Using Machine Learning to Support Student Attendance

Eric Ekholm, Ph.D.

Patricia Fox, Ph.D.

Background & Rationale

In a given year, about 12% of CCPS students are chronically absent

Effects of Missing School

Who Misses School?

Why a predictive model?

Although CCPS has several strategies currently in place to combat absenteeism, a predictive model can complement these.

• Most current attendance practices are reactive rather than proactive
• Once students fall into the “habit” of missing school, it may be more difficult to correct
• Providing principals and other school leaders with predictions about which students are likely to be chronically absent in the future can lead to more efficient strategies for students and schools.

Model Development

Overarching Approach

We set out to develop a model that answers the question: “given what I know about a student this year, what is the likelihood they will be chronically absent next year?”

• Pulled historical data from 16-17, 17-18, and 18-19 school years*
• Used machine learning techniques to identify meaningful patterns in data
• Iterated through several approaches to arrive at a final model
• Interpreted results from final model
• Apply learned patterns to new data

​

*excluded 19-20 and 20-21 due to influence of COVID

Data Used

We used several types of data to help our model learn patterns related to student absenteeism

• Student demographics
• Student academic achievement and course-taking
• Student geography (e.g. long/lat of home address)
• Student mobility
• School demographics
• Community demographics (from Census Bureau)

​

Overall, we used 110+ variables and 25,000+ student observations

Model Fitting

We tested several types of models and data transformations before arriving at our final model.

• Tried MARS, random forest, and boosted tree models before arriving at penalized logistic regression
• Power transformed and normalized numeric variables
• Natural splines to transform longitude and latitude variables
• Used SMOTE algorithm to address imbalance in outcome variable (chronic absenteeism)

Model Results

Model Results: Accuracy

Our model correctly predicted chronic absenteeism for 83% of students

• True Negative Rate: 87% (students not chronically absent correctly predicted as not chronically absent)
• True Positive Rate: 53% (students chronically absent correctly predicted as chronically absent)
• Although this may seem low, keep in mind that the base rate of absenteeism is only 12%, and our model does considerably better than that.
• This rate will likely improve in future iterations.

Model Output

Probabilities

​

Yes/No Predictions

School attendance teams receive a list of students who have a high probability (>85%) of being chronically absent.

Takeaways

Takeaways

This prototype improves knowledge of factors contributing to absenteeism and provides actionable student-level predictions.

​

• A variety of factors play a role in student absenteeism, including demographics, geography, achievement, coursetaking, and more.
• Using historical data, we can accurately predict which students are likely to be absent before the first day of school
• Controlling for numerous confounding variables allows for more equitable comparisons

Contact:

eric_ekholm@ccpsnet.net

patricia_fox@ccpsnet.net

​