Predictive analytics in child
Prof Eileen Munro, Durham University and London School of
CHESS Working Paper No. 2019-03
[Produced as part of the Knowledge for Use (K4U) Research Project]
CHESS working paper (Online) ISSN 2053-2660
The K4U project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and
innovation programme (grant agreement No 667526 K4U) The above content reflects only the author's view and that the ERC is not re-
sponsible for any use that may be made of the information it contains
Predictive analytics in child protection
Prof Eileen Munro
50 Old Elvet
Durham DH1 3HN
Emeritus Professor of Social Policy
Department of Social Policy
London School of Economics and Political Science
London WC2A 2AE
In child protection work, difficult risk assessments need to be made when deciding what actions to
take when helping a child who has been abused or is likely to suffer abuse. Predicting the future is
difficult and errors can be false negatives (leaving a child in danger) or false positives (removing a
child who would not have been harmed). Both types of errors in child protection work have a
high cost, firstly to the child and family and secondly to the professionals involved. Society is
harshly critical when a child is left with parents or carers and is subsequently killed but there is
also, though less frequently, severe criticism when children are thought to be removed without
Prediction has also become more important in recent years because of the political interest in
‘social investment’ – providing early help to those seen as likely to be problematic in some way in
later life. Prevention services tackle a range of problems, including being maltreated, behaving
badly and having poor health and educational outcomes. One political option is to provide
universal services so there is no need to work out which children, families or communities to
target. This however is less widely used nowadays and there is a shift towards targeted services. It
therefore becomes necessary to find some means of determining who should receive additional
help to flourish. One solution is to make services available and leave parents to seek help but this
raises concerns that some families may not be willing to come forward.
Improving risk predictions has therefore attracted a lot of research attention. In the past,
instruments to help professionals have either been a form of guided professional judgement or an
actuarial instrument. In recent years, however, there has been a surge of interest in using
predictive analytics to address important decision points with the hope that they will increase
accuracy. This is facilitated by scale of datasets now available for data mining as more and more
agencies develop computerized records that can be linked. Data mining in child welfare and
protection has linked family members’ datasets from health, education, police, income and
housing, building a detailed set that enables profiling.
Predictive analytics are seen as having the ability or potential ability to identify the children (or
even foetuses) who should be targeted for additional help.
I am using the term predictive analytics to refer to decision making systems that use data mining
to identify patterns in large datasets and use algorithmic processes, including machine learning, to
automate or support human decision making. Machine learning is the process by which a
computer system trains itself to spot patterns and correlations in (usually large) datasets and to
infer information and make predictions based on those patterns and correlations without being
specifically programmed to do so. Typically, these systems involve ‘profiling’, the processing of
personal data about an individual in order to evaluate personal characteristics relating to their
behavior, preferences, economic situation, health etc.
Data mining has been considered valuable in other sectors. In healthcare, considerable work is
going into using predictive analytics to improve decision making. But the possibility of using it in
child protection causes considerable discussion and disagreement. There is general recognition
that it raises serious technical, legal and ethical challenges but there are differing views on
whether these can be overcome for particular decision tasks. Some are more optimistic (Cuccaro-
Alamin, Foust, Vaithianathan, & Putnam-Hornstein, 2017; Schwartz, York, Nowakowski-Sims, &
Ramos-Hernandez, 2017). Others are concerned that it will be used in ways that have a negative
impact on children and their families ((Church & Fairchild, 2017; Eubanks, 2017; Keddell, 2015;
There has also been a mixed pattern of development in usage and the decision they are designed
to support. Some places have developed decision tools based on predictive analytics and then
dropped them because of concerns about accuracy or ethics (New Zealand’s Predictive Risk Model
for early identification of future harm; Illinois Department of Children and Family Services Rapid
Safety Feedback for rating referrals to the agency hotline). Others however are in use (for
example Allegheny Family Screening Tool (2017); London Councils Children’s Predictive
There are many decision points at which predictive analytics might be used in child protection so
there is no single answer to this debate. My aim here is to identify the range of factors that need
to be considered in deciding whether they are useful, legal and ethical in a specific decision
context and to make some comments on the distinctive features of the child protection context.
