spT.Gibbs: MCMC sampling for the spatio-temporal models.

Description

This function is used to draw MCMC samples using the Gibbs sampler.

Usage

spT.Gibbs(formula, data=parent.frame(), model="GP", time.data=NULL, 
	coords, knots.coords, newcoords=NULL, newdata=NULL, priors=NULL, 
	initials=NULL, nItr=5000, nBurn=1000, report=1, tol.dist=0.05, 
	distance.method="geodetic:km", cov.fnc="exponential", scale.transform="NONE", 
	spatial.decay=spT.decay(distribution="FIXED"), annual.aggrn="NONE")

Arguments

formula

The symnbolic description of the model equation of the regression part of the space-time model.

data

An optional data frame containing the variables in the model. If omitted, the variables are taken from environment(formula), typically the environment from which spT.Gibbs is called. The data should be ordered first by the time and then by the sites speci

model

The spatio-temporal models to be fitted, current choices are: "GP", "AR", and "GPP" with the first one as the default.

time.data

Defining the segments of the time-series set up using the function spT.time.

coords

The n by 2 matrix or data frame defining the locations (e.g., longitude/easting, latitude/northing) of the fitting sites, where n is the number of fitting sites. One can also supply coordinates through a formula argument such as ~Longitude+Latitude.

knots.coords

The locations of the knots in similar format to coords above, only required if model="GPP".

newcoords

The locations of the prediction sites in similar format to coords above, only required if fit and predictions are to be performed simultaneously. If omitted, no predictions will be performed.

newdata

The covariate values at the prediction sites specified by newcoords. This should have same space-time structure as the original data frame.

priors

The prior distributions for the parameters. Default distributions are specified if these are not provided. If priors=NULL a flat prior distribution will be used with large variance. See details in spT.priors

initials

The preferred initial values for the parameters. If omitted, default values are provided automatically. Further details are provided in spT.initials.

nItr

Number of MCMC iterations. Default value is 5000.

nBurn

Number of burn-in samples. This number of samples will be discarded before making any inference. Default value is 1000.

report

Number of reports to display while running the Gibbs sampler. Defaults to number of iterations.

distance.method

The preferred method to calculate the distance between any two locations. The available options are "geodetic:km", "geodetic:mile", "euclidean", "maximum", "manhattan", and "canberra". See details in dist. The

tol.dist

Minimum separation distance between any two locations out of those specified by coords, knots.coords and pred.coords. The default is 0.005. The programme will exit if the minimum distance is less than the non-zero specified value. This will ensure non-si

cov.fnc

Covariance function for the spatial effects. The available options are "exponential", "gaussian", "spherical" and "matern". If "matern" is used then by default the smooth parameter ($\nu$) is estimated from (0,1) uniform distribution using discrete sampl

scale.transform

The transformation method for the response variable. Currently implemented options are: "NONE", "SQRT", and "LOG" with "NONE" as the deault.

spatial.decay

Provides the prior distribution for the spatial decay parameter $\phi$. Currently implemented options are "FIXED", "Unif", or "Gamm". Further details for each of these are specified by spT.decay.

annual.aggrn

This provides the options for calculating annual summary statistics by aggregating different time segments (e.g., annual mean). Currently implemented options are: "NONE", "ave" and "an4th", where "ave" = annual average, "an4th"= annual 4th highest. Only a

