📊 statistics

Joint Count Transformation Models with Covariate-dependent Correlations

This paper introduces joint count transformation models, a novel framework that combines distribution-free marginal count transformations with covariate-dependent latent Gaussian copulas to efficiently estimate species-specific abundance patterns and dynamic interspecific correlations, offering a computationally feasible and accurate alternative to existing parametric methods for analyzing joint species distribution data.

Original authors: Lukas Graz, Luisa Barbanti, Roland Brandl, Torsten Hothorn

Published 2026-06-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lukas Graz, Luisa Barbanti, Roland Brandl, Torsten Hothorn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand a busy bird sanctuary. You want to know two things:

How many of each bird species are there on any given day?
How do they relate to each other? Do they hang out together, or do they avoid each other?

For a long time, scientists had tools to answer the first question (how many birds) and tools to answer the second (how they relate), but they couldn't easily do both at the same time, especially when the rules of the game changed with the seasons.

This paper introduces a new, smarter tool called Joint Count Transformation Models (JCTMs). Here is how it works, using simple analogies:

1. The Problem with Old Tools

Think of traditional models like a rigid plastic mold. If you try to pour wet clay (real bird data) into a mold designed for perfect spheres (standard statistical shapes like the Poisson distribution), the clay doesn't fit right.

The Issue: Bird counts are messy. Sometimes there are zero birds (a dry patch of clay), sometimes there are hundreds (a huge lump), and the "shape" of the data changes depending on the season. Old molds assumed the data always looked the same, which led to inaccurate predictions.
The "Static" Trap: Old models also assumed that if two bird species were friends in the winter, they were friends in the summer too. They treated their relationship as a fixed rule, ignoring that birds might compete for food in summer but huddle for warmth in winter.

2. The New Solution: A "Shape-Shifting" Mold

The authors created a new framework that acts like smart, liquid clay instead of a rigid mold.

Distribution-Free: Instead of forcing the bird counts into a pre-made shape, the model learns the shape directly from the data. It's like having a mold that instantly reshapes itself to fit whatever lump of clay you pour in, whether it's a tiny speck (zero birds) or a giant blob (hundreds of birds).
The "Transformation": The model uses a mathematical "translator" to convert the messy bird counts into a smooth, understandable language. This allows it to handle the "zeros" and the "huge numbers" without breaking a sweat.

3. The Secret Sauce: The "Weather-Dependent" Friendship Map

The most exciting part of this new tool is how it handles relationships between species.

Old Way: Imagine a static map showing that Bird A and Bird B are always 50% friends.
New Way (JCTM): Imagine a live, weather-dependent map.
- In the Winter, the map might show a thick, red line connecting the birds, meaning they are huddled together for warmth (high positive correlation).
- In the Summer, that line might fade or even turn blue, indicating they are avoiding each other because they are fighting over the same fish (negative correlation).
- The model realizes that the "friendship" between species isn't a fixed number; it changes based on the "covariates" (in this case, the time of year).

4. The Real-World Test: The Lake Seehamer Birds

The authors tested this on three types of fish-eating birds in a German lake: the Great Crested Grebe, the Great Cormorant, and the Goosander.

What they found:
- Seasonal Patterns: The model perfectly captured that these birds are rare in summer but flock to the lake in winter. It handled the fact that the Goosander is almost invisible in summer (lots of zeros) but very common in winter.
- Changing Relationships: The model showed that in winter, all three species were highly correlated (they were all there at the same time). But in summer, their relationships shifted, sometimes even showing they were competing.
Speed vs. Accuracy: They compared their new tool to a very famous, powerful, but slow competitor called HMSC.
- HMSC was like a super-precise but incredibly slow artisan. It took 10 days to crunch the numbers and still had trouble "settling" on a final answer (convergence issues).
- JCTM was like a high-speed 3D printer. It solved the same problem in 10 minutes on a single computer core and gave a clear, stable answer.

5. Why This Matters

This paper doesn't claim to cure diseases or predict stock markets. It claims to solve a specific problem in ecology: How do we accurately count animals and understand how their relationships change with the environment, without making up fake rules about how the data "should" look?

By using this new "liquid mold" approach, scientists can finally see the true, shifting dynamics of nature—seeing not just who is there, but how they interact as the seasons turn, all while doing the math much faster than before.

Technical Summary: Joint Count Transformation Models with Covariate-Dependent Correlations

Problem Statement
Joint Species Distribution Models (JSDMs) are critical for understanding how ecological covariates shape species communities and for disentangling shared environmental responses from true biotic interactions. However, existing approaches face two primary limitations:

Rigid Parametric Assumptions: Most JSDMs rely on Multivariate Generalized Linear Mixed Models (MGLMMs) or Generalized Linear Latent Variable Models (GLLVMs) that assume specific parametric distributions (e.g., Poisson or Negative Binomial) for count data. These assumptions often fail to capture complex features such as overdispersion, underdispersion, and high frequencies of zero counts without arbitrary transformations (e.g., $\log(y+1)$ ) that incorrectly assume continuous likelihoods for discrete data.
Static Correlation Structures: Current methods typically model interspecific associations as static residual correlations. They lack the capacity to model how these associations change dynamically in response to environmental covariates (e.g., seasonality), a phenomenon observed in ecological literature where competition or correlation patterns shift with environmental conditions. Furthermore, many existing copula-based methods rely on two-step estimation procedures (estimating marginals before copula parameters), leading to incorrect standard errors, or are restricted to continuous data.

