Teenagers Pathways Through Childrens Social Care
Teenagers Pathways Through Childrens Social Care
Teenagers Pathways Through Childrens Social Care
pathways
through
children's social
care
May 2022
In collaboration with
Contents
Summary 4
Acknowledgements 4
Main points 4
Introduction 6
Data, definitions, and methodology 6
Children’s social care definitions 6
Overview of the dataset 7
Clustering methodology 7
Limitations 8
Section 1: Overview of journey types 9
1.1 Sizes of different cohorts 9
1.2 Characteristics of journey types 10
Length and frequency of interventions 10
Proportion of children experiencing interventions 12
Section 2: Minimal contact cluster 13
2.1 Prototypical case 13
2.2 Child characteristics 13
Gender 13
Ethnicity 14
Free School Meals 15
2.3 Needs 16
3.4 Outcomes 17
Section 3: Escalating risk cluster 18
3.1 Prototypical case 18
3.2 Child characteristics 19
Gender 19
Ethnicity 20
3.3 Needs 21
3.4 Outcomes 22
Section 4: De-escalating risk cluster 23
4.1 Prototypical case 23
4.2 Child characteristics 23
Gender 23
2
Ethnicity 24
4.3 Needs 25
4.4 Outcomes 26
Section 5. Long term care cluster 27
5.1 Prototypical case 27
5.2 Child characteristics 27
Gender 27
Ethnicity 28
5.3 Needs 29
5.4 Outcomes 30
Section 6. Long term child in need cluster 31
6.1 Prototypical case 31
6.2 Child characteristics 31
Gender 31
Ethnicity 32
Special Educational Needs (SEND) 33
6.3 Needs 34
6.4 Outcomes 34
Annex A. Additional findings 36
Does the length of interventions change over time? 36
What does prior attainment for those children look like? 37
Annex B. Model validation 38
Bibliography 39
Logo of Data Science Laboratory has been designed using resources from Flaticon.com
3
Summary
Acknowledgements
This analysis was carried out by the Department for Education’s Data Science
Lab based on a brief prepared by the review. A Memorandum of Understanding
covering ways of working and governance was put in place.
The review is grateful to members of the Data Science Lab for their work on the
project and continued support in preparing this publication.
Main points
The following annex is an experimental and exploratory analysis of the journeys
of teenagers through the social care system.
We focused exclusively on two birth cohorts: children born in 2000 and 2001,
who went from the age of 12 to 16 during the period covered by our data.
4
● there is a group of children who spend most of their teenage years in
long term care. They have the second highest average progress
between Key Stage 2 and Key Stage 4
● the smallest cluster of children are those who experience long term
children in need plans. This group is characterised by a high proportion
of children with special educational needs and disabilities (SEND)
5
Introduction
6
o is provided with accommodation, for a continuous period of more
than 24 hours [Children Act 1989, section 20 and 21]
o is subject to a Care Order [Children Act 1989, Part IV]
o is subject to a Placement Order
The unique identifier of children in this dataset was used to match to the
National Pupil Dataset – this allowed for exploration of children’s Key Stage 4
outcomes, socio-economic status (with Free School Meal eligibility as a proxy)
and Special Educational Need and Disability status.
Clustering methodology
The following factors were used in defining journeys of children through the
system:
● lengths of statutory “interventions” (children in need plans, child
protection plans and being looked after) in each year between the ages
of 12 and 16
● number of children’s social care “events” in each year between the ages
of 12 and 16. Events include number of assessments, number of
referrals flagged as “no further action”, number of section 47 inquiries
and initial child protection conferences
A data led approach – “K-means clustering” – was used to define types of
journeys based on these factors. K-means clustering is an unsupervised
learning mechanism, which is a family of machine learning tools used to
discover underlying patterns and information in the data (in this case, length of
interventions and numbers of events) without using any user defined labels.
In our case, a decision was made to split the data into five clusters. Fewer than
five resulted in groups that, on qualitative inspection, were too broad to add
7
much insight to the evidence base. Each additional cluster beyond five would
increase the granularity of the analysis and potentially add further insight, but
gives progressively smaller clusters, with less power for follow up analysis, and
increases complexity and hinders our original purpose of simplifying the overall
picture. To balance interpretability of results with robustness of the model, five
clusters were selected.
Limitations
No attempt has been made to establish whether reported differences are the
consequences of journeys, children’s characteristics and circumstances or any
other factors.
8
Section 1: Overview of journey types
9
1.2 Characteristics of journey types
Length and frequency of interventions
10
Figure 3 illustrates the scaled frequency of events experienced by children in
each of the clusters.
11
Proportion of children experiencing interventions
For example, whilst the average length of a child in need plan for a 13 year old
child in the escalating risk group is less than one month, figure 4 shows a very
small proportion of children that age experience the intervention, resulting in an
average of less than one month. For children aged 16 in that cohort, more than
85% have a child in need plan, which results in an average of around seven
months for children in that group.
12
Section 2: Minimal contact cluster
The minimal contact group is characterised as having little contact with the
children’s social care system throughout their adolescence.
Figure 5 illustrates the gender split in this cohort compared to the national
average of children in Year 11 in 2018.
Figure 5
13
Ethnicity
Figure 6
14
Free School Meals
Figure 7 illustrates Free School Meals eligibility for children in each cluster.
Figure 7
15
2.3 Needs
For analysis of children’s needs, we used a variable which corresponds to
factors identified by social workers at the end of assessment. It should be noted
that not all children have an assessment in this period, and not all of those
assessments have complete information about the factors involved. A social
workers may identify multiple factors at the end of assessment.
