Bayesian inference for spatial compositional data with exact zeros using gradient-informed MCMC
Pirzamanbin, Behnaz LU
(2026)
30th Nordic Conference in Mathematical Statistics
p.40-40
- Abstract
- Spatial compositional data with exact zeros arise in many applications
but remain challenging for models that often assume strictly positive
components. We develop a Bayesian spatial model that combines the
Dirichlet Composition Distribution, which accommodates exact zeros through
component-specific zero probabilities, with a Gaussian Markov random field representation
of the latent compositional field. The resulting framework supports
joint inference on zero occurrence, compositional structure, and spatial dependence.
For efficient posterior computation, we derive gradient-based expressions
in additive log-ratio coordinates and investigate how MCMC performance depends
on the update strategy for both the... (More) - Spatial compositional data with exact zeros arise in many applications
but remain challenging for models that often assume strictly positive
components. We develop a Bayesian spatial model that combines the
Dirichlet Composition Distribution, which accommodates exact zeros through
component-specific zero probabilities, with a Gaussian Markov random field representation
of the latent compositional field. The resulting framework supports
joint inference on zero occurrence, compositional structure, and spatial dependence.
For efficient posterior computation, we derive gradient-based expressions
in additive log-ratio coordinates and investigate how MCMC performance depends
on the update strategy for both the high-dimensional latent field and the
parameters.
We compare four latent-field samplers, MALA, pCN, gradient-informed pCN
(gpCN), and the auxiliary-gradient sampler, combined with joint and decoupled
updates for the Dirichlet precision parameter α, and random-walk Metropolis
updates for the spatial range parameter κ.
In a simulation study on a 35×35 grid with three compositional components
and structural zeros, only the Q-preconditioned samplers, gpCN and auxiliarygradient, reliably recover the spatial structure. In addition, decoupling α via slice sampling is essential: joint updates lead to slow mixing in α and downward bias in κ, while decoupled updates recover the true parameters more accurately.
We apply the best-performing strategy to the LANDCLIMII pollen-based REVEALS dataset for Europe, comprising 303 sites and 13 plant functional types
with zero rates ranging from 0.3% to 78.2%. Ten-fold cross-validation confirms
out-of-sample predictive performance, and stable parameter estimates across
folds. These results show that accurate spatial inference for compositional data
with exact zeros requires both a likelihood that respects zero structure and
MCMC updates adapted to the geometry of the latent field. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/c8efc44f-19d0-44d5-a7f8-86360decc257
- author
- Pirzamanbin, Behnaz
LU
- organization
- publishing date
- 2026-06
- type
- Contribution to conference
- publication status
- published
- subject
- pages
- 40 - 40
- conference name
- 30th Nordic Conference in Mathematical Statistics
- conference location
- Helsinki, Finland
- conference dates
- 2026-06-01 - 2026-06-04
- language
- English
- LU publication?
- yes
- id
- c8efc44f-19d0-44d5-a7f8-86360decc257
- date added to LUP
- 2026-06-04 10:52:31
- date last changed
- 2026-06-04 12:49:36
@misc{c8efc44f-19d0-44d5-a7f8-86360decc257,
abstract = {{Spatial compositional data with exact zeros arise in many applications<br/>but remain challenging for models that often assume strictly positive<br/>components. We develop a Bayesian spatial model that combines the<br/>Dirichlet Composition Distribution, which accommodates exact zeros through<br/>component-specific zero probabilities, with a Gaussian Markov random field representation<br/>of the latent compositional field. The resulting framework supports<br/>joint inference on zero occurrence, compositional structure, and spatial dependence.<br/>For efficient posterior computation, we derive gradient-based expressions<br/>in additive log-ratio coordinates and investigate how MCMC performance depends<br/>on the update strategy for both the high-dimensional latent field and the<br/>parameters.<br/>We compare four latent-field samplers, MALA, pCN, gradient-informed pCN<br/>(gpCN), and the auxiliary-gradient sampler, combined with joint and decoupled<br/>updates for the Dirichlet precision parameter α, and random-walk Metropolis<br/>updates for the spatial range parameter κ.<br/>In a simulation study on a 35×35 grid with three compositional components<br/>and structural zeros, only the Q-preconditioned samplers, gpCN and auxiliarygradient, reliably recover the spatial structure. In addition, decoupling α via slice sampling is essential: joint updates lead to slow mixing in α and downward bias in κ, while decoupled updates recover the true parameters more accurately.<br/>We apply the best-performing strategy to the LANDCLIMII pollen-based REVEALS dataset for Europe, comprising 303 sites and 13 plant functional types<br/>with zero rates ranging from 0.3% to 78.2%. Ten-fold cross-validation confirms<br/>out-of-sample predictive performance, and stable parameter estimates across<br/>folds. These results show that accurate spatial inference for compositional data<br/>with exact zeros requires both a likelihood that respects zero structure and<br/>MCMC updates adapted to the geometry of the latent field.}},
author = {{Pirzamanbin, Behnaz}},
language = {{eng}},
pages = {{40--40}},
title = {{Bayesian inference for spatial compositional data with exact zeros using gradient-informed MCMC}},
year = {{2026}},
}