K-12 Predictive Analytics: Time for a Better Dropout Diagnosis
Real Clear Education

By: Dr. Joel Boyd
Original post from Real Clear Education

Imagine going to your family doctor with a cough, fever and chills. You explain that you’ve had symptoms for a few days. Most of us already know what questions the doctor will ask, and with good reason. Doctors are trained in differential diagnosis. They don’t hear a cough and assume that you’ve come down with pneumonia. They look for the simplest, most common explanation to a problem first and utilize blood work, X-rays or MRIs to guide them in the search for rare or complicated explanations.

Unfortunately, school districts generally have far fewer tools to identify students at risk of dropping out. Unlike blood work or X-rays, K-12 early warning systems typically rely on as few as four lagging indicators  —considered in isolation— to diagnose problems. The risk piles up before teachers and administrators are called to action. Sadly, most states and districts use this very same approach to allocate and target precious resources.

Pioneered in the late nineties, so-called “threshold” or “static” early warning systems, which identify correlations in nationally aggregated data to identify cut points for risk indicators (using data like grades or attendance rates), are not without utility. But they tend toward over-identification of high school students because they rely on data for what brings students off track later in their academic careers. They also under-identify at-risk elementary students, causing educators to miss the opportunity to intervene early. Research suggests that static models can achieve about 50 percent accuracy in identifying students at risk before eighth grade. But a one-in-two chance of identifying the right students means that far too many students can fall through the cracks.

It doesn’t have to be this way. Real-time analysis of disparate data streams now enables the beneficial application of predictive analytics for everything from credit card fraud detection to personalized health care. And, as it turns out, the “digital symptoms” required to identify students at risk already exist in most districts. Districts can access and use that data to inform decision-making. And the advent of lower-cost technologies for data integration means that districts can apply advanced statistical methods and predictive analytics to direct resources toward students in earlier grades and bring them back on track with minimal interventions relative to the extensive efforts required to assist students in middle and high school.

Predictive analytics, which are increasingly used to support college retention, are fundamentally different from static early warning systems in that they move beyond national-level correlations, and use a district’s historical data, along with complex algorithms, machine learning techniques and current student indicators to forecast the likelihood that a student will go off track at some point in their academic careers. Unlike static models, predictive models “learn” over time and can identify risk indicators for sub-populations of students by grade level. The approach not only provides regular and on-going data to school staff to identify at-risk students early but also helps pinpoint the reasons students might be off track.

The deeper data is essential to making smart decisions about how to allocate limited resources. If educators notice a child is struggling with reading comprehension, for example, they can intervene with targeted supports early in a school year rather than relying on intensive remediation of a student who has been off track for a year or more. With more accurate identification, school and district leaders can make better decisions about how to direct resources that are likely to help more students do better.

A predictive analytics model may be able to identify with much more accuracy —about 94 percent in some cases—the students who are truly at risk. As school districts better diagnose challenges, more students have an equitable shot at building the skills they need to succeed in school and beyond. Our old approach to early warning served its purpose, but newer models can get us where we need to go more efficiently. It’s time to retire threshold approaches and open the door to predictive analytics. Our children deserve it.