Geometric Data Analysis, From Correspondence Analysis to Structured Data Analysis (book review)

Review of: Brigitte Le Roux and Henry Rouanet, Geometric Data Analysis, From Correspondence Analysis to Structured Data Analysis, Kluwer, Dordrecht, 2004, xi+475 pp.

💡 Research Summary

The book “Geometric Data Analysis, From Correspondence Analysis to Structured Data Analysis” by Brigitte Le Roux and Henry Rouanet offers a comprehensive, mathematically rigorous yet accessible treatment of correspondence analysis (CA) and its extensions within a unified geometric framework. The authors trace the lineage of “geometric data analysis” back to Patrick Suppes, emphasizing that mathematical structures dictate analytical procedures and that the burden of proof lies in the mathematics, not in ad‑hoc statistical heuristics.

Chapter 2 lays the conceptual foundation by introducing the “measure‑versus‑variable duality.” Rows (measures) and columns (variables) are treated as dual vector spaces, and the authors adopt a transition notation—subscripts for measures, superscripts for variables—that replaces cumbersome matrix algebra. Within this notation, the χ² metric on the dual clouds, the Euclidean metric on the factor space, and the ultrametric induced by hierarchical clustering are presented as three complementary distance concepts. The chapter also connects CA to Fisher’s linear discriminant analysis, canonical correlation, regression, and multidimensional scaling, providing a broad view of its relationships to other multivariate techniques.

Chapter 3 moves from abstract duality to concrete geometry. It explains how a cloud of points in Euclidean space can be decomposed spectrally (eigenvalues/eigenvectors) to reveal concentration ellipsoids, and how Ward’s minimum‑variance agglomerative clustering—referred to as “Euclidean classification”—operates directly on the CA factor scores. The authors discuss nearest‑neighbour algorithms, noting the omission of the more efficient nearest‑neighbour‑chain method and the lack of an explicit treatment of ultrametrics, but they point readers to Murtagh (2005) for those details.

Chapter 4 revisits Principal Components Analysis (PCA) through the lens of CA, showing how the same geometric principles underlie both methods and how PCA can be interpreted as a special case of correspondence analysis on continuous data.

Chapter 5 introduces Multiple Correspondence Analysis (MCA) for questionnaire data with many categorical modalities. The authors detail the construction of Burt tables, the selection of active questions, and the handling of sparse modalities. A full case study of a 1997 French government survey (3,002 respondents) illustrates the complete workflow from coding to interpretation.

Chapter 6 expands the framework to “structured data” by grafting ANOVA, MANOVA, and regression onto the CA geometry. The chapter demonstrates how nesting and crossing of factors (e.g., age, gender, education) can be accommodated, and it presents a striking example involving annotated video of basketball players to identify high‑potential athletes.

Chapter 7 focuses on stability analysis. After reviewing functional‑data approaches, bootstrapping, and combinatorial inference, the authors present the Escofier‑Le Roux perturbation method, which perturbs a cloud relative to a reference cloud and examines the impact on eigenvalues, factor scores, and clustering. Case studies explore the effect of deleting a single observation, removing a variable, or omitting an entire group.

Chapter 8 advocates an “Inductive Data Analysis” philosophy that integrates descriptive statistics (based on relative frequencies, independent of sample size) with inferential statistics (which do depend on sample size). The authors discuss traditional frequentist tests, Bayesian inference, and combinatorial methods, all framed geometrically for clarity.

Chapter 9 showcases three extensive applications: a medical trial on Parkinson’s disease, a study of French political attitudes before the 1997 elections, and an evaluation of Stanford’s Education Programme for Gifted Youth. Each example demonstrates data coding, structuring, CA‑based dimensional reduction, and subsequent inferential modeling, reinforcing the book’s claim that the geometric approach scales from small exploratory studies to large‑scale policy research.

Throughout the volume, the authors favor transition‑based notation, lightly employ Einstein tensor notation, and occasionally reference Dirac notation as used in quantum physics, underscoring the deep mathematical roots of the methodology. They also weave in Pierre Bourdieu’s concept of a “social space,” showing how CA can map social fields into geometric spaces—a rare but powerful bridge between quantitative methods and sociological theory.

In sum, Le Roux and Rouanet deliver a richly illustrated, exercise‑driven, and mathematically sound guide to geometric data analysis. The book is valuable for statisticians, data scientists, and social scientists who seek a unified, geometry‑centric perspective on correspondence analysis, its extensions, and its application to structured, high‑dimensional data.