The Effect of Public Schooling on Test Scores

Newhouse and Beegle 541 point. Meanwhile, despite the theory that increased wealth increases the probability of private school attendance ceteris paribus, there is no consistent pattern between the quality of the floor and the probability of attending private secular or Muslim school. The student’s past academic performance has a larger effect on the probability of attending private school. An increase in the elementary school exam score of one stan- dard deviation lowers the probability of attending private secular school by nine per- centage points; the probability of attending private Muslim school falls by seven percentage points, and the probability of attending a private Madrassah falls by one percentage point. Grade repetition in elementary school also reduces in the probabil- ity of attending a secular private school, although this effect is smaller and not statis- tically significant. Likewise, parental education is generally not a strong determinant of school type. Parental education is presumably correlated with student motivation, and, in results presented later, is shown to be a strong predictor of junior secondary school test score. The relatively small effect of parental education suggests that after controlling for lagged test scores, there may not be a large difference in student moti- vation at different types of schools. These results confirm the widespread impression that public schools in Indonesia tend to benefit from positive selection. If elementary school test score and the lack of repetition are positively correlated with unobservable determinants of learning, like motivation, the estimated effects of public school attendance on test scores will be biased upward. The next section turns to examining estimates of the effect of public schooling on test scores.

VII. The Effect of Public Schooling on Test Scores

Are public or private schools, on average, more effective at raising the test scores of Indonesian junior secondary school students? To address this question, we regressed the respondents’ normalized junior secondary test score on the control variables listed above, with school type represented by a dummy variable for public school attendance. To conserve space, only the coefficient on school type is shown. Column 1 of Table 2 indicates that public school students, in the presence of controls, score 0.19 standard deviations higher than private school students. The second speci- fication includes time-varying control variables measured within a year of junior sec- ondary graduation, which are only available for the junior secondary school subsample. The estimated public school effect, reported in Column 2, increases to 0.23 standard deviations in this specification. When time-varying variables measured around the student’s elementary school graduation are added, the magnitude of the premium rises slightly, to 0.26 standard deviations Column 3. The final specification Column 4 includes family-level fixed effects, which identifies the public school effect using sib- lings that attended different types of schools, and the estimated public school pre- mium is 0.24 standard deviations. 15 15. Results of the fixed effects estimation strategy are not reported for the junior secondary school and ele- mentary subsamples because there is little variation within family in the time-varying variables that are included in these subsamples. The Journal of Human Resources 542 Of course, least squares estimates of the public school effect will be biased if pub- lic school attendance is correlated with unobserved factors that determine test scores. In the Indonesian context, the direction of this endogeneity bias is unclear in theory, as described above. However, the correlation between observable characteristics and school choice suggests that public schools benefit from positive selection, which might bias the estimated public school premium upward. Moreover, because parents choose schools separately for each child within the household, partly on the basis of unobservable child characteristics, the inclusion of family-level fixed effects does not eliminate this bias. 16 Least squares estimates of the public school effect also will be biased if recall error in the test score is correlated with the type of school attended. To infor- mally assess the effect of recall error, we exploit the fact that survey asked respon- dents to produce their official test report if available. Sixty-two percent of the respondents in the final sample showed their test card to the interviewer in either 1997 or 2000. We include a dummy variable in the regression for whether the respondent showed a card. In the regression determining test score in the junior high school sample, the coefficient on the card dummy is −0.09 and statistically sig- nificant, which could reflect a mixture of respondents overstating scores recalled from memory and higher scoring students being less likely to retain their card. Omitting the card dummy from the regression has a negligible effect on the school- type coefficients, however, and coefficients on the interactions of the card dummy 16. In the sample of children for whom two siblings attended different types of schools, within-family vari- ation accounts for 39 percent of the total variation in test scores, which is consistent with intra-household selection into public schools. Table 2 Effect of public school attendance on junior secondary school test score Family fixed OLS OLS OLS effects Junior secondary Elementary Sample and Sample and Full Full sample Variables Variables sample Attended public junior 0.192 0.227 0.264 0.244 secondary school 0.029 0.039 0.044 0.066 Observations 4,382 2,733 1,948 883 R -squared 0.45 0.48 0.51 0.76 Note: Robust standard errors in parentheses. significant at 5 percent; significant at 1 percent. Regression includes other control variables listed below Table 1. and school type were not statistically significant. This suggests that recall error has a weak association with school type, and is unlikely to be a serious concern in our data. Nonrandom sorting of students into different types of schools remains a potential source of bias. To address it, we estimate two-stage-least-squares models of test scores, employing measures of the local availability of public schools as an instrument for public school attendance. 