Statistics Glossary → Parallel analysis (Horn) and the MAP criterion
Parallel analysis (Horn) and the MAP criterionconcept
A method for deciding the number of factors or components by comparing the observed eigenvalues with eigenvalues from random data, and a further criterion based on partial correlations.
Even in data made of completely uncorrelated variables, sampling fluctuation makes the first eigenvalues exceed 1. Horn's parallel analysis generates many random datasets with the same numbers of observations and variables as the real data and computes their eigenvalues. Factors whose observed eigenvalue exceeds the mean (Horn) or the 95th percentile (Glorfeld) of the random eigenvalues are retained. This usually gives a far more accurate answer than the Kaiser rule of keeping factors with eigenvalues above 1, which tends to overstate the number of factors, especially when there are many variables.
Velicer's MAP (minimum average partial) criterion computes the average squared partial correlation remaining among the variables after the first m components are partialled out; the m at which this average is smallest is the point where the systematic part of the common variance has been used up. MAP sometimes underestimates the number of factors, while parallel analysis with the mean criterion can occasionally suggest one too many. The two methods, the scree plot and theoretical interpretability should therefore be weighed together.
An everyday example
A teacher examines answers to a 30-item attitude scale and sees that four eigenvalues are larger than 1. In 100 fake datasets of random answers of the same size, the second eigenvalue averages 1.6, while the third eigenvalue of the real data is only 1.3. According to parallel analysis, only two factors carry structure beyond chance.
In MerQur
'Parallel analysis criterion' (95th percentile (Glorfeld) / Mean (Horn)) in Principal Component Analysis; the 'Number of factors' option (Automatic, Fixed 1-10, Parallel analysis, MAP) in Exploratory Factor Analysis.
Analyses where this concept appears
Principal Component Analysis (PCA)Exploratory Factor Analysis (EFA)