Grade retention — holding a child back versus promoting, and test-based promotion policies

Verdict

This topic is the clearest demonstration in the archive of why the grading rules exist. For thirty years the settled professional view was that holding a child back damages achievement and drives dropout. That view came from meta-analyses of studies that made two independent errors, either of which alone manufactures the result:

1. They compared retained children with promoted children who were better students to begin with. Retention is assigned on low achievement, which is substantially heritable and strongly persistent. Allen and colleagues coded 22 studies for design quality and found the answer moves with the coding: −0.30 in the lowest-quality studies, +0.04 in the medium and high — and the +0.04 "is not practically or statistically significantly different from 0."
2. They compared retained children with younger classmates, or with older age-mates, inconsistently and without saying which. Same-grade comparisons credit retention with an extra year of schooling and an extra year of maturation; same-age comparisons penalise it for material not yet covered. Goos and colleagues quantify the choice across 84 design-gated studies: same-grade +0.05, same-age −0.11, F(1,1326) = 23.60, p < .001. Valbuena and colleagues find 34 of 42 credible causal studies use the flattering comparison, and conclude that same-grade results "could be read as the upper-bound of the true effect."

Remove both errors and the achievement effect is g = −0.04 (95% CI −0.11 to 0.03) — nothing.

But the correction does not vindicate retention; it relocates the question. What the regression-discontinuity studies establish is not that holding children back works, but that when you do it, and what you attach to it, determine the sign. Early retention bundled with intensive remediation produces real short-run achievement gains that fade, delays graduation by less than a full year, and leaves attainment roughly untouched in Florida — while retention in grades seven to nine damages attainment, dropout and, in Louisiana, adult violent crime, in every study that looks.

What the evidence shows

The syntheses, ordered by how much design filtering they do

Achievement under regression discontinuity

Attainment, dropout and crime — where the grade decides the sign

Behaviour

The methodological attack on all of it

Hereditarian-lens assessment

Risk: low for the verdict, and the archive's premise is what explains the literature it replaces.

Retention is assigned on low measured achievement. School achievement is roughly 60% heritable and highly persistent, so any comparison of retained with promoted children that does not condition on prior achievement is comparing two populations that differ in exactly the trait the intervention targets. That is not a hypothetical: it is measurable in the data. Schwerdt and colleagues report that "relative to our preferred IV estimates, OLS estimates of the effects of third grade retention are always lower," and that in first-year reading the OLS-IV gap is more than a third of a standard deviation. Eren and colleagues find "non-negligible positive selection biases" in OLS at both grades. Jacob and Lefgren get OLS +0.02 for grade-6 dropout where the instrument gives −0.04.

The meta-analytic literature reaches the same conclusion without needing an instrument, and it reached it in 1989. Holmes's original meta-analysis, the founding document of the anti-retention consensus, reported −0.28 across its 25 best-matched studies — and −0.01 across the five matched on both achievement and socio-economic status, rising to +0.12 with one questionable study removed. The entire negative finding lived in studies that failed to match on prior achievement, and it was visible in the source, in a stratum nobody quoted. Allen and colleagues found the same in 2009 and Goos and colleagues in 2021. Three independent syntheses across thirty years, and the professional consensus did not move.

What the causal designs buy is that at a test-score cutoff, treated and untreated children have the same expected distribution of ability and of alleles. That is why genetic_confound_risk is low despite the topic sitting on top of one of the most confounded literatures in education.

One thing the RD designs do not fix. They identify effects on children near the cutoff. Two papers extrapolate beyond it and they disagree: Opper and Özek find benefits larger for lower-baseline students in Florida, while Eren, Lovenheim and Mocan find crime harms larger away from the cutoff in Louisiana and warn that "one cannot simply move the passing threshold to eliminate the crime effect." Nobody knows the population effect.

