By linking health records with records of contact with other sectors and services that indicate probable disadvantage (e.g., housing or homelessness services, criminal justice system) we can identify a sampling frame of individuals at risk of social exclusion to apply to administrative health data. This offers an improvement over the status quo, in which socially excluded populations are often unidentifiable in administrative health data, precluding targeted investigation into their health needs. Cross-sectoral data linkage has been used to systematically identify populations experiencing incarceration [17, 22–24] and homelessness [18, 19, 25] in health administrative data. Similar applications of cross-sectoral data linkage would be helpful to identify other difficult-to-reach populations, such as migrants, asylum seekers, and sex workers, who remain underrepresented in health research [15, 26, 27].

Data linkage between multiple sectors or services can also allow for the systematic identification of people experiencing multiple forms of social exclusion, which is likely to have an additive or synergistic effect on morbidity, mortality, and quality of life [15]. For example, in Scotland, Tweed et al [10] linked national health service records with individual-level data from homelessness services, opioid treatment records, and a psychosis care register to examine the health of people affected by the intersection of homelessness, drug use, and serious mental illness. Two Danish studies [18, 19] linked national homelessness and substance use treatment registers, respectively, with psychiatric and vital statistics data to systematically identify these populations in administrative health datasets and examine their health and mortality outcomes.

When information regarding markers of disadvantage or social exclusion is routinely collected in health records, it is often of poor quality (i.e., inaccurately or inconsistently reported). This can lead to inaccurate conclusions regarding the burden of morbidity and mortality experienced by populations characterised by disadvantage or social exclusion; biased estimates of prevalence, incidence, and association; and (to the extent that they are evidence-informed) potentially misguided policy and practice responses.

Cross-sectoral data linkage can markedly improve ascertainment of markers of disadvantage and social exclusion status by permitting the triangulation of data from multiple data sources. One example is ascertainment of Indigenous identification in administrative data, which is often inconsistently reported [5, 11]. While Indigeneity does not in itself indicate social exclusion, many Indigenous people worldwide experience profound health and social inequity due to the ongoing impacts of colonisation and systemic racism [28]. A qualitative synthesis of Australian studies [5] concluded that ascertainment of Aboriginal and Torres Strait Islander status improved when using multiple datasets, with measurable implications for morbidity and mortality estimates. One study found that under-ascertainment of Aboriginal and Torres Strait Islander status led to over-estimation of improvements in life expectancy, and in the prevalence of stigma-associated conditions such as sexually transmitted and blood-borne infections [5]. Another recent Australian study [12] using national death data identified overestimations in life expectancy of 2.3 and 2.1 years for Aboriginal and Torres Strait Islander males and females, respectively, in the absence of additional data linkage to enhance identification of Aboriginal and Torres Strait Islander status. Triangulation of data from multiple sources could be particularly useful for identifying populations that are less likely to disclose information on disadvantage or social exclusion due to stigma, discrimination, or legal consequences [29].

Cross-sectoral data linkage can improve ascertainment of health outcomes experienced by socially excluded populations. Due to the barriers to inclusion described above, people experiencing social exclusion often have low rates of engagement with primary and preventive healthcare services and rely heavily on acute and emergency care [8, 13]. This means that their health needs are often systematically underrepresented in primary and preventive healthcare records. Health and social services that are outside the scope of mainstream healthcare (e.g., government assistance, child protection, migrant services, assertive community outreach, harm reduction, homelessness services) often act as a safety net for individuals experiencing barriers or reluctance towards accessing primary and preventive health services. Inclusion of these broader data sources –often characterised as ‘non-health’ data – can capture other service use and population characteristics, thereby broadening our understanding of the health and social needs of socially excluded populations.

The inclusion of ‘non-health’ data makes it possible to examine the association between social exclusion and health outcomes through a social-ecological framework that more broadly considers individual, social, and structural factors that affect health [30]. By integrating data from health and ‘non-health’ sectors, cross-sectoral data linkage can provide insights into the social and structural determinants of health. This is important because socially excluded populations often experience disadvantage across numerous social and structural determinants that span multiple governmental sectors including housing, education, employment, criminal justice, physical environment, and social support [8]. Such complex health needs necessitate a diversity of data to optimally inform coordinated, person-centred responses across multiple sectors [15]. Research, monitoring, and quality improvement will therefore have greater impact if they can address both the health and non-health contributors to health inequity [13].

