bmstdr (version 0.8.2)

Bmoving_sptime: Model fitting and validation for spatio-temporal data from moving sensors in time.

Description

Model fitting and validation for spatio-temporal data from moving sensors in time.

Usage

Bmoving_sptime(
  formula,
  data,
  coordtype,
  coords,
  prior.sigma2 = c(2, 1),
  prior.tau2 = c(2, 1),
  prior.phi = NULL,
  prior.phi.param = NULL,
  scale.transform = "NONE",
  ad.delta = 0.8,
  t.depth = 12,
  s.size = 0.01,
  N = 2500,
  burn.in = 1000,
  no.chains = 1,
  validrows = 10,
  predspace = FALSE,
  newdata = NULL,
  mchoice = TRUE,
  plotit = FALSE,
  rseed = 44,
  verbose = TRUE,
  knots.coords = NULL,
  g_size = 5
)

Value

A list containing:

• params - A table of parameter estimates

• fit - The fitted model object.

• datatostan - A list containing all the information sent to the rstan package.

• prior.phi.param - This contains the values of the hyperparameters of the prior distribution for the spatial decay parameter \(phi\).

• prior.phi - This contains the name of of the prior distribution for the spatial decay parameter \(phi\).

• validationplots - Present only if validation has been performed. Contains three validation plots with or without segment and an ordinary plot. See obs_v_pred_plot for more.

• fitteds - A vector of fitted values.

• residuals - A vector of residual values.

• package - The name of the package used for model fitting. This is always stan for this function.

• model - The name of the fitted model.

• call - The command used to call the model fitting function.

• formula - The input formula for the regression part of the model.

• scale.transform - The transformation adopted by the input argument with the same name.

• sn - The number of data locations used in fitting.

• tn - The number of time points used in fitting for each location.

• computation.time - Computation time required to run the model fitting.

Arguments

formula

An object of class "formula" (or one that can be coerced to that class): a symbolic description of the model to be fitted.

data

The data frame for which the model formula is to be fitted. The data frame should be in long format having one row for each location and time combination. The data frame must be ordered by time within each site, and should optionally have a column, named s.index, providing the site indices. Thus the data, with n sites and T times within each site, should be organized in the order: (s1, t1), (s1, t2), ... (s1, T), ... (sn, t1), ... (sn, T). The data frame should also contain two columns giving the coordinates of the locations for spatio temporal model fitting.

coordtype

Type of coordinates: utm, lonlat or plain with utm (supplied in meters) as the default. Distance will be calculated in units of kilometer if this argument is either utm or lonlat. Euclidean distance will be calculated if this is given as the third type plain. If distance in meter is to be calculated then coordtype should be passed on as plain although the coords are supplied in UTM.

coords

A vector of size 2 giving the column numbers of the data frame which contain the coordinates of the data locations. Here the supplied data frame must contain a column named `time` which should indicate the time index of the data row. The values in the column `time` should be positive integers starting from 1.

prior.sigma2

Shape and scale parameter value for the gamma prior on 1/sigma^2, the precision.

prior.tau2

Shape and scale parameter value for the gamma prior on tau^2, the nugget effect.

prior.phi

Specifies the prior distribution for \(\phi\) only when package is one of Stan, spTimer or spTDyn. Distribution options uniform specified by "Unif" and gamma specified by "Gamm" have been implemented in both Stan and spTimer. Additionally a half-Cauchy prior distribution specified as "Cauchy" has been implemented in Stan. In the case of spTimer the uniform distribution is discrete while in the case of Stan the uniform distribution is continuous. In the case of spTimer the option "FIXED" can be used to keep the value fixed. In that case the fixed value can be given by by a scalar value as the argument prior.phi.param below or it can be left unspecified in which case the fixed value of \(\phi\) is chosen as 3/maximum distance between the data locations. The "FIXED" option is not available for the Stan package.

