Environmetrics最新文献

We contribute to the discussion of the insightful article “Assessing predictability of environmental time series with statistical and machine learning models” by Bonas et al. (2024), in which the authors commend their effort in comparing a wide range of methodologies for the challenging task of predicting environmental time series data. We focus our discussion on two topics of interest to us. First, we consider extensions of the explored methodologies that allow for heteroscedastic error terms. Second, we consider non-Gaussianity and fitting models on transformed data. For both of these points, we will make use of the authors' supplied code and data in order to extend their examples. Ultimately, we find that modeling of heteroscedasticity error terms has the potential to improve both point and interval estimates for these environmental time series. We also find that the use of transformations to handle non-Gaussianity can improve interval estimates.

In dose-response modeling, several models can often yield satisfactory fits to the observed data. The current practice in risk assessment is to use model averaging, which is a way to combine multiple models in a weighted average. A key parameter in risk assessment is the benchmark dose, the dose resulting in a predefined abnormal change in response. Current practice when applying frequentist model averaging is to use weights based on the Akaike Information Criterion (AIC). This paper introduces stacking weights as an alternative for dose-response modeling and generalizes a Diversity Index from dichotomous to continuous responses for model space selection. Three simulation studies were conducted to evaluate the new methods. They showed that, in three realistic scenarios, recommended strategies generally performed well, with stacking weights outperforming AIC weights in several cases. Strategies involving model selection were less effective. However, in a challenging scenario, none of the methods performed well. Due to the promising results of stacking weights, they have been added to the R package “bmd.”

The relative merits of machine learning and statistical methods are discussed recently by Bonas et al. 2004, who raise important open questions for the statistical community regarding the value-added benefits of data science and the future role of environmental statistics. Specifically, they identify three major knowledge gaps where statistics is seen as crucial to strengthening inference in machine learning (ML): to provide an ML model-based framework amenable to explainability, to determine the best approach for quantifying uncertainty in relation to complex environmental dynamics, and to comprehensively identify ML's value-added benefits. We continue this discussion by exploring these general questions and sharing our perspective and insights from our modeling of marine and terrestrial ecosystem dynamics. We propose several lines of inquiry where environmental statisticians and data scientists could collaboratively advance predictive analytics.

Focusing on a sample of 39 countries in the period 1996–2017, we analyze whether the relationship between environmental taxes and CO₂ emissions depends on the quality of political institutions. Our results show that an increase in the environmental tax revenue is related to a reduction in CO₂ emissions only in countries with more consolidated democratic institutions, higher civil society participation, and less corrupt governments. Moreover, the relationship between CO₂ emissions and revenue neutral shifts to different tax sources depends not only on the quality of political institutions, but also on the kind of externality the policymaker aims at correcting.

In environmental science, practitioners often deal with data recorded sequentially along time, such as time series of wind direction or wind speed. In this context, clustering of time series is a useful tool to carry out exploratory analyses. While most of the proposals are focused on real-valued time series, very few works consider circular time series, despite the frequent appearance of these objects in many disciplines. In this manuscript, a dissimilarity for circular time series is introduced and used in combination with a soft clustering method. The metric relies on a measure of serial dependence considering circular arcs, thus taking advantage of the directional character inherent to the series range. The clustering approach is able to group together time series generated from similar stochastic processes. Some simulations show that the method exhibits a reasonable clustering effectiveness, outperforming alternative techniques in many contexts. Two interesting applications involving time series of wind direction in Saudi Arabia show the potential of the proposed approach.

