Contextual Consent: Ethical Mining of
Social Media for Health Research
Social media are a rich source of insight for data mining and user-centred research, but the question of consent arises when studying such data without the express knowledge of the creator. Case studies that mine social data from users of online services such as Facebook and Twitter are becoming increasingly common. This has led to calls for an open discussion into how researchers can best use these vast resources to make innovative findings while still respecting fundamental ethical principles. In this position paper we highlight some key considerations for this topic and argue that the conditions of informed consent are often not being met, and that using social media data that some deem free to access and analyse may result in undesirable consequences, particularly within the domain of health research and other sensitive topics. We posit that successful exploitation of online personal data, particularly for health and other sensitive research, requires new and usable methods of obtaining consent from the user.
Contextual Consent: Ethical Mining of
Social Media for Health Research
|School of Computer Science|
|University of St Andrews|
|St Andrews, Fife, UK|
|School of Computer Science|
|University of St Andrews|
|St Andrews, Fife, UK|
ethics; privacy; health; consent; social media
Social media services, such as Facebook and Twitter, are a near ubiquitous part of people’s lives. Facebook, for example, boasts over 1.7 billion monthly active users [?]. As a result we are increasingly sharing more and more data about our lives online: 68% of UK-based online adults look at social media sites or apps each week, 35% upload photos or videos, 28% share links to websites or online articles, and 24% contribute comments to a website or blog [?].
The vast quantities of personal data shared through such platforms can offer new insights into important areas of research [?]. Academics from various fields have capitalised on these data, leading to the comprehension of complex social behaviours and trends in society. Within a health context alone, such social data has been found to act as a predictor for depression [?, ?], suicide risk factors [?, ?], mood changes [?], flu outbreaks [?] and problem drinking in US college students [?]. Such findings can help researchers and practitioners identify important markers in health and in society.
Despite these advantages, those who created the social data in question are not always informed about the process. While some researchers outline clear methods of obtaining informed consent when conducting research on social media data (e.g. [?]), only around 5% of the literature that use such data describe their consent procedures [?].
Some have argued that social data posted online are freely available for research [?]; however, others argue that just because personal information is made available online does not mean it is appropriate to capture and analyse [?, ?, ?, ?]. Conway and O’Connor argue that the potential challenge to privacy occurs “not in the reading or accessing of individual materials (publicly available as they are), but rather in the processing and dissemination of those materials in a way unintended" [?].
While it is important for such research using social media data to continue, the informed consent of all participants is an important milestone to strive for. Firstly, consent decisions are strongly tied to the expected audience [?, ?]. Research has suggested that users have differing online behaviour based on their perceived audience, including more self-censorship [?, ?], and sharing smaller proportions of both positive and negative emotions [?] when the perceived audience is more sparse, and less controlled.
Secondly, research into privacy settings on social media have suggested that most users may be significantly oversharing what is intended [?, ?]. When analysing social media data online, data can be taken from individuals who are unaware that it is accessible, and privacy violations may be occurring. The likelihood of privacy violations increases with the number of people involved, with many such studies indiscriminately scraping the data of huge numbers of individuals [?], or not even reporting how they gather data [?].
And finally, participants may find it intrusive to discover that their data is used in a way which is outside the boundaries of what was originally expected [?, ?]. An example of this is Samaritans Radar, a well-intentioned Twitter app designed to monitor the tweets of the user’s contacts for potentially suicidal messages [?]. This app received criticism from the Twitter community over privacy concerns, and was shut down days after launching.
Therefore, with all previous points considered, participants may be sharing more sensitive and emotional content with a far wider audience than intended, they may not have otherwise provided consent for their data to be studied, and would likely find a violation of this to be intrusive. This suggests a disconnect between the researcher and the participant; a situation in which the autonomy of the participant should be respected. Safer and more explicit methods of consent are therefore necessary to best protect the needs of potential participants, while ensuring that we can make the best use of the vast quantities of social media data for health research.
In this paper, we take the position that new forms of obtaining meaningful consent are necessary for successful online health research. We outline current methods of obtaining consent, discuss new potential methods that seek to mitigate these problems, and outline some of the challenges that need to be overcome for such a method to be developed.