On one aspect, the technical argument for predictive analytics is persuasive. Computers can
analyse much larger datasets than humans can manage. This analysis also comes with a degree of
accuracy and speed that outstrips human capabilities. It can uncover trends and insights a human
might discount or not even consider. Machine learning systems are trained using large datasets
provided by the system designer. Once trained, it can infer information or make predictions based
on additional data inputted to the system and processed according to the algorithm.
However computers work with the data created by humans in specific social, historical and
political conditions and consequently do not avoid biases and prejudices that may be buried in
that data. Caliskan et al’s study shows that ‘standard machine learning can acquire stereotyped
biases from textual data that reflect everyday human culture’ (Caliskan, Bryson, & Narayanan,
2017 p.183). Artificial intelligence and machine learning may perpetuate cultural stereotypes and
they conclude: ‘caution must be used in incorporating modules constructed via unsupervised
machine learning into decision-making systems’ (Caliskan et al., 2017 p.185).
The capacity of predictive modeling to contain hidden biases is a major concern in child protection
because of the nature of the datasets used. Their reliability and completeness are open to serious
challenge. As discussed in Chapter 6, the core concept of child maltreatment is problematic and
there is no universal, fixed and detailed definition of what it means. Professional judgments about
whether a child is experiencing or is likely to experience maltreatment have low reliability, i.e. low
inter-rater agreement. Several studies show this not just in making judgments about what counts
as maltreatment but also which cases reach the threshold for initial investigation or for removal
from the family and also whether or not practitioners are using decision support tools (e.g. Arad-
Davidzon & Benbenishty, 2008; Britner & Mossler, 2002; Jergeby & Soydan, 2002; Regehr, Bogo,
Shlonsky, & LeBlanc, 2010; Schuerman, Rossi, & Budde, 1999; Spratt, 2000). Therefore data is
influenced by the particular practitioner who entered it.
Gillingham (2015) raises further concerns about the way the data is constructed:
‘As information is entered into an information system, it has to be categorized according to the
fields built into the information system, which may or may not fit the circumstances the
practitioner has observed. This can happen in many ways (see Author’s own) but an obvious
example is the level of detail required by the information system. For example, a common
question in risk assessment tools concerns illicit drug use by caregivers. Ticking a yes/no box in
response to such a question is not only overly simplistic but confounding. In terms of data,
caregivers who occasionally smoke marijuana after the children have gone to sleep are put in
the same category as caregivers who inject heroin two or three times a day and spend much of
their time finding the means to do so. Dick (2017) calls this the ‘flattening effect’ of categorizing
data. Clearly there are different levels of risk of harm or neglect posed by each scenario’.
The degree of unreliability in the dataset has significance for the overall accuracy of any
predictions: ‘When associations are probed between perfectly measured data (e.g. a genomic
sequence) and poorly measured data (e.g. administrative claims health data), research accuracy is
dictated by the weakest link’ (Khoury & Ioannidis, 2014 p.1054).
The incompleteness of many of the datasets that are being used for predictive analytics is also a
concern since child protection datasets, in particular, are known to be incomplete in a non-
random way. They include children and families who have been referred to the service and these
are known to cover only a percentage of children who suffer maltreatment. Studies of people’s
self reporting of maltreatment in childhood reveal a much higher number than official statistics of
cases known to child protection services (Stoltenborgh, Bakermans‐Kranenburg, Alink, & van
Ijzendoorn, 2015). However, these studies also have wide variation among their results (Radford,
2011 Appendix D). Getting a reliable measure of a phenomenon such as child maltreatment is
very difficult. For example, in England, Gilbert (2009) estimates that only 10% of cases are
reported. Jud (2018) summarises a number of studies of the incidence of maltreatment which
show not only that a large majority of cases are not known to services but also that the incidence
varies depending on whether minor and moderate maltreatment are included as well as serious.
