Confidentiality and variance estimation in complex surveys

A variance estimator in a large survey based on jackknife or balanced repeated replication typically requires a large number of replicates and replicate weights. Reducing the number of replicates has important advantages for computation and for limiting the risk of data disclosure in public use data files. In the first part of this thesis, we propose algorithms adapted from scheduling theory to reduce the number of replicates. The algorithms are simple and efficient and can be adapted to easily account for analytic domains. An important concern with combining strata is that the resulting variance estimators may be inconsistent. We establish conditions for the consistency of the variance estimators and give bounds on attained precision of the variance estimators that are linked to the consistency conditions. The algorithms are applied to both a real sample survey and to samples from simulated populations, and the algorithms perform very well in attaining variance estimators with precision levels close to the upper bounds. Another important issue in survey sampling is the conflict of interest between information sharing and disclosure control. Statistical agencies routinely release microdata for public use with stratum and/or cluster indicators suppressed for confidentiality. For the purpose of variance estimation, pseudo-cluster indicators are sometimes produced for use in linearization methods or replication weights for use in resampling methods. If care is not taken these can be used to (partially) reconstruct the stratum and/or cluster indicators and thus inadvertently break confidentiality. In the second part of this thesis, we will demonstrate the dangers and adapt algorithms used from scheduling theory and elsewhere to attempt to reduce this danger.