Software for Cox Regression Analysis of Interval-Censored Data

项目总结

Project Summary Interval-censored data arise frequently in clinical and epidemiological studies, because the time to the devel- opment of an asymptomatic disease (e.g., tumor occurrence, HIV infection, onset of diabetes or hypertension) cannot be observed exactly but rather is known to lie in a time interval between two consecutive clinical exam- inations. Recent theoretical and computational advances in nonparametric maximum likelihood estimation of semiparametric regression models with interval-censored data promise far more efficient and reliable analysis than what is currently possible. The broad, long-term objective of this SBIR proposal is to create a suite of commands, along with a companion text, in the widely used commercial software package Stata for performing cutting-edge nonparametric maximum likelihood estimation of the familiar Cox proportional hazards model with time-dependent covariates for interval-censored event times and for extending this methodology to multivariate interval-censored event times, which arise when several asymptomatic diseases or recurrences of a particu- lar disease are of interest or when study subjects are sampled in clusters (e.g., families, litters). The recently completed Phase I of this project has successfully established the scientific merit and technical feasibility of the proposed research and development effort by producing a prototype command for nonparametric maximum like- lihood estimation of the Cox proportional hazards model with potentially time-dependent covariates for univariate interval-censored data (i.e., a single event time for unrelated subjects) and by certifying the correctness of the estimation results from the new command against results from published papers and research code. The Phase II project will build on the success of the Phase I effort to develop a suite of reliable, robust, user-friendly, speed- and memory- efficient commands for semiparametric regression analysis of interval-censored data. Specifically, the Phase I code will be expanded substantially to incorporate stratification factors and likelihood ratio statistics (as an alternative to the Wald statistics implemented in Phase I) for univariate interval-censored data, to fit marginal Cox proportional hazards models for multiple diseases and clustered data and proportional rates/means models for recurrent events, and to provide model-checking procedures for both univariate and multivariate models. The correctness of the results will be certified in five clinical and epidemiological studies, and the sped-up code will be converted into a commercial-grade program with a graphical user interface and comprehensive documentation. Finally, a companion text will be written to document the software itself and serve as a substantive reference for researchers new to the field. The software program produced by this SBIR project will be a part of the Stata pack- age. This powerful and convenient software will enable biomedical investigators to analyze interval-censored data in a statistically efficient and unbiased manner. As such, this new tool will facilitate the search for effective inter- vention and prevention strategies for many common diseases (e.g., cancer, HIV/AIDS, diabetes, hypertension), thereby leading to improvements in public health.