Software for Fitting Non-Gaussian Random Effects Models

用于拟合非高斯随机效应模型的软件

• 批准号: 6998525
• 负责人: TIM C HESTERBERG
• 金额: $ 36.75万
• 依托单位: INSIGHTFUL CORPORATION
• 依托单位国家: 美国
• 财政年份: 2004
• 资助国家: 美国
• 起止时间: 2004-04-01 至 2006-11-30
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/6998525
• 关键词: clinical research; computer program /software; computer system design /evaluation; data collection methodology /evaluation; data quality /integrity; epidemiology; human data; mathematical model; statistics /biometry

DESCRIPTION (provided by applicant): We propose a software implementation of recent methodological advances for efficiently computing Maximum likelihood estimates in multilevel mixed effects models in the context of generalized linear and parametric survival models. Such models are often used in the analysis of longitudinal and cluster sample data arising in Epidemiological and other studies. The recent methodological advances we propose to implement make it possible to compute consistent and asymptotically unbiased maximum estimates in a much wider variety of problems, and we also propose to compute statistics validating these estimates. The "Preliminary Results" section of this proposal shows that it easy to encounter situations with Epidemiological data in which the usual mixed effect model algorithms fail, even in problems with large sample sizes if the 'within cluster' sample sizes are small. These failures are made more troublesome by the fact that the user seldom has any warning that the computational algorithm has failed. We propose to provide such a warning. Most mixed effect model software assumes a multivariate normal random effect density. We propose to allow other densities, including user specified densities, in the random effects model. We also propose to develop software with adaptive MARS like model fitting capabilities.

描述（由申请人提供）：我们提出了在多层次混合效应模型中有效计算最大似然估计的最新方法学进步的软件实现，该模型在广义线性和参数生存模型的背景下。这种模型通常用于流行病学和其他研究中产生的纵​​向和簇样品数据的分析。我们建议实施的最新方法学进步使得在更广泛的问题中计算一致和渐近公正的最大估计是可能的，我们还建议计算验证这些估计值的统计数据。该提案的“初步结果”部分表明，使用流行病学数据遇到的情况很容易，其中通常的混合效应模型算法失败了，即使在“群集内部”样本尺寸较小的情况下，样本大小的问题也很小。由于用户很少警告计算算法失败，因此使这些失败变得更加麻烦。我们建议提供这样的警告。大多数混合效应模型软件都假定多元正常随机效应密度。我们建议在随机效果模型中允许其他密度，包括用户指定的密度。我们还建议开发具有自适应火星（例如模型拟合功能）的软件。