Bayesian Species Distribution Models using Hierarchical Decomposition Priors

This paper introduces a Hierarchical Decomposition prior framework for Bayesian species distribution models that reparametrizes variance components to enable transparent, ecologically meaningful control over the relative contributions of environmental, spatial, and temporal processes, as demonstrated through improved interpretability and comparable predictive performance on fish distribution data.

The Gist

Imagine you are trying to understand why a specific type of fish shows up in certain spots in the ocean and not others. Is it because of the water temperature? The depth? The time of year? Or maybe it's just random chance, or because of how that fish interacts with other creatures?

This is the big question ecologists face. To answer it, they use a statistical tool called a Species Distribution Model (SDM). Think of an SDM as a giant recipe that tries to predict where a species will be based on a list of ingredients (environmental factors).

However, there's a problem with how these recipes are usually written. The "ingredients" (the math behind the model) are hard to control. It's like trying to bake a cake where you can't easily tell how much sugar versus how much flour is actually doing the work. You might guess, "I think the flour is important," but the math doesn't let you say that clearly.

This paper introduces a new, smarter way to bake that cake. Here is the breakdown in simple terms:

1. The Problem: The "Black Box" of Variance

In the old way of doing things, the model has a "variance" knob for every single ingredient. "Variance" is just a fancy word for "how much this factor changes the outcome."

• The Issue: Setting these knobs is like guessing the weight of a feather in a hurricane. You don't know if the "Temperature" knob should be turned up high or low. If you guess wrong, your model might think the fish loves cold water when it actually doesn't.

2. The Solution: The "Hierarchical Decomposition" (HD) Tree

The authors (Luisa, Massimo, and Alex) suggest a new way to set these knobs called Hierarchical Decomposition.

Imagine you have a giant pie representing 100% of the mystery of why the fish are where they are.

• Old Way: You just throw darts at the pie to guess how big each slice is.
• New Way (HD): You use a Tree to cut the pie logically.
  1. First Cut: You split the pie in half. One side is "Environment" (Temperature, Depth, etc.), and the other is "Everything Else" (Time, Space, Randomness). You can now say, "I think 70% of the mystery is about the environment."
  2. Second Cut: You take the "Environment" slice and split it again. Maybe you decide Depth is the biggest slice, and Temperature is smaller.
  3. Third Cut: You keep splitting until you get down to the specific ingredients.

This tree structure lets you talk to the model in plain English: "I believe Depth is the most important factor, and Temperature is less important." The math then translates your English into the correct "knob settings."

3. The Secret Sauce: Standardization (The "Ruler")

There was a catch. If you measure a slice of pie in inches and another in centimeters, you can't compare them fairly. The authors realized that the old math didn't use a consistent "ruler."

They introduced a Standardization step. Before they even start cutting the pie, they make sure every ingredient is measured on the same scale.

• Analogy: Imagine you are comparing the weight of a feather and a bowling ball. If you weigh the feather in grams and the ball in tons, the numbers are useless. They convert everything to "grams" first. This ensures that when you say "Depth is 50% of the pie," you actually mean 50% of the real impact, not just a math trick.

4. The Real-World Test: The Fish Story

To prove this works, they tested it on 39 different types of fish in the North Atlantic (using data from NOAA).

• The Result: The new method predicted where the fish would be just as well as the old, complicated methods.
• The Bonus: But unlike the old methods, the new method told them exactly what was driving the fish.
  • Discovery: They found that Depth and Bottom Temperature were the biggest drivers (the biggest slices of the pie). This makes perfect sense because these are "bottom-dwelling" fish, and they care deeply about how deep the water is and how warm the floor is.

5. Why This Matters: The "Sensitivity" Check

The best part of this new method is transparency.

• Old Way: If you wanted to test if your model was too sensitive to a guess, you had to tweak complex math formulas and hope for the best.
• New Way: You can simply say, "Let's see what happens if I make the 'Depth' slice smaller." The model instantly shows you how the prediction changes. It's like adjusting the volume on a radio to see how the music sounds, rather than rewiring the speakers.

Summary

This paper is about giving ecologists a better set of tools to understand nature.

• Before: They were guessing how much each environmental factor mattered, using confusing math.
• Now: They can use a logical "Tree" to decide, "I think Depth matters most," and the math handles the rest.
• The Outcome: We get predictions that are just as accurate, but we finally understand why the model thinks what it thinks. It turns a black box into a clear, transparent window.

Technical Summary

Here is a detailed technical summary of the paper "Bayesian Species Distribution Models using Hierarchical Decomposition Priors" by Ferrari, Ventrucci, and Laini.

1. Problem Statement

Species Distribution Models (SDMs) are fundamental in ecology for understanding how environmental, spatial, and temporal processes drive species occurrence. A critical challenge in Bayesian SDMs is the specification of priors for variance components.

• The Issue: Traditional approaches often assign independent, vague priors (e.g., Inverse-Gamma) or fixed large values to variance parameters ($\sigma^2$). This makes it difficult to incorporate expert knowledge regarding the relative importance of different factors (e.g., abiotic vs. biotic effects) and hinders the interpretability of variance partitioning.
• The Gap: There is a lack of a principled framework that allows ecologists to directly control the proportion of variance attributed to specific model effects (environmental covariates, spatial, temporal) in a transparent manner.

