Longitudinal threshold model analysis of clinical mastitis using linear splines

Abstract

This study evaluated random regression threshold models using linear splines (RRLS) versus Legendre polynomials (RRLP) as basis functions for the genetic analysis of clinical mastitis (CM) in first-lactation Finnish Ayrshire cows. Estimates of genetic parameters from RRLS models with 4 and 6-knots were compared to estimates from RRLP model fitting Legendre polynomials of third-order in modelling the genetic trajectory of CM over lactation. Co-variance components were estimated using Bayesian threshold models and MCMC procedures. Posterior means of heritability of CM before calving to one month after calving was from 0.04 to 0.14 and 0.04 to 0.05 from RRLP and RRLS models, respectively. From calving to 300dim after calving, heritability ranged from 0.03 to 0.09 and 0.04 to 0.06, respectively. Estimates from RRLP were particularly high and erratic at both ends of the lactation trajectory whilst estimates from RRLS had no marked end-of range problems. However, with RRLS models concavity in heritability curves and angularity in genetic variance estimates at position of knots were observed. This could partly be due to the sparse and uneven distribution of data points resulting in weak genetic correlations between knots. During lactation, genetic correlations from RRLP ranged from 0.27 to 0.98. Whereas correlations between adjacent knots from RRLS were in part weak and ranged from 0.01 to 0.52 except between the first two knots (rg=0.97). Rank correlations between estimated breeding values from both models were relatively high and ranged from 0.96 to 0.98. However, compared to RRLS 6-knots and RRLP models, RRLS 4-knots model had the lowest Deviance Information Criterion indicating that it is best fit and less parameterised model.