Students Leaving Before They Finish Their Degree

Keeping admitted students enrolled and on track to graduate, by reaching the ones who are drifting off course while there is still time to help, without labelling whole groups of students as likely to fail.

AI is suitable with conditions

Colleges hold years of records on which students persisted and graduated, and models trained on them sort students by risk well, with c-statistics of about 0.9 in research on Virginia's community colleges. It is conditional rather than strong because accurate prediction has not shown up reliably as more graduates. The largest randomised trial of analytics-enabled proactive advising, across 11 universities and more than 10,000 students, found no overall effect on graduation or persistence after six years, with a gain only at Georgia State, the one site that ran it for the full six years. Georgia State's own gains began nearly a decade before its alert system existed, and its alerts are specific triggers, such as registering for a course outside the major or a low grade in a prerequisite, that its analysts derived from years of its own grade data. The fairness record is also poor, with commercial risk scores rating Black students high risk at two to four times the rate of White students at the universities investigated, and research finding that the same models would direct fewer resources to at-risk Black students than to lower-risk White ones. A model can earn a place ranking students for an adviser's outreach when there are advisers with time to act, but specific alerts on the degree map, registration, and finances name problems an adviser can fix, and a score must never decide anything about a student on its own.

Problem Statement

A university admits students it believes can graduate, and a large share of them leave without a degree.

Some leave because the course was not what they wanted, which is not a failure, but many leave over problems the institution could have seen and helped with, such as a failed prerequisite, a wrong course choice, a small unpaid bill, or a term in which they stopped turning up.

The institution has to find those students early enough to help, with a limited number of advisers and tutors, and without treating students as likely failures because of who they are.

Business Context

Georgia State University is the case most often cited.

Ithaka S+R reports that its six-year graduation rate rose from 32% in 2003 to 54% in 2014, while the share of students eligible for Pell grants, a marker of low income, nearly doubled.

The same case study is explicit that no single programme explains the gain, which it attributes to a dozen or more modest ones.

Its advising system, GPS, holds more than 800 alerts, such as registering for a course that does not count towards the student’s major or getting a low grade in a prerequisite for it, and was introduced in 2012 alongside a consolidation of advising that more than doubled the number of advisers.

The case study points out that GPS arrived nearly a decade after the university’s improvement began.

Georgia State also gives small Panther Retention Grants to juniors and seniors who are on track academically but would otherwise be dropped from their classes under a state rule over a small unpaid balance, about 4,200 students in the first years of the programme.

Vendors sell predictive risk scores to hundreds of institutions, and The Markup reported in 2021 that more than 500 universities used the risk algorithms in EAB’s Navigate platform.

In England, the Office for Students sets a minimum continuation rate of 80% and completion rate of 75% for full-time first degree students under its condition B3, which turns retention into a regulatory matter.

Why This Is Hard

The first difficulty is that prediction is not the bottleneck.

Models sort students by risk well, but the largest randomised trial of analytics-enabled proactive advising, MAAPS, found no overall effect on graduation or persistence across 11 universities after six years.

Ithaka S+R, its evaluator, reports that most sites struggled to implement the protocol, with advisers leaving, caseloads climbing well above the intended 150 students each, students not taking up meetings, and early alert systems not ready in the first half of the trial.

The second difficulty is that the model you choose changes who you help.

Researchers using data on more than 330,000 Virginia community college students found that a complex model was only modestly more accurate than a linear regression, yet only 60% of the students the linear model rated high risk were also rated high risk by the complex one.

The third difficulty is fairness, and removing race from the inputs does not solve it.

The same research found accuracy almost unchanged without demographic inputs, and students of colour still flagged more often, because grades and other records carry the same information.

Later work by the team found that the models would leave at-risk Black students with fewer resources than lower-risk White students, partly because the available records predict Black students’ success less well, especially for new students, and that making a model more attuned to race reduced this bias in one of their models but increased it in the other.

The fourth difficulty is that a score is a label, and labels change how people see a student.

The Markup found that, at the universities it examined, Navigate rated Black students high risk at two to four times the rate of White students, showed the rating prominently in advisers’ lists, and offered predicted risk by major, which experts warned could steer students out of demanding fields.

How The Solutions Fit Together

Rules on the degree map, registration, grades in courses that predict success in the major, and balances owed are the foundation, because each alert names a specific, fixable problem and an adviser who owns it.

Engagement screening from the learning management system adds a timely prompt for a personal tutor when a student’s activity falls well below their course’s, without using demographic inputs, though its flags still fall more often on students from disadvantaged backgrounds.

