The Latent Dirichlet Allocation model with covariates (LDAcov): A case study on the effect of fire on species composition in Amazonian forests.

Denis Valle,Divino Vicente Silverio,Leandro Maracahipes,Rafael Izbicki,Paulo Brando,Gilson Shimizu,Lucas N Paolucci,Yusuf Jameel

doi:10.1002/ece3.7626

Abstract

Understanding and predicting the effect of global change phenomena on biodiversity is challenging given that biodiversity data are highly multivariate, containing information from tens to hundreds of species in any given location and time. The Latent Dirichlet Allocation (LDA) model has been recently proposed to decompose biodiversity data into latent communities. While LDA is a very useful exploratory tool and overcomes several limitations of earlier methods, it has limited inferential and predictive skill given that covariates cannot be included in the model. We introduce a modified LDA model (called LDAcov) which allows the incorporation of covariates, enabling inference on the drivers of change of latent communities, spatial interpolation of results, and prediction based on future environmental change scenarios. We show with simulated data that our approach to fitting LDAcov is able to estimate well the number of groups and all model parameters. We illustrate LDAcov using data from two experimental studies on the long‐term effects of fire on southeastern Amazonian forests in Brazil. Our results reveal that repeated fires can have a strong impact on plant assemblages, particularly if fuel is allowed to build up between consecutive fires. The effect of fire is exacerbated as distance to the edge of the forest decreases, with small‐sized species and species with thin bark being impacted the most. These results highlight the compounding impacts of multiple fire events and fragmentation, a scenario commonly found across the southern edge of Amazon. We believe that LDAcov will be of wide interest to scientists studying the effect of global change phenomena on biodiversity using high‐dimensional datasets. Thus, we developed the R package LDAcov to enable the straightforward use of this model.

Highlights

Understanding and predicting how species composition has been and will be altered by global change phenomena is key to sustaining biodiversity and ecosystem functioning
We have described LDAcov, a novel model that can provide inference and prediction by embedding a regression structure within the standard Latent Dirichlet Allocation (LDA) model
We illustrate the use of this model on data from two fire experiments in the Brazilian Amazon, enabling inference on how fragmentation and fire jointly affect species composition of these forests

Summary

Introduction

Understanding and predicting how species composition has been and will be altered by global change phenomena is key to sustaining biodiversity and ecosystem functioning. Clustering and ordination have been the main approaches in ecology to reduce the dimensionality of biodiversity data (Legendre & Legendre, 2012). Hard-clustering approaches have dominated the field, few ecological theories predict the sharp delineations implied by these methods (Legendre & Legendre, 2012). These hard-clustering methods will assign a given location to a single group, limiting the ability of these approaches in detecting gradual changes in species composition across environmental gradients

Objectives

Findings

Conclusion