---
title: "Wrangling Categories with forcats"
---
# Wrangling Categories with forcats {#forcats}
```{r}
library(tidyverse)
knitr::opts_chunk$set(echo = TRUE, warning = FALSE, message = FALSE)
```
## Factors = Categories with Order
If you have ever made a bar chart in R and the bars came out in alphabetical order instead of something more useful — that is exactly the problem factors are designed to solve.
A **factor** is just R's fancy word for a categorical variable: things like survey responses ("Agree", "Neutral", "Disagree"), product categories, regions, departments, or customer segments. Under the hood, it is a text variable with a defined set of allowed values (called **levels**) in a specific order.
Why should you care about the order? One reason: **ggplot2 uses factor order to arrange your charts**. Alphabetical bars are the data visualization equivalent of an unsorted resume --- technically complete, but nobody wants to read it.
The **forcats** package (part of the tidyverse, loaded automatically) gives you simple tools to fix that. Let's see the problem first.
## The Problem: Alphabetical Everything
```{r}
# A character vector
satisfaction <- c("Low", "Medium", "High", "Low", "High")
# When you make it a factor, R defaults to alphabetical order
factor(satisfaction)
```
The levels are "High", "Low", "Medium" --- alphabetical, not logical. You want "Low", "Medium", "High." That is the entire reason this chapter exists.
```{r}
# Fix it by specifying levels
factor(satisfaction, levels = c("Low", "Medium", "High"))
```
This manual approach works for three levels. When you have real data with dozens of categories, `forcats` makes it manageable.
::: {.callout-warning}
## AI Pitfall: AI reorders factor levels and breaks downstream code
Factor levels matter beyond the cosmetics of chart order. Statistical models use the *first* level as the reference category — so changing the order of levels changes which group every other coefficient is compared against. AI assistants happily reorder factor levels for visualization purposes without flagging that this could change a downstream regression's interpretation.
A concrete failure mode: you fit a logistic regression where `region = "Northeast"` is the reference category, get a coefficient table interpreted relative to Northeast, and then ask AI to "make the bar chart show regions in order of average revenue." The AI calls `fct_reorder(region, revenue)`. Your chart looks great. If you re-fit the model on the reordered factor, every coefficient now compares to a different reference and the interpretation in your write-up no longer matches the output.
The discipline: keep one canonical factor ordering for modeling and apply visualization-only reorderings inside the chart code (`aes(x = fct_reorder(region, revenue))`) rather than mutating the underlying data. That way models and charts each get the ordering they need without stepping on each other.
:::
## The Dataset: General Social Survey
We will use `gss_cat`, a dataset built into forcats with survey data from the General Social Survey. It has several factor columns --- perfect guinea pig.
```{r}
glimpse(gss_cat)
```
## `fct_reorder()` --- THE Most Important Function
If you learn one thing from this chapter, make it this one. `fct_reorder()` reorders a factor by the values of another variable. In plain English: **it sorts your bar chart bars so they actually make sense.**
Think of it this way: if your bar chart were a PowerPoint slide, `fct_reorder()` is the difference between your manager saying "what am I looking at?" and "great insight." Here is the before-and-after.
Without reordering:
```{r}
gss_cat %>%
mutate(marital = fct_infreq(marital)) %>%
ggplot(aes(x = marital)) +
geom_bar(fill = "steelblue") +
labs(title = "Marital Status (sorted by frequency)", x = NULL, y = "Count") +
theme_minimal()
```
Combine it with `fct_rev()` for a horizontal bar chart that reads top-to-bottom (the way humans naturally read):
```{r}
gss_cat %>%
mutate(marital = fct_rev(fct_infreq(marital))) %>%
ggplot(aes(y = marital)) +
geom_bar(fill = "darkorange") +
labs(title = "Marital Status (highest count at top)", x = "Count", y = NULL) +
theme_minimal()
```
## `fct_relevel()` --- Manual Ordering
Sometimes you need a specific order that is not alphabetical or by frequency. Maybe "Never married" is your reference group in a regression, or you need income brackets from low to high. `fct_relevel()` lets you move specific levels wherever you want --- like rearranging slides in a deck.
