Correspondence Analysis
Also known as: CA, Simple Correspondence Analysis, Reciprocal Averaging, Karşılıklı Uyum Analizi
Correspondence Analysis (CA) is an exploratory multivariate technique for visualizing the association structure of a two-way contingency table. Developed systematically by Jean-Paul Benzécri in France during the 1960s–1970s and brought to an English-language audience by Michael Greenacre in 1984, CA decomposes the chi-square statistic of a cross-tabulation to produce a low-dimensional joint display — called a biplot — in which rows and columns are represented as points whose proximities reflect their associations.
Key highlights
- Produces an intuitive joint visual display (biplot) of row and column categories in the same low-dimensional space.
- Scales naturally to large contingency tables with many categories without requiring distributional assumptions.
- Total inertia provides a chi-square-based global measure of association, giving statistical grounding to the geometric solution.
- Handles asymmetric data structures well; both symmetric and asymmetric biplots can be derived from the same decomposition.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Correspondence Analysis when you have a two-way contingency table of count data and wish to explore associations between categorical row and column variables. The method assumes the data are non-negative counts (or frequencies) and that an exploratory, rather than confirmatory, goal is primary. CA is appropriate when categories are nominal or ordinal; for ordinal variables consider non-linear CA variants. It is not suitable for continuous or interval-level data — use Principal Component Analysis instead. When more than two categorical variables are involved, Multiple Correspondence Analysis is the appropriate extension.
Strengths & limitations
- Produces an intuitive joint visual display (biplot) of row and column categories in the same low-dimensional space.
- Scales naturally to large contingency tables with many categories without requiring distributional assumptions.
- Total inertia provides a chi-square-based global measure of association, giving statistical grounding to the geometric solution.
- Handles asymmetric data structures well; both symmetric and asymmetric biplots can be derived from the same decomposition.
- Interpretation relies on proximity between points in a reduced space, which can be misleading if the retained dimensions capture only a small fraction of total inertia.
- CA is purely exploratory; it does not produce p-values or confidence intervals for specific category pairs without additional resampling procedures.
- Results are sensitive to rare categories or cells with very small counts, which can disproportionately pull the solution.
- The method applies only to two-way tables; richer multi-table designs require Multiple Correspondence Analysis or other extensions.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How is Correspondence Analysis different from Principal Component Analysis?
PCA operates on continuous variables using Euclidean distance and a covariance or correlation matrix. CA operates on count data in a contingency table using chi-square distance between row or column profiles. The two methods share an SVD backbone but differ fundamentally in the data type they accept and the distance metric they optimize.
How many dimensions should I retain in a CA solution?
Retain enough dimensions so that the cumulative inertia (percentage of total chi-square explained) is acceptably high — commonly 70–80% or more. A scree plot of singular values helps identify the elbow beyond which additional dimensions contribute little. The maximum number of meaningful dimensions is min(I-1, J-1) where I and J are the numbers of rows and columns.
Can I apply CA to a table that contains zeros?
Structural zeros (cells that cannot logically contain observations) are problematic because CA relies on log-ratio geometry implicitly and zero cells produce undefined profile coordinates. Sampling zeros (rare but possible events) are generally tolerable if sparse; a common remedy is to merge rare categories or add a small constant, though the latter alters the chi-square structure and should be reported transparently.
Sources
- 1.Greenacre, M. J. (1984). Theory and Applications of Correspondence Analysis. Academic Press.ISBN 978-0-12-299050-2
You have read it. What now?
Cite this page
ScholarGate. (2026, June 2). Correspondence Analysis. ScholarGate. https://scholargate.app/statistics/correspondence-analysis