Informed consent, a declaration that the participant understands the consequences of participating in the study, is widely seen as fundamental to conducting research with human participants [?]. From the point of view of the researcher, it is seen as fulfilling ethical responsibilities with regards to the protection of data, privacy, and autonomy of the participant [?]. While there appears to be broad agreement when it comes to interventional research, the subject of whether informed consent is necessary in a study of online communications is the subject of debate [?, ?]. As a result, new models for consent have been proposed [?, ?].
Studies into user attitudes are often subject to selection bias, given that participants have opted in and are therefore not necessarily representative of the wider population. Despite this, it is worth consulting the viewpoints of those who may be potentially affected by such research, and how they approach the trade-off between societal benefit and personal intrusion. Mikal et al. investigated user attitudes toward the analysis of social media data for research, discovering “equivocal findings" in the literature [?].
Mikal et al. asked participants about their expectations of privacy with regard to monitoring depression at the population level, finding that most were accepting of it [?]. This was, however, largely conditional on the analysis being conducted in an aggregated and anonymised way, with no way of targeting a particular individual. This is echoed by Conway and O’Connor, who argue the need for a differentiation between automatic identification at the individual level and at the population level in order to protect the privacy of those involved [?].
Mikal et al. found that “many respondents felt as though a failure to protect online data constituted consent to have that data systematized and analyzed", however, the authors go on to consider what kinds of individuals would be less likely to protect their data [?]. Such a model would leave those with limited Internet literacy skills or those mistakenly oversharing their social data at risk of unwittingly providing implied consent.
Research has questioned the suitability of informed consent as it currently stands [?, ?, ?, ?, ?]. Luger and Rodden argue that consent, as currently conceived, cannot hope to meet the challenges posed by ubiquitous computing systems, and tensions exist between what is expected of consent and what is expected of a ubiquitous computer system. They describe consent as having been “stretched thin to the point of breaking" [?].
Kaye et al. argue that current proposals for the European Data Protection Regulations111since the writing of that paper, the GDPR has been ratified and will come into effect in 2018. require explicit consent, questioning the legality of broad consent methods, and that static systems for obtaining consent are no longer fit for purpose due to the changing nature of research and technology. New approaches are therefore needed to meet the ethical and legal requirements for consent while accomodating the dynamic nature of modern research [?].
Some of these concerns have been echoed by Morrison et al., who found that few participants were aware of participating in an academic trial, as specified in the Terms and Conditions document of a mobile application. Further means of informing users of researchers’ intentions to collect and analyse user data are necessary to behave in an ethical manner [?]. Such findings call into question the suitability of broad, one-off methods of collecting consent when dealing with online research. As Steinsbekk et al. articulate, at the core of the debate is “what it means to be ‘adequately informed’ and whether giving consent based on broader premises is valid or not" [?].
Researchers have investigated different mechanisms for coping with the above issues, such as using dynamic consent [?, ?, ?]. Dynamic consent is a framework for allowing participants to grant access to their personal data to researchers in a way that they can control. For example, a participant may decide to revoke access to their data, or customise their opt-in/opt-out preferences for participating in research.
Kaye et al. argue that, for participants, some of the advantages of dynamic consent include the ability to easily consent to new projects, alter consent preferences in real time, find out how their data has been used, and to set preferences about how they are kept informed. Researchers are said to gain more engaged participants, streamlined recruitment, improved public trust, and the knowledge that their research conforms to high legal standards [?].
While researchers have outlined the potential benefits of dynamic consent, Steinsbekk et al. dispute these, arguing that a convincing case has not been made. Their criticisms of dynamic consent include more frequent (and therefore more trivial) requests for re-consent, and an increased risk of the relationship between researcher and participant breaking down due to unmet expectations and a lack of reciprocity [?]. The authors argue that “broad consent combined with competent ethics review and an active information strategy is a more sustainable solution". Hutton and Henderson have also raised issues with dynamic consent, arguing that frequent requests for consent can lead to a significantly greater burden on the participant [?], which could frustrate participants and lead them to withdraw from the research.
