Using Data from Schools and Child Welfare Agencies to Predict Near-Term Academic Risks

School districts want to know how to identify students who may be at risk of academic problems in the near future. A new REL Mid-Atlantic study shows how school and child welfare agency data can be used to identify students at risk for absenteeism, suspensions, course failure, poor grades, and low performance on state tests. By identifying students who are likely to experience these precursors to dropout, educators can target resources for the most at-risk students and intervene before problems become more serious. Using data from two local education agencies in Allegheny County, Pennsylvania, as well as data from the county’s Department of Human Services, the study examines types of predictors that relate to each type of academic problem. Predictive models use these data to identify students who are at risk for academic problems in the coming months. Key findings include: Predictive models using machine learning algorithms identify at-risk students with moderate to high accuracy. Prior academic problems are strongly related to future academic problems. Some out-of-school variables, as reflected in human services data (for example, child welfare system events and justice system involvement), are strongly related to future academic problems. This study provides information to administrators, research offices, and student support offices in local education agencies (LEAs) interested in identifying students who are likely to have near-term academic problems such as absenteeism, suspensions, poor grades, and low performance on state tests. It describes an approach for developing a predictive model and assesses how well the model identifies at-risk students using data from two LEAs in Allegheny County, Pennsylvania. It also examines which types of predictors—including those from school, social services, and justice system data systems—are individually related to each type of near-term academic problem to better understand the causes of why students might be flagged as at risk by the model and how best to support them. The study finds that predictive models which apply machine-learning algorithms to the data are able to identify at-risk students with a moderate to high level of accuracy. Data from schools are the strongest predictors across all outcomes, and predictive performance is not reduced much when excluding social services and justice system predictors and relying exclusively on school data. However, some out-of-school events are individually related to near-term academic problems, including child welfare involvement, emergency homeless services, and juvenile justice system involvement. The models are more accurate in a larger LEA than in a smaller charter network, and they are better at predicting low GPA, course failure, and below basic performance on state assessments in grades 3-8 than they are for chronic absenteeism, suspensions, and below basic performance on end-of-course high-school standardized assessments. Results suggest that many LEAs could apply machine-learning algorithms to existing school data to identify students who are at-risk of near-term academic problems that are known to be precursors to dropout.