Introduction to regression models with spatial and temporal correlation using R-INLA
R-INLA is an R package, specifically designed to solve a wide range of models using integrated nested Laplace approximation to account for spatial and spatio-temporally correlated datasets.
The last few years have seen a dramatic increase in the availability of spatial and spatio-temporally correlated data, due to advances in computational tools that allow the collection of real time data from e.g. satellites, GPS data, loggers or long term data studies.
For example, for epidemiological studies considering cases of disease incidence across a country, spatial similarity due to geographic features may be correlated with disease incidence. Cases may also be temporally correlated due to the number of cases at time t being dependent on the number of cases at time t-1. Therefore, most data collected by ecologists, epidemiologists or behaviourists is subject to dependency or autocorrelation, resulting in structure within the data. This is a problem, because many regression models assume that data points are independent, yet in real life this clearly is not the case.
Non-Bayesian approaches to handle dependency within models are limited, especially when non-normally distributed data is considered (which is common in ecology datasets which often consist of either proportional, bounded data or count data). The advent of powerful computation methods means that Bayesian modelling approaches such as Markov Chain Monte Carlo simulation are now possible on complex, large datasets. The advantage of Bayesian approaches is that of taking uncertainty in the estimates / predictions into account (important for models used for prediction), and flexibility when dealing with missing data. However, MCMC simulations are still computationally expensive, particularly when dealing with large, complex datasets. Integrated nested Laplace approximation (INLA) is an alternative, deterministic algorithm which provides accurate, and fast results.
How to handle spatial and spatio - temporally correlated data
The course is a thorough and intensive training in how to handle spatial and spatio – temporally correlated data.
Part of the course:
- mixed regression models (the mainstay of ecological data analysis);
- the fundamentals of Bayesian analysis;
- how to use R-INLA to account for dependency in models with a range of distributions.
This not only gives the grounding necessary to understand model construction and interpretation but also gives practical insight into how to construct the models in R.
This will allow participants to quickly and confidently move onto analysing their own data, which will result in considerable time savings and extend the complexity of analysis that individuals are able to carry out. For PhD students and post-doctoral researchers with their own data, this also offers an opportunity to get expert insight into individual problems.
Highland Statistics Ltd is a statistical consulting company based in the UK. They provide a number of courses teaching a range of different statistical approaches and have authored 10 statistical textbooks. Dr Alain Zuur and Dr Elena Iono are recognised as experts within the field, particularly in making complicated modelling approaches clear and accessible to ecologists without a strong mathematical background. During the course, several case studies are presented in which statistical theory is integrated with applied analysis.
The course is taught over 5 days: Monday 9 to Friday 13 April 2018.
Monday – Thursday involve a combination of theory and practical exercises from 9:00 h. til 16:00 h., while Friday will finish at 12:45 h.
Exercises and theory are introduced with a number of different case studies and minimal lectures. Although the book: “Beginners Guide to Spatial, Temporal and Spatio – temporal Ecological Data Analysis with R-INLA” is not included, copies of all the relevant chapters, the R code and the case studies are included. The teaching is done by a series of lectures and practical sessions, concentrating on practical application and ensuring that participants are able to understand, program and interpret R-INLA models.