Skip to contents


Spatially varying coefficient models (SVCs, Gelfand et al. 2003) are often used to model data when relationships between dependent and independent variables are not uniform across space, a common situation when exploring phenomena across large spatial extents (Finley 2011). Meehan et al. (2019) described an SVC model to evaluate continent-scaled variation in bird abundance trends. The SVC model used in that analysis employed discrete aerial units (100 km grid cells), with spatial structure described by neighborhood matrices and spatial relationships described by an intrinsic conditional autoregressive model (Besag 1974). The online supplement for the paper included code for building the model using the R-INLA package (Rue et al. 2009) for the R statistical programming language (R Core Team 2021). Both the manuscript and code can be accessed at

In this vignette, we will describe how to build an SVC model similar to that described in Meehan et al. (2019), but within a continuous-space framework. This model will be computed using the stochastic partial differential equation (SPDE) approach of Lindgren et al. (2011, 2022), implemented with the inlabru interface to the R-INLA package for R. The SPDE approach employs a computationally efficient approximation of a Gaussian random field with parameters directly comparable to those of a Matérn covariance function. The benefits of a continuous-space versus a discrete-space SVC include the potential for finer resolution estimation and prediction, a better understanding of the range of spatial correlation, and a reduction in boundary effects associated with discrete-space analyses.

We build the model using a subset of the data described in Meehan et al. (2019). Specifically, we use counts of American Robin (Turdus migratorius) from south central North America collected between 1987 and 2016 during the Audubon Christmas Bird Count (CBC). The overall goal of the analysis is to produce spatially explicit estimates of annual relative abundance as well as long-term relative abundance trends for robins that account for spatial and temporal variation in count effort.


The model used to analyze these data assumes that counts come from a negative binomial distribution with an expected count and dispersion parameter. The expected count has a log-linear predictor:

\[\log(\lambda_{st}) = \kappa_{s} + \alpha_s + \epsilon_{s} \log[\text{Effort}_{st}] + \tau_{s} \text{Year}_{st}\]

where the natural log of expected count, \(\log(\lambda_{st})\), at site \(s\) during year \(t\), is modeled with a zero-centered, normally distributed intercept per site, \(\kappa_s\), a spatially varying intercept, \(\alpha_s\), a spatially varying effect of the log of count effort in hours, \(\epsilon_s\), and a spatially varying linear effect of year, \(\tau_s\). The spatially structured effects are modeled as Gaussian random fields with Matérn covariance functions with range and variance parameters.

Model parameters \(\kappa_s\), \(\alpha_s\), \(\epsilon_s\) and \(\tau_s\) are analogous to those in Meehan et al. (2019). For example, \(\kappa_s\) is included to account for site-level differences in counts, possibly due to habitat availability or observer experience. \(\alpha_s\) can be interpreted as an effort-corrected abundance index at year zero. \(\epsilon_s\) is the exponent for a power-law effort-correction function. And \(\tau_s\) is the long-term temporal trend at a given site.


To get started with data analysis, we set up the environment, by loading some packages, and setting some options.

# libraries
library(tidyterra) # raster plotting
# Note: the 'splancs' package also needs to be installed,
# but doesn't need to be loaded

# set option
select <- dplyr::select
options(scipen = 99999)
options(max.print = 99999)
options(stringsAsFactors = FALSE)

Next we define a coordinate reference system (CRS) for spatial analysis and create a base map for later use. The CRS uses the USA Contiguous Albers Equal-Area Conic projection, and is identified by the EPSG code 6703. We modify the CRS slightly to have units of kilometers, so that distances between widespread count sites are not especially large numbers (Krainski et al. 2018).

# define a crs
epsg6703km <- paste(
  "+proj=aea +lat_0=23 +lon_0=-96 +lat_1=29.5",
  "+lat_2=45.5 +x_0=0 +y_0=0 +datum=NAD83",
  "+units=km +no_defs"

# make a base map
states <- st_as_sf(maps::map("state", plot = FALSE, fill = TRUE)) %>%
  filter(ID %in% c(
    "texas", "oklahoma", "kansas", "missouri",
    "arkansas", "louisiana"
  )) %>%

Import data

Next we import some bird count data from the GitHub repository associated with Meehan et al. (2019), and turn the data set into spatially referenced points. We use a subset of the data (30 years, 6 US states) for this analysis to reduce computing time (~ 1 min). Note that site selection and zero filling, important components of trend analyses, have already been conducted and this is the resulting data set.

