
经济试验的协变量平衡自适应设计
Covariate-adjusted Randomization Design for Economic Experiment
在经济学领域的随机对照试验中,在经济学领域的随机对照试验中,受试者的分配通常按照完全随机的方式进行.然而,完全随机化可能无法使基线协变量在试验组和对照组间的分布均衡可比,导致试验的解释性与准确性降低,甚至会得出错误的分析结果.本文在经济学随机对照试验中引入协变量平衡自适应设计,该设计在分配过程中自适应地对协变量的平衡性进行调整,从而能够获得协变量在组间分布相对均衡的分配方案.本文基于一项探究个性化信息是否能够影响养老金个人账户储蓄的随机对照试验案例,分析比较了不同随机化设计对于在随机化试验中的协变量平衡以及处理效应估计等方面的影响.实证分析表明,相比于完全随机化,考虑协变量平衡调整的随机化设计能够降低组间的协变量的不平衡程度,并提高后续对于平均处理效应的估计精度和检验功效.
In economic randomized controlled trials, subjects are often assigned by complete randomization. However, under complete randomization, the distribution of baseline covariates between treatment and control groups is usually incomparable, which decreases interpretability and accuracy of the experiment, or even distorts the results. In this paper, we introduce the covariate-adjusted randomization design for the economic randomized controlled trial. The covariate-adjusted randomization design adaptively adjusts the covariates balance during the allocation process so as to achieve the better covariate balance. Based on a randomized controlled trial investigating whether personalized information can affect pension savings, we compare the impact of three different randomizations on the covariates balance and estimation of the average treatment effect. Empirical analysis results show that, compared to complete randomization, covariate-adjusted randomization design can significantly reduce the covariate imbalance and thus improve the subsequent estimation precision and testing power.
随机对照试验 / 自适应随机化 / 马氏距离 / 因果推断 / 检验功效 {{custom_keyword}} /
randomized controlled trial / adaptive randomization / Mahalanobis distance / causal inference / test power {{custom_keyword}} /
表1 协变量的描述与检验 |
协变量 | 定义 | 对照组均值 | 试验组均值 | K-S检验 | |
edad | 年龄 | 39.021 | 37.451 | 0.002 | 0.044 |
rem_actual | 前六个月的平均工资 | 443.443 | 482.189 | 0.025 | 0.000 |
deseada_pens | 希望获得的养老金 | 504.510 | 580.046 | 0.325 | 0.995 |
esperada_pens | 预期能获得的养老金 | 249.946 | 293.929 | 0.309 | 0.601 |
saldo_uf | 强制账户 | 379.505 | 423.898 | 0.126 | 0.878 |
comodidad | 对养老金系统的熟悉程度 | 4.784 | 4.724 | 0.405 | 0.989 |
TotCotVol | 过去一年自愿缴纳养老金金额 | 20.731 | 32.783 | 0.388 | 0.965 |
TotCot | 过去一年强制缴纳养老金金额 | 428.165 | 441.202 | 0.524 | 0.562 |
NCotVol | 过去一年自愿缴纳养老金月数 | 0.396 | 0.446 | 0.549 | 0.983 |
NCot | 过去一年强制缴纳养老金月数 | 7.884 | 8.038 | 0.434 | 0.770 |
res_sim_1 | 模拟器计算能获得的养老金 | 263.551 | 276.069 | 0.322 | 0.034 |
ErrorPension | 预期与模拟的差距 | -17.860 | 13.605 | 0.476 | 0.160 |
表2 不同随机化方法下的马氏距离均值和差异性检验结果(以年龄为例) |
随机化方法 | 马氏距离均值 | 差异显著性检验 | |
K-S检验 | |||
CR | 21.84 | 3.63% | 5.08% |
RR | 7.97 | 0.00% | 0.00% |
ARM | 1.82 | 0.00% | 0.00% |
表3 平均处理效应估计与协变量调整 |
均值差异估计 | 协变量校正的回归估计 | ||||||
处理效应 | 处理效应 | TotCot | edad | rem_actual | saldo_uf | ||
估计 | -0.367 | -2.129 | 0.066 | -0.105 | 0.017 | 0.002 | |
0.842 | 0.012 | 0.000 | 0.009 | 0.000 | 0.019 | ||
0.000 | 0.794 |
表4 平均处理效应的渐近标准差、MSE与检验功效 |
随机化 | power(with cov) | power(no cov) | ||||
CR | ||||||
RR | ||||||
ARM |
刘雅琦, 刘曼, 刘玉秀, 陈文松, 袁阳丹, 《新英格兰医学杂志》临床试验随机化方法应用状况评价[J]. 中国循证医学杂志, 2021, 21 (3): 303- 307.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
殷赫然, 刘健, 樊小农, 沈燕, 王舒, 最小随机化在国内针灸临床研究中的应用与思考[J]. 中华中医药杂志, 2021, 36 (1): 76- 79.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
Pinto J M, (2012). Does Personalized Pension Projection Affect the Retirement Decision?[R]. Superintendencia de Pensiones.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
Qin Y, Li Y, Ma W, Yang H, Hu F, (2022). Adaptive Randomization via Mahalanobis Distance[J/OL]. Statistica Sinica. Doi: 10.5705/ss.202020.0440.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
Schlosser A, Shanan Y, (2022). Fostering Soft Skills in Active Labor Market Programs: Evidence from a Large-Scale RCT[R]. CESifo Working Paper Series.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_ref.label}} |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
/
〈 |
|
〉 |