Navigating ethical challenges in psychological research involving digital remote technologies and people who use alcohol or drugs

Abstract

Digital and remote technologies (DRT) are increasingly being used in scientific investigations to objectively measure human behavior during day-to-day activities. Using these devices, psychologists and other behavioral scientists can investigate health risk behaviors, such as drug and alcohol use, by closely examining the causes and consequences of monitored behaviors as they occur naturalistically. There are, however, complex ethical issues that emerge when using DRT methodologies in research with people who use substances. These issues must be identified and addressed so DRT devices can be incorporated into psychological research with this population in a manner that comports to the ethical standards of the American Psychological Association. In this paper, we discuss the ethical ramifications of using DRT in behavioral studies with people who use substances. Drawing on allied fields with similar ethical issues, we make recommendations to researchers who wish to incorporate DRT into their own research. Major topics include (a) threats to and methods for protecting participant and nonparticipant privacy; (b) shortcomings of traditional informed consent in DRT research; (c) researcher liabilities introduced by real-time continuous data collection; (d) threats to distributive justice arising from computational tools often used to manage and analyze DRT data; and (e) ethical implications of the “digital divide”. We conclude with a more optimistic discussion of how DRT may provide safer alternatives to gold standard paradigms in substance use research, allowing researchers to test hypotheses that were previously prohibited on ethical grounds.

The “Internet of Things” refers to the interconnected network of physical devices embedded with sensors, software, and other technologies that communicate and exchange data with each other over the internet, enabling a seamless integration of the physical and digital worlds (Gubbi et al., 2013). This technological shift has transformed how humans utilize technology to track behavior and physiology through rapid advances in mobile computer technologies. The miniaturization of computers has led to the wide availability of unobtrusive portable devices that people can use to monitor various aspects of their behavior, physiology, and environment. These devices record momentary information using a variety of sensors, process it using onboard components, and communicate it through robust wireless networks. The result is continuous, detailed data streams providing rich descriptions of a person’s experience in real time. Digital remote technologies (DRT) are poised to fundamentally shift psychological research by allowing objective measurement of experiences previously only assessed using self-report questionnaires or in laboratory settings. DRT devices have been used by psychologists to objectively measure the lived experiences of human participants in research involving both clinical (Low et al., 2021) and nonclinical groups (Fairbairn et al., 2018), offering a method for studying the dynamics of naturalistic behavior (Insel, 2017).

The availability of DRT is already facilitating psychological research with people who use alcohol or other drugs; it allows researchers to capture subjective and objective momentary data from people as they engage in myriad behaviors in their natural environments (see online supplement for examples of studies using DRT in substance use research). DRT can solve many of the challenges inherent in research on substance use behavior by facilitating the collection of rich temporally sequenced data that can elucidate the environmental and personal causes and consequences of substance use. DRT even provides a method for objectively measuring substance use in real time (e.g., transdermal alcohol sensors; Carreiro et al., 2020). While the development of a new set of methodological tools to advance the study of human behavior is exciting, DRT-involved research is fraught with ethical conflicts that must be reconciled with the American Psychological Association’s (APA) Ethical Code (APA, 2016).

In this paper, we describe key ethical conflicts inherent to human research involving DRT and recommend ways to resolve them. We do so from the perspective of a team of clinical psychologists who use DRT in our research aimed at improving the lives of people who engage in unhealthy drug or alcohol use. Embedded in this perspective is a commitment to protecting research participants’ rights to privacy and their physical and emotional wellbeing, with specific attention to the medico-legal issues that complicate recovery from SUD. Although we write from the perspective of research psychologists studying substance use, many of the ethical issues raised — and solutions suggested — will apply to other research areas where psychologists wish to utilize DRT in their research. We build on previous reviews that focused broadly on the ethics of digital behavioral health (e.g., Nebeker, Bartlett Ellis, & Torous, 2019; Nebeker, Torous, & Bartlett Ellis, 2019).

Ethical Codes Governing Psychological Research

The American Psychological Association (APA) Ethics Code is composed of 5 aspirational principles and 10 enforceable ethical standards (APA, 2016). Only one standard (Section 8) specifically addresses ethical conduct in research activities, although most other standards have indirect bearing on psychologists’ behavior in research contexts. The APA’s ethical standards are not legally binding but generally comport with legally binding federal policy for human participant protection (i.e., the Common Rule; Protection of Human Subjects [2018]). Section 8.01 of the Ethical Standards (and federal law) formalizes the requirement that psychologists who wish to conduct research with humans will, in most cases, be required to seek approval from Institutional Review Boards (IRB), which are responsible for ensuring that researchers protect the welfare of human participants as described in the Common Rule. Consequently, the Common Rule – and the Belmont Report from which it is derived – guides most psychological research, although the APA Ethics Code can provide useful heuristic guidance for resolving ethical dilemma in cases of conflicting rules (Sales & Folkman, 2000). In this paper, we will refer to both sets of codes and their moral foundations in our discussion of relevant ethical issues.

Sources of risk in DRT research involving people who use drugs or alcohol

The following sections review five categories of ethical risk inherent in DRT research with human participants. Table 1 provides an overview of these topics and summarizes how risk in each category can be conceptualized and assessed. These five categories are not meant to be an exhaustive discussion of all ethical considerations germane to this type of research. We do not discuss, for example, linkage of DRT with electronic health records or ethical issues related to the governmental regulation of digital health tools. Rather, we selected five topics that we see as both highly relevant to behavioral substance use research and potential sources of incongruence between currently accepted research regulations and the higher-order ethical principles on which they are based. These topics were also chosen because of their correspondence with public interest in the ethics of digital health surveillance. It is important that the field works proactively to resolve pressing ethical issues – particularly those of high public interest – to maintain public trust and continue to lead by example.

Table 1.

