Predictive analytics is the use of big data to feed machine learning and statistical models which then predict the likely occurance and character of future events. In humanitarian action, this could mean, for example, combining historical famine records with mobile phone operator data and social media activity, feeding these into an algorithm or statistical model, and using this to predict malnutrition rates among different population groups. These predictions could then be used to improve targeting of nutrition programmes.
The UK’s Institute of Development Studies recently published an excellent mapping of predictive analytics applications in the humanitarian sector. It’s really clearly written, with a short executive summary, and would serve as a great intro for any non-experts wanting to know more. It was published by the Knowledge, Evidence and Learning for Development (K4D) programme, funded by the UK’s Foreign, Commonwealth and Development Office. You can read the full article here.
To my mind, the article triggers a couple of crucial questions for humanitarian evaluators and evaluation managers, which I address below. But first, I’ll give a brief flavour of the article itself…
What can we say about the use of predictive analytics in humanitarian action?
The authors conducted a rapid review of 49 projects worldwide, provided a typology of predictive analytics in digital humanitarianism, and sought to answer some key questions on patterns of current use, ethical risks and future directions in the field.
What they found was really interesting. As an all-too brief summary:
- Predictive analytics is being used for human resources, fundraising, logistics, disaster mitigation, preparedness and response. It’s less clear how predictive analytics is being used for the recovery phase.
- The questions receiving most attention from predictive analytics tools are where a disaster will occur and who it will affect. Questions of what will happen and when are less commonly addressed.
- One of the most commonly stated benefits of predictive analytics is improved response efficiency by saving time and money.
- Of the sectors covered in the sample, disease outbreak (9 projects) and migration (9) were the most common. Followed by conflict (7), disaster risk reduction (6) and food security (4).
- Most applications of predictive analytics complement, rather than replace, pre-existing forecasting approaches.
- Predictive analytics is typically being used by large INGOs and small start-up companies.
- Very little evidence exists of affected populations engaging in the design, implementation or evaluation of prediction systems in humanitarian contexts.
- Some clear risks can already be seen, including: the use of partial, biased or incomplete data will only lead to flawed predictions and can, particularly in the case of social media posts, exacerbate pre-existing marginalisation; the use of expensive computing hardware and skilled data scientists can exclude local actors and community groups from benefitting equally; poorly implemented digitalisation can amplify historic inequalities and community tensions.
Like all aspects of humanitarian action, predictive analytics has positive and negative aspects. Trade-offs will always be made in any humanitarian activity, including the use of predictive analytics. Evaluation practitioners, therefore, must be in a position to assess the success and measure the effects of predictive analytics in humanitarian action.
But it also speaks to the need for the humanitarian community to develop a clearer understanding of what the objectives of predictive analytics are for our sector. What does good look like here? What types of results are we trying to achieve? What types of trade-offs are problematic for principled humanitarian action? What types of trade-offs are acceptable? Normative standards are needed for each of these questions. The work of the OCHA Centre for Humanitarian Data and the Data Science & Ethics Group are instructive. But more work needs to be done to guide evaluators in the assessment of results for predictive analytics in the humanitarian field.
Lastly, perhaps the stand-out sentence in the Institute of Development Studies report is the following:
We did not find any future plans which mentioned the inclusion of affected populations in the design, implementation or evaluation of predictive analytics. This raises the question of accountability and the issue of ‘Whose predictions?’, ‘Whose priorities?’ and ‘Whose interests are being served by predictive analytics in humanitarian aid?’
(Hernandez and Roberts, 2020, p.25. Emphasis added)
Evaluation practitioners can and must carry some of the burden here. If affected populations are not centrally involved in the assessment of predictive analytics, the understanding of their trade-offs, and the measurement of their results; then the humanitarian sector will miss another opportunity to improve accountability towards the people they exist to serve.
Reference:
- Hernandez, K. and Roberts, T. (2020). Predictive Analytics in Humanitarian Action: a preliminary mapping and analysis. K4D Emerging Issues Report 33. Brighton, UK: Institute of Development Studies.