The inlabru package contains a pregenerated version of the data subset, called robins_subset, that can be accessed with data(robins_subset), to avoid accessing the full data online. The following code was used to generate the subset:

robins_subset <- read.csv(paste0(
)) %>%
    circle, bcr, state, year, std_yr, count, log_hrs,
    lon, lat, obs
  ) %>%
  mutate(year = year + 1899) %>%
    state %in% c(
    year >= 1987

We load the data, filter out observation sites with less than 20 years of data, add index variables to uniquely index site and year information, and transform the coordinates to the epsg6703km CRS:

count_dat <- robins_subset %>%
  mutate(site_idx = as.numeric(factor(paste(circle, lon, lat)))) %>%
  group_by(site_idx) %>%
  mutate(n_years = n()) %>%
  filter(n_years >= 20) %>%
  ungroup() %>%
    std_yr = year - max(year),
    obs = seq_len(nrow(.)),
    site_idx = as.numeric(factor(paste(circle, lon, lat))),
    year_idx = as.numeric(factor(year)),
    site_year_idx = as.numeric(factor(paste(circle, lon, lat, year)))
  ) %>%
  st_as_sf(coords = c("lon", "lat"), crs = 4326, remove = FALSE) %>%
  st_transform(epsg6703km) %>%
    easting = st_coordinates(.)[, 1],
    northing = st_coordinates(.)[, 2]
  ) %>%
  arrange(circle, year)

A rough view of changes in robin relative abundance, which does not account for variation in count effort, can be seen by plotting raw counts per site and year.

# map it
ggplot() +
    data = count_dat %>% filter(year_idx %in% seq(1, 30, 3)),
    aes(col = log(count + 1))
  ) +
  geom_sf(data = states, fill = NA) +
  coord_sf(datum = NA) +
  facet_wrap(~year) +
  scale_color_distiller(palette = "Spectral") +

Make spatial data

Next we use the count data to make a map of distinct count sites and save the coordinates of the sites, unique and across all years, for later spatial modeling and plotting.

# make a set of distinct study sites for mapping
site_map <- count_dat %>%
  select(circle, easting, northing) %>%
  distinct() %>%
  select(circle, easting, northing)

SPDE components

Computing a continuous-space model with R-INLA using the SPDE approach requires construction of four distinct sets of data and model objects (Blangiardo and Cameletti 2015, Krainski et al. 2018). First, we create a modeling mesh, which is used to provide a piecewise linear representation of the continuous spatial surface, based on a triangulation of the modeled region. Here, the same mesh will get reused for each of the spatial terms in the model. Second, we construct an SPDE model object that specifies properties of the spatial model. Again, we will use the same SPDE object for each of the spatial terms in the model.

In plain R-INLA, we would also need to create index vectors and projector matrices (or A matrices as they are often called). However, in inlabru, these objects are created automatically, and the user does not need to deal with them directly. This qorks by an automatic creation of a bru_mapper object that knows how to map between mesh nodes and spatial data locations. Thus, the only two R-INLA functions needed in the user code of the SPDE modelling steps are inla.mesh.2D() and inla.spde2.pcmatern(), and the inla.spde.make.index() and inla.spde.make.A() are only called internally by the inlabru code itself.

Modeling mesh

There are various things to consider when constructing a mesh (Lindgren and Rue 2015, Blangiardo and Cameletti 2015, Krainski et al. 2018, Bakka et al. 2018). In constructing one, we balance a trade-off between capturing fine-scaled features of the Gaussian random field and computing times. Here, we create a non-convex hull around the count sites, and then build a triangular mesh by specifying minimum and maximum edge lengths within and slightly outside the hull.

