Predicting IQ change from brain structure: A cross-validation study

Procedures that can predict cognitive abilities from brain imaging data are potentially relevant to educational assessments and studies of functional anatomy in the developing brain. Our aim in this work was to quantify the degree to which IQ change in the teenage years could be predicted from structural brain changes. Two well-known k-fold cross-validation analyses were applied to data acquired from 33 healthy teenagers – each tested at Time 1 and Time 2 with a 3.5 year interval. One approach, a Leave-One-Out procedure, predicted IQ change for each subject on the basis of structural change in a brain region that was identified from all other subjects (i.e., independent data). This approach predicted 53% of verbal IQ change and 14% of performance IQ change. The other approach used half the sample, to identify regions for predicting IQ change in the other half (i.e., a Split half approach); however – unlike the Leave-One-Out procedure – regions identified using half the sample were not significant. We discuss how these out-of-sample estimates compare to in-sample estimates; and draw some recommendations for k-fold cross-validation procedures when dealing with small datasets that are typical in the neuroimaging literature.     

似乎不相关回归模型及其在老年认知功能研究中的应用

目的 探讨似乎不相关回归模型在医学研究中的应用。方法 利用似乎不相关回归模型对老年人认知功能减退的影响因素进行分析。结果 特定的非独立结构情形下似乎不相关回归估计方法比传统最小二乘估计方法更合理，并可以提高参数估计的效率。结论 似乎不相关回归模型在医学科研领域具有良好的应用前景。     

稳健Poisson和log.binomial的GEE模型应用于非独立数据的研究

探讨流行病学资料中非独立数据的RR/患病率比(PR)的合适估计方法 。采用计算机模拟实验和实例分析观察稳健Poisson.GEE和log-binomial.GEE模型的适用性并进行比较。结果 表明log.binomial-GEE模型与稳健Poisson-GEE模型的收敛率基本均为100%,两模型估计各参数的平均值均与真值接近;在类内聚集性变小或类别数增加时,两模型估计各参数的95%CI覆盖率均有所提高;稳健Poisson.GEE模型对参数估计的稳健性较好,应用到实例时可正确评价暴露对结局的影响。稳健Poisson和log.binomial的GEE模型很少存在收敛问题,且有较高的准确率,可用于流行病学资料中非独立数据的RR/PR值估计。