inlasloo: A function to run a spatial leave-one-out cross-validation in...

In timcdlucas/INLAutils: Utility Functions for 'INLA'

A function to run a spatial leave-one-out cross-validation in R using INLA.

Description

This function will apply a spatial leave-one-out cross-validation (SLOO-CV) of one or several models running on INLA. The function requires a dataframe that includes a response, coordinates (longitude, latidude). Covariates can be inluced as well in the dataframe. Furthermore, the user needs to provide a radius around which points will be removed during the SLOO-CV process. The number of iterations needs to be provided as well. Also the user needs to provide INLA inputs, including: formula, family (and ntrials for Binomial), and mesh In addition, one can provide further INLA specifications, such as int.strategy='eb'. SLOO-CV computes the root mean square error (rmse) with a tolerance for confidence interval defined by the user. In option, other scores can be computed, such as the mean absolute error, dawid-sebastiani or using square root of the data.

Usage

inlasloo(
  dataframe,
  long,
  lat,
  y,
  ss,
  rad,
  modform,
  family,
  mesh,
  ntrials = NULL,
  int.strategy = "eb",
  alpha = 0.05,
  mae = FALSE,
  ds = FALSE,
  sqroot = FALSE,
  print = FALSE,
  plot = TRUE,
  ...
)

Arguments

dataframe	dataframe including the variables to be used
long	Character string giving the name of the longitude variable
lat	Character string giving the name of the latitude variable
y	Character string giving the names of the response
ss	Numeric value giving the sample size to process (number of SLOO runs)
rad	Numeric value giving the radius for the spatial buffer around left-out point's location
modform	Formula or list of Formula for the model or models, respectively in INLA format
family	Character string or list of Character string giving the family or families, respectively of the response in INLA format
mesh	inla.mesh object, consisting in triangle mesh to be defined using INLA.
ntrials	vector of positive integers c(1,2,...) setting the number of trials for a Binomial family
int.strategy	Character string giving INLA integration strategy
alpha	Numeric value (0,...,1) giving the threshold for computing confidence intervals (1-alpha) of rmse and mae estimation
mae	If TRUE, compute the mean absolute error (mae) and the root mean square error (rmse). If FALSE, compute the rmse only.
ds	If TRUE, compute the Dawid-Sebastiani score (ds). If FALSE, does not compute ds.
sqroot	If TRUE, compute the square root of the observed and predicted values to generate the rmse and/or mae. If FALSE, the rmse and/or mae are computed without transformation of the data.
print	Logical to determine whether to print a summary of the results.
plot	Logical to determine whether to print the plot of the results.
...	other arguments passed to inla

Examples

 ## Not run: 
require(INLA)
require(sp)
require(grDevices)

# generate a dataframe and INLA output for the function
set.seed(10)
coords <- data.frame(long = c(rnorm(70), rnorm(30, 3)), lat = rnorm(100))
PA <- rep(c(0, 1), each = 50)
x <- data.frame(x1 = rnorm(100),x2 = c(rnorm(70), rnorm(30, 2)))# x1 no relat., x2 pos. relat.
dataf1 <- sp::SpatialPointsDataFrame(coords = coords, data = data.frame(y = PA, x))
mesh <- INLA::inla.mesh.2d(loc = sp::coordinates(dataf1),max.edge = c(3, 3),cutoff = 1.3)
spde <- INLA::inla.spde2.matern(mesh, alpha=2)#SPDE model is defined
A <- INLA::inla.spde.make.A(mesh, loc = sp::coordinates(dataf1))#projector matrix
dataframe <- data.frame(dataf1) #generate dataframe with response and covariate
modform<-stats::as.formula(paste('y ~ -1+ x1 + x2 + y.intercept + f(spatial.field, model=spde)'))
stk <- INLA::inla.stack(data=list(y=dataframe$y),A=list(A, 1),
effects=list(list(spatial.field=1:spde$n.spde),
list(y.intercept=rep(1,length(dataframe$y)),covariate=dataframe[c(-1)])),tag='est')
out <- INLA::inla(modform, family='normal',Ntrials = 1, data=INLA::inla.stack.data(stk, spde=spde),
                  control.predictor = list(A=INLA::inla.stack.A(stk),link=1),
                  control.compute = list( config=TRUE),control.inla = list(int.strategy='eb'))
out.field <- INLA::inla.spde2.result(out,'spatial.field', spde, do.transf=TRUE)
range.out <- INLA::inla.emarginal(function(x) x, out.field$marginals.range.nominal[[1]])

# parameters for the SLOO process
ss <- 1#sample size to process (number of SLOO runs)
rad <- range.out*0.15#define the radius of the spatial buffer surrounding the removed point
mesh <- mesh#use the mesh of the model
dataframe <- dataframe#dataframe with response 'y' and covariates 'x1', 'x2'
dataframe$y <- round(runif(length(dataframe$y), 1, 12))#create positive discrete response
modform <- stats::as.formula(paste('y ~ -1+ y.intercept + x1 + f(spatial.field, model=spde)'))
family <- list('gamma')#one model
ntrials <- rep(round(max(dataframe$y,na.rm=TRUE)*2),length(dataframe$y))# create ntrials for Binomial family
alpha <- 0.05#rmse and mae confidence intervals (1-alpha)

# run the function
inlasloo(dataframe=dataframe, long='long', lat='lat',y= 'y', ss=ss, rad=rad,
modform=modform,mesh=mesh,family=family,alpha=0.05,mae=TRUE,ds=TRUE,sqroot=FALSE)

# SLOO function with one model formula (Binomial) and linear terms
sloo1<-inlasloo(dataframe=dataframe, long='long', lat='lat', y='y', ss=1, rad=0.1,
ntrials=ntrials,modform='y ~ -1+ x1 + x2 + y.intercept + f(spatial.field, model=spde)',mesh=mesh,
family='binomial',alpha=0.05,mae=FALSE,ds=FALSE,sqroot=FALSE)

# SLOO function with two families (Binomial and Poisson) and linear terms
sloo2<-inlasloo(dataframe=dataframe, long='long', lat='lat', y='y', ss=1, rad=0.1,
ntrials=ntrials,modform='y ~ -1+ x1 + x2 + y.intercept + f(spatial.field, model=spde)',mesh=mesh,
family=list('binomial','poisson'),alpha=0.05,mae=TRUE,ds=FALSE,sqroot=FALSE)

# SLOO function with one family (Binomial) and two model formulas
sloo3<-inlasloo(dataframe=dataframe, long='long', lat='lat', y='y', ss=1,
rad=0.1, ntrials=ntrials,modform=list('y ~ -1+ y.intercept + f(spatial.field, model=spde)',
'y ~ -1+ x1 + x2 + y.intercept + f(spatial.field, model=spde)'),
mesh=mesh,family='binomial',alpha=0.05,mae=TRUE,ds=FALSE,sqroot=FALSE)
 
# SLOO function with two families (Binomial and Poisson), two model formulas, 
# and MAE, DS, and sqroot activated
sloo4<-inlasloo(dataframe=dataframe, long='long', lat='lat', y='y', ss=1,
rad=0.1, ntrials=ntrials,modform=list('y ~ -1+ y.intercept + f(spatial.field, model=spde)',
'y ~ -1+ x1 + x2 + y.intercept + f(spatial.field, model=spde)'),
mesh=mesh,list('binomial','poisson'),alpha=0.05,mae=TRUE,ds=TRUE,sqroot=TRUE)

## End(Not run)

timcdlucas/INLAutils documentation built on Nov. 29, 2022, 5:41 a.m.