Ecological Time Series Analysis and Forecasting in R

Time series analysis and forecasting are standard goals in applied ecology. In this course, you will learn how to wrangle, visualise and explore ecological time series. You will also learn to use the mvgam package to analyse a diversity of ecological time series to gain useful insights and produce accurate forecasts. This workshop will cover time series and time series visualization, Generalised Linear Models (GLMs) and hierarchial models (GLMMs), Generalized Additive Models (GAMs) for nonlinear effects and complex random effects, dynamic GLMs and dynamic GAMs, multivariate modeling strategies, and forecasting and forecast evaluation. This workshop is aimed at higher degree and early career ecologists who are interested in making better predictions with their statistical models. The strategies to be covered are extendable well beyond time series and participants will leave this workshop with a better understanding of strategies to capture the types of complex, nonlinear effects that dominate ecological data.

Time series modeling with Bayesian Dynamic Generalized Additive Models

In this talk I introduce Bayesian Dynamic Generalized Additive Models (DGAMs) and illustrate their advantages for analyzing and forecasting real-world time series. I discuss mvgam, an open-source R package that can fit DGAMs with nonlinear effects, hierarchical effects and dynamic processes to data from a wide variety of observation distributions. These models are especially useful for analysing multiple series, as they can estimate hierarchical smooth functions while learning complex temporal associations with latent vector autoregressive processes or dimension-reduced dynamic factor processes. Because the package uses Hamiltonian Monte Carlo inference through Stan, it is straightforward to create Stan code and all necessary data structures so that additional stochastic elements can be added to suit the user’s bespoke needs. Other key features of {mvgam} are functions to critique models using rolling window forecasts and posterior predictive checks, online data augmentation via a recursive particle filter and graphical tools to visualise probabilistic uncertainties for smooth functions and predictions. I hope show how models that describe real-world complexity, both through nonlinear covariate functions and multi-series dependence, are useful to ask targeted questions about drivers of change.

Ecological forecasting with dynamic GAMs

Time series analysis and forecasting are standard goals in applied ecology. But ecological forecasting is difficult because ecology is complex. The abundances of species, for example, fluctuate for many reasons. Food and shelter availability limit survival. Biotic interactions affect colonization and vital rates. Severe weather events and climate variation alter habitat suitability. These sources of variation make it difficult to understand, let alone predict, ecosystem change. Moreover, most available time series software cannot handle features that dominate ecological data, including overdispersion, clustering, missingness, discreteness and nonlinear effects. In this talk, I will introduce Dynamic Generalized Additive Models (DGAMs) as one solution to meet this complexity. I illustrate a number of models that can be tackled with the mvgam R package, which builds Stan code to specify probabilistic Bayesian models that include nonlinear smooth functions, random effects and dynamic processes, all with a simple interface that is familiar to most R users.

Ecological forecasting with dynamic Generalized Additive Models (DGAMs)

Time series analysis and forecasting are standard goals in applied ecology. But ecological forecasting is difficult because ecology is complex. The abundances of species, for example, fluctuate for many reasons. Food and shelter availability limit survival. Biotic interactions affect colonization and vital rates. Severe weather events and climate variation alter habitat suitability. These sources of variation make it difficult to understand, let alone predict, ecosystem change. Moreover, most available time series software cannot handle features that dominate ecological data, including overdispersion, clustering, missingness, discreteness and nonlinear effects. In this talk, I will introduce Dynamic Generalized Additive Models (DGAMs) as one solution to meet this complexity. I illustrate a number of models that can be tackled with the mvgam R package, which builds Stan code to specify probabilistic Bayesian models that include nonlinear smooth functions, random effects and dynamic processes, all with a simple interface that is familiar to most R users.

Ascertainment and near-term forecasting of tick paralysis admissions

Tick paralysis is a leading cause of emergency veterinary admissions for Australian companion animals, often resulting in death if left untreated. Availability of timely information on periods of increased risk can help modulate behaviours that reduce exposures to ticks and improve awareness of owners for the need of preventative ectoparasite treatment. However, detection of trends in risk is hampered by the lack of clearly annotated historical records of tick paralysis. Natural Language Processing (NLP) of clinical records is required to first ascertain historical cases. Here we describe a platform to perform NLP on VetCompass Australia’s veterinary clinical records to accurately identify historical cases of canine tick paralysis where we make use of combine bespoke spellchecking and tokenization routines with a domain-expertise inspired clinical dictionary to identify important terms in free text indicative of a tick paralysis diagnosis. Resulting time series of tick paralysis cases are then analysed using Dynamic Generalised Additive Models to jointly estimate nonlinear distributed lag effects of environmental predictors and dynamic latent temporal processes that facilitate probabilistic near-term forecasts of tick paralysis risk. Our models forecast tick paralysis cases with exceptional accuracy while preserving epidemiological interpretability. We have designed an interactive online dashboard to showcase our data and modelling results so that we can refine the way we present probabilistic predictions to meet end-user requirements. We expect our data acquisition / modelling pipeline to act as a platform for developing early warning systems that can notify clinicians and pet owners about heightened risks of environmentally driven veterinary conditions.