recipes step_zv() in R: Remove Zero-Variance Predictors

The recipes step_zv() function in R removes zero-variance predictors, columns that contain only a single unique value, so they cannot destabilize a model or a downstream preprocessing step.

step_zv(rec, all_predictors())            # drop every constant column
step_zv(rec, x, y)                        # screen named columns only
step_zv(rec, all_numeric_predictors())    # numeric predictors only
step_dummy(x) |> step_zv(all_predictors()) # after dummy encoding
prep(rec); bake(rec, new_data = NULL)     # train, then apply
tidy(prep(rec), number = 1)               # list removed columns

What step_zv() does

step_zv() deletes any column that has zero variance, meaning every row holds the same value. A constant column carries no information a model can learn from. Worse, it can trigger errors: many fitting routines fail when a predictor has no spread, and operations like scaling divide by a standard deviation of zero.

step_zv() screens the columns you select while the recipe is prepared. During prep() it counts the distinct values in each selected column. Any column with exactly one distinct value is marked for removal, and bake() then drops it from every data frame the recipe touches, training and new data alike.

step_zv() syntax and arguments

step_zv() takes a recipe and a column selection, and almost never needs anything else. It is a selection step, so you can name columns directly or use tidyselect helpers such as all_predictors() or all_numeric_predictors().

Most workflows pass only the recipe and a selector. The removals slot is filled automatically at prep() time with the names of the constant columns. The group argument is for the rare case where a column should be considered constant only within levels of another variable. The full reference lives in the recipes documentation.

step_zv() examples

The pattern is always prep then bake. prep() finds the constant columns; bake() removes them. The examples below start from a small dataset where one predictor never changes.

The region column held "east" in every row, so step_zv() dropped it. Only age and income survive. To confirm exactly which columns were removed, call tidy() on the prepared recipe and pass the step number.

The terms column lists every removed predictor. An empty tibble means no column was constant. This is the quickest audit when a recipe silently changes the number of columns it returns.

step_zv() is most valuable right after step_dummy(). Dummy encoding can create indicator columns that are all zeros when a factor level never appears in the data, and those columns have zero variance.

The factor grade declares level B, but no row uses it, so step_dummy() produced an all-zero grade_B column. step_zv() then removed that dead column before it could reach the model. You can also restrict screening to named columns instead of every predictor.

Only const_a and varies were screened, so const_a was dropped while const_b stayed, even though it is also constant. Selection scopes the check; an unselected constant column passes through untouched.

step_zv() vs step_nzv() vs step_corr()

Three recipes steps remove unhelpful columns, but each targets a different defect. Choosing the wrong one either keeps noise you wanted gone or deletes columns you meant to keep.

Use step_zv() for strictly constant columns: it is cheap, safe, and rarely surprises you. Reach for step_nzv() when a column is overwhelmingly one value but not perfectly constant, such as 999 zeros and a single one. Use step_corr() to thin out redundant numeric predictors. Many pipelines run step_zv() first as a guaranteed-safe filter, then add the others.

Common pitfalls

Three mistakes account for most step_zv() confusion.

1. Expecting it to catch near-constant columns. step_zv() removes a column only when every value is identical. A column that is 99 percent one value still has variance and survives. Use step_nzv() for that case.
2. Running it before step_dummy(). All-zero dummy columns do not exist until step_dummy() runs. Place step_zv() after the dummy step so it can see and remove them.
3. Assuming the removed set is fixed. Zero variance is judged from the data given to prep(). Under resampling, a column constant in one fold may vary in another, so the removal list can differ across resamples.

Try it yourself

Related recipes functions

step_zv() is one filter in a larger column-cleaning toolkit. These steps frequently appear together in the same recipe:

• step_nzv() removes near-zero-variance columns that are almost, but not exactly, constant.
• step_corr() drops predictors that are highly correlated with another column.
• step_lincomb() removes columns that form exact linear combinations of others.
• step_rm() deletes specific columns you name by hand.
• step_dummy() converts factors into indicator columns, often paired before step_zv().

FAQ

What does step_zv() do in R?

step_zv() is a recipes step that removes zero-variance predictors, columns where every observation has the same value. When the recipe is prepared with prep(), it counts the distinct values in each selected column and marks any column with a single unique value for removal. bake() then drops those columns from training and new data alike. It is a safe, fast filter commonly used to clean a dataset before modeling.

What is the difference between step_zv() and step_nzv()?

step_zv() removes only strictly constant columns, those with exactly one unique value. step_nzv() is more aggressive: it removes near-zero-variance columns that are overwhelmingly one value but not perfectly constant, judged by a frequency ratio and the percentage of unique values. A column with 999 identical entries and one different entry survives step_zv() but is caught by step_nzv(). Use step_zv() when you only want to delete truly dead columns.

Should step_zv() come before or after step_dummy()?

After. step_dummy() builds indicator columns from a factor, and if a level never appears in the data, the corresponding dummy column is all zeros and therefore zero variance. If step_zv() runs first, those dead columns do not exist yet and slip through. Placing step_zv() after step_dummy() lets it see and remove the all-zero indicators before they reach the model.

How do I see which columns step_zv() removed?

Call tidy() on the prepared recipe and pass the step number, for example tidy(prepped_recipe, number = 1). It returns a tibble whose terms column lists every predictor that was removed for zero variance. An empty tibble means no column was constant. This is the fastest way to audit a recipe when the baked output unexpectedly has fewer columns than the input.