Summary of selected critical ethical issues in the use of DRT in research with people who use alcohol or drugs

Topic	Relevant APA Ethical Code principles	Key threats to and limitations of current ethical principles and norms	Illustrative questions for evaluating risk
1. Privacy and confidentiality	• Beneficence and nonmaleficence • Integrity • Respect for People’s Rights and Dignity	• Data sharing may threaten participant anonymity • Partnerships with consumer technology companies reduce researcher control over data • Privacy of community members may be compromised	• Who will have access to participants’ DRT data? • Does the DRT data theoretically allow for reidentification of participants? • How much harm would participants incur if linked to their DRT data?
2. Informed consent	• Respect for People’s Rights and Dignity	• Consent provided at time of collection may not cover future uses of data • Bystander private data may be recorded without their consent	• How might the collected DRT data be used in the future? • How much exposure will bystanders have to the data collection systems? • What harm could befall bystanders whose data are collected?
3. Analytic bias in DRT data processing	• Justice • Beneficence and nonmaleficence	• Machine learning techniques may perpetuate preexisting biases	• How susceptible to hidden bias are the planned analytic techniques? • To what extent would unfair predictions negatively impact non-majority group members?
4. Real time data collection and safety monitoring	• Beneficence and nonmaleficence	• Continuous data acquisition may necessitate logistically challenging intervention to protect participants • Uncertainty regarding what constitutes evidence of risk to wellbeing	• Do any collected data have potential to legally or ethically compel the researcher to intervene to protect participant safety? • What is the risk level of the participant population?
5. Distributive justice	• Justice	• Interventions and other knowledge gained through DRT research may unequally benefit those who already engage with technology and have access to digital systems	• How will the resulting knowledge or deliverable product be distributed among the population? • Does the research design allow for identification of barriers of use in vulnerable groups?

Privacy and Confidentiality

Respect for privacy is a core ethical tenet of human research, and this is especially true of research on substance use. This right is rooted in Principles A (Beneficence and Nonmaleficence), C (Integrity), and E (Respect for People’s Rights and Dignity) of the APA Ethics Code and the corresponding principles of the Belmont Report (i.e., Respect for Persons, Beneficence). The right to privacy affirms a person’s right to control how information describing them is distributed by choosing what aspects of their thoughts, behaviors, and physiology are shared with others. Confidentiality is a closely related concept that describes, in the research context, the right of the participant to maintain private any privileged information that is shared with the researcher as part of their research participation (Sales & Folkman, 2000). Efforts to protect participant privacy are of paramount importance for research involving people who use substances due to the legal status of many drugs and general stigma against substance use, especially in marginalized groups for whom the risks of privacy breaches are greater (e.g., people of color who use drugs; Walters et al., 2023). Qualitative research examining concerns associated with DRT among people who use drugs indicates that legal and social problems that may result from privacy breaches are a top concern for this population (Roth et al., 2017).

How are privacy and confidentiality typically protected in behavioral research involving people who use substances? Security requirements are typically based on federal or state regulations (e.g., HIPAA Privacy Rule) and institutional interpretation thereof. The specifics of these regulations are beyond the scope of this paper, but generally involve systems to protect against unauthorized data access and the strategic separation of personal identifying information from physical or electronic data records that include sensitive information. Requirements specify minimal necessary security measures for both electronic (e.g., transport layer security, advanced encryption standard) and physical (e.g., storage under 2 sets of locks) storage or transfer. Certificates of Confidentiality, which protect researchers (and participants) from forced disclosure of identifiable research data, are currently issued for all research funded through the National Institutes of Health (NIH, 2017).

Protecting the privacy and confidentiality interests of participants in DRT-involved research is comparatively complicated and requires additional safeguards. Descriptions of participant behavior that could place them at risk for serious harm may emerge from data streams that initially appeared to be low risk. For example, many of the information systems that support DRT incur additional security risks as data are transmitted across multiple devices with relatively higher exposure to online systems, which can raise the risk of data breach (McLeod & Dolezel, 2018). Because of the comprehensive and often continuous nature of DRT-based data collection, researchers will be required to think flexibly about what constitutes “sensitive” data. For many potential uses of DRT in substance use research, data collection extends beyond the individual participant to nonparticipants and their community at large.

Data sharing.

In recent decades, research funding organizations have instituted data sharing policies, requiring data availability for funding. The NIH implemented a policy mandating data sharing for all research generating scientific data (Office of the Director, NIH, 2020). This policy reflects a broader cultural shift toward open science (Jorgenson, Wolinetz, & Collins, 2021), driven by ethical arguments that data sharing maximizes insights and societal benefits while increasing scientific rigor and transparency, aligning with Principal C (Integrity) of the APA Ethics Code. Standard 8.14 requires that psychologists “do not withhold data from competent professionals who seek to verify the substantive claims through reanalysis…”, although historically this standard is inconsistently applied (Wicherts, Borsboom, Kats, & Molenaar, 2006).

There are benefits to making data collected using DRT generally available to the research community. Many types of DRT produce theoretically rich and highly flexible data records that can be analyzed using sophisticated computational approaches to answer important research questions that were not yet imagined when the data were originally collected. This is especially true for DRT that produce detailed records of fundamental aspects of human behavior (e.g., movement, sleep) that have applications in numerous subfields of psychology. This idea is exemplified by recent studies using existing data resources that include DRT records. For example, Lyall and colleagues (Lyall et al., 2018) recently used accelerometer data from a subset of UK Biobank participants (n = 91,105) to relate disruptions in circadian rhythm with psychopathology and subjective wellbeing. These types of “big data” investigations are only possible with biobank-style data resources that are made available to the broader research community. Data collected via DRT will likely become more common in these types of large-scale investigations aimed at producing data resources for the research community, especially as appropriate types of DRT (e.g., activity trackers) move towards data standardization.

