基于g-computation控制纵向数据未测混杂因素的因果推断模型构建及应用研究

At present, unmeasured confounders and time-dependent confounders are the main influence factors of accuracy of causal inference. There are few models that can effectively control the effects of these confounders. As a result, they have become the important bottleneck problems for causal inference in the longitudinal study. In this study, a new causal inference approach was proposed: g-computation multivariate joint mixed-effects model. In this model, unmeasured confounders were controlled as random effects. Series correlation was employed to address the cross impact of time-dependent confounders. Multivariate joint mixed-effects model was constructed with outcome variable model, time-dependent exposure model, and time-dependent confounder model sharing the common random effect. Adaptive Gauss-Hermite quadrature was used to estimate the parameters of model. Then, g-computation was used to construct the causal function of estimators, which could infer the causal effect by integrating outcome expected values over covariate histories. Lastly, this model was applied to the longitudinal data of follicle-stimulating hormone commonly regulating and controlling the bone mass and fat energy metabolism in postmenopausal women. The causal effect of follicle-stimulating hormone was approximatively unbiasedly estimated. Therefore, the theoretical innovation would be translated into actual value. In conclusion, this study could not only extend the causal inference approaches in the longitudinal study, but also provide the common prevention target of osteoporosis and visceral fat accumulation. Therefore, the results of this study were very significant for theoretical research and valuable for practical application.

未测混杂因素和依时混杂因素是影响因果推断准确性的主要因素，目前尚无有效控制其影响的统计模型，使其成为因果推断亟待解决的重要瓶颈问题。本研究提出一种新的因果推断方法——g-computation多元联合混合效应模型：该模型将未测混杂因素作为随机效应进行控制，通过序列相关系数控制依时混杂因素的交叉影响，将结果变量模型、依时暴露因素模型和依时混杂因素模型共享同一随机效应构建多元联合混合效应模型，利用自适应高斯求积法进行参数估计；通过g-computation对模型参数估计值构建因果关系函数，对结果变量条件期望值进行积分运算作出因果推断；最后，将该模型应用于估计绝经期女性FSH综合调控骨量及脂肪能量代谢的因果效应，将理论创新转化为实际价值。因此，本研究拓展纵向研究因果推断方法的同时为骨质疏松和内脏脂肪堆积提供共同预防靶点，具有重要的理论意义和实际应用价值。

在新时代的精准医学和健康医疗大数据背景下，真实世界研究下的观察性研究的近似无偏因果推断策略越来越受到重视。纵向研究中，因果推断主要受以下几种因素影响：i. 依时暴露因素；ii. 依时混杂因素；iii. 缺失数据和失访；iv. 未测混杂因素。因此，如何有效控制因果推断的主要影响因素，尤其是依时混杂因素和未测混杂因素，给出近似无偏因果推断是纵向研究统计模型亟待解决的重要瓶颈问题！针对纵向研究中因果推断存在的问题，本研究提出一种新的基于g-computation的因果推断方法——g-computation多元联合混合效应模型(multivariate joint mixed-effects model，MJMM)，该模型利用随机效应和序列相关等措施控制未测混杂因素和依时混杂因素的影响，通过g-computation对模型参数估计值构建因果关系函数，从而作出近似无偏因果推断；同时将该模型用于估计卵泡刺激素调控骨量及脂肪能量代谢的因果效应，为骨质疏松和内脏脂肪堆积提供共同预防靶点。本研究通过模拟实验评价g-computation多元联合混合效应模型在控制纵向研究中未测混杂因素进行因果推断方面的性能特点，结果显示g-computation多元联合混合效应模型可有效控制纵向研究中的未测混杂因素进行因果推断，且因果效应估计具有较好的准确性，但稳定性略差。利用该模型分析卵泡刺激素与骨密度和肥胖的因果关系，证实血清卵泡刺激素水平可促进骨量流失，增加骨折的风险，且在绝经后期女性中更为明显；另一方面，血清卵泡刺激素和黄体生成素水平可增加身体脂肪含量，促进脂肪堆积，同时可促进身体肌肉含量流失，从而改变身体成分，增加肥胖的发生风险。因此，本研究针对因果推断中存在的主要问题提出新的统计模型，为纵向研究因果推断方法提供新思路，同时将其应用于卵泡刺激素综合调控骨量及脂肪能量代谢的因果推断中，将理论创新转化为实际应用。本研究既可拓展纵向研究因果推断方法，也能为预防骨质疏松和内脏脂肪堆积提供共同靶点，此为统计学方法与医学实际问题的交叉和融合，具有重要的理论意义和实际应用价值。