Effects of Poverty, Funding Structure and Scale on Public School System Performance John Mackenzie FREC/CANR, University of Delaware May, 2010 BACKGROUND: Coleman Report (1966): finds link between money and school performance to be tentative at best. (Basic problem: lack of data!) Eric Hanushek (1981,1986,1997): “Throwing Money at Public Schools” meta-analyses of studies relating funding to school performance. (Methodological error in counting studies as datapoints: 7 heads in 10 coin tosses does not prove a coin is biased, but combining 100 trials of 10 tosses and getting 7 or more heads in 60% of the trials does.) Jay Greene (Manhattan Institute): Real per-pupil spending “almost doubled” from 1972 to 2002, while NAEP scores did not improve much at all. (So what? Real per-capita disposable incomes “almost doubled” too. Compare rising costs of college!) Public schools today are… more inclusive of minorities, immigrants, etc.; offer a broad array of non-traditional services; serve an expanded population of poor children; deliver more remedial & special education. Education is typically a “luxury” good: as incomes rise, households invest larger proportions of their incomes in education. Education confers status; may be a positional good. US income inequality continues to increase, and the relative economic return to a HS diploma is falling. If education lifts families from poverty for multiple generations, is residual poverty becoming more intractable? Does the rising real cost of public schools simply reflect the rising marginal cost of eliminating poverty? School Spending/Pupil vs. Median Household Income, by State Is Public Education a “Luxury” Good? $14,000 y = 0.1242x + 2793.5 K-12 Public School Spending/Pupil, 2004 R2 = 0.2252 $12,000 US: $8,287 per pupil $44,231 median income Elasticity = +0.66 $10,000 $8,000 $6,000 $4,000 $30,000 $35,000 $40,000 $45,000 Median Household Income, 2003 $50,000 $55,000 $60,000 A favorite neo-con hobbyhorse: “Throwing Money at Schools” SAT05 vs. Total 2004 Per-Pupil Spending, by State 1250 1200 1150 1100 1050 1000 950 $4,000 $5,000 $6,000 $7,000 $8,000 $9,000 $10,000 $11,000 $12,000 $13,000 $14,000 NAÏVE UNIVARIATE MODELS [t-statistics in brackets] SAT02 = 1187.16 – 0.0156*TXPP02 {25.70] [-2.59] N=50; R-square=0.123 SAT03 = 1185.38 – 0.0143*TXPP03 [26.06] [-2.51] N=50; R-square=0.116 SAT04 = 1181.65 – 0.0131*TXPP04 {26.85] [-2.49] N=50; R-square=0.114 SAT05 = 1196.88 – 0.0117*TXPP05 [28.45] [-2.87] N=50; R-square=0.146 Those kids get dumber every year, Jill! They sure do, Bob! Is my hair OK? More waste and bloat in public education? School spending rises and SAT scores fall! State mean SAT scores depend on test participation rates 1250 2004 SAT verbal+math 1200 1150 SAT04 = 993.23 - 59.574Ln(%Participation) 2 R = 0.8435 1100 1050 1000 950 0% 20% 40% 60% Percent of seniors taking the SAT in 2004 80% 100% A textbook example of an omitted variable problem: REGRESSION MODELS CONTROLLING FOR PARTICIPATION SAT02 = 1062.5 + 0.0135*TXPP02 – 244.58*Partic02 [43.29] [3.46] [-12.82] N=50; R-square=0.806 SAT03 = 1074.5 + 0.0113*TXPP03 – 234.60*Partic03 [43.29] [3.46] [-12.82] N=50; R-square=0.809 SAT04 = 1079.4 + 0.0104*TXPP04 - 230.50*Partic04 [48.65] [3.37] [-13.00] N=50; R-square=0.807 SAT05 = 1095.9 + 0.0070*TXPP05 - 221.17*Partic05 [46.60] [2.63] [-11.47] N=50; R-square=0.775 ECONOMETRIC DIGRESSION A formal treatment of the participation bias problem: When only high-performing students take the SAT the mean score is biased upward. Expanding participation to include more typical students educes the bias. Low SAT participation boosts the mean scores of low-performing states above the mean scores of high-performing states. 