New research finds privacy protections preserve most social science findings amid growing census data debate

In the largest study of its kind, researchers from MIT Sloan School of Management and Carnegie Mellon University found that roughly 91% of published research conclusions remain intact under commonly used privacy-preserving methods.

Cambridge, MA, Oct. 13, 2026 (GLOBE NEWSWIRE) -- As policymakers debate whether government statistical agencies should continue using differential privacy and other disclosure avoidance techniques to protect individuals' privacy in social science research, new research from MIT Sloan School of Management professor Alessandro Acquisti and co-authors Ryan Steed and Eduardo Abraham Schnadower Mustri at Carnegie Mellon University provides the largest study of its type to date that privacy protections can indeed safeguard confidentiality without substantially compromising scientific insight.

Published in Proceedings of the National Academy of Sciences (PNAS), "Reproducibility of Social Science Research Using Aggregate Statistics with Noise Infused for Differential Privacy" examined 303 findings from 93 published social science papers to test whether those conclusions would hold if researchers relied on data protected with differential privacy — a framework for releasing statistical information about datasets while protecting the privacy of individual data subjects — the same class of techniques used by the U.S. Census Bureau and other organizations to safeguard sensitive information.

Federal agencies publish statistics that underpin research and policymaking on public health, education, housing, labor markets, elections, and economic development. As agencies reconsider privacy practices and disclosure avoidance methods, the researchers’ findings indicate that policymakers may benefit from treating privacy and utility as empirical questions rather than assuming a fixed tradeoff applies across all datasets and products. \

How does differential privacy affect social science research findings?

The researchers found that approximately 91% of social science research findings remained consistent with the original published conclusions under privacy settings commonly used in practice, while 77% remained consistent even under substantially stronger privacy protections. Changes that result from adding privacy noise — the practice of adding controlled, random values to a dataset or query result — were generally smaller than discrepancies commonly observed in large-scale replication studies, and existing sources of uncertainty, including measurement error and data quality limitations, often had a greater impact on research findings than privacy protections themselves.

The findings arrive at a pivotal moment. Social science findings often inform public policy as well as decisions that can affect people’s lives, so it is vital for researchers to be able to reproduce published findings and verify that the results hold up. Such decisions are being made by statistical agencies, technology companies, and data stewards to protect confidential information, and those choices will shape the data available to researchers and policymakers for decades, Acquisti noted. Despite years of vigorous debate by policymakers, researchers and practitioners in industry over differential privacy, there has been surprisingly little empirical evidence measuring how much privacy protections actually alter the conclusions researchers draw from published data.

"The debate around differential privacy has proceeded with surprisingly limited empirical evidence, and, until now, there has been no large-scale estimate of how much differential privacy noise changes the conclusions researchers can draw from published data,” said Acquisti. "Our research showed that most published social science findings can remain reproducible under privacy protections commonly deployed in practice."

Why is this research important for census and federal data policy?

The study challenges a central assumption underlying many criticisms of differential privacy: that protecting privacy necessarily comes at a steep cost to data usefulness. For the types of research examined in this study, like social science research with published aggregate statistics, the results suggest the privacy-versus-utility tradeoff may be less severe than often portrayed by critics.

The authors argue that eliminating privacy protections may offer only limited improvements in data quality while potentially reducing data availability and foregoing important safeguards against reconstruction and re-identification attacks. Rather than forcing a choice between privacy and scientific value, the findings suggest both objectives can often be achieved simultaneously.

One of the study's most notable findings is that privacy-related distortions were often smaller than the variability researchers already encounter in the social sciences in general. Stronger original findings were particularly resilient to privacy protections, while weaker findings were more likely to be affected, noted the authors. This suggests that privacy-aware research practices could further strengthen standards of evidence across the field.

What can policymakers, researchers, and data stewards do next?

The authors suggested that data curators in government and industry can tailor privacy mechanisms and privacy budgets to the specific risks and sensitivities of different datasets rather than applying a one-size-fits-all approach. Policymakers can approach privacy and utility as empirical questions that should be evaluated product by product rather than through broad assumptions about disclosure avoidance methods. Researchers can adopt privacy-aware analytical techniques, evaluate how results perform under realistic data perturbations, and report those findings transparently.

To support that effort, the research team has made its benchmark and code publicly available and hopes other researchers will extend the analysis to additional datasets, disciplines, and policy contexts in the future.

As growing computational power and the proliferation of linked datasets increase the risk of re-identification from ostensibly anonymous statistics, the authors concluded that the debate should move beyond whether privacy protections and research quality can coexist. The evidence, they argue, increasingly points toward a more practical question: how to design data systems that achieve both.

"Our findings suggest that differential privacy may not be the scientific obstacle it is often portrayed to be," said Acquisti. "The real opportunity is to develop research methods that account for privacy protections while continuing to produce trustworthy, policy-relevant evidence."

Attachment

• New research finds privacy protections preserve most social science findings amid growing census data debate

Matthew Aliberti
MIT Sloan School of Management
7815583436
malib@mit.edu