Factor Regression Analysis: A New Method for Weighting Predictors.
Abstract
The optimum weighting of variables to predict a dependent/criterion variable is an important problem in nearly all of the social and natural sciences. Although the predominant method, multiple regression analysis (MR), yields optimum weights for the sample at hand, these weights are not generally optimum in the population from which the sample was drawn. A method was developed that sacrifices some 'prediction' in the sample at hand in order to achieve a more reliable and stable predictor composite. The method developed, Factor Regression Analysis (FRA), is based on the first principle component of the predictor intercorrelation matrix with validities in the diagonal cells. FRA yielded very stable predictor composites and weights--the weights themselves varied less from sample to sample than did MR weights from the same samples. These differences were marked for low sample sizes (e.g., N = 25), regardless of the number of variables in regression. With regard to prediction, FRA composites were substantially more valid in the population than the MR composites based on the same samples. The number of predictors in the subset did not turn out to be very important. FRA weights based on samples of 25 were about as valid as MR weights based on samples of 100. With samples of 200 the two methods yielded roughly equivalent prediction. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1976
- Accession Number
- ADA035441
Entities
People
- Ervin W. Curtis
Organizations
- Bureau of Naval Personnel