Protecting participants’ privacy is a key goal in developing ethical data sharing guidelines, and the risk of privacy loss increases as data are made more easily accessible. Data with minimal access limitations, such as publicly available data, are vulnerable to deductive reidentification (Kaye, Heeney, Hawkins, de Vries, & Boddington, 2009). This type of privacy breach occurs when a deidentified record is linked to a participant’s identity by matching aspects of that record to publicly available datasets containing identifying information, ultimately creating a single linked record that reveals both the participant’s identity and personal sensitive information that was initially collected under the promise of anonymity. To protect participants from privacy loss associated with reidentification, there are strict regulations which govern open sharing of research data. Most types of data sharing require that they be shared in anonymized (i.e., data record cannot be linked to a participant’s identity) or deidentified form (i.e., data record contains no explicit identifying information but can theoretically be linked back to a participant). Anonymized data can be made publicly available, while sharing deidentified data typically requires additional safeguards (e.g., restricted access). However, effectively anonymizing data is not a simple task and requires deep domain and technical knowledge; it is not clear to what extent data that are anonymized today will withstand reidentification techniques developed in the future (Rothstein, 2010).

Deidentification or anonymization of DRT data is complicated when the data record contains detailed accounts of behavior. It is difficult to anticipate the computational techniques that might be used in pursuit of deductive identification (Gürsoy et al., 2022). This is not just a hypothetical problem – publicly released data records which were thought to be anonymous at the time of release have been linked to people’s identities using only information available online (Zimmer, 2020), and researchers have shown that physical activity data collected using accelerometers can be used to reidentify participants in a large national dataset (Na et al., 2018). We therefore encourage using controlled data access strategies rather than releasing DRT data publicly in cases where deductive identification cannot be ruled out with certainty. Controlled access strategies that are commensurate to the risk level of the data being protected should be applied. For example, data access can be granted to vetted individuals, or stricter measures can be taken such as requiring users to analyze data remotely within virtual computing environments (McGuire & Beskow, 2010). If alternative measures can be applied that positively eliminate any risk of identification, then public release is preferable in the interest of open science.

What strategies can be used to protect against reidentification if data are released publicly? We can look to associated research fields (e.g., genomics) that have grappled with balancing the conflicting values of privacy protection and open science. Researchers in these fields have developed methods of systematically modifying data such that it retains informational value but obscures any information that could be used to identify the participant (Bonomi, Huang, & Ohno-Machado, 2020). Algorithmic solutions that strategically degrade data in a way that protects participant privacy while retaining its informational value have been developed (e.g., Lin & Wei, 2009). These algorithms are already being applied to DRT data (Abdrashitov & Spivak, 2016), although bringing this approach to maturity in all cases will be challenging given the relative diversity of DRT data compared to genetic data. Simpler data transformations may suffice in some cases (e.g., recording relative rather than absolute GPS location data). Although these tools are beneficial in that they facilitate open access to data that could not be ethically shared otherwise, they place additional regulatory, financial, and computational burdens on researchers.

Estimating the likelihood of a privacy breach and the resulting harm will vary widely between studies and must be considered on a case-by-case basis. Factors relevant to determining the level of risk include the nature and extent of DRT data that will be collected, sensitivity of associated (non-DRT) data records, vulnerability of the participants, extent of planned data sharing, data security plans, and security of IT systems that will be used. For example, consider an observational study examining the association between alcohol use and physical movement in which wrist-worn accelerometers are used to track minute-to-minute activity patterns in a group of healthy adult social drinkers. The DRT data collected (minute-to-minute changes in physical activity) is presumably low risk in that a participant would be unlikely to incur significant harm if their identity were linked to the data record. It also is unlikely that a participants’ identity could be deduced simply by using activity level data. Even if a participant were identified through their data, the behavior under study (normal social drinking) is not illegal and the resulting harm likely would be minimal. Compare this with a hypothetical study in which GPS data are used to identify location-based risk factors for relapse in a group of people who inject drugs and are HIV positive. In this case, the DRT data collected are high risk because participant’s identity can be easily deduced from a comprehensive GPS record, and linking participants to records of injection drug and HIV status would heighten their risk of harm considerably.

Maintaining anonymity is extremely important in behavioral research on substance use because the data often describe sensitive or illegal behaviors that might cause harm to participants if confidentiality is broken. Even if safeguards can ensure that participants’ anonymity is maintained, it is possible that some participants may not wish for data describing some of their more private/personal behaviors to be distributed. It is important that participants are given as much information as possible during the informed consent process about how their data could be used (e.g., in what form might their data be shared [e.g., raw or minimally processed form], who will have access to their data, and can it be used for purposes not explicitly stated in the initial consent form).

Nonparticipant privacy.

IRBs in the United States evaluate the ethicality of research according to a Eurocentric value system that emphasizes individualism and self-determination (Parker, Pearson, Donald, & Fisher, 2019). Consequently, protocols are evaluated in terms of the risk incurred by the individual participant; the potential risks to the community in which participants are embedded are not routinely considered (Sieber & Tolich, 2013). This individualist framework fits well for many of the common methods used in psychological research on substance use: participants self-report on their own history of use or complete laboratory sessions in which a drug is administered or behavioral tasks are performed. Rarely is specific, identifiable information about other members of the community shared with the researcher.

Currently, there is little nonparticipant protection included in regulatory codes with authority over research activity. In the Common Rule, nonparticipants are not considered the subject of investigation (CRF 46.101) and therefore may not be afforded the protections afforded to designated research participants. The APA ethical standards do not explicitly require psychologists to protect nonparticipants in their research activities, although we are encouraged to protect “the community”. The APA standards make specific mention of requiring specialized consent from participants when voice or image recordings are made, provided that these records are not made in a public place (8.03). Use of some DRT devices, particularly those capturing speech or digital images, will be regulated by relevant federal and state laws that limit the conditions under which a person can be legally recorded (Robbins, 2017).