2005 ACT Participation vs. 2005 SAT Participation, by State 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 2005 SAT Participation vs. Total 2004 Per-Pupil Spending, by State 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% $4,000 $6,000 $8,000 $10,000 $12,000 $14,000 $16,000 FORMAL HECKMAN MODEL: PROBIT: NORMSINV(Partic%05) = ExpPP04, ACTPart%05 Regression Statistics Multiple R 0.9153 R Square 0.8378 Adjusted R Square 0.8309 Standard Error 0.3970 Observations 50 ANOVA df Regression Residual Total Intercept ExpPP04 ACTPart% SS 38.245 7.406 45.651 MS 19.122 0.158 F 121.357 Coefficients Standard Error -0.093569 0.374798 0.000107 0.000038 -2.531849 0.211647 t Stat -0.249651 2.790748 -11.962623 P-value 0.803945 0.007576 0.000000 2 47 49 PROBIT Residuals State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Predicted -1.34361 Residuals 0.06205 Lambda 0.17768 PredPart% 9.0% 0.32769 0.06969 -1.29832 0.37901 -1.80895 0.80478 0.89688 -0.40756 -0.00240 0.41217 -0.91858 -1.67618 0.25853 -0.95007 -1.21531 -1.28251 -1.47621 0.67091 0.58593 -0.27753 -0.50960 -0.25645 -0.37901 1.16560 0.27554 -0.25354 0.79288 0.67689 -0.13285 0.11216 0.39463 0.15393 -0.69479 -0.12544 0.10753 0.07114 0.00358 -0.03254 1.01753 0.84277 0.19021 1.05379 0.08052 1.37111 1.44319 0.55781 0.79636 1.07747 0.31873 0.10272 0.96943 0.30646 0.21470 0.19472 0.14428 1.26839 1.20453 62.8% 52.8% 9.7% 64.8% 3.5% 79.0% 81.5% 34.2% 49.9% 66.0% 17.9% 4.7% 60.2% 17.1% 11.2% 10.0% 7.0% 74.9% 72.1% SAT05PCT 10% 52% 33% 6% 50% 26% 86% 74% 65% 75% 61% 21% 10% 66% 5% 9% 12% 8% 75% 71% Lambda = NORMDIST(PredY,0,1,FALSE)/(1-NORMDIST(PredY,0,1,TRUE)) = f(X)/[1-F(X)] where f and F are standard normal density and distribution 2005 SAT %Participation, by State -- Actual vs. Predicted from Probit 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Predicted & Actual Participation vs. LAMBDA 1.80 1.60 1.40 1.20 1.00 0.80 0.60 0.40 0.20 0.00 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Incorporate LAMBDA as participation bias correction instrument: Heckman 2nd Stage: SAT05 = ExpPP04, Lambda Regression Statistics Multiple R 0.8222 R Square 0.6760 Adjusted R Square 0.6622 Standard Error Observations 38.9339 50 ANOVA df SS MS F 2 148622.07 74311.03 49.02 Residual 47 71244.91 1515.85 Total 49 219866.98 Coefficients Std Error t Stat P-value 1054.6446 30.8222 34.2170 0.0000 ExpPP04 0.0148 0.0045 3.3095 0.0018 LAMBDA -145.1194 15.9954 -9.0726 0.0000 Regression Intercept Or construct the bias correction instrument directly from the observed participation rates: Fast Heckman: SAT05 = ExpPP04, L* L* = (NORMDIST(NORMSINV(%Partic05),0,1,FALSE)/(1-%Partic05) Regression Statistics Multiple R 0.86749 R Square 0.75254 Adjusted R Square 0.74201 Standard Error 34.02397 Observations 50 ANOVA Regression Residual Total Intercept ExpPP04 Heckman df 2 47 49 SS 165458.35 54408.63 219866.98 MS 82729.17 1157.63 F 71.46 Coefficients 1064.9402 0.0128 -132.3417 Std Error 26.1859 0.0037 11.9657 t Stat 40.6684 3.4751 -11.0601 P-value 0.0000 0.0011 0.0000 Or use the equivalent logit-based instrument (McFadden): Fast McFadden: SAT05 = ExpPP04, MU where MU = -%Part05*ln(%Part05)/ln(1-%part05) - ln(1-%Partic05) Regression Statistics Multiple R 0.85742 R Square 0.73517 Adjusted R Square 0.72390 Standard Error 35.19792 Observations 50 ANOVA Regression Residual Total Intercept ExpPP04 MU df 2 47 49 SS 161639.0 58228.0 219867.0 MS 80819.5 1238.9 F 65.2 Coefficients 1060.261 0.01335 -74.0986 Std Error 27.343 0.00386 7.0262 t Stat 38.777 3.45532 -10.5460 P-value 0.0000 0.0012 0.0000 Fast McFadden vs. Fast Heckman 4.00 3.50 y = 1.7858x - 0.0017 R2 = 0.9985 3.00 2.50 2.00 1.50 1.00 0.50 0.00 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 2.00 “50 experiments in public school system finance and structure” Data Sources: US Census Bureau, Public Elementary–Secondary Education Finance Data (annual State-level tables) http://www.census.gov/govs/school US Dept. of Education, 2003, 3005, 2007 and 2009 National Assessments of Education Progress (NAEP) http://nces.ed.gov/nationsreportcard/statecomparisons/ NAEP09 vs. Total 2007 Per-Pupil Spending, by State 1080 1060 1040 1020 1000 y = 0.0035x + 967.43 2 R = 0.1343 980 960 940 $6,000 $8,000 $10,000 $12,000 $14,000 $16,000 $18,000 $20,000 NAEP09 vs. 2007 FEDERAL Funding per Pupil, by State 1080 1060 1040 1020 y = -0.0312x + 1038.3 R2 = 0.1488 1000 AK 980 LA 960 940 $500 $700 $900 $1,100 $1,300 $1,500 $1,700 $1,900 $2,100 $2,300 NAEP09 vs. 2007 STATE Funding Per Pupil, by State 1080 1060 VT 1040 1020 1000 y = -5E-05x + 1007.4 2 R = 0.00002 980 HI 960 940 $2,000 $4,000 $6,000 $8,000 $10,000 $12,000 $14,000 $16,000 NAEP09 vs. 2007 LOCAL Funding per Pupil, by State 1080 1060 MA NJ CT 1040 PA 1020 NY 1000 y = 0.0062x + 978.52 2 980 R = 0.2697 960 940 $0 $2,000 $4,000 $6,000 $8,000 $10,000 $12,000 NAEP09 vs. 2007 FEDERAL, STATE and LOCAL Per-Pupil Funding, by State Regression Statistics Multiple R 0.62032 R Square 0.38480 Adjusted R Square 0.34468 Standard Error 20.55471 Observations 50 ANOVA df Regression Residual Total Intercept FED07PP STATE07PP LOCAL07PP SS 12156.15 19434.82 31590.97 MS 4052.05 422.50 Coefficients Standard Error 992.89010 15.28438 -0.02583 0.00964 0.00211 0.00137 0.00599 0.00144 t Stat 64.96109 -2.67844 1.54618 4.16168 3 46 49 F 9.59 P-value 0.00000 0.01022 0.12891 0.00014 NAEP09 vs. Percent of Students Eligible for FRPL, by State 1080 1060 1040 1020 y = -249.32x + 1112 1000 R2 = 0.8076 980 960 940 920 20% 30% 40% 50% 60% 70% 80% 2009 NAEP 4th grade MATH 2009 NAEP 8th grade MATH 2009 NAEP 4th grade READING 2009 NAEP 8th grade READING NAEP vs. Log2(Avg. District Size), by State 1080 y = -7.123x + 1090.6 1060 2 R = 0.2322 1040 1020 1000 980 960 940 6 8 10 12 14 16 18 State funding replacing property tax Targeted Federal dollars (Title I, etc.) Less reliance on property taxes in larger districts Less reliance on local funding in higher-poverty states Larger average district size in higher-poverty states NAEP08 vs. Per-Pupil Funding by Source and Log2 of Avg. District Size Regression Statistics Multiple R 0.7959 R Square 0.6335 Adjusted R Square 0.5918 Standard Error 16.2225 Observations 50 ANOVA Regression Residual Total Intercept FED07PP STATE07PP PROPTX07PP OTHRLOC07PP L2AVDISTSZ df 5 44 49 SS 20011.50 11579.47 31590.97 MS 4002.30 263.17 F 15.21 Coefficients Standard Error 1089.5879 21.7298 -0.0318 0.0077 0.0028 0.0011 0.0046 0.0015 0.0059 0.0013 -7.7268 1.4171 t Stat 50.1425 -4.1269 2.5162 3.1430 4.6607 -5.4526 P-value 0.0000 0.0002 0.0156 0.0030 0.0000 0.0000 NAEP08 vs. Federal, State, Property Tax and Other Local Funding, Percent Poverty and log2 of Dis of Avg. District Size Regression Statistics Multiple R R Square Adjusted R Square Standard Error Observations 0.9315 0.8676 0.8491 9.8625 50 ANOVA df Regression Residual Total Intercept FED07PP STATE07PP PROPTX07PP OTHRLOC07PP PCTPOV L2AVDISTSZ SS 27408.45 4182.52 31590.97 MS 4568.07 97.27 Coefficients Standard Error 1127.24317 13.89842 -0.00488 0.00561 0.00061 0.00071 0.00129 0.00096 0.00241 0.00087 -197.56627 22.65538 -3.74497 0.97503 t Stat 81.10587 -0.87019 0.85619 1.34205 2.77650 -8.72050 -3.84086 6 43 49 F 46.96 P-value 0.00000 0.38903 0.39664 0.18662 0.00811 0.00000 0.00040 Similar results are obtained with 2007, 2005 and 2003 NAEP scores analyzed against contemporaneous or lagged funding, average district size, poverty, etc. (These are large, slowly-evolving systems) Poverty is the strongest predictor of NAEP performance, but is collinear with federal funding (+), local funding (-) and average district size (+). Structural Variables: Federal funding is mainly targeted to low-performing populations—Title I and special needs. State funding does not drive NAEP performance much at all. Local funding drives NAEP performance most efficiently. Average district size is negatively correlated with NAEP performance, suggesting scale diseconomies. Implications of funding and average district size: HI, FL*, MD*, NV, UT, NC, LA: large average district size, generally more state funding, lower NAEP scores VT*, MT, ND, ME, SD, NE, OK, NH: many small districts, generally more local funding, higher NAEP scores Likely sources of public dissatisfaction with schools: Displacement of local funding with state funding, and consequent loss of local control. Legacy of desegregation, Serrano, consolidation, bureaucratization, political intervention and accumulation of regulation, union protection of teacher seniority, etc. These results suggest the efficiency of public education can be improved by restoring local autonomy of school systems. Why would local dollars drive school system performance more efficiently than state dollars? Local funding implies stronger local governance and accountability.