Application of classification trees and logistic regression to determine factors responsible for lamb mortality

The aim of the presented research was to statistically analyse the survival of 20,044 Polish Merino lambs between birth and 100 day of their life, using classification trees and logistic regression. The lamb survival trait was expressed in binomial scale: 1 for survival, 0 for mortality. Two different models of the trees were developed, depending on the division criterion: they were the function of entropy and the Gini index. For comparison purposes, an additional statistical analysis was carried out using a multiple logistic regression. The quality of decision tree models and multiple regressions was compared taking into consideration the following criteria: average error function, average squared error, lift cumulative, Kolmogorov–Smirnov statistics and the area under the Receiver Operating Characteristic curve. A statistical analysis was conducted using the Enterprise Miner 6.2 software included in the SAS package. The calculated quality criteria of four models that were developed lead to the conclusion that the classification trees established based on the Gini index, and on the function of entropy, are the most accurate in defining the variability of characteristics under examination, i.e. survival of lambs up to 100 days of age. In the case of the best classification model available, i.e. a tree built using the Gini index, the ranking of variable importance, which was developed based on the “Importance” measure, leads to the conclusion that the flock, type, and the year of a lamb's birth are the most significant differentiating factors.