Nonparticipant privacy in research involving DRT involves larger ethical, and potentially legal, concerns because some of these devices produce data records describing the behavior of nonparticipants who have not consented to participate in the research. Nonparticipants might be included in the research record simply because of their physical or relational proximity to the participant. To illustrate this point, we will consider a study published in the Journal of Abnormal Psychology in which nonparticipant characteristics were explicitly measured using DRT and used to test a hypothesis on the social determinants of alcohol reward. Fairbairn et al (2018) used 2 DRT devices to track alcohol consumption (transdermal alcohol sensor), social circumstance (digital camera), and mood (self-report on smart phone) in a group of moderate drinkers. Participants were asked to take a picture of their surroundings when they were randomly prompted to complete a mood assessment. Images were stored on the smart phone and later coded on various characteristics by the research team and the participant, including their level of familiarity with the nonparticipant. These images were key parts of the research record and contained identifying information (facial images) of nonparticipants.

The above study is an innocuous example of how nonparticipants’ private data can be included in data records but nicely illustrates how such investigation may place community members at risk for breached privacy without their awareness or consent. The investigators used several important safeguards to reduce risk to nonparticipants. Participants were cued to manually take the pictures, so they likely conformed to social privacy norms for picture taking (e.g., not taking pictures in public restrooms). Further, the behavior under study (alcohol use) is well-accepted in the United States, so it is unlikely a nonparticipant would incur even moderate harm from the release of images of them drinking alcohol in a public space. Other studies, however, may put nonparticipants at higher risk for harmful breach of privacy. Small changes to the protocol (e.g., using this methodology to examine the effects of social circumstance on cocaine use) would have profound implications for the level of risk incurred by nonparticipants. It is therefore important that researchers using these methods consider 1) the likelihood that the set of DRT data collected could be used to identify members of the community, 2) the amount and type of information collected on the nonparticipant, and 3) the amount of harm that would result from release of the information.

Several considerations are relevant for evaluating these factors and determining level of risk. Video and audio recording via DRT devices clearly produce data that could be used to identify nonparticipants; however, the risk associated with other DRT data could be less obvious. For example, studies have used specialized software to record data from sensors in participants’ smartphones (Bae, Chung, Ferreira, Dey, & Suffoletto, 2018), which can digitally tag the participant as being near or communicating with others via Bluetooth handshakes, shared WIFI connections, or keyboard inputs on text-based messaging apps. All these data streams can potentially contain identifying information of nonparticipants who are communicating with or in close physical proximity to the participant (Givehchian et al., 2022). In terms of the informational content of data streams and associated harms to nonparticipants if their identity were tied to the data, images and audio recordings are again associated with the highest level of risk, particularly if samples are recruited based on their engagement in risky or illegal behaviors. It also is important to consider what can be inferred if data streams are combined. For example, the combination of a proximity-based identifier (i.e., smartphone Bluetooth handshake) and a drug use biosignal in the primary participant may imply that the nonparticipant also used the drug.

We can again look to allied fields for guidance on how to protect the privacy of nonparticipants. One methodological solution is to provide participants with a method for temporarily pausing data collection, although this relies on the participants’ judgment to protect nonparticipants’ privacy. Technology-based solutions involve modifying the devices or software to collect data in a way that the offending data are not stored or are quickly deleted. For example, facial recognition algorithms can be applied on an “allow list” basis so that all faces who have not opted in to being recorded are blurred immediately after an image is collected (Ra, Lee, Miluzzo, & Zavesky, 2017). Predictive models developed using deep learning can automatically identify photographs depicting specific activities, which can then be degraded to preserve privacy during sensitive activities (Dimiccoli, Marín, & Thomaz, 2018). Rule-based location guidelines (i.e., geofencing) can disable data collection in specific areas. These computational solutions can be applied automatically, making them attractive for large-scale data collection efforts. However, they do require considerable computational resources and technical expertise to implement. Further, some facial and voice recognition algorithms show racial and gender biases; implementing these methods to protect nonparticipant privacy may confer unequal protection to racial minority groups and women (Koenecke et al., 2020).

Privacy and technology companies.

Many DRT devices used in behavioral research come from industry partners who develop and market them as consumer wellness devices, such as heart-rate monitors, step counters, and sleep monitors. These devices often lack regulation as health devices, leaving them exempt from privacy laws that protect consumers’ health data (Klosowski, 2021). Incompatibility exists between privacy prioritization in health and research sectors and the data economy of technology companies. User data is frequently treated as a product, sold and resold without informing the original seller or the individual described, increasing the risk of data breaches. There have been documented cases of technology companies releasing deidentified sensitive data that is linked back to individuals, causing significant harm (Klosowski, 2021).

Using consumer DRT in research raises ethical concerns related to informed consent and participant privacy, as data protection policies by private companies may not meet minimum ethical standards for human research. Participants should be informed about the researcher’s lack of ultimate control over their data and made aware of the product’s commercial terms of use. Implementing strong deidentification practices, including using random identifiers, is essential. A long-term solution is for psychologists and researchers to advocate for stricter privacy policies regarding consumer data, aligning scientific ethics with technology companies’ practices. The European Union’s General Data Protection Regulation (GDPR) exemplifies broad privacy protections that have been replicated in other regions. Advocates, such as the Electronic Frontier Foundation, continue to push for stronger privacy protections, and we encourage psychologists to support these effort through professional advocacy.

Informed Consent

Informed consent is the formal process by which psychologists ensure that research participants understand the purpose, procedures, risks, benefits, and requirements of the research in which they are agreeing to participate (Sales & Folkman, 2000). This is a cornerstone of research ethics and a prerequisite for supporting participants’ right to self-determination. The exchange of information that occurs during the informed consent process allows volunteers to weigh the risks and benefits of their participation through the lens of their own value system. In traditional hypothesis-driven research, the informed consent process involves a participant reviewing and signing a formal agreement (i.e., informed consent form for adults; assent form for minors). This operationalization of establishing informed consent works well for the traditional laboratory and questionnaire-based research described in previous sections. There are, however, several shortcomings of this model when applied to substance use research involving DRT that must be rectified if informed consent is to be faithfully enacted in these types of studies.

Protecting agency.

