存在混杂时高维数据的随机森林分析

目的：探讨存在混杂因素时高维数据中随机森林（random forest，RF）的分析方法。方法：通过模拟实验和实例数据分析对单纯随机森林分析、增加节点候选变量为最大值以及基于广义线性模型的残差校正混杂因素的结果进行比较，以重要变量的重要性评分排序情况进行评价。结果：模拟实验表明，增加节点候选变量的方法对混杂因素的校正效果不明显，而基于广义线性模型残差的方法能有效校正混杂效应；实际数据分析结果显示单纯随机森林分析rs3754686和rs2322660分别排在第一和第二位。增加节点候选变量后rs3754686排序变化较小，而基于残差的方法校正人群分层后这两个单核苷酸多态位点（SNPs）的排序大幅度降低，从而打破乳糖酶（LCT）基因与身高之间的虚假关联。结论：随机森林分析需要考虑混杂因素问题，基于广义线性模型的残差能有效校正混杂因素，适用于高维数据的变量筛选。

Random Forest analysis of high-dimensional data with the confounding effects

Objective This project explored the Random Forest analysis of high-dimensional data with the confounding effects. Methods We used computer simulations and real data validation to evaluate the performance of 2 methods which can potentially account for the confounding effects in RF analysis: RF analysis with maximum candidate variables at each split (RFMCV) and RF with glm-based correction. The distribution of ranks of the causal variable was used to evaluate these approaches. Results Simulation experiments suggested that RF with glm-based correction was more effective than the RFMCV to correct the confounding effects. Analysis results of GWAS data confirmed that RF with glm-based correction can effectively remove the spurious association between the LCT gene and height. Conclusion The confounding effects should be correctly adjusted in RF analysis. RF with glm-based correction was applicable to adjust the confounding effects and variable selection in high-dimensional data.