In this vignette we will introduce how to fit Whittle–Matérn fields
with general smoothness based on finite element and rational
approximations. The theory for this approach is provided in Bolin,
Kovács, et al. (2023) and Bolin, Simas, et al. (2023). For the
implementation, we make use of the rSPDE
package for the rational approximations.
These models are thus implemented using finite element approximations. Such approximations are not needed for integer smoothness parameters, and for the details about the exact models we refer to the vignettes
For details on the construction of metric graphs, see Working with metric graphs
For further details on data manipulation on metric graphs, see Data manipulation on metric graphs
We begin by loading the rSPDE and
MetricGraph packages:
As an example, we consider the following metric graph
edge1 <- rbind(c(0,0),c(1,0))
edge2 <- rbind(c(0,0),c(0,1))
edge3 <- rbind(c(0,1),c(-1,1))
theta <- seq(from=pi,to=3*pi/2,length.out = 20)
edge4 <- cbind(sin(theta),1+ cos(theta))
edges = list(edge1, edge2, edge3, edge4)
graph <- metric_graph$new(edges = edges)
graph$plot()To construct a FEM approximation of a Whittle–Matérn field with general smoothness, we must first construct a mesh on the graph.
In the command build_mesh, the argument h
decides the largest spacing between nodes in the mesh. As can be seen in
the plot, the mesh is very coarse, so let’s reduce the value of
h and rebuild the mesh:
We are now ready to specify the model \[
(\kappa^2 - \Delta)^{\alpha/2} \tau u = \mathcal{W}
\] for the Whittle–Matérn field \(u\). For this, we use the
matern.operators function from the rSPDE
package:
sigma <- 1.3
range <- 0.15
nu <- 0.8
rspde.order <- 2
op <- matern.operators(nu = nu, range = range, sigma = sigma,
parameterization = "matern",
m = rspde.order, graph = graph) As can be seen in the code, we specify \(\kappa\) via the practical correlation
range \(\sqrt{8\nu}/\kappa\). Also, the
model is not parametrized by \(\tau,
\alpha\) but instead by \(\sigma,
\nu\). Here, sigma denotes the standard deviation of
the field and nu is the smoothness parameter, which is
related to \(\alpha\) via the relation
\(\alpha = \nu + 1/2\). The object
op contains the matrices needed for evaluating the
distribution of the stochastic weights in the FEM approximation.
Let us simulate the field \(u\) at the mesh locations and plot the result:
u <- simulate(op)
df_u <- data.frame(u = as.vector(u), edge_number = graph$mesh$VtE[,1],
distance_on_edge = graph$mesh$VtE[,2])
df_u <- graph$process_data(data = df_u, normalized = TRUE)
graph$plot_function(data = "u", newdata = df_u, type = "plotly")If we want to evaluate \(u(s)\) at
some locations \(s_1,\ldots, s_n\), we
need to multiply the weights with the FEM basis functions \(\varphi_i(s)\) evaluated at the locations.
For this, we can construct the observation matrix \(\boldsymbol{\mathrm{A}}\), with elements
\(A_{ij} = \varphi_j(s_i)\), which
links the FEM basis functions to the locations. This can be done by the
function fem_basis in the metric graph object. To
illustrate this, let us simulate some observation locations on the graph
and construct the matrix:
obs.per.edge <- 100
obs.loc <- NULL
for(i in 1:graph$nE) {
obs.loc <- rbind(obs.loc,
cbind(rep(i,obs.per.edge), runif(obs.per.edge)))
}
n.obs <- obs.per.edge*graph$nE
A <- graph$fem_basis(obs.loc)In the code, we generate \(100\) observation locations per edge in the graph, drawn at random. It can be noted that we assume that the observation locations are given in the format \((e, d)\) where \(e\) denotes the edge of the observation and \(d\) is the position on the edge, i.e., the relative distance from the first vertex of the edge.
To compute the precision matrix from the covariance-based rational
approximation one can use the precision() method on object
returned by the matern.operators() function:
As an illustration of the model, let us compute the covariance
function between the process at \(s=(2,0.1)\), that is, the point at edge 2
and distance on edge 0.1, and all the other mesh points. To this end, we
can use the helper function cov_function_mesh that is
contained in the op object:
c_cov <- cov_function_mesh(op, matrix(c(2,0.1),1,2))
df_c_cov <- data.frame(cov = as.vector(c_cov), edge_number = graph$mesh$VtE[,1],
distance_on_edge = graph$mesh$VtE[,2])
df_c_cov <- graph$process_data(data = df_c_cov, normalized = TRUE)
graph$plot_function(data = "cov", newdata = df_c_cov, type = "plotly")There is built-in support for computing log-likelihood functions and
performing kriging prediction in the rSPDE package which we
can use for the graph model. To illustrate this, we use the simulation
to create some noisy observations of the process. We generate the
observations as \(Y_i = 1 + 2x_{i1} - 3 x_{i2}
+ u(s_i) + \varepsilon_i\), where \(\varepsilon_i \sim N(0,\sigma_e^2)\) is
Gaussian measurement noise, \(x_1\) and
\(x_2\) are covariates generated from
the relative positions of the observations on the graph.
sigma.e <- 0.1
x1 <- obs.loc[,1]
x2 <- obs.loc[,2]
Y <- 1 + 2*x1 - 3*x2 + as.vector(A %*% u + sigma.e * rnorm(n.obs))Let us now fit the model. To this end we will use the
graph_lme() function (that, for the finite element models,
acts as a wrapper for the rspde_lme() function from the
rSPDE package). To this end, let us now assemble the
data.frame() with the observations, the observation
locations and the covariates:
df_data <- data.frame(y = Y, edge_number = obs.loc[,1],
distance_on_edge = obs.loc[,2],
x1 = x1, x2 = x2)Let us now add the data to the graph object and plot it:
## Adding observations...
## Assuming the observations are normalized by the length of the edge.
We can now fit the model. To this end, we use the
graph_lme() function and set the model to
'WM’.