# make a hull and mesh for spatial model
hull <- inla.nonconvex.hull(
  points = as.matrix(st_coordinates(count_dat)),
  convex = 200, concave = 350
mesh <- inla.mesh.2d(
  boundary = hull, max.edge = c(100, 600), # km inside and outside
  cutoff = 50, offset = c(100, 300)
) # cutoff is min edge

# plot it
ggplot() +
  gg(data = mesh) +
  geom_sf(data = site_map, col = "darkgreen", size = 1) +
  geom_sf(data = states, fill = NA) +
  theme_bw() +
  labs(x = "", y = "")

SPDE model object

Next we create an SPDE object to define the model smoothness, with prior distributions for the variance and range parameters, and the mesh. Here we assume a Gaussian random field characterized with a Matérn covariance function with penalized complexity priors (Simpson et al. 2017) for the practical range (distance where spatial correlation approaches 0.1) and variation explained by the function (Fuglstad et al. 2019). The prior for the spatial range is set such that the probability of a range exceeding 500 km is 0.5. The prior for the variance explained by the spatial effect is set such that the probability of a standard deviation exceeding 1 is 0.5 (Krainski et al. 2018). If one wants to constrain this kind of spatial effect to integrate to zero, constr=TRUE should be added at this stage.

# make spde
spde <- inla.spde2.pcmatern(
  mesh = mesh,
  prior.range = c(500, 0.5),
  prior.sigma = c(1, 0.5)

Weighted spatial effects in inlabru

In ordinary R-INLA, it would be necessary to construct projector matrices for each model component, that would include the different covariate weightings for each spatial model component. In inlabru, this can instead be handled by either explicitly multiplying the spatial random fields with the spatial covariates in the model formula expression, or by specifying the weights argument for each model component specification. We will here use the latter approach, since it is the easiest.

In our model, \(\epsilon_s\) is a spatially varying effect of the log of count effort. Similarly, \(\tau_s\) is a spatially varying effect of year, so standardized year (with 1987 = 0) is also specified in the weights argument. \(\alpha_s\) is also an SVC, but it is a spatially varying intercept. For intercepts it is not necessary to specify a constant weight of 1. The model component \(\kappa_{s}\) is linked to each observation site and is not modeled on the mesh.

Note that the current use of the term ‘weights’ is different from that often encountered when defining mixed effect models in R. Here it is used to define covariate value multiplication, as opposed to importance values for likelihoods in other contexts.

Data stack for model fitting

In plain R-INLA, we would need to bundle all the model data and component projector matrix information using the inla.stack() function. In inlabru, this is done automatically, so we skip that step.

Model formula

The last required input for the analysis is the model formula, which includes information on the prior for explained variation for the unstructured random intercept. We define the prior for \(\kappa_s\) as a penalized complexity prior (Simpson et al. 2017), set such that the probability of the standard deviation associated with the random effect exceeding 1 is 0.01.

# iid prior
pc_prec <- list(prior = "pcprec", param = c(1, 0.1))

Notice that if one wants to constrain a spatial spde effect to integrate to zero, it should be added constr=TRUE in the SPDE model definition rather than in the component definition. As we want to constrain \(\kappa_s\) and it is a non-spatial term we can use constr=TRUE in its corresponding label() component definition below, which imposes a sum-to-zero constraint.

The model described above is translated to inlabru modeling syntax as:

# components
svc_components <- ~ -1 +
  kappa(site_idx, model = "iid", constr = TRUE, hyper = list(prec = pc_prec)) +
  alpha(geometry, model = spde) +
  eps(geometry, weights = log_hrs, model = spde) +
  tau(geometry, weights = std_yr, model = spde)
# formula, with "." meaning "add all the model components":
svc_formula <- count ~ .

Specifying the sf column geometry as input causes inlabru to textract the spatial coordinate information for the observation locations from. Alternatively, we could specify the function st_coordinates instead, but that would extract the raw coordinates, potentially losing important coordinate reference system information.. Alternatively, we can specify that CRS-free information explicitly as input, with cbind(easting, northing) or st_coordinates(.data.).

Here, we define the response as count, remove the automatic global intercept with a -1, and then specify the other terms in the model with label() statements. The first kappa() statement defines \(\kappa_s\), the site effect, as a normally distributed and independent (model="iid"), globally zero-centered (sum to zero constr=TRUE), deviation from \(\alpha_s\). The second alpha() statement defines \(\alpha_s\) as a spatially varying intercept with spatial structure described by the SPDE object called ‘spde’. Remember that if one wants to constrain this kind of spatial effect to integrate to zero, constr=TRUE should be added in the SPDE model definition rather than in the label() arguments. The third eps() statement defines \(\epsilon_s\) as an SVC for the effect of count effort, with spatial structure also described by the SPDE object. The weights for this spatially structured random slope are specified with the weights=log_hrs argument. The fourth tau() statement defines \(\tau_s\) as an SVC for the year effect, with spatial structure described in the SPDE object. The weights for this spatially structured random slope are specified with the weights=std_yr argument.