Value

acceptThe acceptance rate for the $\phi$ parameter if the "MH" method of sampling is chosen.
phipMCMC samples for the parameter $\phi$.
nupMCMC samples for the parameter $\nu$. Only available if "matern" covariance function is used.
sig2epsMCMC samples for the parameter $\sigma^2_\epsilon$.
sig2etapMCMC samples for the parameter $\sigma^2_\eta$.
betapMCMC samples for the parameter $\beta$.
rhopMCMC samples for $\rho$ for the AR or GPP model.
opMCMC samples for the true observations.
fittedMCMC summary (mean and sd) for the fitted values.
tol.distMinimum tolerance distance limit between the locations.
distance.methodName of the distance calculation method.
cov.fncName of the covariance function used in model fitting.
scale.transformName of the scale.transformation method.
sampling.sp.decayThe method of sampling for the spatial decay parameter $\phi$.
covariate.namesName of the covariates used in the model.
Distance.matrixThe distance matrix.
coordsThe coordinate values.
nTotal number of sites.
rTotal number of segments in time, e.g., years.
TTotal points of time, e.g., days within each year.
pTotal number of model coefficients, i.e., $\beta$'s including the intercept.
initialsThe initial values used in the model.
priorsThe prior distributions used in the model.
PMCCThe predictive model choice criteria obtained by minimising the expected value of a loss function, see Gelfand and Ghosh (1998). Results for both goodness of fit and penalty are given.
iterationsThe number of samples for the MCMC chain, without burn-in.
nBurnThe number of burn-in period for the MCMC chain.
computation.timeThe computation time required for the fitted model.
modelThe spatio-temporal model used for analyse the data.
Text OutputThis option is only applicable when fit and predictions are done simultaneously. For GP models: OutGP_Values_Parameter.txt: (nItr x parameters matrix) has the MCMC samples for the parameters, ordered as: beta's, sig2eps, sig2eta, and phi. OutGP_Stats_FittedValue.txt: (N x 2) matrix of fitted summary, with 1st column as mean and 2nd column as standard deviations, where N=nrT. OutGP_Stats_PredValue.txt: ((predsites*r*T) x 2) matrix of prediction summary, with 1st column as mean and 2nd column as standard deviations. OutGP_Values_Prediction.txt: (nItr x (predsites*r*T)) matrix of MCMC predicted values in the predicted sites. If annual.aggregation="ave" then we get text output as: OutGP_Annual_Average_Prediction.txt: (nItr x (predsites*r)) matrix. If annual.aggregation="an4th" then we get text output as: OutGP_Annual_4th_Highest_Prediction.txt: (nItr x (predsites*r)) matrix. For AR models: OutAR_Values_Parameter.txt: (nItr x parameters matrix) has the MCMC samples for the parameters, ordered as: beta's, rho, sig2eps, sig2eta, mu_l's, sig2l's and phi. OutAR_Stats_TrueValue.txt: (N x 2) matrix of true summary values, with 1st column as mean and 2nd column as standard deviations. OutAR_Stats_FittedValue.txt: (N x 2) matrix of fitted summary, with 1st column as mean and 2nd column as standard deviations. OutAR_Stats_PredValue.txt: ((predsites*r*T) x 2) matrix of prediction summary, with 1st column as mean and 2nd column as standard deviations. OutAR_Values_Prediction.txt: (nItr x (predsites*r*T)) matrix of MCMC predicted values in the predicted sites. If annual.aggregation="ave" then we get text output as: OutAR_Annual_Average_Prediction.txt: (nItr x (predsites*r)) matrix. If annual.aggregation="an4th" then we get text output as: OutAR_Annual_4th_Highest_Prediction.txt: (nItr x (predsites*r)) matrix. For models using GPP approximations: OutGPP_Values_Parameter.txt: (nItr x parameters matrix) has the MCMC samples for the parameters, ordered as: beta's, rho, sig2eps, sig2eta, and phi. OutGPP_Stats_FittedValue.txt: (N x 2) matrix of fitted summary, with 1st column as mean and 2nd column as standard deviations. OutGPP_Stats_PredValue.txt: ((predsites*r*T) x 2) matrix of prediction summary, with 1st column as mean and 2nd column as standard deviations. OutGPP_Values_Prediction.txt: (nItr x (predsites*r*T)) matrix of MCMC predicted values in the predicted sites. If annual.aggregation="ave" then we get text output as: OutGPP_Annual_Average_Prediction.txt: (nItr x (predsites*r)) matrix. If annual.aggregation="an4th" then we get text output as: OutGPP_Annual_4th_Highest_Prediction.txt: (nItr x (predsites*r)) matrix.

