Demand Analysis with Many Prices: Methods and Application

本研究开发了机器学习方法，从大数据中估计经济福利。在杂货店和其他零售商店收集的扫描仪数据提供了可用于估计经济福利的大数据。研究者基于大数据开发了新的双机器学习经济福利估计器。这些估计器将特定经济权重的新型机器学习与需求函数的机器学习估计相结合，实现了福利的双重机器学习估计。这些估计量也被推广并应用于许多其他问题。此外，这项研究利用了扫描仪数据随时间跟踪个体的事实。因此，对个人需求函数进行估计和平均，以构建改进的福利措施。本研究的目的是为包括许多价格的大型数据集（如扫描仪数据）开发和应用经济需求分析。扫描器数据的一个共同特征是交叉价格效应往往很小，通常比自身价格效应小一个数量级。这一特征表明，在大多数交叉价格效应很小的情况下，允许近似稀疏性的机器学习方法可能在实践中有用。本研究开发了精确的消费者剩余和其他福利效应的双重机器学习估计器。研究者在Riesz表示中使用了新的机器学习对象，而不是条件期望。本文提出了一种具有一阶级数估计的广义矩量方法的双机器学习的一般方法。扫描仪数据通常是面板数据，其中个人或家庭被跟踪一段时间。研究者进一步在具有一般异质性的面板数据中推导出需求的识别结果，并分析了可用于需求估计的平均效应的正则化固定效应面板数据估计。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。

This research develops machine learning methods to estimate economic welfare from big data. Scanner data, as collected in grocery and other retail stores, provides big data that can be used to estimate economic welfare. The investigator develops new double machine learning estimators of economic welfare based on big data. These estimators combine novel machine learning of certain economic weights with machine learning estimators of demand functions to do double machine learning estimation of welfare. These estimators are also generalized and applied to many other problems. In addition, this research uses the fact that scanner data follows individuals over time. Hence, individual demand functions are estimated and averaged to construct improved welfare measures.The objective of this research is to develop and apply economic demand analysis for large data sets that include many prices, such as scanner data. A common feature of scanner data is that cross price effects tend to be small, often an order of magnitude smaller than own price effects. This feature suggests that machine learning methods that allow for approximate sparsity, where most cross price effects are small, might be useful in practice. This research develops double machine learning estimators of exact consumer surplus and other welfare effects. The investigator uses novel machine learning of objects in Riesz representations that are not conditional expectations. This research produces a general method of double machine learning for generalized method of moments with first step series estimators. Scanner data is often panel data, where individuals or households are followed over time. The investigator further derives identification results for demand in panel data with general heterogeneity, and analyzes regularized fixed effect panel data estimators of average effects that can be applied to demand estimation.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.