效应改变法和有向无环图在线性回归变量筛选中的应用及实现

Applications and realizations of change-in-estimate and directed acyclic graph in linear regression variable screen

目的 通过实例展示如何应用在线工具DAGitty和R实现有向无环图(DAG)联合效应改变法(CIE)筛选线性回归中需要调整的混杂。方法 基于DATADRYAD的Jung, Su-Young J eettaall.(2019)数据包,假设暴露因素为BMI_ICU,结局为Death_28D,混杂包括Sex等9个变量为混杂因素。本文以逻辑回归模型为例,首先应用在线工具DAGitty将变量间的因果关系可视化,并自动识别需要调整的混杂;然后使用R批量筛选出使暴露因素的效应量变化超过10%的混杂。结果 经DAG未发现中介变量,在单因素模型中逐一加入混杂,或在全变量模型中逐一剔除混杂后BMI_ICU的效应量变化均未超过10%,综上,暂没有需要调整的混杂。结论 通过DAGitty和R可以简便、有效、快速的实现DAG联合CIE筛选线性回归中的混杂,其他线性回归类型可在此基础上稍作修改。

Objective To display how to use online tools-DAGitty and R to realize directed acyclic graph(DAG)combined with change-in-estimate(CIE)for screening confounders need to be adjusted in linear regression.Methods It was assumed that exposure factor was BMI_ICU,outcome was Death_28D and there were 9 confounder factors including Sex based on data packages of Jung and Su-Young J et al.(2019)from DATADRYAD.Taken aet al logistic regression model as an example,firstly an online tool,DAGitty,was applied to visualize the cause-and-effect relationships between variables,and to automatically identify the confounders needed to be adjusted.R was then used for batch screening confounders which changed estimate size of exposure factor over 10%.Results There was no mediating variable found by DAG.when confounders were sequentially added to the univariate model or removed from the full-variable model,the effect size of BMI_ICU changed by less than 10%.In summary,there were no confounders that required to be adjusted.Conclusion The screen of confounders in linear regression can be easily,effectively and quickly conducted by DAG combined with CIE through DAGitty and R.The code and method of other linear regression models need be slightly modified on this basis.