Developing synthetic data methods for large confidential administrative databases

社会科学家需要获得高质量的研究数据，传统上是大型调查。这些都是昂贵的，所以已经有了一个转变，使常规收集的行政数据更容易为研究人员所用。政府的开放数据访问政策也导致了一项倡议，使其部门持有的行政数据更广泛地可用。这些数据库通常包含大量具有潜在敏感信息的记录的信息，并且严格限制访问。这导致研究如何在不损害机密性的情况下改进对政府行政数据库的访问。合成数据是解决这个问题的一种日益流行的方法。该方法用从适合原始数据的统计模型中提取的合成值替换数据。这通常会多次执行，以生成多个合成数据集。由于数据现在只包含合成值，因此应该保护机密性，并且使用了合理的模型，因此应该保留统计属性。合成数据将使研究人员能够在分析原始数据之前在合成版本上测试他们的方法。这个项目将为管理数据库开发综合数据方法，从而有可能获得更容易获得的综合版本。近年来，对部分合成的SDC数据的使用越来越多。美国正在开发多种综合数据产品，例如收入和计划参与调查（https://ecommons.cornell.edu/handle/1813/43924）和纵向商业数据库（Kinney et al., International Statistical Review, 2011: 79(3)）。在欧洲，随着德国的IAB调查合成数据以保护德国企业调查（Drechsler和Reiter， Journal of Official Statistics, 2009: 25(4)），这种呼吁也在增长。在英国，生产综合数据的活动相对较少，唯一的例外是考虑综合纵向数据方法的项目（https://sls.lscs.ac.uk/projects/view/2013_012/）。到目前为止，在联合王国还没有在生成综合行政数据库方面进行实质性的工作。

There is a demand from social scientists to access high quality data for research, traditionally large surveys. These are costly and so there has been a shift to making routinely collected administrativedata more available to researchers. The government open data access policy has also led to an initiative to make the administrative data their departments hold, available more widely. Thesedatabases typically contain information on a large number of records with potentially sensitive information, and have severely restricted access. This has led to investigating ways to improve access to government administrative databases without compromising confidentiality.Synthetic data is an increasingly popular approach to address this problem. The approach replaces the data with synthetic values drawn from a statistical model fit to the original data. This is typicallydone multiple times to generate multiple synthetic data sets. As the data now comprise only synthetic values, confidentiality should have been protected, and providing a plausible model hasbeen used, statistical properties should be preserved. Synthetic data would give researchers the ability to test their methodology on a synthetic version prior to analysis of the original data. This project will develop synthetic data methods for administrative databases leading to the potential for more accessible synthetic versions.The use of partially synthetic data for SDC has been increasing in recent years. There are multiple examples of synthetic data products being developed in the US, such as the Survey of Income and Program Participation (https://ecommons.cornell.edu/handle/1813/43924), and the Longitudinal Business Database (Kinney et al., International Statistical Review, 2011: 79(3)). The appeal is also growing in Europe with the IAB in Germany investigating synthetic data to protect the German Establishment Survey (Drechsler and Reiter, Journal of Official Statistics, 2009: 25(4)). There is relatively little activity in producing synthetic data in the UK, the one exception being a project considering methods for synthesising longitudinal data (https://sls.lscs.ac.uk/projects/view/2013_012/). To date there has been no substantive work on generating synthetic administrative databases in the UK.