spConjNNGP {spNNGP}

Function for Fitting Univariate Bayesian Conjugate Spatial Regression Models

Description

The function spConjNNGP fits Gaussian univariate Bayesian conjugate spatial regression models using Nearest Neighbor Gaussian Processes (NNGP).

Usage

    spConjNNGP(formula, data = parent.frame(), coords, knots, n.neighbors = 15,
               theta.alpha, sigma.sq.IG, cov.model = "exponential",
               k.fold = 5, score.rule = "crps",
               X.0, coords.0, n.omp.threads = 1, search.type = "cb",
               ord, return.neighbor.info = TRUE, 
               neighbor.info, fit.rep = FALSE, n.samples, verbose = TRUE, ...)

Arguments

formula

a symbolic description of the regression model to be fit. See example below.

data

an optional data frame containing the variables in the model. If not found in data, the variables are taken from environment(formula), typically the environment from which spConjNNGP is called.

coords

an n \times 2 matrix of the observation coordinates in R^2 (e.g., easting and northing), or if data is a data frame then coords can be a vector of length two comprising coordinate column names or indices. There can be no duplicate locations.

knots

an r \times 2 matrix of the observation coordinates in R^2 (e.g., easting and northing). Adding the knots argument invokes SLGP, see Shin et al. (2019) below.

n.neighbors

number of neighbors used in the NNGP.

theta.alpha

a vector or matrix of parameter values for phi, nu, and alpha, where \alpha=\tau^2/\sigma^2 and nu is only required if cov.model="matern". A vector is passed to run the model using one set of parameters. The vector elements must be named and hold values for phi, nu, and alpha. If a matrix is passed, columns must be named and hold values for phi, nu, and alpha. Each row in the matrix defines a set of parameters for which the model will be run.

sigma.sq.IG

a vector of length two that holds the hyperparameters, shape and scale respectively, for the inverse-Gamma prior on \sigma^2.

cov.model

a quoted keyword that specifies the covariance function used to model the spatial dependence structure among the observations. Supported covariance model key words are: "exponential", "matern", "spherical", and "gaussian". See below for details.

k.fold

specifies the number of k folds for cross-validation. If theta.alpha is a vector then cross-validation is not performed and k-fold and score.rule are ignored. In k-fold cross-validation, the data specified in model is randomly partitioned into k equal sized subsamples. Of the k subsamples, k-1 subsamples are used to fit the model and the remaining k samples are used for prediction. The cross-validation process is repeated k times (the folds). Root mean squared prediction error (RMSPE) and continuous ranked probability score (CRPS; Gneiting and Raftery, 2007) rules are averaged over the k fold prediction results and reported for the parameter sets defined by theta.alpha. The parameter set that yields the best performance based on the scoring rule defined by score.rule is used to fit the final model that uses all the data and make predictions if X.0 and coords.0 are supplied. Results from the k-fold cross-validation are returned in the k.fold.scores matrix.

score.rule

a quoted keyword "rmspe" or "crps" that specifies the scoring rule used to select the best parameter set, see argument definition for k.fold for more details.

X.0

the design matrix for prediction locations. An intercept should be provided in the first column if one is specified in model.

coords.0

the spatial coordinates corresponding to X.0.

n.omp.threads

a positive integer indicating the number of threads to use for SMP parallel processing. The package must be compiled for OpenMP support. For most Intel-based machines, we recommend setting n.omp.threads up to the number of hyperthreaded cores. Note, n.omp.threads > 1 might not work on some systems.

search.type

a quoted keyword that specifies type of nearest neighbor search algorithm. Supported method key words are: "cb" and "brute". The "cb" should generally be much faster. If locations do not have identical coordinate values on the axis used for the nearest neighbor ordering (see ord argument) then "cb" and "brute" should produce identical neighbor sets. However, if there are identical coordinate values on the axis used for nearest neighbor ordering, then "cb" and "brute" might produce different, but equally valid, neighbor sets, e.g., if data are on a grid.

ord

an index vector of length n used for the nearest neighbor search. Internally, this vector is used to order coords, i.e., coords[ord,], and associated data. Nearest neighbor candidates for the i-th row in the ordered coords are rows 1:(i-1), with the n.neighbors nearest neighbors being those with the minimum euclidean distance to the location defined by ordered coords[i,]. The default used when ord is not specified is x-axis ordering, i.e., order(coords[,1]). This argument should typically be left blank. This argument will be ignored if the neighbor.info argument is used.

return.neighbor.info

if TRUE, a list called neighbor.info containing several data structures used for fitting the NNGP model is returned. If there is no change in input data or n.neighbors, this list can be passed to subsequent spNNGP calls via the neighbor.info argument to avoid the neighbor search, which can be time consuming if n is large. In addition to the several cryptic data structures in neighbor.info there is a list called n.indx that is of length n. The i-th element in n.indx corresponds to the i-th row in coords[ord,] and holds the vector of that location's nearest neighbor indices. This list can be useful for plotting the neighbor graph if desired.

neighbor.info

see the return.neighbor.info argument description above.

fit.rep

if TRUE, regression fitted and replicate data will be returned. The argument n.samples controls the number of fitted and replicated data samples.

n.samples

gives the number of posterior samples returned. Note, point and associated variance estimates for model parameters are not based on posterior samples. Only specify n.samples if you wish to generate samples from parameters' posteriors (this is an exact sampling algorithm). If fit.rep is TRUE, then n.samples also controls the number of fitted and replicated data samples.

verbose

if TRUE, model specification and progress is printed to the screen. Otherwise, nothing is printed to the screen.

