as a binary model, which allows researchers to use established dynamic generalized linear models for estimation and prediction in complex spatio-temporal settings.
Sequential Ordinal Model Likelihood
This is the mathematical framework used to describe how the data follows an ordinal pattern. Instead of imposing strict ordering constraints on category break points, this likelihood considers the transition probabilities between adjacent levels, simplifying estimation for high-dimensional spacetime applications.
Bayesian Estimation in Spatio-Temporal Models
This involves using Bayesian inference to estimate parameters within a model that incorporates both spatial and temporal random effects. This approach is particularly useful for complex biological data because it allows for the incorporation of informative prior knowledge, which can be challenging in other estimation methods.
Spatio-Temporal Dependence
This refers to the correlation between observations based on both their location (space) and the time they were measured. The study found that all four invasive species shared similar magnitudes of dependence with short effective spatial ranges and strong temporal autocorrelations.

Terminology

Summary

The multivariate sequential ordinal model is investigated for use in the Bayesian analysis of spatio-temporal ordinal data and is applied to estimate foliage cover abundance of four actively managed invasive alien species. p. 0

Where the paper says it

Moreover, the sequential ordinal model avoids the ordering constraints on the threshold parameters that determine the category break points required by cumulative ordinal models, and so simplifies the estimation procedure for high-dimensional spacetime applications using Bayesian inference. The dynamic spatio-temporal sequential ordinal model is applied to estimate foliage cover abundance of four actively managed invasive alien species.

Model Formulation and Theoretical Basis

The dynamic spatio-temporal sequential ordinal model (DSTSOM) is proposed as a way to fit spatio-temporal applications for ordinal data. The sequential ordinal model likelihood is equivalent to a binary model conditional on unknown regression coefficients and spatio-temporal random effects, allowing estimation and prediction in the space-time context using the dynamic generalised linear model framework. p. 0

Where the paper says it

The sequential ordinal model likelihood is equivalent to a binary model conditional on unknown regression coefficients and spatio-temporal random effects. Therefore, estimation and prediction in the space-time context can proceed using the well-established dynamic generalised linear model framework.

It avoids the ordering constraints on the threshold parameters that determine category break points required by cumulative ordinal models, which simplifies estimation for high-dimensional spacetime applications using Bayesian inference. p. 0
Where the paper says it

Therefore, estimation and prediction in the space-time context can proceed using the well-established dynamic generalised linear model framework. Moreover, the sequential ordinal model avoids the ordering constraints on the threshold parameters that determine the category break points required by cumulative ordinal models, and so simplifies the estimation procedure for high-dimensional spacetime applications using Bayesian inference.

Data and Application Context

The model is applied to estimate foliage cover abundance of four actively managed invasive alien species observed by means of a modified Braun-Blanquet score that constitutes ordinal data. These invasive weed species are sparsely distributed in space and time with few observations recorded in high foliage cover categories. p. 0

Where the paper says it

These invasive weed species are observed by means of a modified Braun-Blanquet score that is commonly used in vegetation studies and constitutes ordinal data. The multivariate ordinal data for the managed weeds species are sparsely distributed in space and time with few observations recorded in high foliage cover categories.

The study uses weed monitoring and routine control activity data from the Australian Capital Territory (ACT) for the years 2019—2023. p. 2
Where the paper says it

In this paper, weed monitoring and routine control activity (e. , herbicide, manual removal, mowing) data from the ACT for the years 2019—2023 were investigated for three perennial invasive grasses and a perennial herb species.

Key Model Features

The DSTSOM is defined by extending dynamic generalised linear models to account for the multivariate ordinal response data in spatio-temporal models. The likelihood of a sequential ordinal GLM can be recast as a binary regression model, accommodating the sequential ordinal model into the spatio-temporal GLM of Eq. p. 5

Where the paper says it

(5) and (6) is multivariate. The likelihood of a sequential ordinal GLM, however, can be recast as a binary regression model (Armstrong and Sloan, 1989; Berridge and Whitehead, 1991; Bender and Benner, 2000; Tutz, 2012).

(2). A special case of the DSTSOM that is equivalent to a spatio-temporal cumulative ordinal proportional hazards model is shown in Section 5. p. 2
Where the paper says it

DYNAMIC SPATIAL-TEMPORAL ORDINAL MODELS 3 implications of the sequential ordinal model and the global effects structure are interpreted for dynamic growth models of biological populations in Section 4. A special case of the DSTSOM that is equivalent to a spatio-temporal cumulative ordinal proportional hazards model is shown in Section 5.

