Is it possible for a computer model to predict with more accuracy than a human caseworker which people are likely to become homeless? Marybeth Shinn, the chair of the Department of Human and Organizational Development at Vanderbilt University, wanted to find out. So, she and fellow researchers followed 11,000 low-income families in New York City, who were seeking services designed to target at-risk people and keep them from becoming homelessness. They tracked which of the families caseworkers decided to assist and which they passed over — based on the caseworkers’ own judgment — and the eventual outcomes for everyone. The Vanderbilt team also devised a computer model to predict the risk of homelessness, based on variables like childhood trauma, past evictions, previous need for public shelter and low education.
When the researchers pitted man versus machine, they found that their computer model would have selected the correct families to help 26% more often and reduced the number of families turned away, who later become homeless, by two-thirds. Shinn and her team aren’t suggesting that their model could, or should, replace caseworkers, but they propose that having the right data could help them more effectively identify those who need help the most. The New York Department of Homeless Services agreed, and will now use the model in its HomeBase homelessness prevention program for families.