Home-Schooling in the US

6,033 self-identified homeschooled SAT takers in 2001, against 975,117 public, 54,682 private-independent and 137,671 private-religious school test-takers (ETS data, foreign nationals and missing income excluded, ages 14-24)

US SAT test-takers in 2001 — college-aspirant students only. Homeschoolers were 0.5% of test-takers against roughly 1.5% of the school-age population, so this cohort is the college-bound top slice of home education.

NCSPE Occasional Paper 88 (Teachers College, Columbia). A policy review with an original cross-sectional analysis of the SAT data — the SAME 2001 College Board homeschool cohort that Ray & Eagleson 2008 analysed, but at the INDIVIDUAL level and with covariates, which makes the pair a natural comparison of what an independent analyst and an advocacy institute do with one dataset. Belfield fits OLS on SAT scores with mother's and father's education (6 categories each), gender, grade level, disability, first language, ethnicity, citizenship, religion (10 categories), state-level SAT participation rate and its square, state test requirements, county poverty and income, and district funding and expenditure variables, then compares raw to predicted means by school type. There is no prior achievement measure, no instrument, and no correction for selection into SAT-taking, which Belfield flags himself: the negative correlation between a state's participation rate and its mean score is long documented, and homeschoolers' 0.5% participation share means their raw scores are inflated by selection. Belfield states explicitly that these data cannot identify a treatment effect and that he is reporting relative differences and the size of the biases.

Raw SAT-Total for homeschoolers was 1093.1, about 0.4 SD above public school (1012.6) and 0.2 SD above private-religious (1055.6) — but 0.15 SD BELOW private-independent (1123.8). The premium is almost entirely Verbal, not Math: raw Verbal 566.6 vs 502.6 public, raw Math 526.5 vs 510.1 public. After covariate adjustment the school-type differences collapse: predicted totals converge (home 1054.5, public 1021.1, private-independent 1064.4, private-religious 1050.5), so the home-school premium over religious schooling falls almost to zero. Comparing each group's raw score to its own prediction, homeschoolers beat their predicted Verbal by ~40 points and came in fractionally BELOW their predicted Math (526.5 observed vs 527.7 predicted) — Belfield's reading is that any treatment effect is much larger for Verbal than Math, plausibly reflecting parental subject competence. The equity finding is the one nobody quotes: SES gradients are STEEPER for homeschoolers than for any other sector on SAT-Total (Q2 vs Q1 = 0.32 SD for homeschoolers vs 0.16 public, 0.14 private-religious, 0.21 private-independent), i.e. under home education family background does more of the work, not less. Belfield also reports that in five of the nine "high regulation" states (WA, UT, WV, PA, NY), homeschooled students' assessment results are either not required or not recorded, so the data needed to evaluate regulation do not exist.

High but partially bounded by the covariate set, which is unusually rich for this literature (both parents' education, income, religion, geography, state testing policy). What remains unadjusted is everything that makes a family choose home education and then choose to sit the SAT, plus the child's own ability — no prior achievement measure exists. The steeper SES gradient is direct evidence that heritable family characteristics do MORE explanatory work in the home-schooled arm, which is the opposite of what the equalisation story predicts.

The individual-level SAT analysis has never been repeated with better data; the College Board stopped releasing comparable homeschool files. Ray & Eagleson 2008 worked the same 2001 cohort with state aggregates and no covariates and reported the regulation null; Belfield's covariate-adjusted convergence has not been contested empirically, only ignored. The Verbal-not- Math pattern independently recurs in Rudner 1999's subtest profile and in Cardus-based reports of homeschool graduates' college selectivity, which is weak convergent support rather than replication.