The effects of test-based retention on student outcomes over time: Regression discontinuity evidence from Florida

about 75,000 students in the pooled discontinuity sample (9,981-15,687 per cohort); six cohorts entering grade 3 in 2002-03 through 2007-08, followed to 2013-14

Florida public school third graders scoring at or near the FCAT reading Level 1/Level 2 cutoff; graduation and college outcomes use the 2003 and 2004 cohorts (n = 27,724 and 15,687)

Journal of Public Economics 152, 154-169; full text read from the April 2017 revision of NBER working paper 21509. Fuzzy regression discontinuity with two-stage least squares on Florida's third-grade reading cutoff, 10-point Imbens-Kalyanaraman bandwidth. First stage 0.28 pooled (range 0.20-0.37 by cohort, F > 100). Retention among Florida third graders jumped from 2.8% to 13.5% when the policy began in 2003 and fell back to 5.6% by 2008. THE TREATMENT IS A BUNDLE AND THE AUTHORS SAY SO ON PAGE 2: the policy requires a summer reading programme, assignment to a 'high-performing' teacher, and intensive reading intervention including ninety uninterrupted minutes daily, so 'our estimates of the policy's impact will therefore capture the combined effect of retention and these additional measures'. The outcome test is the FCAT, the same instrument family used to assign retention. The paper reports BOTH same-grade and same-age comparisons, which is why it is the anchor for this topic - most of the literature reports only one.

The trajectory is the finding. Averaged over the first three years, retained students outperform their SAME-AGE peers by 31% of a standard deviation in reading and 23% in maths - then 'these initial benefits fade out... statistically insignificant in both subjects within five years', and maths turns significantly NEGATIVE by years 5 and 6 (-0.075 SD and, on the rescaled metric, -0.107 SD). On SAME-GRADE comparisons the same students look enormously better for a decade - 0.73 SD in grade 4 reading, still 0.19 SD in grade 8 and 0.22 SD on the grade-10 FCAT 2.0 - and the authors themselves refuse the reading: this 'is likely due to the effects of age and schooling and cannot necessarily be interpreted as a long-run effect of grade retention'. ON ATTAINMENT THE ANSWER IS A PRECISE ZERO, which is the single most important corrective to the received wisdom that retention causes dropout: high school entry -0.006 (0.020), any diploma -0.003 (0.036), regular diploma among graduates +0.005, GED +0.006, highest grade completed -0.116 (0.121) with a standard error small enough 'to rule out negative effects as small as one third of a grade'. College enrolment is likewise null on every measure. What retention does buy is TIME: graduation is delayed by 0.628 years, less than a full year, because third-grade retention substitutes for later retention (repeating a grade two years later falls 11 percentage points) and five years on retained students are only 0.73 grade levels behind. Middle-school course-taking improves - GPA +0.067 (12% of an SD) and remedial courses -1.61 against a mean of 5.35 - while college-prep course-taking is null. Differential later retention explains at most 38% of the reading fadeout and 25% of the maths fadeout, so the fade is real, not an artefact of the comparison group catching up by being held back themselves.

Low within the design and this is the paper's reason for existing. Retention is assigned on a test score, so comparing retained with promoted children anywhere but at the cutoff compares children who differ in the heritable trait the policy selects on. The authors report the size of that contamination directly: OLS estimates are always lower than the instrumented ones, and in first-year reading the OLS-IV gap is more than a third of a standard deviation. At the cutoff, treated and untreated students have the same expected distribution of ability and of alleles.

Mixed, and the disagreements are informative rather than noise. The short-run positive achievement effect replicates in Chicago, Indiana, Florida's English learners and Texas. The long-run null on attainment does NOT replicate in Texas, where Zhong finds third-grade retention lowers ever graduating by 9.1 percentage points and annual earnings at 26 by $3,477 (19%) - a difference Zhong attributes to Texas retaining about 2% of third graders with three test attempts and no mandated post-retention support, against Florida's 12% with a full remediation bundle. Two methodological attacks matter. Cimpian's 2015 NEPC review argued the exclusion restriction fails because students below the cutoff receive extra services whether or not they are retained, so scaling the reduced form by the first stage inflates the estimate; the review was widely treated as advocacy. A decade later Berne, Jacob, Weiland and Strunk published the same objection with survey evidence, naming this literature and concluding that 'LATE estimates (but not ITT estimates) in the prior literature might be biased upward'. With a 28-percentage-point first stage the inflation factor here is about 3.6, which is moderate by the standards of this literature but not negligible.