Why Are Single-Sex Schools Successful?

The randomization is exceptional — mandatory, computerized, enforced for three decades, with documented near-zero dropout rates. For the first question (do single-sex schools outscore coed schools?), this is one of the cleanest designs available anywhere. For the second question (what happens when a school converts?), the difference-in-differences approach with school fixed effects is standard, with event-study graphs showing no pre-trends for boys.

The design for the third question — separating classroom-level from school-level effects — is the paper's most creative contribution and its weakest link. It compares adjacent cohorts in converting schools: one had same-sex classmates but attended a newly coed school, the other had mixed-gender classmates. These cohorts overlapped in the same building for two of three years, so the paper assumes school-wide effects hit both equally. But the cohorts had different doses of school-level exposure (two years versus three), so some of the apparent class-level effect could reflect extra school-level exposure. The event study for girls partially mitigates this — school-level exposure alone shows no effect — but the mitigation is indirect. The entire girls' result rests on four schools, and the boys' null result rests on seven, which may lack power to detect a moderate effect.

The econometric framework is correctly specified and the identification logic is clear. Effect sizes are plausible and reported in interpretable units (fractions of a standard deviation). The between-school estimates replicate prior work.

The problem is inference. With four treated clusters for girls, standard cluster-robust standard errors reject the null far more often than the nominal rate. The paper reports girls' language results at high confidence but does not use wild cluster bootstrap or randomization inference — the standard tools for this situation. The leave-one-out analysis, while showing no single school drives the result, leaves three treated clusters after each exclusion, which provides no basis for formal inference. No power analysis is reported for the boys' null finding, so we cannot tell whether the study could have detected a meaningful class-level effect for boys. Multiple comparisons across three subjects, two genders, and three specifications are not addressed.

The between-school finding is well-measured: a national standardized exam, externally administered, taken by 96 percent of students. The paper explicitly states that this composite advantage vanishes with observable controls and that gender type is confounded with other school characteristics.

The conversion estimate measures what conversion does — but conversion changes many things at once (classroom gender mix, teacher hiring, scheduling, institutional culture), so calling it a gender-composition effect is a stretch the paper sometimes makes. The class-level estimate maps conceptually to what changed (classroom gender composition), but the statistical foundation is too thin for the claimed precision. The policy conclusion — that gender type is a lever policymakers can pull — extends beyond what eleven self-selected schools in one city under unique institutional conditions can establish.

The paper is more careful about its limitations than most work in this area. It warns against interpreting between-school estimates as peer-composition effects, acknowledges the few-switcher problem, and hedges its mechanism claims. The three-way decomposition (between-school, within-school, class-level) is intellectually clear.

Where the reasoning stretches: the title and abstract foreground the between-school advantage without immediate qualification, even though Table 3 shows it is inseparable from other school characteristics. The policy conclusion that gender type is "replicable" goes beyond what eleven schools in Seoul's unusual context can demonstrate. The language-versus-math pattern (girls harmed in languages but not math) contradicts the stereotype-threat explanation the paper discusses, and no satisfying alternative is offered. The "distraction" interpretation for boys is speculative — the event-study pattern is equally consistent with organizational disruption from the conversion process.

The paper is intellectually transparent about what its design can and cannot identify — the careful distinction between three causal parameters, the honest treatment of the between-school finding, and the clear statement of identifying assumptions are strengths. Standard open-science practices are absent: no public data, no analysis code, no pre-registration, no funding or conflict-of-interest disclosure, no ethics statement. The limitations discussion covers specific methodological issues (optional math exam, non-random school selection) but does not address whether inference is reliable with four treated clusters.

The empirical findings stand on the paper's design, and the key predecessor (Park et al. 2013) is accurately represented. Mechanism citations are offered as interpretive rather than conclusive. Two widely cited contrary contributions are absent: a meta-analysis by Pahlke, Hyde, and Allison (2014) finding negligible advantages of single-sex schooling, and a commentary by Halpern et al. (2011) in Science arguing the case was unsupported. Their omission leaves the paper's policy framing without the context a reader would need to assess generalizability.

No identifiable conflicts. The authors are independent academics at UCL and Korea University. No funding source is declared. The government that provided data access had actively promoted coeducation — the paper's findings run against that policy's premise.