In the United Kingdom, children looked after are defined as children under the age of 18 years who are looked after by the local authority and are in its care or provided with accommodation for a continuous period of more than 24 hours [1]. Children can become looked after under voluntary arrangements or under a compulsory care order granted by the courts. Throughout the last decade, there has been an upward trend in the number of children looked after in the UK, with the exception of Scotland where rates of children in care have been falling [2, 3]. These children are considered one of the most vulnerable groups in society, with previous work showing increased risk of poor outcomes across social, education and health domains [4–6].

In recent years, there has been growing interest in the use of administrative data to advance our understanding and insights into children in receipt of care and support, and children looked after by the state, as well as the implementation and effectiveness of social policies [7–9]. Current Government policy strongly supports the need to reduce the number of children in need of care and improve their outcomes, including the recent policy paper Data Saves Lives which promotes how it can improve children’s social care services [10, 11]. Administrative data relates to individual or organisational level information, which has been derived from data routinely collected by public sector agencies for organisational purposes including, but not limited to, education, healthcare and social care agencies [12]. It contains abundant, detailed information that has significant potential for research.

As part of the UK’s statutory guidance, local authorities are required to submit information about families and children who interact with children’s statutory social services to the Government on an annual basis; this enables service provisions and outcomes to be monitored [13]. These returns, known as the Children Looked After (CLA) Census, contain individual-level data for all children and young people placed under the care of a local authority and are available for research purposes within Trusted Research Environments (a centralised, secure data platform) across all four UK nations.¹ Information about the variables within the datasets have been described elsewhere [14–18].

Within these TREs, person-level data can be linked between datasets using a unique identifier, allowing for a more comprehensive understanding about the needs of a population [19]. Within health datasets, the linkage variable is often based on unique identifiers such as the National Health Service (NHS) number (England and Wales), the Community Health Index (CHI) (Scotland) or the Health and Care number (HCN) (Northern Ireland). However, for the children’s social care datasets (England, Scotland and Wales²), including the CLA dataset, the NHS/CHI number is not available. Instead, quasi-identifiers are used – these are attributes that are not themselves unique identifiers but in combination can be used to identify an individual (e.g. name, gender, date of birth, address and postcode) [20].

Demographic data collected within the CLA Census is limited; alongside the Unique Pupil Number (UPN) (England and Wales) or Scottish Candidate Number (SCN) (Scotland), only the date of birth, gender and ethnicity are collected. Other identifiable information such as the child’s name, address and postcode are stored separately in the education dataset³. The UPN/SCN is used to link the two datasets to access the child’s name and address. Children are automatically allocated a UPN/SCN on their entry to a maintained school. The majority of children in the UK will not attend school until after they turn four years old, although the compulsory school age in England and Wales is five years old. In the absence of the UPN/SCN, very few children under the age of four have sufficient quasi-identifiable information for probabilistic matching, meaning they do not have a unique identifier which allows linkage to other datasets. In some cases, children aged two or three may have an identifier if they attend a maintained nursery [21]. This is particularly problematic for researchers who wish to link the CLA dataset because it is well established that most children looked after are less than one years old [22].

Within the CLA dataset held in the Secure Anonymised Information Linkage (SAIL) Databank, only 37% of the CLA population have a unique identifier, the Anonymised Linkage Field (ALF), assigned via the routine algorithm (see Methods and Supplementary Appendix 1), drastically reducing the cohort sample that can be linked to other datasets [19, 23]. Previous researchers using the CLA dataset have also reported inconsistencies in the data, including discrepancies with the week of births, duplicate ALFs and year-on-year changes in quasi-identifiers amongst individuals in North Wales⁴. Lastly, the use of direct and indirect identifiers is particularly problematic for children looked after. These children are subject to recurring changes in quasi-identifiers e.g. frequent address changes which will contribute to the lack of a postcode.

To overcome the issue of limited routine ALF availability, we have developed a six-step CLA matching algorithm for the SAIL CLA dataset, with the main aim being to improve research quality by producing a standardised data table with additional matched ALFs. This tool utilises de-identified quasi-identifiers from several datasets already available to obtain the routine pre-assigned ALF held in SAIL. The improved ALF table has been made available as an additional resource to all researchers working with the CLA data to ensure projects can benefit from a benchmarked and reliable solution. Whilst this algorithm was developed specifically for the data held within SAIL Databank, it can be adapted to be applied to other anonymised population databases that use similar techniques to those in SAIL.

