Logistic joinpoint regression model in cohort studies

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): In studying trend data such as cancer mortality and incidence data one is frequently concerned with detecting a change in recent trend. The ability to identify such changes in a cohort is an important problem for both retrospective and prospective cohort studies when looking for disease patterns. The proposed research seeks to develop a method to broaden the applicability of the join point regression model in detecting changes in disease trends. Specifically, we shall consider the extension of the simple Gaussian joinpoint regression model to logistic regression with K responses and possibly non-homogenous dispersion parameters. We shall derive and implement with the software the method for estimation and testing of the model parameters on the basis of the conditional maximum likelihood. Since the location of the joinpoints (change points) in the model is unknown the method would employ the iterative conditional maximization algorithm in seeking the solutions of the likelihood equations. In order to test the validity of the final model as well as to assess the significance of the final set of detected change points we shall sequentially apply the parametric bootstrap method. The conditions for consistency and general appropriateness of all the resulting estimation and testing procedures in our setting shall be also derived. Additionally, we shall compare via simulation studies the performance of the joinpoint logistic regression model versus that of penalized splines (P-splines) and multivariate adaptive regression splines (MARS) models. Finally we shall also apply the developed model to the longitudinal dataset on cancer mortality among the members of the Louisville VC cohort of now retired chemical workers up until 1996. The dataset is available via the University of Louisville Health Surveillance Program. Using the joinpoint logistic regression model we shall determine the pattern of longitudinal changes (time change-points) in the cohort cancer occurrences as compared with the state reference population, adjusting for the temporal clustering of the disease in the different production areas. The use of the Louisville VC cohort data shall allow us to illustrate our approach in an innovative application to monitoring occupational diseases and to compare its effectiveness with that of the standard methodology in view of the multiplicity of different analysis of this dataset available in the literature.