In fseaton/jsdmstan: Fitting jSDMs in Stan

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(jsdmstan)
library(ggplot2)
set.seed(8264372)

Joint Species Distribution Models

Joint Species Distibution Models, or jSDMs, are models that model an entire community of species simultaneously. The idea behind these is that they allow information to be borrowed across species, such that the covariance between species can be used to inform the predictions of distributions of related or commonly co-occurring species.

In plain language (or as plain as I can manage) jSDMs involve the modelling of an entire species community as a function of some combination of intercepts, covariate data and species covariance. Therefore the change of a single species is related to not only change in the environment but also how it relates to other species. There are several decisions to be made in how to specify these models - the standard decisions on which covariates to include, whether each species should have its own intercept (generally yes) and how to represent change across sites - but also how to represent the covariance between species. There are two options for representing this species covariance in this package. First, the original way of running jSDMs was to model the entire covariance matrix between species in a multivariate generalised linear mixed model (MGLMM). However, more recently there have been methods developed that involve representing the covariance matrix with a set of linear latent variables - known as generalised linear latent variable models (GLLVM).

The jsdmstan package aims to provide an interface for fitting these models in Stan using the Stan Hamiltonian Monte Carlo sampling as a robust Bayesian methodology.

Underlying maths

Feel free to skip this bit if you don't want to read equations, it is largely based on Warton et al. (2015). We model the community data $m_{ij}$ for each site $i$ and taxon $j$ as a function of a species intercept, environmental covariates and species covariance matrix:

$$ g(m_{ij}) = \beta_{0j} + \mathbf{x}i^\intercal\beta_j + u{ij} $$

where $g(\cdot)$ is the link function, $\mathbf{x}i^\intercal$ is the transpose of vector $\mathbf{x}$, and for each taxon $j$, $\beta{0j}$ is an intercept and $beta_j$ is a vector of regression coefficients related to measured predictors.

A site effect $\alpha_{i}$ can also be added to adjust for total abundance or richness:

$$ g(m_{ij}) = \alpha_{i} + \beta_{0j} + \mathbf{x}i^\intercal\beta_j + u{ij} $$

Multivariate Generalised Linear Mixed Models

The entire matrix of covariance between species is modelled in MGLMMs.

$$ u_{ij} \sim \mathrm{N}(\mathbf{0},\mathbf{\Sigma}) $$

Fitting the entire covariance matrix means that the amount of time required to fit these models scales with the number of species cubed, and the data required scales with the number of species squared. This makes these models both computationally and data intensive.

Generalised Linear Latent Variable Models

In response to some of these issues in fitting MGLMMs, GLLVMs were developed in which $u_{ij}$ is now specified as a linear function of a set of latent variables $\mathbf{z_i}$:

$$ y_{ij}|\mathbf{u}i \sim \mathrm{F}(m{ij},\phi_j) $$ $$ u_{ij} = \mathbf{z}_i^\intercal \lambda_j $$

The latent variables $\mathbf{z_i}$ are treated as random by assuming:

$$y_{ij}|\mathbf{z_i} \sim \mathrm{F}(m_{ij},\phi_j)) $$ $$\mathbf{z_i} \sim \mathrm{N}(\mathbf{0},\mathbf{1}) $$ Treating the species covariance as pulling from a set of latent variables greatly reduces the computational time required to fit these models.

Relationship to environmental covariates

Within jsdmstan the response of species to environmental covariates can either be unstructured (the default) or constrained by a covariance matrix between the environmental covariates. This second option (specified by setting beta_param = "cor") assumes that if one species is strongly positively related to multiple covariates then it is more likely that other species will either also be positively related to all these covariates, or negatively related. Mathematically this corresponds to:

$$\beta_j \sim \mathrm{N}(\mathbf{0},\mathbf{\Sigma})$$

Fitting a MGLMM

First we can use the in-built functions for simulating data according to the MGLMM model - we'll choose to simulate 15 species over 200 sites with 2 environmental covariates. The species are assumed to follow a Poisson distribution (with a log-link), and we use the defaults of including a species-specific intercept but no site-specific intercept. At the moment only default priors (standard normal distribution) are supported. We can do this using either the jsdm_sim_data() function with method = "mglmm" or with the mglmm_sim_data() function which just calls jsdm_sim_data() in the background.

nsites <- 75
nspecies <- 8
ncovar <- 2
mglmm_test_data <- mglmm_sim_data(N = nsites, S = nspecies, 
                                  K = ncovar, family = "pois")

This returns a list, which includes the Y matrix, the X matrix, plus also the exact parameters used to create the data:

names(mglmm_test_data)
dat <- as.data.frame(mglmm_test_data$X)

Now, to fit the model we can use the stan_jsdm() function, which interfaces to Stan through the rstan package. There are multiple ways to supply data to the stan_jsdm() function, one is to supply the data as a list with the appropriate named components (the jsdm_sim_data() functions supply data in the correct format already), the second way is to specify the Y and X matrices directly, and the third way is to use a formula for the environmental covariates and supply the environmental data to the data argument, which is what we'll use here:

mglmm_fit <- stan_jsdm(~ V1 + V2, data = dat, Y = mglmm_test_data$Y, 
                       family = "pois", method = "mglmm", refresh = 0)

If we print the model object we will get a brief overview of the type of jSDM and the data, plus if there are any parameters with Rhat > 1.01 or effective sample size ratio (Neff/N) < 0.05 then they will be printed:

mglmm_fit

To get a summary of all the model parameters we can use summary(), there are many parameters in these models so we just include a few here:

summary(mglmm_fit, pars = "cor_species")

To get a better overview of the R-hat and effective sample size we can use the mcmc_plot() function to plot histograms of R-hat and ESS.

mcmc_plot(mglmm_fit, plotfun = "rhat_hist")

mcmc_plot(mglmm_fit, plotfun = "neff_hist")

We can also examine the output for each parameter visually using a traceplot combined with a density plot, which is given by the default plot() command:

plot(mglmm_fit, ask = FALSE)

By default the plot() command plots all of the parameters with sigma or kappa in their name plus a random selection of 20 other parameters, but this can be overridden by either specifying the parameters by name (with or without regular expression matching) or changing the number of parameters to be randomly sampled. Use the get_parnames() function to get the names of parameters within a model - and the jsdm_stancode() function can also be used to see the underlying structure of the model.