Moreover, there is evidence that the dataset has persistent biases in the over-representation of
low income families and ethnic minorities (Cawson, Wattam, Brooker, & Kelly, 2000).
The dataset of referrals to child protection also includes large numbers who, on subsequent
consideration, are deemed not to need a child protective service. In the US, 57.6 % referrals were
screened in during 2017 and 42.4% were screened out (Children's Bureau, 2018). In England,
37.9% of referrals were deemed not to need a service during 2017-18 (Department for Education,
Testing the accuracy of predictive tools is limited by the imperfect feedback available. If a
prediction that a child is in too much danger to remain with their family leads to the child’s
removal into alternative care, the prediction is never tested. If a child gets a low rating and stays
at home, testing is limited to the feedback from repeat referrals to child protection. Therefore,
the ability to learn and rectify any errors in the predictive algorithm is weak and limits its technical
There is considerable discussion in the literature on two key legal matters: problems of
anonymising data and the transparency and accountability of decision making.
Developers and users of predictive analytics need to pay attention to the local laws on privacy and
sharing of confidential and sensitive information without consent. In many jurisdictions,
confidentiality restrictions can be lifted in child protection cases. In the English law, the threshold
is if there is concern that a child is suffering or likely to suffer significant harm. This restriction has
different impact on the various decision tasks that predictive analytics are designed to support. If
the decision relates to assessing the risk of maltreatment of a child referred to child protection
then the tools developed have typically been using the set of data that is already available to the
professional decision maker. However, the growing interest in broadening the range of risk
assessment to preventive services and of combining data from a wider range of datasets from
other public and private services raises new legal questions.
One solution that is offered is of anonymising the data so that it cannot be linked to an identified
or identifiable individual; the individual’s privacy is still preserved. However, the research need is
for linked data – enabling development of a rich profile of an individual – and this creates a
Paul Ohm (2009) conducted a major review of the literature and reached the daunting conclusion:
‘Data can be either useful or perfectly anonymous but never both’ (2009 p.1704) . A similar point is
made in the Royal Society Report on ‘Science as an Open Enterprise’ (2012):
‘It had been assumed in the past that the privacy of data subjects could be protected by
processes of anonymisation such as the removal of names and precise addresses of data
subjects. However, a substantial body of work in computer science has now demonstrated that
the security of personal records in databases cannot be guaranteed through anonymisation
procedures where identities are actively sought’.
Korff and Georges (2015) clarify why this is so:
‘The main problem is that effective anonymisation does not just depend on stripping away
direct identifiers (name, address, national identification number, date of birth) from a data set.
Instead, the relevant measure is the size of the “anonymity set” – that is, the set of individuals
to whom data might relate. If you’re described as “a man” the anonymity set size is three and a
half billion, but if you’re described as “a middle-aged Dutchman with a beard” it is maybe half a
million and if you’re described as “a middle-aged Dutchman with a beard who lives near
Cambridge” it might be three or four ‘ (Korff & Georges, 2015).
Pseudonymisation is offered as a partial solution. It is defined in the EU General Data Protection
Regulation (GDPR) as ‘the processing of personal data in such a way that the data can no longer be
attributed to a specific data subject without the use of additional information’. The problem lies
in the phrase ‘without the use of additional information’. As databases increase, additional
information is becoming increasingly available.
An added danger for children comes from the potential linkages between welfare-related datasets
and others so that the profiles of children and their parents can be more detailed and hence more
readily de-anonymised. In the UK, for example, there are companies that pull together numerous
datasets and offer a service to help you understand the profiles of households and postcodes and
have also developed classifications covering health, retail and leisure activities. One such company
says it offers:
‘a geodemographic segmentation of the UK’s population. It segments households, postcodes
and neighbourhoods into 6 categories, 18 groups and 62 types. By analysing significant social
factors and population behaviour, it provides precise information and an in-depth
understanding of the different types of people’ (Acorn, 2019)
With such detailed additional information, identifying individuals becomes more probable.
