Searching For Life Data Models

Censored data arise from laboratory life tests and field tracking studies as well as in biological testing and medical clinical trials. The statistical literature contains an enormous number of models and methods for making inferences from such data. Choosing among these models can be a daunting and complex task for the non-specialist. We describe a method for searching through these models, by systematically decreasing the space of possible models for the dataset. This is done by structuring the space of models into a hierarchy of model descriptions, with most general models towards the root, and most specific models towards the leaves. The hierarchy defines a search tree, with maximum likeihood providing the search criterion. The tree is specified in a new language, implemented on top of the S statistical system. The language is designed to allow for generation of further such strategies for descriptions of statistical analysis tasks.