Correspondence Analysis

Description

Turn a cross-tabulation of categorical variables into a map on which associated row and column categories sit close together, with inertia showing how much association the map retains.

Correspondence analysis is the categorical counterpart of principal components analysis. Counts in a contingency table are converted to row and column profiles and masses and chi-square based association measures; dimensions are extracted with the same matrix algebra as PCA, and row and column categories are plotted in one space. Total inertia (chi-square divided by n) measures how much association there is; Mazzocchi suggests above 0.20 for a useful map, and each dimension's share of inertia shows how much the plot keeps. Normalisation (row principal, column principal, symmetrical) decides which distances can be compared, and row-to-column distances are not direct association measures. Rare or irrelevant categories can be made supplementary. Multiple correspondence analysis extends the method to three or more variables via an indicator or Burt matrix. Axis units have no intuitive meaning: the value is in seeing which categories cluster.

Attributes

Loss type
Siloed
Effort
Hours

Steps

  1. Build a contingency table of counts for two categorical variables, categorising continuous ones first.
  2. Compute row and column profiles, masses and chi-square association measures, and extract dimensions (for example with ca in R or prince in Python).
  3. Check total inertia (ideally above 0.20) and the share of inertia on the first two dimensions.
  4. Choose a normalisation: row principal to compare rows, column principal to compare columns, symmetrical otherwise.
  5. Mark sparse or irrelevant categories as supplementary points.
  6. Read the contributions of points to dimensions and of dimensions to points, and name the dimensions.
  7. Plot rows and columns together and interpret proximities within each set; for three or more variables run multiple correspondence analysis.

Sources

  • Statistics for Marketing and Consumer Research (Mazzocchi, 2008)
  • Practical Statistics for Data Scientists (Bruce et al., 2020)
  • Quantitative Social Research Methods (Singh, 2007)
  • The SAGE Handbook of Quantitative Methods in Psychology (Millsap & Maydeu-Olivares, 2009)
  • A User's Guide to Principal Components (Jackson, 1991)
  • Methods of Multivariate Analysis (Rencher, 2002)

Also known as

  • CA
  • Simple correspondence analysis
  • Multiple correspondence analysis (MCA)
  • Dual scaling
  • Homogeneity analysis
  • Biplot of a contingency table

Categories

  • Analytics & Behaviour
  • Statistics & Modelling