Data sensitivity shaped Chinese adults’ health data-sharing choices

In a discrete choice experiment with Chinese adults presented with hypothetical health data-sharing arrangements, the type of information mattered more than any other tested feature: general health information was substantially more acceptable than sensitive information. The strongest preferences emerged from the combination of data sensitivity, custodial responsibility, rights protection and incentives, indicating that the same governance model could elicit different choices depending on the information involved. Sensitive records prompted stronger demand for operational safeguards and a preference for public hospitals, whereas routine health information was more often selected with a government-affiliated neutral third party and tangible benefits.

The findings come from a discrete choice experiment conducted online in March and April 2026. Researchers recruited 1,481 adults through the Credamo platform and divided them randomly among three questionnaire blocks. Each respondent completed five forced choices between two data-sharing profiles containing different levels of five attributes: custodian, purpose, rights protection, incentive and data type.

General information included routine diagnostic records, test results and medication use, while sensitive information included genetic data, mental health records and infectious disease history. The profiles spanned 144 possible combinations, from which the investigators selected 15 sets using a D-efficient design. Across the five tasks per person, the study produced 7,405 valid choice observations, or 14,810 option-level records for analysis.

Ordinary least squares models estimated average marginal component effects, meaning the change in the probability of choosing one profile when an attribute changed while the experimental design held other features constant. Standard errors were clustered by respondent to account for repeated choices. General health information increased the probability of selection by 32.05 percentage points relative to sensitive health information, the largest effect among the five attributes.

The government-affiliated neutral third party increased choice probability by 5.9 percentage points compared with a public medical institution, while strong rights protection added 7.8 points relative to basic protection. A RMB 50 cash reward increased selection by 14.5 points and redeemable health points by 12.9 points, whereas an honorary certificate had no statistically detectable effect in the overall model. The data context altered not only the size but, in several cases, the direction of these effects.

A neutral third-party custodian reduced the probability of choosing a sharing profile for sensitive data by 4.8 percentage points compared with a public hospital, but increased it by 8.0 points for general data. The difference between those contexts was statistically significant after correction for multiple comparisons. Use for individual diagnosis and treatment also performed differently: relative to medical research, it reduced selection by 33.3 points for sensitive information and by 10.5 points for general information.

Epidemic control was less acceptable than medical research for sensitive data, but no longer differed significantly for general data. These are stated-choice associations within the experimental design, not evidence that any particular custodian or use would improve real-world health outcomes. Rights protection showed one of the clearest boundaries between routine and sensitive information.

The strongest package allowed participants to withdraw authorization at any time, provided an annual report on data use, required fresh confirmation before each use and subjected sharing to regular third-party audit. Compared with basic protection, this package increased the probability of selection by 22.1 percentage points for sensitive data, but had no detectable effect for general data. Moderate protection showed a small positive association for sensitive information and essentially no effect for general data.

Incentives also changed direction. Health points reduced selection for sensitive data by 3.0 points but increased it by 24.4 points for general data. An honorary certificate reduced selection by 11.0 points in the sensitive condition and increased it by 5.8 points in the general condition.

Cash remained positive in both conditions, though its estimated effect was larger for general information. Trust in government organizations had a narrower and more conditional influence than data sensitivity. The study measured trust on a seven-point scale, with a mean of 5.756, and included nine interaction tests that were adjusted using the Benjamini–Hochberg false-discovery-rate procedure.

Only two remained statistically significant after correction. Higher trust strengthened the preference for a government-affiliated neutral third party over a public hospital; the estimated effect was 6.91 percentage points among respondents above the trust-score median and 4.73 points among those below it. The direction remained positive in both groups.

Higher trust also weakened the effect of health points, whose estimated effect fell from 15.26 points in the lower-trust group to 11.01 points in the higher-trust group. Trust did not significantly change preferences involving use, rights protection, cash, honorary incentives or general versus sensitive data. A supplementary latent class model suggested that the averages may conceal conflicting preferences within the sample.

The investigators used penalized estimation because an unpenalized model did not have a finite maximum with five tasks and nine class-specific parameters. Model selection was inconclusive, so the most stable three-class solution was reported as an exploratory summary rather than a definitive segmentation of the population. The largest class, comprising 51.3% of respondents, responded positively to strong rights protection, health points and cash.

A second class, 29.0%, responded negatively to most arrangements and especially strongly against use for individual diagnosis. A third class, 19.7%, had coefficients near zero for most attributes. Average posterior assignment probabilities ranged from 0.85 to 0.89, but the class labels describe statistical coefficient patterns rather than validated psychological types.

For governance, the results favor differentiation rather than a uniform framework. Strong rights protection and public medical custody were most strongly associated with selection of profiles involving sensitive data, while a neutral third party and tangible incentives performed better for routine information. The study therefore supports the tested proposition that data sensitivity can act as the risk context through which people evaluate every other feature of a sharing arrangement.

It does not establish that public hospitals are objectively safer than neutral custodians, or that incentives necessarily improve data quality or public health. Those outcomes were not measured. The findings instead concern which hypothetical arrangements an online sample of Chinese adults selected and may help policymakers test whether different proposals receive different responses when risk, accountability and control are made visible.

Several limitations constrain generalization. The online panel was not probability sampled and no population weights were applied because selection probabilities were unavailable. Respondents were predominantly young, urban and highly educated: 73.6% were aged 19 to 35, 92.2% lived in urban areas and 91.0% held at least a bachelor’s degree.

Self-selection, professional respondents and digital-literacy bias may therefore have influenced the results, and the findings may not represent older, rural or less-educated Chinese adults. Forced binary choices identify relative preferences but provide no opt-out option, estimate no absolute participation rate and do not capture actual sharing behavior. Trust was self-reported and concentrated toward the upper end of the scale, while proposed mechanisms involving accountability, perceived risk or motivation were not directly tested.

The latent class solution was also exploratory.

Subject of Research: Public preferences for context-dependent health data-sharing governance in China

Article Title: Public preferences for health data sharing governance: a discrete choice experiment in China

Article References: Li, J., Jin, Y., Hui, H., & Huang, X. (2026). Public preferences for health data sharing governance: a discrete choice experiment in China. Frontiers in Public Health, 14, Article 1883099. https://doi.org/10.3389/fpubh.2026.1883099

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1883099

Keywords: health data sharing, data sensitivity, discrete choice experiment, China, public preferences, rights protection, incentive design, institutional trust, data governance, privacy, Public, preferences