You may have heard about predictive analytics – but what does this mean for those of us who are shaping or delivering services for vulnerable people?
Let’s start with some definitions:
- The term ‘predictive analytics’ is used for assessing large quantities of information to see if there are trends. Those indicators can then be used to identify more families who might be at risk.
- Big data means the huge amount of data that can now be searched by computers – with the potential to add new data sources such as from surveys.
- We talk about risk and protective factors which are identified by predictive analytics. For example risk factors for a child in care becoming homeless include having a relative in prison. Protective factors would include good educational outcomes.
- And you’ll have heard the term ‘machine learning algorithms’ or ‘artificial intelligence’. This means the apps used to trawl through the data don’t need to be told exactly what to look for, they can learn about connections between multiple indicators.
So why does this help us? For me it represents effectively designed services for vulnerable people – it’s not beyond the realms of possibility that one day we’ll know exactly what services or interventions work, who needs our help and how to support them earlier.
The science of predictive analytics being applied to people services is brand new, we will discuss later some of the ethical issues that emerge, and also tantalising results that might help us to offer better services. In Troubled Families world, our new Earned Autonomy pilots are among local authorities and partners who are pioneering predictive analytics — developing both the intelligence and new interventions.
There are a few things that improvements in predictive analytics promise, that we don’t get through conventional techniques:
- New services – with the right data we can see what combination of factors (for example environment, family, income, attainment, history) have the biggest impact on outcomes and the likelihood of needs escalating. Using these risk factors and protective factors we can change services to focus on the most impactful – a commissioning redesign.
- Who to help – predictive analytics is personal. Not only do we get a commissioning design but also information about children or families who are likely to need help, enabling professionals to target their limited time in the most effective way.
- Intelligence – up to now we have relied on methods such as the joint strategic needs analysis (JSNA) to inform local strategies. But this data can be two or three years out of date and more often than not shows an escalating need in a geographic area that we may not be able to resource. Predictive analytics will, over time, give us more immediate, local and targeted management information.
- Time – the most important aspect of predictive analytics might be the time it gives us to respond to need. Most complex need in society can be hidden until it escalates above a threshold and triggers an intervention. If instead we could predict need, then we might have two years to respond and offer more cost-effective support to families, whilst reducing demand to our more expensive services.
This can feel a bit academic without hard examples —here are some local and international approaches that have interested me over the years:
- The Behavioural Insight Team recently tested whether computer analysis of social care case notes might identify children that will be re-referred after the case is closed. The team was able to spot half of these cases with a low level of errors. Here the implications are tantalising, both for better safeguarding and more efficient interventions.
- In Auckland using health and care data, researchers were able to identify children most at risk of maltreatment by five years old. The accuracy was 50% for the top decile most at risk. What’s remarkable is that the data was for children under two, opening the potential to target support to families much earlier.
- In a shire authority early analysis of youth offending needs shows that some young people are exposed to higher risk factors than others. It’s therefore possible to tweak the time spent with specific young people based on who will benefit most from help.
- Durham police are using artificial intelligence to predict the risk of future offending. The algorithms are trained on data from 2008 to 2013, with 88% accuracy in predicting high-risk cases, and 98% accurate for low-risk cases. This is used to inform bail applications.
- New York Fire Service have identified 60 risk factors to show which buildings are most at risk of fire and to target inspections. There are similar projects in the UK to identify schools or GP surgeries which may benefit from an earlier inspection.
- And to show the longer-term potential, Google have detected breast cancer from pathology images using artificial intelligence trained on high resolution scans. The best professionals have an accuracy of 73%, Google has an accuracy of 92%.
There are however serious concerns about this future digital world – particularly the ethics of using families’ data – and the worry that without understanding the implications we are opening Pandora’s Box:
- Ethics — the first concern is about aggregating data from different government databases. Will this significantly improve our picture of families until it’s intrusive, and if people knew would they try to stop it? This is about ethical collection and use of data and is, to an extent, addressed by the forthcoming General Data Protection Regulation (GDPR) that requires explicit consent about the use of data and who it will be shared with. I think there’s more we need to do to test the ethics and what is socially acceptable and supported by citizens in this new age, for example setting up Ethics Committees with local residents considering the risks and benefits.
- Accuracy — professionals are also concerned about the accuracy of the analytics. Will the algorithms used to detect future needs be accurate enough to be useful, and can a computer really capture the complexity and emotion of family life? These are the key questions being tested by other public sectors and it’s fair to say the application of predictive analytics is immature. Early evaluations and international evidence are promising, but we’ve yet to see this applied to the point of making a substantial difference to families’ lives. Academics talk about analysis that underpins professional decision making, but algorithms will only ever give a likelihood of a need, not a certainty.
- Inequality — then there’s the issue that algorithms may make our services less equal. Lazy analytics could single out poverty, race or disabilities and treat these individuals unfairly, for example confusing being poor with poor parenting. And statistics which flag the disadvantaged, simply because we have more data about these families. I think there are genuine challenges here; as a sector how will we safeguard emerging models?However, accepting the very real challenges to predictive analytics that we must address, there are also elements of protectionism that creep into the debate. It’s clear that public sector roles (in general) will change because of new technology – I hope for our sector it’s about making more time to build relationships with families, and better supporting more people who need help.We recently invited proposals for earned autonomy — offering up-front funding in exchange for an ambitious transformation plan. In most of these proposals there is a cutting-edge project to develop predictive analytics — potentially being able to support families earlier and make the limited family support staffing go further.
This all sounds exciting but there are also practical difficulties that the earned autonomy pilots will work on. Such as integrating the right data sets; developing the hypotheses and cohorts to test; skilling up intelligence teams and public health; creating learning algorithms to identify risk and protective factors; finding the right way to engage families and offer early help; changing the way teams work; engaging universal services or developing new markets of lower cost (non-case holding) support; and evaluating the impact on families.
But despite challenges the prize is significant: sustaining the impact of Troubled Families post-2020 and continuing to improve families’ lives.
Detecting Cancer Metastases on Gigapixel Pathology Images https://arxiv.org/abs/1703.02442
Durham Police http://www.bbc.co.uk/news/technology-39857645
University of Auckland https://www.msd.govt.nz/documents/about-msd-and-our-work/publications-resources/research/vulnerable-children/auckland-university-can-administrative-data-be-used-to-identify-children-at-risk-of-adverse-outcome.pdf
Big data in the big apple https://policyexchange.org.uk/wp-content/uploads/2016/09/big-data-in-the-big-apple.pdf
First published by MHCLG: https://supportingfamilies.blog.gov.uk/2018/05/14/predictive-analytics/