Statistics Glossary → Kaiser criterion (eigenvalue > 1)
Kaiser criterion (eigenvalue > 1)concept
The rule, in principal component or factor analysis, of retaining the components (factors) whose eigenvalue in the correlation matrix exceeds 1.
In an analysis of the correlation matrix, each standardized variable has a variance of 1, and a component's eigenvalue shows the total variance it explains. The reasoning behind the Kaiser-Guttman rule is that a component with an eigenvalue below 1 explains less variance than a single original variable and so is not worth keeping. For an analysis of the covariance matrix, the equivalent rule is to keep components whose eigenvalue exceeds the average eigenvalue.
Although simple and widely used, the rule is of limited reliability: with many variables and a small sample, random sampling fluctuation alone produces eigenvalues above 1, so the rule often suggests too many components, and sharply separating near-identical eigenvalues such as 1.01 and 0.99 is arbitrary. It should therefore be weighed together with the scree plot, the cumulative variance explained and, above all, parallel analysis (Horn), which compares the observed eigenvalues with those from random data. The final decision should also take the interpretability of the components into account.
An everyday example
In a principal component analysis of a 20-item job-satisfaction survey, six components have eigenvalues above 1: 5.2, 2.4, 1.6, 1.1, 1.05 and 1.02. The Kaiser criterion suggests six components. Parallel analysis, however, shows that even random data produce first eigenvalues of up to about 1.5 and supports only the first three; reading the survey as three-dimensional (pay, colleagues, management) is more realistic.
In MerQur
It is one of the 'Number-of-components criterion' options in Principal Component Analysis (Manual, Kaiser, Cumulative variance target, Parallel analysis); for the covariance matrix the average-eigenvalue rule is applied. In Exploratory Factor Analysis, the Automatic setting of 'Number of factors' takes the number of factors with eigenvalues above 1 (Kaiser), and in Variable Clustering the 'MAXEIGEN (divisive)' threshold is a Kaiser-like stopping criterion.
Analyses where this concept appears
Variable Clustering (VARCLUS)Principal Component Analysis (PCA)Exploratory Factor Analysis (EFA)
Related concepts
Eigenvalue and eigenvectorParallel analysis (Horn) and the MAP criterion