Testing the Significance of Class Refinement

Generalized linear modeling (GLM) is becoming a regular tool for insurance ratemaking. Actuaries and underwriters have begun to realize that classes may not simply interact, whether additively or multiplicatively. Some class combinations may synergize, or more than simply interact; others may counteract, or less than simply interact. But lest actuaries be tempted by abundant computer power and affordable GLM software to over-refine rating classes, they must know how to test whether class refinement is statistically significant. This paper provides the theory for this testing, and performs an illustrative test on a small dataset of automobile physical-damage claims.