Data privacy in official statistics

[brubinstein]

Might tools like differential privacy complement work in the ABS-MD, whether in linkage or official stats more generally? In a sentence: DP concerns release of aggregate statistics/parameter estimates to untrusted third parties, where observations might be of a sensitive nature. DP doesn't solve the same problem as cryptographic protocols which are concerned about untrusted communication channels/computational platforms but ultimately trusted recipients (DP complements such tools). DP has been studied primarily in the computer science community (starting in 2006, and this year winning the Godel Prize - a test of time award in theoretical computer science - to the group proposing the framework lead by Dwork now at Harvard). DP has had some traction with the US Census Bureau (a press release/article from Duke - my latest RHD is starting a postdoc in the group in a few weeks), and some notable personalities in statistics such as John Abowd (Cornell/US Census Bureau and a co-author of the Duke work), Stephen Fienberg and Larry Wasserman (CMU stats).

Shameless plug, a recent shipped R package diffpriv available on CRAN picks up some ideas of Wasserman to help make DP easier to use.