The second major legal concern is the lack of transparency in decisions made according to an
algorithm and the consequent difficulties this causes if anyone wishes to challenge a judgment
made about them. One US state requires details of any algorithm to be made public but many are
being developed by private companies who refuse to publish on the grounds of it being their
intellectual property, and commercial concerns. Even if the details are available, few people
would be able to scrutinize it or understand the computation. The increasing number of ‘expert’
systems that create feedback loops to continuously improve the underlying algorithm create
another barrier to transparency. The problem comes in several forms:
This problem is often termed ‘algorithmic opacity’, of which three distinct forms have been
identified. The first is intentional opacity, where the system’s workings concealed to protect
intellectual property. The second is illiterate opacity, where a system is only understandable to
those with the technical ability to read and write code. And the third is intrinsic opacity, where a
system’s complex decision-making process itself is difficult for any human to understand. More
than one of these may combine – for example, a system can be intentionally opaque and it be
the case that even if it wasn’t then it would still be illiterately or intrinsically opaque. The result
of algorithmic opacity is that an automated system’s decision-making process may be difficult
to understand or impossible to evaluate even for experienced systems designers and engineers’
(Cobbe, 2018 p.5).
The use of predictive analytics is creating new challenges for legal systems as they alter the
transparency of decision making and the protection of privacy.
Most jurisdictions are now implementing regulations on predictive analytics. Transparency,
accountability and a ‘positive impact on society’ are among the key values. However, Zeuderveen
Borgesias offers a word of caution:
‘Several caveats are in order regarding data protection law’s possibilities as a tool to fight AI-
driven discrimination. First, there is a compliance and enforcement deficit. Data Protection
Authorities have limited resources. And many Data Protection Authorities do not have the
power to impose serious sanctions (in the EU, such authorities received new powers with the
GDPR [the EU General Data Protection Regulation]). Previously, many organisations did not take
compliance with data protection law seriously. It appears that compliance improved with the
arrival of the GDPR, but it is too early to tell’ (Zuiderveen Borgesius, 2018 p.24) .
How will the predictive tool be used?
A tool cannot be appraised in isolation. It will be used by people with human abilities and
limitations in a physical and cultural context. Will the interaction between these be constructive
A key problem will be in people’s understanding of how to interpret the results. The ‘base rate
fallacy’ is well evidenced as a common intuitive error. For professionals using predictive
instruments, the practical issue is how much confidence they should have in the results. If this
instrument predicts that Parent X is likely to harm her child, how likely is this to be true? If positive
results are often false positives, then professionals know they need to treat the result with
A famous study, ‘The Harvard Medical School Test’, illustrates the prevalence of the base rate
fallacy in evaluating predictive tests. Staff and students at Harvard Medical School were told of a
diagnostic test that had a high sensitivity of 95 per cent (of accurately identifying those with the
disease) and a superb specificity of 100 per cent (no one with the disease would test negative).
They were asked the probability of someone who tested positive actually having the disease. The
majority of respondents gave the answer of 0.95 – the rate of true positives – overlooking the
significance of the base rate in determining the accuracy (Casscells et al., 1978). As the following
section will explain, these estimates are far from accurate and, depending on whether the illness
being diagnosed was common or rare, this test might or might not be clinically valuable.
Bayes theorem is the formal probability calculation to work out how likely it is that the positive or
negative result is accurate but it is not intuitively obvious. When the underlying calculations are
presented in terms of probability formulae, people tend to find them hard to follow but
Gigerenzer and his colleagues at the Max Planck Institute for Human Development in Germany
have found that people are well able to understand the reasoning when it is presented in more
familiar ways (Gigerenzer, 2002).