Results and Interpretation

Bayesian estimation and prediction shows that the four invasive weed species differentially respond to habitat type, control effort and accessibility, and share similar magnitudes of dependence with short effective spatial ranges and strong temporal autocorrelations. The analysis suggests that for some species, the inclusion of spatial dependence led to a 95% CI that was greater than zero for duration. p. 1

Where the paper says it

Weed distribution has important spatial and temporal dimensions that must be accounted for by specific weed management activities undertaken to control weed distribution and limit their negative environmental and economic impacts. Spatial dependence is anticipated due to unobserved covariates, such as fine-scale habitat characteristics, that lead to greater similarity among nearby locations compared to those further apart.

The high autocorrelations for all species indicate that the Gompertz growth model is approaching an exponential growth model. p. 0
Where the paper says it

Bayesian estimation and prediction shows that the four invasive weed species differentially respond to habitat type, control effort and accessibility, and share similar magnitudes of dependence with short effective spatial ranges and strong temporal autocorrelations. Keywords: collapsibility, Gompertz population growth, grouped Cox model, proportional hazards model, species distribution model 1.

Short effective spatial ranges relative to the study area were estimated for all species. p. 0
Where the paper says it

Estimation and prediction is demonstrated using integrated nested Laplace approximation (INLA) methods developed for univariate spatio-temporal models. Bayesian estimation and prediction shows that the four invasive weed species differentially respond to habitat type, control effort and accessibility, and share similar magnitudes of dependence with short effective spatial ranges and strong temporal autocorrelations.

The inclusion of spatio-temporal dependence led to a 95% CI that was greater than zero for duration. p. 1
Where the paper says it

Weed distribution has important spatial and temporal dimensions that must be accounted for by specific weed management activities undertaken to control weed distribution and limit their negative environmental and economic impacts. Spatial dependence is anticipated due to unobserved covariates, such as fine-scale habitat characteristics, that lead to greater similarity among nearby locations compared to those further apart.

Model Validation and Extensions

The model was evaluated against alternative models of dependence, including M1, sequential ordinal GLM without random effects; M2, spatial sequential ordinal model; and M3, dynamic spatio-temporal sequential ordinal model. The DSTSOM was shown to be particularly apt choice for biological populations. p. 2

Where the paper says it

DYNAMIC SPATIAL-TEMPORAL ORDINAL MODELS 3 implications of the sequential ordinal model and the global effects structure are interpreted for dynamic growth models of biological populations in Section 4. A special case of the DSTSOM that is equivalent to a spatio-temporal cumulative ordinal proportional hazards model is shown in Section 5.

Extensions to non-separable latent space-time process models could be implemented by including more complex structure in the state transition matrix of Eq. p. 4
Where the paper says it

, 2003), where the random effects follow an order-one autoregressive process with v × v state transition matrix M. The k-dimensional vector of process innovations, ϵ(t), accounts for process uncertainty in the latent autoregressive process at time t.

(2). The sequential ordinal model allows for estimation and prediction in INLA as with univariate spatio-temporal models. p. 2
Where the paper says it

This model has the useful property of collapsibility that maintains the interpretability of regression coefficients across aggregated ordinal categories. The DSTSOM allows estimation and prediction in INLA as with univariate spatio-temporal models (Section 6).

The inclusion of informative priors into Bayesian analysis is an important challenge for INLA. p. 0
Where the paper says it

FROESE7 1CSIRO Data61, Hobart, Tasmania, Australia 2CSIRO Data61, Dutton Park, Queensland, Australia 3CSIRO Health and Biosecurity, Townsville, Queensland, Australia 4Office of Nature Conservation, Australian Capital Territory Government, Canberra, Australian Capital Territory, Australia 5Parks and Conservation Service, Australian Capital Territory Government, Canberra, Australian Capital Territory, Australia 6CSIRO Health and Biosecurity, Canberra, Australian Capital Territory, Australia 7CSIRO Health and Biosecurity, Dutton Park, Queensland, Australia Abstract. The multivariate sequential ordinal model is investigated for use in the Bayesian analysis of spatio-temporal ordinal data.

Extensions are also possible to multivariate models that incorporate other response data types, such as continuous and Poisson count data. p. 1
Where the paper says it

The dynamic spatio-temporal sequential ordinal model (DSTSOM) is defined in Section 3. This ordinal model formulation extends dynamic generalised linear models to account for the multivariate ordinal response data in spatio-temporal models.

