Maximum Likelihood Fitting of General Risk Models to Stratified Data
将用于逻辑回归的递归算法推广到任意相对风险模型，通过示例比较分层与非分层分析、加性与乘性风险模型，建议在非标准模型中避免使用基于观测信息的Wald检验和得分检验。
A recursive algorithm (Howard, 1972; Gail et al., 1981) useful for maximum conditional likelihood fitting of logistic regression models with large strata can be generalized to arbitrary relative risk models. An example is presented which permits comparison between fitting methods vis a vis stratified vs. unstratified analysis, additive vs. multiplicative risk model, and use of expected vs. observed information. On the basis of results from this comparison we suggest that Wald's test and the score test computed with observed information be avoided in non‐standard models. An interactive computer program is available for fitting multiplicative, additive and general risk models to stratified data.