Single-sex schools, student achievement, and course selection: Evidence from rule-based student assignments in Trinidad and Tobago

The quasi-experimental design represents a genuine advance over prior observational studies. The simulated-assignment instrument is carefully constructed, and the exogeneity tests (38 balance regressions, density tests, preference-selectivity checks) are thorough. Externally graded national exams as outcomes avoid measurement problems from teacher-assigned grades.

Several issues limit the design's reach. The paper separates "preferred school effect" from "single-sex effect" by conditioning on which choice rank a student attained — introducing four additional endogenous variables with no reported test of instrument validity for the expanded system. The residual variation identifying the single-sex effect after this conditioning is never quantified and may be thin, since single-sex schools dominate the preferred-school category. The course-selection finding rests on the untested assumption that all schools offer the same courses — stated as fact but supported by no school-level data. The gender comparison cannot distinguish whether girls respond more to single-sex instruction or whether all-girls schools are simply better institutions than all-boys schools. The "86 percent" null figure counts 40 percent of students who were excluded from analysis entirely (they listed no single-sex schools and the instrument provides no variation for them) as showing null effects, when their effects are simply unestimated.

The identification strategy is well-constructed and the progressive conditioning approach — moving from unconditional to conditional estimates — is the right way to probe what drives the initial positive results. Standard errors are clustered appropriately, and the agreement between the discontinuity and difference-in-differences strategies is reassuring.

The problems are concentrated in the subgroup analysis. The strongest results appear in the smallest subgroup — students listing all four choices as single-sex (3 percent of the sample, just 6,597 students). The analysis runs roughly 48 gender-specific tests across four preference groups and six outcomes with no correction for multiple comparisons. No subgroup-level first-stage diagnostics are reported, so the reader cannot assess whether the null results for weak-preference students reflect true zeros or weak instruments. No power analysis addresses whether the subgroups can reliably detect the effects being tested, which matters when the central conclusion is that most effects are null. The paper does not report how much identifying variation remains after conditioning on choice rank — essential for interpreting the finding that the single-sex effect disappears when you control for attending a preferred school.

The paper's core methodological contribution — demonstrating that selection bias inflates naive comparisons — is well-established. Students who choose single-sex schools score 0.56 standard deviations higher on entrance exams even after accounting for school selectivity, and instrumented estimates shrink to roughly one-third of the naive estimates. This finding is robust across specifications.

The further claims are progressively less well-supported. The finding that the single-sex effect disappears after conditioning on preferred-school status is consistent with the data but also consistent with the model simply running out of variation. The "86 percent null" figure covers students who were never tested. The gender difference cannot be separated from institutional differences between all-girls and all-boys schools. The course-selection finding depends on an unverified assumption about equal course supply. The policy conclusion — that expanding single-sex schools would not help and would not increase female STEM representation — extrapolates from local treatment effects in one Caribbean nation's secondary system to claims about career-level outcomes, with no downstream educational or employment data.

The paper reasons carefully about its main finding: that accounting for selection dramatically reduces the apparent single-sex advantage. But several inferential steps overreach. The "86 percent" claim combines a group with estimated null effects (46 percent) and a group with no estimates at all (40 percent). Treating all-girls and all-boys schools as a single category and attributing the gender difference to differential responses to single-sex instruction ignores that these are separate institutions. The "match quality" interpretation of the preference gradient has simpler alternatives — geography, religion, family tradition — that the paper cannot rule out. The STEM-representation conclusion extrapolates from exam choices at age 16 to career-level outcomes without any intervening data.

The paper describes its identification strategy in detail and discusses many limitations. But the simulation code implementing the school-assignment algorithm — on which every causal estimate depends — is not shared, and the administrative data are not publicly available. The study was not peer-reviewed (it is an NBER working paper). No ethics disclosure addresses the use of identifiable records on minors (students were matched across datasets using name, gender, and date of birth). No power analysis is reported for any specification, including the subgroup comparisons where power is most in question.

The prior empirical literature is presented with its conflicts intact — the paper walks through specific disagreements between Lee and Bryk (1986), Marsh (1989), and LePore and Warren (1997) on the same Catholic school data. The finding that contradicts the paper's theoretical setup (girls taking fewer, not more, science courses) is reported straightforwardly. The decorative neuroscience citations in the introduction are presented without noting that sex-difference brain claims were already contested by 2011, but since the paper never tests these mechanisms, this doesn't affect the conclusions. The key interpretive step — that selectivity explains the single-sex advantage — draws on the author's prior work (Jackson 2010) using the same data system, which is worth noting but not improper.

No funding source is disclosed. The author's affiliations are academic (Northwestern University, NBER). No identifiable industry, advocacy, or policy organization has a financial stake in the findings. The conclusions — discouraging expansion of single-sex schools — cut against the policy trend the paper describes.