MerQuR+Statistics Glossary

Statistics Glossary → Partial and semi-partial correlation

Partial and semi-partial correlationconcept

Correlations that show how strong the relationship between two variables is after the influence of one or more third variables has been removed.

The partial correlation is the correlation between the residuals left after the linear effect of Z has been removed from both X and Y; for a single control variable, r_XY·Z = (r_XY - r_XZ·r_YZ) / √((1 - r²_XZ)(1 - r²_YZ)). In the semi-partial (part) correlation, the effect of Z is removed from only one variable (the predictor, in multiple regression) and the other is left as it is. The squared semi-partial correlation equals the unique increase in R² obtained by adding that predictor to the model.

The partial correlation answers the question 'is there still a relationship between X and Y when Z is held constant?' and helps expose spurious associations produced by a third variable. It removes only linear effects, however, and cannot control for unmeasured variables. If the control variable lies on the causal path between X and Y (a mediator), controlling for it can hide a real effect, so the choice of control variables should be made on theoretical grounds. The zero-order correlation is the simple correlation with nothing controlled.

An everyday example

In a city, ice-cream sales and drownings are strongly and positively correlated. That does not mean ice cream causes drowning: both rise in hot weather. Once air temperature is controlled for, the partial correlation comes out close to zero, showing that the relationship is largely due to temperature.

In MerQur

In Correlation Analysis the 'Control variables (partial r)' parameter computes the partial and semi-partial correlations; in Multiple Linear Regression the 'Zero-order / partial / part correlations' extra output gives them for each predictor. Hierarchical Clustering includes the semi-partial R² among its cluster-number criteria, and Exploratory Factor Analysis offers the MAP criterion, based on the squared partial correlation, for the number of factors.

Analyses where this concept appears

Correlation AnalysisMultiple Linear RegressionHierarchical ClusteringExploratory Factor Analysis (EFA)

Related concepts

Pearson correlation coefficientCovariate