Doctoral Dissertation Research: Synthetic Data Generation for Small Area Estimation

各种研究人员、分析师、决策者和社区规划者对小面积估计的需求正在大幅增长，他们使用这些数据来推动对影响社区及其居民生活的问题的现有知识。统计机构定期收集小地理区域的调查数据，但由于发布小区域标识符有保密风险，往往无法以微观数据形式公开发布这些数据。这项建议的主要目标是制定一种方法，用于生成完全合成的微观数据集，允许有效估计小面积的统计数据，同时保护答卷人的机密性？的数据。拟议的方法将使用众所周知的贝叶斯分层建模技术，根据公共数据集中常用变量集的假设预测模型生成模拟（或估算）值。建模方法将考虑到不同程度的变化发生在县和州一级的目的，生成合成数据，产生有效的推论，为两个层次的地理。将考虑参数和非参数建模策略，并将比较基于实际数据和合成数据的小面积推断，以评估合成数据方法的实用性。此外，在合成数据应用中通常被忽视的两个现实世界数据复杂性将在本研究中得到解决，包括：1）为家庭和个人级别的属性生成合成数据，并维护家庭内的组成结构; 2）考虑复杂的样本设计特征（例如，选择、分层和聚类的不相等概率）。拟议的研究将通过测试一种传播适用于小面积估计的公共使用数据的替代方法，同时加强保密保护，开辟新天地。如果所提出的方法被证明是成功的，则可以避免要求数据用户访问受限数据中心设施内的小区域数据的当前实践。通过发布带有小区域标识符的合成微观数据，用户将能够对目前不允许不受限制地访问数据的地理级别进行定制的小区域分析。这一创新可能有助于满足对小面积估计的日益增长的需求，并增加利用统计机构的微观数据编制的小面积估计的绝对数量。

Demand for small area estimates is growing heavily among a variety of researchers, analysts, decision-makers, and community planners, who use these data to advance current knowledge on issues affecting communities and the lives of their residents. Statistical agencies regularly collect survey data from small geographic areas, but are often prevented from publicly releasing these data in microdata form because of confidentiality risks associated with releasing small area identifiers. The main objective of this proposal is to develop a methodology for generating fully-synthetic micro-level datasets that permit valid estimation of small area statistics while protecting the confidentiality of respondent?s data. The proposed methodology will use well-known Bayesian hierarchical modeling techniques to generate simulated (or imputed) values based on an assumed prediction model for a commonly used set of variables found in public-use datasets. The modeling approach will account for different levels of variation occurring at the county- and state-level for the purposes of generating synthetic data that produce valid inferences for both levels of geography. Parametric and nonparametric modeling strategies will be considered, and small area inferences based on the actual data and synthetic data will be compared to evaluate the utility of the synthetic data methodology. In addition, two real-world data complexities typically ignored in synthetic data applications will be addressed in this research, including: 1) generating synthetic data for household- and individual-level attributes and maintaining the within-household composition structure; and 2) accounting for complex sample design features (e.g., unequal probabilities of selection, stratification, and clustering). The proposed research will break new ground by testing an alternative method of disseminating public-use data suitable for small area estimation while enhancing confidentiality protection. If the proposed methods prove to be successful, then the current practice of requiring data users to access small area data within restricted data center facilities may be avoided. By releasing synthetic microdata with small area identifiers, users will be able to perform customized small area analysis for levels of geography that are not currently permitted without restricted data access. This innovation may help meet the growing demand for small area estimates and increase the sheer volume of small area estimates produced using microdata from statistical agencies.