CLA data is currently available from 2002/03 to 2020/21 and can be accessed by researchers through the SAIL Databank. During this period, we identified 38,260 individuals with a record in the CLA dataset, with 14,024 (37%) having been allocated an ALF via the standard SAIL algorithm (Supplementary Appendix 2).

Following the implementation of our algorithm, the overall match rate increased by 25%, with a total of 23,409 (61%) individuals having an ALF (via our algorithm and the routine SAIL algorithm). Figure 1 shows the number of ‘new’ ALFs assigned at each step of our developed algorithm. A total of 9,842 individuals were assigned an ALF via the algorithm. Note, an inidividual could be assigned an ALF for one or more care episodes where their system ID had changed if they had moved local authorities. This is why the unique number of ALFs assigned via the alogorithm, as per Supplementary Appendix 2 (n = 9,385), is lower.

The algorithm was also able to correct data quality issues. We also report the cohort demographics for the entire CLA cohort at the time of first entry into CLA before (routine ALF) and after (boosted ALF) applying the algorithm (Supplementary Appendix 2).

We were able to benchmark our algorithm by comparing routine ALFs assigned via the SAIL process to those assigned by our algorithm. Our algorithm demonstrated a high sensitivity rate of 90% when compared to the ALFs already assigned in SAIL (13,253/14,725)⁵. Only a small proportion were ‘incorrect’ or mis-matched (1.2%, 170/14,725). The algorithm was unable to assign the remaining 8.8% of children an ALF (1302/14,725). We were unable to calculate specificity because we do not have access to the de-anonymised data, and therefore we do not have a set of children that we can confidently say who should not be linked/recorded within the CLA dataset.

Individuals who were born before 1994 had the fewest pre- and post-match ALFs allocated, likely due to fewer datasets being available (e.g. CINW in 2009/10, EDUW in 2002/2003 (academic year)) and poorer data coverage. For example, although the WLGP data pre-dates 2000, the transition to electronic records was not widespread across practices until the late 1990s [32], and records are known to be less complete before 1 January 2000 [33].

As expected, there were also more children under the age of four who did not have an ALF when they entered care for the first time, having the most ‘unlinked’ ALFs (Figure 2). As previously alluded to, this is likely due to the UPN being assigned at entry into school but could also be explained by the lack of placement stability (poor recording of postcode) and continuity in the healthcare system that children looked after experience. The recognition of children looked after as a highly mobile population, together with limited experience of continued general practice care may contribute to the reduction in attributable ALFs [34]. However, the former explanation appears to be most probable, supported by comparable improvements in ALF percentages amongst those less than and greater than four years of age. There was also a high proportion of children aged 15+ years who remained ‘unlinked’. One plausible explanation could be because these children have ‘aged out’ of care and/or left education, and therefore, they have no further contact with these sectors.

There appeared to be no difference in ALF improvement by sex; males and females both had around a 25% increase. Again, deprivation did not appear to impact ALF assignment, although there were fewer ALFs matched for individuals who were missing a WIMD score (Figure 3). This is to be expected and reflects the lack of a recorded postcode, an identifier used in the ALF linkage process. In line with previous studies, variation in data completeness was more prominent for ethnicity [35, 36]. Specifically, ethnic minority groups, with the exception of ‘Mixed’ ethnic group, were less likely to be successfully matched (Figure 4) [37, 38]. It has been speculated that this reflects inconsistencies in the quality and distinctiveness of registered names [39, 40]. Since those with asylum seeker status (a person who has left their own country and is seeking protection in another country) are likely to be born outside of the UK, this may also impact linkage success (e.g. lack of NHS number, postcode history).

Our CLA ALF improvement algorithm increased ALF matching from 37% to 61%, allowing an additional 9,385 (25%) individuals to be assigned an ALF. The algorithm also amended incorrect weeks of birth and gender, and incorrect/duplicate identifiers. Benchmarking against the SAIL ALF assignment algorithm demonstrated high levels of sensitivity (90%) for our CLA ALF algorithm.