Run model

We estimate the model with a call to bru(). First we set the option to use the (new) experimental way to do internal computations, see Van Niekerk et. al. (2022), for the sake of computing speed and better numerics. In the call to bru(), we give the model components, and specify an observation likelihood with the model formula for the negative binomial distribution for the counts, and define the estimation data. Then we ask inla() to compute WAIC and CPO to evaluate model fit, and to save the information necessary for posterior sampling (config=TRUE, but this will otherwise be set by default by bru(), as inlabru::predict() relies on it for posterior sampling and prediction). For computing speed, we choose to use the simplified laplace approximation strategy and Empirical Bayes estimation. The inla() run for this model takes about 1 minutes on a standard laptop computer. Another option is to use the Variational Bayes approximation as detailed in Van Niekerk and Rue (2021) and Van Niekerk et. al. (2022).

res <- bru(svc_components,
    formula = svc_formula,
    family = "nbinomial",
    data = count_dat
  options = list(
    control.compute = list(waic = TRUE, cpo = TRUE),
    control.inla = list(int.strategy = "eb"),
    verbose = FALSE

Model summaries

Once computation is complete, we look at the initial results to see how things went. First we check the posterior means for the hyperparameters of the model, mainly the variance components and the spatial ranges of the spatially structured parameters.

# view results
res$summary.hyperpar[-1, c(1, 2)]
#>                              mean            sd
#> Precision for kappa    2.17689882    0.40782427
#> Range for alpha     1039.41326452  357.19395244
#> Stdev for alpha        2.19127220    0.54454709
#> Range for eps       5328.11285299 3703.36443079
#> Stdev for eps          0.25354483    0.12234129
#> Range for tau       1167.36808250  525.11789696
#> Stdev for tau          0.07834748    0.02440056

Next we examine some summaries of the random effect estimates, starting with \(\exp(\alpha_s)\), which is effort-corrected relative abundance at year = 0 (1987), given 1 hour of count effort (i.e., log[1]=0).

summary(exp(res$summary.random$alp$"0.5quant")) # exp(alpha) posterior median
#>     Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
#>  0.03151  1.13932  3.71693  7.61947  9.57583 59.80462

Note that to avoid issues due to \(E[\exp(x)|y]\neq\exp[E(x|y)]\), we use the posterior median instead of the posterior mean.

The summary for \(\epsilon_s\) shows variation in the exponent for the effort correction function across space.

summary(res$summary.random$eps$mean) # epsilon
#>    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#>  0.7527  0.8715  0.9643  0.9529  1.0229  1.1609

The \(\tau_s\) summary shows how long-term, log-linear trends of robin relative abundance have varied across space, from annual decreases of around 10% to annual increases of around 10%.

summary((exp(res$summary.random$tau$"0.5quant") - 1) * 100) # (exp(tau)-1)*100
#>    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#> -14.262  -5.945  -1.997  -1.501   3.059  13.291

SVC maps

Next we create maps of \(\alpha_s\), \(\epsilon_s\), and \(\tau_s\) to inspect the spatial structure of these parameter estimates. We start by creating a 25-km mapping grid, and then projecting this mapping grid to the modeling mesh.

# get easting and northing limits
xlim <- range(hull$loc[, 1])
ylim <- range(hull$loc[, 2])
grd_dims <- round(c(x = diff(range(xlim)), y = diff(range(ylim))) / 25)

# make mesh projector to get model summaries from the mesh to the mapping grid
mesh_proj <- fm_evaluator(
  xlim = xlim, ylim = ylim, dims = grd_dims

Then we populate the mapping grids with parameter estimates (posterior median and range95), turn them into a raster stack, and mask the raster stack to the study area.