Methodology
The authors introduce Joint Count Transformation Models (JCTMs), a novel framework that combines distribution-free marginal models with a covariate-dependent latent Gaussian copula.

Marginal Models: Instead of assuming a parametric distribution, JCTMs employ count transformation models (Siegfried and Hothorn, 2020). These models estimate a monotone transformation function $h_j(y_j)$ directly from the data using Bernstein polynomials. This allows the model to accommodate complex distributional features (skewness, zero-inflation) without pre-specified distributional assumptions. The framework extends to location-scale transformation models (Siegfried et al., 2023), where both the location (shift) and scale of the transformed counts depend on covariates via linear predictors.
Covariate-Dependent Copula: The joint distribution is constructed using a latent Gaussian copula $\Phi_{0, \Sigma(\mathbf{x})}$ . The correlation matrix $\Sigma(\mathbf{x})$ is not static; it varies as a function of covariates $\mathbf{x}$ . To ensure $\Sigma(\mathbf{x})$ remains a valid correlation matrix during estimation, the authors parameterize it using a scaled inverse Cholesky factor $\Lambda(\mathbf{x})$ , where the off-diagonal elements are modeled as linear predictors of the covariates.
Interpretation: The correlations on the latent transformed scale can be converted to Spearman's rank correlation ( $\rho$ ) or Kendall's $\tau$ , providing interpretable measures of association on the observed count scale.
Inference: All parameters (marginal transformation, shift, scale, and copula parameters) are estimated simultaneously via joint maximum likelihood estimation (MLE). The likelihood is derived from the exact discrete log-likelihood of interval-censored data, utilizing Genz's algorithm for multivariate normal integration. This joint estimation ensures asymptotically correct standard errors, avoiding the pitfalls of two-step procedures.

Key Contributions

Distribution-Free Marginals: The framework eliminates the need for rigid parametric assumptions (e.g., Poisson) for count data, offering a flexible, semi-parametric approach that naturally handles over/underdispersion and zero-inflation.
Dynamic Interspecific Associations: JCTMs are the first JSDM framework to explicitly model covariate-dependent correlations for count data, allowing interspecific associations to vary across environmental gradients (e.g., time/season).
Joint Estimation: The method provides a unified MLE framework that estimates marginals and dependencies simultaneously, ensuring valid statistical inference and avoiding the bias of two-step estimation.
Computational Feasibility: The authors demonstrate that joint MLE is computationally feasible for communities of up to 10 species, implemented in the R package tram.

Results
The framework was applied to daily count data of three fish-eating bird species (Great Cormorant, Great Crested Grebe, and Goosander) at Lake Seehamer See over 14 years.

Marginal Performance: The model successfully captured complex seasonal abundance patterns, including high zero-count periods in summer for the Goosander and seasonal shifts in variance for the other species.
Correlation Dynamics: The model revealed strong, seasonally varying correlations. It identified high positive correlations among all three species during winter months and periods of moderate negative correlation in specific months (e.g., between Great Crested Grebe and Great Cormorant in June).
Comparison with HMSC: The results were compared against the Hierarchical Modeling of Species Communities (HMSC) framework, a state-of-the-art Bayesian approach.
- Accuracy: The JCTM correlation curves closely matched the empirical trajectories and the HMSC results.
- Efficiency: The JCTM converged in 10 minutes on a single core, whereas the HMSC model required 10 days and 5 hours on 4 cores.
- Convergence: The HMSC model exhibited poor mixing and convergence diagnostics for covariance parameters (median potential scale reduction factor of 5.53), whereas the JCTM provided stable estimates.
Simulation: A simulation study confirmed the consistency and accuracy of JCTMs for up to 10 species, though computational complexity scales roughly exponentially with the number of species ( $J$ ).

Significance and Claims
The paper claims that JCTMs offer a significant advancement over existing JSDMs by providing a flexible, distribution-free alternative that does not sacrifice the ability to model dynamic interspecific relationships.

Interpretability: Unlike MGLMMs where random effects obscure marginal interpretations, JCTMs provide directly interpretable marginal SDMs as a by-product of the fitting process.
Robustness: By avoiding strong parametric assumptions, JCTMs are better suited for ecological data that often violates standard distributional requirements (e.g., sparse counts, complex zero-inflation).
Feasibility: The authors demonstrate that joint maximum likelihood estimation for covariate-dependent correlations is now feasible for small-to-medium sized communities (up to 10 species), offering a computationally efficient alternative to Bayesian MCMC methods like HMSC.

The authors note limitations, including the sensitivity of parameter estimates to the ordering of species in the Cholesky factor (requiring sensitivity analysis) and the current computational infeasibility for very large communities ( $J > 15$ ) without dimensionality reduction or two-step approximations. They also caution that residual correlations in phenomenological models may capture shared responses to unmeasured environmental factors rather than direct biotic interactions.