Continued participation in a research study is contingent on the participant’s continued consent over the entire course of the research project. In the traditional human research paradigm, the right to withdraw consent without penalty meant that participants were free to stop participating in a study prior to its completion, consistent with the emphasis placed on volunteerism and freedom of choice in foundational research ethics frameworks (McGuire & Beskow, 2010).

As previously described (see Data Sharing), many studies that utilize DRT produce large quantities of data that describe foundational aspects of human behavior, making them appealing data resources for subsequent analysis. Detailed records of physiological and behavioral data collected using DRT wearables (e.g., Fitbits) have been included in major research initiatives that are designed to produce multimodal data records that can be used for subsequent investigation (e.g., Adolescent Brain Cognitive Development study; Bagot et al., 2018). Through collaborations with major technology companies, researchers have accessed DRT data from large (> 100,000) cohorts of users and analyzed data to answer research questions to which users did not explicitly agree in the end user agreement (Radin, Wineinger, Topol, & Steinhubl, 2020). Actigraphy data in the UK Biobank have been used in wide-ranging health applications such as characterizing sleep phenotypes (Katori, Shi, Ode, Tomita, & Ueda, 2022). Importantly, participants did not consent to these specific research questions. Rather, they provided broad consent agreeing to unspecified future use of their data.

“Behavioral biospecimens” collected using DRT and included in persistent datasets share similarities with tissue samples in biobank records, leading to comparable ethical issues. Informed consent models in biobank-style repositories pose ethical concerns regarding participants’ continued control over their data (Rothstein, 2005). Participants are not informed of all research studies that are being conducted using their data; it is inevitable that some participants’ data will be used for research that they do not approve of due to deeply held beliefs (Rothstein, 2005), such as religious or cultural beliefs that are incompatible with scientific investigation into certain aspects of human behavior (e.g., sexuality). Some investigations might even work against the participants’ financial interests (e.g., raising health insurance premiums; Raber, McCarthy, & Yeh, 2019). The enduring nature of these data resources means participation continues long after initial data collection, making it difficult or even impossible to withdraw a participant’s data from future research due to technical challenges, such as multiple existing copies (McGuire & Beskow, 2010).

To increase the amount of control that participants have over their data in this context, bioethicists have developed novel models of informed consent that deal directly with the challenges imposed by large-scale enduring biosamples. We provide description and list advantages and disadvantages of these different models in an online supplement. Attitude research finds that many people prefer more control over their data than is provided with broad consent (Shen et al., 2022); several of the alternative models presented in the online supplement (e.g., dynamic consent, meta consent) address this issue by providing mechanisms for participants to continue to communicate their preferences to researchers about how their data are used after collection is complete. Participants are generally supportive of this continued engagement and appreciate having persistent control of their data, although some participants regard the ongoing communication as unnecessary (Spencer et al., 2016). Determining the level of continued engagement necessary to protect participant agency will likely depend on the scope of data collected and its potential for use in data mining applications.

Expanding consent to match the scope of DRT research.

Many applications of DRT methods in substance use research entail some risk to individuals with whom participants interact directly (see Nonparticipant Privacy) and members of the communities in which participants are embedded. Moreover, because some DRT methods involve collection of data that describe communities themselves (e.g., images of community spaces and members; descriptions of geographical regions via GPS), findings may have profound implications for the well-being of the community in which the research takes place. This again highlights the mismatch between individualist models of consent – wherein consent is sought only from the participant who is directly involved in the research process – and the reality of diffuse data records created in many DRT applications. It is necessary to rethink conceptual models of informed consent and create flexible methods for scaling the informed consent process to better accommodate this aspect of DRT research. Ignoring this issue is tantamount to presumed consent, where it is assumed that people are obligated to participate in the research.

Alternatives to individualist models of informed consent include community-based models that were developed and used by community-based participatory researchers (Sieber & Tolich, 2012). Communities – defined in this framework as groups of people sharing a common characteristic, such as a shared trait, interest, or geographical location – are treated as the unit of investigation. Informed consent is sought from the community; this can be established through partnerships with community leaders, community advisory boards, focus groups with community members, or through numerous other strategies. This ensures that research is conducted in a way that incorporates the values of the community and safeguards against research being designed or disseminated in a way that harms a community, particularly among vulnerable and marginalized communities (Parker et al., 2019; Wensley & King, 2008).

The optimal approach for seeking adequate informed consent from individuals or groups whose interests may be affected by a research project will depend on the methods and goals of the research project in question. In some cases, a dyadic informed consent agreement between the participant and researcher will be sufficient. For example, a study in which DRT data collection is limited to psychophysiological outcomes would expose nonparticipants to no more risk than a traditional laboratory study. In other cases, DRT research may require deep engagement with a specific community from the ground up. A study involving passive collection of visual images to study temporal characteristics of health risk behaviors among people experiencing homelessness who inject drugs would expose community members to significant risk. This study would require extensive community engagement to ensure that the research is acceptable to those involved and placed at risk.

In many DRT use cases, the appropriate approach to consent will fall somewhere between these extremes, so there will be an identified primary participant assuming proportionally greater risk than nonparticipants whose risks are derived from their relationship with or physical proximity to the primary participant. Researchers should consider the degrees of involvement and distribution of risks broadly, with specific attention to those outside of the traditional research/participant dyad. Study design choices might also be used to limit the scope of data collection and eliminate the need to seek consent from the community. Following is a case study in which the need to obtain informed consent from the community is managed using a combination of design choices, community engagement, and expanded consent-seeking.

One member of the authorship group (omitted for anonymous review) with colleagues¹ is currently conducting Project Connect, a field study examining the role of peer group social influence on alcohol use in young adults. Primary participants in this study are recruited and consent to responding to smartphone-based surveys for 21 days. They will complete a baseline assessment battery during which they will identify same-age peers with whom they drink alcohol regularly.

