index.md

mvgam

MultiVariate (Dynamic) Generalized Addivite Models

The goal of mvgam is to fit Bayesian (Dynamic) Generalized Additive Models. This package constructs State-Space models that can include highly flexible nonlinear predictor effects for both process and observation components by leveraging functionalities from the impressive brms and mgcv packages. This allows mvgam to fit a wide range of models, including hierarchical ecological models such as N-mixture or Joint Species Distribution models, as well as univariate and multivariate time series models with imperfect detection. The original motivation for the package is described in Clark & Wells 2022 (published in Methods in Ecology and Evolution), with additional inspiration on the use of Bayesian probabilistic modelling coming from Michael Betancourt, Michael Dietze and Sarah Heaps, among many others.

Installation

Install the stable version from CRAN using: install.packages('mvgam'), or install the development version from GitHub using: devtools::install_github("nicholasjclark/mvgam"). Note that to actually condition models with MCMC sampling, the Stan software must be installed (along with either rstan and/or cmdstanr). Only rstan is listed as a dependency of mvgam to ensure that installation is less difficult. If users wish to fit the models using mvgam, please refer to installation links for Stan with rstan here, or for Stan with cmdstandr here. You will need a fairly recent version of Stan (preferably 2.29 or above) to ensure all the model syntax is recognized. We highly recommend you use Cmdstan through the cmdstanr interface as the backend. This is because Cmdstan is easier to install, is more up to date with new features, and uses less memory than Rstan. See this documentation from the Cmdstan team for more information.

Introductory seminar

<iframe src="https://www.youtube.com/embed/0zZopLlomsQ" width="533" height="300" frameborder="0" allowfullscreen="" data-external="1"></iframe>

Cheatsheet

Getting started

mvgam was originally designed to analyse and forecast non-negative integer-valued data (counts). These data are traditionally challenging to analyse with existing time-series analysis packages. But further development of mvgam has resulted in support for a growing number of observation families that extend to other types of data. Currently, the package can handle data for the following families:

• gaussian() for real-valued data
• student_t() for heavy-tailed real-valued data
• lognormal() for non-negative real-valued data
• Gamma() for non-negative real-valued data
• betar() for proportional data on (0,1)
• bernoulli() for binary data
• poisson() for count data
• nb() for overdispersed count data
• binomial() for count data with known number of trials
• beta_binomial() for overdispersed count data with known number of trials
• nmix() for count data with imperfect detection (unknown number of trials)

See ??mvgam_families for more information. Below is a simple example for simulating and modelling proportional data with Beta observations over a set of seasonal series with independent Gaussian Process dynamic trends:

set.seed(100)
data <- sim_mvgam(family = betar(),
                  T = 80,
                  trend_model = GP(),
                  prop_trend = 0.5, 
                  seasonality = 'shared')

Plot the series to see how they evolve over time

plot_mvgam_series(data = data$data_train, series = 'all')

Visualizing multivariate proportional time series using the mvgam R package #rstats

Fit a State-Space GAM to these series that uses a hierarchical cyclic seasonal smooth term to capture variation in seasonality among series. The model also includes series-specific latent Gaussian Processes with squared exponential covariance functions to capture temporal dynamics

mod <- mvgam(y ~ s(season, bs = 'cc', k = 7) +
               s(season, by = series, m = 1, k = 5),
             trend_model = GP(),
             data = data$data_train,
             newdata = data$data_test,
             family = betar())

Plot the estimated posterior hindcast and forecast distributions for each series

layout(matrix(1:4, nrow = 2, byrow = TRUE))
for(i in 1:3){
  plot(mod, type = 'forecast', series = i)
}

Forecasting multivariate time series with Dynamic Generalized Additive Models

Various S3 functions can be used to inspect parameter estimates, plot smooth functions and residuals, and evaluate models through posterior predictive checks or forecast comparisons. Please see the package documentation for more detailed examples.

Vignettes

You can set build_vignettes = TRUE when installing but be aware this will slow down the installation drastically. Instead, you can always access the vignette htmls online at https://nicholasjclark.github.io/mvgam/articles/

Other resources

A number of case studies and step-by-step webinars have been compiled to highlight how GAMs and DGAMs can be useful for analysing multivariate data:

• Time series in R and Stan using the mvgam package
• Ecological Forecasting with Dynamic Generalized Additive Models
• State-Space Vector Autoregressions in mvgam
• How to interpret and report nonlinear effects from Generalized Additive Models
• Phylogenetic smoothing using mgcv
• Distributed lags (and hierarchical distributed lags) using mgcv and mvgam
• Incorporating time-varying seasonality in forecast models

Please also feel free to use the mvgam Discussion Board to hunt for or post other discussion topics related to the package, and do check out the mvgam changelog for any updates about recent upgrades that the package has incorporated.

Interested in contributing?

I’m actively seeking PhD students and other researchers to work in the areas of ecological forecasting, multivariate model evaluation and development of mvgam. Please reach out if you are interested (n.clark’at’uq.edu.au). Other contributions are also very welcome, but please see The Contributor Instructions for general guidelines. Note that by participating in this project you agree to abide by the terms of its Contributor Code of Conduct.