PSIII-24 Establishment of multilevel linear model and analysis of factors affecting piglet litter performance at birth in central China.

This study aimed to establish a multilevel linear model and analyze the factors affecting piglet litter performance at birth. A total of 17,906 litter performance at birth from 16 commercial pig farms were collected from January 2010 to December 2012 in central China. The general linear regression model (PROC GLM), multilevel Poisson regression model (PROC GLMMIX and PROC NLMIXED), and multilevel linear model (PROC MIXED) were established in SAS software to compare the goodness of fit among the three models. Results showed the ICC of total born piglet (TBP), piglet born alive (PBA), low birth weight piglet (LBW), and average birth weight (ABW) were 27.89%, 23.88%, 24.66%, and 22.27%, respectively (P < 0.05). The AIC, AICC, BIC, and -2LL in the multilevel linear models of TBP, PBA, LBW, and ABW were all lower than those in the general linear regression models. Moreover, the Pearson residuals of TBP, PBA, and LBW increased to nearly 1 after introducing discrete scale factor into models. The P values were all similar between the multilevel Poisson regression models and multilevel linear models for TBP, PBA, and LBW. Furthermore, multilevel analysis revealed the litter performance at birth was significantly influenced by management at farm level, and breed, parity, gestation diet, year, and season at litter level (P < 0.05). In conclusion, the multilevel linear model is better fit for the data of litter performance at birth than the general linear regression model. To simplify the analysis of discrete data, the multilevel Poisson regression model can be replaced by the multilevel linear model. Importantly, factors affecting litter performance at birth from the multilevel linear model provides valuable information on sow production management.