A risk score can order an adviser’s outreach across a large caseload, but only as a supporting tool, built without race, measured for who it misses, and kept off the student’s record.

Course-specific chatbot outreach supports students between meetings with reminders and answers to routine questions, and hands anything personal to a person.

Implementation Considerations

Staff the response before buying the signal.

The advising successes in the evidence pair alerts with advisers who have time to act, and the MAAPS sites with high turnover and caseloads well above the intended 150 saw little effect.

Some MAAPS institutions raised take-up by placing a registration hold until the student met their adviser, which is a design choice with real costs for students and should be decided deliberately.

Record each alert, whether it led to contact, and what happened next, so the institution can tell whether its alerts lead to action and whether the action helps.

Build any model only from what was known at the time, and hold race, ethnicity, and income separately to measure who it flags and misses.

The evidence on whether to include race as an input is mixed, so this playbook keeps it out on legal and trust grounds, as Texas A&M asked its vendor to after The Markup’s reporting, and relies on measurement rather than omission for fairness.

Georgia State’s early alert models do not use race, and in MAAPS its Black students received more of the intervention and showed the largest gains, which Ithaka S+R suggests may partly explain the result.

Evaluate against a fair comparison.

The critique of Purdue’s Course Signals showed how a claim that students who took more courses using the tool were retained at higher rates can arise simply because students who stay longer take more courses.

A randomised or rotating holdout from outreach is the most reliable way to know whether a model’s ranking beats the advisers’ own judgement.

ROI Considerations

The return is counted in students who stay, and for the institution in tuition and funding retained, against the cost of advisers and systems.

Bird and colleagues cite one estimate that predictive analytics cost an institution around $300,000 a year, and their analysis suggests most of the accuracy is available from simpler models.

The larger costs are usually staff, because outreach without capacity does not change outcomes.

Some of the best returns in the evidence are administrative, such as Georgia State’s retention grants, which it started in 2011 after finding that more than 1,000 students on track to graduate were being dropped each semester, some over small balances, where the fix was known and needed no prediction.

Security, Privacy, and Regulation

Student records, financial aid, and any risk score derived from them are personal data, and some of it, such as disability and ethnicity, is special category data under UK and EU law.

In the United States, FERPA allows an institution to share education records with a vendor that performs an institutional function only where the vendor is under the institution’s direct control over how the records are used and maintained.

Under the EU AI Act, Annex III lists as high-risk the AI systems used to determine admission or assign students to institutions, to evaluate learning outcomes including when they steer a student’s learning, to assess the level of education a student can access, and to monitor students during tests.

A dropout risk score is not named in that list, but a model whose output steers a student’s course or programme choices could fall within it, which is a judgement each institution should take advice on.

Regulation (EU) 2026/1744, the Digital Omnibus on AI, set the obligations for Annex III systems to apply from 2 December 2027.

In England, the Office for Students’ condition B3 thresholds make continuation and completion regulated outcomes, which raises the pressure to act on retention but does not make any particular method required.

Where The Evidence Disagrees

Georgia State’s results and MAAPS disagree about whether analytics-enabled advising works in general.

Georgia State reports large gains, and in MAAPS its students in the treatment group graduated at a rate seven percentage points higher than the control group, with Black students fifteen points higher, yet across all 11 universities the effect was not significant, and at one site persistence was five points lower.

Wisconsin’s dropout early warning system for school students shows that a score can sort well and still be a poor label.

The Berkeley study of it found that it sorted students accurately, with an AUC of 0.80, yet The Markup reported that it was wrong nearly three quarters of the time when it predicted a student would not graduate on time, at a threshold set to cast a wide net.

The Markup reported the study as finding no effect on graduation, while the study itself estimated a possible single-digit gain that it could not distinguish from no effect, and warned that targeting by the score could de-prioritise marginalised students.

The same study argues that targeting interventions by the school or community a student is in could be as efficient as individual risk scores, which questions whether individual scores are needed at all.

Research on Georgia State’s advisers also questions the weight put on the score, finding that two out of three advising interventions in the trial were plausibly well targeted using context no model held, which credits human judgement.

