See How Your Categories Map Together Not Just Whether They're Related

How it works

Classical correspondence analysis. The two mapped columns are cleaned (blanks become a 'Missing' category, categories below a stated observation threshold or outside the eight most common are pooled into 'Other') and cross-tabulated, weighted by the count column when one is mapped. The correspondence matrix P = N/n gives row and column masses; the matrix of standardized residuals, D_r^(-1/2) (P - r c') D_c^(-1/2), is decomposed by base singular value decomposition. The squared singular values are the principal inertias, summing to the total inertia — which equals the chi-square statistic divided by the sample size, the identity that ties the map back to the test of independence. Both sets of categories are placed in principal coordinates (the symmetric map). Per-category contributions (mass times squared coordinate over the dimension's inertia) and cos-squared quality (squared coordinate over the category's total squared distance from the average profile) are reported so that poorly represented points can be excluded from interpretation. Dimension signs are oriented deterministically so repeated runs give the same picture.

Use it when a chi-square or Cramér's V has told you two categorical variables are related and you need to see the shape of that relationship — which categories go with which, and along what underlying dimensions.

Not for testing whether two categorical variables are related at all (use the categorical-association tool — chi-square, Fisher, Cramér's V), not for two-category columns (a single dimension, no map), not for numeric variables (use correlation or principal components), and not for more than two categorical variables at once (that is multiple correspondence analysis).

What data do you need?

One row per observation of two categorical variables. For example, survey respondents naming an attribute for a brand:

Minimum 30 rows · Best with 200-20,000 rows, 3-8 categories per column

What's in the report?

Standard-library analysis: how do two categorical variables map together? Map two categorical columns — brands against attributes, segments against behaviours, products against complaint reasons — and get the classical correspondence analysis: the contingency table's total inertia with a chi-square test of independence, the dimension table showing how much of the association each dimension explains, row and column coordinates on the first two dimensions, the symmetric biplot with both sets of categories on one map, and per-category contribution and quality-of-representation (cos-squared) tables so you know which points are actually well enough represented to interpret. The report states, in your own column names, the rule almost every reader breaks: in a symmetric biplot the distance from a row point to a column point is not a measure of association.