Public Preferences About Secondary Uses of Electronic Health Information

Importance  As health information technology grows, secondary uses of personal health information offer promise in advancing research, public health, and health care. Public perceptions about sharing personal health data are important for establishing and evaluating ethical and regulatory structures to oversee the use of these data.

Objective  To measure patient preferences about sharing their electronic health information for secondary purposes (other than their own health care).

Design, Setting, and Participants  In this conjoint analysis study, we surveyed 3336 adults (568 Hispanic, 500 non-Hispanic African American, and 2268 non-Hispanic white); participants were randomized to 6 of 18 scenarios describing secondary uses of electronic health information, constructed with 3 attributes: uses (research, quality improvement, or commercial marketing), users (university hospitals, commercial enterprises, or public health departments), and data sensitivity (whether it included genetic information about their own cancer risk). This design enabled participants to reveal their preferences for secondary uses of their personal health information.

Main Outcomes and Measures  Participants responded to each conjoint scenario by rating their willingness to share their electronic personal health information on a 1 to 10 scale (1 represents low willingness; 10, high willingness). Conjoint analysis yields importance weights reflecting the contribution of a dimension (use, user, or sensitivity) to willingness to share personal health information.

Results  The use of data was a more important factor in the conjoint analysis (importance weight, 64.3%) than the user (importance weight, 32.6%) and data sensitivity (importance weight, 3.1%). In unadjusted linear regression models, marketing uses (β = −1.55), quality improvement uses (β = −0.51), drug company users (β = −0.80), and public health department users (β = −0.52) were associated with less willingness to share health information than research uses and university hospital users (all P < .001). Hispanics and African Americans differentiated less than whites between uses.

Conclusions and Relevance  Participants cared most about the specific purpose for using their health information, although differences were smaller among racial and ethnic minorities. The user of the information was of secondary importance, and the sensitivity was not a significant factor. These preferences should be considered in policies governing secondary uses of health information.

During the past 4 years, the federal government has made an unprecedented public investment in health information technology. By the end of 2012, 72% of office-based physicians had adopted an electronic health record system.¹ Policy discussions have generally focused on how adoption of health information technology can affect the quality and value of health care.² However, increasingly digitized health information also enables new and potentially far-reaching opportunities for secondary uses of electronic health information. Secondary uses, which we define as uses other than personal medical care, fall into several categories, including research (eg, comparative effectiveness studies), public health (eg, public health surveillance), health care quality improvement (eg, measurement of provider performance), and commercial marketing (eg, pharmaceutical marketing).

There are strong ethical traditions (ie, autonomy) and regulatory structures (ie, privacy laws) that emphasize the privacy of this information and the rights of patients to know about and approve its use.³ However, in some circumstances privacy is overridden to advance societal interests, as with contagious or otherwise reportable illnesses.⁴ Ethicists and regulators have debated other circumstances when the proper balance between individual privacy interests and public good is less clear. For example, comparative effectiveness research could greatly advance scientific understanding of health and health care if it could be performed on clinically detailed and broadly representative information from interoperable electronic health records. The social benefits of such research might be large enough to justify more relaxed approaches to consent if inviolable elements of personal privacy were preserved.^4-7 Deidentifying data does not necessarily eliminate these tensions between personal privacy and public good, both because the highly protective systems of deidentification imposed under HIPAA (the Health Insurance Portability and Accountability Act) often eliminate information that is important to retain^8,9 and because patients often express concern about data stewardship and control separate from issues of privacy.^10,11

In this context, we sought to investigate public attitudes about the secondary use of electronically available personal health information. We recognized that these attitudes differ according to context, so we deployed conjoint analysis in an experimental survey design to examine patients’ willingness to share their personal health information. Conjoint analysis is commonly used in marketing research to disentangle preferences for individual attributes of consumer products when presented in combination,^12-14 for example, automobiles that vary according to price, performance, comfort, gas mileage, safety, and style. By asking individuals to rate different combinations of these attributes, one can measure the unique importance of each attribute to the consumer.

In the current context, the attributes are not consumer product characteristics but attributes of how health information might be used. We examined different uses of this information (medical research, quality improvement, or commercial marketing), different users of this information (university hospitals, public health departments, or commercial enterprises), and data sensitivity (eg, whether it included genetic information about their own cancer risk).

