The Department has a policy emphasis on ensuring that all high school graduates in Tennessee are prepared to pursue a post-secondary credential. To that end, it has introduced a ready graduates metric in accountability which currently captures on-time graduates with an ACT composite score of 21 or higher. In future years, it will also credit students for completing early postsecondary opportunities, earning industry certification, or meeting specified criteria on the AFQT.
In light of the Department's emphasis on ready graduates, I use machine learning to predict ready graduates based on observable information about students as they are in middle school.
This work follows a similar initiative in Wisconsin to create a Dropout Early Warning System.
We use the following information about students in 6th, 7th, 8th grades:
- TCAP Scores
- Disciplinary Issues (Suspensions/Expulsions/Reasons)
- Absenteeism (# days)
- Mobility (# of schools enrolled)
- School aggregates of above
The outcome of interest is on-time graduation (within four years and a summer) with an ACT composite score of 21 or higher. As data become available over time via the state's longitudinal data system, we will also look to predict measures of enrollment and persistence in postsecondary education.
I currently use the following models (available via R's caret package):
gbm(gradient boosting machine)rpart(recursive partitioning/decision tree)rlda(regularized linear discriminant analysis)nnet(neural network)xgbLinear(extreme gradient boosting w/ logistic regression)xgbTree(extreme gradient boosting w/ trees)
For a more detailed look at methods and results, please see my latest presentation.