Software for Missing Covariate Data in Cancer Trials.

描述（由申请人提供）： 这是一个第二阶段SBIR提案，旨在完成LogXact、EGRET、SAS（作为PROC）和SPLUS（作为功能）中嵌入的统计工具的全面集合的开发。这组工具将计算广义线性模型（GLM）和具有缺失分类协变量的参数生存模型的最大似然估计值，其中缺失协变量假设为随机缺失（MAR）。在第二阶段的工作中，我们将扩展原型软件XMISS中可用的当前版本的工具，以处理： (i)具有Iogit、probit或互补双对数链接的二项响应模型缺失分类协变量， (ii)匹配病例对照数据的条件logistic回归缺失分类协变量， (iii)泊松回归模型的分类协变量缺失， (iv)正态线性回归模型的分类协变量缺失， (v)有序响应回归模型的分类协变量缺失， (vi)指数、Weibull和对数正态回归模型的分类协变量缺失，允许响应变量的右删失。 在为上述所有GLM和生存模型开发II期软件时，我们将允许缺失协变量最多有5个类别，其中任何一个类别都可能有缺失值。此外，在第二阶段开发中，我们将允许总共多达50个协变量，其中10个二元协变量可能缺失。此外，我们将研究加快EM算法的方法，以及开发新的算法，以获得良好的EM算法的初始值。 缺失协变量数据是癌症临床试验中常见的问题。对于上述模型范围，尚无通过最大似然法处理缺失协变量数据的商业软件。

DESCRIPTION (provided by applicant): This is a Phase II SBIR proposal for completing the development of a comprehensive collection of statistical tools embedded in LogXact, in EGRET, in SAS as PROCs and in SPLUS as functions. This set of tools will compute maximum likelihood estimates for generalized linear models (GLMs) and parametric survival models with missing categorical covariates, where the missing covariates are assumed to be missing at random (MAR). In this Phase II effort, we will expand the current version of tools available in prototype software XMISS to handle: (i) missing categorical covariates for binomial response models with Iogit, probit, or complementary log-log links, (ii) missing categorical covariates for conditional logistic regression for matched case-control data, (iii) missing categorical covariates for Poisson regression models, (iv) missing categorical covariates for normal linear regression models, (v) missing categorical covariates for ordinal response regression models, (vi) missing categorical covariates for exponential, Weibull and log-normal regression models allowing for right censoring in the response variable. In the development of the Phase II software for all of the GLM's and survival models considered above, we will allow a missing covariate to have up to 5 categories, any of which may have missing values. Also, in Phase II development, we will allow up to 50 covariates total, of which 10 binary covariates can be missing. In addition, we will investigate methods for speeding up the EM algorithm as well as develop new algorithms for obtaining good starting values for the EM algorithm. Missing covariate data is very common problem with cancer clinical trials. There exists no commercial software to handle missing covariate data by maximum likelihood method for the range of models listed above.