Moussa M A
Comput Programs Biomed. 1980 Dec;12(2-3):161-71. doi: 10.1016/0010-468x(80)90062-8.
The likelihood ratio classification rule based on the location mode is estimated given: (1) data consist of both binary and continuous variables; (2) some states have either zero frequency or too few observations, the case that usually happens the practice. An iterative proportional fitting of convenient approximation of the log-linear models as well as a linear additive model are utilized in estimating the rule's parameters. Performance of the obtained rule is then assessed by estimated error rates.
（1）数据由二元变量和连续变量组成；（2）某些状态的频率为零或观测值过少，这种情况在实际中经常发生。在估计规则参数时，采用了对数线性模型的便捷近似的迭代比例拟合以及线性加性模型。然后通过估计的错误率来评估所获得规则的性能。
