存在时依性混杂时的G方法

仇沁晓; 尤东方; 赵杨

doi:10.16462/j.cnki.zhjbkz.2021.06.002

存在时依性混杂时的G方法

doi: 10.16462/j.cnki.zhjbkz.2021.06.002

仇沁晓¹,
尤东方¹,
赵杨^{1, 2, ,}

1.
211166 南京，南京医科大学公共卫生学院生物统计学系
2.
211166 南京，南京医科大学生物医学大数据重点实验室，肿瘤个体化医学协同创新中心

基金项目:

国家重点研发计划 2016YFC1000207

国家自然科学基金 81872709

江苏省高等学校自然科学研究重大项目 18KJA110004

详细信息

通讯作者:
赵杨，E-mail：yzhao@njmu.edu.cn

中图分类号: R195.1
计量
- 文章访问数: 436
- HTML全文浏览量: 307
- PDF下载量: 109
- 被引次数: 0
出版历程
- 收稿日期: 2021-04-28
- 修回日期: 2021-05-28
- 刊出日期: 2021-06-10

G-methods in the existence of time varying confounding

1.
Department of Biostatistics, School of Public Health, Nanjing Medical University, Nanjing 211166, China
2.
Key Laboratory of Biomedical Big Data, Nanjing Medical University, Collaborative Innovation Center for Individual Medicine in Cancer, Nanjing 211166, China

Funds:

National Key Research and Development Program of China 2016YFC1000207

National Natural Science Foundation of China 81872709

Key Project of University Natural Scientific Research of Jiangsu Province 18KJA110004

More Information

Corresponding author: ZHAO Yang, E-mail: yzhao@njmu.edu.cn

摘要

摘要: 目的介绍处理时依性混杂的G方法，并对不同G方法进行探讨和比较。方法通过4个情境的模拟试验验证不同G方法在不同情境下对时依性混杂的处理效果，并应用英国生物样本库(UK Biobank)的数据集进行实例分析。结果模拟试验和实例分析结果均显示G方法能有效处理时依性混杂。模拟试验显示3种方法效果类似，G-computation易受G-null paradox的影响。随着时依性混杂因素数量增加，相比于G-computation和G-estimation，逆概率加权法(inverse probability of treatment weighting, IPTW)的效果波动较大。结论不同G方法都能适当地处理时依性混杂，降低统计分析过程中的偏倚大小。
- 时依性混杂 /
- 碰撞偏倚 /
- G方法
Abstract: Objective To introduce and compare different G-methods which can deal with time varying confounding. Methods The simulation experiments of four scenarios were carried out to verify the effects of different G-methods on time varying confounding in different situations. Dataset from UK Biobank was then analyzed using different G-methods. Results All three G methods can effectively deal with time varying confounding with similar performance, while G-computation was vulnerable to G-null paradox. However, with the increasing number of time varying confounders, the estimated effects of inverse probability of treatment weighting (IPTW) were more variable. Conclusion All of the three G-methods can remove the bias resulted from time varying confounding appropriately.
- Time varying confounding /
- Collider bias /
- G-methods

HTML全文

图 1 时依性混杂的有向无环图(A：肥胖对糖尿病影响的研究；B：存在时依性混杂因素的研究)

注：U是未知混杂因素，L是时依性混杂因素，A是处理或暴露因素，Y是结局。

Figure 1. A directed acyclic graph of time varying confounding

(A: Study on the effect of obesity on diabetes mellitus; B: Study on the existence of time varying confounding)

下载: 全尺寸图片幻灯片

图 2 模拟试验的有向无环图(A：情境一；B：情境二；C：情境三；D：情境四)

Figure 2. Directed acyclic graphs of simulation (A: Scenario 1; B: Scenario 2; C: Scenario 3; D: Scenario 4)

