Milk yield prediction at late lactation in reproductive rabbit does

Separating kits and mother to determine milk yield at 4^th wk of lactation (MY4) could have negative consequences on the training and development of young rabbits. In this work, a total of 313 lactation curves (28 d long), taken from 2 different trials, were used to fit regression models to estimate MY4. In both trials, females were subjected to a semi-intensive reproductive rhythm [insemination at 11 d post-partum (dpp) and weaning at 28 dpp], but diets, genetic types, parity order and day of controls were slightly different. The models included variables which, according to the bibliography, are related to milk yield and are often recorded in joint management (without separation of litters  rom mothers), such as litter size at weaning (LSW; both linear and quadratic), joint energy intake of doe plus litter at 4th wk of lactation (JEI; both linear and quadratic), perirenal fat thickness change (ΔPFTd) and milk yield at 3rd wk (MY3). The overlapping degree (OL) between current lactation and next pregnancy was included as a dummy variable, as well as their interactions with quantitative traits. To fit these models, 3  procedures were proposed to obtain accurate equations with biological meaning: Eq1, multiple linear regression (MLR) of data; Eq2, MLR with previous smoothing of sample distribution; and Eq3, MLR with previous smoothing and avoiding redundant samples and collinearities among variables. MY3 had a positive and relevant linear effect on MY4 for the 3 equations obtained (responsible for 39 to 50% of MY4 prediction). JEI had also a relevant role in MY4 prediction (28 to 61%), its positive effect being linear on Eq1, quadratic on Eq2 and both linear and quadratic on Eq3. ΔPFTd and LSW related traits were only included in Eq3, with a low relative weight, and OL inclusion did not improve prediction in any equation. Predicting MY4 was possible with the variables used, although certain precautions must be taken. Traditional MLR seems to predict central values properly, but extreme values poorly, whereas pre-treatment of data to smooth the dependent variable distribution appears to improve prediction of extreme values.