Data are increasingly being used for child welfare purposes. Across the country, caseworkers are already relying on large administrative datasets collected through various social service systems to make evidence-based decisions about their clients’ safety, risk, and service planning. The challenge now is to create data-driven tools to ensure all this new information effectively enhances and supports—but does not replace—the real-life expertise and intuition of highly-seasoned caseworkers. One set of promising tools is machine-learning-driven predictive analytics and recommender systems.
What Are Machine-Learning-Driven Predictive Analytics and Recommender Systems?
Let’s start with predictive analytics. In its simplest form, predictive analytics is about using existing information to make predictions about future outcomes, like whether a person’s job application is likely to lead to a job offer, or the chances a consumer shopping online will ultimately purchase the item they’re reading about. Predictive analytics allows researchers to tell sophisticated algorithms a lot of past information, and the algorithms can provide an educated guess about the likely outcomes for a particular future event.
Recommender systems are a type of predictive analytics that provide the researcher a recommended set of actions that optimize the chances of a beneficial outcome. For example, a student could complete a training program to maximize skills identified by the recommender system, or an online retailer could tailor an ad to convince a consumer to make a specific purchase. In fact, many of the recommended products you see while shopping online are the result of recommender systems, which tap into your browsing and buying history to “nudge” you towards a purchase.
Machine learning is one way to make all these predictions and recommendations happen. The term is often used as a catch-all to describe computer algorithms that learn from data and adjust their models over time in response to changing data; they range from simple statistical models to advanced decision tree algorithms, such as gradient boosting.
Applications for Child Welfare
Broadly speaking, these tools can be used in child welfare in two ways: to make predictions (e.g., about the scale of a child’s endangerment), or to make recommendations (e.g., what services and referrals can be provided to parents to maximize child well-being). Predictions of child endangerment can help caseworkers flag the most urgent cases, focusing agencies’ limited time and resources on higher-risk families. Recommender systems can provide a “to-do” list of services and training to help families achieve the best outcomes for parents and children.
There are barriers, of course. Any applications of predictive analytics in child welfare will require large amounts of historical and demographic client data. Better predictions are directly related to the quantity and quality of data, and incorporating advanced machine learning techniques will generally necessitate datasets with more than a million observations.
In practice, it is vital that predictive results don’t outweigh other important information, such as caseworker assessment or ongoing case supervision. Predictive analytics and recommender systems are assistive tools and do not take over the caseworker’s decision making. These tools are additional pieces of the puzzle for a caseworker and can be used to triage support or increase the duration or intensity of services, as opposed to giving a single “yes/no” answer.
Abt Associates Uses Data to Predict Graduation Outcomes
Abt Associates has implemented gradient boosting, a powerful machine-learning algorithm, on data from community college students to predict an individual’s likelihood of graduating, based on readily available contextual information.
Abt used non-identifiable, multi-year administrative data from approximately 18,000 individuals to help make degree predictions (e.g., will the individual have an associate’s degree within five years?), based on personal characteristics and baseline academic assessments. The predictive accuracy of graduation outcomes using this algorithm currently stands at around 93 percent and is expected to improve in accuracy as we are able to gather additional data points and fine-tune the algorithm.
This work can be applied to other predictions or recommendations—including important applications in child welfare—provided there is enough longitudinal administrative data to train the machine-learning algorithm. A key advantage to using these machine-learning-driven predictive analytics and recommender systems is that they are easily adaptable from one topic area to the next. The specific types of variables are not as important as how manyvariables one has. The more data, the better the machine-learning algorithm is able to make connections within the data to make more accurate predictions and recommendations.
Machine learning isn’t the answer to every problem. But it can help the humans using it to think critically about the information at their fingertips and make the hard choices they face every day from a position of strength. Child welfare is ready to harness the potential of data.