Population growth models are essential tools for natural resources management and conservation since they provide understanding on factors affecting renewal of natural animal populations. However, we still do not properly understand how the processes underlying reproduction of natural animal populations are affected by the environment at the spatial scale at which reproduction actually happens. A particular challenge for analyzing these processes is that observations from different life cycle stages are often collected at different spatial scales, and there is a lack of statistical methods to link local and spatially aggregated information. We address this challenge by developing spatially explicit population growth models for annually reproducing fish. Our approach integrates mechanistic Ricker and Beverton–Holt population growth models with a zero-inflated species distribution model and utilizes the hierarchical Bayesian approach to estimate the model parameters from data with varying spatial support: local scale count data on offspring and environment, and areal data from commercial fisheries informing about a spawning stock size. We show, both theoretically and empirically, that our models are identifiable and have good inferential performance. As a proof of concept application, we used the proposed models to analyze the drivers of whitefish Coregonus laveratus (L.) s.l.) reproduction along the Finnish coast of the Gulf of Bothnia in the Baltic Sea. The results show that the proposed model provides novel understanding beyond what would be attainable with earlier methods. The distributions of the reproduction areas, spawner density, and maximum proliferation rate were strongly dependent on local environmental conditions, but the effects and the relative importance of the covariates varied between these processes. The proposed models can be extended to other systems and organisms and enable ecologists to extract a better understanding of processes driving animal reproduction.

Spatiotemporal causal inference methods are needed to detect the effect of interventions on indirectly measured epidemiological outcomes that go beyond studying spatiotemporal correlations. Chronic wasting disease (CWD) causes neurological degeneration and eventual death to white-tailed deer (Odocoileus virginianus) in Wisconsin. Targeted culling involves removing deer after traditional hunting seasons in areas with high CWD prevalence. The evaluation of the causal effects of targeted culling in the spread and growth of CWD is an important unresolved research and CWD management question that can guide surveillance efforts. Reaction–diffusion partial differential equations (PDEs) can be used to mechanistically model the underlying spatiotemporal dynamics of wildlife diseases, like CWD, allowing researchers to make inference about unobserved epidemiological quantities. These models indirectly regress spatiotemporal covariates on diffusion and growth rates parameterizing such PDEs, obtaining associational conclusions. In this work we develop an innovative method to obtain causal estimators for the effect of targeted culling interventions on CWD epidemiological processes using an inverse-probability-of-treatment-weighted technique by means of marginal structural models embedded in the PDE fitting process. Additionally we establish a novel scheme for sensitivity analysis under unmeasured confounder for testing the hypothesis of a significant causal effect in the indirectly measured epidemiological outcomes. Our methods can be broadly used to study the impact of spatiotemporal interventions and treatment exposures in the epidemiological evolution of infectious diseases that can help to inform future efforts to mitigate public health implications and wildlife disease burden.

Motivated by empirical case studies and discussions of Bonas et al. (2024), this discussion paper critically examines challenges in the predictability of environmental processes, focusing on three key spheres: (a) predictability and interpretability, (b) predictability in dynamic environments, and (c) predictability into unknown spaces. These spheres highlight the responsibilities within environmetrics to ensure that predictive models, particularly advanced machine learning and deep learning methods, are applied thoughtfully. First, we discuss the trade-off between interpretability and predictive complexity, contrasting the transparency of traditional statistical models with the “black-box” nature of machine learning but also highlighting their enormous potential for exploiting new data sources and types. Second, we address real-time adaptability, where models must handle concept drift and should, therefore, be continuously monitored. Finally, we consider the challenges of extrapolating predictions into unknown/nontrained areas, underscoring the risks of model overreach. This paper aims to contribute to the discussion in the field, emphasizing the critical role environmetricians play in advancing responsible, interpretable, and scientifically sound predictive practices.

Building on the insights from Bonas et al. (2024), we explore the relationship between statistical and machine learning models in the analysis of environmental time series. We specifically address the unique challenges of environmental time series data, including the need to consider the multivariate approach and account for spatial dependence. Emphasizing the importance of various types of statistical inference in environmental studies—not limited to forecasting—we propose that viewing statistical and machine learning approaches as complementary rather than alternative methods can unlock innovative modeling strategies that enhance both predictive accuracy and interpretive power. To illustrate these concepts, we present a case study that highlights the key points raised in the discussion.