Predictive ability of alternative models for genetic analysis of clinical mastitis

Abstract

Mastitis in cows can be defined as a binary trait, reflecting presence or absence of clinical mastitis (CM), or as a count variable, number of mastitis cases (NCM), within a defined time interval. Many different models have been proposed for genetic analyses of mastitis, and the objective of this study was to evaluate the predictive ability and sire predictions of a set of models for genetic evaluation of CM or NCM. Linear- and threshold liability models for CM, and linear, censored ordinal threshold, and zero-inflated Poisson (ZIP) models for NCM were compared in a cross-validation study. To assess the ability of these models to predict future data, records from 620492 first-lactation Norwegian Red cows, which were daughters of 3064 sires, were evaluated in a fourfold cross-validation scheme. The mean squared error of prediction was used for model comparison. All models but ordinal threshold model equally performed when comparing the overall predictive ability. This result was on average, across sick and healthy cows; however, the models behaved differently for each category of animals. For example, healthy cows were predicted better by the threshold and linear models for binary data and ZIP model, whereas for mastitic cows, the ordinal threshold model was by far the best model. Predicted sire effects and rankings of sires were highly correlated across all models. For practical purposes, the linear models are very competitive with the nonlinear models.

Multiple trait genetic evaluation of clinical mastitis in three dairy cattle breeds

“…They found that heritability estimates were fairly similar for the first three lactations. When comparing different CM measures, Vazquez et al (2012) found heritability estimates ranging from 3.2% to 8% for binary CM and number of CM cases. In our case, differences in heritability observed between breeds may be at least partly related to different mean incidences of the trait and to the quality of the data.…”

“…So on the long term, consequences of CM for breeders are increasingly detrimental. Although management and hygiene practices on farms have a large impact on CM (Pérez-Cabal et al, 2009;Vazquez et al, 2012), genetic selection for mastitis resistance is also a solution to be considered for improvement of UH. For example, Nordic countries have included mastitis in their genetic evaluation since 1978 and have shown that improvement of resistance to mastitis as a result of genetic selection is effective (Heringstad et al, 2003).…”

Validation of genomic predictions for wellness traits in US Holstein cows

“…Additionally, validation of marker effects is a common practice in genomics research within breed and also when the validation breed is different from the reference breed. Cross validation with random assignment of records to the training and validation populations is a robust approach commonly employed when validating a genetic evaluation (Vazquez et al, 2012;Haugaard et al, 2013). However, this approach can fail to properly account for clustering of animals within herds, and it does not remove the inherent bias of the methodology when herds are allowed to contribute phenotypes and pedigree to both the training and validation populations.…”

Modelling genetic evaluation for dressage in Pura Raza Español horses with focus on the rider effect

“…The lowest deviance information criterion (DIC) value (Spiegelhalter et al 2002) was used as a criterion of fitness, and a cross-validation approach was used to evaluate their prediction ability (Efron & Tibshirani 1993). Cross-validation was originally employed to evaluate the predictive validity of linear regression equations for forecasting a performance criterion from scores on a battery of tests (Mosier 1951), and nowadays, it is usually used in quantitative genetics (Olsen et al 2012;Vazquez et al 2012;Andonov et al 2013). For each model, the entire data set was randomly tenfold split into a training data set containing 6650 records (75%) to estimate the parameters and solve each model, and a validation data set with 2217 records (25%) to test the predictive ability of the model using the solutions obtained with the training set.…”