Statistical results overview Updated: 2026-09-10 Purpose - Summarize the audited public Evidence Lab models of socioeconomic context, attendance, and achievement. - Keep the model population, outcome identity, weights, validation, and interpretation visible beside the headline numbers. Primary scope and modeling setup - Primary unit: one ordinary ODE Regular elementary school site, classified from its cross-subject tested-grade profile (grades 3-5 only). - High-school and combined-grade sites are excluded. Middle-school results are separate sensitivities because a school-address tract is a weaker proxy for a wider catchment. - Outcome: each subject's official Total Population All Grades row for 2024-25. - Achievement rates and student weights use exact Level 1-4 counts and the Scored Performance Denominator. Participant and attendance counts are never achievement denominators. - Attendance models require the 2024-25 tested-grade rollup and exact included/regular-attender count identity. - BA+ is an ACS tract adult share; income is ACS median household income; SEP is ODE Students Experiencing Poverty for the enrolled school population. - Results are school-level associations and held-out predictive checks, not causal effects or student-level relationships. 1) Income and adult BA+ Standardized two-predictor results (income + BA+): - ELA: beta income 0.209, beta BA+ 0.577, in-sample R^2 0.557, repeated-CV R^2 0.551 (470 schools). - Math: beta income 0.239, beta BA+ 0.550, in-sample R^2 0.556, repeated-CV R^2 0.549 (469 schools). - Science: beta income 0.166, beta BA+ 0.578, in-sample R^2 0.505, repeated-CV R^2 0.496 (438 schools). Interpretation: BA+ has the larger controlled association in every elementary-primary subject. Repeated half-sample ranges preserve that ordering; those ranges describe split stability, not confidence intervals. 2) School poverty and adult BA+ Standardized two-predictor results (SEP + BA+): - ELA: beta SEP -0.649, beta BA+ 0.273, joint R^2 0.747; added R^2 BA+ after SEP 0.037, SEP after BA+ 0.211. - Math: beta SEP -0.614, beta BA+ 0.294, joint R^2 0.719; added R^2 BA+ after SEP 0.043, SEP after BA+ 0.188. - Science: beta SEP -0.560, beta BA+ 0.302, joint R^2 0.646; added R^2 BA+ after SEP 0.045, SEP after BA+ 0.153. Interpretation: SEP is stronger in these specifications, while BA+ retains distinct explanatory information in all three subjects. 3) Attendance in joint models Standardized income + BA+ + attendance results: - ELA: beta income 0.065, beta BA+ 0.455, beta attendance 0.438, in-sample R^2 0.684, repeated-CV R^2 0.677 (457 schools). - Math: beta income 0.088, beta BA+ 0.430, beta attendance 0.446, in-sample R^2 0.687, repeated-CV R^2 0.680 (456 schools). - Science: beta income 0.110, beta BA+ 0.487, beta attendance 0.290, in-sample R^2 0.571, repeated-CV R^2 0.561 (433 schools). Poverty-aware SEP + BA+ + attendance results: - ELA: beta SEP -0.502, beta BA+ 0.259, beta attendance 0.228, repeated-CV R^2 0.768. - Math: beta SEP -0.443, beta BA+ 0.275, beta attendance 0.267, repeated-CV R^2 0.749. - Science: beta SEP -0.486, beta BA+ 0.292, beta attendance 0.137, repeated-CV R^2 0.652. 4) Income-versus-poverty predictive comparison Mean repeated five-fold CV R^2 (BA+ and attendance are included in every model): - ELA: income 0.6767, poverty 0.7676, income + poverty 0.7727. - Math: income 0.6788, poverty 0.7475, income + poverty 0.7485. - Science: income 0.5636, poverty 0.6538, income + poverty 0.6546. The poverty-aware model predicts better in all three subjects. Adding income back produces only a small further gain. SEP midpoint, lower endpoint, upper endpoint, and exact-only checks preserve the qualitative ordering. 5) Interaction checks Mean paired held-out delta R^2 after adding all pairwise interactions: - ELA: income model 0.005; poverty model 0.000. - Math: income model 0.016; poverty model 0.012. - Science: income model 0.000; poverty model 0.004. Interpretation: the clearest held-out interaction gain is in Math. ELA's income-model gain is small, and Science ranges include zero in the principal specifications. The old broad claim that interactions improved every subject is not supported by the corrected held-out test. 6) Other public exploratory families - Outcome-level, spending, class-size, geography, and historical analyses remain separate model families. Their claims should not be treated as verified by this core-SES rebuild unless their own audited report says so. - The pre-pandemic comparison is intentionally deferred pending the historical-data parameterization audit. Overall interpretation - The large statewide association between socioeconomic context and achievement remains. - SEP is the strongest enrolled-student hardship measure in these elementary-primary specifications. - Adult BA+ is the strongest tract-based measure and adds information beyond SEP. - Attendance adds substantial information, especially in Math, but its standardized contribution is smaller after SEP replaces income. - Income remains useful neighborhood description but adds comparatively little after BA+, attendance, and SEP. - Local estimates require extra caution because tract-to-school linkage, enrollment choice, and small samples vary by place. Replication The Evidence Lab package includes the audited source scripts, machine-readable results, fixed input hashes, coverage diagnostics, and public report companions.