References

1. Bakar, K.S. & Sahu, S.K. (2015). spTimer: Spatio-Temporal Bayesian Modelling Using R. Journal of Statistical Software. 63(15). 1-32. 2. Sahu, S.K. & Bakar, K.S. (2012). Hierarchical Bayesian Autoregressive Models for Large Space Time Data with Applications to Ozone Concentration Modelling. Applied Stochastic Models in Business and Industry, 28, 395-415. 3. Sahu, S.K., Gelfand, A.E., & Holland, D.M. (2007). High-Resolution Space-Time Ozone Modelling for Assessing Trends. Journal of the American Statistical Association, 102, 1221-1234. 4. Bakar, K.S. (2012). Bayesian Analysis of Daily Maximum Ozone Levels. PhD Thesis, University of Southampton, Southampton, United Kingdom.

Examples

Run this code

##

###########################
## Attach library spTimer
###########################

library(spTimer)

###########################
## The GP models:
###########################

##
## Model fitting
##

# Read data 
data(NYdata)

# MCMC via Gibbs using default choices
set.seed(11)
post.gp <- spT.Gibbs(formula=o8hrmax ~cMAXTMP+WDSP+RH,   
        data=NYdata, model="GP", coords=~Longitude+Latitude, 
        scale.transform="SQRT")
print(post.gp)

# MCMC via Gibbs not using default choices
# Read data 
s<-c(8,11,12,14,18,21,24,28)
DataFit<-spT.subset(data=NYdata, var.name=c("s.index"), s=s, reverse=TRUE) 
DataFit<-subset(DataFit, with(DataFit, !(Day %in% c(30, 31) & Month == 8)))
DataValPred<-spT.subset(data=NYdata, var.name=c("s.index"), s=s) 
DataValPred<-subset(DataValPred, with(DataValPred, !(Day %in% c(30, 31) & Month == 8)))

# define the time-series 
time.data<-spT.time(t.series=60,segment=1)

# hyper-parameters for the prior distributions
priors<-spT.priors(model="GP",inv.var.prior=Gamm(2,1),
        beta.prior=Norm(0,10^4))

# initial values for the model parameters
initials<-spT.initials(model="GP", sig2eps=0.01, 
            sig2eta=0.5, beta=NULL, phi=0.001)

# input for spatial decay, any one approach from below
#spatial.decay<-spT.decay(distribution="FIXED", value=0.01)
spatial.decay<-spT.decay(distribution=Gamm(2,1), tuning=0.08)
#spatial.decay<-spT.decay(distribution=Unif(0.01,0.02),npoints=5)

# Iterations for the MCMC algorithms
nItr<-5000

# MCMC via Gibbs
set.seed(11)
post.gp <- spT.Gibbs(formula=o8hrmax ~ cMAXTMP+WDSP+RH, 
         data=DataFit, model="GP", time.data=time.data, 
         coords=~Longitude+Latitude, priors=priors, initials=initials, 
         nItr=nItr, nBurn=0, report=nItr, 
         tol.dist=2, distance.method="geodetic:km", 
         cov.fnc="exponential", scale.transform="SQRT", 
         spatial.decay=spatial.decay)
print(post.gp)

# Summary and plots
summary(post.gp)
summary(post.gp,pack="coda")
plot(post.gp)
plot(post.gp,residuals=TRUE)

coef(post.gp)
confint(post.gp)
terms(post.gp)
formula(post.gp)
model.frame(post.gp)
model.matrix(post.gp)

# Model selection criteria
post.gp$PMCC 

##
## Fit and spatially prediction simultaneously
##

# Read data 
s<-c(8,11,12,14,18,21,24,28)
DataFit<-spT.subset(data=NYdata, var.name=c("s.index"), s=s, reverse=TRUE) 
DataFit<-subset(DataFit, with(DataFit, !(Day %in% c(30, 31) & Month == 8)))
DataValPred<-spT.subset(data=NYdata, var.name=c("s.index"), s=s) 
DataValPred<-subset(DataValPred, with(DataValPred, !(Day %in% c(30, 31) & Month == 8)))

# Define the coordinates
coords<-as.matrix(unique(cbind(DataFit[,2:3])))
pred.coords<-as.matrix(unique(cbind(DataValPred[,2:3])))

# MCMC via Gibbs will provide output in *.txt format  
# from C routine to avoide large data problem in R
set.seed(11)
post.gp.fitpred <- spT.Gibbs(formula=o8hrmax ~cMAXTMP+WDSP+RH,   
         data=DataFit, model="GP", coords=coords, 
         newcoords=pred.coords, newdata=DataValPred,
         scale.transform="SQRT")
print(post.gp.fitpred)
summary(post.gp.fitpred)
coef(post.gp.fitpred)
plot(post.gp.fitpred)
names(post.gp.fitpred)