The binary likelihood representation of the dynamic spatio-temporal sequential ordinal model therefore also allows application of these multivariate techniques to joint models of ordinal datasets. p. 9
Where the paper says it

As shown in Section 3, the recasting of the sequential ordinal model by means of the binary likelihood representation allows its incorporation into an effectively univariate estimation framework without need for approximation in an ordinal GLM. Univariate implementations of the dynamic spatio-temporal sequential ordinal model developed for INLA are then directly applicable to the multivariate ordinal data with this approach.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI can estimate foliage cover abundance of four actively managed invasive alien species using Bayesian analysis by employing a dynamic spatio-temporal sequential ordinal model (DSTSOM). This allows for robust and accessible prediction of weed abundance and spread, facilitating more effective targeting and evaluation of invasive weed control efforts.

  2. Improved AI can handle high dimensional latent variable settings common in spatial and spatio-temporal applications without requiring computationally demanding Markov chain Monte Carlo simulations. The sequential ordinal model's likelihood is equivalent to a binary model conditional on unknown regression coefficients and spatio-temporal random effects, allowing estimation and prediction using the well-established dynamic generalised linear model framework.

  3. Improved AI can simplify the estimation procedure for high dimensional spacetime applications by avoiding the ordering constraints on threshold parameters required by cumulative ordinal models. The sequential ordinal model avoids this constraint because it does not impose strict ordering of cut levels and instead considers the transition probabilities from one level to another.

  4. Improved AI can perform analysis that maintains parameter interpretability even in the presence of aggregated ordinal categories through a separable model for space-time dependence. This is achieved by developing a separable model that maintains parameter interpretability in the presence of aggregated ordinal categories.

  5. Improved AI can incorporate ecological domain knowledge into Bayesian prior elicitation for ordinal GLMs through expert elicitation, which is an important challenge for INLA. This allows the system to utilize multivariate normal or Student’s t prior elicitation for unknown coefficients in ordinal GLMs.

  6. Improved AI can extend the DSTSOM to model biological population growth using a Gompertz growth model with time-varying intrinsic rates of growth. The linear predictor is formulated such that the linear predictor of Eq. (2) for category c of observation j with the sign reversal convention of Eq. (10) is then eta cj(t) = beta c - w Tj log xi(t), linking ordinal categories to population size.

  7. Improved AI can assess the influence of spatio-temporal dependence on control covariate coefficients by comparing model structures, such as M1, M2, and M3. This allows researchers to see how the inclusion of spatial dependence led to a 95% CI that was greater than zero for duration and how the presence of a control event was positively associated with coverage of St. John’s wort but not duration since control.

  8. Improved AI can utilize the binary likelihood representation of the DSTSOM to apply standard univariate spatio-temporal modelling procedures, such as those developed for a separable autoregressive space-time model for the latent random effects. This allows its incorporation into an effectively univariate estimation framework without need for approximation in an ordinal GLM.

  9. Improved AI can facilitate extensions to multivariate models by allowing the application of techniques from Poisson point process models with binary likelihoods. The complementary log-log link function used in the sequential ordinal model also maps into Poisson point process models with binary likelihoods, which is useful for joint modeling of ordinal datasets.

Abstract

The multivariate sequential ordinal model is investigated for use in the Bayesian analysis of spatio-temporal ordinal data. The sequential ordinal model likelihood is equivalent to a binary model conditional on unknown regression coefficients and spatio-temporal random effects. Therefore, estimation and prediction in the space-time context can proceed using the well-established dynamic generalised linear model framework. Moreover, the sequential ordinal model avoids the ordering constraints on the threshold parameters that determine the category break points required by cumulative ordinal models, and so simplifies the estimation procedure for high-dimensional space-time applications using Bayesian inference. The dynamic spatio-temporal sequential ordinal model is applied to estimate foliage cover abundance of four actively managed invasive alien species. These invasive weed species are observed by means of a modified Braun-Blanquet score that is commonly used in vegetation studies and constitutes ordinal data. The multivariate ordinal data for the managed weeds species are sparsely distributed in space and time with few observations recorded in high foliage cover categories. A separable model for space-time dependence that maintains parameter interpretability in the presence of aggregated ordinal categories is therefore developed. Estimation and prediction is demonstrated using integrated nested Laplace approximation (INLA) methods developed for univariate spatio-temporal models. Bayesian estimation and prediction shows that the four invasive weed species differentially respond to habitat type, control effort and accessibility, and share similar magnitudes of dependence with short effective spatial ranges and strong temporal autocorrelations.

Related papers