To judge predictive accuracy – i.e. to judge among those who get positive results on the test, how
many are cases of abuse, we need to have the values of three variables:
Sensitivity: among many cases of abuse, how many will it predict accurately (true
Specificity: among non-abusive families, how many will it identify correctly (true
Base rate or prevalence of the phenomenon: how common it is in the population in
Each of these three variables plays a distinctive part in working out the overall usefulness of an
instrument, but it is the final one – the base rate – that is most often over-looked or
misunderstood. Put briefly, the rarer the phenomenon being assessed, the harder it is to develop
an instrument with a clinically useful level of accuracy. Conversely, the higher the base rate, the
easier it is. Hence, researchers face a harder task trying to develop a risk assessment instrument to
screen the general population, where the incidence of abuse is relatively low, than if their target
population was specifically families known to child protection agencies, where the base rate will
be much higher.
Let us take a practical example as illustration of the impact of the base rate and show how it leads
to different results even when the sensitivity and specificity remain the same and are fairly high.
Suppose we have an instrument where, the sensitivity is 90 per cent, and the specificity is 80 per
cent and the base rate is 10 per cent
Ten out of every 100 families in this population are abusive (the base rate). Of these 10
families, 9 will get a positive result on using the instrument (the sensitivity of 90 per
Of the other 90 families, around 72 will accurately get a negative result but some 18 will
get a (false) positive result (the specificity of 80 per cent).
Imagine the instrument has given a positive result for a group of families. How many of
these families with a positive result will actually be abusive? This tree diagram helps to
make the calculation clearer.
The calculation shows that, in total, 27 families will get a positive result, of which 9 will
be true positives and 18 false positives. Thus the probability of a positive result being a
true positive is: 9 divided by 27 = 0.33.
In short, about two thirds of the families judged dangerous by the instrument will not be.
In contrast, if the tool is used on a sub-group of the population where we have reason to assign a
higher base rate of 40% then the equivalent figures are:
Forty out of every 100 families in this population are abusive (the base rate). Of these 36
families will get a positive result on using the instrument (the sensitivity of 90 per cent).
Of the other 60 families, around 48 will accurately get a negative result but some 12 will
get a (false) positive result (the specificity of 80 per cent).
In total, 48 families will get a positive result, of which 36 will be true positives and 12 false
positives. Thus the probability of a positive result being a true positive is: 36 divided by 48 = 0.75.
In short, a quarter of the families judged dangerous by the instrument will not be so, i.e. false
The crucial message is that it is surprisingly hard to develop a high accuracy rate in predicting a
relatively rare event. Even instruments with what seem to be impressively high statistics about
how many families they will accurately identify as abusive or safe have a disappointingly low
overall accuracy: the majority of the families the instrument identifies as abusive will, in fact, be
non-abusive; that is, they will be false positives.
The danger is that the base rate fallacy will adversely influence people’s use of the results of
predictive analytics. Added to this risk is ‘automation bias’ - the tendency for people to have
undue confidence in the results produced by computers so they are more likely to discount
contradictory evidence than people who are making judgments. There is evidence that this is a
significant source of error in aviation and medicine (Goddard, Roudsari, & Wyatt, 2011) and so it
may be a problem in child protection. A lack of understanding of the importance of base rates is
likely to lead to over-confident use and the prevalence of defensive practice in societies that react
very punitively when child protection workers fail to protect a child from being killed may increase
the automation bias.
Many of the existing tools are, as their name suggests, designed to support decisions by
professionals. The designers of the tools generally stress that it should be treated as one among
several factors that the professional considers. The Allegheney Family Screening Tool, for
instance, reports the score for a referral along with text explaining that the system ‘is not intended
to make investigative or other child welfare decisions’. However, some may be reluctant to use
their professional expertise to reach a different decision than the one recommended by the tool.
In the event of an adverse outcome, there is a safety in blaming the tool for the decision and a
fear that they would struggle to justify going against its recommendation. The more defensive the
work culture, the more likely automation bias will occur.
