CARrampsOcl.fit: Fit Bayesian normal conditional autoregressive model

In CARrampsOcl: Reparameterized and marginalized posterior sampling for conditional autoregressive models, OpenCL implementation

Description

This function fits CAR models to data. It draws independent samples from the posterior distribution of precision parameters and regression coefficients for certain polynomial trend surfaces. If the randeffs argument is set to true, it will also produce the estimated mean and standard deviation of the marginal posterior density of each random effect.

Usage

CARrampsOcl.fit(alpha, beta, Q, y, nsamp, seed, fixed = FALSE, coefs = FALSE, 
randeffs = FALSE, designMat = NULL, mult = 20, filename = "params.txt")

Arguments

alpha	Vector of alpha parameters for gamma prior densities on precisions. The first element is for the prior density on the measurement error precision, and the remaining entries are for the the prior density(ies) on the random effects precision(s), in the same order as the Q matrices in the list argument Q.
beta	Vector of beta parameters for gamma prior densities on precisions. The first element is for the prior density on the measurement error precision, and the remaining entries are for the the prior density(ies) on the random effects precisions, in the same order as the Q matrices in the list argument Q.
Q	List of specifications of structure matrices for the CAR model. Each element of the list Q is itself a list with 2 elements: \ type: one of c("CAR1","RW1","Gen") indicating a conditional autoregressive structure of order 1 on a rectangular grid, a random walk 1 structure on a line, or a general (any other) structure, and \ content: for "CAR1," a two-vector giving the dimensions of the rectangular grid; for "RW1," a scalar giving the dimension; and for "Gen," a matrix – the symmetric structure matrix.
y	Vector of observed data values. If there is more than one structure matrix in the list called Q, then the data must be ordered accordingly. For example, suppose the data are measured on a rectangular lattice with r rows and c columns, and that Q[[1]] represents the within-row neighborhood structure while Q[[2]] represents the within-column neighborhood structure. Then the data must be in row-major order.
nsamp	Integer representing the number of desired draws from the joint posterior density of the model parameters.
seed	Seed for the random number generator. Must be a positive integer.
fixed	Logical value (TRUE or FALSE). CARrampsOcl.fit utilizes rejection sampling to draw from the marginalized joint posterior density. This involves generating a batch of candidate values, of which some are rejected and some are accepted. If fixed is TRUE, then a single batch of candidates of size nsamp is generated and the number of accepted candidates returned is likely to be much smaller than nsamp. If fixed is FALSE (the default), then additional batches of candidates will be generated iteratively until at least nsamp samples have been accepted.
coefs	Logical value. If coefs is TRUE, then regression coefficients for a polynomial trend surface will be estimated. The design matrix must be provided in the argument designmat.
randeffs	Logical. If true, random effects corresponding to each observation in the dataset will be calculated and returned.
designMat	The design matrix for regression. It must be set to NULL if coefs is FALSE. The only regression coefficients that CARramps can estimate are for polynomial trend surfaces. The degree of the polynomial trend surface must agree with the structure matrix or matrices in the list Q. For a CAR(1) or RandomWalk(1) structure matrix, the appropriate design matrix is an intercept (column of ones) only. RandomWalk(2) structure matrices correspond to linear trend surfaces, and the appropriate design matrix consists of an intercept and a column of consecutive integers. The example code for this function demonstrates how to construct the design matrix for a two-dimensional linear trend surface, which corresponds to a Q list consisting of two RandomWalk(2) structure matrices.
mult	The mult argument and the nsamp argument together determine the size of the batches of candidates generated for rejection sampling, which is mult times nsamp. Leaving mult at its default value of 20 is generally safe. Decreasing mult for very small datasets with high acceptance rates, or increasing it for large datasets with low acceptance rates, may speed computing.
filename	The name for the file in which batches of accepted samples of precision parameters will be saved while the function is running. Future releases of the CARrampsOcl package will be able to use this output file to resume sampling after interruption.

Value

params	Matrix of samples drawn from joint posterior density of the precision parameters. The first column is the measurement error precision, and subsequent columns are the spatial precisions in the same order as the Q matrices.
phi	List with two components: phimean is vector of means of posterior densities random effects; phisd is vector of standard deviations of marginal posterior densities of random effects. Both are in the same order as the observations the y vector.
preds	Always NULL in current release of CARramps. Future releases will enable estimation of means and standard deviations from posterior predictive distributions, which will be output here.
regcoefs	Matrix of samples drawn from joint posterior density of regression coefficients. There is one column for each coefficient.
y	Data vector
acptrate	Rejection sampling acceptance rate
n	Number of observations in y

Author(s)

Kate Cowles

Examples

# load data
  data(iowaSW06)

# construct random walk 2 structure matrix for each dimension
  Q1<- makeRW2Q(33)       # for rows
  Q2<- makeRW2Q(24)       # for columns


# dimensions of Q1, Q2,  in that order
  na<- nrow(Q1)
  nb<- nrow(Q2)

  Q <- list( list(type="Gen",content=Q1), list(type="Gen",content=Q2) )

# construct the design matrix with with as many columns as there are
# in null space of kronecker prod of Q's

  X2 <- cbind( rep(1,nb), 1:nb)
  X1 <- cbind( rep(1,na), 1:na)
  X <-  kronecker( X2, X1)

# parameters of gamma prior densities on tausqy, tausqphi[1], tausqphi[2]
  alpha2 = beta2 <- c(.1, .1, .1)

# number of samples
  nsamp = 100

#random seed
  myseed = 314

  output <- CARrampsOcl.fit(alpha=alpha2,
            beta=beta2, Q=Q, y=iowaSW06,  nsamp=nsamp,
            seed=myseed,
            fixed = FALSE, randeffs=TRUE, coefs=TRUE,designMat=X,
            mult= 50) 

# summarize marginal posterior densities of precision parameters
  library(coda)
  summary(as.mcmc( output$params ))

# summarize marginal posterior densities of regression coefficients
#    intercept, slope within rows (west-to-east linear trend), 
#    slope for columns (north to south linear trend), 
#    interaction between rows  and columns
  summary(as.mcmc( output$regcoefs))

# summary statistics for site-specific random effects at first 10 sites
  print( cbind( output$phi$phimean, output$phi$phisd)[1:10,] )

# plot the raw data and the posterior means of the site-specific random effects
  plot2Q( output, numcols=32, col = rev(terrain.colors(32)), 
  rev.inds = c(FALSE, TRUE))

CARrampsOcl documentation built on May 2, 2019, 3:27 a.m.