CSC services in England are run by 153 local authorities, exercising various powers and duties under the Children Act 1989. Children can be referred to CSC by a number of agencies including self-referral by the family, though the most common referral sources are police, schools and healthcare [4]. Different local authorities have different service configurations but, broadly speaking, among children with officially recorded referrals, most receive a social work assessment and, of children who receive an assessment, most are considered “in need” [4, 8]. A subset of children is referred due to safeguarding concerns. The result of a safeguarding investigation may be that the child is made subject to a CPP, which should specify what should be done to keep the child safe, or the local authority may instigate care proceedings to receive the child into state care [1]. Some children may be “referred” to CSC but are seen through early help or other local authority services, and are not recorded in the CiN census [8, 10]. Additionally, where a child is referred but not assessed (or assessed and found not to be in need), they may be eligible for other services, though data on such services are also not available.

Our deidentified copy of the CiN census [8] was a complete extract from 1 October 2008 to 31 March 2018 covering data from all CSC departments in England. Data were available on referrals, whether children were assessed by a social worker, whether the child was recognised as a CiN and whether the child was subject to a CPP. Children who are legally recognised as CiN because they are looked after in state care are included in the CiN census, though information identifying them as such is not available. Data are collected annually and can be analysed longitudinally using identifiers described below. We cleaned the data according to Supplementary File 1.

Because the CiN census began in October 2008, there was not a single cohort that could be followed from birth to age 18. Additionally, analyses of data quality (Supplementary File 2) revealed an undercount of events in the years before financial year 2011/12, especially for under-5s. Consequently, reliable data for birth cohort approaches are only available from 2011/12. We therefore adopted an approach akin to an accelerated longitudinal design [11] by combining cumulative incidence estimates in three separate open cohorts (Figure 1). Children did not have to be born in England (birth place is not collected in the CiN census). These cohorts were: 1) children born in 2012/13 (all financial years, April to March), followed to age 5 (i.e., to the day before their 6^th birthday); 2) children born in 2005/06, followed from age 6 age to 12; and 3) children born in 2000/01, followed from age 13 to 17.

We examined the cumulative incidence of (all to age 18): (a) being referred to CSC; (b) being assessed by a social worker; (c) being recorded as a CiN; and (d) being as subject to a CPP. Whether a referral of a child resulted in an assessment was determined by reference to a variable in the CiN census that indicated further action (i.e., an assessment) was required following referral. CiN status for a child was determined with reference to variables indicating that further action from the referral was required and that the reason for closure was not code RC8 (assessed but found not to be in need). In other words, the child must have been referred to CSC, assessed by a social worker and found to be in need. Being subject to a CPP was determined according to a non-missing record of a CPP start date.

Details for how the cumulative incidences were estimated can be found in Supplementary File 3. In summary, we counted the observed number of first events in each cohort as numerators and we used Office for National Statistics mid-year population estimates as denominators. For cohorts 2 and 3, we adjusted the numerators to account for left censoring, that is, potential recurrent events from the unobserved periods. These were based on a combination of estimated recurrence rates and observed events as shown in Supplementary File 4. Estimates from the three cohorts were then combined.

To avoid double counting children with events at different ages or in different parts of England, we linked children’s longitudinal records using one of two identifiers within the dataset. Children’s records can be linked using an identifier assigned by the local authority (encrypted before being shared with researchers), the LA-child ID, or, alternatively, the anonymised Pupil Matching Reference (PMR). Neither is perfect. The LA-child ID is unique only to the local authority that assigned it and is not transferred if a child moves between authorities. The PMR, by contrast, is usually assigned when a child starts school or state-funded nursery (pre-school). It is nationally unique and persists when a child moves but only 70% of children in the dataset had a PMR (higher for school-aged children and lower for children under 5 and over 15: Supplementary File 2).