For example, in a Canadian study of 13,318 individuals who had experienced opioid overdose between 2014 and 2016, surveillance data were linked with national data on immigration, employment, social assistance, and police contact [31]. Considering these broader determinants of health revealed that markers of vulnerability to social exclusion were common in the year prior to first overdose, with 66% experiencing unemployment, 50% receiving social assistance, and 39% having at least one contact with police. These findings provide useful evidence to inform targeted preventive interventions across multiple sectors. Vertical approaches that consider health as distinct from the broader social issues faced by socially excluded populations will be less effective because they overlook other important contributing factors that influence health outcomes [9].

Much of the existing data linkage research on social exclusion centers on populations for which data tend to be more readily available and reliable. Systematic data collection typically hinges on the existence of sectors or services dedicated to supporting a given population (e.g., criminal justice systems, housing services) or collection of information on markers of disadvantage or social exclusion at service contact (e.g., substance use disorder recorded using standard coding conventions such as the International Classification of Diseases (ICD) [33]). Both reviews in the Lancet Inclusion Health Series primarily identified research on people with substance dependence, with fewer studies on homeless populations or people with a history of incarceration [9, 13]. Migrants, asylum seekers, and sex workers are examples of difficult-to-reach populations with extremely limited data capture, which contributes to their exclusion from health research [26]. Almost no studies were found on these groups in either Lancet review.

One way to improve data availability for hard-to-reach populations is by supporting the sectors and services working with these populations to initiate or improve data collection and cross-sectoral linkage for research and quality improvement purposes. For example, low-barrier sexual health clinics often collect confidential data on sex worker status to inform clinical care and risk mitigation. These data have been used for epidemiological research [34] and have contributed to our understanding of the health burden associated with sex work. At a clinic or systems level, linkage of these de-identified data to other health and non-health datasets would provide critical insights into the causes and contributors of documented health equities [27, 34, 35]. Similarly, records of migrants and asylum seeking populations are sometimes collected by governments [26] or through community-based health and ‘social’ support services [36].

Coalitions formed between researchers and service providers, premised on a shared interest in improving health outcomes and service delivery, can help support ad-hoc or repeated linkages of these data to other health and non-health datasets. Although resource intensive, the novelty of this work and importance for health equity may be attractive to funding agencies or governments. Where sufficient data are not yet collected, such coalitions could also work towards the establishment and quality assurance of a minimum required data collection within sectors and services targeted at difficult-to-reach populations that would allow for linkage of these identifiers to additional administrative data.

Health service use is an imperfect indicator of health need, particularly for socially excluded populations that often have low rates of access to primary and preventive healthcare, and relatively high rates of acute and emergency healthcare contact. Administrative health data therefore tend to under-ascertain the upstream health needs of socially excluded populations, capturing only symptomatic reasons for presentation. Furthermore, stigmatised behaviours and associated health outcomes that are more common among socially excluded populations, such as self-harm, suicidality, and non-fatal overdose, often do not lead to health service contact. They are therefore often under-reported in administrative health data. Socially excluded populations may also withhold information on these health experiences due to fear of stigmatisation, differential treatment, or legal consequences [37–41], an adaptive behavior documented among people with substance dependence [42–44] and people living with HIV [42–46]. Consequently, under-ascertainment of health service use and health outcomes can lead to inaccurate conclusions regarding health burden, missed opportunities for early intervention, and poorly informed policy and practice change.

Linking data from multiple sources can enable more accurate ascertainment of the health needs of socially excluded populations (e.g., linking data from low-barrier assertive community outreach services with other types of health and non-health data [47]). Triangulation of data collected from multiple sources, including combining administrative records with patient self-report, can also enhance ascertainment of health outcomes that are not well captured in administrative health data. One study of adults recently released from prisons in Australia found moderate to good agreement between administrative health records and patient self-report across primary care (kappa = 0.69, positive predictive value (PPV) = 80.2%), emergency department (kappa = 0.41, PPV = 64.0%) and hospital services (kappa = 0.62, PPV = 76.0%), and for the use of prescribed medications (kappa = 0.62, PPV = 68.3%) [48]. These findings complement existing research examining the validity of combining administrative and self-reported data to verify experiences of self-harm [49] and non-fatal drug overdose [50] among people released from prison.