Let us obtain a summary of the model:
##
## Latent model - Whittle-Matern
##
## Call:
## graph_lme(formula = y ~ x1 + x2, graph = graph, model = "WM")
##
## Fixed effects:
## Estimate Std.error z-value Pr(>|z|)
## (Intercept) 0.2234 0.7177 0.311 0.75553
## x1 2.1522 0.2009 10.712 < 2e-16 ***
## x2 -2.1239 0.7104 -2.990 0.00279 **
##
## Random effects:
## Estimate Std.error z-value
## alpha 1.25577 NaN NaN
## tau 0.05903 NaN NaN
## kappa 15.32439 2.28257 6.714
##
## Random effects (Matern parameterization):
## Estimate Std.error z-value
## nu 0.75577 NA NA
## sigma 1.32640 0.13695 9.686
## range 0.16046 0.02528 6.348
##
## Measurement error:
## Estimate Std.error z-value
## std. dev 0.097207 0.006301 15.43
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Log-Likelihood: -126.5583
## Number of function calls by 'optim' = 56
## Optimization method used in 'optim' = L-BFGS-B
##
## Time used to: fit the model = 15.2974 secs
An improved estimate of the Hessian can be obtained by setting
improve_hessian to TRUE, which improves the
precision of the standard errors.
Let us obtain a summary of the model fitted with the improved Hessian:
##
## Latent model - Whittle-Matern
##
## Call:
## graph_lme(formula = y ~ x1 + x2, graph = graph, model = "WM",
## improve_hessian = TRUE)
##
## Fixed effects:
## Estimate Std.error z-value Pr(>|z|)
## (Intercept) 0.2234 0.7198 0.310 0.75623
## x1 2.1522 0.2021 10.649 < 2e-16 ***
## x2 -2.1239 0.7105 -2.989 0.00279 **
##
## Random effects:
## Estimate Std.error z-value
## alpha 1.25577 0.06205 20.237
## tau 0.05903 0.01667 3.541
## kappa 15.32439 3.11678 4.917
##
## Random effects (Matern parameterization):
## Estimate Std.error z-value
## nu 0.75577 0.06205 12.179
## sigma 1.32640 0.13695 9.686
## range 0.16046 0.02528 6.348
##
## Measurement error:
## Estimate Std.error z-value
## std. dev 0.097207 0.006311 15.4
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Log-Likelihood: -126.5583
## Number of function calls by 'optim' = 56
## Optimization method used in 'optim' = L-BFGS-B
##
## Time used to: fit the model = 15.03612 secs
## compute the Hessian = 2.27229 secs
We can also obtain additional information by using the function
glance():
## # A tibble: 1 × 9
## nobs sigma logLik AIC BIC deviance df.residual model alpha
## <int> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <chr> <dbl>
## 1 400 0.0972 -127. 267. 295. 253. 393 WhittleMatern 1.26
Let us compare the values of the parameters of the latent model with the true ones:
print(data.frame(sigma = c(sigma, fit$alt_par_coeff$coeff["sigma"]),
range = c(range, fit$alt_par_coeff$coeff["range"]),
nu = c(nu, fit$alt_par_coeff$coeff["nu"]),
row.names = c("Truth", "Estimates")))## sigma range nu
## Truth 1.300000 0.1500000 0.8000000
## Estimates 1.326403 0.1604563 0.7557713
Given that we have estimated the parameters, let us compute the kriging predictor of the field given the observations at the mesh nodes.
We will perform kriging with the predict() method. To
this end, we need to provide a data.frame containing the
prediction locations, as well as the values of the covariates at the
prediction locations.
df_pred <- data.frame(edge_number = graph$mesh$VtE[,1],
distance_on_edge = graph$mesh$VtE[,2],
x1 = graph$mesh$VtE[,1],
x2 = graph$mesh$VtE[,2])
u.krig <- predict(fit, newdata = df_pred, normalized = TRUE)The estimate is shown in the following figure
df_krig <- data.frame(mean = as.vector(u.krig$mean), edge_number = graph$mesh$VtE[,1],
distance_on_edge = graph$mesh$VtE[,2])
df_krig <- graph$process_data(data = df_krig, normalized = TRUE)
graph$plot_function(data = "mean", newdata = df_krig, type = "plotly")We can also use the augment() function to easily plot
the predictions. Let us a build a 3d plot now and add the observed
values on top of the predictions:
graph_lmeThe graph_lme function provides flexibility in model
fitting by allowing users to fix certain parameters at specific values
or set custom starting values for the optimization process. This can be
useful when you have prior knowledge about some parameters or when you
want to improve convergence by providing better starting points.
Parameters can be fixed by using the model_options
argument with elements of the form fix_parname = value,
where parname is the name of the parameter you want to fix.
The parameters that can be fixed in the model_options list
are:
fix_sigma_e: Fix the standard deviation of the noise
parameter \(\sigma_\varepsilon\)fix_sigma: Fix the standard deviation parameter \(\sigma\)fix_range: Fix the range parameterfix_alpha: Fix the fractional power \(\alpha\)Similarly, you can set starting values for parameters using elements
of the form start_parname = value in the
model_options list. This is particularly useful when the
default starting values might be far from the optimal values, which
could lead to slow convergence or convergence to a local minimum.
start_sigma_e: Starting value for the standard
deviation of the noise parameter \(\sigma_\varepsilon\)start_sigma: Starting value for the standard deviation
parameter \(\sigma\)start_range: Starting value for the range
parameterstart_alpha: Starting value for the fractional power
\(\alpha\)Let us demonstrate how to use these parameters with a graph model example. We will fix the standard deviation \(\sigma\) to 1 and set the starting value for the range parameter to 0.5, on the previous example.