Figure 8
16
3.4 Outcomes
Figure 9 illustrates an average Attainment 8 score by children representing
each journey type.
Figure 9
This shows that children with minimal contact have the highest average
Attainment 8 scores among all clusters.
17
Section 3: Escalating risk cluster
This journey is characterised by little contact with the system prior to the age of
14, with a substantial increase in activity after this point.
18
3.2 Child characteristics
Gender
Figure 10
Escalating risk cluster has the highest over-representation of girls among all
groups.
19
Ethnicity
Figure 11
This shows that children from the White, Mixed ethnicity and Black ethnic
groups are over-represented in this cohort.
20
3.3 Needs
Figure 12 shows factors identified at the end of assessment for this cluster
with the highest over-representation score. Over-representation has been
calculated by comparing the proportion of all assessment with a specific factor
identified at the end for this cohort compared to all assessments in this
sample.
Figure 12
Whilst mental health concerns about the health of the child are the most
represented factor for this cohort, it should be noted that several of the over-
represented factors related to harms outside the home (such as socially
unacceptable behaviour and child sexual exploitation).
21
3.4 Outcomes
Figure 13 shows average Progress 8 scores by journey types.
Figure 13
Given a high proportion of these children are in contact with the system in their
Key Stage 4 year, this finding appears to be consistent with the DfE analysis
that shows having contact with the social care system in an exam year was
associated with lower likelihood of pupils achieving a strong pass in English
and maths after controlling for other characteristics (Department for Education,
2019).
22
Section 4: De-escalating risk cluster
This journey typically involves a fairly large amount of contact between the ages
of 12 and 14, with much less intervention between the ages of 14 and 16.
Figure 14
23
Ethnicity
Figure 15
Children from the White and Mixed ethnic groups are over-represented in this
group.
24
4.3 Needs
Figure 16 shows factors identified at the end of assessment for this cluster with
the highest over-representation score.
Figure 16
The plot above shows that four of the five most over-represented factors for this
group relate either to neglect, or to issues affecting the child’s parent or carer.
Whilst further research is needed, one of the hypotheses is that these kinds of
factors become less of a concern to social workers as a teenager gets older
and potentially becomes more independent from their parent/carer.
25
4.4 Outcomes
Figures 17 and 18 show that whilst these children’s average Progress 8
scores are middlemost compared to all clusters, they have the second highest
Attainment 8 score (after minimum contact cluster).
Figure 17
Figure 18
26
Section 5. Long term care cluster
Figure 19
27
Ethnicity
Figure 20
White children are the most over-represented group in this cluster and children
from the Mixed ethnicity group are more over-represented in this group than
any other ethnic group. Whilst children from the Asian ethnic group are under-
represented in all clusters, they are most under-represented in this cluster.
28
5.3 Needs
Figure 21 shows factors identified at the end of assessment for this cluster with
the highest over-representation score for this cluster.
Figure 21
Missing is the most over-represented factor for this group – given the
prevalence of children in care in this cohort, a possible explanation is that this
relates to episodes of these children missing from the care placement itself.
29
5.4 Outcomes
Figure 22 shows this group achieves the second highest average progress
score in the sample. This is consistent with the wider literature which shows
that long term care has a protective effect on educational outcomes (Sebba et
al., 2015).
Figure 22
30
Section 6. Long term child in need cluster
This journey typically involves long periods on a child in need plan throughout
the child’s teenage years, with other kinds of intervention relatively rare.
Figure 23
31
Ethnicity
Figure 24
Children from the Black ethnic groups are significantly over-represented in this
cluster, more than in any other group compared to the national average of
children in Year 11 in 2018.
32
Special Educational Needs (SEND)
Figure 25 shows the proportion of children with a SEND status in Key Stage 4
by cluster.
Figure 25
Just over 75% of this cluster are recorded as having SEND when they take their
exams – the highest proportion among all groups.
33
6.3 Needs
Figure 26 shows factors identified at the end of assessment for this cluster with
the highest over-representation score for this cluster.
Figure 26
This shows that factors related to child’s learning and physical disabilities are
overwhelmingly over-represented in this group, as well as factors related to
disabilities and illnesses of the child’s parents or carers.
6.4 Outcomes
As per figures 27 and 28, children in this cluster have the lowest average
Attainment 8 and the second lowest Progress 8 score – this likely reflects the
large proportion of children with a Special Educational Needs status in this
cluster.
34
Figure 27
Figure 28
35
Annex A. Additional findings
Figure 29
36
What does prior attainment for those children look
like?
Figure 30
Distributions for the escalating risk, minimal contact and de-escalating risk
clusters are fairly similar, with band 2 being the most common. Long term
clusters are somewhat unusual, as they have the highest proportion of children
from the top prior attainment band out of all of these groups, and also the
second-highest proportion in the lowest band.
37
Annex B. Model validation
To assess the quality of the clustering model, we created silhouette scores for
each cluster. This measure uses compactness of individual clusters and
separation among clusters to measure an overall representative score of how
well the algorithm has performed.
A silhouette score close to 1 means that the clusters are very dense and well
separated from other clusters. A score close to 0 means that clusters are
overlapping. A score close to -1 means that the data might have been assigned
to the wrong cluster.
Figure 31
38
Bibliography
Sebba, J., Berridge, D., Luke, N., Fletcher, J., Bell, K., Strand, S., Thomas, S.,
Sinclair, I., & O’Higgins, A. (2015). The Educational Progress of Looked After
Children in England: Linking Care and Educational Data. Rees Centre &
University of Bristol. https://research-information.bris.ac.uk/en/publications/the-
educational-progress-of-looked-after-children-in-england-link
39