Variance Partitioning

Variance partitioning is the statistical practice of decomposing the total variance of a dependent variable into components attributed to different sources — treatments, blocks, random effects, genetic factors, environmental factors. The technique, originating in Ronald Fisher's analysis of variance and extended through mixed-effects models and structural equation modeling, is one of the most powerful and most abused tools in quantitative science.

The mathematical validity of variance partitioning is not in question. Under the assumptions of the linear model, the decomposition is exact and the expected mean squares have the properties Fisher proved. The epistemological validity is another matter. Variance partitioning presupposes that the causes of a phenomenon can be represented as additive, independent, and non-interacting components. When these assumptions are violated — as they almost always are in biological and social systems — the decomposition produces numbers that are mathematically correct and causally meaningless.

The most common abuse is the interpretation of variance components as measures of causal importance. A variance component is a measure of spread, not of mechanism. It tells us how much a source contributes to differences between observations, not how it produces those differences. The conflation of statistical partitioning with causal attribution has produced decades of confusion in heritability research, educational evaluation, and program assessment.

Modern alternatives — including variance components models with structured covariance, Bayesian hierarchical models, and causal graphical models — attempt to preserve the inferential power of partitioning while relaxing its restrictive assumptions. Whether they succeed depends on whether the field is willing to trade the simplicity of a single "proportion explained" for the complexity of a properly specified generative model.

Variance partitioning is a statistical telescope: it reveals how much variation exists, not what produces it. Treating it as a causal microscope is the original sin of quantitative social science.