# Fit model with fixed sigma and start_range
fit_fixed <- graph_lme(y ~ x1 + x2, graph = graph, model = "WM",
model_options = list(
fix_sigma = 1, # Fix sigma to 1
start_range = 0.5 # Set starting value for range
))
# Print summary of the model
summary(fit_fixed)##
## Latent model - Whittle-Matern
##
## Call:
## graph_lme(formula = y ~ x1 + x2, graph = graph, model = "WM",
## model_options = list(fix_sigma = 1, start_range = 0.5))
##
## Fixed effects:
## Estimate Std.error z-value Pr(>|z|)
## (Intercept) 0.2461 0.4741 0.519 0.604
## x1 2.1407 0.1340 15.973 < 2e-16 ***
## x2 -2.1519 0.5063 -4.250 2.14e-05 ***
##
## Random effects:
## Estimate Std.error z-value
## nu 0.72823 0.01055 69.00
## sigma (fixed) 1.00000 NA NA
## range 0.11440 0.00872 13.12
##
## Random effects (SPDE parameterization):
## Estimate Std.error z-value
## alpha 1.22823 0.01055 116.4
## tau 0.06768 NA NA
## kappa 21.09912 NA NA
##
## Measurement error:
## Estimate Std.error z-value
## std. dev 0.097303 0.006312 15.42
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Log-Likelihood: -132.2996
## Number of function calls by 'optim' = 63
## Optimization method used in 'optim' = L-BFGS-B
##
## Time used to: fit the model = 14.77062 secs
## Log-likelihood with fixed sigma: -132.2996
## Log-likelihood with all parameters estimated: -126.5583
Let us now illustrate how to simulate a data set with replicates and
then fit a model to such data. To simulate a latent model with
replicates, all we do is set the nsim argument to the
number of replicates.
Now, let us generate the observed values \(Y\):
Note that \(Y\) is a matrix with 20
columns, each column containing one replicate. We need to turn
y into a vector and create an auxiliary vector
repl indexing the replicates of y:
y_vec <- as.vector(Y.rep)
repl <- rep(1:n.rep, each = n.obs)
df_data_repl <- data.frame(y = y_vec,
edge_number = rep(obs.loc[,1], n.rep),
distance_on_edge = rep(obs.loc[,2], n.rep),
repl = repl)Let us clear the previous observations and add the new data to the graph:
## Adding observations...
## Assuming the observations are normalized by the length of the edge.
We can now fit the model in the same way as before by using the
rspde_lme() function. Note that we can optimize in parallel
by setting parallel to TRUE. If we do not
specify which replicate to consider, in the which_repl
argument, all replicates will be considered.
Now, let us see a summary of the fit:
##
## Latent model - Whittle-Matern
##
## Call:
## graph_lme(formula = y ~ -1, graph = graph, model = "WM", parallel = TRUE)
##
## No fixed effects.
##
## Random effects:
## Estimate Std.error z-value
## alpha 1.289673 0.006275 205.52
## tau 0.052023 0.001459 35.65
## kappa 15.899673 0.507568 31.32
##
## Random effects (Matern parameterization):
## Estimate Std.error z-value
## nu 0.789673 0.006275 125.84
## sigma 1.314022 0.024740 53.11
## range 0.158081 0.004865 32.49
##
## Measurement error:
## Estimate Std.error z-value
## std. dev 0.3014 0.0030 100.5
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Log-Likelihood: -9837.691
## Number of function calls by 'optim' = 118
## Optimization method used in 'optim' = L-BFGS-B
##
## Time used to: fit the model = 57.37242 secs
## set up the parallelization = 2.54467 secs
Let us also take a glance of the fit:
## # A tibble: 1 × 9
## nobs sigma logLik AIC BIC deviance df.residual model alpha
## <int> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <chr> <dbl>
## 1 12000 0.301 -9838. 19683. 19713. 19675. 11996 WhittleMatern 1.29
Let us compare the values of the parameters of the latent model with the true ones:
print(data.frame(sigma = c(sigma, fit_repl$alt_par_coeff$coeff["sigma"]),
range = c(range, fit_repl$alt_par_coeff$coeff["range"]),
nu = c(nu, fit_repl$alt_par_coeff$coeff["nu"]),
row.names = c("Truth", "Estimates")))## sigma range nu
## Truth 1.300000 0.1500000 0.8000000
## Estimates 1.314022 0.1580813 0.7896728
Let us do kriging. We will use the same prediction locations as in
the previous example. Let us get prediction for replicate 10, then add
the original observations on top of them. Observe that even though the
group/replicate might be a number, it is stored as character in the data
frame, so we need to pass it as a character in the
which_repl argument.
We also have an R-INLA implementation of the rational
SPDE approach for metric graphs.
We begin by defining the model by using the
rspde.metric_graph() function. This function contains the
same arguments as the function rspde.matern(). We refer the
reader to the R-INLA
implementation of the rational SPDE approach vignette for further
details.
We begin by clearing the previous observations and adding the observations (for the case without replicates) to the graph:
## Adding observations...
## Assuming the observations are normalized by the length of the edge.
Let us create the model object:
By default, the order of the rational approximation is 1.