The objective of this research is to evaluate the feasibility, acceptability, and technical success of passively detecting participants’ interactions with their peers in the interest of studying the social context of participant alcohol use – although these methods could be useful for other research in which social contact is relevant. The work will attempt to create a record of interactions – operationalized in terms of physical proximity – between the primary participant and their peers who were nominated during the baseline appointment (termed “peer participants”). There are several possible methods for recording such interactions, such as general recording of Bluetooth handshakes between smartphones, continuous audio or video recording, or collating GPS records of the primary and peer participants. These solutions, however, would entail extensive collection of identifiable data from unconsenting bystanders, or require collection of sensitive private data from the peer participants. Rather than utilize methods that place nonparticipants at risk, the investigators will utilize a system that achieves the goals of the study while limiting the amount of data collected. To measure participants’ proximity to peers, the study will utilize “beacons” that transmit a unique identifier via Bluetooth that can be identified by an application on the primary participant’s Smartphone when the beacon is in proximity.

Because this method requires that the peers carry these Bluetooth beacons, it is important they complete informed consent for participation. However, because peers are unaware of the research and have not volunteered to be contacted, study staff are prohibited from directly contact the peers. They therefore ask the primary participant to facilitate contact at the time of the baseline appointment. This requires some triangulation in which the participant contacts the peer and asks if they would be interested in learning about the study. If the peer agrees, then they are put in contact with the researcher. The peer participant then receives information about the aspects of the study that concern them, completes informed consent, is provided with the Bluetooth beacon and receives compensation that is commensurate to their role in the study. Importantly, the process of seeking the consent of and working with peer participants occurs separately from the primary participants to maintain the privacy of peer participants.

In preparation for this study, the principal investigators conducted focus groups to examine how potential peer participants understood the devices and potential sources of discomfort. Initially, members of these groups reported variable attitudes towards the devices, with some group members reporting discomfort due to the misunderstanding that their location would be tracked electronically (there is no location record) or that somehow the beacon would convey their location to the participant (it does not). Facilitators in these groups were able to provide education on the extremely limited scope of the data collected using the devices. The group format proved advantageous as it helped identify concerns or points of confusion shared by participants and their peers.

Analytic Considerations: Machine Learning and Justice

Many types of DRT produce data records that are orders of magnitude larger than those obtained in most laboratory or questionnaire-based research. It is impractical or, in many cases, impossible for humans to directly interpret the resulting data records, especially if a platform utilizes multiple sensor streams. Rule-based transformations and other data reduction techniques have their place in specific DRT applications; however, researchers often turn to high-complexity computational modeling to manage these massive datasets. These sophisticated analytic techniques are critical for realizing many of the advantages of DRT: they can be used to identify patterns in multivariate time-series or other high-dimensional data. In consumer devices, predictive models derived from machine learning methods drive the powerful detection and evaluation functions of the devices (e.g., step-counts, sleep quality indications). In bespoke applications relevant for research and niche clinical purposes, predictive models can be developed from the ground up to detect study-defined outcomes using machine learning approaches, informing targeted deployment of individualized interventions. In substance use research involving DRT, machine learning is being used along these lines to, for example, anticipate heavy drinking episodes using previous-day smartphone data (Bae et al., 2018).

Despite the powerful capabilities of machine learning methods for dealing with the types of data often collected using DRT, there is considerable risk that the resulting models will produce unfair or biased predictions.² Many of the most powerful machine learning algorithms are complex and create “black box” models that generate predictions using inscrutable decision rules. Unlike classical statistical techniques, a researcher using these models cannot simply examine model coefficients to understand how the model generates a certain prediction. In the cases of state-of-the-art deep learning approaches, which are favored for the types of complex time series data structures produced by many types of DRT (Dempster, Petitjean, & Webb, 2020), variable inputs used in these models are transmitted across multiple hidden layers in which the original inputs are combined and modified using nonlinear transformation functions. The resulting decision rules are complex and through these high-level interactions and transformations the algorithms may learn to base decisions on emergent variables that are highly correlated with protected class membership. Predictive models can perpetuate unfairness is by generating less accurate predictions for members of specific groups.

How might unfairness in machine learning models affect research involving DRT with people who use substances? Consider a hypothetical example. An academic consulting group develops a DRT-based monitoring platform for individuals recently abstaining from opioid use after completing residential treatment. The platform, using a low-cost smart wristband, aims to detect both opioid intoxication and overdose events in real time through physiological (i.e., pulse measured via PPG) and GPS data to facilitate timely intervention. If predictions meet certain thresholds, the system triggers a clinical contact with an on-call clinician (intoxication) or informs emergency services (overdose). Training data are collected from clients wearing the equipment during their transition back to the community, with critical events tracked via self-report and medical records. The group successfully trains predictive models for detecting both intoxication and overdose events with high accuracy (See Roth et al 2021 for an approximate example of this approach) within a geographically limited municipality.

When tested in an independent sample, the consultants find higher rates of false positives for opioid intoxication and false negatives for opioid overdose in Black clients compared to White clients. The model is audited (Landers & Behrend, 2022) and 2 key issues are identified. First, in a visual examination of PPG activity prior to confirmed episodes of opioid overdose, it is determined that there are higher rates of PPG data loss among clients with darker skin tones (Colvonen, DeYoung, Bosompra, & Owens, 2020). In the intoxication model, the consultants find that model accuracy is primarily driven by features derived from GPS data. By visually inspecting GPS data against a city map, the researchers find that the model bases predictions partly on features that are correlated with time spent in parts of town that are economically depressed and have high rates of illegal drug sales. However, due to historical economic marginalization of minority groups in this community, Black clients reside in these parts of town at a higher rate than do their White counterparts. The former issue directly results in a higher false negative rate for overdose detection for racial minority groups, and the latter results in higher false positive rates for relapse prediction. This differential performance of the algorithm harms Black clients by providing them with less effective overdose protection and more unnecessary medical contacts with concerned providers.

