forcats fct_lump_prop() in R: Lump Rare Factor Levels

The forcats fct_lump_prop() function lumps factor levels in R that appear below a proportion threshold into a single "Other" category, so the cutoff scales with your data instead of being a fixed count.

fct_lump_prop(x, prop = 0.1)                       # keep levels above 10% of rows
fct_lump_prop(x, prop = 0.05)                      # stricter 5% threshold
fct_lump_prop(x, prop = -0.1)                      # keep the rare tail instead
fct_lump_prop(x, prop = 0.1, w = wt)               # rank levels by a weighted share
fct_lump_prop(x, prop = 0.1, other_level = "Misc") # rename the catch-all bucket
fct_count(fct_lump_prop(x, prop = 0.1))            # tally the lumped result

What fct_lump_prop() does in one sentence

fct_lump_prop() collapses factor levels by their share of the data, not by a raw count. It comes from the forcats package in the tidyverse. You pass a factor and a proportion prop, and the function keeps every level that appears more often than prop of the time while merging the rest into one labelled bucket. Because the threshold is a fraction, the same prop = 0.05 rule means "below 5 percent" whether the factor has 100 rows or 100,000.

Syntax

fct_lump_prop() takes a factor and a proportion cutoff. The signature has just four arguments:

The arguments are:

• f: a factor, or any vector that can be coerced to one (character, numeric, or logical).
• prop: the proportion threshold, a number between -1 and 1. A positive prop keeps levels that appear more than prop of the time; a negative prop keeps levels that appear less than -prop of the time.
• w: an optional numeric weight vector, one value per observation. Each level's share is then its summed weight divided by the total weight, instead of a plain row proportion.
• other_level: the name of the catch-all level. Defaults to "Other".

fct_lump_prop() never drops rows. It only rewrites the levels attribute, so the result has the same length as the input and slots straight into a pipeline.

fct_lump_prop() examples

Example 1 lumps every religion below 10 percent of survey responses. The relig column of gss_cat, a dataset bundled with forcats, has 15 religion levels with a long tail.

Only three religions clear the 10 percent line; the remaining twelve levels collapse into a single Other level holding 1,990 rows.

Example 2 shows that prop judges share, not raw count. A level with the exact same number of rows can survive in a small factor and get lumped in a large one.

The "rare" level has 5 rows in both factors. In the small factor that is 33 percent and survives; in the big factor it is 2.4 percent and gets lumped.

Example 3 keeps the rare tail with a negative prop. A negative value inverts the rule: it keeps the uncommon levels and lumps the frequent ones.

With prop = -0.20, only fox appears in fewer than 20 percent of rows, so it stays named while dog and cat merge into Other. This is handy when the rare categories are the signal, such as unusual error codes or fraud flags.

Example 4 ranks levels by a weighted share and renames the bucket. Pass a numeric w so a level is judged by its summed weight rather than its row count.

By row share, cat and dog win. By spend share, fox and owl dominate even though each appears once, and the catch-all carries the friendlier Minor label.

fct_lump_prop() vs the other fct_lump_*() functions

fct_lump_prop() is one of four lumping variants, each naming a different keep rule. Pick the variant whose rule matches the decision you actually want to make.

Use fct_lump_prop() when the cutoff should scale with the data, fct_lump_n() when you need a predictable category count, and fct_lump_min() when a raw frequency floor matters more than a share.

Common pitfalls

prop is a fraction, not a percentage. Passing prop = 10 to mean "10 percent" sets the threshold to 1,000 percent. No level can clear that, so every level lumps into one Other bucket.

Write prop = 0.10 for a 10 percent cutoff. Values of prop always live between -1 and 1.

Here cat is 30 percent of the data, a meaningful group, yet it merges into Other because it fell short of the 50 percent line.

Try it yourself

Related forcats functions

These forcats functions pair naturally with fct_lump_prop() for level management.

• fct_lump_n(): keep a fixed number of levels instead of a proportion.
• fct_lump_min(): keep levels by a raw frequency floor.
• fct_lump(): the umbrella lumping function that switches on n, prop, or a heuristic.
• fct_infreq(): order levels by frequency, the natural follow-up after lumping.
• Categorical Data in R: the full guide to working with factors.

See the forcats reference for the official documentation.

FAQ

What does fct_lump_prop() do in R?

fct_lump_prop() collapses the infrequent levels of a factor into a single "Other" category based on a proportion threshold. You give it a factor and a prop value, and it keeps every level whose share of the rows exceeds prop, merging the rest. Because the cutoff is a fraction of the data rather than a fixed count, the same rule keeps working as the dataset grows or shrinks. It is part of the forcats package in the tidyverse.

What is the difference between fct_lump_prop() and fct_lump_n()?

They answer different questions. fct_lump_prop() keeps every level above a proportion share, so the number of surviving levels depends on how the data is distributed. fct_lump_n() keeps a fixed count of the most common levels, so you always get at most n named levels plus "Other". Use fct_lump_prop() when the rule is "drop anything under 5 percent" and fct_lump_n() when a chart or model needs a known, predictable number of categories.

How does a negative prop work in fct_lump_prop()?

A negative prop inverts the keep rule. Instead of keeping the common levels, fct_lump_prop(x, prop = -0.1) keeps the levels that appear in fewer than 10 percent of rows and lumps the frequent ones into "Other". This is useful when the rare categories carry the signal you care about, such as uncommon survey answers, rare diagnoses, or anomalous error codes that you want to isolate from the bulk of the data.

Why is fct_lump_prop() putting everything into "Other"?

Almost always because prop is too high. Remember prop is a fraction between 0 and 1, so passing prop = 10 instead of prop = 0.10 sets an impossible 1,000 percent threshold and lumps every level. Even a valid value like prop = 0.5 lumps any level without an outright majority. Check the level shares first with fct_count(f, prop = TRUE), then pick a prop below the share you want to preserve.

How do I rename the "Other" level from fct_lump_prop()?

Pass the other_level argument with your preferred label, for example fct_lump_prop(x, prop = 0.1, other_level = "Smaller groups"). The catch-all level then carries that name instead of the default "Other". This is worth doing whenever the factor appears on a chart axis or in a report, where a generic "Other" reads poorly. The other_level argument behaves identically across fct_lump_prop(), fct_lump_n(), and fct_lump_min().