The Evaluation of Charter School Impacts: Final Report

2,904 consented applicants; 2,330 in the primary analysis sample (1,400 lottery winners / 930 losers); the Year-1/Year-2 achievement analysis uses 2,150 students across 29 sites. 36 charter middle schools in 15 states.

Applicants to oversubscribed charter middle schools across 15 US states. Study schools were MORE advantaged than charter middle schools nationally (44% vs 62% FRPL-eligible, 47% vs 62% minority, 34% vs 49% below proficient in math), and applicants were more likely to be proficient than their neighbours (73% vs 57% reading).

The national lottery-based evaluation (NCEE 2010-4029) — the only multi-state randomized charter study. An IES report with no DOI; the stable reference is the ERIC full text (ED510573). Only charter middle schools that were oversubscribed and held lotteries could be included, so results do not generalise to charter middle schools as a whole. Benjamini-Hochberg multiplicity correction within each outcome domain. IMPORTANT: the entire school-characteristics analysis (enrollment, ability grouping, urbanicity, revenue per student, hours) is explicitly CORRELATIONAL and exploratory — impacts are randomized, but the moderators are not, and the report says so in terms: "we could not determine whether the school characteristics themselves directly influenced charter school effectiveness."

Nationally, the charter LABEL buys nothing on average — and the point estimates are consistently NEGATIVE (ITT -0.04 to -0.07 SD; TOT -0.06 to -0.09 SD; the report equates this to roughly a quarter-year less instruction), none surviving multiplicity correction, with worst-case attrition bounds running to -0.15/-0.21 and significant. Impacts vary enormously across sites (-0.78 to +0.65 SD in Year 2 math). The variation is strongly associated with intake: schools serving more disadvantaged or lower-achieving pupils show +0.18 and +0.12 SD in math, schools serving advantaged pupils -0.24 and -0.21. Two cautions the sector-average framing usually loses: the favourable half is MATH ONLY (reading is 0.00 for disadvantaged intakes), and the moderator analysis is correlational and exploratory, so "who the school serves and what model it runs" is a hypothesis generated by this study, not a causal finding of it. On resources specifically, revenue per student correlates with math impacts bivariately (p=0.034) but not after controls, and longer hours and lower student-teacher ratios likewise wash out. The one domain that moves large and robustly is satisfaction: parents rating the school excellent +33pp.

Randomized admission; genes cannot differ between lottery winners and losers.

The heterogeneity pattern matches the Massachusetts urban/non-urban split, and the report itself cites Abdulkadiroglu 2009 and Hoxby 2009 as consistent. Note the study's own qualification — urbanicity stops being an influential factor once intake characteristics are controlled, so the replicated regularity is about WHO IS SERVED, not about cities.