Proponents of predictive analytics point to the increased accuracy of the decisions made with the
support of the automated analysis. Critics worry about how it will fit into the whole process of
working with families. A balanced assessment of what is the best action to take for a child requires
a positive assessment of the rewards in a situation not just the risks. Also, most, if not all, child
protection practice approaches involve the worker building a relationship with family members in
order to understand their problems and help them provide safer care. How and whether
professionals can integrate the predictive tools constructively into this relationship is a concern
raised by some (Broadhurst, Hall, Wastell, White, & Pithouse, 2010; Oak, 2015).
Oak also raises the question of whether having the risk assessment performed by a tool will ‘lead
to the erosion of critical thinking and professional judgment skills, including the ability to define
key concepts such as ‘risk’ or ‘abuse’ and to recognise that they are socially constructed and
contested entities’ (2015 p.1215). This seems to overstate the role that predictive analytics are
intended to play. Predictive analytics are being developed for the major decision points such as
whether to investigate an allegation of harm but workers make many decisions every day. The
development of decision support systems does not eliminate the need for professionals to assess
risk and make decisions on how best to manage it in their daily work. In some respects, these
seem small matters. For instance, workers with heavy workloads (as most are) have to make
decisions about how to use their time, which families or other activities need to be prioritized.
These decisions will involve risk assessments in deciding which families to prioritise visiting. It is
only with hindsight that some of these decisions may be seen to be pivotal in the management of
the case – an unplanned home visit seeing evidence of harm or a visit being delayed and swiftly
being followed by the child suffering injury.
The final questions to ask about the use of such predictive analytics relate to whether they are
morally acceptable. What benefits will they produce for children and their families? What harm
might they do? How do you balance these out?
When used preventively to screen families to identify those children who are likely to develop
problems, they raise the standard questions of any screening method. How accurate is the
screening tool? Do we have effective services for resolving the predicted problem? Do we have
sufficient resources to provide those services?
Judging the accuracy of the screening tool is not just a technical matter but also requires making a
judgment about the risk threshold. As discussed earlier, the accuracy of a predictive tool is related
on the base rate of the phenomenon you are seeking to predict plus the sensitivity and specificity
of the tool. Decision support systems derived from predictive analytics are not 100% accurate and
will never be so but this means that decisions need to be made about the risk threshold for action
– the balance between the sensitivity and specificity of the predictions. As I discussed earlier in
Chapter Four, these are inversely related: if we want to improve the sensitivity (have a low rate of
false negatives, of missing children) then automatically we lower the specificity (we increase the
number of false positives, inaccurately identifying children).
Do we have effective enough methods to deal with identified needs? A key principle of health
screening is that it has benefits for those screened because effective methods area available to
mitigate the potential harmful outcome that the screening identifies. In child protection, it is not
enough to say that some intervention has been shown to be effective (usually in comparison with
another intervention or no treatment in an RCT). We also need to know what percentage of
people showed benefit, how great that benefit was, and whether for some there were negative
Do we have sufficient resources to provide the services? Typically, preventive services need to
provide help to a large number of families (the false positives) in order to include those who might
otherwise have developed serious problems.
In discussing the growing interest in screening for adverse childhood experiences (ACEs), Finkelhor
puts the counter argument:
‘We are going to argue here that it is still premature to start widespread screening for ACES
until we have answers to several important questions: 1) what are the effective interventions
and responses we need to have in place to offer for positive ACE screening, 2) what are the
potential negative outcomes and costs to screening that need to be buffered in any effective
screening regime, and 3) what exactly should we be screening for?’ (Finkelhor, 2018 p.175).
What do we know about the actual or potential negative effects of being profiled? Predictions
may be carried out by well-motivated professionals who want to help families but that does not
necessarily mean that they will have beneficial effects. Problems around parenting and child
development are all too easily seen negatively by others. The mere fact of being known to
children’s services can be stigmatizing and be interpreted by some as a damaging mark against
you whether you are an adult or a child. In an English trial of a national database on all children,
including all services with which they were in contact, one school Head used the database to
screen out all applicants who had a history of being known to Children’s Social Care. This was, of
course, an illegal use but it still had a harmful effect on the children involved. It would be naïve to
assume that criminality would be rare when detailed databases are becoming of increasing
practical and commercial value.
