Correlation Estimates from Discontinuous Data

Open Access
Publisher Website
Google Scholar
Abstract
Continuous data of 40,416 1st-lactation Guernsey records of milk, fat and fat percent were stratified into categories to simulate discrete traits. Phenotypic correlations on the continuous scale generally were underestimated in absolute value when 1 or both traits were discrete. Underestimation became greater as the number of discrete classes decreased and became substantial (50%) when data on both traits were binomial-like. Genetic correlations on the continuous scale from paternal halfsibs were closely estimated when 1 or both traits were discrete. There were slight indications of overestimations, and fewer categories caused greater overestimation. Overestimation was greater when the underlying continuous genetic correlation was negative or small positive than when the correlation was high positive. When the underlying correlation was high positive, neither over- nor underestimation occurred. In all cases, the fewer the number of categories, the greater the sampling variation in correlation estimates.