# validation criteria
spT.validation(DataValPred$o8hrmax,c(post.gp.fitpred$prediction[,1]))  


######################################
## The GP model for sp class data
######################################

# Creating sp class data
library(sp)
data(meuse)
summary(meuse)
coordinates(meuse) <- ~x+y
class(meuse)
out<-spT.Gibbs(formula=zinc~sqrt(dist),data=meuse,
               model="GP", scale.transform="LOG")
summary(out)

# Create a dataset with spacetime class
library(spTimer)
site<-unique(NYdata[,c("Longitude","Latitude")])
library(spacetime)
row.names(site)<-paste("point",1:nrow(site),sep="")
site <- SpatialPoints(site)
ymd<-as.POSIXct(seq(as.Date("2006-07-01"),as.Date("2006-08-31"),by=1))
# introduce class STFDF
newNYdata<-STFDF(sp=site, time=ymd, data=NYdata) # full lattice
class(newNYdata)
out <- spT.Gibbs(formula=o8hrmax ~cMAXTMP+WDSP+RH,   
      data=newNYdata, model="GP", scale.transform="SQRT")
summary(out)


###########################
## The AR models:
###########################

##
## Model fitting
##

# Read data 
data(NYdata)

# Define the coordinates
coords<-as.matrix(unique(cbind(NYdata[,2:3])))

# MCMC via Gibbs using default choices
set.seed(11)
post.ar <- spT.Gibbs(formula=o8hrmax ~cMAXTMP+WDSP+RH,   
         data=NYdata, model="AR", coords=coords, 
         scale.transform="SQRT")
print(post.ar)

# MCMC via Gibbs not using default choices
# define the time-series 
time.data<-spT.time(t.series=62,segment=1)

# hyper-parameters for the prior distributions
priors<-spT.priors(model="AR",inv.var.prior=Gamm(2,1),
        beta.prior=Norm(0,10^4))

# initial values for the model parameters
initials<-spT.initials(model="AR", sig2eps=0.01, 
            sig2eta=0.5, beta=NULL, phi=0.001)

# Input for spatial decay
#spatial.decay<-spT.decay(distribution="FIXED", value=0.01)
spatial.decay<-spT.decay(distribution=Gamm(2,1), tuning=0.08)
#spatial.decay<-spT.decay(distribution=Unif(0.01,0.02),npoints=5)

# Iterations for the MCMC algorithms
nItr<-5000

# MCMC via Gibbs
set.seed(11)
post.ar <- spT.Gibbs(formula=o8hrmax~cMAXTMP+WDSP+RH, 
         data=NYdata, model="AR", time.data=time.data, 
         coords=coords, priors=priors, initials=initials, 
         nItr=nItr, nBurn=0, report=nItr, 
         tol.dist=2, distance.method="geodetic:km", 
         cov.fnc="exponential", scale.transform="SQRT", 
         spatial.decay=spatial.decay)
print(post.ar)

# Summary and plots
summary(post.ar)
plot(post.ar)

# Model selection criteria
post.ar$PMCC 

##
## Fit and spatially prediction simultaneously
##

# Read data 
s<-c(8,11,12,14,18,21,24,28)
DataFit<-spT.subset(data=NYdata, var.name=c("s.index"), s=s, reverse=TRUE) 
DataFit<-subset(DataFit, with(DataFit, !(DayDataValPred<-spT.subset(data=NYdata, var.name=c("s.index"), s=s) 
DataValPred<-subset(DataValPred, with(DataValPred, !(Day
# Define the coordinates
coords<-as.matrix(unique(cbind(DataFit[,2:3])))
pred.coords<-as.matrix(unique(cbind(DataValPred[,2:3])))

# MCMC via Gibbs will provide output in *.txt format  
# from C routine to avoide large data problem in R
set.seed(11)
post.ar.fitpred <- spT.Gibbs(formula=o8hrmax ~cMAXTMP+WDSP+RH,   
         data=DataFit, model="AR", coords=coords, 
         newcoords=pred.coords, newdata=DataValPred,
         scale.transform="SQRT")
print(post.ar.fitpred)
summary(post.ar.fitpred)
names(post.ar.fitpred)

