标题： 重新思考分布性工具变量：KAN驱动的D-IV-LATE与模型选择 显示英文标题

标题： Rethinking Distributional IVs: KAN-Powered D-IV-LATE & Model Choice

摘要： 双倍/去偏机器学习（DML）框架已成为现代因果推断的基石，使研究者能够利用灵活的机器学习模型来估计非关键函数，而不会在最终参数估计中引入一阶偏差。 然而，对于非关键函数的机器学习模型选择通常被视为一个次要的实现细节。 在本文中，我们主张这一选择可能对分析的实质性结论产生深远影响。 我们通过呈现和比较两种不同的分布工具变量局部平均处理效应（D-IV-LATE）估计量来证明这一点。 第一个估计量利用标准机器学习模型（如随机森林）来估计非关键函数，而第二个估计量是一种新颖的估计量，使用了科尔莫戈罗夫-阿诺德网络（KANs）。 我们建立了这些估计量的大样本性质，并通过蒙特卡洛模拟评估它们的性能。 一项分析401(k)参与对净金融资产分布影响的经验应用显示，非关键函数的机器学习模型选择可以显著改变实质性结论，基于KAN的估计量表明治疗效果的异质性更为复杂。 这些发现强调了一个重要的警告：非关键函数估计器的选择不是一个简单的实现细节。 相反，这是一个关键的选择，可能对因果推断的研究结果产生深远影响。 显示英文摘要

摘要： The double/debiased machine learning (DML) framework has become a cornerstone of modern causal inference, allowing researchers to utilise flexible machine learning models for the estimation of nuisance functions without introducing first-order bias into the final parameter estimate. However, the choice of machine learning model for the nuisance functions is often treated as a minor implementation detail. In this paper, we argue that this choice can have a profound impact on the substantive conclusions of the analysis. We demonstrate this by presenting and comparing two distinct Distributional Instrumental Variable Local Average Treatment Effect (D-IV-LATE) estimators. The first estimator leverages standard machine learning models like Random Forests for nuisance function estimation, while the second is a novel estimator employing Kolmogorov-Arnold Networks (KANs). We establish the asymptotic properties of these estimators and evaluate their performance through Monte Carlo simulations. An empirical application analysing the distributional effects of 401(k) participation on net financial assets reveals that the choice of machine learning model for nuisance functions can significantly alter substantive conclusions, with the KAN-based estimator suggesting more complex treatment effect heterogeneity. These findings underscore a critical "caveat emptor". The selection of nuisance function estimators is not a mere implementation detail. Instead, it is a pivotal choice that can profoundly impact research outcomes in causal inference.