Nissenbaum’s model of contextual integrity is a theoretical framework that attempts to prescribe “specific restrictions on collection, use, and dissemination of information about people” depending on presiding norms of information appropriateness and distribution [?]. Contextual integrity focuses on whether a flow of information is appropriate within a particular context, or whether a violation of privacy has occurred [?]. It has previously been used to explore privacy implications of social networking sites [?, ?, ?, ?, ?].
Hutton and Henderson have used contextual integrity to explore a new method of obtaining consent for social media research, described as ‘contextual integrity consent’ [?]. This middle-ground approach seeks the flexibility of dynamic consent, allowing users to choose what data is accessible and when, while reducing the burden of such data management. It works by inferring context-specific norms; a ruleset determining the appropriate flow of information. Individuals are asked explicitly about their willingness to share unless they clearly conform to or deviate from such a norm. In situations where a violation is found to have occurred, the framework rejects the practice in question and further actions can be taken, such as re-requesting consent from the user.
These norms depend on a number of factors, such as what the data are, to whom the data are flowing, and for what purpose they are being requested. For example, McNeilly et al. found that participants of a location sharing study were more likely to give the researchers access to their location data than a health study, and in both studies, photos were much less likely to be shared than ‘liked’ pages [?]. These norms can change over time, and individuals, organisations, and sections of society can each have their own expectations of what is appropriate.
Much of the work in dynamic consent has studied mobile or social network applications. An open question is how to apply these techniques to mining online health data. Our aim is to investigate contextual consent for health; can we capture, interpret, and act on context-specific norms within a healthcare domain? Further, can we improve on contextual consent by blending contextual integrity with machine learning techniques? Finally, such techniques are only useful if available to practitioners, and feedback from such deployments will inform further development of tools and models.
By predicting when social media users find it appropriate to share different types of data with different stakeholders, such as researchers and clinicians, the consent process will, firstly, better reflect the context in which data were created, and secondly, respect users’ preferences about which data should be made available and with whom. For instance, someone seeking support from their peers because they are anxious about an upcoming medical procedure might not want this shared with medical researchers, while others may want reports about side-effects of their medication to be viewed by clinicians.
We are focusing specifically within a health context for three reasons. Firstly, we believe that health data is an area where presiding norms exist. Secondly, getting these expectations of appropriate data flow correct is important due to the potential sensitivity of the data. Finally, given the vast number of studies using social media data within a medical domain, we believe this work is timely, and there is an opportunity to provide guidance and tools for the appropriate handling of social media data in health research to academics and practitioners alike.
We are currently planning two studies to further this investigation. The first will involve collecting a large corpus of data to use in a training and model evaluation phase. We will work with participants to determine what the main predictors are of appropriate data flow to different medical stakeholders. With this data, we will then derive machine-learning models for predicting consent with these different stakeholders. Finally, we will evaluate our model’s effectiveness in a follow-up study with participants in a social media medical support community.
Gomer et al. propose a similar semi-autonomous method of obtaining consent. Their proposed system trains a consent agent, uses this model to receive and accept or reject requests for participation, and allows users to review past decisions with the results re-training the model [?]. This method differs from ours as it trains and builds a model on a person-by-person basis, rather than building on the collective norms which shape contextual integrity. As such, it would require user training before it could be used. At the time of writing, no follow-up work has since been published.
Throughout this research, we aim to engage with clinicians, members of support communities and researchers to understand the concerns and interests of all parties. Due to the potential for sensitive topics and the extent in which a sharing violation may impact individuals, it is important that steps are put in place to greatly reduce risk. Such a predictive model will likely serve as a recommender system until the technique can be proven, the risks are understood, and confidence in such a method is developed. Working closely with all relevant stakeholders will help us to document and understand the risks involved and the challenges that will need to be overcome.
One of the largest technical challenges we face is the contextual nature of consent over items of data; One shared status update or photo may be seen as appropriate for researchers to access and utilise, whereas another may not be, making it difficult to design broad rules for what is acceptable and what is not [?, ?]. Furthermore, the given rules for any one individual may change over time, as social relationships and opinions evolve [?].
Participants may have concerns about sharing sensitive data with any autonomous system, given that it is an extension of the researcher. This raises new challenges about how such a classifier could be used, or even trained, if some participants are not willing to share certain data with researchers. Approaching this project, we must strike a balance between accuracy and appropriateness, as participants may or may not be comfortable with having the actual content of their social data being mined, however, it is unclear how effective prediction techniques can be using metadata alone.