# validation criteria
spT.validation(DataValPred$o8hrmax,c(post.ar.fitpred$prediction[,1]))  

#################################
## The GPP approximation models:
#################################

##
## Model fitting
##

# Read data 
data(NYdata); 

# Define the coordinates
coords<-as.matrix(unique(cbind(NYdata[,2:3])))
# Define knots
knots<-spT.grid.coords(Longitude=c(max(coords[,1]),
              min(coords[,1])),Latitude=c(max(coords[,2]),
              min(coords[,2])), by=c(4,4))

# MCMC via Gibbs using default choices
set.seed(11)
post.gpp <- spT.Gibbs(formula=o8hrmax ~cMAXTMP+WDSP+RH,   
         data=NYdata, model="GPP", coords=coords, 
         knots.coords=knots, scale.transform="SQRT")
print(post.gpp)

# MCMC via Gibbs not using default choices
# define the time-series 
time.data<-spT.time(t.series=62,segment=1)

# hyper-parameters for the prior distributions
priors<-spT.priors(model="GPP",inv.var.prior=Gamm(2,1),
        beta.prior=Norm(0,10^4))

# initial values for the model parameters
initials<-spT.initials(model="GPP", sig2eps=0.01, 
            sig2eta=0.5, beta=NULL, phi=0.001)

# input for spatial decay
#spatial.decay<-spT.decay(distribution="FIXED", value=0.001)
spatial.decay<-spT.decay(distribution=Gamm(2,1), tuning=0.05)
#spatial.decay<-spT.decay(distribution=Unif(0.001,0.009),npoints=10)

# Iterations for the MCMC algorithms
nItr<-5000 

# MCMC via Gibbs
set.seed(11)
post.gpp <- spT.Gibbs(formula=o8hrmax~cMAXTMP+WDSP+RH, 
         data=NYdata, model="GPP", time.data=time.data, 
         coords=coords, knots.coords=knots,
         priors=priors, initials=initials, 
         nItr=nItr, nBurn=0, report=nItr, 
         tol.dist=2, distance.method="geodetic:km", 
         cov.fnc="exponential", scale.transform="SQRT", 
         spatial.decay=spatial.decay)
print(post.gpp)

# Summary and plots
summary(post.gpp)
plot(post.gpp)

# Model selection criteria
post.gpp$PMCC 

##
## Fit and spatially prediction simultaneously
##

# Read data 
s<-c(8,11,12,14,18,21,24,28)
DataFit<-spT.subset(data=NYdata, var.name=c("s.index"), s=s, reverse=TRUE) 
DataFit<-subset(DataFit, with(DataFit, !(Day %in% c(30, 31) & Month == 8)))
DataValPred<-spT.subset(data=NYdata, var.name=c("s.index"), s=s) 
DataValPred<-subset(DataValPred, with(DataValPred, !(Day %in% c(30, 31) & Month == 8)))

# Define the coordinates
coords<-as.matrix(unique(cbind(DataFit[,2:3])))
pred.coords<-as.matrix(unique(cbind(DataValPred[,2:3])))
knots<-spT.grid.coords(Longitude=c(max(coords[,1]),
              min(coords[,1])),Latitude=c(max(coords[,2]),
              min(coords[,2])), by=c(4,4))

# MCMC via Gibbs will provide output in *.txt format  
# from C routine to avoide large data problem in R
set.seed(11)
post.gpp.fitpred <- spT.Gibbs(formula=o8hrmax ~cMAXTMP+WDSP+RH,   
         data=DataFit, model="GP", coords=coords, knots.coords=knots,
         newcoords=pred.coords, newdata=DataValPred,
         scale.transform="SQRT")
print(post.gpp.fitpred)
summary(post.gpp.fitpred)
plot(post.gpp.fitpred)

names(post.gpp.fitpred)

# validation criteria
spT.validation(DataValPred$o8hrmax,c(post.gpp.fitpred$prediction[,1]))  

##

Run the code above in your browser using DataLab