Use of (co)variance structures in multivariate mixed models may improve predictions

Speaker: Tormod Ådnøy, Associate Professor, Department of Animal and Aquacultural Sciences, Norwegian University of Life Sciences.

Abstract:

Goat breeding values for milk fat percentage improved 3-4% using heritable part of FTIR specter compared to using heritability of fat percentage.

Take home message:

Modelling known structures multivariately may improve univariate diagnoses derived from multivariate measures.

Fat percentage and other milk contents are routinely measured in Dairy Control using FTIR light specter. Hundreds of wavelength measures are reduced to one fat percentage per sample. (Calibration of analysis is done using chemometrics and comes with the machine.) To breed for changes in fat percentage a mixed model including genetic relationship and environmental covariance structures due to farm and test-day etc. is used for the univariate fat percentage, and breeding values of selection candidates predicted.

We propose to use the same mixed model for the multivariate FTIR specter to find the ‘breeding values’ of the multivariate specter, and then use the linear combination the multivariate specter breeding values to calculate univariate fat percent breeding values. The covariance structures of the multivariate specter should improve the prediction.

The challenge of the method is estimation of variance components for the multivariate specter. Available methods limited the number of variables to less than 30-40. We found that with eight principle components very much of the total (co)variance structure of the around 300 FTIR wavelengths was kept, and estimated their (co)variance components.

We have shown that accuracies of fat percentage breeding values in goats might improve by 3-4% by predicting the heritable part of the specter compared to predictions based on available univariate phenotypic fat percentage in Norwegian Dairy Goat. This was from the same FTIR specters as the fat percentage. We would prefer an independent chemical assession of fat percentage for verification. We have found a dataset in Poland with blood assessed ketosis in dairy cows and corresponding milk FTIR specters and will try if our multivariate method better predicts blood ketosis than a univariate prediction.

In a nested case-control study, controls are selected among those at risk at the failure times of the cases (corresponding to matching on study time). Usually additionally matching is also performed to control for potential confounders. If one breaks the matching, nested case-control data may be considered as case-cohort data with a non-standard sampling design for the subcohort. The data may then be analysed using inverse probability weighting or by methods that make use of the data for the full cohort (ML-estimation and multiple imputation). In the talk, methods for analysing nested case-control data will be reviewed, and it will be discussed when it may be useful to break the matching and when bias will be introduced by doing so.

 

Published May 2, 2015 2:39 PM - Last modified May 7, 2015 1:49 PM