We can now create the auxiliary quantities that will be needed with
the graph_data_rspde() function:
The remaining is standard: we create the formula object, the stack
object, and then fit the model by using the inla()
function. So, first we create the formula object:
Now we create the inla.stack object. To such an end,
observe that data_rspde contains the dataset as the
data component, the index as the index
component and the so-called A matrix as the
basis component. We will now create the stack using these
components:
stk.dat <- inla.stack(
data = data_rspde[["data"]]["y"], A = list(data_rspde[["basis"]],1), tag = "est",
effects =
list(c(
data_rspde[["index"]],
list(Intercept = 1)), list(x1 = data_rspde[["data"]]["x1"] ,
x2 = data_rspde[["data"]]["x2"])
)
)Finally, we can fit the model:
rspde_fit <- inla(f.s, data = inla.stack.data(stk.dat),
control.inla = list(int.strategy = "eb"),
control.predictor = list(A = inla.stack.A(stk.dat), compute = TRUE),
num.threads = "1:1"
)We can use the same functions as the rspde fitted models
in inla. For instance, we can see the results in the
original scale by creating the result object:
## mean sd 0.025quant 0.5quant 0.975quant mode
## std.dev 1.382670 0.1418910 1.131230 1.372790 1.687810 1.349520
## range 0.176908 0.0407764 0.111529 0.171823 0.270939 0.161874
## nu 0.700832 0.0970332 0.518154 0.698452 0.897807 0.694333
Let us compare with the true values:
result_df <- data.frame(
parameter = c("std.dev", "range", "nu"),
true = c(sigma, range, nu),
mean = c(
result_fit$summary.std.dev$mean,
result_fit$summary.range$mean,
result_fit$summary.nu$mean
),
mode = c(
result_fit$summary.std.dev$mode,
result_fit$summary.range$mode,
result_fit$summary.nu$mode
)
)
print(result_df)## parameter true mean mode
## 1 std.dev 1.30 1.3826685 1.3495203
## 2 range 0.15 0.1769083 0.1618744
## 3 nu 0.80 0.7008316 0.6943329
We can also plot the posterior marginal densities with the help of
the gg_df() function:
posterior_df_fit <- gg_df(result_fit)
library(ggplot2)
ggplot(posterior_df_fit) + geom_line(aes(x = x, y = y)) +
facet_wrap(~parameter, scales = "free") + labs(y = "Density")R-INLA implementationWe will do kriging on the mesh locations:
Let us now add the observations for prediction:
graph$add_observations(data = data.frame(y=rep(NA,nrow(pred_loc)),
x1 = graph$mesh$VtE[,1],
x2 = graph$mesh$VtE[,2],
edge_number = pred_loc[,1],
distance_on_edge = pred_loc[,2]),
normalized = TRUE)## Adding observations...
## Assuming the observations are normalized by the length of the edge.
Let us now create a new model and, then, compute the auxiliary
components at the prediction locations. To this end, we set the argument
only_pred to TRUE, in which it will return the
data.frame containing the NA data.
rspde_model_prd <- rspde.metric_graph(graph)
data_rspde_prd <- graph_data_rspde(rspde_model_prd, only_pred = TRUE)Let us build the prediction stack using the components of
data_rspde_prd and gather it with the estimation stack.
ef.prd <-
list(c(data_rspde_prd[["index"]], list(Intercept = 1)),
list(x1 = data_rspde_prd[["data"]][["x1"]],
x2 = data_rspde_prd[["data"]][["x2"]]))
stk.prd <- inla.stack(
data = data.frame(y = data_rspde_prd[["data"]][["y"]]),
A = list(data_rspde_prd[["basis"]],1), tag = "prd",
effects = ef.prd
)
stk.all <- inla.stack(stk.dat, stk.prd)Let us obtain the predictions:
rspde_fitprd <- inla(f.s,
data = inla.stack.data(stk.all),
control.predictor = list(
A = inla.stack.A(stk.all),
compute = TRUE, link = 1
),
control.compute = list(
return.marginals = FALSE,
return.marginals.predictor = FALSE
),
control.inla = list(int.strategy = "eb"),
num.threads = "1:1"
)Let us now extract the indices of the predicted nodes and store the means:
id.prd <- inla.stack.index(stk.all, "prd")$data
m.prd <- rspde_fitprd$summary.fitted.values$mean[id.prd]Finally, let us plot the predicted values. To this end we will use
the plot_function() graph method.
R-INLA implementation to fit models with
replicatesLet us begin by cloning the graph and clearing the observations on the cloned graph:
We will now add the data with replicates to the graph:
graph_rep$add_observations(data = data.frame(y=as.vector(Y.rep),
edge_number = rep(obs.loc[,1], n.rep),
distance_on_edge = rep(obs.loc[,2], n.rep),
repl = rep(1:n.rep, each = n.obs)),
group = "repl",
normalized = TRUE)## Adding observations...