A significant ethical consideration is the potential for erroneous and unacceptable norm violations taking place as a result of this research. In short, the sharing of a particular bit of data against the will of the owner may erode trust and call into question the entire project. Such a system may not be given a second chance. Our interim solution to this is to focus on a system which suggests appropriate sharing levels to the user for approval, rather than explicitly sharing the data autonomously. While this increases the burden of sharing content, it would still be less burdensome than asking the user to specifically choose the acceptable audience on each bit of content shared, such as with groups, lists, or circles. We believe that this limitation exists in any system that aims to reduce the burden of consent by automating the user’s decision.
The collection of data for such research will also likely prove to be a barrier to overcome. Collecting any data in an ethically aware way will introduce selection bias. Obtaining large sample sizes may also be difficult. We believe that this challenge is also inherent in such research, although there may be ways to minimise the impact without compromising the principle of informed consent.
People are increasingly sharing more and more information via social media, and as a result, such platforms are seen as rich data sources for researchers from various fields and backgrounds. This can lead to interesting research and discoveries, however, it also raises significant questions over informed consent in the age of social media and data mining. We argue that the successful exploitation of online personal health data requires new and usable methods of obtaining consent from the content creators. By deriving a new and usable method of obtaining consent in such circumstances, we hope that researchers can continue safe in the knowledge that the research is transparent and the participants are informed about how their data is being used and why.
We are currently planning studies to explore the contextual norms of data sharing within the context of health research in order to investigate whether or not this process can be automated. To develop a new method of obtaining consent, we need a broad range of input in the areas of social media, privacy, ethics, and law. We would like to invite experts, researchers and practitioners in these fields, as well as users of such services, to provide thoughts and contribute to discussions about how such methods may work.
Acknowledgements This work was supported by the Wellcome Trust [UNS19427].
-  L. Barkhuus. The Mismeasurement of Privacy: Using Contextual Integrity to Reconsider Privacy in HCI. In Proc. CHI 2012, pages 367–376. ACM, 2012.
-  d. boyd and K. Crawford. Critical questions for big data. Info., Comm. & Soc., 15(5):662–679, 2012.
-  M. Burke and M. Develin. Once more with feeling: Supportive responses to social sharing on Facebook. In Proc. CSCW 2016, pages 1462–1474. ACM, 2016.
-  M. Conway and D. O’Connor. Social media, big data, and mental health: current advances and ethical implications. Current Opinion in Psychology, 9:77–82, 2016.
-  S. Das and A. D. I. Kramer. Self-Censorship on Facebook. In Proc. ICWSM 2013, 2013.
-  M. De Choudhury, E. Kiciman, M. Dredze, G. Coppersmith, and M. Kumar. Discovering Shifts to Suicidal Ideation from Mental Health Content in Social Media. In Proc. CHI 2016, pages 2098–2110. ACM, 2016.
-  Company Info | Facebook Newsroom. http://newsroom.fb.com/company-info. Accessed: 2016-11-08.
-  R. Gomer, M. C. Schraefel, and E. Gerding. Consenting agents: Semi-autonomous interactions for ubiquitous consent. In Proc. UbiComp 2014 Adjunct, pages 653–658. ACM, 2014.
-  J. M. Hudson and A. Bruckman. “Go Away": Participant Objections to Being Studied and the Ethics of Chatroom Research. The Information Society, 20(2):127–139, 2004.
-  G. Hull, H. R. Lipford, and C. Latulipe. Contextual gaps: Privacy issues on facebook. Ethics and Inf. Technol., 13(4):289–302, 2011.
-  L. Hutton and T. Henderson. “I didn’t sign up for this!”: Informed consent in social network research. In Proc. ICWSM 2015, pages 178–187, 2015.
-  L. Hutton and T. Henderson. Towards reproducibility in online social network research. IEEE Transactions on Emerging Topics in Computing, page 1, 2015.
-  J. Jashinsky, S. H. Burton, C. L. Hanson, J. West, C. Giraud-Carrier, M. D. Barnes, and T. Argyle. Tracking suicide risk factors through Twitter in the US. Crisis, 35(1):51–59, 2014.
