What Is FAMD? A Clear, Beginner-Friendly Guide

What Is FAMD? A Clear, Beginner-Friendly Guide

What FAMD stands for

FAMD = Factor Analysis of Mixed Data.

Purpose

FAMD is a dimension-reduction technique designed to handle datasets that contain both continuous (numeric) and categorical variables. It produces a low-dimensional representation that captures the main patterns and relationships across mixed-type features.

When to use it

• Your dataset mixes numeric and categorical variables.
• You want to visualize structure (clusters, gradients) in 2–3 dimensions.
• You need to reduce dimensionality before clustering or visualization while preserving contributions from both variable types.

How it works (overview)

• Numeric variables are centered and scaled.
• Categorical variables are converted to indicator (dummy) variables and weighted so each categorical variable contributes comparably to the analysis.
• A singular value decomposition (SVD) or equivalent eigen-decomposition is applied to the combined, weighted matrix to extract principal components (dimensions).
• The resulting components are interpreted similarly to PCA: coordinates for observations and loadings for variables, but adjusted to account for mixed types.

Output and interpretation

• Individual coordinates: each observation gets coordinates on principal dimensions (useful for scatterplots, clustering).
• Variable contributions: numeric variables have loadings; categorical variables show category coordinates and contribution measures.
• Explained variance: each dimension has an associated eigenvalue indicating how much variance it explains (interpreted with caution because of mixed scaling).

Practical tips

• Standardize numeric variables if they have different units or scales.
• Rare categories can dominate; consider combining rare levels.
• Use biplots to visualize individuals and variable contributions together.
• Retain only the first few dimensions that explain substantial variance for downstream tasks.
• Implementations available in R (FactoMineR::FAMD) and Python (prince library).

Example use cases

• Customer datasets with demographics (categorical) and spending (numeric).
• Survey data combining Likert scales and categorical responses.
• Medical records with lab values and diagnosis codes.

Quick workflow (steps)

1. Clean data, handle missing values.
2. Encode categorical variables (most FAMD implementations handle this internally).
3. Standardize numeric variables.
4. Run FAMD and inspect eigenvalues.
5. Plot individuals on first two dimensions; examine variable contributions.
6. Use coordinates for clustering or predictive models.

Limitations

• Interpretation of mixed-variable variance is less straightforward than PCA.
• Sensitive to scaling and rare categories.
• Computational cost grows with many categories (high-dimensional dummy encoding).

If you want, I can run an FAMD example on a sample dataset (R or Python) and show code + plots.