Using publicly available UK datasets to identify recruitment sites to maximise inclusion of under-served groups: three case studies

Background

In response to the substantial body of evidence that those recruited into research are not always representative of the population for whom the research is most relevant^1–4, there is increasing recognition of the imperative to include under-served communities. In this context under-served communities have been described as groups that are ‘less well represented in research than would be desirable from population prevalence and healthcare burden’ (https://www.nihr.ac.uk/about-us/our-key-priorities/under-served-communities.htm). More inclusive research, addressing ethnicity, diversity and inclusivity (EDI), supports generalisability of results, and adoption into practice⁵.

Whether a group is under-served is context and study specific. Groupings for consideration include by: demographic factors (e.g. age, sex, ethnicity, education); social and economic factors (e.g. employment, socio-economic status, geographic location, language, digitally excluded); health status (e.g. mental health condition, cognitive impairment, physical disabilities); and disease specific factors (e.g. rare diseases)⁵.

The UK National Institute for Health and Care Research (NIHR) and other health and social care research funders want to ensure that the research they fund meets the needs of the whole population for which it is relevant⁶. The NIHR-INCLUDE project roadmap shows potential points in a trial’s lifecycle where considerations and decisions about inclusion of under-served populations may most usefully be made⁵. These include development of the study design by researchers through to research funding decisions by panels, all being made with consideration to ensuring inclusion of a study population that reflects the target population. Details of how under-served groups have been identified and showing how trial recruitment will target where the relevant groups are situated, needs to be reported in funding applications if funding panels are to make decisions about whether inclusivity has been appropriately considered.

There are studies that have identified context and condition specific barriers to inclusion of under-served groups in trials and proposed solutions^7–9. Current practical guidance to address these barriers focuses on designing and modifying participant recruitment and retention approaches once a trial is up and running. Available guidance includes consideration of eligibility criteria and the recruitment pathway, participant facing materials, cultural competency and building community partnerships¹⁰. An area that has not been addressed is recruitment site selection. Identifying sites that include the relevant under-served population should enhance the chances of their inclusion in a study.

In this paper, we use three NIHR funded trials as case studies for demonstrating how publicly available UK datasets can be used to facilitate the identification of under-served communities for inclusion in trials. Two of the three grant applications were commended by the funding panels for their approach to inclusivity.

Overview of case studies

The case studies we describe here are of three randomised controlled trials (RCTs), where we piloted use of UK national, freely available datasets to identify recruitment sites with consideration of health status, socio-economic and ethnicity characteristics to support inclusivity of under-served populations. In planning a trial we consider equality, diversity and inclusion across several aspects of the trial design and delivery including the eligibility criteria, the primary outcome measure, public involvement plans, site and participant recruitment strategy. Our focus here is site selection.

The example trials have different methodological approaches, address different needs and demonstrate recruitment planning across both health and social care settings. An overview of the key aspects of the studies is presented in Table 1. All three trials have internal pilots and an economic evaluation; TIDE (https://www.york.ac.uk/healthsciences/research/trials/ytutrialsandstudies/trials/tide/) and INTERACT (https://www.york.ac.uk/healthsciences/research/trials/ytutrialsandstudies/trials/interact/) also include qualitative process evaluations. The trials were all funded by the NIHR Health Technology Assessment (HTA) programme. For each of the trials we used the site selection approach with a different level of intensity: i) as the main method of selecting the recruitment sites for the trial (INTERACT), ii) to shortlist/prioritise sites from an existing group of sites we had previously worked with on a related study (TIDE) and iii) to check the suitability of a pre-identified group of recruitment sites (DIDACT) (https://www.york.ac.uk/healthsciences/research/trials/ytutrialsandstudies/trials/didact/).

Pre-identified potential recruitment sites

Standard, pragmatically focussed approaches to identifying sites include using the networks of clinical leaders in their field (as with all three of our case studies), an existing relationship with specific NHS Trusts through a previous study, as we utilised with TIDE¹¹, or a survey of potential Principal Investigators to gauge level of interest. The latter approach identified a high level of interest from educational settings and speech and language therapists in the INTERACT trial. Likewise, a national survey of British Elbow and Shoulder Society (BESS) surgeons, secured expressions of interest from 23 sites to be involved in the DIDACT trial. However, we were aware that we had identified the sites based on level of interest from the clinical teams and before prioritising these sites for recruitment we needed to have a fuller understanding of whether recruiting from these sites would facilitate recruiting participants reflecting the range of people who have a distal clavicle fracture. The same principle was applied to the pre-existing list of NHS Trusts for TIDE, and the interest from educational settings for INTERACT.

Identifying risk factors

The next step therefore was to collate information on the epidemiology of the condition to fully understand the target population. For DIDACT, the information available was very sparse. Clinical experience is that this fracture can be experienced by anyone but it is a more common injury amongst younger physically active people, particularly men of working age¹². For TIDE, our literature searches found a systematic review which identified risk factors associated with increased risk of the clinical problem being addressed by the trial – specifically risk factors of MRSA colonisation¹³. Of particular interest to our study were antibiotic use, previous MRSA colonisation, and care and nursing home residents. For INTERACT we found that over 10,000 children and young people in England have PMLD. All have multiple disabilities, the most significant being profound intellectual disability and great difficulty communicating. Higher rates of PMLD have been found among Traveller children of Irish heritage, Pakistani, Bangladeshi, ‘other’ Asian and ‘other’ black heritage and ‘any other’ ethnic group¹⁴ (https://www.gov.uk/government/publications/people-with-learning-disabilities-in-england/chapter-1-education-and-childrens-social-care-updates).

