Adjusting for missing data due to culling before testing in genetic evaluations of swine

Abstract

The aim of this study was to evaluate data augmentation in genetic evaluations as a method of adjusting for missing data due to culling of pigs before testing. A stochastic simulation was used to generate 10 yr of data for age at test (AGE) and fat thickness (FAT) in a breeding unit with 100 sows and 15 boars. Culling was performed at random (C-RAND), within litters (C-W/IN) or over litters (C-OVER), by deleting two-thirds of the records from the simulated data sets. The culling variate (CVAR) used had genetic and phenotypic correlations of 1.00, .75, .50, and .25 with AGE [r(CVAR, AGE)], whereas culling was uncorrelated with AGE in C-RAND. Missing records for AGE were replaced with their expectations (dummy records), based on the phenotypic average of the tested animals and selection intensities applied. With missing records, predictions were seriously biased for AGE in C-W/IN and especially in C-OVER, when r(CVAR, AGE) exceeded .50 and .25, respectively. The ranking of the animals was more affected in C-OVER than in C-W/IN. With dummy records, bias was removed effectively in cases with a high r(CVAR, AGE) in C-W/IN and C-OVER, whereas a larger bias was created in the opposite direction when r(CVAR, AGE) was less than .50 and for C-RAND. In conclusion, this method was beneficial for adjusting missing data owing to culling, when the correlation between CVAR and AGE was .50 or greater.