R语言中含分类变量的PCA分析：是否有对应工具包？

Great question! Standard PCA is built for numerical data, so when your dataset includes categorical variables (like factors, group-label strings), you need specialized approaches or packages to get meaningful results. Let’s break down the most practical options:

1. Multiple Correspondence Analysis (MCA) – For All-Categorical Data #

If your data is entirely categorical, MCA is the go-to method (think of it as PCA tailored for categorical variables). Two R packages make this straightforward:

FactoMineR (Core Analysis) #

This package has a dedicated MCA() function that handles categorical variables natively. Here’s a quick example:

# Install and load the package
install.packages("FactoMineR")
library(FactoMineR)

# Load sample categorical data (the built-in tea dataset works perfectly)
data(tea)

# Run MCA on selected categorical columns, excluding supplementary variables
mca_result <- MCA(tea, quanti.sup = 19, quali.sup = c(20:23))

# View a detailed summary of results
summary(mca_result)

FactoExtra (Visualization) #

Pair FactoMineR with FactoExtra to create clean, interpretable plots of your MCA results:

install.packages("FactoExtra")
library(FactoExtra)

# Plot MCA individuals and variables with repelled labels for readability
fviz_mca_biplot(mca_result, repel = TRUE, ggtheme = theme_minimal())

2. Dummy Variable PCA – For Mixed Numerical/Categorical Data #

If you have a mix of numerical and categorical variables, one common approach is to convert categorical variables into dummy variables (binary 0/1 columns) then run standard PCA. Here’s how:

# Load sample mixed data (iris has numerical measurements + a categorical species column)
data(iris)
iris$Species <- as.factor(iris$Species)

# Convert categorical variables to dummy variables (exclude intercept with `-1`)
dummy_data <- model.matrix(~ . -1, data = iris)

# Run standard PCA (remember to scale numerical variables for balanced influence!)
pca_result <- prcomp(dummy_data, scale. = TRUE)

# View how much variance each component explains
summary(pca_result)

# Plot the PCA biplot to visualize variable and sample relationships
biplot(pca_result)

Note: Be cautious with high-cardinality categorical variables (many levels) – this can lead to a huge number of dummy columns and "curse of dimensionality" issues.

3. Factor Analysis of Mixed Data (FAMD) – For Mixed Data (Better Than Dummies!) #

FAMD is designed specifically for datasets with both numerical and categorical variables, balancing the influence of each type so neither dominates the analysis. It’s implemented in FactoMineR too:

# Run FAMD on mixed data (tea has both categorical and numerical variables)
famd_result <- FAMD(tea, ncp = 5, graph = FALSE)

# Visualize results with a biplot
fviz_famd_biplot(famd_result, repel = TRUE, ggtheme = theme_minimal())

4. ade4 Package – Alternative for Mixed Data #

The ade4 package offers dudi.mix(), another flexible tool for analyzing mixed numerical/categorical data:

install.packages("ade4")
library(ade4)

# Split iris data into numerical and categorical parts
num_vars <- iris[, 1:4]
cat_vars <- iris[, 5, drop = FALSE]

# Run mixed data analysis
mix_result <- dudi.mix(num_vars, cat_vars, scannf = FALSE, nf = 2)

# Plot the results
scatter(mix_result)

Quick Recap #

• All categorical data: Use MCA (FactoMineR + FactoExtra)
• Mixed data: Prefer FAMD (balanced, avoids dummy variable bloat) or dummy variable PCA (simple but risky for high-cardinality vars)
• Alternative: ade4’s dudi.mix() for more customizable mixed-data analysis

内容的提问来源于stack exchange，提问作者udAyKumArVermA