DEPENDENCE BETWEEN LIANAS AND TREE HEALTH IN MINAS GERAIS: RIDGE-PENALIZED LOGISTIC AND MULTINOMIAL COPULA MODELS

Abstract

Introduction: This study presents an integrated statistical framework for the joint modeling of mixed ecological variables, simultaneously addressing issues of multicollinearity and dependence between response variables. Data from the National Forest Inventory of Minas Gerais were used to model liana presence (), a binary variable, and tree health status (TH), an ordinal categorical variable with four levels.

Results: The marginal distribution of  was fitted using logistic regression with the penalized AURPLCE estimator, which showed satisfactory predictive performance (AUC = 0.771). For the TH variable, multinomial logistic regression was applied. Dependence between the responses was modeled using copulas, with the Clayton copula selected based on the Akaike Information Criterion (AIC = −72.0142). Estimates of Kendall’s coefficient (τ = 0.1109) and Spearman’s rank correlation (ρₛ = 0.1707) indicated a positive but weak dependence. The joint structure, represented by a probability table and heat map, revealed co-occurrence patterns, highlighting a higher probability of healthy trees in the absence of lianas (estimated probability of 0.3655).

Conclusion: Overall, the results underscore the potential of combining penalized estimators and copula-based models for the robust analysis of forest data, providing relevant quantitative insights into the relationship between liana presence and tree phytosanitary condition.