This example illustrates how unfair predictive models might propagate preexisting societal injustices and deny specific groups equal benefit from the systems. Unequitable distribution of benefits from research violates the Justice Principles of the APA Ethics Code and the Belmont Report. Although unfairness should be avoided in all research applications of machine learning, characteristics of people who are involved in and benefit from research on substance use should be considered. People who use substances are involved in the criminal justice system at a relatively higher rate; machine learning models in criminal justice have been used to inform potentially life-altering decisions, such as sentencing decisions (Angwin, Larson, Mattu, & Kirchner, 2016). Other sources of unfairness that are currently present in institutions that frequently interact with people who use substances (e.g., treatment providers, law enforcement) could be carried forward in predictive models.

There has been concerted effort in the multidisciplinary machine learning field to develop computational tools for ferreting out unfairness in predictive models (Rajkomar, Hardt, Howell, Corrado, & Chin, 2018). Content-matter expertise (in the above example, prior knowledge of skin-tone related differences in PPG accuracy) can help researchers select fair feature (variable) sets and outcome labels, and post-processing algorithms can be used to recalibrate unfair predictive models to maximize accuracy across all defined groups (Barda et al., 2020). Landers and Behrend (2022) recently published an excellent review and programmatic guide to identifying and reducing unfairness in machine learning models. We recommend this resource as a guide for reducing unfairness in predictive models used in DRT research.

While the availability of these tools to improve algorithmic fairness is encouraging, it is important that they are consistently applied in DRT research so that the benefits of psychological research are distributed more equitably. Some institutions that have grappled with racial bias in machine learning and automated intelligence systems have attempted to protect against bias by using a “colorblind” approach wherein legally protected characteristics of users (e.g., race, sex, physical disability) are not explicitly used to train the model or make predictions (Benthall & Haynes, 2019). This approach has proven inadequate for mitigating unfairness in predictive models. Unfairness in machine learning and artificial intelligence systems is a pervasive problem that requires active solutions – advocacy groups are currently working to raise awareness and ensure thatappropriate safeguards against unfairness are included in these systems (https://www.ajl.org).

Continuous Real Time Data Acquisition and Investigator Liabilities

Principle A (Beneficence and Nonmaleficence) of the APA Ethics Code requires that psychologists minimize the amount of harm that befalls research participants with whom we work. One consequence of this obligation is that we are bound to intervene when we learn that a participant (or, in some cases, a person not directly involved in the research) is at imminent risk of harm regardless of whether than risk occurs as a direct result of the research being conducted. However, because interventions to protect participants often involve breaching participants’ confidentiality and limiting their autonomy, doing so can be ethically suspect. In some cases, the requirements that must be satisfied to justify breaking confidentiality and limiting autonomy are explicitly codified in relevant state laws. For example, all states have laws detailing legal requirements for involuntary hospitalization when participants report acute suicidality. Such laws are helpful as they provide specific guidance regarding the conditions under which psychologists are required to intervene and preventative actions that are allowed. In other cases, psychologists must weigh their ethical obligations against relevant federal and state law to determine the appropriate course of action.

In research involving people who use substances, there are medical, legal, and psychiatric complications that risk the well-being of participants and nonparticipants that occur at a higher rate than in the general population. Examples include: 1) medical emergencies such as drug overdose or life-threatening withdrawal symptoms; 2) psychiatric emergencies such as acute suicidality or psychotic symptoms; and 3) socio-legal emergencies such as driving under the influence or child neglect/abuse. Strategies for dealing with these eventualities have been developed by the research community over time. Generally, these strategies are based on clear guidelines that describe how crisis is identified (e.g., routine psychiatric screenings during research appointments) and handled, and optimally balance the need to protect participants’ and others’ well-being while maintaining confidentiality and participant autonomy to the greatest extent possible.

Research involving DRT adds additional layers of complication to researcher obligation that must be considered. For some devices, it is not practical to actively monitor for critical events given the volume and format of data that are produced; a researcher may find themselves in possession of data creating a legal obligation for intervention long before they become aware that intervention is warranted. In traditional research settings, participants provide information to researchers, either verbally or in writing. The transfer of data from participant to researcher clearly defines the point at which the investigator comes to be in possession of the information. The process of information transfer in DRT research is murkier and depends on the systems supporting data collection and storage. Some devices store data locally for days or weeks before it is transferred to storage at the end of the project. Other applications of DRT (e.g., intervention delivery, prompted remote ecological momentary assessment; Roberts & McKee, 2019) require that data be continuously uploaded to cloud storage in real time so that it can be automatically analyzed and contingent events can be triggered. These variations raise questions about balancing beneficence and feasibility in participant safety monitoring.

The use of DRT will likely impact the degree to which participants can maintain control over what aspects of their data are communicated to the researcher. In a traditional research setting, a participant may understandably choose not to disclose drug use – this can be achieved simply by denying any use when completing a questionnaire or clinical interview. Along with a well-conducted informed consent interview, the participants’ ability to exercise discretion in their reporting provides a layer of protection in which a participant may choose to exercise their autonomy to avoid legal repercussions that may be overly punitive and unjust. For example, in some states, any drug use during pregnancy constitutes criminal child abuse and may result in civil commitment or even termination of parental rights (Guttmacher Institute, n.d.). A participant who is pregnant may wish to avoid disclosing recent drug use to researchers, as they may determine that the social and legal costs resulting from this disclosure outweigh the benefit. In contrast, research involving DRT systems that detect substance use automatically and objectively makes it considerably more difficult for participants to decide what personal information is provided to researchers.

A final issue pertaining to investigator liability concerns the predictive models that often support DRT use. An increasingly popular application of DRT is using data to predict clinically relevant events. These systems operate by continuously feeding batches of data through an analytic pipeline to produce near real-time estimates of the probability that the wearer is currently or will soon experience the outcome that the system was designed to detect. In cases where the outcome being detected requires intervention to protect the wellbeing of the participant (e.g., suicidality, drug overdose), there are many unanswered questions. Does an alert generated from passive data fed through an imperfect predictive model justify breaching participant confidentiality or restricting autonomy? Does the level of certainty of the model’s prediction impact these decisions? There is currently limited guidance on these issues available from regulators and decisions may therefore depend on the nuances of local IRBs (Sieber & Tolich, 2013).