# pull data
kappa <- data.frame(
  median = exp(res$summary.random$kappa$"0.5quant"),
  range95 = exp(res$summary.random$kappa$"0.975quant") -
alph <- data.frame(
  median = exp(res$summary.random$alpha$"0.5quant"),
  range95 = exp(res$summary.random$alpha$"0.975quant") -
epsi <- data.frame(
  median = res$summary.random$eps$"0.5quant",
  range95 = (res$summary.random$eps$"0.975quant" -
taus <- data.frame(
  median = (exp(res$summary.random$tau$"0.5quant") - 1) * 100,
  range95 = (exp(res$summary.random$tau$"0.975quant") -
    exp(res$summary.random$tau$"0.025quant")) * 100

# loop to get estimates on a mapping grid
pred_grids <- lapply(
  list(alpha = alph, epsilon = epsi, tau = taus),
  function(x) as.matrix(fm_evaluate(mesh_proj, x))
# make a terra raster stack with the posterior median and range95
out_stk <- rast()
for (j in 1:3) {
  mean_j <- cbind(expand.grid(x = mesh_proj$x, y = mesh_proj$y),
    Z = c(matrix(pred_grids[[j]][, 1], grd_dims[1]))
  mean_j <- rast(mean_j, crs = epsg6703km)
  range95_j <- cbind(expand.grid(X = mesh_proj$x, Y = mesh_proj$y),
    Z = c(matrix(pred_grids[[j]][, 2], grd_dims[1]))
  range95_j <- rast(range95_j, crs = epsg6703km)
  out_j <- c(mean_j, range95_j)
  terra::add(out_stk) <- out_j
names(out_stk) <- c(
  "alpha_median", "alpha_range95", "epsilon_median",
  "epsilon_range95", "tau_median", "tau_range95"
out_stk <- terra::mask(out_stk, states, touches = FALSE)

Finally, we plot the SVCs with the following code. We plot the posterior median and 95% uncertainty width (“range95”) for \(\exp(\kappa_s)\), \(\exp(\alpha_s)\), \(\epsilon_s\), and \(100(\exp(\tau_s)-1)\).

make_plot_field <- function(data_stk, scale_label) {
  ggplot(states) +
    geom_sf(fill = NA) +
    coord_sf(datum = NA) +
    geom_spatraster(data = data_stk) +
    labs(x = "", y = "") +
      palette = "Spectral",
      na.value = "transparent"
    ) +
    theme_bw() +
    geom_sf(fill = NA)
make_plot_site <- function(data, scale_label) {
  ggplot(states) +
    geom_sf() +
    coord_sf(datum = NA) +
    geom_sf(data = data, size = 1, mapping = aes(colour = value)) +
    scale_colour_distiller(scale_label, palette = "Spectral") +
    labs(x = "", y = "") +
    theme_bw() +
    geom_sf(fill = NA)

# medians
# fields alpha_s, epsilon_s, tau_s
pa <- make_plot_field(
  data_stk = out_stk[["alpha_median"]],
  scale_label = "posterior\nmedian\nexp(alpha_s)"
pe <- make_plot_field(
  data_stk = out_stk[["epsilon_median"]],
  scale_label = "posterior\nmedian\nepsilon_s"
pt <- make_plot_field(
  data_stk = out_stk[["tau_median"]],
  scale_label = "posterior\nmedian\n100(exp(tau_s)-1)"
# sites kappa_s
ps <- make_plot_site(
  data = cbind(site_map, data.frame(value = kappa$median)),
  scale_label = "posterior\nmedian\nexp(kappa_s)"
# range95
# fields alpha_s, epsilon_s, tau_s
pa_range95 <- make_plot_field(
  data_stk = out_stk[["alpha_range95"]],
  scale_label = "posterior\nrange95\nexp(alpha_s)"
pe_range95 <- make_plot_field(
  data_stk = out_stk[["epsilon_range95"]],
  scale_label = "posterior\nrange95\nepsilon_s"
pt_range95 <- make_plot_field(
  data_stk = out_stk[["tau_range95"]],
  scale_label = "posterior\nrange95\n100(exp(tau_s)-1)"
# sites kappa_s
ps_range95 <- make_plot_site(
  data = cbind(site_map, data.frame(value = kappa$range95)),
  scale_label = "posterior\nrange95\nexp(kappa_s)"
# plot together
multiplot(ps, pa, pe, pt, cols = 2)

The map for the posterior mean of \(\tau_s\) shows that robins have decreased in the southern part of the study area and increased in the northern part. This demonstrates how the wintering range of robins is shifting northward as winters become warmer due to climate change.

