forcats Exercises in R: 28 Real Factor Practice Problems

Explanation: factor() builds a factor with levels sorted alphanumerically by default, so the levels come out as 4 6 8 regardless of the order rows appear in the data. The underlying integer codes (1, 2, 3) point at those levels, which is why categorical regressors take less memory than character vectors. Use as_factor() from forcats when you want the levels to follow first appearance instead of alphanumeric sort.

Explanation: Setting ordered = TRUE upgrades a factor from nominal to ordinal, which is what statistical models need when ranks matter (think polynomial contrasts in lm()). The < separators in the printed levels confirm the ordering took. The diamonds version of clarity is already ordered, but rebuilding it manually gives you the recipe for any custom ordinal scale (Likert, risk rating, severity).

Explanation: fct_count() is the forcats equivalent of table() returning a tibble instead of a named integer vector, which plays nicely with the rest of the tidyverse. The sort = TRUE flag reorders rows by count descending. Add prop = TRUE to get a third column with relative shares, useful for class-imbalance audits before training a classifier.

Explanation: When you supply an explicit levels vector to factor(), any value not in that list is silently coerced to NA. That makes factor() double as a validation step: a quick sum(is.na(ex_1_4)) after construction flags how many rows fell off the allowed list. The alternative forcats::fct_match() would throw an error on the typo, which is preferable in a pipeline you want to fail loudly.

Explanation: fct_inorder() reorders levels by the position of their first occurrence, which matters when the natural row order encodes information (chronological events, ranked respondents, geographic adjacency). Useful right before plotting so bar order or line color order tracks the data instead of the alphabet. The sibling fct_inseq() does the same trick for numeric-looking factors.

Explanation: fct_infreq() reorders levels by descending frequency, which is the single most useful default for categorical bar charts: a Pareto-style ranking emerges automatically. Pair with fct_rev() when you flip to a horizontal bar chart and want the tallest bar on top. For grouped frequencies (per facet) reach for fct_reorder() instead.

Explanation: In treatment contrasts (R's default), the first factor level becomes the intercept and all other coefficients are offsets from it. Choosing a clinically meaningful reference, like the standard-of-care arm in a trial, makes coefficients directly interpretable. fct_relevel() only moves the named level; the rest stay in their original order. For position-based moves use the after argument: fct_relevel(x, "4", after = 2).

Explanation: fct_reorder() ranks levels by a numeric summary computed within each level, which is the cleanest way to align bar or boxplot order with the metric you are plotting. Pass any function, including custom ones: fct_reorder(x, y, function(z) quantile(z, 0.9)) ranks by the 90th percentile. The two-argument cousin fct_reorder2() is built for line plots where you want the legend order to match the rightmost endpoint.

Explanation: fct_rev() flips level order end-to-end, the standard trick when ggplot lays out a horizontal bar chart bottom-up and you want largest-on-top. fct_shift() rotates levels by n positions, which is handy for cyclical scales like weekdays where you want Monday to lead instead of Sunday. Both leave the underlying values intact; only the level vector permutes.

Explanation: fct_recode() takes new-name = old-name pairs (note the direction), updates the levels in place, and preserves the underlying integer codes so the data layout is unchanged. If a target name already exists in the factor, the levels merge automatically, which makes it a one-call tool for both rename and minor collapse. Misspelled old-names trigger a warning rather than silent miss, so misnames surface immediately.

Explanation: fct_collapse() is the bulk-rename tool: each new name maps to a vector of old names that get merged. It is the right pick when you want to reduce cardinality by domain logic (versus fct_lump_* which reduces by frequency). The order of the new names in the call becomes the new level order, which controls plotting and contrast coding downstream.

Explanation: Passing NULL as the replacement converts the named level to NA for every row that held it, while keeping the row count intact. Use this when you want a level temporarily ignored by downstream summaries (na.rm = TRUE will then exclude it) without slicing rows away. To also remove the now-empty level from the factor, chain a fct_drop() call afterward.

Explanation: The other_level argument is a catch-all for levels not named in the mapping. Passing NA routes the unhandled levels to missing, which is the safest behavior when you cannot interpret a value; passing a string like "Unknown" keeps them visible as an explicit bucket. This recipe scales: list every clean target and let other_level absorb the long tail of typos and rare codes.

Explanation: fct_lump_n() keeps the n most frequent levels and merges everything else into the other_level bucket (default name "Other"). It is the standard tool for taming high-cardinality categoricals before a bar chart or a tree-based model, where dozens of tiny levels dilute signal. Negative n flips the logic: n = -3 keeps the three rarest levels and lumps the rest.

