|
Abstract:
|
Multinomial data may be "incomplete" if there exist observations that are only partially classified. That is, for some observations, while it is known that the response belongs to a given subset of the possible response categories, the true, precise classification is unknown. The incomplete multinomial distribution and incomplete multinomial regression have special relevance for the field of meta-analysis, in that studies may report multinomial outcomes using slightly different sets of response categories, complicating direct comparison of their results. For example, while many clinical trials for cancer patients use the Response Evaluation Criteria in Solid Tumors (RECIST) criteria to assess tumor response, investigators often choose to combine categories when reporting results. We discuss the extension of incomplete multinomial regression to the meta-analytic setting and demonstrate its use in a meta-analysis of tumor response in patients receiving either placebo or no anti-cancer therapy.
|
