Online Charter School Study 2015

158 full-time online (virtual) charter schools in 17 states plus the District of Columbia, 2008-09 through 2012-13. Tested online enrolment grew from about 35,000 students in 2009-10 to over 65,000 in 2012-13 - about 0.5% of students in the data. Panel regressions run on roughly 39 million student-year records for reading and 38 million for maths (online students plus the full comparison pool).

US K-12 public school students, all tested grades (roughly ages 8-18), attending full-time online charter schools - schools whose entire enrolment is full-time online, with a state school ID distinct from any brick-and-mortar school. Blended and part-time online programmes are explicitly EXCLUDED, as are the far larger number of students who take one or two online courses inside a physical school. This is the only source in the cluster measuring what happens when school itself, not a single course, is replaced by software and a home internet connection - and it is squarely inside the archive's 4-18 band.

DESIGN - THIS IS NOT AN EXPERIMENT AND MUST NOT BE READ AS ONE. CREDO's Virtual Control Record (VCR) method builds a synthetic "virtual twin" for each online charter student by averaging records of up to seven students in the schools that actually feed the online charter, matched on grade, gender, race/ethnicity, free/reduced lunch, English-learner status, special education status, and prior test score within 0.1 SD in the same subject. The outcome is year-to-year GROWTH in state-standardised test scores. A second VCR set uses brick-and-mortar CHARTER students as the comparison, to separate the online-ness from the charter-ness. THE SELECTION PROBLEM: families choose full-time online charters, and matching on observables cannot capture why. The report is candid about this - it notes that students may move to online schools precisely because of academic, behavioural or social problems in conventional schools, in which case prior scores are not predictive of future scores regardless of where the student goes. DIRECTION OF THE LIKELY BIAS: the most plausible reasons for choosing a virtual school (illness, bullying, prior school failure, family instability, high mobility, disengagement) are themselves predictors of falling achievement, so the estimate most likely OVERSTATES the harm - the arrow points toward too negative, not too positive. The report partially quantifies this: in the year BEFORE transferring, future online students already had negative growth of -0.06 SD in reading and -0.08 SD in maths. That is a real pre-trend, and it is small next to the -0.10 and -0.25 measured after transfer. Working the other way, some families choose virtual schooling for reasons that would predict better outcomes (a homeschooling philosophy, an athlete or performer's schedule), which would bias toward too positive. On the reported evidence the pre-trend correction is the better-supported one, and it does not come close to closing a -0.25 SD gap in maths. A second, under-discussed threat: online charter students change schools far more often (15% non-structural moves per year even after excluding the move into the online school, against 8% for comparison students), which both disrupts learning and complicates the growth measure. UNITS - THE REPORT GIVES BOTH: every headline is stated as a standard-deviation effect size AND converted to "days of learning" using a flat rule of 0.01 SD = 7.2 days, derived from Hanushek, Peterson and Woessmann's 0.25 SD per school year. The report itself concedes the conversion is approximate and that Hill, Bloom, Black and Lipsey (2008) grade-differentiated growth rates would yield smaller day counts, but keeps the flat rule for consistency with earlier CREDO reports. The archive records the SD estimates as primary and treats "180 days of lost learning in maths" as a restatement of -0.25 SD, not an independent finding. STANDARD ERRORS: the main body reports effect sizes with significance stars and day conversions but no standard errors; SEs appear in the appendix tables (0.01-0.05 for the brick-charter subgroup estimates, 0.002-0.003 for the panel OLS regressions on tens of millions of observations, where statistical significance is guaranteed and only magnitude is informative). FUNDING: Walton Family Foundation, a major charter-school philanthropy - a funder whose priors run the opposite way from the result, which strengthens rather than weakens it. CREDO led the impact analysis; Mathematica Policy Research ran the school-characteristics survey and the Center on Reinventing Public Education the state-policy review, in companion volumes.

Compared with matched virtual twins in traditional district schools, students in full-time online charter schools grew -0.10 SD less in reading and -0.25 SD less in mathematics per year (both p<.01) - among the largest negative estimates anywhere in education research, and the source of the widely quoted "180 days of lost learning in maths". Comparing them instead with brick-and-mortar CHARTER twins gives -0.12 reading and -0.25 maths, so it is the online delivery and not the charter status that does the damage. The estimates are stable across five cohort years and negative in nearly every state, and the specification designed to be most robust to selection - restricted to a student's first year in an online charter and requiring a pre-enrolment observation - gives the LARGEST negative estimates (-0.239 reading, -0.445 maths). The essential caveat is that this is a matched-comparison design, not an experiment - students who choose virtual schools were already losing ground before they transferred (-0.06 reading, -0.08 maths in the prior year), so some unknown part of the gap is selection, and the honest reading is a large negative effect of unknown but almost certainly smaller-than-reported magnitude. Nothing in the report identifies a school model, network or state policy that reverses the sign.

HIGH by the archive's standard, and the reason this is graded C rather than B. Enrolment is a family choice, and the traits that lead a family to withdraw a child into a full-time virtual school are heritable and correlated with achievement growth. Matching on prior score, demographics and programme status attenuates but cannot remove this. The pre-transfer growth trend (-0.06 reading, -0.08 maths) is the report's own measurement of how much confounding is visible in the data, and it accounts for well under half the estimated gap in maths.

The direction and rough magnitude have been reproduced repeatedly. Within this report the estimate survives four analytic approaches (virtual-twin matching against district students, virtual-twin matching against brick-and-mortar charter students, unrestricted panel OLS, and a restricted panel OLS limited to a student's first year in an online charter with a pre-enrolment observation required), with the more selection-robust restricted specification giving the LARGEST negative estimates (-0.239 reading, -0.445 maths). State-level estimates are negative in 16 of 17 states in maths and 15 of 17 in reading. Independent state audits and subsequent virtual-school analyses have found the same pattern, and no credible study finds full-time virtual charter attendance neutral or positive on tested achievement. What is NOT replicated is a randomised estimate - the handful of online charters that hold admissions lotteries are too few and too unrepresentative to generalise from, and the report says so.