spOccupancy fits single-species, multi-species, and integrated spatial
occupancy models using Markov Chain Monte Carlo (MCMC). Models are fit
using Póly-Gamma data augmentation. Spatial models are fit using either
Gaussian processes or Nearest Neighbor Gaussian Processes (NNGP) for
large spatial datasets. The package provides functionality for data
integration of multiple single-species occupancy data sets using a joint
likelihood framework. For multi-species models, spOccupancy provides
functions to account for residual species correlations in a joint
species distribution model framework while accounting for imperfect
detection. As of v0.4.0, spOccupancy
provides functions for
multi-season (i.e., spatio-temporal) single-species occupancy models.
Below we provide a very brief introduction to some of the package’s
functionality, and illustrate just one of the model fitting funcitons.
For more information, see the resources referenced at the bottom of this
page.
You can install the released version of spOccupancy
from
CRAN with:
install.packages("spOccupancy")
spOccupancy Function |
Description |
---|---|
PGOcc() |
Single-species occupancy model |
spPGOcc() |
Single-species spatial occupancy model |
intPGOcc() |
Single-species occupancy model with multiple data sources |
spIntPGOcc() |
Single-species spatial occupancy model with multiple data sources |
msPGOcc() |
Multi-species occupancy model |
spMsPGOcc() |
Multi-species spatial occupancy model |
lfJSDM() |
Joint species distribution model without imperfect detection |
sfJSDM() |
Spatial joint species distribution model without imperfect detection |
lfMsPGOcc() |
Multi-species occupancy model with species correlations |
sfMsPGOcc() |
Multi-species spatial occupancy model with species correlations |
intMsPGOcc() |
Multi-species occupancy model with multiple data sources |
tPGOcc() |
Single-species multi-season occupancy model |
stPGOcc() |
Single-species multi-season spatio-temporal occupancy model |
svcPGBinom() |
Single-species spatially-varying coefficient GLM |
svcPGOcc() |
Single-species spatially-varying coefficient occupancy model |
svcTPGBinom() |
Single-species spatially-varying coefficient multi-season GLM |
svcTPGOcc() |
Single-sepcies spatially-varying coefficient multi-season occupancy model |
postHocLM() |
Fit a linear (mixed) model using estimates from a previous model fit |
ppcOcc() |
Posterior predictive check using Bayesian p-values |
waicOcc() |
Compute Widely Applicable Information Criterion (WAIC) |
simOcc() |
Simulate single-species occupancy data |
simTOcc() |
Simulate single-species multi-season occupancy data |
simBinom() |
Simulate detection-nondetection data with perfect detection |
simTBinom() |
Simulate multi-season detection-nondetection data with perfect detection |
simMsOcc() |
Simulate multi-species occupancy data |
simIntOcc() |
Simulate single-species occupancy data from multiple data sources |
simIntMsOcc() |
Simulate multi-species occupancy data from multiple data sources |
To get started with spOccupancy
we load the package and an example
data set. We use data on twelve foliage-gleaning birds from the Hubbard
Brook Experimental Forest, which is
available in the spOccupancy
package as the hbef2015
object. Here we
will only work with one bird species, the Black-throated Blue Warbler
(BTBW), and so we subset the hbef2015
object to only include this
species.
library(spOccupancy)
data(hbef2015)
sp.names <- dimnames(hbef2015$y)[[1]]
btbwHBEF <- hbef2015
btbwHBEF$y <- btbwHBEF$y[sp.names == "BTBW", , ]
Below we fit a single-species spatial occupancy model to the BTBW data
using a Nearest Neighbor Gaussian Process. We use the default priors and
initial values for the occurrence (beta
) and regression (alpha
)
coefficients, the spatial variance (sigma.sq
), the spatial range
parameter (phi
), the spatial random effects (w
), and the latent
occurrence values (z
). We assume occurrence is a function of linear
and quadratic elevation along with a spatial random intercept. We model
detection as a function of linear and quadratic day of survey and linear
time of day the survey occurred.
# Specify model formulas
btbw.occ.formula <- ~ scale(Elevation) + I(scale(Elevation)^2)
btbw.det.formula <- ~ scale(day) + scale(tod) + I(scale(day)^2)
We run the model using an Adaptive MCMC sampler with a target acceptance
rate of 0.43. We run 3 chains of the model each for 10,000 iterations
split into 400 batches each of length 25. For each chain, we discard the
first 6000 iterations as burn-in and use a thinning rate of 4 for a
resulting 3000 samples from the joint posterior. We fit the model using
5 nearest neighbors and an exponential correlation function. We also
specify the k.fold
argument to perform 2-fold cross-validation after
fitting the full model. Run ?spPGOcc
for more detailed information on
all function arguments.
