inla.sens: Testing code for sensitivity

In INBO-BMK/INLA: Full Bayesian Analysis of Latent Gaussian Models using Integrated Nested Laplace Approximations

Description

TODO: Write a description

Usage

inla.sens(inlaObj)

Arguments

inlaObj

The result from a run of inla.

Value

TODO: This is an EXPERIMENTAL function!

Author(s)

Geir-Arne Fuglstad [e-mail]

See Also

inla

Examples

## Case 1: Simple linear regression on simulated data
  # Number of observations 
  nObs = 100
  
  # Measurement noise
  sdNoise = 0.1
  
  # Coefficients
  mu = 2
  beta = 1
  
  # Covariate
  x = runif(nObs)
  
  # Generate data
  y = mu + beta*x + rnorm(nObs)*sdNoise

  # Make some data unobserved
  nUnObs = 20
  y[(nObs-nUnObs+1):nObs] = NA
  
  # Fit the model
  mod = inla(y ~ x,
           data = list(x = x, y = y))
  
  # Calculate sensitivites
  inla.sens(mod)
  
## Case 2: Time series
  # Length of time-series
  nObs = 100
  
  # Measurement noise
  sdNoise = 0.1

  # Autoregressive process
  rho = 0.6
  sdProc = 0.1
  arP = matrix(0, nrow = nObs, ncol = 1)
  for(i in 2:nObs)
      arP[i] = rho*arP[i-1] + rnorm(1)*sdProc
  tIdx = 1:nObs

  # Coefficients
  mu = 2
  
  # Generate data
  y = mu + arP + rnorm(nObs)*sdNoise

  # Make some data unobserved
  nUnObs = 20
  y[(nObs-nUnObs+1):nObs] = NA
  idx = 1:nObs
  
  # Run INLA
  mod = inla(y ~ f(tIdx, model = "ar1"),
             data = list(y = y, tIdx = tIdx),
             control.inla = list(reordering = "metis"))
  
  # Calculate sensitivities
  inla.sens(mod)

## Case 3: Epil dataset
  data(Epil)
  my.center = function(x) (x - mean(x))

  # make centered covariates
  Epil$CTrt    = my.center(Epil$Trt)
  Epil$ClBase4 = my.center(log(Epil$Base/4))
  Epil$CV4     = my.center(Epil$V4)
  Epil$ClAge   = my.center(log(Epil$Age))
  Epil$CBT     = my.center(Epil$Trt*Epil$ClBase4)

  # Define the model
  formula = y ~ ClBase4 + CTrt + CBT+ ClAge + CV4 +
            f(Ind, model="iid") + f(rand,model="iid")

  mod = inla(formula,family="poisson", data = Epil)
 
  # Calculate sensitivities
  inla.sens(mod)

## Case 4: Spatial data
  # Number of observations
  nObs = 100
  
  # Measurement noise
  sdNoise = 0.2

  # Spatial process
  sdProc = 1.0
  rho0 = 0.2

  # Coefficients
  beta0 = 1
  beta1 = 2
  
  # Generate spatial data + measurement noise
  loc = cbind(runif(nObs), runif(nObs))
  dd = as.matrix(dist(loc))
  Sig = sdProc^2*inla.matern.cov(nu = 1, kappa = sqrt(8)/rho0, dd, corr = TRUE)
  L = t(chol(Sig))
  u = L
  
  # Generate Covariate
  x = runif(nObs)-0.5

  # Combine to observations
  y = beta0 + beta1*x + u
  
  # Number of unobserved
  nUnObs = 2
  y[1:nUnObs] = NA
  
  # Mesh
  mesh = inla.mesh.2d(loc, max.edge = 0.05, cutoff = 0.05)
  
  # Make SPDE object
  spde  = inla.spde2.matern(mesh)
  spde2 = inla.spde2.matern(mesh, constr = TRUE)
  
  # Make A matrix
  A = inla.spde.make.A(mesh, loc)
  
  # Stack
  X = cbind(1, x)
  stk = inla.stack(data = list(y = y), A = list(A, 1),
                   effects = list(field = 1:spde$n.spde,
                                  X = X))
  
  # Run INLA
  mod1 = inla(y ~ -1 + X + f(field, model = spde),
              data = inla.stack.data(stk),
              control.predictor = list(A = inla.stack.A(stk)),
              control.family = list(prior = "pcprec",
                                    param = c(3, 0.05)))
  mod2 = inla(y ~ -1 + X + f(field, model = spde2),
              data = inla.stack.data(stk),
              control.predictor = list(A = inla.stack.A(stk)),
              control.family = list(prior = "pcprec",
                                    param = c(3, 0.05)))

  # Calculate sensitivities
  res1 = inla.sens(mod1)
  res2 = inla.sens(mod2)

INBO-BMK/INLA documentation built on Dec. 4, 2019, 11:43 p.m.