All the mcmc plot types within bayesplot are supported by the mcmc_plot() function, and to see a full list either use bayesplot::available_mcmc() or run mcmc_plot() with an incorrect type and the options will be printed.

We can also view the environmental effect parameters for each species using the envplot() function.

envplot(mglmm_fit)

Posterior predictions can be extracted from the models using either posterior_linpred() or posterior_predict(), where the linpred function extracts the linear predictor for the community composition within each draw and the predict function combines this linear predictor extraction with a random generation based on the predicted probability for the family. Both functions by default return a list of length equal to the number of draws extracted, where each element of the list is a sites by species matrix.

mglmm_pp <- posterior_predict(mglmm_fit)
length(mglmm_pp)
dim(mglmm_pp[[1]])

As well as the MCMC plotting functions within bayesplot the ppc_ family of functions is also supported through the pp_check() function. This family of functions provides a graphical way to check your posterior against the data used within the model to evaluate model fit - called a posterior retrodictive check (or posterior predictive historically and when the prior only has been sampled from). To use these you need to have set save_data = TRUE within the stan_jsdm() call. Unlike in other packages by default pp_check() for jsdmStanFit objects extracts the posterior predictions then calculates summary statistics over the rows and plots those summary statistics against the same for the original data. The default behaviour is to calculate the sum of all the species per site - i.e. total abundance.

pp_check(mglmm_fit)

The summary statistic can be changed, as can whether it is calculated for every species or every site:

pp_check(mglmm_fit, summary_stat = "mean", calc_over = "species",
         plotfun = "ecdf_overlay")

We can examine the species-specific posterior predictive check through using multi_pp_check(), or examine how well the relationships between specific species are recovered using pp_check() with plotfun = "pairs".

As we have run the above model on simulated data and the original data list contains the parameters used to simulate the data we can use the mcmc_recover_ functions from bayesplot to see how the model did:

mcmc_plot(mglmm_fit, plotfun = "recover_hist",
          pars = paste0("sigmas_species[",1:8,"]"),
          true = mglmm_test_data$pars$sigmas_species)

mcmc_plot(mglmm_fit, plotfun = "recover_intervals",
          pars = paste0("cor_species[",rep(1:nspecies, nspecies:1),",",
                        unlist(sapply(1:8, ":",8)),"]"),
          true = c(mglmm_test_data$pars$cor_species[lower.tri(mglmm_test_data$pars$cor_species, diag = TRUE)])) +
  theme(axis.text.x = element_text(angle = 90))

Fitting a GLLVM

The model fitting workflow for latent variable models is very similar to that above, with the addition of specifying the number of latent variables (D) in the data simulation and model fit. Here we change the family to a Bernoulli family (i.e. the special case of the binomial where the number of trials is 1 for all observations), make the covariate effects on each species draw from a correlation matrix such that information can be shared across species, and change the prior to be a Student's T prior on the predictor-specific sigma parameter.

set.seed(3562251)
gllvm_data <- gllvm_sim_data(N = 50, S = 12, D = 2, K = 1,
                             family = "bernoulli",
                             beta_param = "cor",
                             prior = jsdm_prior(sigmas_preds = "student_t(3,0,1)"))

gllvm_fit <- stan_jsdm(Y = gllvm_data$Y, X = gllvm_data$X,
                       D = gllvm_data$D,  
                       family = "bernoulli",
                       method = "gllvm", 
                       beta_param = "cor",
                       prior = jsdm_prior(sigmas_preds = "student_t(3,0,1)"),
                       refresh = 0)
gllvm_fit

Again, the diagnostic statistics seem reasonable:

mcmc_plot(gllvm_fit, plotfun = "rhat_hist")

mcmc_plot(gllvm_fit, plotfun = "neff_hist")

For brevity's sake we will not go into the detail of the different functions again here, however there is one plotting function specifically for GLLVM models - ordiplot(). This plots the species or sites scores against the latent variables from a random selection of draws:

ordiplot(gllvm_fit, errorbar_range = 0.5)

ordiplot(gllvm_fit, type = "sites", geom = "text", errorbar_range = 0) +
  theme(legend.position = "none")

You can change the latent variables selected by specifying the choices argument, and alter the number of draws or whether you want to plot species or sites with the other arguments.

References

Warton et al (2015) So many variables: joint modeling in community ecology. Trends in Ecology & Evolution, 30:766-779. DOI: 10.1016/j.tree.2015.09.007.

Wilkinson et al (2021) Defining and evaluating predictions of joint species distribution models. Methods in Ecology and Evolution, 12:394-404. DOI: 10.1111/2041-210X.13518.

Vehtari, A., Gelman, A., and Gabry, J. (2017). Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Statistics and Computing. 27(5), 1413–1432. DOI: 10.1007/s11222-016-9696-4.

fseaton/jsdmstan documentation built on June 12, 2025, 7:02 a.m.