Driver importance

Ranks how much each predictor drives an outcome such as the NPS rating, as importance scores summing to 100. Relative weights analysis is the default; a random forest and a plain correlation are available as alternatives.

Usage

calc_importance(
  data = NULL,
  outcome,
  predictors,
  method = c("rwa", "forest", "correlation"),
  recode = TRUE,
  likert_levels = c("Very bad", "Bad", "Ok", "Good", "Very good")
)

Arguments

data

A data frame.

outcome

The outcome column (unquoted), e.g. nps_value.

predictors

The predictor columns, using tidyselect (e.g. starts_with("ratings_")).

method

How to measure importance. "rwa" (default) is relative weights analysis via rwa::rwa(); "forest" is random-forest permutation importance via randomForest::randomForest(); "correlation" is the absolute correlation with the outcome. See Details.

recode

If TRUE (default), character predictor columns are mapped to 1-5 with recode_likert(); numeric columns are left as-is.

likert_levels

Scale wordings passed to recode_likert() when recoding.

Value

A tibble with feature and importance, the latter rescaled to sum to 100 whichever method produced it, strongest driver first.

Details

All three methods answer the same question and are scaled the same way, so each feature's importance reads as "this feature accounts for X% of what drives the outcome" and any of them can be fed to plot_ipm(). They differ in what they can see:

• "rwa", relative weights analysis (Johnson's epsilon), shares out the linear model's explained variance between correlated predictors, which ordinary regression coefficients handle poorly. This is the default because rating batteries are almost always heavily correlated.

• "forest" grows a random forest and measures how much prediction error rises when each predictor is permuted, so it picks up non-linear effects and interactions that relative weights cannot. It is the useful cross-check: if a feature ranks high here and low under "rwa", its effect is not linear. Being a random method, it moves a little between runs; call set.seed() first for a figure you intend to publish.

• "correlation" is the crude one, ignoring the other predictors entirely, so correlated features double-count. Its merit is that it needs no suggested package and cannot fail to converge, which makes it a reasonable sanity check on the other two.

Only complete cases are used. Worded rating columns are recoded to 1-5 automatically (recode = TRUE). Negative importances (which permutation importance can produce for a predictor that is pure noise) are floored at zero before rescaling. Most users call ipm_model(), which pairs this with performance for the importance/performance matrix.

See also

ipm_model(), plot_ipm().

Other modelling: calc_nps(), ipm_model()

Examples

calc_importance(podracing_survey, nps_value, starts_with("ratings_"))
#> Parsing `nps_value` as a non-binary variable.
#> Applying multiple regression to calculate relative weights...
#> # A tibble: 6 × 2
#>   feature            importance
#>   <chr>                   <dbl>
#> 1 ratings_value           24.9 
#> 2 ratings_atmosphere      23.8 
#> 3 ratings_speed           18.0 
#> 4 ratings_safety          14.1 
#> 5 ratings_venue           12.2 
#> 6 ratings_commentary       6.99

# a cross-check that can see non-linear effects
set.seed(1)
calc_importance(podracing_survey, nps_value, starts_with("ratings_"),
                method = "forest")
#> # A tibble: 6 × 2
#>   feature            importance
#>   <chr>                   <dbl>
#> 1 ratings_value           27.5 
#> 2 ratings_atmosphere      24.3 
#> 3 ratings_speed           16.5 
#> 4 ratings_safety          15.6 
#> 5 ratings_venue           12.5 
#> 6 ratings_commentary       3.58