MerQuR+Statistics Glossary

Statistics Glossary → Proportion of explained variance

Proportion of explained varianceconcept

The share of the total variability in the data captured by a component, factor or model.

In principal component analysis, the proportion of variance explained by each component is its eigenvalue divided by the sum of all eigenvalues; when working with the correlation matrix, that sum equals the number of variables. Because components are ordered from largest to smallest, the cumulative proportion shows how much information the first m components carry together; a cumulative target such as 70-80% can be used to choose the number of components. In factor analysis only the common variance (the part shared by the common factors) is explained, and the proportions are usually smaller than in PCA; after rotation the distribution across factors changes but the total stays the same.

The proportion of explained variance alone does not show that a component is meaningful or interpretable; a small component can be theoretically important, and a large one can arise from scale effects. In canonical correlation and partial least squares regression, how much variance each canonical variate or latent component explains in its own set of variables and in the other set (or the response) is reported separately. R² in regression is a form of the same idea: the proportion of the outcome's variance explained by the model.

An everyday example

A music platform wants to summarize users' listening time across 10 genres. In a principal component analysis the first component (overall listening) explains 42% of the variance, the second (pop and rap versus classical and jazz) 21% and the third 11%. Since the first three reach 74% together, the platform can work with three summary scores instead of ten genres.

In MerQur

In principal component analysis the 'Variance target (%)' option sets the cumulative explained-variance criterion for the number of components; in variable clustering the 'R² and 2nd eigenvalue' output gives the explained proportions of the clusters. It is also reported in the output of exploratory factor analysis, canonical correlation and partial least squares regression.

Analyses where this concept appears

Canonical CorrelationVariable Clustering (VARCLUS)Principal Component Analysis (PCA)Exploratory Factor Analysis (EFA)Partial Least Squares Regression (PLS)

Related concepts

Eigenvalue and eigenvectorCommunality