Symptoms

First-year retention or continuation falls short of the institution's target, or of a regulator's threshold, and nobody can say which students left or why until the next year's figures arrive.
Advisers see the students who ask for help and the ones already failing, while the students in the middle who are quietly falling behind are not seen until they stop enrolling.
Students are dropped from classes or blocked from registering over small unpaid balances, and some never come back.
Students take courses that do not count towards their major, and discover in their final year that they cannot graduate on time.
Leaders are offered predictive analytics platforms that promise higher retention, while the press reports on risk scores that label Black students as high risk at two to four times the rate of their White peers, and Latino students at substantially higher rates too.

Root causes

Most reasons students leave, such as money, work, family, and not feeling they belong, are not recorded anywhere the institution can see until the student has already gone.
Advising caseloads commonly run to several hundred students per adviser, and even the 300 that advising bodies treat as a benchmark allows about four hours a year per student, so outreach goes to whoever is flagged or asks.
Degree requirements are complex and course choices compound, so a small wrong turn in the first year can add a semester or more.
Administrative rules, such as dropping students with a balance owed, act on students regardless of how close they are to finishing.
Signals that a student is struggling sit in separate systems, such as the learning management system, the registrar, and student accounts, owned by different teams.

Business impact

Students who leave with debt and no degree, which is the worst financial outcome higher education produces for the student.
Lost tuition and, where funding follows enrolment or completion, lost public funding.
Regulatory risk where outcomes are regulated, such as the Office for Students' minimum continuation threshold of 80% for full-time first degree students in England.
Lower published graduation rates, which affect rankings, recruitment, and the institution's reputation with the communities it serves.
Reputational and legal damage when risk scores are found to rate students by race, or are used to steer students out of demanding majors or out of the institution.

Traditional approaches

Mandatory advising appointments at set points, such as before first registration and after a poor first term.
Faculty-raised early alerts for missed classes or low grades, routed to an adviser.
Midterm grade reports and academic probation rules after a low term GPA.
Degree maps that set out the courses for each major term by term, with checks when a student registers off the map.
Small completion or retention grants for students close to finishing who owe a modest balance.
Learning communities, supplemental instruction, and summer bridge programmes for incoming students.

Solution approaches

Each of these is written for this problem specifically, not borrowed from a generic catalogue.

Recommendation

Start with the rules the institution already understands, alerts when a student registers off their degree map, earns a low grade in a course that predicts success in their major, does not register for next term, or is about to be dropped for a small balance, each routed to a named adviser with time to act and a record of what was done. Fix the administrative causes those alerts reveal, such as dropping students over small balances. Add engagement signals from the learning management system as a prompt for a personal tutor, not as a judgement on the student. Use a risk score only to order an adviser's outreach, built without race and measured for who it misses by group, and never show it as a label on the student's record, use it in a conversation about changing major, or use it to decide admission, aid, or whether a student should leave. Advisers, tutors, and faculty make every decision about a student.

Data requirements

Student information system records high sensitivity: Enrolment status, registered courses each term, major and declared degree programme, term and cumulative GPA, credits attempted and earned, and holds, kept as they were each term so a model only learns from what was known at the time.
Degree maps and course requirements low sensitivity: The courses each programme requires, in the order and by the term they should be taken, and the courses whose grades are known to predict success in that programme.
Learning management system activity medium sensitivity: Logins, assignment submissions, quiz attempts, and content views per course and week, and where used, attendance and library or building access, compared with the course cohort rather than an absolute standard.
Student accounts and financial aid high sensitivity: Balances owed, payment plans, aid awarded and its conditions, and holds that would block registration or drop the student from classes.
Advising and intervention records high sensitivity: Each alert raised, who raised it, who it was routed to, whether the student was contacted and met, what was done, and the student's later enrolment, needed to tell whether alerts lead to action and whether action helps.
Protected characteristics for fairness monitoring high sensitivity: Race and ethnicity, sex, disability, first-generation status, and Pell or equivalent low-income status, held separately from any model's inputs and used only to measure who is flagged, contacted, and retained by group.

Evaluation metrics

First-year retention, continuation, and six-year graduation rates, by cohort and by group, against a comparison that accounts for changes in who enrolled.
Share of alerts that led to contact with the student within the agreed time, and share that led to a meeting.
Credits that did not count towards the student's degree, and time to degree.
Number of students dropped or blocked from registering over a balance, and how many of them re-enrolled.
For any risk score, the share of students who left that it had flagged, and the share of flagged students who left, measured by group, and compared with a random or rotating holdout of outreach.

Human oversight

An academic adviser or personal tutor decides whether and how to contact a student and what to recommend, using the alert as a prompt alongside what they know of the student. No change of major, withdrawal, probation, financial decision, or admission decision is ever taken because of a risk score. Faculty decide grades and academic standing. The student success leadership owns any predictive model, reviews who it flags and misses by group at least once a year, and stops it if outreach based on it does not beat outreach without it.

