Perform stepwise selection on a linear model

Usage

tl_step_selection(
  data,
  formula,
  direction = "backward",
  criterion = "AIC",
  trace = FALSE,
  steps = 1000,
  ...
)

Arguments

data: A data frame containing the training data
formula: A formula specifying the initial model
direction: Direction of stepwise selection: "forward", "backward", or "both"
criterion: Criterion for selection: "AIC" or "BIC"
trace: Logical; whether to print progress
steps: Maximum number of steps to take
...: Additional arguments to pass to step()

Value

A tidylearn_model object of class tidylearn_linear wrapping the selected lm model. Access the underlying model via $fit and the selected formula via $spec$formula. $data is the data passed in.

Details

Every candidate model is fitted on the same rows: those with no missing value in any variable of formula. When that leaves out rows a candidate could otherwise use, a message gives the count, and the fit records the rows in its na.action, as lm() records the rows it drops itself.

Forward and both-direction selection start from a model holding the intercept, or none if formula removes it with - 1, and any offset() terms in formula. Neither is ever dropped.

Examples

# \donttest{
model <- tl_step_selection(mtcars, mpg ~ ., direction = "backward")
summary(model)
#> tidylearn Model
#> ===============
#> Paradigm: supervised 
#> Method: linear 
#> Task: Regression 
#> Formula: mpg ~ wt + qsec + am 
#> 
#> Training observations: 32 
#> 
#> Training Performance:
#> # A tibble: 3 × 2
#>   metric value
#>   <chr>  <dbl>
#> 1 rmse   2.30 
#> 2 mae    1.93 
#> 3 rsq    0.850
# }