Collaborative Research: Deep Inference - Artificial Intelligence for Structural Estimation

为了评估经济政策的影响，如扩大医疗保健覆盖范围或最低工资的变化，经济学家使用结构模型，它有力地描述了工作机制，但估计结构模型通常具有挑战性。估计这类结构模型的主要工具是通过模拟进行推理。对于模型的不同参数化，通过模型生成合成数据，并使用最接近观测数据的参数生成数据作为估计。最近的现代人工智能方法，如用于图像识别的深度学习，就是基于同样的原理。这些方法在过去几年中取得了令人印象深刻的成果。因此，本研究利用现代模式识别中这种强大的工具来进行经济学中的结构估计。本研究考虑了一种设置，其中个体结果是外生变量的已知函数，其分布已知到参数的有限维向量。目标是估计有限维参数。研究者采用生成对抗网络方法（GANs）来寻找参数值，使得给定一个鉴别器，一个可以准确区分使用模型生成的数据与实际数据的设备，在根据该参数值生成数据时无法做到这一点。该方法与其他基于仿真的最小距离估计方法的不同之处在于距离是自适应的。也就是说，鉴别器学习最能区分真实数据和合成数据的数据特征，而不是硬编码要匹配的数据特征。这种自适应特性在模式识别任务中被证明是强大的。在结构估计中，适应性可以转化为减轻维度的诅咒，并获得能够更紧密地匹配整个数据分布的参数，而不是一组预先指定的矩。这种估计框架在分配效应和异质性是第一阶的应用中应该是有用的，以评估特定政策的效果。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

In order to evaluate the effect of economic policies such as extended health care coverage or changes in the minimum wage, economists use structural models which powerfully describe the mechanism at work, but estimating structural models is typically challenging. A main tool for estimating such structural models is inference via simulation. For different parametrizations of the model, synthetic data is generated via the model, and the parameters generating data that most closely resembles observed data are used as estimates. Recent modern artificial intelligence methods such as deep learning for image recognition are based on this same principle. These methods have been achieving impressive results over the past years. Therefore, this research takes advantage of such powerful tools in modern pattern recognition for structural estimation in economics. This research considers a set-up where individual outcomes are a known function of exogenous variables and an error whose distribution is known up to a finite dimensional vector of parameters. The goal is to estimate the finite dimensional parameter. The investigators adopt the generative adversarial network approach (GANs) to find the parameter value such that given a discriminator, a device that can accurately distinguish data generated using the model from real data, is unable to do so when the data is generated according to such parameter value. The method developed in this research differs from other simulation-based minimum distance estimators in that the distance is adaptive. That is, the discriminator learns the features of the data that are best at distinguishing real from synthetic data as opposed to hard-coding what features of the data to match. This adaptability property has proven powerful in pattern recognition tasks. In structural estimation, adaptability can translate into alleviating the curse of dimensionality, and obtaining parameters that are able to more closely match entire distributions of data, as opposed to a set of pre-specified moments. This estimation framework should be useful in applications were distributional effects and heterogeneity are first order to evaluate the effect of a particular policy.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.