In our main analyses, where events were of children aged 5+, we used data only from children who had a PMR. Where events were under 5 years of age, we used the PMR if it was available, and the LA-child ID where it was not. Because about 10% of records of school-aged children had no PMR recorded (Supplementary File 2), this method excludes a proportion of school-aged events from the numerator. We therefore carried out sensitivity analyses whereby we re-estimated the cumulative incidences using the PMR where it was available and the LA-child ID where it was not, regardless of age. This ensures that records are not excluded due to missing PMR but introduces a risk of missed links leading to inflation of the numerator.

All analyses were conducted in R with the data.table and ggplot2 packages. Statistical disclosure rules stipulated a minimum cell count of 10 (below which values are suppressed), that all frequencies be rounded to the nearest 10 and that all percentages be rounded to one decimal place.

Our findings agree closely with Bilson and Martin’s [5] prior estimates to age 5 of 14% being recognised as CiN and 3.5% subject to a CPP. Our respective estimates to age 5 were 14.2% and 4.1%. We are aware of no other published estimates of cumulative receipt of CSC services in England using administrative data across childhood [12], with the exception of Troncoso’s estimate of referral to age 5 of 19% using the CiN census [9], which is similar to our estimate of 20.4%.

Using freedom of information requests to local authorities in Scotland, Bilson and Macleod [13] estimated that 3.8% of children in Scotland born August 2012 to July 2013 were on the child protection register by age 5. This is the similar to our estimate for children being made subject to child protection plans in England by age 5 of 4.1% and to Bilson and Martin’s [5] and Bilson and Munro’s [14] estimates for England using freedom of information requests.

There is limited evidence from cohort and panel studies. Maxwell and others [15] in reviewing research data resources for social work, found that a range of cohorts, such as the Longitudinal Study of Young People in England, the Millenium Cohort Study and the National Child Development Study, contained questions relating to CSC services. In all cases, numbers of children in contact with CSC services were in the 100s and studies only collected limited retrospective data at specific waves. For example, Henderson et al [16] used the Longitudinal Study of Young People in England, in which parents were asked whether they had been in contact with their council’s social services in relation to their children’s behaviour at home or school (aged 13-14 years at study inception). Parents were asked the same question once per year over three waves, with 387 to 615 (3%-4% of 15,770) reporting such contact each wave, and 1,184 (8%) ever reporting contact over the three waves. In addition to relatively small samples, such studies are also subject to selection and attrition biases, which may be severe. Teyhan et al [17], for example, report that, of 12,868 participants in the Avon Longitudinal Study of Parents and Children, 137 linked to the children looked after return and 209 to the CiN census. Among those without linkage to these datasets, just under 60% had maternal questionnaires returned at age 12. This compares to around 35% for those who linked to the CiN census and under 20% for those who linked to the children looked after return.

Studies from Australia, the USA, New Zealand and Canada have used administrative data to estimate the cumulative incidence or cumulative prevalence of contact with children’s services (Table 2). Given differences in service configurations, direct comparison with England is difficult. In all studies, the focus was on the child protection system and/or substantiated maltreatment. In England, however, the remit of CSC is much broader: child protection is only a subset of England’s CSC activity. Nonetheless, the international studies, like ours, find high cumulative incidences of contact with their respective services. For example, the latest data from Australia show that in New South Wales and South Australia [18], around 40% of children are subject to a child protection notification before age 14 and a quarter to a third are subject to a screened-in notification. Over 10% were subject to an actual child protection investigation and 6% to 8% were recorded as having substantiated maltreatment (which is perhaps closest to England’s child protection plan). Therefore, although no other study provides a directly comparable estimate to ours, our findings show that England is similar to other countries in having very high proportions of children receiving some form of social care intervention across childhood and early adolescence.

The Department for Education collects only limited personal identifiable information in the CiN census [34], with the consequence that longitudinal linkage must rely on one of two encrypted identifiers in the dataset, the LA-child ID and PMR. As noted above, neither is perfect. Without personal identifiers such as name, address and date of birth, it is not possible to examine the reliability of the LA-child ID. Whereas all data in the CiN census may be subject to missed links, thereby biasing incidence estimates upwards, this is likely to particularly affect data on referrals where local authorities may possess poorer quality data. This could be caused, for example, where there is no follow up required or carried out and therefore the local authority does not hold accurate identifiable information to link an apparently new referral to an old one already held for the same child in their database. This may be particularly challenging where a child or parent changes name (e.g., due to marriage) or moves property. The result is that the same child within an authority could be assigned different identifiers and therefore erroneously counted more than once. On the other hand, it is known that not all “referrals” are recorded as referrals in the CiN census [8], which would serve to depress incidence estimates. For children who do receive services, these factors are less likely to be an issue as more detailed identifiable information will be held by the authority.