Boundaries & what critics say

• The regression-discontinuity studies do not identify retention. They identify a policy bundle, and every author team says so. Florida's treatment is retention plus a summer reading camp plus assignment to a "high-performing" teacher plus ninety uninterrupted minutes of daily reading instruction plus an academic improvement plan. Schwerdt and colleagues state this on page 2. Winters and Greene write that "we cannot completely separate the differential effects of the treatments attached to the policy." Mumma and Winters say their estimates "reflect the overall treatment under the policy, not only retention." Goos and colleagues put a number on the distinction: "retention plus" g = +0.20 (CI 0.04 to 0.35) against "mere rehearsal" g = −0.10 (−0.17 to −0.02), F = 11.57, p < .001. The bundle may be doing all the work.
• Louisiana separates the bundle, and the two halves have opposite signs. Eren and colleagues run two discontinuities on one policy: the March cutoff triggers summer remediation and the July retest triggers retention. Qualifying for summer remediation lowers dropout by 2 to 2.5 percentage points and felony conviction by 0.4; retention itself raises dropout by 5.1 points in grade 4 and 3.3 in grade 8. This is the cleanest evidence in the topic that the remediation helps and the holding-back hurts.
• They may not even identify the bundle-for-the-retained. Berne, Jacob, Weiland and Strunk — with Brian Jacob, who wrote the founding RD study, as second author — show that in Michigan, being flagged for retention raises reading by 0.05 SD while raising actual retention by only 3.4 percentage points, implying "an implausibly large 1.3 SD" effect. Survey data show flagged-but- promoted students also receive intensive support, and similar effects appear in districts that retained nobody. The exclusion restriction fails, and "LATE estimates (but not ITT estimates) in the prior literature might be biased upward." The bias scales as one over the first stage, which gives a clean exposure ranking: Mississippi 5.8pp (severe), Florida 28pp and Texas 35.5pp (moderate), Uruguay ~43pp (small), New York City ~64pp (small, and pre-empted by its authors). Note that Joseph Cimpian made essentially this argument against the Florida anchor in a 2015 NEPC review and was dismissed as advocacy.
• The outcome test is usually the assignment test. Chicago's outcome is the ITBS, which assigned retention; Florida's is the FCAT, which assigned retention. Only three studies break this — Winters and Greene's norm-referenced test (effects survive at 55-70% of magnitude), Mariano and colleagues' Regents exam, and Hwang and Koedel's ISTEP+. For a literature this consequential that is a thin independent-verification base.
• Two grade-3 RD studies reach adulthood and disagree completely. Florida: precisely estimated zeros on graduation and college. Texas: ever graduating −9.1pp and annual earnings at 26 −$3,477 (−19%). Zhong's own reconciliation is the bundle — Texas mandates no post-retention support, gives three test attempts, and retained about 2% of third graders against Florida's 12%, so Texas retainees are far more negatively selected and stigma is plausibly larger.
• Same-grade persistence is not a long-run effect, and every author who reports one says so. Schwerdt: same-grade gains "cannot necessarily be interpreted as a long-run effect of grade retention." Hughes and colleagues attribute their d = 0.99 to curriculum "dosage" — four years of exposure against three. Mariano and colleagues call their Regents result "suggestive evidence at best" and note it "could also reflect an age effect." Mumma and Winters can only do same-grade, because Mississippi's scales are not comparable across grades.
• The one number in the old literature that should be quoted more. Roderick found 69% of retained students dropped out against 27% of the never-retained — the statistic that built the case against retention. In the same study, 58% of students who were overage-for-grade and never retained also dropped out. Almost the whole gap is reproduced by being behind, without any retention having occurred.
• Abolishing retention is not free either. Zhang and Huang exploit China's staggered provincial ban on repetition: exposure in grades 1-3 cut junior-high completion by 6.42 percentage points (3.15 under the Callaway-Sant'Anna estimator) and raised adult property-crime rates by 32.7% of the mean. And the psychological case for abolition did not show up on the outcome that predicts it — depression was unmoved.
• Behaviour is the outcome class where the received wisdom was always weakest, including in its own founding source: Jimerson's meta reported that 86% of behavioural analyses showed no significant difference. The modern designs agree — null in New York after multiple-comparison correction, null in Indiana, and transient in Florida with a fifth to a third attributable to simply being the oldest child in the room.

Practical guidance

• Never retain an adolescent. This is the single most robust causal finding in the topic and it holds in every study spanning both age ranges: grade-8 retention raises dropout 8pp in Chicago and 10.8pp in New York, costs 0.76-0.91 years of schooling in Uruguay, and raises adult violent-crime conviction 58% in Louisiana. The mechanism is time — a retained eighth grader starts high school over 0.4 years late with no remaining opportunity to catch up, while a retained sixth grader starts 0.21 years late and is less likely to be retained again.
• If you retain in early primary, attach the remediation and treat the remediation as the intervention. The meta-analytic difference between bundled and bare retention (+0.20 versus −0.10) is larger than the effect of retention itself, and Louisiana's two-cutoff design shows the summer programme lowering dropout while the retention raises it.
• Budget for a fading achievement gain and a one-year delay. Expect roughly 0.2-0.3 SD in the first three years on an honest same-age comparison, gone within five. Expect graduation about 0.63 years late (Florida) or exactly one year late (Texas). Do not expect a dropout effect in either direction from early retention; the estimates are null and precise.
• Ask what fraction of flagged students you will actually retain. If the answer is under about 10%, you are running a flagging-and-remediation policy, not a retention policy, and you should expect and evaluate it as such. That is also the regime in which the published effect sizes are most inflated.
• Do not quote the old literature. "Retained students are five to eleven times more likely to drop out" is a correlation between being a struggling student and dropping out. The only design-gated estimate for that outcome class is g = −0.10 (95% CI −0.19 to −0.01) — real, and an order of magnitude smaller.
• The counterfactual worth pricing is a second delivery of the curriculum without the lost year. Most of what retention buys is exposure to the same material twice, and the archive already holds cheaper ways to deliver that: high-dosage tutoring, double-dose subject blocks, and mastery pacing.

Open questions

• Confidence is capped at medium by process and by a live methodological dispute. Several independent grade-B sources agree on the age gradient, which would otherwise support high; but status: surveyed means the adversarial pass has not run, and Berne and colleagues' exclusion- restriction result is too recent for the field to have absorbed.
• Nobody has separated retention from remediation outside Louisiana. The single most useful experiment in this topic would randomise flagged children to retention-with-support, promotion-with-support, and promotion-alone. The variation exists in state policy and has never been exploited.
• Why do Florida and Texas disagree about a whole life? Precise zeros on college versus a 19% earnings loss at 26, from the same intervention at the same grade. The bundle explanation is plausible and untested.
• The RD estimates are local, and the two extrapolations point in opposite directions. Until someone reconciles Opper and Özek with Eren, Lovenheim and Mocan, nothing here licenses a statement about children far from the cutoff — which is most of the children a teacher is actually deciding about.
• There is no genetically informative evidence at all. No within-family or discordant-sibling study of retention exists, and given that retention is assigned on a heritable trait, a sibling-fixed-effects analysis would be unusually informative about the non-cutoff population.
• Jimerson's same-age/same-grade split was never obtainable. The single number that would show exactly how much of the received wisdom was the comparison-group artifact is not in any accessible source.