## Assuming the observations are normalized by the length of the edge.
Let us create a new rspde model object:
To fit the model with replicates we need to create the auxiliary
quantities with the graph_data_rspde() function, where we
set the repl argument in the function
graph_data_spde to .all since we want to use
all replicates:
data_rspde_rep <- graph_data_rspde(rspde_model_rep,
name = "field", repl = ".all",
repl_col = "repl")Let us now create the corresponding inla.stack
object:
st.dat.rep <- inla.stack(
data = data_rspde_rep[["data"]],
A = data_rspde_rep[["basis"]],
effects = data_rspde_rep[["index"]]
)Observe that we need the response variable y to be a
vector. We can now create the formula object, remembering
that since we gave the name argument field, when creating
the index, we need to pass field.repl to the
formula:
We can, finally, fit the model:
rspde_fit_rep <-
inla(f.rep,
data = inla.stack.data(st.dat.rep),
family = "gaussian",
control.predictor =
list(A = inla.stack.A(st.dat.rep)),
num.threads = "1:1"
)We can obtain the estimates in the original scale with the
rspde.result() function:
## mean sd 0.025quant 0.5quant 0.975quant mode
## std.dev 1.323770 0.02535830 1.274860 1.323450 1.374470 1.322720
## range 0.169494 0.00767116 0.154821 0.169369 0.184935 0.169178
## nu 0.701141 0.02445980 0.654210 0.700719 0.750242 0.699540
Let us compare with the true values of the parameters:
result_rep_df <- data.frame(
parameter = c("std.dev", "range", "nu"),
true = c(sigma, range, nu),
mean = c(
result_fit_rep$summary.std.dev$mean,
result_fit_rep$summary.range$mean,
result_fit_rep$summary.nu$mean
),
mode = c(
result_fit_rep$summary.std.dev$mode,
result_fit_rep$summary.range$mode,
result_fit_rep$summary.nu$mode
)
)
print(result_rep_df)## parameter true mean mode
## 1 std.dev 1.30 1.3237738 1.3227246
## 2 range 0.15 0.1694942 0.1691785
## 3 nu 0.80 0.7011413 0.6995398
We can also plot the posterior marginal densities with the help of
the gg_df() function:
posterior_df_fit_rep <- gg_df(result_fit_rep)
ggplot(posterior_df_fit_rep) + geom_line(aes(x = x, y = y)) +
facet_wrap(~parameter, scales = "free") + labs(y = "Density")inlabru implementationThe inlabru package allows us to fit models and do
kriging in a straighforward manner, without having to handle
A matrices, indices nor inla.stack objects.
Therefore, we suggest the reader to use this implementation when using
our implementation to fit real data.
Let us clear the graph, since it contains NA
observations we used for prediction, add the observations again, and
create a new rSPDE model object:
## Adding observations...
## Assuming the observations are normalized by the length of the edge.
Let us now load the inlabru package and create the
component (which is inlabru’s formula-like object). Let us
begin by building the auxiliary data to be used with the
graph_data_rspde() function, where we pass the name of the
location variable in the above formula as the loc_name
argument, which in this case is "loc":
Now, we create the component to be used in inlabru, in
which we pass the index element from the
data_rspde_bru object as index locations:
library(inlabru)
cmp <-
y ~ -1 + Intercept(1) + x1 + x2 + field(
cbind(.edge_number, .distance_on_edge),
model = rspde_model
) Now, we can directly fit the model, by using the data
element of data_rspde_bru:
Let us now obtain the estimates of the parameters in the original
scale by using the rspde.result() function:
## mean sd 0.025quant 0.5quant 0.975quant mode
## std.dev 1.394100 0.1607360 1.111050 1.382380 1.741500 1.355700
## range 0.187895 0.0585566 0.101581 0.178094 0.329511 0.160171
## nu 0.690710 0.1121820 0.478779 0.688632 0.916853 0.687062
Let us compare with the true values of the parameters:
result_bru_df <- data.frame(
parameter = c("std.dev", "range", "nu"),
true = c(sigma, range, nu),
mean = c(
result_bru_fit$summary.std.dev$mean,
result_bru_fit$summary.range$mean,
result_bru_fit$summary.nu$mean
),
mode = c(
result_bru_fit$summary.std.dev$mode,
result_bru_fit$summary.range$mode,
result_bru_fit$summary.nu$mode
)
)
print(result_bru_df)## parameter true mean mode
## 1 std.dev 1.30 1.3941002 1.3556990
## 2 range 0.15 0.1878952 0.1601711
## 3 nu 0.80 0.6907102 0.6870621
We can also plot the posterior marginal densities with the help of
the gg_df() function:
posterior_df_bru_fit <- gg_df(result_bru_fit)
ggplot(posterior_df_bru_fit) + geom_line(aes(x = x, y = y)) +
facet_wrap(~parameter, scales = "free") + labs(y = "Density")inlabru implementationIt is very easy to do kriging with the inlabru
implementation. We simply need to provide the prediction locations to
the predict() method.
In this example we will use the mesh locations. To this end we will
use the get_mesh_locations() method. We also set
bru=TRUE to obtain a data frame suitable to be used with
inlabru. In this case, the mesh locations will be returned
as a data.frame with the location columns
.edge_number and .distance_on_edge. We will,
then, add the covariates x1 and x2 to the data
frame:
prd_loc <- graph$get_mesh_locations(bru = TRUE)
prd_loc[["x1"]] <- prd_loc[,1]
prd_loc[["x2"]] <- prd_loc[,2] Now, we can simply provide these locations to the
predict method along with the fitted object
rspde_bru_fit:
Let us now prepare the predictions so we can plot them easily by
using the process_rspde_predictions() function:
Finally, let us plot the predicted values. To this end we will use
the plot() method on y_pred:
We can also create the 3d plot, together with the true data:
We can also use our inlabru implementation to fit models
with replicates. We will consider the same data that was generated
above, where the number of replicates is 30.
For this implementation we will use the rspde_model_rep
object.
We can now create the component, passing the vector with the indices
of the replicates as the replicate argument. To obtain the
auxiliary data, we will pass repl argument we use the
function graph_data_rspde(), where we set it to
.all, since we want all replicates. Further, we also set
the argument bru to TRUE.