A remedy to these problems would be for the Department for Education to mandate collection of personal identifiers within the CiN census, and to do so retroactively. This would enable creation of a single, nationally unique, encrypted identifier within the CiN census, making it easier to eliminate double counting. Retroactive collection of personal identifiers would also enable assessment of linkage quality across children’s records within CiN, which would improve the quality not only of research using the CiN census, but also the Department for Education’s own uses of the data, such as their re-referral statistics [4] and use of early help [35]. This could further result in positive feedback loops to local authorities to improve their source data, especially as validation checks are applied at data submission from the local authority to the Department for Education [34].

Holding identifiers would also enable linkage of the CiN census to other datasets for all children, which is currently only possible for the subset of children with a PMR via the National Pupil Database (NPD) [8, 36–38]. Indeed, NPD data have been shown to be of very high quality as evidenced by 99% linkage rates between NPD and the Hospital Episode Statistics using personal identifiers in the Education and Child Health Insights using Linked Data project [39], suggesting that where PMR is currently available in the dataset, it is likely to represent a high quality matching variable.

Another limitation lies in the fact that we combined estimates from three separate cohorts. Our method assumes that there was no change in the underlying cumulative incidence of CSC service receipt across these cohorts. This assumption is supported by general stability in the proportions of children who are referred to social care and assessed as being in need since 2013 [3]. In terms of child protection, while there was a significant increase in the numbers of children investigated over this time period, the rate at which children were made subject to child protection plans increased only slightly [3, 5]. Similarly, while we made efforts to avoid double counting induced by left-censoring by estimating recurrence and adjusting numerators accordingly, these estimates, too, assume consistency across cohorts and age (the latter assumption is supported by Troncoso’s analysis of re-referrals [9]). Further, we were unable to investigate how our cumulative incidence estimates would vary by deprivation (it is expected to be significantly higher among poorer children [40]) as no such data are available in the CiN census. Additional research could focus on sex/gender and ethnicity, though this would require an investigation into the completeness and reliability of the gender/sex and ethnicity data in the CiN census and for the denominator population derived from the decennial census, especially given known problems with ethnicity data quality in other datasets such as the Hospital Episode Statistics [41–43].

Future estimates will be affected by a decrease in CSC activity during the COVID-19 pandemic [44], which would reduce estimates for children living through this period. Estimates for future cohorts will also be affected by changes in CSC practice, especially due to very large increases in the use of child protection investigations found by Bilson and Munro between 2011/12 and 2016/17 [14].

Finally, we were unable to account for death and emigration as potential competing risks. While death after infancy is very rare and unlikely to substantially affect results, not accounting for emigration may have upwardly biased estimates.

The consequence of the limitations in the identifiers is that it is difficult to estimate with exactitude the cumulative incidence of CSC activity in England using the CiN census. This will still be the case even when sufficient data have been collected to follow a single cohort from birth to age 18 (which, counting from 2011/12, will be in 2029/30). We recommend that the Department for Education review what personal identifiers it collects from local authorities on all children referred to CSC to enable complete future and retroactive data linkage and improved understanding of their circumstances through childhood.

Nonetheless, it is certain that very high proportions of children are referred to, assessed and recognised as needing services by CSC in England. Our findings support a shift of CSC services towards early help [45]. The high cumulative incidence of CiN also raises questions about the extent to which government policies on employment, benefits, housing, healthcare, education, social care and justice support or challenge nurturing and upbringing of children. Needs might be more effectively addressed by targeting the socio-economic and structural factors beyond individual or family circumstances that lead a large minority of children to become in contact with CSC services.