Virtual Classrooms: How Online College Courses Affect Student Success

2,323,023 student-by-course observations for the course-grade estimates (2,360,645 for persistence, 1,974,831 for next-term GPA); drawn from more than 230,000 students in over 168,000 sections of 750 courses, Spring 2009 - Fall 2013. Professor-variance sub-sample - approximately 78,000 students and 5,000 professors, Spring 2012 - Fall 2013.

POST-SECONDARY ADULTS, not school-age children. Undergraduates at DeVry University, a large US for-profit institution (~130,000 undergraduates, 102 physical campuses); 80% seeking a bachelor's degree, mostly in business, technology or health. Mean age ~31 (33.0 online vs 28.4 in-person), 44.5% female, 25.5% Black, 18.7% Hispanic. Two-thirds of all DeVry courses are taken online. The archive's topic covers ages 4-18, so every estimate here must be discounted for age transfer - these are working adults choosing course formats, not children.

WHAT "ONLINE" MEANS HERE: asynchronous discussion-board sections in which the professor's lecturing is largely replaced by standardised pre-recorded video, and all interaction - lecturing, discussion, group projects - happens in online discussion boards. It is NOT live synchronous video and NOT self-paced software with no instructor. Everything else is held constant by institutional design: the same syllabus, textbook, class size, assignments, quizzes, tests and grading rubrics are used in the online and in-person sections of each course, and in many courses the quizzes and tests are standardised across sections. This is the cleanest active-alternative contrast in the literature - the only thing that varies is the mode of communication. IDENTIFICATION: students self-sort, so OLS is not causal. The authors instrument for taking a course online with (i) an indicator that the course was offered in person at the student's nearest campus in the current, prior or following term, (ii) home-to-campus distance, and (iii) their preferred instrument, the interaction of the two (Card 1995), with only the interaction excluded. All specifications carry course, term and home-campus fixed effects plus prior cumulative GPA, gender and age; standard errors cluster on section. First-stage F ranges 19-193. Compliers are 27.4% of the sample, always-online 47.8%, always-in-person 24.8%, so this is a LATE for students who genuinely had both options - the paper says explicitly that for students with no campus nearby, online may be the only route to any college at all, and no welfare conclusion follows. OLS and IV are close on grades (-0.380 vs -0.442), so selection into online is not doing the work. MEASUREMENT: the course-grade outcome is teacher-assigned, and the authors take grading discretion seriously - DeVry constrains it with common rubrics and shared tests, and the persistence outcomes are administrative records untouched by professor judgment. That combination is the paper's answer to "maybe professors just grade online sections harder". FUNDING: IES grant R305B090016 to Stanford. DeVry supplied the data and named staff support but did not run or author the evaluation - independent under the archive's rule. READ NOTE: the published AER article is paywalled; the numbers below were read from the full text of the identical study circulated as CEPA Working Paper No. 15-10 (October 2015), titled "Changing Distributions: How Online College Classes Alter Student and Professor Performance". Its abstract and the published AER abstract state the same result set.

Taking a course online rather than in person lowers the grade in that course by about 0.44 grade points on a 0-4 scale (SE 0.023), roughly a quarter to a third of a standard deviation given the 1.33-point grade SD. The result that matters is what happens next - the same student's GPA in the FOLLOWING term falls by 0.17 grade points, and the probability of still being enrolled a year later falls by about 10.6 percentage points against a base of 68.6%. A grading artefact cannot travel forward into other, mostly in-person courses or into re-enrolment records, so the loss is in learning and momentum, not in marking. The harm is concentrated on the weakest students - the online penalty shrinks by 0.123 grade points for each additional point of prior GPA - and online sections also widen the spread of outcomes while narrowing the spread of professor effects, which is exactly the trade the technology promises and exactly the wrong side of it for students who most need a good teacher.

Not randomised, but heritable ability cannot drive the estimate - identification comes from where and when the university happened to schedule in-person sections, and prior cumulative GPA plus course, term and campus fixed effects absorb student ability directly. The residual threat is institutional, not genetic - campuses might schedule in-person sections in response to local demand from stronger students - and the preferred distance-interacted instrument is the response to it.

The direction is one of the better-replicated results in ed-tech. Randomised trials point the same way (Figlio, Rush & Yin 2013; Kofoed et al. 2024, -0.215 SD at West Point), and the paper's own literature review lists four independent quasi-experimental studies of online course-taking in community colleges finding 7-8.4 percentage-point drops in completion and lower grades (Xu & Jaggars 2013, 2014; Hart, Friedmann & Hill 2014; Streich 2014). The only non-negative contemporary is Bowen et al. (2014), which compared HYBRID (not fully online) with face-to-face and found no difference - a different treatment. What is NOT replicated is the persistence result specifically - no other study has followed the same students into subsequent in-person courses, which is what makes this paper irreplaceable.