We can now define the bru component formula, passing the
repl as the replicate argument:
cmp_rep <-
y ~ -1 + field(cbind(.edge_number, .distance_on_edge),
model = rspde_model_rep,
replicate = repl)Now, we are ready to fit the model:
rspde_bru_fit_rep <-
bru(cmp_rep,
data=data_rspde_rep[["data"]],
options=list(
family = "gaussian",
num.threads = "1:1")
)We can obtain the estimates in the original scale with the
rspde.result() function:
result_bru_fit_rep <- rspde.result(rspde_bru_fit_rep, "field", rspde_model_rep)
summary(result_bru_fit_rep)## mean sd 0.025quant 0.5quant 0.975quant mode
## std.dev 1.323770 0.02535830 1.274860 1.323450 1.374470 1.322720
## range 0.169494 0.00767116 0.154821 0.169369 0.184935 0.169178
## nu 0.701141 0.02445980 0.654210 0.700719 0.750242 0.699540
Let us compare with the true values of the parameters:
result_bru_rep_df <- data.frame(
parameter = c("std.dev", "range", "nu"),
true = c(sigma, range, nu),
mean = c(
result_bru_fit_rep$summary.std.dev$mean,
result_bru_fit_rep$summary.range$mean,
result_bru_fit_rep$summary.nu$mean
),
mode = c(
result_bru_fit_rep$summary.std.dev$mode,
result_bru_fit_rep$summary.range$mode,
result_bru_fit_rep$summary.nu$mode
)
)
print(result_bru_rep_df)## parameter true mean mode
## 1 std.dev 1.30 1.3237738 1.3227246
## 2 range 0.15 0.1694942 0.1691785
## 3 nu 0.80 0.7011413 0.6995398
We can also plot the posterior marginal densities with the help of
the gg_df() function:
posterior_df_bru_fit_rep <- gg_df(result_bru_fit_rep)
ggplot(posterior_df_bru_fit_rep) + geom_line(aes(x = x, y = y)) +
facet_wrap(~parameter, scales = "free") + labs(y = "Density")Let us now do prediction for observations of replicate
10. We start by building the data list with the prediction
locations:
data_prd_repl <- graph$get_mesh_locations(bru = TRUE)
data_prd_repl[["repl"]] <- rep(10, nrow(data_prd_repl))Let us now obtain predictions for this replicate:
y_pred <- predict(rspde_bru_fit_rep,
newdata=data_prd_repl,
~field_eval(cbind(.edge_number, .distance_on_edge),
replicate = repl))Let us now process the predictions:
We can now plot the predictions along with the observed values for
replicate 10:
Our goal now is to show how one can fit model with non-stationary \(\sigma\) (std. deviation) and non-stationary \(\rho\) (a range parameter). One can also use the parameterization in terms of non-stationary SPDE parameters \(\kappa\) and \(\tau\).
We follow the same structure as INLA. However,
INLA only allows one to specify B.tau and
B.kappa matrices, and, in INLA, if one wants
to parameterize in terms of range and standard deviation one needs to do
it manually. Here we provide the option to directly provide the matrices
B.sigma and B.range.
The usage of the matrices B.tau and B.kappa
are identical to the corresponding ones in
inla.spde2.matern() function. The matrices
B.sigma and B.range work in the same way, but
they parameterize the stardard deviation and range, respectively.
The columns of the B matrices correspond to the same
parameter. The first column does not have any parameter to be estimated,
it is a constant column.
So, for instance, if one wants to share a parameter with both
sigma and range (or with both tau
and kappa), one simply let the corresponding column to be
nonzero on both B.sigma and B.range (or on
B.tau and B.kappa).
For this example we will consider the pems data
contained in the MetricGraph package. The data consists of
traffic speed observations on highways in the city of San Jose,
California. The variable y contains the traffic speeds.
pems_graph <- metric_graph$new(edges = pems$edges)
pems_graph$add_observations(data = pems$data)
pems_graph$prune_vertices( )
pems_graph$build_mesh(h=0.1)The summary of this graph:
## A metric graph object with:
##
## Vertices:
## Total: 347
## Degree 1: 11; Degree 2: 16; Degree 3: 315; Degree 4: 5;
## With incompatible directions: 17
##
## Edges:
## Total: 504
## Lengths:
## Min: 0.01040037 ; Max: 7.663457 ; Total: 470.6126
## Weights:
## Columns: .weights
## That are circles: 0
##
## Graph units:
## Vertices unit: degree ; Lengths unit: km
##
## Longitude and Latitude coordinates: TRUE
## Which spatial package: sp
## CRS: +proj=longlat +datum=WGS84 +no_defs
##
## Some characteristics of the graph:
## Connected: TRUE
## Has loops: FALSE
## Has multiple edges: TRUE
## Is a tree: FALSE
## Distance consistent: FALSE
## Has Euclidean edges: FALSE
##
## Computed quantities inside the graph:
## Laplacian: FALSE ; Geodesic distances: TRUE
## Resistance distances: FALSE ; Finite element matrices: FALSE
##
## Mesh:
## Max h_e: 0.0999534 ; Min n_e: 0
##
## Data:
## Columns: y
## Groups: .group
##
## Tolerances:
## vertex-vertex: 0.001
## vertex-edge: 0.001
## edge-edge: 0
Observe that it is a non-Euclidean graph.
We now define a non-stationary covariate based on the position along each edge. This covariate is designed to indicate whether the traffic speed observation was taken close to an intersection. More precisely, we identify points whose relative position on the edge is either below 0.1 or above 0.9, corresponding to locations near the endpoints of the edge. Thus, the covariate is equal to TRUE for points close to intersections and FALSE otherwise.