Finally, for all decisions in which predictive analytics may be used, there is a significant danger of
preserving existing biases and prejudices in professional practice but making them more
dangerous because they are hidden from sight in the performance of an apparently neutral
scientific mechanism for reaching judgments.
To summarise, predictive analytics may be used for different decision tasks in child protection, the
major ones being early identification of families likely to become problematic, decisions on
whether to investigate a referral, and decisions on removing or returning a child to their home.
Benefits and problems with predictive analytics need to be appraised in relation to the specific
decision task they are aiming to support.
Their introduction raises many technical, legal, and ethical concerns.
‘Machine learning systems are known to have various issues relating to bias, unfairness, and
discrimination in outputs and decisions6, as well as to transparency, explainability, and
accountability in terms of oversight7, and to data protection, privacy, and other human rights
issues, among others’ (Cobbe, 2018 p.5)
A concluding point is that even the most accurate, legal and ethical tools only cover a small part of
the task of improving children’s safety and well-being. They omit the assessment of the positive
aspects of families. Working with a family to provide safe enough care or providing good
alternative care will continue to absorb most professional time.
Despite the many counterarguments and concerns about using predictive analytics, many
jurisdictions are introducing decisions support systems derived from them to tackle urgent
practical problems in targeting limited services. Perhaps they should consider the advice given by
Zuiderveen Borgesius on proceeding with caution:
‘The public sector could adopt a sunset clause when introducing AI systems that take decisions
about people. Such a sunset clause could require that a system should be evaluated, say after
three years, to assess whether it brought what was hoped for. If the results are disappointing, or
if the disadvantages or the risks are too great, consideration should be given to abolishing the
system’ (Zuiderveen Borgesius, 2018 p.29).
When viewed from the narrow point of improving decisions relating to children’s safety and well
being, predictive analytics look appealing, harnessing the information buried in vast databases to
guide professional decision making. However, when this task is placed in the wider context of the
technical processes involved and the social situations in which the tools are used, a large number
of problems emerge - the hidden bias in the algorithms, the incompleteness and unreliability of
the datasets, the lack of transparency, and the impact upon families. Considerable work is going
on in artificial intelligence and in improving the law and regulation relating to its use and these
may make sufficient progress to reduce some of the difficulties. However, at present, the use of
predictive analytics in child protection seems to introduce too many new problems that outweigh
their potential benefits
Acorn. (2019). Acorn User Guide. Retrieved from https://acorn.caci.co.uk/downloads/Acorn-User-
Allegheny County. (2017). Frequently Asked Questions about the Allegheny Family Screening Tool.
Retrieved from Allegheny County, Pennsylvania:
questions-allegheny-family-screening-tool/ downloaded 30.01.19
Arad-Davidzon, B., & Benbenishty, R. (2008). The role of workers' attitudes and parent and child
wishes in child protection workers' assessments and recommendation regarding removal
and reunification. Children and Youth Services Review, 30(1), 107-121.
Britner, P. A., & Mossler, D. G. (2002). Professionals’ decision-making about out-of-home
placements following instances of child abuse. Child Abuse & Neglect, 26(4), 317-332.
Broadhurst, K., Hall, C., Wastell, D., White, S., & Pithouse, A. (2010). Risk, Instrumentalism and the
Humane Project in Social Work: Identifying the Informal Logics of Risk Management in
Children's Statutory Services. British Journal of Social Work, Advance ac.
Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language
corpora contain human-like biases. Science, 356(6334), 183-186.
Cawson, P., Wattam, C., Brooker, S., & Kelly, G. (2000). Child Maltreatment in the United Kingdom:
A Study of the Prevalence of Child Abuse and Neglect. London: NSPCC.
Children's Bureau. (2018). Child Maltreatment 2017. Retrieved from Washington, DC:
Church, C. E., & Fairchild, A. J. (2017). In Search of a Silver Bullet: Child Welfare's Embrace of
Predictive Analytics. Juvenile and Family Court Journal, 68(1), 67-81.
