Predictive risk modelling: on rights, data and politics.

One of the items included in the scope of the current New Zealand government’s review of the Child, Youth and Family services (CYFS) is this one: ‘The potential role of data analytics, including predictive risk modelling, to identify children and young people in need of care and protection’.

Predictive risk modelling (PRM) is a simple and seductive idea. If we can predict with accuracy who is likely to abuse children before they have done so, then we can target services to those families, fulfilling the dual objectives of preventing harm before it occurs, and being uber efficient with taxpayer dollars. Such seductive ideas, especially in an age where access to the ‘big data’ required to attempt such a proposition is viable, are often worth investigating. Enormous datasets can be mined, a large number of variables can be included, and patterns of particular combinations of risk factors for certain populations can be identified. In the case of the proposed Ministry for Social Development (MSD) PRM tool, however, there a number of issues. In particular, the level of accuracy of the PRM tool is overstated, the data it relies on has serious problems, its use as a practice decision-making tool is minimal, it has significant rights implications, and using it to decide who should be offered preventive services may not be any more effective than the current state of affairs (although to be fair this is difficult to ascertain – but needs to be).

Let’s start with rights. Human rights are supposed to be universal, that is, they apply to everyone equally. Normally, the only time a person might be subject to restrictions on their rights is if they commit a crime or harm another. However, the utility of the PRM rests on the premise that identifying in advance that a person might commit a crime or harm another justifies the overriding of ordinary rights to privacy and the stigma that accompanies it. One unintended consequence of this override is the creation of a category of people to whom it’s okay to assign a lesser level of rights than others, because they are ‘high risk’ and we have the science to prove it. McQuilan (2015) argues that data mining that attempts to predict future outcomes rests on a confusion between causation and correlation, and this creates a ‘state of exception’ for certain groups of people. Such a mode of governance suspends ordinary civil rights to privacy because of the risk of future harm – despite no current wrongdoing or certainty about that future harm. A hierarchy of rights is created, with ‘high risk’ people (unknowingly) trading their rights to privacy and fair treatment for the greater good of the future protection of children. Is this a fair trade? Weighing this up requires careful consideration of accuracy and effectiveness.

Firstly, let’s examine claims about the accuracy of PRM. These claims tend to be relative to the technical accuracy of other statistical models, rather than ‘accuracy’ in any kind of common understanding of that word. For example, despite statements that the PRM tool is ‘more accurate than anything seen before’, or is accurate compared to ‘other models of its type’, in the top decile of risk [1] it’s 48% accurate, and over the top two deciles, it’s 37% accurate [2] (Vaithianathan, 2012). I think if you got on a plane and were told that there was only a 48% chance of reaching your destination, or told that the TV guide you were looking at was only 37% accurate on the programmes that were screening that night, calling this ‘accurate’ would be a stretch. In an abstract sense, a statistical modeller comparing one level of accuracy with another another less accurate model might be satisfied.  However, in the context of identifying families at risk, offering preventive services and, potentially, intrusive CYFS investigations, is this accurate enough?

One issue affecting the accuracy of the PRM is that it relies on substantiation data as the outcome variable. The whole model is built around predicting this outcome, and risk factors are associated with this decision point, despite widespread acceptance in research domains that substantiation does not represent a realistic way of measuring child abuse incidence (Mumpower, 2010; Putnam-Hornstein, Webster, Needell, & Magruder, 2011). Particular groups of people are notoriously overrepresented in child abuse statistics gathered from child protection services, while other abuse is not picked up at all. ‘Visibility bias’ affects initial notifications to child protection services, and results in the over-identification of people in contact with referrers most likely to contact CYFS: those receiving state benefits, those attending low-decile schools with SWIS social workers, and those receiving other social services. This process tends to over-identify those who are poor and those overrepresented within the poor – Maori, Pasifika and women. While it’s true that poverty does increase the risk of child abuse, the level of overrepresentation is probably overstated in child protection services notifications. A national incidence study using a population-based representative sample would help: we don’t have one.

After cases are notified, they pass through several major decision points, one of which is substantiation – ostensibly a decision about whether abuse has occurred or not. However, the substantiation decision is notoriously variable. For example, in New Zealand, the rate of notification to substantiation varies between 5% and 48% depending on the site office (Wynd, 2013). International studies have found that numerous factors contribute to substantiation variability, such as: differing perceptions of risk, differing attitudes towards family preservation, different levels of neighbourhood deprivation and different community/caregiver resources (see Keddell, 2014). This means that children in similar circumstances in different towns may get different substantiation decisions. Put simply, neither the initial notification ‘pool’, nor the subsequent subset of substantiation decisions, provides an accurate picture of the actual incidence of child abuse in the population. Therefore, developing a PRM tool that claims to predict who will have a substantiated finding of abuse in the CYFS data is not the same as saying it can predict child abuse across a whole population.  Understanding this devilish detail challenges recent claims that the PRM is ‘free of race or class biases’ as well as being more ‘objective’ than social work decision-making. Instead, it is likely to reproduce whatever ethnic or class biases already exist in the system at notification, and reflect existing social work decision-making patterns at substantiation.