We will now build the non-stationary matrices to be used:
Let us also obtain the same covariate for the observations:
Let add this covariate to the data:
## Adding observations...
## Assuming the observations are NOT normalized by the length of the edge.
## The unit for edge lengths is km
## The current tolerance for removing distant observations is (in km): 3.83172861400444
graph_lmeWe are now in position to fit this model using the
graph_lme() function. We will also add cov_obs
as a covariate for the model.
fit <- graph_lme(y ~ 1, graph = pems_graph, model = list(type = "WhittleMatern",
B.sigma = B.sigma, B.range = B.range, fem = TRUE))Let us now obtain a summary of the fitted model:
##
## Latent model - Generalized Whittle-Matern
##
## Call:
## graph_lme(formula = y ~ 1, graph = pems_graph, model = list(type = "WhittleMatern",
## B.sigma = B.sigma, B.range = B.range, fem = TRUE))
##
## Fixed effects:
## Estimate Std.error z-value Pr(>|z|)
## (Intercept) 51.526 2.492 20.68 <2e-16 ***
##
## Random effects:
## Estimate Std.error z-value
## nu 2.27550 0.08519 26.710
## theta1 2.87528 0.23846 12.058
## theta2 1.85945 0.17770 10.464
## theta3 0.39400 1.28364 0.307
## theta4 0.12797 0.57730 0.222
##
## Measurement error:
## Estimate Std.error z-value
## std. dev 6.9554 0.3518 19.77
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Log-Likelihood: -1206.891
## Number of function calls by 'optim' = 501
## Optimization method used in 'optim' = Nelder-Mead
##
## Time used to: fit the model = 31.87703 secs
Let us plot the range parameter along the mesh, so we can see how it is varying:
est_range <- exp(B.range[,-1]%*%fit$coeff$random_effects[2:5])
df_range <- data.frame(range = est_range, edge_number = pems_graph$mesh$VtE[,1],
distance_on_edge = pems_graph$mesh$VtE[,2])
df_range <- pems_graph$process_data(data = df_range, normalized = TRUE)
pems_graph$plot_function(data = "range", newdata = df_range, vertex_size = 0,
type = "mapview", mapview_caption = "Range")Similarly, we have for sigma:
est_sigma <- exp(B.sigma[,-1]%*%fit$coeff$random_effects[2:5])
df_sigma <- data.frame(sigma = est_sigma, edge_number = pems_graph$mesh$VtE[,1],
distance_on_edge = pems_graph$mesh$VtE[,2])
df_sigma <- pems_graph$process_data(data = df_sigma, normalized = TRUE)
pems_graph$plot_function(data = "sigma", newdata = df_sigma, vertex_size = 0,
type = "mapview", mapview_caption = "Sigma")Our goal now is to plot the estimated marginal standard deviation of
this model. To this end, we start by creating the non-stationary Matérn
operator using the rSPDE package:
rspde_object_ns <- rSPDE::spde.matern.operators(graph = pems_graph,
parameterization = "matern",
B.sigma = B.sigma,
B.range = B.range,
theta = fit$coeff$random_effects[2:5],
nu = fit$coeff$random_effects[1])Now, we compute the estimated marginal standard deviation:
est_cov_matrix <- covariance_mesh(rspde_object_ns)
est_std_dev <- sqrt(Matrix::diag(est_cov_matrix))We can now plot:
df_std <- data.frame(std = est_std_dev, edge_number = pems_graph$mesh$VtE[,1],
distance_on_edge = pems_graph$mesh$VtE[,2])
df_std <- pems_graph$process_data(data = df_std, normalized = TRUE)
pems_graph$plot_function(data = "std", newdata = df_std, vertex_size = 0,
type = "mapview", mapview_caption = "Std. dev")In non-stationary models, the parameters are labeled as theta1,
theta2, etc., corresponding to the coefficients in the B matrices.
Similar to the stationary case, we can fix individual parameters or set
starting values, but these options must be set in the
model_options list argument using fix_theta1,
fix_theta2, etc., or start_theta1,
start_theta2, etc.
For example, if we want to fix the first coefficient theta1 to 0 in the previous example:
# Fit model with fixed theta1 parameter
fit_ns_fixed_theta1 <- graph_lme(y ~ 1,
graph = pems_graph,
model = list(type = "WhittleMatern",
B.sigma = B.sigma,
B.range = B.range,
fem = TRUE),
model_options = list(fix_theta1 = 0.5)) # Fix theta1 to 0.5## Warning in rSPDE::rspde_lme(formula = formula, loc =
## cbind(df_data[[".edge_number"]], : optim method L-BFGS-B failed to provide a
## positive-definite Hessian. Another optimization method was used.
##
## Latent model - Generalized Whittle-Matern
##
## Call:
## graph_lme(formula = y ~ 1, graph = pems_graph, model = list(type = "WhittleMatern",
## B.sigma = B.sigma, B.range = B.range, fem = TRUE), model_options = list(fix_theta1 = 0.5))
##
## Fixed effects:
## Estimate Std.error z-value Pr(>|z|)
## (Intercept) 50.8495 0.8783 57.9 <2e-16 ***
##
## Random effects:
## Estimate Std.error z-value
## nu 0.266500 0.004959 53.743
## theta1 (fixed) 0.500000 NA NA
## theta2 2.087741 0.511435 4.082
## theta3 2.481165 0.378907 6.548
## theta4 0.924256 0.885356 1.044
##
## Measurement error:
## Estimate Std.error z-value
## std. dev 13.7723 0.5875 23.44
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Log-Likelihood: -1328.889
## Number of function calls by 'optim' = 443
## Optimization method used in 'optim' = Nelder-Mead
##
## Time used to: fit the model = 17.18008 secs
Similarly, we can provide starting values for the entire theta vector
with start_theta:
# Fit model with starting values for theta parameters
fit_ns_start <- graph_lme(y ~ 1,
graph = pems_graph,
model = list(type = "WhittleMatern",
B.sigma = B.sigma,
B.range = B.range,
fem = TRUE),
model_options = list(start_theta = c(0.4, 0.7, 1.0, 0.2))) # Starting values for theta vector## Warning in rSPDE::rspde_lme(formula = formula, loc =
