中文摘要：

在非寿险分类费率厘定中,广义线性模型的应用十分普遍,但当某些费率因子的水平数很多时（本文称之为多水平因子）,广义线性模型的估计结果将不可靠。解决此类问题的一种方法是把多水平费率因子作为随机效应处理。将多水平费率因子作为随机效应处理可以采取下述三种方法：（1）分别用广义线性模型和信度模型估计普通费率因子和多水平因子,通过广义线性模型与Buhlmann-Straub信度模型的迭代应用预测索赔频率和索赔强度;（2）应用广义线性混合模型分别预测索赔频率和索赔强度;（3）直接对经验纯保费数据建立Tweedie混合效应模型。本文把上述模型应用于中国车损险实际数据的研究结果表明,这三种方法比较接近,但从总体上看,广义线性混合模型的估计结果更加可取。

英文摘要：

Generalized linear models are widely used in non-life insurance ratemaking, but they may produce unreliable parameter estimates when some factors have too many levels （called multi-level fac- tors）. One approach to solving this problem is treating a multi-level factor as a random effect. The paper discusses three methods of treating a multi-level factor as a random effect： 1） combining generalized linear models and Buhlmann-Straub credibility model; 2） Building generlaized linear mixed models for claim frequency and claim severity respectively; 3） Building Tweedie mixed models for experience pure premium. Based on a claim data set of vehicle damage insurance from a Chinese property insurance com- pany, the paper shows that the three methods produce the similar results, but GLMM is more advisable.