...

currently no additional arguments.

Value

An object of class NNGP and conjugate, and, if knots is provided, SLGP. Among other elements, the object contains:

theta.alpha

the input theta.alpha vector, or best (according to the selected scoring rule) set of parameters in the theta.alpha matrix. All subsequent parameter estimates are based on this parameter set.

beta.hat

a matrix of regression coefficient estimates corresponding to the returned theta.alpha.

beta.var

beta.hat variance-covariance matrix.

sigma.sq.hat

estimate of \sigma^2 corresponding to the returned theta.alpha.

sigma.sq.var

sigma.sq.hat variance.

k.fold.scores

results from the k-fold cross-validation if theta.alpha is a matrix.

y.0.hat

prediction if X.0 and coords.0 are specified.

y.0.var.hat

y.0.hat variance.

n.neighbors

number of neighbors used in the NNGP.

neighbor.info

returned if return.neighbor.info=TRUE see the return.neighbor.info argument description above.

run.time

execution time for parameter estimation reported using proc.time(). This time does not include nearest neighbor search time for building the neighbor set.

Author(s)

Andrew O. Finley [e-mail],
Abhirup Datta [e-mail],
Sudipto Banerjee [e-mail]

References

Datta, A., S. Banerjee, A.O. Finley, and A.E. Gelfand. (2016) Hierarchical Nearest-Neighbor Gaussian process models for large geostatistical datasets. Journal of the American Statistical Association, doi: 10.1080/01621459.2015.1044091.

Finley, A.O., A. Datta, S. Banerjee (2022) spNNGP R Package for Nearest Neighbor Gaussian Process Models. Journal of Statistical Software, doi: 10.18637/jss.v103.i05.

Finley, A.O., A. Datta, B.D. Cook, D.C. Morton, H.E. Andersen, and S. Banerjee. (2019) Efficient algorithms for Bayesian Nearest Neighbor Gaussian Processes. Journal of Computational and Graphical Statistics, doi: 10.1080/10618600.2018.1537924.

Gneiting, T and A.E. Raftery. (2007) Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, doi: 10.1198/016214506000001437.

Shirota, S., A.O. Finley, B.D. Cook, and S. Banerjee (2019) Conjugate Nearest Neighbor Gaussian Process models for efficient statistical interpolation of large spatial data. https://arxiv.org/abs/1907.10109.

Examples

rmvn <- function(n, mu=0, V = matrix(1)){
  p <- length(mu)
  if(any(is.na(match(dim(V),p))))
    stop("Dimension problem!")
  D <- chol(V)
  t(matrix(rnorm(n*p), ncol=p)%*%D + rep(mu,rep(n,p)))
}

##Make some data
set.seed(1)
n <- 2000
coords <- cbind(runif(n,0,1), runif(n,0,1))

x <- cbind(1, rnorm(n))

B <- as.matrix(c(1,5))

sigma.sq <- 5
tau.sq <- 1
phi <- 3/0.5

D <- as.matrix(dist(coords))
R <- exp(-phi*D)
w <- rmvn(1, rep(0,n), sigma.sq*R)
y <- rnorm(n, x%*%B + w, sqrt(tau.sq))

ho <- sample(1:n, 1000)

y.ho <- y[ho]
x.ho <- x[ho,,drop=FALSE]
w.ho <- w[ho]
coords.ho <- coords[ho,]

y <- y[-ho]
x <- x[-ho,,drop=FALSE]
w <- w[-ho,,drop=FALSE]
coords <- coords[-ho,]

##Fit a Conjugate NNGP model and predict for the holdout
sigma.sq.IG <- c(2, sigma.sq)

cov.model <- "exponential"

g <- 10
theta.alpha <- cbind(seq(phi,30,length.out=g), seq(tau.sq/sigma.sq,5,length.out=g))

colnames(theta.alpha) <- c("phi", "alpha")

m.c <- spConjNNGP(y~x-1, coords=coords, n.neighbors = 10,
                  X.0 = x.ho, coords.0 = coords.ho,
                  k.fold = 5, score.rule = "crps",
                  n.omp.threads = 1,
                  theta.alpha = theta.alpha, sigma.sq.IG = sigma.sq.IG, cov.model = cov.model)

m.c