How Behavioral Science Improves Student Retention: Interventions for College Persistence and Course Completion

Last updated October 30, 2026

Of the nearly 2.62 million students who started college in the US in fall 2024, 77.1% were enrolled at any college the following fall, according to the National Student Clearinghouse Research Center's 2026 persistence and retention report. That leaves roughly 600,000 first-year students who did not return anywhere.

Behavioral science improves student retention by finding the specific moments where students who intend to stay end up leaving, such as an unfinished loan form or a missed deadline in a gateway course, and redesigning those moments so the next step is easy to take. The interventions with the strongest evidence are timely reminders from the student's own college, defaults that favor full-time enrollment, short exercises that address whether students feel they belong, and personal coaching.

At The Decision Lab, we apply behavioral science across the education system, from how districts choose curricula to how university students reach mental health support. Higher education asks more of students than K-12 does, because students make most of these decisions alone, often with money at stake and nobody checking that they followed through. The research is now specific enough to say which interventions work for which students, and where they stop working once rolled out at scale.

• Text outreach from Georgia State University, timed to specific summer tasks, cut summer melt among students who had committed to enroll by about a fifth compared with a control group. National financial aid campaigns that reached more than 800,000 students had no measurable effect on enrollment.
• In a randomized evaluation, CUNY's ASAP program nearly doubled three-year graduation at three community colleges, from 22% of students in the control group to 40%, by pairing full-time enrollment defaults with close advising and tuition support.
• Chatbot messages tied to one course's deadlines and content raised the share of students earning a B or higher by 4 percentage points, from 61% in the control group, across two large online courses.
• A one-hour belonging exercise cut the grade gap between Black and white students roughly in half over three years, in a small randomized study at one university.
• Personal coaching raised the share of students still enrolled 18 months later from 28.6% to 32.9%, compared with uncoached students, in a randomized trial at eight institutions.
• Interventions most often fail when every student gets the same message, regardless of which barrier they face.

1. Enrollment. Students who have been admitted and have committed to a college fail to arrive in the fall, usually because summer paperwork went unfinished. Researchers call this summer melt.
2. Financial aid. Students leave aid unclaimed or stall on loan counseling and verification, and a hold on their account stops them from registering.
3. Course selection and registration. Students take too few credits or put off required courses, and each choice lengthens the path to a degree.
4. Course completion. Students fall behind in large or online courses, and the first warning sign often arrives after it is too late to recover the grade.
5. Persistence to graduation. Students who doubt they belong, or who have no one checking in, stop enrolling between terms, usually without ever deciding to leave.

In summer 2016, Georgia State University tested an AI text-message assistant called Pounce with 7,489 admitted students, half of them randomly assigned to receive it. Among students who had already committed to Georgia State, those who received Pounce were 3.3 percentage points more likely to enroll on time than the control group, whose on-time enrollment rate was 84.6%. Lindsay Page and Hunter Gehlbach, who ran the randomized trial of Georgia State's AI enrollment assistant, describe that as cutting summer melt by about a fifth. Among admitted students who had not yet committed, the assistant made no measurable difference.

That split should shape how institutions buy and deploy chatbots. Pounce helped students who had already decided to come and were stuck on tasks, and it did little for students still deciding. A follow-up trial with 4,442 incoming students at East Carolina University found no change in enrollment across the full group. Among first-generation students, enrollment rose by 3.4 points over a control rate of 89%, a gain the authors could only weakly distinguish from chance, and they concluded that the technology works best when it is aimed at the students who need it.

The same Georgia State trial measured the summer tasks behind enrollment. Committed students who received Pounce were 6.3 percentage points more likely than the control group to complete the loan counseling required before borrowing, and 4.8 points less likely to have an immunization hold blocking their registration.

Scale changes the picture. Kelli Bird, Benjamin Castleman and colleagues ran financial aid nudging campaigns that reached more than 800,000 students through the Common Application and a large state system, using emails, texts and mailers that encouraged students to file the federal aid application (FAFSA) early. They found no effect on aid received or on college enrollment, overall or for any subgroup. In the Common Application arm, they could rule out enrollment gains larger than half a percentage point. Earlier studies that found large effects had usually been run with a local organization serving a few thousand students, and the authors point to weaker ties between students and a distant sender, along with more generic messages, as likely reasons for the gap.

The City University of New York's Accelerated Study in Associate Programs (ASAP) is the strongest evidence that changing the structure of course-taking changes completion. MDRC randomly assigned low-income students who needed one or two developmental courses at three CUNY community colleges to either ASAP or the colleges' usual services. ASAP required full-time enrollment, grouped first-year students into blocked or linked courses, gave them advisers with small caseloads, and covered the gap between financial aid and tuition. It also handed out free transit cards, conditional on students using key services.

After three years, 40% of ASAP students had earned a degree, compared with 22% of the control group, according to MDRC's evaluation of CUNY ASAP. ASAP students had also earned 9 more credits on average. ASAP is a full program with real money behind it, so it should not be read as a nudge. Several of its parts still work through behavior. Full-time study becomes the default and the first-year schedule is settled in advance, while the transit card turns showing up for advising into something students are rewarded for. MDRC also found that the cost per degree was lower under ASAP than under the colleges' usual services.

A later Georgia State trial moved the chatbot into the classroom. Katharine Meyer, Lindsay Page and colleagues randomly assigned 2,483 students in two large online courses, Introduction to American Government and Principles of Microeconomics, to receive two or three scheduled texts a week about deadlines and course content. In their randomized trial of chatbot outreach in online courses, the share of students earning a B or higher rose by 4 percentage points, from 61% in the control group. Effects on withdrawing from or dropping the course were small and not reliably different from zero.