The siloed nature of service delivery provides additional challenges for cross-sectoral data linkage. The delivery of services by different levels of government or non-governmental sectors means that data collection is often decentralised. This makes it challenging to link data across sectors [3, 51]. Deterministic linkage using unique personal identifiers (e.g., personal health number) that are available across all datasets is simpler and more accurate. This is most common in centralised service delivery and data collection systems, such as in Scandinavia [52–54]. Decentralised data collection results in imperfect linkage because data must be linked based on the probability that two or more records match, using non-unique identifiers (e.g., name, sex, date of birth) [55]. Notably, some datasets will not have sufficient information to conduct probabilistic linkages, precluding data linkage altogether [55]. While probabilistic linkage methods are widely used and can produce high quality linkages [56, 57], the impact on linkage error and bias in study findings can vary across datasets and can disproportionately impact socially excluded populations who may be less likely to have identification and whose identifiers may be more prone to misspelling or inaccuracy [57, 58]. It is therefore important that analysts carefully consider the quality of data linkage on research outputs, and seek out opportunities for validation studies to inform future research [59].

It is largely for this reason that much of the research using cross-sectoral data linkage to examine the health consequences of social exclusion has taken place in high-income countries with more advanced, centralised data collection systems and more liberal approaches to privacy legislation that permit the appropriate use of these data for research, monitoring, and quality improvement. For example, Sweden, Denmark, and Norway maintain numerous nation-wide registries that can be linked for research purposes using unique personal identifiers [52–54, 60, 61]. This has contributed to a favourable environment for routine linkage of both existing datasets and primary (e.g., survey) data [18, 19]. In Australia, the Australian Institute of Health and Welfare maintains over 150 national databases that include housing, homelessness, criminal justice, disability, alcohol and other drugs, mental health, healthcare, education, and vital statistics [61], many of which are available for research purposes and can be linked using a National Linkage Map. In the UK, Administrative Data Research (ADR UK) facilitates access to public sector data from England, Northern Ireland, Scotland, Wales, and the Office for National Statistics for use in public interest projects led by accredited researchers. For example, the Welsh Secure Anonymised Information Linkage (SAIL) Databank (one component of ADR UK) includes over 80 datasets across health, justice, education, and child protection sectors [62]. In British Columbia (BC), Canada, Population Health BC provides researchers with linkage services and access to over 90 datasets across health, justice, education, housing, social care benefits, child welfare, employment, and immigration sectors [63].

Cross-sectoral data linkage is more complex in countries with marked decentralisation of service delivery, including countries with two-tier service delivery systems. In such settings, data are collected from numerous public and private health and social service systems that often lack consistent personal identifiers, resulting in piecemeal data collection and complex linkage processes that introduce avoidable linkage error [55]. In these environments, researchers should work closely with data custodians to understand populations and services represented within individual datasets, identify additional datasets that may help fill gaps in population or service coverage of interest, carefully consider impact on internal and external validity, and conduct validation studies to assess the impact of linkage methods and quality on research outputs.

The routine collection of electronic health and non-health data remains limited in many low- and middle-income countries (LMICs), and the use of these data for research, monitoring, and quality improvement purposes has been restricted due to concerns over their quality, completeness, and public availability [64]. A 2020 review of studies using routine health information in LMICs [64] provided evidence of increased uptake and investment in routine health information systems in recent decades. For example, the implementation of the District Health Information Management System 2 (DHIMS-2) in sub-Saharan Africa has allowed more efficient, integrated, and timely health data collection from local to national levels [65]. DHIMS-2 has permitted data integration from community-based health registers, population surveys, and meteorological departments for health system performance measurement and research [64–66]. It is conceivable that such systems could be used to integrate data from other non-health sectors as they evolve in LMICs. However, of a mere 132 studies identified in the LMIC review [64], none focussed on socially excluded populations as defined by the Lancet Inclusion Health Series [9]. These findings may reflect both the infancy of routine data collection in LMICs for inclusion health research and differing research priorities in LMICs, which were overwhelmingly focussed on malaria and maternal health. Nonetheless, over time, these routine health information systems may provide the foundation for future cross-sectoral data linkage involving socially excluded populations in these settings.

Despite exceptionally high standards of ethical clearance required for health research on vulnerable populations [67], many data custodians remain reluctant to share data for research, monitoring, and quality improvement purposes. This is particularly common in sectors not typically involved in data linkage research [32]. Data custodians are responsible for maintaining the privacy and confidentiality of data in their possession. Thus, sharing these data require a high degree of trust that privacy and confidentiality will be maintained – particularly when they are focussed on vulnerable and socially excluded populations. Concerns around data security result in a heightened sense of data protection, particularly for those new to data linkage. The emergence of data safe haven (DSH) environments that monitor secure storage and use, establish robust data governance protocols, and prohibit the export of individual-level data are premised on the principles of data safety and trust [68]. Additional advancements including the Five Safes Framework for planning, assessing, and managing risks associated with data sharing and release [69]; secure transfer of encrypted data across data custodians and linkage teams [70]; linkage methods that preclude the need to share full identifiers across sectors [71, 72]; and application of the separation principle [73] – such that no one can view both personal identifying information (used for data linkage) and content data (e.g., clinical or service use information) at any point during linkage and analysis – should in theory address data custodian concerns around data security. Importantly, these advancements can build public trust in the management and protection of sensitive administrative data, which can in turn support data custodian decisions to participate in data linkage research [74–76]. However, it is currently unknown to what degree these advancements have impacted the willingness of data custodians to share data for health research purposes. This is an important area for future research.