2. Methodology

The authors adapt the Hierarchical Decomposition (HD) prior framework (Fuglstad et al., 2020) to Latent Gaussian Models (LGMs) used in SDMs. The methodology involves three core technical steps:

A. Standardization of Effects

For HD priors to be interpretable, the variance parameters must accurately reflect the actual contribution of each effect to the total variance. The authors implement a standardization procedure (based on Ferrari and Ventrucci, 2025):

• Reparameterization: Effects are scaled so that their variance parameters correspond to the variance of the effect over the covariate space.
• Distributional Assumptions: The authors assume a Uniform distribution for covariates (environmental, spatial, and temporal) rather than relying on the empirical data distribution. This ensures the variance contribution is defined over the entire support of the domain, making it robust and comparable across studies.
• P-Spline Decomposition: For continuous covariates modeled with P-Splines (which typically only measure deviation from a linear trend), the authors split the effect into two components: a linear effect ($f_L$) and a non-linear effect ($f_N$). Both are standardized separately.

B. Hierarchical Decomposition Tree

The total variance ($V$) is reparametrized into a total scale and a set of proportion parameters ($\omega$) using a decomposition tree:

1. Total Variance ($V$): The sum of all variance components.
2. Proportions ($\omega$): The tree recursively splits $V$ into groups.
   • Level 1: Splits total variance into Abiotic (environmental) vs. Biotic/Residual (spatio-temporal) effects.
   • Level 2/3: Further splits based on specific covariates (e.g., Depth, Temperature) or effect types (e.g., Spatial vs. Temporal).
   • Level 4: Splits linear vs. non-linear components for P-Spline effects.

• Default Tree: The authors propose a "default tree" structure tailored to SDMs, which can be pruned or expanded based on specific model architectures.

C. Prior Specification

The framework assigns priors to the new parameters ($V$ and $\omega$):

• Total Variance ($V$): A scale-invariant Jeffreys prior or a Penalized Complexity (PC) prior.
• Proportions ($\omega$):
  • Dirichlet Priors: Used for splits with no strong prior preference (e.g., splitting among multiple environmental covariates). The concentration parameter $q$ controls sparsity.
  • Penalized Complexity (PC0) Priors: Used for binary splits (e.g., Linear vs. Non-linear) to induce shrinkage toward a simpler base model (e.g., favoring linearity unless data suggests non-linearity). The authors derive a simplified closed-form expression for the PC0 prior on proportions: $\pi(\omega) \propto \frac{\exp(-\lambda\sqrt{\omega})}{\sqrt{\omega}}$.

3. Key Contributions

1. Adaptation of HD Priors to SDMs: This is the first application of the HD framework specifically to Species Distribution Models, addressing the unique challenge of separating abiotic and biotic/spatio-temporal variance.
2. Standardization Protocol: The paper provides a rigorous method to standardize LGM effects (including P-Splines) using Uniform covariate assumptions, ensuring that variance parameters are interpretable as true contributions to total variance.
3. Default Tree Design: A proposed, ecologically grounded decomposition tree that allows researchers to intuitively specify priors based on the relative importance of environmental vs. spatial/temporal factors without needing complex mathematical derivations.
4. Simplified PC0 Prior: The derivation and validation of a simplified functional form for PC0 priors on variance proportions, which significantly reduces computational complexity and implementation barriers.

4. Results

The method was applied to a dataset of 39 demersal fish species from the NOAA Northeast Fisheries Science Center (2000–2019).

• Predictive Performance: The HD prior model achieved predictive performance (measured by Log-Likelihood, Brier Score, Tjur $R^2$, and Accuracy) comparable to standard priors (Independent Inverse-Gamma and standard PC priors). It showed slight improvements over Inverse-Gamma, particularly in accuracy.
• Variance Partitioning: The model successfully quantified the relative contributions of different factors.
  • Key Drivers: The analysis identified Spatial effects, Depth, and Bottom Temperature as the primary drivers of occurrence variability for the studied species.
  • Interpretability: The results aligned with ecological theory (e.g., depth acting as a proxy for pressure and substrate), validating the model's ability to capture known ecological gradients.
• Prior Sensitivity Analysis: The HD framework demonstrated superior transparency. By varying the Dirichlet concentration parameter ($q$), the authors showed how the model smoothly shifts toward sparser solutions (shrinking less important covariates to zero). This allowed for intuitive, direct control over model complexity, which is difficult to achieve with traditional variance priors.

5. Significance

• Enhanced Transparency: The framework shifts the focus from abstract variance scales to intuitive variance proportions, allowing ecologists to incorporate expert knowledge directly into the prior specification.
• Robustness: By using Uniform distributional assumptions for covariates rather than empirical data distributions, the variance contributions become portable and comparable across different studies and regions.
• Future Directions: The authors highlight that while the current study uses marginal (single-species) models, the HD framework is well-suited for extension to Joint Species Distribution Models (JSDMs), which account for biotic interactions and residual correlations between species.

In conclusion, this paper provides a principled, flexible, and interpretable Bayesian framework for SDMs, solving the long-standing issue of prior specification for variance components and enabling more robust ecological inference.