-  J. Kaye, E. A. Whitley, D. Lund, M. Morrison, H. Teare, and K. Melham. Dynamic consent: a patient interface for twenty-first century research networks. European Journal of Human Genetics, 23(2):141–146, 2014.
-  J. A. Lee, C. Efstratiou, and L. Bai. OSN Mood Tracking: Exploring the Use of Online Social Network Activity As an Indicator of Mood Changes. In Proc. UbiComp 2016, pages 1171–1179. ACM, 2016.
-  N. Lee. Trouble on the radar. The Lancet, 384(9958):1917, 2014.
-  J. Li and C. Cardie. Early Stage Influenza Detection from Twitter. arXiv preprint arXiv:1309.7340, 2013.
-  Y. Liu, K. P. Gummadi, B. Krishnamurthy, and A. Mislove. Analyzing Facebook Privacy Settings: User Expectations vs. Reality. In Proc. IMC 2011, pages 61–70. ACM, 2011.
-  E. Luger and T. Rodden. An informed view on consent for UbiComp. In Proc. UbiComp 2013, pages 529–538. ACM, 2013.
-  M. Madejski, M. Johnson, and S. M. Bellovin. A study of privacy settings errors in an online social network. In Proc. SESOC 2012, pages 340–345. IEEE, 2012.
-  N. C. Manson and O. O’Neill. Rethinking Informed Consent in Bioethics. Cambridge University Press, 2007.
-  S. McNeilly, L. Hutton, and T. Henderson. Understanding ethical concerns in social media privacy studies. In Proc. CSCW 2013 Workshop on Measuring Networked Social Privacy: Qualitative & Quantitative Approaches, 2013.
-  J. Mikal, S. Hurst, and M. Conway. Ethical issues in using Twitter for population-level depression monitoring: a qualitative study. BMC Medical Ethics, 17(1), 2016.
-  M. A. Moreno. Associations Between Displayed Alcohol References on Facebook and Problem Drinking Among College Students. Archives of Pediatrics & Adolescent Medicine, 166(2):157–163, 2012.
-  M. A. Moreno, L. A. Jelenchick, K. G. Egan, E. Cox, H. Young, K. E. Gannon, and T. Becker. Feeling bad on Facebook: depression disclosures by college students on a social networking site. Depression and anxiety, 28(6):447–455, 2011.
-  A. Morrison, D. McMillan, and M. Chalmers. Improving consent in large scale mobile HCI through personalised representations of data. In Proc. NordiCHI 2014, pages 471–480. ACM, 2014.
-  C. Munteanu, H. Molyneaux, W. Moncur, M. Romero, S. O’Donnell, and J. Vines. Situational ethics: Re-thinking approaches to formal ethics requirements for human-computer interaction. In Proc. CHI 2015, pages 105–114. ACM, 2015.
-  H. F. Nissenbaum. Privacy as contextual integrity. Washington Law Review, 79(1):119–157, 2004.
-  Ofcom. Adults’ media use and attitudes Report 2016, 2016.
-  A. G. Reece and C. M. Danforth. Instagram photos reveal predictive markers of depression. CoRR, abs/1608.03282, 2016.
-  R. Sar and Y. Al-Saggaf. Contextual integrity’s decision heuristic and the tracking by social network sites. Ethics and Information Technology, 2013.
-  M. Sleeper, R. Balebako, S. Das, A. L. McConahy, J. Wiese, and L. F. Cranor. The post that wasn’t: Exploring self-censorship on Facebook. In Proc. CSCW 2013, pages 793–802. ACM, 2013.
-  L. Solberg. Data mining on Facebook: A free space for researchers or an IRB nightmare? University of Illinois Journal of Law, Technology & Policy, 2010(2), 2010.
-  K. S. Steinsbekk, B. Kare Myskja, and B. Solberg. Broad consent versus dynamic consent in biobank research: Is passive participation an ethical problem?;. European Journal of Human Genetics, 21(9):897–902, 2013.
-  M. Zimmer. “But the data is already public”: on the ethics of research in Facebook. Ethics and Information Technology, 12(4):313–325, 2010.