# plot together
multiplot(ps_range95, pa_range95, pe_range95, pt_range95, cols = 2)

More information

More information on building spatial models using the SPDE approach in R-INLA can be found in Lindgren and Rue (2015), Blangiardo and Camaletti (2015), Bakka et al. (2018), Krainski et al. (2018), and Moraga (2019).


Bakka, H., Rue, H., Fuglstad, G.A., Riebler, A., Bolin, D., Illian, J., Krainski, E., Simpson, D. and Lindgren, F., 2018. Spatial modeling with R‐INLA: A review. Wiley Interdisciplinary Reviews: Computational Statistics, 10(6), p.e1443.

Besag, J., 1974. Spatial interaction and the statistical analysis of lattice systems. Journal of the Royal Statistical Society: Series B (Methodological), 36(2), pp.192-225.

Blangiardo, M., Cameletti, M., Baio, G. and Rue, H., 2013. Spatial and spatio-temporal models with R-INLA. Spatial and spatio-temporal epidemiology, 4, pp.33-49.

Finley, A.O., 2011. Comparing spatially‐varying coefficients models for analysis of ecological data with non‐stationary and anisotropic residual dependence. Methods in Ecology and Evolution, 2(2), pp.143-154.

Fuglstad, G.A., Simpson, D., Lindgren, F. and Rue, H., 2019. Constructing priors that penalize the complexity of Gaussian random fields. Journal of the American Statistical Association, 114(525), pp.445-452.

Gelfand, A.E., Kim, H.J., Sirmans, C.F. and Banerjee, S., 2003. Spatial modeling with spatially varying coefficient processes. Journal of the American Statistical Association, 98(462), pp.387-396.

Gómez-Rubio, V., 2020. Bayesian inference with INLA. CRC Press.

Krainski, E., Gómez-Rubio, V., Bakka, H., Lenzi, A., Castro-Camilo, D., Simpson, D., Lindgren, F. and Rue, H., 2018. Advanced spatial modeling with stochastic partial differential equations using R and INLA. Chapman and Hall/CRC.

Lindgren, F., Rue, H. and Lindström, J., 2011. An explicit link between Gaussian fields and Gaussian Markov random fields: the stochastic partial differential equation approach. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73(4), pp.423-498.

Lindgren, F. and Rue, H., 2015. Bayesian spatial modelling with R-INLA. Journal of statistical software, 63, pp.1-25.

Lindgren, F., Bolin, D. and Rue, H., 2022. The SPDE approach for Gaussian and non-Gaussian fields: 10 years and still running. Spatial Statistics, p.100599.

Link, W.A., Sauer, J.R. and Niven, D.K., 2006. A hierarchical model for regional analysis of population change using Christmas Bird Count data, with application to the American Black Duck. The Condor, 108(1), pp.13-24.

Meehan, T.D., Michel, N.L. and Rue, H., 2019. Spatial modeling of Audubon Christmas Bird Counts reveals fine‐scale patterns and drivers of relative abundance trends. Ecosphere, 10(4), p.e02707.

Moraga, P., 2019. Geospatial health data: Modeling and visualization with R-INLA and shiny. CRC Press.

R Core Team (2021). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria.

Rue, H., Martino, S. and Chopin, N., 2009. Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations. Journal of the royal statistical society: Series b (statistical methodology), 71(2), pp.319-392.

Simpson, D., Rue, H., Riebler, A., Martins, T.G. and Sørbye, S.H., 2017. Penalising model component complexity: A principled, practical approach to constructing priors. Statistical science, 32(1), pp.1-28.

Soykan, C.U., Sauer, J., Schuetz, J.G., LeBaron, G.S., Dale, K. and Langham, G.M., 2016. Population trends for North American winter birds based on hierarchical models. Ecosphere, 7(5), p.e01351.

Van Niekerk, J. and Rue, H., 2021. Correcting the Laplace Method with Variational Bayes. Journal of Machine Learning Research, Under review.

Van Niekerk, J. and Krainski, E. T. and Rustand, D. and Rue, H., 2022. A new avenue for Bayesian inference with INLA. Submitted.