What Difference does a Decision Make?
Annie Moore, the first person to be processed by the Ellis Island immigration centre in the US, and whose name is given to the first algorithm designed to help decision-makers assign localities to refuge cases, based upon the likelihood of finding employment within 90 days.
In the United States technology is beginning to play a significant role in decision-making and helping agencies to arrive at improved long-term life outcomes for resettled refugees. In New York refugees have been resettled in a particular location based on the suggestion of an algorithmic tool known as Annie Moore.
Named after the first immigrant registered at Ellis Island in New York, Annie Moore (Matching and Outcome Optimization for Refugee Empowerment) is a matching algorithm used by decision-makers to place resettled refugees in locations where they are most likely to find employment. Despite its proven potential to improve matching and its intention to increase ‘refugee empowerment’, Annie Moore is complicated by a lack of data in two important regards: first, what are refugees’ preferences? While the algorithm is able to utilise a variety of inputs to make recommendations, the actual preferences of refugees to choose the type of environments they prefer (e.g. rural or urban), and second, other important variables that could likely play a role in their choice of or success in livelihoods (e.g. in-demand sectors in certain parts of the US, and housing type) are not currently being considered. Allowing for refugees’ preferences to be built into the algorithm would not entail refugees being offered the option to pick specific locations within which to be resettled, but instead would extend the computational and decision-making process to take into consideration certain basic variables that, if matched, may increase the likelihood that refugees thrive – or, indeed, at the very least stay – in their site of resettlement.
The interface used at HIAS in New York to prepare and upload case data and run matching to find optimised solutions fo each case.
But because these variables have not been inputted into the algorithm, the algorithm quite obviously cannot use them to make recommendations of resettlement location. Some refugee agency employees argue that doing so would be ‘too complicated’, but given the range of algorithmic computing possible, this seems doubtful.
The Annie Moore algorithm is designed to place resettled refugees in locations where they are most likely to find employment. Designed by academics and computer scientists at the University of Oxford, University of Lund, and Worcester Polytechnic, Annie Moore matches refugees to locations through a set of indicators drawn from past employment, nationality, and language data. Increasing the changes of helping refugees find employment quickly after resettlement matters because as one HIAS employee explained,
“Many people think that once a family is resettled, they continue to be supported by the government, but in the United States this simply isn’t true. Refugees are expected to obtain employment very quickly and start supporting themselves. This technology has been key to helping our regional offices connect relatively straightforward resettlement cases with new homes and communities where they are more likely to thrive in their jobs.” (Quote from HIAS, 2018).
Instead, based on the work of PPA, including extensive interviews, the decision not to give refugees a say in the matter of their own new home location seems more rooted in immigration politics of ‘deservingness’ (as in, refugees are lucky to be given a home in the US at all) and veiled paternalism than of ease, as well as a rigid bureaucratic system of resettlement that is reliant on the wide dispersal of refugees to cities, towns, and areas they have likely never heard of and which may well not have the characteristics they would choose for themselves.
Case tiles can be moved by dragging to alternate tiles, whereupon the match scores dynamically update. The background of the case tike changes, at right, to show that this is not an optimised state. Image adapted from Trapp, A. C., and Teytelboym, A., and Martinello, A., and Andersson, T., and Ahani, N., 2018. Placement Optimization in Refugee Resettlement, Working Papers 2018:23, Lund University, Department of Economics, revised 20 Mar 2020.
While the lack of data about refugees’ preferences within Annie Moore has practical as well as ethical implications, there is a wider data gap that calls into question the effectiveness of the algorithm in overall terms: what data illustrates ‘successful matching’? The US is notorious for a lack of long-term resettlement data; while fairly comprehensive data exists for the first 90 days of refugees’ arrival, there is no national (or in cases even regional or municipal) data collection after that point. This means that Annie Moore is limited by datasets and is built on outcome data at the 90-day mark. While some other data exists through ongoing case management programmes, it is not comprehensive and is not currently used by the algorithm. Data collected by the Bureau of Population, Refugees, and Migration (PRM), the humanitarian bureau of the State Department, is notoriously difficult to obtain.
The Annie Moore matching algorithm is aimed at making placement recommendations to enhance the chances of finding employment within 90 days of arrival. What if such algorithms were able to take a much wider selection of data into consideration?