Perform stepwise and full search variable selection with INLA

Run forwards or backwards stepwise and full search variable selection with INLA regression. Only fixed effects will be considered for removal. Use holdout data or within sample data to test new models. Choose the level of interactions and power terms to expand the model formula to. Force part of the model to remain in the model (e.g. the spatial term and intercept) or limit the number of predictor variables during a full search selection.

Usage

INLAstep(
  fam1 = "gaussian",
  dataf,
  spde = NULL,
  in_stack = NULL,
  invariant = "0 + Intercept",
  direction = c("forwards", "backwards", "none"),
  y = NULL,
  y2 = NULL,
  include = NULL,
  include_max = NULL,
  powerl = 1,
  inter = 1,
  thresh = 2,
  num.threads = 1,
  ...
)

Arguments

fam1: String defining the likelihood familiy.
dataf: A data.frame including covariates and response data.
spde: An spde model object for the spatial component.
in_stack: An inla.data.stack object containing all needed data.
invariant: The part of the formula that should not change (e.g. the intercept and the spatial component.)
direction: string 'forwards' for forward variable selection, 'backwards' for backwards variable elimination and 'none' for a full search across all possible formulas. To limit the number of covariates and computational efforts for a large number of columns in dataf set include_max.
y: String determining the response variable.
y2: String determining the name of the test response data.
include: Vector of integers to determine which columns in dataf should be used. If NULL, use all columns except y and y2.
include_max: Integer determining the maximum number or covariates for a full search selected with direction = 'none'. If NULL, use all columns except y and y2.
powerl: Integer up to 3 determining which power terms to include.
inter: Integer up to 3 determining how many levels of interactions to include.
thresh: Threshold difference in DIC for whether a new model should replace the old model in 'forward' or 'backward' variable selection.
num.threads: How many threads to use for INLA computation.
...: Further arguments to INLA::inla function.

Value

An object of class INLAstep.

References

redding_2017elbe1d

Examples

if (FALSE) { # \dontrun{
  library(INLA)
  data(Epil)
  stack <- INLA::inla.stack(data = list(y = Epil$y),
                           A = list(1),
                           effects = list(data.frame(Intercept = 1, Epil[3:5])))
                      
 
  result <- INLAstep(fam1 = "poisson", 
                     Epil,
                     in_stack = stack,
                     invariant = "0 + Intercept",
                     direction = 'backwards',
                     include = 3:5,
                     y = 'y',
                     y2 = 'y',
                     powerl = 1,
                     inter = 1,
                     thresh = 2)

  autoplot(result$best_model, which = c(1, 5), CI = TRUE)
} # }