Explanation: fct_lump_prop() works on relative share rather than absolute count, which is the right knob when sample sizes vary between cohorts. With 234 rows in mpg, a 10 percent threshold lumps anything below 24 rows: 2seater (5) and minivan (11) fall under and collapse into "Other". The proportion variant scales cleanly across small and large datasets without retuning a threshold.

Explanation: fct_lump_min() keeps any level meeting the minimum count and lumps the rest. Here every color in diamonds has at least 2808 rows, so no "Other" bucket is created and all seven colors survive unchanged. Push min to a higher value (say 5000) and J, then I, then H peel off into "Other". This is the variant to reach for when you have a hard analytical floor like "need at least 30 samples per cell".

Explanation: fct_other() lumps by name rather than by frequency. Use keep = for an allowlist of survivors or drop = for a denylist of levels to collapse into "Other". The two arguments are mutually exclusive. This is the right tool for stakeholder-driven cuts ("show me Apple, Google, Microsoft, and lump the rest") where the survivors are not necessarily the most frequent.

Explanation: Wrapping the factor in fct_infreq() inside aes() reorders bars by frequency without mutating the underlying data, which keeps the rest of the pipeline ignorant of the cosmetic ordering. This is the cleanest pattern for one-shot plots. For repeated use across multiple plots, mutate the column once with mutate(class = fct_infreq(class)) and keep aes() simple.

Explanation: fct_reorder() defaults to median as the summary, which is robust to outliers and matches the line drawn in a boxplot. Plotting categories in a meaningful order, rather than alphabetical, removes a cognitive step for the reader. Use .desc = TRUE to flip the direction, or chain with fct_rev() for the same effect.

Explanation: fct_reorder2() ranks levels by the y-value at the largest x within each level, which is exactly what the eye tracks on a line plot. Pair with geom_line(aes(color = ex_5_3)) and the legend ordering matches the rightmost line endpoints top-to-bottom. The standard fct_reorder() would use a single summary like median across all x and miss this ordering.

Explanation: ggplot draws the first factor level at the bottom of the y axis after coord_flip(), so a frequency-ordered factor produces an inverted Pareto. Composing fct_rev(fct_infreq(...)) flips it back to the canonical largest-on-top reading order. Composition order matters: fct_infreq(fct_rev(...)) would frequency-sort first then reverse, also valid but the inner-first reading is clearer.

Explanation: Subsetting rows never trims the level set on its own because factors store levels independently of values. The leftover empty levels show up as zero-count rows in table() and as empty bars in geom_bar(), which is rarely what you want. fct_drop() removes levels with zero observations; pass only = "x" to target a specific level, useful when you need to keep some empty levels intentionally.

Explanation: fct_expand() adds new levels without changing any observation. The added level has zero counts until rows are appended later. This matters when you want stable factor levels across multiple data batches (e.g. so a fitted model can score future data with a previously-unseen treatment arm), or to lock in the level ordering before plotting an empty category as a baseline.

Explanation: Hiding NA values can mask a serious data-quality issue; turning them into an explicit level forces the missingness to show up in every plot and table. fct_na_value_to_level() (formerly fct_explicit_na()) is the right step before passing a factor into a model that cannot accept NA, or before a stakeholder chart where missingness is part of the story.

Explanation: fct_unify() walks a list of factors, computes the union of their levels, and applies that union (in a stable order) to each factor in the list. The result safely stacks under bind_rows() or any column-wise combine because both factors now agree on their level set. Without this step, dplyr quietly coerces mismatched factors to character, losing the ordinal contract you may rely on later.

Explanation: Real survey cleanup chains three forcats verbs: collapse for the bucketing rule (industry-standard NPS thresholds), then explicit NA handling, then a final lumping pass. Building the pipeline left-to-right makes each step auditable in isolation, so a reviewer can swap the collapse cutoffs or relabel NoResponse without rewriting the chain. The same skeleton fits any Likert recoding workflow.

Explanation: Changing the reference level changes the intercept and the names of the dummy coefficients but not the model fit (residual SS and adjusted R-squared are identical). The published intercept now equals the mean weight for diet 3, and the three Diet1/Diet2/Diet4 coefficients are offsets from it, often the form a paper or regulator expects. Reach for relevel() from base R if you cannot add a tidyverse dependency.

Explanation: Combining slice_max() for the per-group top-N selection with fct_reorder() for plot-friendly factor ordering produces a clean ranked panel ready for a small-multiples bar chart. The .desc = TRUE flag puts the bestseller first in level order so legends and axes flow large-to-small. For a true rolling Other bucket across all brands, add a summarise(units = sum(units)) per quarter before slicing.