```{r}
# Default level order
levels(gss_cat$marital)
```
```{r}
# Move "Never married" to the front
gss_cat %>%
mutate(marital = fct_relevel(marital, "Never married")) %>%
pull(marital) %>%
levels()
```
```{r}
# Move "Widowed" to the end
gss_cat %>%
mutate(marital = fct_relevel(marital, "Widowed", after = Inf)) %>%
pull(marital) %>%
levels()
```
And the chart payoff:
```{r}
gss_cat %>%
mutate(marital = fct_relevel(marital,
"Never married", "Married", "Separated",
"Divorced", "Widowed", "No answer")) %>%
ggplot(aes(x = marital)) +
geom_bar(fill = "steelblue") +
labs(title = "Marital Status (life-stage order)", x = NULL, y = "Count") +
theme_minimal()
```
Now the bars tell a life-stage story: never married, married, separated, divorced, widowed. Much more intuitive than whatever "alphabetical" was trying to communicate.
## `fct_recode()` --- Rename Your Categories
Survey data often arrives with labels that look like someone's internal database codes. `fct_recode()` lets you rename levels with a simple `"new name" = "old name"` syntax. Think of it as "Find and Replace" for category labels.
```{r}
gss_cat %>%
mutate(partyid = fct_recode(partyid,
"Republican (strong)" = "Strong republican",
"Republican (weak)" = "Not str republican",
"Independent (near rep)" = "Ind,near rep",
"Independent (near dem)" = "Ind,near dem",
"Democrat (weak)" = "Not str democrat",
"Democrat (strong)" = "Strong democrat"
)) %>%
count(partyid)
```
Also great when your data has labels like "1", "2", "3" that should really say "Strongly Disagree", "Neutral", "Strongly Agree." Your stakeholders should never have to decode numbers on a chart.
## `fct_collapse()` --- Merge Categories Together
Sometimes you have too many categories and need to consolidate. `fct_collapse()` merges multiple levels into new groups. Think of it as the "department reorganization" of your data --- fewer groups, cleaner org chart.
```{r}
gss_cat %>%
mutate(partyid = fct_collapse(partyid,
"Republican" = c("Strong republican", "Not str republican"),
"Independent" = c("Ind,near rep", "Independent", "Ind,near dem"),
"Democrat" = c("Not str democrat", "Strong democrat"),
"Other" = c("No answer", "Don't know", "Other party")
)) %>%
count(partyid)
```
And the chart payoff --- a clean, executive-summary-ready political affiliation plot:
```{r}
gss_cat %>%
mutate(party = fct_collapse(partyid,
"Republican" = c("Strong republican", "Not str republican"),
"Independent" = c("Ind,near rep", "Independent", "Ind,near dem"),
"Democrat" = c("Not str democrat", "Strong democrat"),
"Other" = c("No answer", "Don't know", "Other party")
)) %>%
mutate(party = fct_infreq(party)) %>%
ggplot(aes(x = party, fill = party)) +
geom_bar(show.legend = FALSE) +
scale_fill_manual(values = c(
"Democrat" = "blue", "Independent" = "gray50",
"Republican" = "red", "Other" = "gray80"
)) +
labs(title = "Political Affiliation (simplified)", x = NULL, y = "Count") +
theme_minimal()
```
Went from 10 confusing categories to 4 clean ones. That is a chart your VP can actually read from across the conference table.
## `fct_lump()` --- "Show Me the Top N, Lump the Rest"
This is the function you reach for when you have a variable with 30+ categories. Nobody wants a bar chart with 30 bars --- that is not a chart, that is a picket fence. `fct_lump_n()` keeps the top N categories and lumps everything else into "Other."