Socially excluded populations are often over-represented in particularly sensitive data, such as notifiable disease registries (e.g., HIV, HCV) and other heavily anonymised health data (e.g., abortion records, police contacts). The highly sensitive nature of these data can result in an even greater sense of reluctance to share these data for health record linkage, additional administrative requirements that can reduce research capacity, and more conservative approaches to interpreting legislation that protects (but rarely precludes) their use [77]. This can inhibit research on issues disproportionately impacting socially excluded populations, despite the potential benefits for health equity. Nonetheless, several jurisdictions maintain population-level disease registries (for example, on HIV in Australia and Scotland) that can be linked to additional administrative datasets by data custodians or researchers [78, 79]. The advancements listed above may be of particular value when planning and undertaking data linkage using sensitive and anonymised data. However, linkage to sensitive data remains challenging. Researchers must not undervalue the importance of establishing strong partnerships with data custodians to understand data sharing concerns and legislative requirements, and to jointly identify linkage and analytic arrangements that address them [80, 81]. One option is storing linked data with the data custodian and arranging secure data access for external analysts.

A perennial challenge for researchers interested in conducting cross-sectoral data linkage is the reluctance of many data custodians to participate in rigorous, independent research or quality improvement, and to have research findings made publicly available, whether or not they seem favourable to the custodian [82]. Furthermore, some ‘non-health’ sectors may form the view that health is not part of their remit, impeding cooperation for health data linkage because it is perceived to be irrelevant. This is often manifest in legislative and/or administrative barriers that explicitly or implicitly limit data access, linkage, or public reporting. Data custodians may also actively try to block data access because they do not want independent scrutiny, which in turn can compound health inequities among socially excluded populations. Overcoming this kind of “epistemophobia” requires the development of meaningful partnerships between researchers and data custodians that prioritise research benefit for the population(s) under study, communicate the value of cross-sectoral data linkage from a policy and practice perspective, and shift focus away from concerns over scrutiny and towards opportunities for evidence-informed service and health improvement [32, 82]. Decisions about data access may also be better placed in the hands of centralised data governance bodies that are pre-authorised to release data for agreed purposes based on public benefit, an established and transparent decision-making framework, and high ethical and data security standards. An example of this is the Confidential Advisory Group of the Health Research Authority in the UK [83].

Finally, engaging decision makers to understand their perspectives on the value of cross-sectoral data linkage for informing policy and practice responses within inclusion health [84], and sharing this information with data custodians, can help close the loop on data sharing hesitancy by focussing on real-world, cross-sectoral benefits.

Cross-sectoral data linkage is complex in theory and practice. It involves sharing and linking data from multiple sectors and jurisdictions, all of which may have their own unique requirements and protocols. Cross-sectoral data linkage is therefore prone to extensive requirements for ethical, data custodian, and legislative approval, contributing to its time-consuming and resource-intensive nature [3]. These factors alone substantially reduce capacity for cross-sectoral data linkage and inhibit evidence-informed action on emergent health concerns that disproportionately impact socially excluded populations (e.g., COVID-19 [85]). Centralised bodies dedicated to the coordination of cross-sectoral data linkage can help researchers navigate complex linkage processes, communicate research requirements across multiple data custodians and jurisdictions, and streamline duplicative administrative protocols. Importantly, a key purpose for such bodies must be to reduce often prohibitive administrative burden, rather than introducing an additional layer of bureaucracy.