## cbind(df_data[[".edge_number"]], : optim method L-BFGS-B failed to provide a
## positive-definite Hessian. Another optimization method was used.
##
## Latent model - Generalized Whittle-Matern
##
## Call:
## graph_lme(formula = y ~ 1, graph = pems_graph, model = list(type = "WhittleMatern",
## B.sigma = B.sigma, B.range = B.range, fem = TRUE), model_options = list(start_theta = c(0.4,
## 0.7, 1, 0.2)))
##
## Fixed effects:
## Estimate Std.error z-value Pr(>|z|)
## (Intercept) 51.190 2.789 18.36 <2e-16 ***
##
## Random effects:
## Estimate Std.error z-value
## nu 1.4767 0.6955 2.123
## theta1 3.0923 0.4429 6.982
## theta2 2.2318 0.5233 4.265
## theta3 -0.4515 1.0539 -0.428
## theta4 -0.3982 0.8348 -0.477
##
## Measurement error:
## Estimate Std.error z-value
## std. dev 7.1685 0.3959 18.11
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Log-Likelihood: -1207.766
## Number of function calls by 'optim' = 502
## Optimization method used in 'optim' = Nelder-Mead
##
## Time used to: fit the model = 24.24923 secs
Let us then fit the same model using inlabru now. We
start by defing the rSPDE model with the
rspde.metric_graph() function. Since the smoothness
estimated by graph_lme() is above 2, we increase the upper
bound for \(\nu\) with the
nu.upper.bound argument. We also use a slightly more
informative prior for the coefficients of the non-stationary parameters
(theta.prior.prec = 1) to stabilize the estimation:
rspde_model_nonstat <- rspde.metric_graph(pems_graph,
B.sigma = B.sigma,
B.range = B.range,
parameterization = "matern",
nu.upper.bound = 4,
theta.prior.prec = 1)## Warning in rspde.matern(mesh = mesh, nu.upper.bound = nu.upper.bound, rspde.order = rspde.order, : In dimension 1 you can have unstable results
## for nu.upper.bound > 2. Consider changing
## nu.upper.bound to 2 or 1.
Let us now create the data.frame() and the vector with
the replicates indexes:
Let us create the component and fit.
cmp_nonstat <-
y ~ -1 + Intercept(1) + field(
cbind(.edge_number, .distance_on_edge),
model = rspde_model_nonstat
)
rspde_fit_nonstat <-
bru(cmp_nonstat,
data = data_rspde_bru_ns[["data"]],
family = "gaussian",
options = list(num.threads = "1:1")
)We can get the summary:
## inlabru version: 2.15.0
## INLA version: 26.08.22
## Latent components:
## Intercept: main = linear(1)
## field: main = cgeneric(cbind(.edge_number, .distance_on_edge))
## Observation models:
## Model tag: <No tag>
## Family: 'gaussian'
## Data class: 'metric_graph_data', 'data.frame'
## Response class: 'numeric'
## Predictor: y ~ Intercept + field
## Additive/Linear/Rowwise: TRUE/TRUE/TRUE
## Used components: effect[Intercept, field], latent[]
## Time used:
## Pre = 0.153, Running = 23.1, Post = 0.229, Total = 23.5
## Fixed effects:
## mean sd 0.025quant 0.5quant 0.975quant mode kld
## Intercept 50.996 1.503 48.05 50.996 53.942 50.996 0
##
## Random effects:
## Name Model
## field CGeneric
##
## Model hyperparameters:
## mean sd 0.025quant 0.5quant
## Precision for the Gaussian observations 8.22e+19 1.02e+22 0.087 11720.075
## Theta1 for field 4.08e+00 2.48e-01 3.676 4.050
## Theta2 for field -3.09e+00 2.73e-01 -3.645 -3.080
## Theta3 for field 1.58e+00 2.30e+00 -1.941 1.301
## Theta4 for field -5.16e+00 6.57e+00 -20.186 -4.428
## Theta5 for field -5.81e-01 7.43e-01 -1.727 -0.671
## 0.975quant mode
## Precision for the Gaussian observations 7.80e+13 0.002
## Theta1 for field 4.63e+00 3.924
## Theta2 for field -2.57e+00 -3.050
## Theta3 for field 6.88e+00 -0.144
## Theta4 for field 5.03e+00 -0.560
## Theta5 for field 1.12e+00 -1.120
##
## Marginal log-Likelihood: -1537.63
## is computed
## Posterior summaries for the linear predictor and the fitted values are computed
## (Posterior marginals needs also 'control.compute=list(return.marginals.predictor=TRUE)')
We can obtain outputs with respect to parameters in the original
scale by using the function rspde.result():
result_fit_nonstat <- rspde.result(rspde_fit_nonstat, "field", rspde_model_nonstat)
summary(result_fit_nonstat)## mean sd 0.025quant 0.5quant 0.975quant mode
## Theta1.matern 4.07466 0.248344 3.67568 4.04991 4.63394 3.924410
## Theta2.matern -3.08687 0.273411 -3.64548 -3.08009 -2.56938 -3.049660
## Theta3.matern 1.57764 2.303590 -1.94074 1.30138 6.87614 -0.144056
## Theta4.matern -5.15924 6.565460 -20.18620 -4.42820 5.02893 -0.559532
## nu 1.47728 0.646191 0.60818 1.33460 2.99961 0.912100
We can also plot the posterior densities. To this end we will use the
gg_df() function, which creates ggplot2
user-friendly data frames:
posterior_df_fit <- gg_df(result_fit_nonstat)
ggplot(posterior_df_fit) + geom_line(aes(x = x, y = y)) +
facet_wrap(~parameter, scales = "free") + labs(y = "Density")graph_lmeR-INLA implementation to fit models with
replicatesinlabru
implementation