Humanitarian agencies sometimes struggle to jump into action when new situations of internal displacement occur. They need to raise funding, move supplies, negotiate access and establish a presence in areas where people seek refuge. Knowing as soon as possible when displacement occurs could help them to provide humanitarian assistance earlier. The same goes for donors and funds that provide the financial resources to aid agencies - like the UN’s Central Emergency Response Fund (CERF) where I’m Head of Programme.
Conflict often leads to displacement of people. Hence, can we use the occurrence of conflict to predict where displacement will occur? During a sabbatical with the Harvard Humanitarian Initiative, I wrote a paper on this research question, which I presented at the Midwest Political Science Association’s annual conference in Chicago in April.
This study shows that it is possible to identify countries with a high risk of seeing large-scale displacement over the next 1 to 3 months. The models in this study could not produce accurate forecasts but further studies – with more disaggregated data, additional variables or more sophisticated methods – may find that the predictive power can be improved. Still, being able to identify a high-risk group of countries could be valuable for humanitarian agencies, funds and donors in responding swiftly to new conflicts and displacement crises. They could better prepare for future displacement, for example by putting contingency stocks in place or increase their presence in a high-risk country. This, in turn, could allow aid agencies to provide better and earlier humanitarian assistance and protection to displaced people and host communities.
For my exploratory study, I used accurate and up-to-date data on displacement and conflict that is currently accessible. The Internal Displacement Monitoring Centre (IDMC), Internal Displacement Updates (IDU), maintains a database of every reported instance of displacement across the globe, known as event data. This database offers preliminary estimates but with a high level of detail, global coverage and updated almost in real time The Armed Conflict Location & Event Data Project (ACLED) and the Uppsala Conflict Data Program (UCDP) provide similar event data on conflicts. I used these datasets together with the INFORM indices, which give an overview of humanitarian situations, for my exploratory study.
However, even with these datasets, it’s not easy to predict displacement. Just looking at the link between conflict (ACLED and Uppsala datasets) and displacement (IDU data), data shows that there are many months when conflict was recorded in some countries, but no displacement was reported. More surprisingly, there were also months when displacement occurred, but no conflict was reported. Those are all the dots on the vertical axis in the chart below, where each point represents one month in one country between 2019 and 2022.
As a next step on my research, I built more sophisticated statistical models – logistic regression models and so-called random forest models which rely on machine learning – with more inputs than just conflict.
These models include measures of the severity of humanitarian crises, such as how many people are the worst affected by crises, and an index that measures a country’s risk to face humanitarian needs, as well as a measure of ethnic fractionalization - all of this data comes from the INFORM indices.
To measure the impact of conflict in a more granular way, I distinguish between different types of conflict and violence, and I look specifically at how many civilians were killed.
With these models, I tested a wide range of different ways to forecast displacement, such as the number of people displaced or whether any displacement occurs in the following month. In the end, I found that the models performed best when I tried to forecast large-scale internal displacement over a relatively short timeframe. I defined this as at least 1,000 people being forced to flee their homes in the 1 to 3 months after this had not happened for at least one month. The models are trained on data from 2019 or 2020 until June 2022 and tested on the second half of 2022 to simulate a real-world application. This is the timeframe for which detailed data are available. The data sources cover all countries globally but to make the test harder, I excluded countries that almost never experience conflict displacement. It's impossible to define this group of countries in a clear-cut way; I ended up excluding the OECD countries except those with a recent history of displacement.
These models have a number of limitations: For example, the data is aggregated to the country and month and could be further disaggregated in future attempts. The data is noisy, with some displacement being recorded when there is no conflict. The number of variables that are included in the model is limited to data sources that are updated in near-real-time. As a result, the models do not produce perfect forecasts. They miss some situations where displacement increased to above 1,000 in the following month and issued false alarms for other cases where this didn’t happen. The table below shows how the best model performs, a logistic model using ACLED conflict data, which is more comprehensive and detailed than the UCDP conflict data, to predict large-scale displacement over a 3-month window.
Correct predictions are in green and wrong predictions in red. Across 138 countries, the model predicts 91% of the cases correctly and it can identify 12 of the 15 cases where large-scale displacement did occur over the following 3 months. But it also incorrectly flags another 16 cases where this did not happen. Thus, the model cannot generate accurate predictions.
At the same time, the model is able to identify a group of countries where the risk of future displacement, at 43%, is 27 times higher than in the baseline group where it is less than 2%. Twenty-seven times is a massively higher risk faced by these countries. Thus, the model can identify a high-risk group of countries, which would then have to be closely monitored to distinguish those that will see large-scale displacement from those that won’t.
By leveraging timely and disaggregated data on emerging crises and enhancing our analytical methods, it would be possible to highlight the risk of future displacements resulting from conflict situations. This information could then be used to support early responses and assistance to internally displaced persons.
I would like to thank my colleagues at the Centre for Humanitarian Data and at the IDMC for their help with the data and analysis for this study, and my colleagues at the Harvard Humanitarian Initiative for their comments and support. The views expressed in this blog post are those of the author and do not necessarily reflect the views of the United Nations.
Note from IDMC:
This exploratory analysis exemplifies the potential of IDMC's Internal Displacement Updates (IDU) data for the prediction of emerging crises in conflict contexts. Each day, IDMC adds data to the IDU new displacement situations from a variety of sources.
Explore the IDU dataset.
For access to the data or inquiries, contact us.
