What Is Correspondence Analysis and How Does It Turn a Contingency Table into a Perceptual Map?

Correspondence analysis (CA) is a multivariate statistical technique used to visualize the associations between two categorical variables in a contingency table. Developed by Jean-Paul Benzécri in the 1960s, CA decomposes the chi-square distance between rows and columns into low-dimensional coordinates, producing a perceptual map that reveals how categories relate to one another.

What It Is

Correspondence analysis is essentially a principal component analysis applied to categorical data. Instead of summarizing continuous variables, CA works on frequency counts—the cells of a contingency table—and transforms them into a graphical display where both row and column categories are plotted as points in the same space.

The key output is a perceptual map (also called a biplot or symmetric map), where the proximity between points indicates similarity in profile. Points close together share a similar pattern of association; points far apart are dissimilar. This makes CA especially valuable in market research, customer satisfaction surveys, and any field where you need to explore relationships between categorical responses.

How It Works: The Steps

The mathematics of CA is built on the chi-square metric. Here is the standard procedure:

1. Start with a contingency table of dimensions \(r \times c\), where rows are categories of one variable and columns are categories of the other.

1. Compute the correspondence matrix \(P\) by dividing each cell count \(n_{ij}\) by the total \(N\). Let \(p_{ij} = n_{ij}/N\).

1. Calculate row and column masses (marginal proportions):

- Row mass: \(r_i = \sum_j p_{ij}\)
- Column mass: \(c_j = \sum_i p_{ij}\)

1. Compute the standardized residuals (the core of CA):

\[
z_{ij} = \frac{p_{ij} - r_i c_j}{\sqrt{r_i c_j}}
\]
This matrix measures how much each cell deviates from the independence model, scaled by the expected value.

1. Perform singular value decomposition (SVD) on the residual matrix \(Z\). This yields eigenvalues (called inertia in CA), which quantify how much variance each dimension explains.

1. Derive principal coordinates for rows and columns. The first two dimensions (with the largest inertia) are typically plotted to form the perceptual map.

The total inertia equals \(\chi^2 / N\), where \(\chi^2\) is the Pearson chi-square statistic of the table. Each dimension's contribution to total inertia tells you how much of the association structure is captured by that axis.

A Worked Illustrative Example

Example data (illustrative only). Suppose a small survey asks 100 customers which of three product brands (A, B, C) they prefer, segmented by age group (Young, Middle, Senior). The contingency table is:

Brand A | Brand B | Brand C | Row total

• Young | 25 | 10 | 5 | 40
• Middle | 10 | 20 | 10 | 40
• Senior | 5 | 10 | 5 | 20
• Column total | 40 | 40 | 20 | 100

Following the CA steps:

• Total \(N = 100\), so \(p_{ij} = n_{ij}/100\).
• Row masses: Young = 0.40, Middle = 0.40, Senior = 0.20.
• Column masses: Brand A = 0.40, Brand B = 0.40, Brand C = 0.20.

Computing standardized residuals and performing SVD yields two principal dimensions. In this illustrative case, the first dimension might explain about 85% of the inertia, and the second about 15%. The perceptual map would show:

• Young customers close to Brand A (strong association).
• Middle customers close to Brand B.
• Senior customers sitting near the origin, indicating a weak or average preference profile across all brands.

The map lets you read at a glance that brand preference is age-related, with the strongest association between Young and Brand A.

Common Pitfalls

• Overinterpreting proximity across different axes. In a symmetric map, distances between a row point and a column point are not directly interpretable as chi-square distances—only relative positions matter.
• Ignoring inertia contributions. If the first two dimensions explain less than 70–80% of total inertia, the 2D map may be misleading. Always report the explained inertia.
• Using CA on raw counts without checking sample size. CA is sensitive to sparse cells; very small expected counts can distort the map.
• Forgetting that CA describes association, not causation. A perceptual map shows patterns, never cause-and-effect relationships.

Try It Yourself

Correspondence analysis is a powerful exploratory tool, but the matrix algebra can be tedious by hand. Use the free, no-signup calculator at 6SQ Correspondence Analysis Tool to paste your contingency table and instantly generate the perceptual map with inertia values—no software installation required.