SES explorations summary (student-weighted) Generated: 2026-02-22 Purpose Summarize the analyses used to compare education vs income signals and assess SES proxy quality. All analyses use Total Population rows and drop All Grades when grade-specific rows exist. 1) Joint education + income model (baseline) Source: docs/income_education_joint_model_report.txt - Education consistently has a larger standardized beta than income. - R^2 (both predictors) is ~0.33-0.37 for ELA/Math and ~0.28 for Science. - Bootstrap CIs show education beta is robust; income beta is smaller and can approach 0 in Science. Significance: education appears to explain more unique variation than income in statewide models. 2) Split-sample stability Source: docs/income_education_split_stability_report.txt - Across 200 random half-splits by school, education beta > income beta in 100% of splits. - Locale splits show the strongest signal in suburbs and cities; weaker in towns/rural. Significance: education dominance is stable across resamples, not driven by a small subset. 3) Alternative SES predictors (education, poverty, per-capita income) Source: docs/alternative_ses_metrics_report.txt - Education and per-capita income both correlate strongly with proficiency. - ACS tract poverty rate has weak standalone correlations and minimal impact once the other two are included. - In the 3-predictor model, education still has the largest beta; per-capita income is second. Significance: in the older tract-SES framing, education remains the strongest single SES proxy, with per-capita income adding some signal. In newer 2024-25 poverty-aware models, poverty joins BA+ and attendance in the top tier and sometimes leads. 3a) School-population poverty (ODE) reassessment Sources: - docs/income_poverty_reassessment_note_2026-02-20.txt - docs/why_poverty_outpredicts_income_explainer_2026-02-21.txt - docs/ba_signal_in_high_poverty_summary_2026-02-21.txt Findings: - ODE Students Experiencing Poverty is strongly related to proficiency and often outperforms income-only specs in 2024-25 models. - In non-charter/non-virtual CV checks, poverty-aware models improve fit versus income-based analogs in ELA, Math, and Science. - BA+ remains strong in high-poverty schools, but the BA+ slope is dampened relative to low-poverty schools. Significance: poverty should be treated as a first-tier 2024-25 hardship indicator, while income remains a broader community context variable. 3b) Income metric nuance (median household vs per-capita) Source: docs/income_education_income_metric_note_2026-02-14.md - At the tract level, BA+ correlates more strongly with per-capita income than with median household income. - On matched school-analysis rows, per-capita income is also more correlated with proficiency than median household income (ELA/Math/Science). Significance: income metric choice changes effect strength; per-capita can capture SES gradients that median household income partially compresses. 4) Non-linear (spline) checks Source: docs/ses_spline_report.txt - Splines provide only small R^2 gains for education and per-capita income (Delta R^2 ~0.00-0.02). - Poverty rate shows a larger relative non-linear gain, but overall R^2 stays low. Significance: linear models are adequate for education and per-capita income; strong non-linear effects are limited. 5) Tract centrality sensitivity Sources: docs/tract_centrality_sensitivity_report.txt, docs/tract_centrality_grade_sensitivity_report.txt - Correlations and R^2 increase as schools closer to tract centers are retained. - Elementary grades show the clearest improvement at thresholds ~0.20-0.30. - Middle/high gains are weaker and require high thresholds with small samples. Significance: tract-level SES is more reliable for centrally located schools, especially elementary. 6) District fixed-effects (optional lens) Source: docs/income_education_district_fe_report.txt - Within-district models reduce R^2 to ~0.14-0.18. - Education remains stronger than income but with a smaller gap. Significance: much of the statewide signal reflects between-district variation; within-district signal still favors education. 7) Binned slope check (plateaus and non-linearity) Source: docs/ses_binned_slope_report.txt - Education (BA+ rate) shows a mostly monotonic increase across deciles with no plateau-like segments in ELA/Math and only minor reversals in Science. - Per-capita income remains positive overall but shows more local variability (a few sign reversals), especially in Science. - Poverty rate has the most non-linearity and occasional reversals; overall predictive power remains low. Significance: the education gradient appears continuous across the SES spectrum with limited plateau evidence; non-linearity is more evident for poverty rate. 8) Historical and grade-scope checks (Math) Sources: - docs/income_education_math_2018_2019_comparison.txt - docs/income_education_joint_model_report_math_no_hs.txt - docs/income_education_joint_model_report_math_grade35.txt - docs/income_vs_education_influence_report_2018_2019.txt - docs/income_education_joint_model_report_2018_2019.txt - 2018–2019 Math file enriched with ACS 2018 5-year (tract-level). Findings: - Using era-appropriate ACS for 2018–2019 does not change the relative ranking: education remains the stronger predictor than income. - Excluding high school rows slightly strengthens the SES signal (especially education), consistent with HS attendance areas diluting tract-based SES. - Restricting to grades 3–5 reduces sample size substantially but preserves the same ordering; confidence intervals widen modestly. Significance: the education-vs-income ordering is stable across pre-pandemic data and tighter grade scopes. 8b) Adult-education boundary step-size check (dominant-group lens) Sources: - scripts/report_education_level_boundary_step_sizes.py - artifacts/analysis/education_level_boundary_step_sizes_dominant_group_report.txt - artifacts/analysis/education_level_boundary_step_sizes_dominant_group.csv Findings: - The largest adjacent proficiency drop is consistently at the BA+ -> SomeCollege/Assoc boundary: - ELA: 14.08 points (95% CI 12.64 to 15.52) - Math: 14.99 points (95% CI 13.44 to 16.54) - Science: 10.74 points (95% CI 8.63 to 12.85) - The next boundary (SomeCollege/Assoc -> HSOnly) is much smaller (about 1 to 2 points), and is not statistically clear in Science. - HSOnly -> <HS can be sizable but relies on small-sample groups, so uncertainty is higher. Significance: this supports an apparent outsized BA+ boundary in the current data, with much smaller mid-tier step changes. Caveat: this is an observational decomposition using tract-level dominant education groups, not a causal estimate of education effects. Overall takeaway - Education (BA+ rate) is the most consistent SES predictor in this dataset. - For ACS tract variables, per-capita income adds more signal than ACS poverty rate. - For 2024-25 school-population hardship, ODE Students Experiencing Poverty adds substantial signal and can outperform income-only specs. - Tract-level SES is useful but noisy; centrality filtering can improve reliability at the cost of sample size.