Elusive Longer-Run Impacts of Head Start: Replications Within and Across Cohorts

1,251 (Deming cohorts, 364 HS / 364 other-preschool) / 2,144 (complement) / 3,768 combined (3,738 in the fitted FFE models; 1,195 for earnings)

Same CNLSY sibling design as Deming, extended to later cohorts. Deming's cohorts are broadly 1976-1986 births (Head Start attendance ~1989-90); the complement cohorts add ten later birth years (attendance ~1996-97).

Direct replication and extension of Deming (2009) in three parts: (a) re-estimate Deming's cohorts with ten more years of CNLSY follow-up, adding outcomes Deming could not measure (completed schooling, college graduation, observed earnings); (b) apply the same family-fixed-effects specification to ten later birth cohorts; (c) combine. Robustness includes the Miller-Shenhav-Grosz (2019) reweighting-on-observables correction for selection into the FFE identifying sample, and first-born interaction tests for sibling spillovers. NOTE: the copy read here is the January 2020 working paper, marked "PLEASE DO NOT CITE OR CIRCULATE WITHOUT PERMISSION"; the record's DOI points at the published EEPA version, so published numbers may differ slightly.

The Head Start long-run attainment result replicates for Deming's own cohorts (+0.228 SD, matching him almost exactly; +0.166 SD after ten more years of follow-up) but does NOT generalise to later cohorts (-0.15 SD), and combined across all cohorts the adult summary index is approximately zero (-0.01, SE 0.04). Two things this record previously omitted and that cut the other way: completed schooling still rises +0.31 years (SE 0.145) for Deming's cohorts, and for children of the lowest-AFQT mothers the index (+0.38 SD), schooling (+0.45 years) and EARNINGS (+0.44 log points) are all significant — the only place in the paper where earnings move at all. Anyone citing Deming's +0.23 without this paper is citing half the evidence; anyone citing this paper as a flat failure is citing the other half.

Siblings who attended no preschool; "other preschool" siblings are estimated as a separate arm.

Same sibling-FE limits as Deming. Reweighting for selection into identification attenuates the Deming-cohort estimates for high school graduation, idleness, learning disability and poor health, but leaves ASI and the combined-cohort estimates largely unchanged — so the collapse across cohorts is not a selection-into-identification artefact. Sibling-spillover tests (first-born interactions) find sparse and inconsistent evidence, as Deming did. A separate confound the paper itself surfaces: the effect is concentrated in children of low-AFQT mothers, i.e. exactly where the untreated counterfactual environment is worst.

This IS the replication, and the honest verdict is mixed, not failed. What replicated: Deming's headline +0.23 SD adulthood index reproduces at +0.228 (SE 0.072) on his own cohorts and window, and the school-age estimates are "aligned"; extending the window ten years leaves it significant at +0.166 (SE 0.069); completed schooling is +0.31 years (SE 0.145); and for children of low-AFQT mothers the index (+0.38), schooling (+0.45 years) and earnings (+0.44 log points) are all significant and significantly larger than for the rest of the sample. What failed: the same design on ten later birth cohorts gives -0.15 SD, combined across cohorts ~0, and neither college graduation nor observed earnings moves for anyone outside the low-AFQT subgroup. The pattern is a finding that holds for its original cohorts and does not generalise forward.