prior.phi.param

Lower and upper limits of the uniform prior distribution for phi the spatial decay parameter. For the default uniform distribution the values correspond to an effective range that is between 1% and 100% of the maximum distance between the data locations. For the Gamma distribution the default values are 2 and 1 and for the Cauchy distribution the default values are 0, 1 which specifies a half-Cauchy distribution in \((0, \infty)\).

scale.transform

Transformation of the response variable. It can take three values: SQRT, LOG or NONE. Default value is "NONE".

ad.delta

Adaptive delta controlling the behavior of Stan during fitting.

t.depth

Maximum allowed tree depth in the fitting process of Stan.

s.size

step size in the fitting process of Stan.

N

MCMC sample size.

burn.in

How many initial iterations to discard. Only relevant for MCMC based model fitting, i.e., when package is spBayes or Stan.

no.chains

Number of parallel chains to run in Stan.

validrows

Either a number of randomly selected data rows to validate or a vector giving the row numbers of the data set for validation.

predspace

A 0-1 flag indicating whether spatial predictions are to be made.

newdata

A new data frame with the same column structure as the model fitting data set.

mchoice

Logical scalar value: whether model choice statistics should be calculated.

plotit

Logical scalar value: whether to plot the predictions against the observed values.

rseed

Random number seed that controls the starting point for the random number stream. A set value is required to help reproduce the results.

verbose

Logical scalar value: whether to print various estimates and statistics.

knots.coords

Only relevant for GPP models fitted by either spTimer or spTDyn. Optional two column matrix of UTM-X and UTM-Y coordinates of the knots in kilometers. It is preferable to specify the g_size parameter instead.

g_size

Only relevant for GPP models fitted by either spTimer or spTDyn. The grid size c(m, n) for the knots for the GPP model. A square grid is assumed if this is passed on as a scalar. This does not need to be given if knots.coords is given instead.

See Also

Bsptime for spatio-temporal model fitting.

Examples

# \donttest{
deep <- argo_floats_atlantic_2003[argo_floats_atlantic_2003$depth==3, ]
deep$x2inter <- deep$xinter*deep$xinter
deep$month <- factor(deep$month)
deep$lat2 <- (deep$lat)^2
deep$sin1 <- round(sin(deep$time*2*pi/365), 4)
deep$cos1 <- round(cos(deep$time*2*pi/365), 4)
deep$sin2 <- round(sin(deep$time*4*pi/365), 4)
deep$cos2 <- round(cos(deep$time*4*pi/365), 4)
deep[, c( "xlat2", "xsin1", "xcos1", "xsin2", "xcos2")] <- 
scale(deep[,c("lat2", "sin1", "cos1", "sin2", "cos2")])
f2 <- temp ~ xlon + xlat + xlat2+ xinter + x2inter 
M2 <- Bmoving_sptime(formula=f2, data = deep, coordtype="lonlat", 
coords = 1:2, N=11, burn.in=6, validrows =NULL, mchoice = FALSE)
summary(M2)
plot(M2)
names(M2)
# Testing for smaller data sets with different data pattern  
d2 <- deep[1:25, ]
d2$time <- 1:25 
# Now there is no missing times 
M1 <- Bmoving_sptime(formula=f2, data = d2, coordtype="lonlat", coords = 1:2, 
N=11, burn.in=6,  mchoice = FALSE) 
summary(M1)
d2[26, ] <- d2[25, ]
# With multiple observation at the same location and time 
M1 <- Bmoving_sptime(formula=f2, data = d2, coordtype="lonlat", coords = 1:2, 
N=11, burn.in=6,  mchoice = FALSE) 
summary(M1)
d2[27, ] <- d2[24, ]
d2[27, 3] <- 25
# With previous location re-sampled 
M1 <- Bmoving_sptime(formula=f2, data = d2, coordtype="lonlat", coords = 1:2, 
N=11, burn.in=6,  mchoice = FALSE) 
summary(M1)
# }