# Run the model
out <- spPGOcc(occ.formula = btbw.occ.formula,
det.formula = btbw.det.formula,
data = btbwHBEF, n.batch = 400, batch.length = 25,
accept.rate = 0.43, cov.model = "exponential",
NNGP = TRUE, n.neighbors = 5, n.burn = 2000,
n.thin = 4, n.chains = 3, verbose = FALSE, k.fold = 2)
This will produce a large output object, and you can use str(out)
to
get an overview of what’s in there. Here we use the summary()
function
to print a concise but informative summary of the model fit.
summary(out)
#>
#> Call:
#> spPGOcc(occ.formula = btbw.occ.formula, det.formula = btbw.det.formula,
#> data = btbwHBEF, cov.model = "exponential", NNGP = TRUE,
#> n.neighbors = 5, n.batch = 400, batch.length = 25, accept.rate = 0.43,
#> verbose = FALSE, n.burn = 2000, n.thin = 4, n.chains = 3,
#> k.fold = 2)
#>
#> Samples per Chain: 10000
#> Burn-in: 2000
#> Thinning Rate: 4
#> Number of Chains: 3
#> Total Posterior Samples: 6000
#> Run Time (min): 1.4416
#>
#> Occurrence (logit scale):
#> Mean SD 2.5% 50% 97.5% Rhat ESS
#> (Intercept) 3.9738 0.6010 3.0092 3.9023 5.3748 1.0144 164
#> scale(Elevation) -0.5209 0.2194 -0.9778 -0.5152 -0.0984 1.0003 1101
#> I(scale(Elevation)^2) -1.1552 0.2118 -1.6305 -1.1370 -0.7892 1.0070 238
#>
#> Detection (logit scale):
#> Mean SD 2.5% 50% 97.5% Rhat ESS
#> (Intercept) 0.6642 0.1146 0.4413 0.6634 0.8821 1.0014 5474
#> scale(day) 0.2939 0.0712 0.1524 0.2939 0.4335 0.9999 6000
#> scale(tod) -0.0301 0.0697 -0.1698 -0.0302 0.1063 1.0010 6435
#> I(scale(day)^2) -0.0747 0.0862 -0.2410 -0.0755 0.1033 1.0027 6000
#>
#> Spatial Covariance:
#> Mean SD 2.5% 50% 97.5% Rhat ESS
#> sigma.sq 1.1194 1.0085 0.2165 0.7676 4.1197 1.0419 88
#> phi 0.0072 0.0075 0.0007 0.0039 0.0270 1.1564 58
The function ppcOcc
performs a posterior predictive check on the
resulting list from the call to spPGOcc
. For binary data, we need to
perform Goodness of Fit assessments on some binned form of the data
rather than the raw binary data. Below we perform a posterior predictive
check on the data grouped by site with a Freeman-Tukey fit statistic,
and then use the summary
function to summarize the check with a
Bayesian p-value.
ppc.out <- ppcOcc(out, fit.stat = 'freeman-tukey', group = 1)
summary(ppc.out)
#>
#> Call:
#> ppcOcc(object = out, fit.stat = "freeman-tukey", group = 1)
#>
#> Samples per Chain: 10000
#> Burn-in: 2000
#> Thinning Rate: 4
#> Number of Chains: 3
#> Total Posterior Samples: 6000
#>
#> Bayesian p-value: 0.4917
#> Fit statistic: freeman-tukey
The waicOcc
function computes the Widely Applicable Information
Criterion (WAIC) for use in model selection and assessment (note that
due to Monte Carlo error your results will differ slightly).
waicOcc(out)
#> elpd pD WAIC
#> -681.52155 21.21463 1405.47236
Alternatively, we can perform k-fold cross-validation (CV) directly in
our call to spPGOcc
using the k.fold
argument and compare models
using a deviance scoring rule. We fit the model with k.fold = 2
and so
below we access the deviance scoring rule from the 2-fold
cross-validation. If we have additional candidate models to compare this
model with, then we might select for inference the one with the lowest
value of this CV score.
out$k.fold.deviance
#> [1] 1412.951
Prediction is possible using the predict
function, a set of occurrence
covariates at the new locations, and the spatial coordinates of the new
locations. The object hbefElev
contains elevation data across the
entire Hubbard Brook Experimental Forest. Below we predict BTBW
occurrence across the forest, which are stored in the out.pred
object.
# First standardize elevation using mean and sd from fitted model
elev.pred <- (hbefElev$val - mean(btbwHBEF$occ.covs[, 1])) / sd(btbwHBEF$occ.covs[, 1])
coords.0 <- as.matrix(hbefElev[, c('Easting', 'Northing')])
X.0 <- cbind(1, elev.pred, elev.pred^2)
out.pred <- predict(out, X.0, coords.0, verbose = FALSE)
The vignette("modelFitting")
provides a more detailed description and
tutorial of the core functions in spOccupancy
. For full statistical
details on the MCMC samplers for core functions in spOccupancy
, see
vignette("mcmcSamplers")
. In addition, see our recent
paper that describes the
package in more detail (Doser et al. 2022a). For a detailed description
and tutorial of joint species distribution models in spOccupancy
that
account for residual species correlations, see
vignette("factorModels")
, as well as vignette("mcmcFactorModels")
for full statistical details. For a description and tutorial of
multi-season (spatio-temporal) occupancy models in spOccupancy
, see
vignette("spaceTimeModels")
. For a tutorial on spatially-varying
coefficient models in spOccupancy
, see vignette("svcUnivariateHTML")
and take a look at our recent
pre-print that presents a series of
guidelines and recommendations for using spatially-varying coefficients
in species distribution models.
Doser, J. W., Finley, A. O., Kery, M., and Zipkin, E. F. (2022a). spOccupancy: An R package for single-species, multi-species, and integrated spatial occupancy models. Methods in Ecology and Evolution. https://doi.org/10.1111/2041-210X.13897.
Doser, J. W., Finley, A. O., and Banerjee, S. (2022b). Joint species distribution models with imperfect detection for high-dimensional spatial data. arXiv preprint arXiv:2204.02707.
Doser, J. W., Kery, M., Finley, A. O., Saunders, S. P., Weed, A. S., Zipkin, E. F. (2023). Guidelines for the use of spatially-varying coefficients in species distribution models. arXiv preprint arXiv:2301.05645.