```{r}
# Keep the 4 most common religions, lump the rest
gss_cat %>%
mutate(relig = fct_lump_n(relig, n = 4)) %>%
count(relig, sort = TRUE)
```
The chart version --- before and after lumping:
```{r}
gss_cat %>%
count(denom) %>%
ggplot(aes(x = n, y = denom)) +
geom_point() +
labs(title = "All Denominations (information overload)", x = "Count", y = NULL) +
theme_minimal()
```
```{r}
gss_cat %>%
mutate(denom = fct_lump_n(denom, n = 8)) %>%
count(denom) %>%
mutate(denom = fct_reorder(denom, n)) %>%
ggplot(aes(x = n, y = denom)) +
geom_point(size = 3, color = "steelblue") +
labs(title = "Top 8 Denominations (rest lumped into Other)",
x = "Count", y = NULL) +
theme_minimal()
```
There are a few variants depending on how you want to decide what gets lumped:
- `fct_lump_n(x, n = 5)` --- keep top 5
- `fct_lump_prop(x, prop = 0.10)` --- keep categories with at least 10% of the data
- `fct_lump_min(x, min = 100)` --- keep categories with at least 100 observations
Pick whichever matches your question.
## Bonus: Making mtcars Charts Better with Factors
The `mtcars` dataset has variables like `cyl` (cylinders) and `am` (transmission) coded as numbers. They are really categories wearing a numerical disguise. Converting them to factors with meaningful labels instantly makes your charts presentation-ready.
```{r}
mtcars %>%
mutate(
cyl = factor(cyl),
transmission = factor(am, levels = c(0, 1), labels = c("Automatic", "Manual"))
) %>%
group_by(cyl, transmission) %>%
summarize(mean_mpg = mean(mpg), .groups = "drop") %>%
ggplot(aes(x = cyl, y = mean_mpg, fill = transmission)) +
geom_col(position = "dodge") +
scale_fill_brewer(palette = "Set2") +
labs(title = "Mean MPG by Cylinders and Transmission",
x = "Number of Cylinders", y = "Mean MPG", fill = "Transmission") +
theme_minimal()
```
Without factors, those bars would just say "0" and "1" for transmission. With factors, they say "Automatic" and "Manual." Small change, enormous difference in whether your audience has any idea what they are looking at.
## `fct_reorder2()` --- Aligning Legends with Lines
Quick bonus for line charts. When you have multiple lines, the legend should match the visual order of the lines at the right edge of the plot. `fct_reorder2()` handles this automatically --- a small detail that makes your charts look polished.
```{r}
gss_cat %>%
filter(!is.na(age)) %>%
mutate(marital = fct_lump_n(marital, n = 4)) %>%
count(age, marital) %>%
group_by(age) %>%
mutate(prop = n / sum(n)) %>%
ungroup() %>%
ggplot(aes(x = age, y = prop, color = fct_reorder2(marital, age, prop))) +
geom_line(linewidth = 1) +
scale_color_brewer(palette = "Set1") +
labs(title = "Marital Status by Age",
x = "Age", y = "Proportion", color = "Status") +
theme_minimal()
```
## Quick Reference
Your factor toolkit at a glance:
| What you want to do | Function | When to use it |
|:---------------------|:---------|:---------------|
| Sort bars by a value | `fct_reorder()` | Almost every bar/dot chart |
| Sort bars by count | `fct_infreq()` | "Show me what's most common" |
| Reverse the order | `fct_rev()` | Horizontal bar charts (top to bottom) |
| Set a custom order | `fct_relevel()` | Life stages, survey scales, custom sorts |
| Rename categories | `fct_recode()` | Cryptic labels need human-readable names |
| Merge categories | `fct_collapse()` | Too many groups, need to simplify |
| Keep top N, lump rest | `fct_lump_n()` | 30 categories? Nobody wants that chart |
The single most impactful habit from this chapter: **always use `fct_reorder()` or `fct_infreq()` in your ggplot2 charts.** Alphabetical ordering is almost never the most informative ordering. A well-ordered factor is the difference between a confusing chart and one that gets you promoted. Well, maybe not promoted --- but at least not asked to redo it.