Distributive Justice and the Digital Divide

The digital divide refers to the unequal distribution of access to and knowledge of digital technologies such as those used to access online resources. People who do not have access to digital resources are far less likely to benefit from advances that occur in the digital health arena. Many of the benefits to individuals produced though DRT research require them to have access to online systems and willingness to use such systems. For example, a major thrust of DRT research is developing remote treatment methods for behavioral health conditions such as substance use disorder (SUD; Carreiro et al., 2020). People who do not have access to online connectivity – for reasons ranging from lack of resources to their own preferences – will not derive the benefits of this research to the same extent as those who engage with digital systems. Unequal access to benefit conflicts with the ethical notion of distributive justice and the Justice Principle of the APA Ethics Code, which states that, “fairness and justice entitle all persons to access to and benefit from the contributions of psychology and to equal quality in the processes, procedures, and services being conducted by psychologists.”

Prior research has found that lack of digital access correlates with other types of social disadvantage. People who are older-aged, or have lower incomes, rural residence, or less education are less likely to have internet access in the United States, although racial differences in internet access have declined in recent years (Pew Research Center, 2021). Nonetheless, because many of the groups who do not have online access are already disadvantaged, targeted efforts will be required to ensure that vulnerable populations can reap the benefits of DRT. Studies on usability and acceptability of remote devices are critical for predicting how widely an DRT-supported system might be adopted and identifying barriers for their use in vulnerable populations. Usability research should include diverse participants and actively recruit people from groups who are underrepresented in digital health research. Such studies with people who use substances have identified key barriers to engagement with digital devices, such as difficulty troubleshooting issues with the technology (Guarino, Acosta, Marsch, Xie, & Aponte-Melendez, 2016). Data collected in these types of study will inform how DRT research can be conducted in a way that minimizes barriers of use in vulnerable groups.

There is a frustrating paradox of the digital divide as it applies to DRT research and the resulting mobile health interventions. Many of the demographic groups who are less connected to digital systems (e.g., lower income, rural dwelling; Pew Research Center, 2021) also have limited access to traditional SUD treatment resources (SAMHSA, 2018). Digital systems that can deliver remote interventions have the potential to provide low-cost and evidence-based treatment options to these groups (Carreiro et al., 2020). Strategic design choices to make these systems more appropriate for groups with limited access to traditional, in-person treatment options have the potential to support equitable access to treatment resources.

Ethical Benefits of DRT

In the previous sections, we described the ethical risks that emerge from the use of DRT in research involving people who use substances. It is equally important to consider how DRT can reduce risk and increase potential benefits to participants or society. This is a critical issue for understanding the ethical implications of DRT; a utilitarian evaluation of the relative risk and benefit of a study is the yardstick by which its ethicality is evaluated (Sales & Folkman, 2000). We believe that strategic inclusion of DRT in research studies can improve validity and consequently increase its potential benefit, although systematic evaluation of DRT-based paradigms using established psychometric validation techniques will be necessary to confirm that these devices confer methodological benefit (e.g., Campbell & Fiske, 1959).

DRT also can provide a method for testing novel hypotheses by avoiding previously insurmountable ethical issues. An author of the current manuscript recently conducted a study that was among the first to systematically evaluate subjective cannabis effects in a group of adolescents who use cannabis (omitted for anonymous review). This study utilized smartphones containing ecological momentary assessment (EMA) software to conduct assessments of key pharmacological and emotional outcomes following cannabis self-administration in the natural environment. Historically, scientists have not been able to evaluate the subjective effects of cannabis in adolescents because of the risks associated with experimental administration of cannabis, which is the gold standard method for measuring subjective drug effects (Carter & Griffiths, 2009). By designing an alternative paradigm that leverages the remote assessment capabilities of DRT, we can study these types of behaviors without exposing participants to potential harm through risky laboratory protocols. These benefits of DRT extend well beyond examining pharmacological effects of drugs in adolescents; these devices provide a viable method for studying the immediate antecedents and consequences of other behaviors that are too risky to model in the laboratory. The types of behaviors that cannot be studied in the laboratory on ethical grounds (e.g., suicidal behaviors, youth substance use) are often those with profound implications for the physical and emotional well-being of participants.

Conclusions

Psychological research with people who use substances stands to benefit from DRT. These technologies have advanced to a point where they can be integrated into participants’ lives with minimal burden, providing researchers with rich digital descriptions of day-to-day functioning. Given their ease of use and methodological advantages, we expect rapid expansion of the utilization of DRT in research on substance use in the coming years, either as a primary assessment tool or to supplement more traditional methods. As more psychologists include DRT in their research programs, it is important that ethical issues surrounding these devices are addressed thoughtfully. The safeguards that we recommend in this paper will help protect the privacy and autonomy of participants and their communities. Some solutions can be implemented into individual protocols with relative ease. Strong data hygiene practices and willingness to incorporate privacy and technological experts into research teams will safeguard against many of these ethical threats. Systematic frameworks for ethical use of DRT in health research are currently available and easily adaptable for many types of behavioral research (Torous & Nebeker, 2017). Checklist style resources can be conveniently applied as a study is being developed to avoid ethical pitfalls (Shen et al., 2022). If these issues can be successfully addressed, however, DRT has potential to improve the quality and safety of substance use research and move the field forward by providing safe methods for research on this pervasive societal problem.

Public Significance Statement:

Digital and remote technologies are being used to measure naturalistic human behavior in psychological research on substance use, presenting novel ethical challenges that must be addressed to protect the interests of participants and maintain public trust in research organizations. This paper reviews sources of risk and offers solutions for dealing with the novel ethical dilemmas inherent to research use of these new technologies.