Description Usage Arguments Details Value Author(s) References See Also Examples
Create an inla.spde2
model object for a Matern model,
using a PC prior for the parameters.
1 2 3 4 5 6 7 8 9 10 |
mesh |
The mesh to build the model on, as an |
alpha |
Fractional operator order, 0 < alpha <= 2 supported, for ν=α-d/2>0. |
param |
Further model parameters. Not currently used. |
constr |
If |
extraconstr.int |
Field integral constraints. |
extraconstr |
Direct linear combination constraints on the basis weights. |
fractional.method |
Specifies the approximation method to use for
fractional (non-integer) |
n.iid.group |
If greater than 1, build an explicitly iid
replicated model, to support constraints applied to the combined
replicates, for example in a time-replicated spatial
model. Constraints can either be specified for a single mesh, in
which case it's applied to the average of the replicates
( |
prior.range |
A length 2 vector, with |
prior.sigma |
A length 2 vector, with |
This method constructs a Matern SPDE model, with spatial range ρ and standard deviation parameter σ. In the parameterisation
(kappa^2-Delta)^(alpha/2) (tau x(u)) = W(u)
the spatial scale parameter κ=√{8ν}/ρ, where ν=α-d/2, and τ is proportional to 1/σ.
Stationary models are supported for
0 < alpha <= 2, with spectral approximation
methods used for
non-integer α, with approximation method determined by
fractional.method
.
Integration and other general linear constraints are supported via the
constr
, extraconstr.int
, and extraconstr
parameters, which also interact with n.iid.group
.
The joint PC prior density for the spatial range, ρ, and the marginal standard deviation, σ, and is
p(rho, sigma) = (d R)/2 rho^(-1-d/2) exp(-R rho^(-d/2)) S exp(-S sigma)
where R and S are hyperparameters that must be determined by the analyst. The practical approach for this in INLA is to require the user to indirectly specify these hyperparameters through
P(ρ < ρ_0) = p_ρ
and
P(σ > σ_0) = p_σ
where the user specifies the lower tail quantile and probability for the range (ρ_0 and p_ρ) and the upper tail quantile and probability for the standard deviation (σ_0 and α_σ).
This allows the user to control the priors of the parameters by supplying knowledge of the scale of the problem. What is a reasonable upper magnitude for the spatial effect and what is a reasonable lower scale at which the spatial effect can operate? The shape of the prior was derived through a construction that shrinks the spatial effect towards a base model of no spatial effect in the sense of distance measured by Kullback-Leibler divergence.
The prior is constructed in two steps, under the idea that having a spatial field is an extension of not having a spatial field. First, a spatially constant random effect (ρ = ∞) with finite variance is more complex than not having a random effect (σ = 0). Second, a spatial field with spatial variation (ρ < ∞) is more complex than the random effect with no spatial variation. Each of these extensions are shrunk towards the simpler model and, as a result, we shrink the spatial field towards the base model of no spatial variation and zero variance (ρ = ∞ and σ = 0).
The details behind the construction of the prior is presented in Fuglstad, et al. (2016) and is based on the PC prior framework (Simpson, et al., 2015).
An inla.spde2
object.
Finn Lindgren finn.lindgren@gmail.com
Fuglstad, G.-A., Simpson, D., Lindgren, F., and Rue, H. (2016) Constructing Priors that Penalize the Complexity of Gaussian Random Fields. arXiv:1503.00256
Simpson, D., Rue, H., Martins, T., Riebler, A., and Sørbye, S. (2015) Penalising model component complexity: A principled, practical approach to constructing priors. arXiv:1403.4630
inla.mesh.2d
, inla.mesh.create
,
inla.mesh.1d
, inla.mesh.basis
,
inla.spde2.matern
, inla.spde2.generic
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 | ## Spatial interpolation
n = 100
field.fcn = function(loc) (10*cos(2*pi*2*(loc[,1]+loc[,2])))
loc = matrix(runif(n*2),n,2)
## One field, 2 observations per location
idx.y = rep(1:n,2)
y = field.fcn(loc[idx.y,]) + rnorm(length(idx.y))
mesh = inla.mesh.2d(loc, max.edge=0.05, cutoff=0.01)
spde = inla.spde2.pcmatern(mesh,
prior.range=c(0.01,0.1), prior.sigma=c(100,0.1))
data = list(y=y, field=mesh$idx$loc[idx.y])
formula = y ~ -1 + f(field, model=spde)
result = inla(formula, data=data, family="normal")
## Plot the mesh structure:
plot(mesh)
if (require(rgl)) {
col.pal = colorRampPalette(c("blue","cyan","green","yellow","red"))
## Plot the posterior mean:
plot(mesh, rgl=TRUE,
result$summary.random$field[,"mean"],
color.palette = col.pal)
## Plot residual field:
plot(mesh, rgl=TRUE,
result$summary.random$field[,"mean"]-field.fcn(mesh$loc),
color.palette = col.pal)
}
result.field = inla.spde.result(result, "field", spde)
par(mfrow=c(2,1))
plot(result.field$marginals.range.nominal[[1]],
type="l", main="Posterior density for range")
plot(inla.tmarginal(sqrt, result.field$marginals.variance.nominal[[1]]),
type="l", main="Posterior density for std.dev.")
par(mfrow=c(1,1))
## Spatial model
set.seed(1234234)
## Generate spatial locations
nObs = 200
loc = matrix(runif(nObs*2), nrow = nObs, ncol = 2)
## Generate observation of spatial field
nu = 1.0
rhoT = 0.2
kappaT = sqrt(8*nu)/rhoT
sigT = 1.0
Sig = sigT^2*inla.matern.cov(nu = nu,
kappa = kappaT,
x = as.matrix(dist(loc)),
d = 2,
corr = TRUE)
L = t(chol(Sig))
u = L %*% rnorm(nObs)
## Construct observation with nugget
sigN = 0.1
y = u + sigN*rnorm(nObs)
## Create the mesh and spde object
mesh = inla.mesh.2d(loc,
max.edge = 0.05,
cutoff = 0.01)
spde = inla.spde2.pcmatern(mesh,
prior.range = c(0.01, 0.05),
prior.sigma = c(10, 0.05))
## Create projection matrix for observations
A = inla.spde.make.A(mesh = mesh,
loc = loc)
## Run model without any covariates
idx = 1:spde$n.spde
res = inla(y ~ f(idx, model = spde) - 1,
data = list(y = y, idx = idx, spde = spde),
control.predictor = list(A = A))
## Re-run model with fixed range
spde.fixed = inla.spde2.pcmatern(mesh,
prior.range = c(0.2, NA),
prior.sigma = c(10, 0.05))
res.fixed = inla(y ~ f(idx, model = spde) - 1,
data = list(y = y, idx = idx, spde = spde.fixed),
control.predictor = list(A = A))
|
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.