For example, the Australian Population Health Research Network (PHRN) was established in 2009 as a national coordinating network to support researchers to access and link more than 300 available health and health-related data collections [86]. Australia has a federated system of government comprising six states, two territories, and a Commonwealth government, each with its own data collections and governance. Thus, data linkage projects often rely on several jurisdictional data centres and require multiple approvals and data extractions depending on the number of data collections and jurisdictions involved [86]. A benefit of the PHRN is its national coordinating function: it can coordinate meetings between researchers and all relevant linkage centres and data custodians to jointly establish feasibility of research proposals, identify data access requirements, prepare linkage applications, and organise approval processes. The PHRN also provides publicly available education and training to researchers, data custodians, and regional data linkage units. However, cross-sectoral data linkage in Australia remains challenging and underutilised, due in part to its resource intensity [3, 87]. Despite substantial investment of AUD$55 million in this national data linkage infrastructure between 2009 and 2019, and a 3.5-fold increase in the number of peer-reviewed publications involving ‘health-only’ data linkage during that period, only a small increase in the number of studies involving cross-sectoral data linkage has been observed [6]. Entities such as the PHRN, that have strong links to data custodians and linkage administrators, are well placed to advocate for streamlined administrative processes that improve efficiency of linkage processes. Simultaneously, advances such as the Australian National Mutual Acceptance (NMA) Scheme [88] that aim to streamline scientific and ethical review across data custodians and jurisdictions can help to reduce duplicative processes whist maintaining integrity in ethical and data security protocols [32]. Similar work is being done in the UK via the Health Data Research (HDR UK) institute [89].

Insufficient workforce capacity, skill, and technical expertise contribute to the limited use of cross-sectoral data linkage [3, 32]. Cross-sectoral data linkage is complicated and requires a multi-disciplinary data linkage workforce that 1) is well-rounded in the fundamentals of health research including (at a minimum) basic statistics and epidemiology, 2) is familiar with the intricacies of administrative data collection and linkage across sectors, 3) has commensurate understanding of the social and structural determinants of health, and 4) is prepared to navigate the complex interplay of political sensitivities, epistemophobia, and siloing of within-sector mandates that can impede cross-sectoral data linkage [3, 7]. Furthermore, the complexity of linked administrative data demands skilled analysts that are proficient in both the theoretical and hands-on applications of social science, public health, health data science, biostatistics, and epidemiology. In recent years, there has been an increase in the number of university programs in these domains; in particular, programs relating to health data science. However, it remains unclear whether the supply of graduates with broad skills in these combined fields has kept pace with increased demand for data linkage and associated research. Workforce shortages in both linkage and analytic domains need to be addressed by continuing to expand workforce training opportunities and hands-on experience with cross-sectoral data in postgraduate education, increasing investment in data linkage infrastructure to meet demand for these services, and forming interdisciplinary teams to achieve the required diversity in knowledge and skill [90]. For example, in Ontario, Canada, an innovative partnership was established between ICES (a health service research institute), computer scientists, and machine learning specialists to apply more complex analytic methods to the analysis of province-wide administrative data [91].

Cross-sectoral data linkage provides a unique opportunity for the inclusion of socially excluded populations as experts in their own experiences [92]. Yet, socially excluded populations are only sporadically engaged in cross-sectoral data linkage and associated research. Cross-sectoral data linkage, including the pursuit of improved data collection within sectors and services that predominantly serve socially excluded groups, is currently an underused opportunity to meaningfully engage socially excluded populations and reinforce their right to self-determination in health research. It is imperative that populations who are observed through cross-sectoral data linkage are meaningfully involved as early as possible, from advising on what data to collect and how to collect it from socially excluded groups at the point of service access, to planning research, identifying research questions and the data sources needed to answer them, data analysis, interpretation, knowledge dissemination, and decision making about how to respond to knowledge generated [93–96]. This is essential to ensure that information generated from linked data are contextualised by lived experience, and that socially excluded populations are not further excluded from evidence generation and decision making regarding their health and wellbeing. Engagement of socially excluded populations to inform and participate in data linkage research has been demonstrated [25] but is most often focused on engagement of First Nations communities [97]. In Scotland, the Public Benefit and Privacy Panel for Health and Social Care (HSC-PBPP) is a patient advocacy panel that scrutinises applications requesting access to NHS data with a view to ensuring that linkage projects lead to tangible benefit for populations of interest and that risks are adequately addressed [98]. HDR UK involves patients and the public through a Public Advisory Board, workshops, opinion surveys, and public events to engage and involve underserved communities [89]. Similar applications may benefit inclusion health research and should not preclude the early engagement of populations of interest during development of research proposals. Ongoing evaluation and feedback is essential to ensure that such bodies are capable of understanding the basics of data linkage research, efficient, pragmatic, and do not lead to undue administrative barriers [99]. Better engagement with people experiencing social exclusion can also inform efforts to improve ascertainment of social exclusion status (e.g., by matching data collection with identities held by populations [100]); address concerns regarding (mis)representation in administrative data; educate patient populations on what data are collected and how it will be used to inform policy and practice; build inclusive coalitions between socially excluded populations, data custodians, and researchers; and develop a shared understanding of research benefit [98] – both within existing data collection systems and when developing new ones.