arXiv:2409.11265v1 Announce Kind: cross
Summary: Background: Superior strategies for causal inference, reminiscent of focused most chance estimation (TMLE), require sure situations for statistical inference. Nonetheless, in conditions the place there may be not differentiability as a consequence of knowledge sparsity or near-positivity violations, the Donsker class situation is violated. In such conditions, TMLE variance can endure from inflation of the sort I error and poor protection, resulting in conservative confidence intervals. Cross-validation of the TMLE algorithm (CVTMLE) has been prompt to enhance on efficiency in comparison with TMLE in settings of positivity or Donsker class violations. We purpose to research the efficiency of CVTMLE in comparison with TMLE in varied settings.
Strategies: We utilised the data-generating mechanism as described in Leger et al. (2022) to run a Monte Carlo experiment underneath completely different Donsker class violations. Then, we evaluated the respective statistical performances of TMLE and CVTMLE with completely different tremendous learner libraries, with and with out regression tree strategies.
Outcomes: We discovered that CVTMLE vastly improves confidence interval protection with out adversely affecting bias, significantly in settings with small pattern sizes and near-positivity violations. Moreover, incorporating regression bushes utilizing customary TMLE with ensemble tremendous learner-based preliminary estimates will increase bias and variance resulting in invalid statistical inference.
Conclusions: It has been proven that when utilizing CVTMLE the Donsker class situation is not essential to acquire legitimate statistical inference when utilizing regression bushes and underneath both knowledge sparsity or near-positivity violations. We present by way of simulations that CVTMLE is way much less delicate to the selection of the tremendous learner library and thereby offers higher estimation and inference in instances the place the tremendous learner library makes use of extra versatile candidates and is liable to overfitting.
Supply hyperlink