Tuning LASSO Models for Propensity Score Weighting and Using Synthetic Negative Control Exposures for Residual Bias Detection

调整 LASSO 模型以进行倾向评分加权并使用合成阴性对照暴露进行残余偏差检测

阅读:4

Abstract

The propensity score (PS) is widely used to control for large numbers of covariates in high-dimensional healthcare database studies. In these settings, the least absolute shrinkage and selection operator (LASSO) is commonly used to estimate the PS, with the regularization parameter typically chosen by cross-validation to optimize out-of-sample prediction. For PS weighting, however, theory and simulations have shown that prediction-based tuning can over-regularize the PS model and that less regularization ("undersmoothing") is needed to minimize bias in PS weighted estimators. The optimal degree of undersmoothing can be guided by the efficient influence function of the target causal parameter. In practice, however, the efficient influence function is often unknown or difficult to derive. We therefore investigate balance metrics as a simple, broadly applicable approach for selecting the degree of undersmoothing when the efficient influence function is unavailable. Because balance-based tuning does not guarantee minimal bias in PS weighted estimators-as such metrics are blind to the efficient influence function-we also propose a framework for generating synthetic negative control exposures for bias detection. We show that synthetic negative control exposures can identify analyses that likely violate partial exchangeability due to inadequate control for measured confounding. In numerical studies, balance-based undersmoothing consistently reduced bias relative to standard cross-validation, while synthetic negative control exposures effectively flagged analyses with residual confounding. Together, these approaches provide practical tools for improving PS weighting analyses when theory-driven tuning is infeasible.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。