Skip to contents

Apply step(), taking into account missing values

Source: R/step_with_na.R
step_with_na.Rd

When your data contains missing values, concerned observations are removed from a model. However, then at a later stage, you try to apply a descending stepwise approach to reduce your model by minimization of AIC, you may encounter an error because the number of rows has changed.

Usage

step_with_na(model, ...)

# Default S3 method
step_with_na(model, ..., full_data = eval(model$call$data))

# S3 method for class 'svyglm'
step_with_na(model, ..., design)

Arguments

model

A model object.

...

Additional parameters passed to stats::step().

full_data

Full data frame used for the model, including missing data.

design

Survey design previously passed to survey::svyglm().

Value

The stepwise-selected model.

Details

step_with_na() applies the following strategy:

  • recomputes the models using only complete cases;

  • applies stats::step();

  • recomputes the reduced model using the full original dataset.

step_with_na() has been tested with stats::lm(), stats::glm(), nnet::multinom() and survey::svyglm(). It may be working with other types of models, but with no warranty.

In some cases, it may be necessary to provide the full dataset initially used to estimate the model.

step_with_na() may not work inside other functions. In that case, you may try to pass full_data to the function.

Examples

set.seed(42)
d <- titanic |>
  dplyr::mutate(
    Group = sample(
      c("a", "b", NA),
      dplyr::n(),
      replace = TRUE
    )
  )
mod <- glm(as.factor(Survived) ~ ., data = d, family = binomial())
# step(mod) should produce an error
mod2 <- step_with_na(mod)
#> Error in eval(model$call$data): object 'd' not found
mod2
#> Error: object 'mod2' not found

# \donttest{
## WITH SURVEY ---------------------------------------

library(survey)
#> Loading required package: grid
#> Loading required package: Matrix
#> Loading required package: survival
#> 
#> Attaching package: ‘survey’
#> The following object is masked from ‘package:graphics’:
#> 
#>     dotchart
ds <- d |>
  dplyr::mutate(Survived = as.factor(Survived)) |>
  srvyr::as_survey()
mods <- survey::svyglm(
  Survived ~ Class + Group + Sex,
  design = ds,
  family = quasibinomial()
)
mod2s <- step_with_na(mods, design = ds)
#> Start:  AIC=1471.56
#> Survived ~ Class + Group + Sex
#> 
#>         Df Deviance    AIC
#> - Group  1   1462.6 1469.9
#> <none>       1462.2 1471.6
#> - Class  3   1527.9 1530.3
#> - Sex    1   1712.6 1716.5
#> 
#> Step:  AIC=1469.94
#> Survived ~ Class + Sex
#> 
#>         Df Deviance    AIC
#> <none>       1462.6 1469.9
#> - Class  3   1528.4 1528.8
#> - Sex    1   1713.1 1714.8
mod2s
#> Independent Sampling design (with replacement)
#> Called via srvyr
#> Sampling variables:
#>   - ids: `1` 
#> 
#> Call:  svyglm(formula = Survived ~ Class + Sex, design = design, family = quasibinomial())
#> 
#> Coefficients:
#> (Intercept)     Class2nd     Class3rd    ClassCrew      SexMale  
#>      2.0682      -0.9526      -1.6582      -0.8808      -2.4213  
#> 
#> Degrees of Freedom: 2200 Total (i.e. Null);  2196 Residual
#> Null Deviance:	    2769 
#> Residual Deviance: 2229 	AIC: NA
# }