下载: 全尺寸图片幻灯片

图 3 模拟试验情境二结果

Figure 3. Results of the second scenario of simulation

下载: 全尺寸图片幻灯片

图 4 模拟试验情境四结果

Figure 4. Results of the fourth scenario of simulation

下载: 全尺寸图片幻灯片

表 1 模拟试验情境一结果

Table 1. Results of the first scenario of simulation

样本量	方法	偏倚	MSE	95% CI覆盖率(%)	检验效能(%)
400	调整	0.140 8	0.029 7	0.00	100.00
	未调整	-0.000 3	0.039 7	95.80	100.00
	G-computation	-0.000 6	0.039 6	94.90	100.00
	IPTW	0.105 1	1.138 2	100.00	67.60
	G-estimation ^a	-0.066 3		90.70	99.30
1 000	调整	0.140 6	0.029 9	0.00	100.00
	未调整	-0.000 4	0.039 8	93.40	100.00
	G-computation	0.000 3	0.039 8	94.60	100.00
	IPTW	0.146 1	1.151 8	100.00	97.30
	G-estimation ^a	-0.063 1		86.50	100.00
2 000	调整	0.140 8	0.030 0	0.00	100.00
	未调整	-0.000 1	0.040 0	94.20	100.00
	G-computation	0.000 0	0.040 0	94.80	100.00
	IPTW	0.143 9	1.153 3	100.00	100.00
	G-estimation ^a	-0.062 4		78.30	100.00
注：^a此情境下，G-estimation无法得到预测值，因此无法估计MSE。

下载: 导出CSV

表 2 模拟试验情境三结果

Table 2. Results of the third scenario of simulation

效应	方法	结果1 ^a				结果2 ^b				结果3 ^c
效应	方法	偏倚	MSE	95% CI覆盖率(%)	Ⅰ类错误率(%)	偏倚	MSE	95% CI覆盖率(%)	Ⅰ类错误率(%)	偏倚	MSE	95% CI覆盖率(%)	Ⅰ类错误率(%)
A₀	调整	0.000 8	0.262 6	96.60	4.80	-0.108 0	0.263 2	48.90	51.10	-0.106 0	0.860 2	84.20	15.80
	未调整	0.131 9	0.275 2	35.50	24.60	0.078 4	0.274 5	64.60	35.40	0.080 4	0.872 1	84.30	15.70
	G-computation ^d	0.000 3		93.40	6.60	-0.000 2		94.00	6.00	-0.002 6		94.60	5.40
	IPTW	0.000 3	0.278 9	93.30	6.70	-0.000 2	0.279 7	94.70	5.30	-0.002 6	0.879 9	94.80	5.30
	G-estimation ^d	0.000 0		94.30	5.70	0.003 9		94.80	5.20	0.005 4		94.70	5.20
A₁	调整	-0.002 4	0.262 6	94.10	5.60	-0.007 2	0.263 2	93.00	7.00	-0.008 7	0.860 2	94.90	5.10
	未调整	-0.065 8	0.275 2	74.30	13.10	-0.061 9	0.274 5	75.20	24.80	-0.064 3	0.872 1	87.10	12.90
	G-computation ^d	-0.002 6		93.10	6.90	-0.000 7		95.40	4.60	0.001 6		94.50	5.50
	IPTW	-0.002 6	0.278 9	94.40	5.60	-0.000 7	0.279 4	97.00	3.00	0.001 6	0.879 9	94.60	5.40
	G-estimation ^d	-0.000 8		95.00	5.00	0.001 7		95.40	4.60	0.000 5		94.40	5.40
注：^a结果1：A₀对L的效应为0；^b结果2：A₀对L的效应为0.3，标准差为0.2；^c结果3：A₀对L的效应为0.3，标准差为0.8。^d此情境下，G-computation和G-estimation无法得到预测值，因此无法估计MSE。

下载: 导出CSV

表 3 实例分析结果

Table 3. Results of real data analysis

方法	第一阶段BMI		第二阶段BMI
方法	β值	OR值	β值	OR值
调整	0.513	1.670	-0.195	0.823
未调整	0.534	1.706	-0.171	0.843
G-computation	0.599	1.820	-0.028	0.972
IPTW	0.765	2.149	0.325	1.384
G-estimation	0.108	1.114	0.282	1.326

下载: 导出CSV

参考文献(18)

