文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > expressed are those of the authors and not necessarily those of the U.S. Census Bureau. Mod
侵权投诉

expressed are those of the authors and not necessarily those of the U.S. Census Bureau. Mod

Statistical organizations collect data via survey forms and other methods. The microdata are valuable for modeling and analysis. To produce a public-use file, the organizations mask the data in a manner that may prevent re-identification of data associated

The outline of this paper is as follows. Following this introductory section, we provide additional background on some of the methods for masking data X to put it in a form suitable for public-use data Z. The masking methods are often easily implemented methods that have not been justified in terms of providing a file with analytic properties or preventing re-identification. The exceptions for analytic properties are suitably adjusted additive noise (Kim 1986, 1990; Fuller 1993; Yancey et al. 2002) and synthetic microdata (Raghunathan et al. 2003; Reiter 2002; Kennickell 1999; Abowd and Woodcock 2002, 2004). The clear coefficients are the same (very slight bias) but with slightly inflated coefficients of variation. Kim (1986, 1990) and Kim and Winkler (1995) demonstrated that some analyses such as regression were still possible on some subdomains (that were sufficiently large and with some additional properties) but that most subdomain analyses on Z did not correspond to subdomain analyses of X. Both Fuller (1993) and Lambert (1993) indicated that a masking procedure used in producing public-use data Z from data X should approximately preserve means, covariances, and one additional analytic property. Brand (2002) observed exception for not being re-identifiable is k-anonymity (Samarati and Sweeney 1998, Sweeney 2002, Bayardo and Aggrawal 2005) for which no analytic properties have ever been established. In the third section, we provide some methods that might be used for defining analytic properties and quality of public-use microdata. The primary intuition is that the microdata should be suitable for an intended use. The third section provides some comments and research problems. In the final section, we provide some concluding remarks.

2. Background and Definitions

This section is divided into three subsections. In the first section, we provide general background on some of the general methods that relate to analytic properties of microdata. In the second section, we give more detail of the specific methods used in statistical agencies. We quickly reject many of the methods because they have never been justified as providing valid analytic properties in the masked microdata. In the third section, we describe a crude, but specific, definition of quality that might be applied in evaluating analytic properties of microdata.

2.1. General Analytic Properties of Masked Microdata

Users of public-use microdata are primarily concerned with the analytic properties of data. Given that the original, non-public microdata are of high quality, then the masking procedures used in producing a public-use file will reduce the quality. The masked files may only allow zero (see section 2.2 for examples) or one analyses. Because we are primarily concerned with methods that have demonstrated analytic properties, we begin with an example that involves regression and additive noise.

Kim introduced the idea of additive noise in a form such that data X might be transformed to data Y = X + c ε that is in turn linearly transformed to data Z (also see Tendick and Matloff, 1994). The independent noise ε has mean zero and the same covariance Σ as X. Data Y is the same mean as data X and cov(Y) = (1+c) Σ. The positive constant c is between 0 and 0.5. The data Z has the same means as X and approximately the same covariances and correlations as X. Regression analyses that are possible on X can be approximately reproduced on Z in the sense that regression

that few additional statistical properties were preserved in the files Z due to the effect that the basic additive noise procedure had on the detailed distributional properties of the original data X. Fuller (1993) noted that it might be possible re-identify a small proportion of records in Z that are masked via the basic additive noise procedure. Yancey et al. (2002) demonstrated that a mixture of additive noise approach could still provide the same analytic properties as the Kim approach while significantly reducing re-identification risk. Mixtures of additive noise (e.g., Yancey et al. 2002) would have somewhat greater deleterious effects of the distributions than the effects of basic additive noise.

Several authors have suggested creating synthetic microdata Z that reproduces some of the aggregates obtainable from the original data X or allows reproduction of a model M that could be built on X. Raghunathan et al. (2003) provided the theory for estimating the variances of certain univariate estimates of synthetic data obtained via a valid probabilistic modeling procedure. Reiter (2002) had earlier demonstrated that certain properties of the original data X that were not explicitly included in the model M would never be in the synthetic data Z. Reiter also showed that, due to lack of data, some properties that were in the model M would not necessarily be in the synthetic data Z. Various authors have demonstrated that, if a sufficiently detailed model M of the data X is developed, then some of the records in the synthetic data Z will necessarily be close to some of the real original data records X (Fienberg 1997, Reiter 2002, Winkler 2004).

Certain distributional outliers (that are often easy to detect) may be needed for accurately estimating multiple moments. Kim and Winkler (1995, see appendix to longer web report) provided controlled distortion procedures for changing outliers while (approximately) preserving sets of moments in sufficiently large domains. If only a few outliers need to be changed because they cause records to be easily re-identified, then a large number of aggregates may need to be modified. These types of modifications will yield moderate or substantial changes in a very large number of records.

The Kim-Winkler procedure, like the more general DuMouchel et al. (1999) procedure (see section 3.1), depends on the subdomains where the controlled distortion procedures are applied to be quite large. The values of fields in individual records are the unknowns that must be solved so that the moment equations are satisfied. If there are

上一页第3页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email