This leads to another major issue with the use of PRM. Despite the issues of data integrity, the tool may well identify a group of people who require more assistance and support with parenting. However, from what we know about the relationship between child maltreatment and poverty, this extremely ‘high risk’ group can also be considered to be the product of a society riven with inequalities and lack of social protections. To offer an individualised ‘service’ that does not, indeed cannot, address the broader social policy issues is deeply problematic. It assigns a stigmatised status to individuals in a manner that removes all attention from the wider policy landscape, and implicitly creates a narrative that there are certain people ‘out there’, who are fundamentally different from the rest of us, who have some kind of innate tendency to abuse children that, unless identified and prevented, will eventually express itself like a hidden gene waiting to trigger an inevitable disease process. Pathologising families in this way should trouble us. No one wants children to be abused, and a focus on prevention is welcome, but ‘othering’ a sector of the population with no attention to the social landscape that contributes to their problems sorely challenges the ‘effective intervention’ side of the argument, while strengthening the stigma downsides.

The root of the problem is that we are asking the wrong question: ‘How can we identify the families who will go on to become abusive?’ is fundamentally the wrong question to ask. A better question is ‘how can we more effectively address the risk factors that contribute to abusive behaviour?’ These are well known and exist across the ecological spectrum. If poverty and drug abuse are both highly correlated with child abuse, then why don’t we reduce poverty and provide more effective and accessible drug treatment services? If previous contact with CYFS as a child is associated with future contact as a parent, then let’s provide better therapeutic and support services to care leavers and other children involved in the care system. If poor communities that are highly transient and with high ratios of children are at greater risk, then make a real investment in community development services to create social cohesion and parenting support groups in such communities. If struggling parents are more likely to access secondary prevention services that are attached to universal services, then let’s provide them. This way of re-imagining social services, is well evidenced, and founded on the principle that such approaches are not only likely to prevent children from harm, they are also public goods in their own right, that is, they are likely to raise levels of wellbeing for parents, children and the community as a whole.

An emphasis on social wellbeing across the population reflects the social democratic ideals of earlier times in Aotearoa and elsewhere, where the state viewed its purpose as providing widespread social and economic protection. Most successful responses to child abuse include a similar whole of population approach (see, for example, the recent European Report on Preventing Child Maltreatment).

A PRM, on the other hand, dovetails nicely with a neo-liberal residualist concern with only expending precious resources on those who will ‘definitely’ go on to become abusive, rather than create a social landscape that aims to promote wellbeing for all children and their parents. Without this basis, increasing numbers of families fall through the ever-widening cracks – cracks that neither a PRM or the services offered in  response to a high risk score – can fix.  The wellbeing and safety of New Zealand children will not be secured by algorithms or ‘clever’ schemes to encourage the wealthy to invest in NGO programmes. It will not be obtained by shifting the blame and responsibility for child maltreatment onto individual families. The wellbeing and safety of all New Zealand children can only be won when New Zealanders insist on a government that commits itself to repair the damage of decades of neoliberal neglect.

Emily Keddell, Senior Lecturer in Social Work, University of Otago.

If you liked this, you may also like: Keddell, E. (2014) The ethics of predictive risk modelling in child welfare: child abuse prevention or neo-liberal tool? Critical Social Policy, first published on July 28, 2014 as doi:10.1177/0261018314543224.
Permanent link to archive version: http://hdl.handle.net/10523/5666

The opinions expressed in this post are my own and do not represent the views of my employer, or any association to which I belong.)

References

Keddell, E. (2014) Current debates on variability in child welfare decision-making: a selected literature review, Social Sciences, 3, (1): 916–940; doi:10.3390/socsci3040916 Retrieved from: http://www.mdpi.com/2076-0760/3/4/916

McQuillan, D. (2015). Algorithmic states of exception. European Journal of Cultural Studies, 18(4-5).

Mumpower, J. L. (2010). Disproportionality at the “front end” of the child welfare services system: An analysis of rates of referrals, “hits,” “misses,” and “false alarms”. Journal of Health and Human Services Administration, 33(3), 364-405. doi: 10.2307/25790786

National Ethics Advisory Committee (2013) Sharing Identifiable Health Information for Use in Predictive Risk Modelling Research: Advice to the Minister of Health. Retrieved from: http://neac.health.govt.nz/sharing-identifiable-health-information-use-predictive-risk-modelling-research-advice-minister

Putnam-Hornstein, E., Webster, D., Needell, B., & Magruder, J. (2011). A public health approach to child maltreatment surveillance: Evidence from a data linkage project in the United States. Child Abuse Review, 20(4), 256 -273.

Vaithianathan, R. et al., (2012). Can administrative data be used to identify children at risk of adverse outcomes? Auckland: Business School, Department of Economics, University of Auckland.

Wynd, D. (2013). Child abuse: What role does poverty play? A Child Poverty Action Group Monograph. Auckland: Child Poverty Action Group.

Footnotes

[1] That is the top tenth of risk – 1425 people or 2.5% of the 57 986 babies born over the time period of the study – 48% of this number had a substantiated finding of child abuse.

[2] Interestingly, of all abuse substantiated in the sample, 44% was found in the top two deciles – yep, that’s right, 56% of all abuse in the sample was committed by people in risk deciles 1 – 8.