forcats fct_drop() in R: Drop Unused Factor Levels
The fct_drop() function in forcats drops unused levels from a factor. It removes any level that has zero observations, so the level table matches the data that is actually present.
fct_drop(f) # drop every unused level fct_drop(df$grp) # drop on a data frame column fct_drop(f, only = "b") # drop only the level "b" fct_drop(f, only = c("b", "c")) # drop a chosen set fct_drop(f[f != "b"]) # drop after subsetting a factor levels(fct_drop(f)) # inspect the kept levels
Need explanation? Read on for examples and pitfalls.
What fct_drop() does
fct_drop() removes factor levels that no observation uses. A factor stores its categories as a fixed table of levels. That table does not shrink on its own. When you subset a vector or filter a data frame, rows disappear but every original level stays behind, even if nothing in the data points to it anymore.
Those leftover levels are called unused or orphan levels. They are harmless to the stored values but they leak into output. An unused level still shows up as an empty bar in a chart, a zero-count row in table(), and an extra coefficient slot in a model. fct_drop() is the forcats tool that clears them.
By default the function drops every unused level in one call. The optional only argument narrows that down to a named set, which is useful when you want to keep a deliberate placeholder level while removing the rest.
fct_drop() is the step that re-syncs the table to the codes, which is why it almost always follows a filter or a subset.fct_drop() syntax
fct_drop() takes the factor to clean plus an optional only argument. The full signature is short:
fct_drop(f, only)fis a factor. A character vector is accepted and coerced to a factor first.onlyis an optional character vector. When supplied, only the levels you name are considered for dropping, and any of those that are unused get removed. Levels not named are always kept, used or not.
When only is omitted, fct_drop() drops all unused levels. The code below builds a factor whose level table is wider than its data.
The level L exists in the table but no observation uses it. That is the unused level the examples below remove.
fct_drop() examples
Each example removes one or more unused levels and prints the kept set. fct_drop() returns a new factor, so assign the result to keep it.
1. Drop every unused level
Pass the factor straight to fct_drop(). Every level with zero observations is removed.
The orphan level L is gone and the data values in sizes are untouched. The level table now lists only categories that the data actually contains.
2. Drop unused levels after subsetting
The most common real use is cleaning a factor column after filtering rows. Subsetting keeps every original level, so call fct_drop() to catch up.
After the filter, no row holds grade C, yet the level survived until fct_drop() removed it. Assigning the result back to sub$grade updates the column in place.
3. Drop only a chosen level
The only argument restricts which unused levels are eligible. Here two levels are unused but just one is named.
Both c and d are unused, but only = "c" limits the drop to c. The level d is kept as a deliberate placeholder.
only to keep a known-future level. If a category has no rows yet but you expect it soon, naming the other orphans in only clears the clutter while preserving the placeholder. This keeps chart axes and factor tables stable across data refreshes.4. fct_drop() is a safe no-op when all levels are used
When every level already has data, fct_drop() returns the factor unchanged.
This makes the call safe to run defensively in a pipeline. It never errors and never removes a level that the data still needs.
fct_drop() vs droplevels()
fct_drop() cleans one factor; base R droplevels() also cleans a whole data frame. The two solve the same problem from different angles. droplevels() is built into base R and accepts a data frame, dropping unused levels from every factor column at once.
The table below sums up when to reach for each.
| Function | Works on | Restricting argument | Use when |
|---|---|---|---|
fct_drop() |
a single factor | only keeps a named set |
you clean one factor and want forcats consistency |
droplevels() |
a factor or a data frame | exclude protects named levels |
you clean every factor column at once |
The decision rule is short. For one factor inside a forcats or dplyr pipeline, use fct_drop(). To sweep an entire data frame in one call, use droplevels().
fct_drop() is series.cat.remove_unused_categories(), which drops categories with no observations from a categorical column.Common pitfalls
Pitfall 1: expecting subsetting to drop levels. Indexing a factor with f[f != "b"] or filtering a data frame removes rows but keeps the full level table. The unused level lingers until you call fct_drop(), so a stray empty bar or zero-count row is almost always a missing drop step.
Pitfall 2: forgetting to assign the result. fct_drop() returns a new factor and never edits the original in place. Call it without assigning the output and the cleaned factor is discarded.
only never drops a level that is in use. The only argument restricts which levels are eligible for dropping, but fct_drop() still removes a named level only when it is unused. Naming a level that has observations does nothing, so only cannot delete real categories by mistake.Pitfall 3: confusing dropping with collapsing. fct_drop() removes empty levels only. To merge several small levels into one bucket, use fct_lump() or fct_collapse() instead.
Try it yourself
Try it: The ex_team factor was built with four levels but only three appear in the data. Drop the unused level and save the result to ex_clean.
Click to reveal solution
Explanation: fct_drop() scans the level table for entries with zero observations. The level Gold has no rows, so it is removed, while the three used levels and every data value stay intact.
Related forcats functions
After fct_drop(), these forcats functions cover the rest of factor-level work:
fct_expand(): add new empty levels to a factor, the inverse of dropping.fct_lump(): merge infrequent levels into a singleOthercategory.fct_collapse(): combine named levels into broader groups by hand.fct_recode(): rename level labels without changing their order.fct_relevel(): set the order of the levels that remain.
For the full argument reference, see the forcats documentation at forcats.tidyverse.org.
FAQ
What does fct_drop do in R?
fct_drop() removes unused levels from a factor. An unused level is one that exists in the factor's level table but has no observations pointing to it. You pass a factor and get back a new factor whose level table lists only categories present in the data. The stored values are not changed, only the lookup table is trimmed.
Why does my factor still show levels with no data?
Factors keep a fixed level table. When you subset a vector or filter a data frame, the rows are removed but the level table is left intact, so levels with zero rows remain. This is by design, since R cannot know whether the missing category is gone permanently or just absent from this slice. Call fct_drop() to remove those orphan levels once you are sure.
What is the difference between fct_drop and droplevels?
Both drop unused levels. droplevels() is base R and accepts either a single factor or a whole data frame, cleaning every factor column in one call. fct_drop() is from forcats, works on one factor, and offers the only argument to restrict which levels are eligible. Use fct_drop() inside a forcats pipeline and droplevels() to sweep a full data frame.
Can fct_drop keep some unused levels?
Yes. Pass the only argument a character vector naming the levels you want to consider. fct_drop() will drop only those named levels when they are unused and leave every other level alone. This is useful for keeping a deliberate placeholder category that has no rows yet.
Does fct_drop change my data values?
No. fct_drop() only edits the level table and remaps the integer codes to match the smaller table. Every observation keeps its original category, so counts, means, and other computed results are identical before and after. This makes it safe to run right before plotting or modeling.