Cobbe, J. (2018). Administrative Law and the Machines of Government: Judicial Review of
Automated Public-Sector Decision-Making.
Cuccaro-Alamin, S., Foust, R., Vaithianathan, R., & Putnam-Hornstein, E. (2017). Risk assessment
and decision making in child protective services: Predictive risk modeling in context.
Children and Youth Services Review.
Department for Education. (2018). Characteristics of children in need 2017 to 2018. Retrieved from
Eubanks, V. (2017). Automating Technology, How High-Tech Tools Profile, Police and Punish the
Poor. New York: St Martins Press.
Finkelhor, D. (2018). Screening for adverse childhood experiences (ACEs): Cautions and
suggestions. Child Abuse & Neglect, 85, 174-179.
Gilbert, R., Widom, C. S., Browne, K., Fergusson, D., Webb, E., & Janson, S. (2009). Burden and
consequences of child maltreatment in high-income countries. The Lancet, 373, 68-81.
Gillingham, P. (2015). Implementing Electronic Information Systems in Human Service
Organisations: The Challenge of Categorisation. Practice, 27(3), 163-175.
Goddard, K., Roudsari, A., & Wyatt, J. C. (2011). Automation bias: a systematic review of
frequency, effect mediators, and mitigators. Journal of the American Medical Informatics
Association, 19(1), 121-127.
Jergeby, U., & Soydan, H. (2002). Assessment Processes in Social Work Practice When Children Are
at Risk: A Comparitive, Cross-National Vignette Study. Journal of Social Work Research and
Evaluation, 3(2), 127-144.
Jud, A. (2018). Current research on child maltreatment epidemiology: BioMed Central.
Keddell, E. (2015). The ethics of predictive risk modelling in the Aotearoa/New Zealand child
welfare context: Child abuse prevention or neo-liberal tool? Critical Social Policy, 35(1), 69-
Khoury, M. J., & Ioannidis, J. P. (2014). Big data meets public health. Science, 346(6213), 1054-
Korff, D., & Georges, M. (2015). Passenger Name Records, data mining & data protection: the
need for strong safeguards: T-PD (2015).
Oak, E. (2015). A minority report for social work? The Predictive Risk Model (PRM) and the Tuituia
Assessment Framework in addressing the needs of New Zealand's vulnerable children. The
British Journal of Social Work, 46(5), 1208-1223.
Ohm, P. (2009). Broken Promises of Privacy: Responding to the Surprising Failure of
Anonymization. Colorado: Colorado Law School.
Radford, L. (2011). Child abuse and neglect in the UK today. Retrieved from London:
Regehr, C., Bogo, M., Shlonsky, A., & LeBlanc, V. (2010). Confidence and professional judgment in
assessing children’s risk of abuse. Research on Social Work Practice, 20(6), 621-628.
Schuerman, J., Rossi, P. H., & Budde, S. (1999). Decisions on placement and family preservation:
Agreement and targeting. Evaluation Review, 23(6), 599-618.
Schwartz, I. M., York, P., Nowakowski-Sims, E., & Ramos-Hernandez, A. (2017). Predictive and
prescriptive analytics, machine learning and child welfare risk assessment: The Broward
County experience. Children and Youth Services Review, 81, 309-320.
Society Royal. (2012). Science as an Open Enterprise.
Spratt, T. (2000). Decision making by senior social workers at point of first referral. British Journal
of Social Work, 30(5), 597-618.
Stoltenborgh, M., Bakermans‐Kranenburg, M. J., Alink, L. R., & van Ijzendoorn, M. H. (2015). The
prevalence of child maltreatment across the globe: Review of a series of meta‐analyses.
Child Abuse Review, 24(1), 37-50.
Zuiderveen Borgesius, F. (2018). Discrimination, Artificial Intelligence and Algorithmic Decision
Making. Retrieved from Strasbourg, France: