Using causal inference to achieve better outcomes for refugees
Using causal inference to achieve better outcomes for refugees
Using causal inference to achieve better outcomes for refugees
Using causal inference to achieve better outcomes for refugees

The use of algorithmic decision making by governments is becoming increasingly common. But often, little consideration is given to understanding the knock on effects of the decisions being made.

The use of algorithmic decision making by governments is becoming increasingly common. But often, little consideration is given to understanding the knock on effects of the decisions being made.

The use of algorithmic decision making by governments is becoming increasingly common. But often, little consideration is given to understanding the knock on effects of the decisions being made.

The use of algorithmic decision making by governments is becoming increasingly common. But often, little consideration is given to understanding the knock on effects of the decisions being made.

We were commissioned by a team of researchers at Royal Holloway - part of the University of London - to explore future interfaces for refugee settlement systems with the aim of building a more nuanced and multi- dimensional view of the outcomes of automated decision making - for both refugees and host localities.


How things are done today

We began the project by looking at the current state of the art. In this instance, an application called Annie Moore - named after the first immigrant into the United States to pass through federal immigration.


Screenshot of the "Annie Moore" application


During interviews with the the algorithm designers for Annie Moore, we heard how the system is set up to predict the host locations refugees are most likely to find employment in. To do this, the application factors in information about the refugee, such as levels of education, languages spoken and whether they have any medical conditions, alongside the employment rates of each candidate host location.

These inputs are used to determine an “Expected Employment” score. Essentially, the algorithm is configured to optimise combinations of refugees and hosts to obtain the highest score.



Exploring new possibilities

Whilst it is clear that Annie Moore does help to reduce the amount of time staff spend matching refugees to hosts - and that it likely leads to better life outcomes for refugees, the approach is, by definition, one-dimensional - considering only the likelihood of refugees being employed within 90 days of resettlement.

However, we felt certain that it would be possible to consider a greater variety of factors to produce probabilistic forecasts. With this in mind, we were intrigued by an initial research question:


"What if we could ‘fast forward’ to view potential scenarios for the future ‘worlds’ of these people and the localities that will host them - beyond the simple metrics of today’s systems?"


Intuitively, it made sense to aim for holistic measures of perceived quality of life for refugees and expected 'satisfaction' for the host location – based on various factors such as employment, health and net economic cost. But we were unsure whether these types of outcomes could be represented in an empathetic way within the context of decision support tools.

This challenge is neatly resolved in a self-initiated project that we came across, from LA based design studio, Extraordinary Facility.


Story Shapes, by Extraordinary Facility


Story Shapes was inspired by American author, Kurt Vonnegurt, who set out a theory that the world’s most loved stories can all be plotted on a graph. Whilst this is clearly a reductionist technique, it's one that is intuitive to grasp - an important factor for people with heavy case loads.

When discussing these ideas with the developers of Annie Moore, it appeared that there were no technical reasons why a matching algorithm could not factor in multiple dimensions or be configured to optimise for multiple metrics. The main blocker to this appeared to be a lack of data.

Just as the Annie Moore system calculates the Expected Employment outcome, which in turn becomes a derived data-point for the matching algorithm, it could equally calculate the probability of other outcomes. This would require additional data to be made available and a mathematical model to be developed that would incorporate these new indicators.


A small data, decision intelligence approach

During our research, we were directed towards a 2008 paper called “Understanding Integration: A Conceptual Framework” by Alastair Ager and Alison Strang.

Based on primary fieldwork in settings of refugee settlement in the UK, the paper identifies a number of domains and indicators that are central to perceptions of what constitutes ‘successful’ integration.


Indicators for successful refugee integration, Ager & Strang


The complexity here, of course, lies in the relationships between these indicators. In the absence of statistically significant, longitudinal data, these relationships would need to be estimated. But assuming that the system’s ‘knowledge’ of both the refugee and the host location is enhanced with these indicators, it would be possible (with a higher degree of confidence) to trace a link between the characteristics of a place and the expected experience of a refugee who is settled there.

Research has already been carried out that begins to quantify the relationship between the experiences and activities of refugees and their perceived quality of life. For example, involvement with ethnic communities was found to have a very close relationship to quality of life. If a potential location was known to have a very small population of a their ethnicity, it would be likely that this would have an adverse effect on the refugee’s experience.



By considering a greater variety of indicators and modelling the relationships between them, it becomes possible to build up a picture that describes not only what we might expect life to be like for the refugee, but also attribute this to specific domains or dimensions.

This thinking was the genesis of Voices of Tomorrow, an augmented analytics concept that encodes the output of these calculations into generated narrative outputs that start to bring the future lived experiences of refugees to life.

The concept combines matching algorithms, computational simulations and generative synthetic media, with the aim of building a more nuanced and multi- dimensional view of the outcomes of automated decision making - for both refugees and host localities.

In reality, no one knows for certain whether a match between host and refugee will be successful in terms of the life outcomes or social, political and economic priorities. However, the literature suggests that more multi-dimensional approaches would lead to fairer, more efficient and more humane outcomes. In the absence of the hard data, we believe that approaches based on 'softer' techniques of causal inference can offer greater perspective – and ultimately be more effective than rigid and impersonal algorithms.

Got a project in mind?

To collaborate with us, find out more about our work, or talk to us about a project you have in mind, get in touch.