Prior studies^10,15-22 suggest that each of these factors is a potentially important determinant of public attitudes about data use, but, to our knowledge, no studies have examined them systematically. We focused on genetic information and cancer risk, given the heightened sensitivity to this type of health information and the strong interest in building genomic databases for future research.^23,24 We also focused on 3 racial and ethnic subpopulations: non-Hispanic whites, non-Hispanic African Americans, and Hispanics, using responses as a lens to examine differences across these broad groups, given the historical relationship between race, genetics, and distrust in biomedical research.^25-28

We recruited participants from an online research panel assembled by GfK Knowledge Networks.²⁹ The GfK Knowledge Networks panel members are recruited through probability-based sampling using random digit dialing and address-based sampling to create nationally representative panels.³⁰ This method provides coverage of households with and without telephones, cell phones, and Internet access (97% coverage of US households). Individuals without Internet access are provided with computer hardware and Internet access to complete surveys. Individuals with Internet access are given modest cash compensation. On average, panel members are invited to participate in 4 to 6 surveys per month and complete 3 surveys per month. Online research panels constructed from probability-based samples yield estimates similar to those produced by random digit–dial telephone surveys.³¹

We oversampled African Americans and Hispanics (from English- and Spanish-speaking households) to allow for comparisons across racial and ethnic groups. We administered the survey electronically from November 9, 2012, through December 2, 2012. Before analysis we excluded respondents who completed the survey in less than half the median completion time (<5 minutes) to screen out “speeders,” who were very unlikely to have read the questions before responding and who, in pretest analyses in this population, revealed almost no item-to-item variation in responses, consistent with inattention to question content. The protocol was approved by the University of Pennsylvania Institutional Review Board.

The online survey instrument contained 3 main sections: health status and health care access measures, responses to short scenarios describing secondary data uses, and a 9-item scale measuring health care system distrust.

Participants were presented with the following introduction:

Many doctors and hospitals are starting to use electronic medical records instead of paper charts when they provide care. Electronic medical records can also be used for other health care and public health reasons.

You will be shown some possible uses of your electronic health information. In each case, you will be shown what health information will be used, who will use it, and what they will use it for. Please indicate how willing you would be to share your health information for each situation. Your name would not be released.

This introduction was followed by descriptions of 6 situations in which participants were told what information would be used reflecting the sensitivity of the information (“your medical history” [ie, low sensitivity] or “your medical history and the results of a personal genetic test that predicts your chance of getting cancer” [ie, high sensitivity]), the user of the information (“university hospital,” “drug company,” or “public health department”), and the use of the information (“research new ways to prevent cancer,” “evaluate how well your doctor provides preventive cancer care,” or “identify what kinds of patients will be interested in buying their cancer prevention product”). Each participant was randomly assigned to receive 6 of the possible 18 scenarios. For each item, participants were asked, “In this situation, how willing would you be to share your health information?” and rated their responses on a 10-point scale, where 1 indicates not at all willing and 10, very willing.

Prior research suggests that health status may influence support for sharing health information; however, the evidence is mixed.^15,19 We measured self-rated health with a single item from the Behavioral Risk Factor Surveillance System 2010 questionnaire (short form 1).³² We evaluated insurance status, whether the respondent had a usual source of medical care, and whether the respondent had experienced cost-related barriers to care in the prior year, using items derived from the National Health Interview Survey.³³ We hypothesized that individuals marginalized from the health care system or those with a strained financial relationship would be less willing to share their health information. We measured distrust in the health care system by using the Revised Health Care System Distrust Scale,³⁴ a 9-item scale that ranged from 9 to 45 (45 indicates a high level of distrust) and includes 2 validated subscales: competence distrust and values distrust. Concerns about health information privacy potentially relate to both domains if patients are concerned about mishandling of information as well as uses of information that may bring harm through discrimination or stigma.^35,36 Distrust may also mediate associations between race or ethnicity and willingness to share health information, given the history of distrust among minority populations about biomedical research.^25-28

We asked respondents to rate various institutions and organizations on how much confidence they had in each to protect their health information. Respondents chose from 3 categories: hardly any confidence at all, only some confidence, and very high confidence. This question was adapted from a General Social Survey question measuring confidence in institutions overall.³⁷

GfK Knowledge Networks had previously collected demographic data on panel participants, including age, sex, race, ethnicity, income, educational level, and metropolitan or rural residential status.

We compared characteristics of responders and nonresponders and compared the sociodemographic characteristics between racial or ethnic groups (non-Hispanic white, non-Hispanic African American, and Hispanic) using t tests and analysis of variance for continuous variables and χ² tests for categorical variables.

We conducted a conjoint analysis based on a main effects analysis-of-variance model. In this analysis, we computed a numerical part-worth utility value for each level of each attribute.^13,38 Large part-worth utilities were assigned to the most-preferred levels, and small part-worth utilities were assigned to the least-preferred levels. The attributes with the largest part-worth utility range were considered the most important in explaining the variability in the outcome. Importance weights are reported for each attribute.

Subsequently, to adjust for patient-level characteristics and to investigate potentially important interactions with the attributes, we conducted linear regression between the attribute levels, participant characteristics, and distrust. Correlation of responses within each individual was accounted for using generalized estimating equations assuming an independence correlation structure. Poststratification weights provided by GfK Knowledge Networks were used in all analyses to adjust for any survey nonresponse as well as any noncoverage or undersampling and oversampling resulting from the study-specific sample design.

Models include the 3 test attributes, patient race and ethnicity, a measure of high distrust of the health care system (top quartile), health status, measures of access to care, and other sociodemographic variables, as well as interaction terms between the attributes and patient race and ethnicity. We specified a baseline scenario in the models in which we hypothesized support would be the highest (sensitivity, low; user, university hospital; and use, research). Wald tests were used to assess significance, with a type I error rate of 0.05. All analyses were conducted with SAS software (version 9.3; SAS Institute Inc).

Of the 5119 panel members we invited to participate, 3336 completed the survey (response rate, 65.2%); 272 “speeders” were removed for a final sample of 3064, including 2093 non-Hispanic whites, 455 non-Hispanic African Americans, and 516 Hispanics (250 completed the survey in Spanish). The sociodemographic, health, and health care variables differed across the 3 racial and ethnic categories (Table 1). Responders and nonresponders differed demographically; nonresponders were younger, had lower incomes and educational levels, and were slightly more likely to be female. However, we applied poststratification weights to our analyses to account for nonresponse bias.

The mean willingness to share health information for each conjoint scenario is presented in Table 2 by race or ethnicity. In our baseline scenario in which we hypothesized support would be the highest (sensitivity, low; user, university hospital; and use, research), the overall willingness to share electronic health information was moderately high (mean rating, 6.82 on a scale of 1-10, where 1 indicates not at all willing to share and 10, very willing to share). For this baseline scenario, whites demonstrated greater willingness to share their electronic health information than Hispanics (mean, 6.98 vs 6.38; P = .03). The differences between whites and African Americans were not significant (mean, 6.98 vs 6.58; P = .21).

Using conjoint analysis, we calculated importance weights for the 3 attributes. Importance weights are a measure of the importance of an attribute relative to the other attributes in the model on an individual’s preferences. The use of data was the factor that most influenced participants’ willingness to share their electronic health information (importance weight, 64.3%). The user of the data was less important (importance weight, 32.6%), and the sensitivity of health information was not important (importance weight, 3.1%). Use was more important among all 3 racial/ethnic groups (importance weight, 65.7%, 57.9%, and 56.4% for white, Hispanic, and African American respondents, respectively) than user (importance weight, 31.5%, 38.3%, and 41.3%, respectively) or sensitivity (importance weight, 2.7%, 3.8%, and 2.3%, respectively).

In unadjusted generalized estimating equation models, use and user variables were statistically significant as shown in Table 3. Marketing uses resulted in the largest declines in willingness to share health information compared with research uses (β = −1.55; P < .001). Respondents also reported less support for quality improvement uses than for research uses (β = −0.51; P < .001). The user of health information was also important. Drug companies and public health departments received lower support than university hospitals (β = −0.80 and −0.52, respectively; both P < .001). We tested interactions of race or ethnicity with the conjoint attributes (use, user, and sensitivity) while controlling for sociodemographic characteristics, including educational level and income, health status, and access to care measures. The difference in willingness to share health information for marketing vs research uses was smaller for African Americans and Hispanics than for whites (African American × marketing interaction, β = 0.87; Hispanic × marketing interaction, β = 1.01; both P < .001). There was also a smaller but statistically significant interaction between race or ethnicity and quality improvement uses compared with research uses (African American × quality improvement interaction, β = 0.44; Hispanic × quality improvement interaction, β = 0.59; both P < .001). In Figure 1, we display these interactions for the 3 uses (research, quality improvement, and marketing) and for 2 users (university hospital and drug company) by race or ethnicity.

Respondents lacking a usual source of care (β = −0.44; P < .001) and those experiencing cost barriers to care (β = −0.30; P = .02) were less willing to share their health information, as were those reporting high levels of distrust of the health care system (β = −0.65; P < .001), but distrust did not influence the relationship between race and the conjoint attributes. There were no significant differences by demographic characteristics except that young adults (aged 18-29 years) were slightly more willing to share their health information than middle-aged adults (aged 45-59 years).

We asked respondents how confident they were in various institutions and organizations to protect their health information. The results varied widely and are presented in Figure 2. Health care provider organizations and nonprofit health-related organizations received relatively high support, and Internet and social media companies had very low support.

As clinically detailed elements of personal health information are increasingly coded in electronic health records and as those records become more aggregatable across systems, secondary uses of electronic health information offer enormous promise in advancing research, public health, and health care. Even if we see the use of this information as providing immense social good, its effective use will probably require the acceptance of the public and trust and confidence about information stewardship. Our study begins to describe population support for specific uses of personal health information. It is based largely on revealed preferences derived from an experimental design rather than expressed preferences and has 3 main findings.

First, what our participants care about most is the specific purpose for using information, and among the choices we investigated, the goal they most privilege is research. Although it is not surprising that support was greater for research uses than for marketing uses, it is surprising that support for quality improvement uses was lower than for research because it is the use that is most likely to offer a more immediate benefit to the patient. Participants also revealed they care about the user but less than they care about use. To our surprise, the sensitivity of the health information—at least within the range we presented (personal medical history plus personal genetic test results vs personal medical history alone)—was not important. This finding contrasts with the notion that patients view genetic information as particularly sensitive.^39,40 It may add support to the arguments against privileging genetic information, as some experts have argued.^41,42

Second, we found that compared with white participants, racial and ethnic minorities were no less supportive of their health information being used for any use except for a small but nonsignificant difference for research. We had hypothesized that racial and ethnic minority respondents would be less supportive, given prior work showing lower levels of support for research participation and higher levels of concern for health information technology.^15,25-28,43

Third, relationships with the health care system are associated with support for sharing health information. Participants with high levels of distrust of the health care system, those without a usual source of care, and those who had recently experienced cost barriers to care were less supportive of secondary uses. However, participants also identified the health care delivery system as the most trusted to protect health information relative to other institutions and organizations—although less than half expressed a high level of confidence even in these organizations.

This study has several limitations. First, we presented participants with hypothetical scenarios that may not reflect decisions in more naturalized circumstances. However, responses to hypothetical scenarios have a high level of concordance with actual behavior.⁴⁴ Second, we were not able to test an exhaustive list of attributes describing different data use scenarios. For example, we may have found differences by sensitivity of health information if we explored a broader range of examples with potentially greater (eg, diagnosis of human immunodeficiency virus infection) or lesser (eg, height) sensitivity. It is also possible that respondents think differently about the sensitivity and risks associated with genetic information compared with those associated with other health information considered sensitive. Third, our respondents may not be fully representative of the US population of the 3 racial and ethnic groups we sampled, and there may be important differences between those broad racial or ethnic categories that our study could not uncover. In addition, there may be differences in attitudes between those who are willing to participate in a survey research panel and those who are not.

This study also has strengths. In particular, we used conjoint analysis in a carefully controlled experimental design enabling participants to reveal their preferences rather than merely express them in response to static questions, which might otherwise be subject to attribution bias or social desirability bias. Revealed responses provide a more naturalized context, even when presented in hypothetical settings. Indeed, the experimental design of our survey gives us greater confidence that the effects we observe within the sample are due to true effects of the conjoint attributes rather than other unobserved factors.

Most policy discussions about secondary uses of health information focus on whether data are identifiable or not. If it is deidentified, much of its value can be lost. If it is identifiable, restrictive privacy regulations sometimes limit the potential applications of those data toward social good. Although discussions about data identification are important, our study suggests that populations care about the purpose for data use, so these considerations might be elevated in decisions about personal health data sharing. As the use of health information technology expands, so will the potential value of the health information generated for uses beyond personal medical care. In addition, as more health information is generated outside traditional medical encounters (eg, patient health records), so too will the pressure to define acceptable uses. Our research can help inform these future debates as well as those focused on protecting human subjects, such as proposed changes to the Common Rule governing such protections.⁵ Our findings also suggest that organizations and institutions within the health care delivery system are the most trusted stewards of health information and should play an instrumental role in efforts to extend the uses of population-level health information.

Corresponding Author: David Grande, MD, MPA, Division of General Internal Medicine, Department of Medicine, University of Pennsylvania Perelman School of Medicine, 3641 Locust Walk, Colonial Penn Center 407, Philadelphia, PA 19014 (dgrande@wharton.upenn.edu)

Accepted for Publication: May 28, 2013.

Published Online: August 19, 2013. doi:10.1001/jamainternmed.2013.9166.

Author Contributions:Study concept and design: Grande, Mitra, Shah, Asch.

Acquisition of data: Grande, Asch.

Analysis and interpretation of data: Grande, Mitra, Wan, Asch.

Drafting of the manuscript: Grande, Mitra.

Critical revision of the manuscript for important intellectual content: All authors.

Statistical analysis: Grande, Mitra, Wan.

Obtained funding: Grande, Mitra.

Administrative, technical, and material support: Grande.

Study supervision: Mitra.

Conflicts of Interest Disclosures: None reported.

Funding/Support: This research was supported by grant 5R21HG006047-02 from the National Human Genome Research Institute.

Role of the Sponsors: The funder had no role in the design and conduct of the study; in the collection, management, analysis, and interpretation of the data; and in the preparation, review, or approval of the manuscript.