Broad campus-wide messaging has a weaker record. Philip Oreopoulos and Uros Petronijevic tested online and text-message interventions with nearly 25,000 students at three campuses, including several built around coaching. None of them significantly changed academic outcomes, even for students at higher risk of dropping out, although coaching did improve students' mental health and the time they spent studying. The Georgia State messages were tied to one course's real deadlines and material, and we suspect that specificity explains much of the difference.

Two kinds of intervention have held up for persistence over years rather than weeks. The first addresses whether students feel they belong. In Gregory Walton and Geoffrey Cohen's social-belonging study published in Science, first-year students spent about an hour reading accounts from older students who described worrying about fitting in at first and finding that the worry faded, then wrote their own version. Over the following three years, the gap in grade point average between Black and white students in the study was about half as large among students who did the exercise as among those who did not. The study involved roughly 90 students at one selective university, so the size of the effect is less certain than its direction.

The second is personal coaching. Eric Bettinger and Rachel Baker evaluated a randomized trial of InsideTrack student coaching with 13,555 students at eight institutions. Six months in, coached students were 5.2 percentage points more likely to be enrolled than the control group, whose enrollment rate was 58%. At 18 months, 32.9% of coached students were still enrolled, compared with 28.6% of the control group. At the three four-year colleges with graduation data, 35.2% of coached students completed a degree, compared with 31.2% of students without coaching.

Across these studies, the interventions that moved outcomes came from a student's own institution and arrived when a specific task was due. The campaigns that failed tended to come from a distant sender with general encouragement for everyone on a list. Targeting matters as much as the message: Pounce helped committed students and did nothing for undecided ones, and the East Carolina chatbot showed its only gain among first-generation students.

For a university, this means a retention strategy has to start with which students are stuck, and on what, before anyone writes a message or licenses a tool. An AI assistant that sends the same reminder to every student inherits the weaknesses of the national campaigns, however conversational it sounds. In our own segmentation work for a leading online learning platform, students' attitude toward the product sorted them into clearer and more predictive groups than any demographic did.

When we work on dropout and retention problems, our approach typically involves four steps. Each one below describes what we do and a published project where we have done it, with the results those case studies report.

1. Map where students leave. We pair interviews and surveys with institutional data to trace each decision people make along the way, such as a student's path from admission to graduation, and the barriers at each one. For the Gates Foundation, we used this approach to study why teaching candidates drop out of teacher preparation programs, many of them university-based. Candidates met barriers at every stage from deciding to teach to entering the workforce, and those barriers fell hardest on Black, Latinx and low-income candidates.
2. Group students by the barrier they face. We segment people by attitudes and behavior rather than demographics alone, so each intervention reaches the people it was built for. For a leading online learning platform, we interviewed students from high school through graduate school and surveyed more than 800 of them. Machine learning on the survey data identified six behavioral segments of student learners. The work produced two fully developed product concepts and a dashboard the client still uses to explore its market.
3. Design the intervention at the point of decision. We change the form, message, default or service at the moment someone acts, and leave general awareness campaigns aside. At UC Berkeley, a campus of more than 40,000 students, we helped redesign the university's online student mental health portal so students could find digital support and move more easily to in-person care when they needed it.
4. Predict dropout early and test every change. We build models that flag people at risk of dropping out well before they leave, then test redesigned touchpoints against a comparison group wherever the setting allows. Outside education, we built a model for American Financial Solutions, a nonprofit debt repayment program, that traced program dropout back to the first onboarding call. The barriers it identified explained 80% of dropouts, and the program's dropout rate fell by half once the redesigned counselor calls, check-in texts, mail warnings and marketing materials were in place. The published case study does not describe a control group, so we treat that figure as an operational result rather than a trial estimate.

1. Measure where students leave, by decision point and by group, before choosing an intervention.
2. Send outreach from the student's own institution, timed to a specific task with a specific deadline.
3. Target support at the students facing each barrier, and resist the pull to message everyone.
4. Test every intervention against a comparison group before scaling it, and keep testing after scale-up, since effects often shrink.

How can behavioral science improve student retention?

Behavioral science improves student retention by finding the moments where students who intend to stay drift away, such as an unfinished loan form or a first poor grade, then redesigning those moments. Effective changes include timely, specific reminders from the student's own college, defaults that favor full-time enrollment, belonging exercises and personal coaching.

Do nudges work in higher education?

Some do. Targeted text messages from Georgia State University reduced summer enrollment losses by about a fifth among committed students. National campaigns reaching more than 800,000 students had no effect on aid or enrollment. Results depend on who sends the message, when it arrives, and whether it reaches the students facing that particular task.

Can AI chatbots improve college student persistence?

Randomized trials at Georgia State found that an AI text assistant raised on-time enrollment among committed students and lifted the share earning a B or higher in online courses by 4 percentage points. A trial at East Carolina University found no overall enrollment effect, so chatbots work best when aimed at specific students and tasks.

How do universities increase course completion?

The strongest evidence comes from changing how students take courses. CUNY's ASAP program nearly doubled three-year graduation in a randomized evaluation. It requires full-time study and fixed first-year schedules, and pairs them with close advising, tutoring and help covering tuition. Within individual courses, specific reminders about deadlines and content have raised grades more reliably than general study encouragement.

What is summer melt?

Summer melt describes students who are admitted to college and intend to enroll but never arrive in the fall, often because financial aid steps or registration tasks went unfinished over the summer. It falls hardest on students without family experience of college, and timely outreach from the college itself has reduced it in randomized trials.