A change to the US government’s system for processing the census could improve the accuracy of publicly disclosed data – without compromising the privacy of individual citizens and residents.
The government relies on national census data – gathered every 10 years by the US Census Bureau – to distribute hundreds of billions of dollars in federal funding to states and local communities. Such data also plays a key political role by helping states draw Congressional district boundaries and determining how many seats in the House of Representatives go to each state. Less accurate census data could lead to an unfair distribution of economic resources, as well as distorted representation in the US political system.
But because the Census Bureau publicly releases census statistics, it must process its data to preserve individual privacy. For the 2020 census, the government agency used a “disclosure avoidance” algorithm based on : the algorithm adds statistical noise – small and randomised additions or subtractions – to each published statistic, so that nobody can identify a specific person or household within the public census data. The trade-off is that such statistical noise can also create slight distortions in the data.
Advertisement
at the University of Pennsylvania and his colleagues developed a new method for measuring the privacy level of the processed data. They found that, for the 2020 census, the Census Bureau had introduced “unnecessarily high levels of injected noise” to achieve its privacy guarantee – which potentially meant it was adding unnecessary distortions as well. “Basically, their method is able to more accurately [measure] the level of privacy… than the method the Bureau used,” says at Harvard University.
An improved privacy-measuring method won’t affect the 2020 census data that has already been published – but it could have an impact on the planned 2030 census, says at the University of Chicago. “Something that’s nice about these results is that they don’t require the disclosure avoidance algorithm to be redesigned at all… this enables improvements to the noise without any other changes.”
If the improved method had been used on the 2020 census, the researchers suggest that the Bureau could have achieved the same privacy protection while reducing the amount of statistical noise by about 15 per cent at the national level and nearly 25 per cent at the neighbourhood level. Although such reductions were only demonstrated in an intermediate step of the data processing – census data undergoes further processing before it is finalised and published by the bureau – Cohen expects a similar result could be achieved in the published data used by states’ redistricting commissions and other organisations.
A Census Bureau spokesperson said the agency welcomes suggestions from academic researchers and will examine the paper as part of its “ongoing disclosure avoidance research”. Su says he and his colleagues plan to refine their work based on suggestions from scholars affiliated with the Census Bureau.
Before the Census Bureau changes its privacy protection measures, however, a “more holistic evaluation” is needed to assess how these changes will affect census data, says Imai. For instance, his own has demonstrated that the current algorithm can lead to greater measurement errors for Hispanic and multiracial populations.
arXiv