[1]	Robins JM, Hernán MA, Brumback B. Marginal structural models and causal inference in epidemiology[J]. Epidemiology, 2000, 11(5): 550-560. DOI: 10.1097/00001648-200009000-00011.
[2]	Fewell Z, Hernán MA, Wolfe F, et al. Controlling for time-dependent confounding using marginal structural models[J]. Stata Journal, 2004, 4(4): 402-420. DOI: 10.1177/1536867X0400400403.
[3]	Schisterman EF, Cole SR, Platt RW. Overadjustment bias and unnecessary adjustment in epidemiologic studies[J]. Epidemiology, 2009, 20(4): 488-495. DOI: 10.1097/EDE.0b013e3181a819a1.
[4]	Greenland S. Quantifying biases in causal models: classical confounding vs collider-stratification bias[J]. Epidemiology, 2003, 14(3): 300-306. DOI: 10.1088/0256-307X/25/1/054.
[5]	Mansournia MA, Etminan M, Danaei G, et al. Handling time varying confounding in observational research[J]. BMJ, 2017, 359: j4587. DOI: 10.1136/bmj.j4587.
[6]	Hernán MA, Brumback B, Robins JM. Marginal structural models to estimate the causal effect of zidovudine on the survival of HIV-positive men[J]. Epidemiology, 2000, 11(5): 561-570. DOI: 10.1097/00001648-200009000-00012.
[7]	Daniel RM, Cousens SN, De Stavola BL, et al. Methods for dealing with time-dependent confounding[J]. Stat Med, 2013, 32(9): 1584-1618. DOI: 10.1002/sim.5686.
[8]	Robins JM. A new approach to causal inference in mortality studies with a sustained exposure period-application to control of the healthy worker survivor effect[J]. Mathematical Modelling, 1986, 7(9-12): 1393-1512. DOI: 10.1016/0270-0255(86)90088-6.
[9]	Robins JM, Blevins D, Ritter G, et al. G-estimation of the effect of prophylaxis therapy for pneumocystis carinii pneumonia on the survival of AIDS patients[J]. Epidemiology, 1992, 3(4): 319-336. DOI: 10.1097/00001648-199207000-00007.
[10]	Austin PC. The use of propensity score methods with survival or time-to-event outcomes: reporting measures of effect similar to those used in randomized experiments[J]. Stat Med, 2014, 33(7). DOI: 10.1002/sim.5984.
[11]	Cole SR, Hernán MA. Adjusted survival curves with inverse probability weights[J]. Comput Methods Programs Biomed, 2004, 75(1): 45-49. DOI: 10.1016/j.cmpb.2003.10.004.
[12]	Hernan MA, Robins JM. Estimating causal effects from epidemiological data[J]. J Epidemio Community Health, 2006, 60(7): 578-586. DOI: 10.1136/jech.2004.029496.
[13]	Whitcomb BW, Schisterman EF, Perkins NJ, et al. Quantification of collider-stratification bias and the birthweight paradox[J]. Paediatr Perinat Epidemiol, 2009, 23(5): 394-402. DOI: 10.1111/j.1365-3016.2009.01053.x.
[14]	Hernán M, Robins J. Causal inference: what if[M]. Boca Raton: Chapman & Hall/CRC, 2020.
[15]	National Institutes of Health. Clinical guidelines on the identification, evaluation and treatment of overweight and obesity in adults-the evidence report[J]. Obes Res, 1998, 6(6): 464. http://ci.nii.ac.jp/naid/10018714807
[16]	Obesity S. Obesity: preventing and managing the global epidemic. Report of a WHO consultation[J]. World Health Organ Tech Rep Ser, 2000, 894(1): 18-30. DOI: 10.1002/jps.3080150106.
[17]	Xu S, Ross C, Raebel MA, et al. Use of stabilized inverse propensity scores as weights to directly estimate relative risk and its confidence intervals[J]. Value Health, 2010, 13(2): 273-277. DOI: 10.1111/j.1524-4733.2009.00671.x.
[18]	Funk MJ, Westreich D, Wiesen C, et al. Doubly robust estimation of causal effects[J]. Am J Epidemiol, 2011, 173(7): 761-767. DOI: 10.1093/aje/kwq439.