The effects of single-sex compared with coeducational schooling on students’ performance and attitudes: A meta-analysis.

The meta-analytic approach is appropriate — individual studies are too varied to interpret alone, and the mixed-effects model correctly accounts for between-study heterogeneity. The search was thorough (multiple databases, unpublished dissertations, direct contact with researchers), and the outcomes measured are the right ones for the policy question.

Two serious problems limit the design. First, the binary split between "controlled" and "uncontrolled" is too coarse. Grouping 12 randomized experiments with 45 studies using varying statistical adjustments under one label overstates the quality of the evidence base. Second, "single-sex schooling" is not one intervention — it encompasses programs built on opposing philosophies (biological essentialism versus feminist empowerment), and 92% of studies did not report which type they studied. Pooling opposing treatments and finding the average effect is near zero cannot distinguish "nothing works" from "opposing approaches cancel."

The technical execution is competent: Hedges's correction for small-sample bias, Winsorizing of outliers, both weighted and unweighted means reported, moderator analyses restricted to outcomes with sufficient studies and significant heterogeneity.

The weighted and unweighted effect sizes diverge sharply for several outcomes, and the paper handles this inconsistently. For uncontrolled math performance, one study from Trinidad and Tobago (Jackson 2012, with over 218,000 students and effect sizes of 0.65–0.75) drives the weighted mean to 0.57, making the controlled/uncontrolled contrast look dramatic. The unweighted means tell a quieter story: uncontrolled effects of 0.13–0.17 shrink to 0.09–0.10 in controlled studies. Both are reported, but the abstract and conclusion emphasize the dramatic version. For girls' gender stereotyping, the weighted controlled effect is −0.57 (a medium-sized effect favoring single-sex schools), but the unweighted is 0.02 — the paper flags this and urges caution, applying more skepticism to this non-zero result than to the near-zero results it highlights.

The most important missing analysis is of the 12 randomized studies separately. For a paper whose argument hinges on the controlled/uncontrolled distinction, reporting the strongest tier of evidence would have been the definitive test.

The outcomes are measured with appropriate instruments — standardized test scores for performance, validated scales for attitudes and self-concept. The construct problems lie in what is being compared and what the conclusion claims.

The paper's headline finding — that controlled studies show "trivial" effects — uses a threshold of 0.10, sourced from co-author Hyde's prior work. Under the standard Cohen benchmarks (0.20 = small), the same effects would be described as approaching small rather than trivial. At the scale of millions of students, whether an effect of 0.10 matters is a policy judgment the paper treats as settled.

The deeper problem is treatment heterogeneity. Programs driven by beliefs about biological sex differences and programs driven by feminist empowerment could plausibly produce opposite effects on outcomes like gender stereotyping and academic self-concept. Averaging them and finding zero does not establish that neither works. The paper states this limitation in its methods section but does not carry it into its conclusions.

The directional finding is well-supported: across a large literature, the average controlled difference between single-sex and coeducational schooling is small. But the conclusion — that single-sex schooling "provides no benefits" — goes further. Several controlled results are not trivial: girls' general achievement (0.12), girls' gender stereotyping (weighted −0.57), and U.S. girls' educational aspirations (−0.41, favoring coeducation). The "no benefits" framing requires dismissing or downweighting each of these, and the paper applies skepticism unevenly — the stereotyping result gets a caution flag while the near-zero results do not.

The paper also never considers whether some controlled studies might over-control. It lists "small class size" and "highly motivated teachers" among the confounds that inflate uncontrolled estimates, but if single-sex schools reliably produce these features, they are part of the treatment, not distortions. Controlling for them removes real treatment effects along with confounds.

The paper reports its search process, inclusion criteria, and coding procedures in detail. Table 1 lists all 184 studies with their effect sizes and key characteristics. Both weighted and unweighted means are reported throughout, and the authors flag when these diverge.

Against this: no pre-registration, no open analysis code, no reusable dataset deposit, no conflict-of-interest statement. Readers cannot tell whether the 12 randomized studies agree with or diverge from the 45 statistically adjusted ones.

The study pool is broad and the citations are generally accurate. Three issues merit attention. First, the "trivial" threshold that governs interpretation comes from co-author Hyde's prior publication — defensible but loading the interpretive framework with the senior author's theoretical commitments. Second, at least one randomized study in the pool (Park, Behrman & Choi 2013, studying Seoul high schools) found substantial positive effects of single-sex schooling, including higher college attendance — this is included in the data but never discussed, even though it directly contradicts the headline conclusion using the paper's preferred design. Third, the research team's prior public stance against single-sex education (Halpern et al. 2011, Science, co-authored by Hyde) is not disclosed or cited.

There is no financial conflict. Senior author Hyde originated the gender similarities hypothesis, which holds that males and females are more alike than different. A finding that separating the sexes produces large achievement gains would cut against that hypothesis. Finding no benefit is consistent with it. The analytic choices that most affect the conclusion — the threshold for "trivial," the lumping of all controlled studies into one quality tier, the framing of a heterogeneous null as "no benefit" — all run in the direction that commitment would predict. This does not mean the finding is wrong, but the alignment is systematic rather than incidental.