Aggregate-Level PUF with High Data Confidentiality and Analytic Utility for Descriptive Analyses from Medicare Claims Data

• Activity Number: 638
• Type: Topic Contributed
• Date/Time: Thursday, August 8, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Erkan Erdem*+ and Avi Singh and Joshua M. Borton
• Companies: IMPAQ International LLC and NORC at the University of Chicago and NORC at the University of Chicago
• Keywords: Unit Level vs. Aggregate Level ; Micro Groups and Micro Means ; Subsampling ; Micro Aggregation ; Public Use File

Abstract

Creating a unit level PUF that is analytically useful and disclosure-safe is difficult due to the proliferation of publicly available indirect identifiers from various known and unknown sources that might exist at present or in future. In creating an aggregate level (AL-) PUF for Medicare Claims data, the data structure is transformed from a beneficiary level file with rows representing beneficiaries and columns analytic variables to a file with rows representing small clusters of beneficiaries called micro groups (MGs) and columns representing MG size and various domain means at the MG level (termed micro means--MMs) where domains are subpopulations of beneficiaries defined by variables corresponding to analytic goals. This allows information to be presented without sharing unit level data. Uncertainty in the MG size and associated MMs is introduced by random subsampling followed by weight calibration. The MGMM structure of AL-PUF is somewhat similar to the method of micro-aggregation for unit level PUFs where values of continuous variables deemed to be identifying are blurred by averaging over small clusters of observations based on similarity indices. However, the main difference is that in AL-PUF, MMs are provided for various domains defined by one or more variables and so joint relationships between variables are not distorted unlike the case of micro-aggregation. Moreover, as a result of subsampling, AL-PUF does not require the framework of identifying and sensitive variables used in traditional methods for creating unit level PUFs. For analysis domains, descriptive and analytic parameters of interest can be estimated using the weighted sums of products of suitable domain MMs and MG size over all MGs. Estimation of variance and covariance of point estimates can be obtained using essentially standard sampling methods because sampling errors introduced in MMs and MG size are due to multi-phase sampling. AL-PUF achieves high confidentiality by using MG sizes sufficiently small to reflect adequate uncertainty due to sampling errors but large enough to avoid problems with unit level PUFs. It has high analytic utility because analytic domains could be defined for any subset of variables of interest and the corresponding estimates remain approximately unbiased because uncertainty is only due to random subsampling. Examples from a 15% random sample of the 2010 Medicare claims data on chronic conditions are presented for comparisons between AL-PUF and an existing unit level PUF (termed chronic conditions PUF) based on k-anonymization and micro-aggregation.