Rejoinder to "Reader reaction to 'Outcome-adaptive Lasso: Variable selection for causal inference' by Shortreed and Ertefaie (2017)".
Rejoinder to "Reader reaction to 'Outcome-adaptive Lasso: Variable selection for causal inference' by Shortreed and Ertefaie (2017)".
复制标题
DOI：
10.1111/biom.13681
复制
发表时间：
2023-03
期刊：
影响因子：
1.9
通讯作者：
Shortreed, Susan M.
中科院分区：
文献类型：
--
作者：
Jones, Jeremiah;Ertefaie, Ashkan;Shortreed, Susan M.
我们赞赏Balde，Yand和列费夫尔（BYL）的有用贡献的文献变量选择的因果推理。广义结果自适应套索（GOAL）提供了一个有趣的除了结果自适应套索（OAL）程序，和工作中提出的模拟演示了几个奇怪的趋势。通过实施复制实验，我们已经了解了更多关于OAL的行为，以及它的广义类似物（即，目标）在某些有问题的情况下。由BYL设计的模拟展示了从增加1002惩罚中获得的明显优势。OAL在这些情况下遭受的偏见水平的增加，以及随后通过所提出的方法进行的补救，引发了我们的好奇心。虽然毫无疑问，如果处理不当，高水平的相关性可能会产生有害影响，但在相对温和的相关性水平下存在这种影响是令人惊讶的。在这个反驳，我们证明，这些问题的偏见可能会被解释为存在的近积极的违规行为，这是一个后果的相关参数。我们给出了一个解释，连接额外的脊罚截断，这提供了一个启发式的解释所观察到的行为。最后，我们探讨了可能的方法，结合OAL或目标，这是更强大的积极性问题。
We applaud Balde, Yand, and Lefebvre (BYL) for their useful contribution to the literature on variable selection in causal inference. The generalized outcome adaptive lasso (GOAL) provides an interesting addition to the outcome-adaptive Lasso (OAL) procedure, and the simulations presented within the work demonstrate several curious trends. Through implementing replication experiments, we have come to understand much more about the behavior of the OAL as well as its generalized analog (ie, GOAL) under certain problematic scenarios. The simulations laid out by BYL demonstrated a clear advantage gained from the addition of 𝓁2 penalization. The increased levels of bias suffered by OAL in these scenarios, and their subsequent remediation through the proposed method, sparked our curiosity. While there is little doubt that high levels of correlation may provide deleterious effects if not handled properly, the presence of such effects at relatively mild correlation levels was surprising. In this rejoinder, we demonstrate that these issues of bias may be explained by the presence of near-positivity violations which are a consequence of the correlation parameter. We give an interpretation that connects the additional ridge penalty to truncation, which provides a heuristic explanation for the observed behavior. Finally, we explore possible methods incorporating OAL or GOAL which are more robust to the positivity issue.
登录
查看更多内容
DOI：
10.1080/01621459.2016.1260466
发表时间：
2018-01-01
影响因子：
3.7
作者：
Li, Fan;Morgan, Kari Lock;Zaslavsky, Alan M.
通讯作者：
Zaslavsky, Alan M.
影响因子：
5
作者：
Cole, Stephen R.;Hernan, Miguel A.
通讯作者：
Hernan, Miguel A.
影响因子：
6.1
作者：
ROBINSON, PM
通讯作者：
ROBINSON, PM
影响因子：
1.1
作者：
van de Geer, Sara A.;Buehlmann, Peter
通讯作者：
Buehlmann, Peter
DOI：
10.1111/j.1467-9868.2012.01040.x
发表时间：
2013-03
期刊：
Journal of the Royal Statistical Society. Series B, Statistical methodology
影响因子：
--
作者：
Ma Y;Zhu L
通讯作者：
Zhu L