Selecting datasets and cross referencing

Having established key identifying features for potential sites and participants, we then applied these criteria to relevant accessible datasets. In Table 2 we set out a simple overview of the drivers for site selection identified in the epidemiological preparatory work, and the sources subsequently used to find potential study sites to support inclusivity. Box 1 contains details of the sources used. Depending on the identified trial specific features and the EDI factors to account for, where necessary an order of priority for searching and cross referencing was used. In addition, pragmatic accommodations to facilitate cross referencing between datasets was necessary. For example, the names of NHS Trusts had to be linked to their Clinical Commissioning Group (CCG) before the lists could be applied to the index of multiple deprivation dataset. All the searches used the most up to date information available in each database and for each field.

For this step, the process used for each of our case studies is described below. See Box 1 for the sources used in each case.

TIDE

For TIDE, the search started in Fingertips public health data for any relevant or related data fields as identified in the systematic review, also for additional recognised markers of MRSA colonisation used by this national dataset. For some items, proxies or surrogate markers had to be used due to the limitations of publicly available national datasets. For example, we searched on bacteraemia rates as a proxy for previous MRSA colonisation. This data source gave a value for each of four clinical indicators, for example 147.3 prescribed antibiotics per 1000 registered patients per quarter by CCG. See Table 3 for all markers used.

Each NHS Trust was matched to their corresponding CCG and the value for each indicator mapped to each NHS Trust. For Trusts serving multiple CCGs the mean value for each indicator was calculated. For example, if Trust A is covered by CCGs B and C which had antibiotic prescribing rates of 80 and 100 per 1000 patients, then a rate of 90 was used as the antibiotic prescribing rate for Trust A. It was assumed each CCG carried an equal weight within each Trust. NHS Trusts were then ranked from highest to lowest for each clinical indicator i.e. most antibiotics prescribed, to least. The ranking of the four clinical indicators within each Trust were then combined to give a combined risk of the clinical problem (i.e. combined MRSA colonisation risk) for each NHS Trust.

In a similar manner to the clinical indicators, indicators for underserved populations were searched for in publicly available national datasets. In this instance three indices of deprivation from the ONS and one measure of ethnicity from census data were identified and used. For the indices of deprivation used, each CCG was ranked nationally, most deprived to least. For ethnicity a combined non-white proportion was calculated for each CCG. CCGs were then ranked from highest proportion of non-white population (most diverse) to lowest. The non-white groupings were, Mixed/multiple ethnic, Asian/Asian British, Black/African/Caribbean/Black British and Other ethnicity groupings. Ranking of CCGs by each of the different categories would have been possible had we been seeking to include a specific ethnic group.

These data were not available by NHS Trust but instead by CCG, so the Trust mapping to CCGs was used again. The rankings for each of the four indicators for underserved populations were then merged to give a combined ranking for underserved populations for each Trust. In addition, clinical and underserved population rankings were also combined to give a score for clinical risk and degree of underserved population for each NHS Trust.

INTERACT

With INTERACT, we found routinely available national data for pupils with Special Educational Needs in England, which includes breakdowns by type of SEN provision, and then we identified those schools including children/young people with PMLD. This school level data was combined with other national datasets that included data on ethnicity, index of multiple deprivation, age and gender.

Datasets were combined using the local authority code as that identifier was common across all datasets. Most datasets were from 2021/22 but the index of multiple deprivation dataset was only available from an earlier time point (2019), and the data could only be linked at the local authority level. Schools were ranked based on the combination of high PMLD numbers, low deprivation, and high percentage of ethnic minority groups.

After ranking the schools, we explored the age spread by school. Only a small percentage of state special educational settings include people over 19 and it was felt important to include the full range of the school population where it existed. Hence, we checked whether the schools we ranked to target for recruitment included over 19-year-olds.

DIDACT

For DIDACT we collated the most up to date regional level data from multiple national sources to identify areas with high proportions of young males, alongside the index of multiple deprivation, index of health deprivation and ethnicity. Although there was no specific epidemiological data of the injury being more prevalent in particular socio-economic or ethnic groups, we linked and combined the data on proportion of males with data on ethnicity and deprivation to address the broader concern of ensuring the trial recruited from geographic populations with high disease burden which have been historically underserved by research activity.

Initially we matched each site to CCGs that provided us with an identifier to merge with the health deprivation, age, gender and ethnicity data. We calculated the percentage of the population that were male in the 16–25 age group and combined non-white backgrounds and for each we ranked them across all CCGs. The percentage of interested sites that fell into the top 5%, 10% and 20% nationally for clavicle fractures (young males 16–25) and collectively for the most vulnerable and underserved populations (ethnicity; health deprivation) were calculated. For the combined variables we assumed that variables were equally weighted.

Datasets were combined from different time periods and not all CCGs were available for all variables of interest. These were matched as closely as possible and changes in CCGs were matched as closely as possible using the linked variables within the datasets where available.

Search results and pre-identified potential sites

For DIDACT, examination of the data showed that the 23 sites already identified through a survey of surgeons included one in three of the top 5% of regions with the most vulnerable and underserved populations, and just under half of the top 10% of regions at the highest risk from a clavicle fracture. We thought that this was a reasonable selection of sites, also taking into consideration enthusiasm from the clinical teams.

With TIDE the scores were mapped against a list of NHS Trusts who we had successfully worked with previously to identify the 10 Trusts we planned to target first for recruitment in the pilot phase.

For INTERACT, given this is an under-researched area with no established networks of schools taking part in research, we did not have access to high recruiting sites and thus we used this list of schools as our sampling frame of schools to contact for participation. We created a list of approximately 400 schools serving the target population.

For all three studies, we produced lists of potential recruitment sites in excess of the number anticipated as necessary to meet the recruitment targets.