Common pitfalls

Buying a predictive platform without the advisers to act on its alerts, so a list of at-risk students grows while caseloads stay too large to reach them.
Showing a "high risk" label on the student's record, where it shapes every adviser's first impression, as the investigation of EAB Navigate described.
Using predicted risk by major in a conversation about which major a student should take, which experts warned would push Black students out of science and maths.
Assuming that leaving race out of a model removes its bias, when research found accuracy almost unchanged without demographic inputs and students of colour still flagged more often, because GPA and other inputs carry the same information.
Claiming an intervention works because the students who received it more often stayed longer, the flaw critics found in Purdue's Course Signals retention claim.
Identifying students likely to drop out in order to encourage them to leave before they count in retention figures, which the president of Mount St. Mary's University proposed with a freshman survey, in a controversy that ended with his resignation in 2016.

When not to use AI

The institution cannot staff timely outreach, so a better ranked list changes nothing for the student.
The losses come from causes a rule can find exactly, such as courses off the degree map, missing registration, or small balances that trigger a drop.
The score would be used to decide admission, aid, academic standing, or a change of major, or to encourage a student to leave.
The institution cannot measure who the model flags and misses by race, income, and first-generation status, or will not act when the rates diverge.
Most students who leave share an environment, such as a campus, programme, or schedule, which research on a school dropout warning system in Wisconsin suggests can be targeted as efficiently as with individual scores.

Sources

Agent context

A distilled brief for an AI agent or an engineer scoping this work. It carries the constraints and pitfalls that change an implementation decision, not the whole page.

Get the agent context as Markdown

Objective

Raise the share of admitted students who continue and graduate by finding students who are going off track early enough for an adviser or tutor to help, using alerts tied to specific causes first and any risk model only to order outreach, without labelling students by who they are.

Constraints

A risk score orders an adviser's outreach and nothing else. Never use it to decide admission, financial aid, academic standing, probation, a change of major, or to encourage a student to leave.
Never display a risk score or a "high risk" label on the student's record, to faculty, or to the student, and never show predicted risk by major.
Do not use race, ethnicity, sex, income, or first-generation status as model inputs, a policy choice since the research on race inputs is mixed. Hold them separately and use them only to measure who is flagged, contacted, and retained by group, at the cutoff actually used.
Removing demographic inputs does not remove bias, because grades and other records carry it, and the same holds for engagement flags. Measure the flagged and missed rates per group rather than assuming fairness.
Train only on information known at the time of each prediction, and count students who transferred and graduated elsewhere as completers, not dropouts.
Evaluate any score against a randomised or rotating holdout of outreach in the adviser's usual order. Do not claim effect from comparisons where students who stayed longer simply received more of the intervention.
In the United States, education records and anything derived from them fall under FERPA, and a vendor may receive them only under the institution's direct control. In the EU, an AI system that steers a student's learning or assigns them to programmes may be high-risk under Annex III of the AI Act, with obligations applying from 2 December 2027.

Inputs

Student information system records by term, as known at the time, including registration, major, credits, GPA, and holds.
Degree maps per programme with required courses, their recommended terms, and courses whose grades predict success in the major.
Student accounts balances, holds, and the dates a balance would drop a student from classes.
Learning management system activity per course and week, with course cohort averages and the teaching calendar.
Advising and tutor contact records with each alert, who received it, whether the student was contacted, and the outcome.
Continuation and graduation outcomes, including from the National Student Clearinghouse or equivalent, and protected characteristics held apart from model inputs.

Outputs

Specific alerts, each naming its trigger, such as a course off the degree map or a balance due to drop the student, routed to the student's own adviser with a response deadline.
Weekly engagement flags to personal tutors for students far below their course cohort.
Where justified, an ordering of each adviser's caseload for outreach, with no score visible on the student's record.
A termly report of alerts raised, contacted, and resolved, and of flagged and missed rates and continuation by group.

Success criteria

Continuation and graduation rates rise for the cohorts served, measured against a comparison that accounts for changes in who enrolled.
Most alerts lead to contact within the agreed time, and credits taken off the degree map and drops over small balances fall.
Any score-ordered outreach beats the holdout in continuation, or it is retired.
The gap between groups in the share of at-risk students reached does not widen, and no group is flagged at a rate its outcomes do not justify.

Last updated 2026-10-06