17 This approach has been used to estimate the effect of Catholic schooling effect in the United States see, for example, Neal 1997; Figlio and Ludwig 2000. The importance of availability of private schools in schooling choices has been demonstrated in the developing country context see, for example, Alderman, Orazem, and Paterno 2001. Data on the presence of public and private schools are available at the district level from the school census data collected by the Ministry of Education. We measure the percentage of schools that are public in the district of ju- nior secondary school attendance. The consistency of the two stage least squares estimate is based on the critical assumption that local private school proximity is uncorrelated with unobserved deter- minants of student test scores. This assumption has been questioned in the U.S. con- text, where evidence suggests that proximity to catholic secondary schools is correlated with unobserved determinants of 12th grade math and reading test scores Altonji, Elders, and Taber 2002. However, that conclusion is largely based on the implausibly large differences between OLS and 2SLS estimates of the effect of Catholic schooling on scores, which we do not find in the Indonesia data. Also, the location of American Catholic secondary schools is heavily influenced by historical pattern of past Catholic migration Hoxby 1994, implying that the positive correlation between student unob- servables and proximity to American Catholic secondary schools does not generalize to Indonesian private middle schools. In contrast to selection bias, which likely leads to an overestimate of the public school premium, it is not clear how the location decisions of public and private schools will bias the estimated public school effect. If public schools are spread uni- formly throughout a population that is heterogeneous in its demand for education, then profit-maximizing private schools will locate in areas where demand for educa- tion, and therefore student achievement, is higher. In this case, the estimated public school premium will be biased downward. This downward bias may be mitigated or reversed by two factors. First, the national education department may maximize edu- cational achievement by locating public schools in areas with high student ability for an example of endogenous program placement, see Pitt, Rosenzweig, and Gibbons 1993. Second, the estimated public school effect could be upwardly biased if private andor Muslim schools are more appealing to parents living in areas with undisci- plined students. Likewise, households themselves may make location decisions based on school availability. Concerns regarding endogenous matching of households and schools are lessened to the extent that schools and households locate based on char- acteristics of the population that are included in the model, such as average district- level and student-level test scores. Overall, however, theory provides no clear Newhouse and Beegle 543 17. The IV procedure also eliminating bias due to reporting error in test scores, if reporting error is uncor- related with the local availability of public schools. The Journal of Human Resources 544 guidance as to whether private schools are more common in areas with unobservably stronger or weaker students. To gain insight into the validity of using the percentage of district schools that are public as an instrument for public school attendance, we regress four important determinants of test scores on the instrument, while controlling for all other control variables in the model. These results are reported in Table 3 for mother’s education, father’s education, elementary school test score, and having repeated a grade at the elementary level. The percent of junior secondary schools in the district which are public is not statistically significant in any of the specifications in Table 3. Moreover, the signs of the estimated effects are not consistent. For example, public school access is positively correlated with maternal schooling and negatively corre- lated with paternal schooling. In addition, public school access is negatively correlated Table 3 Effect of public school access: alternative outcome variables Junior Elementary secondary Sample Full Sample and and sample Variables Variables Dependent variables Mother’s education Percent of junior secondary schools 0.081 0.109 0.069 in district that are public 0.223 0.240 0.254 Observations 4,274 2,699 1,934 Father’s education Percent of junior secondary schools −0.053 −0.018 −0.098 in district that are public 0.175 0.199 0.215 Observations 4,252 2,666 1,908 Elementary school test score Percent of junior secondary schools −0.032 0.081 0.191 in district that are public 0.235 0.268 0.277 Observations 4,382 2,733 1,948 R -squared 0.21 0.23 0.25 Repeated a grade in elementary school Percent of junior secondary schools −0.112 −0.119 −0.294 in district that are public 0.236 0.278 0.318 Observations 4,360 2,687 1,923 Note: Each of the four sets of rows refers to a different dependent variable regressed on access to public schools for three different samples. Robust standard errors in parentheses. significant at 5 percent; sig- nificant at 1 percent. Regression includes other control variables listed below Table 1, excluding the depend- ent variable if listed. Parental education regressions are ordered probits. Elementary test scores regressions are OLS. Repeated a grade in elementary school regressions are probits. Newhouse and Beegle 545 with elementary school test score in the full sample but positively correlated in the other two samples. The lack of a clear positive association between the instrument and observable determinants of junior high school test score provides some reas- surance that the instrumental variable estimates are not systematically biased. Table 4 presents the instrumental variables results of the effect of junior secondary school choice on test scores. Using a district-level measure of access to private schools, the public school premium falls slightly to 0.17 for the full sample. The esti- mated premium rises to 0.31 when the junior secondary sample is used, but falls to 0.16 in the elementary school subsample. None of the instrumental variable estimates are statistically significant. The first stage F statistic on these instruments ranges from 25 to 37, meaning that finite sample bias due to weak instruments is not an important concern. 18 Taken as a whole, the results from regressions estimating the average effect of pub- lic schools on test scores are consistent. Least squares estimates suggest that public Table 4 Instrumental variables estimates of the effect of public school attendance on junior secondary school test score Junior Full secondary Elementary sample sample sample Attended public junior secondary school 0.171 0.308 0.163 0.299 0.325 0.361 First stage results F statistic 41.6 32.4 29.2 Partial R-squared 0.021 0.021 0.028 Hausman chi-squared 0.0007 0.0065 0.0184 Hausman P 0.98 0.94 0.89 Observations 4,382 2,733 1,948 R -squared 0.45 0.48 0.51 Note: Robust standard errors in parentheses. significant at 5 percent; significant at 1 percent. Regression includes other control variables listed below Table 1. The instrumental variable is the percent of junior sec- ondary schools in the district that are public. 18. When the village-level presence of private schools is used based on IFLS community data, the results are only estimated for the subsample of students that were interviewed in the subdistrict in which they went to junior secondary school. In comparison with the district access instruments, the estimated public school premium stays roughly the same in the full sample and the junior secondary sample, and rises dramatically in the elementary school sample in results not presented here. The first stage F statistics for the village level instrument are all above 22 in these regressions. We discount the results using village-level instruments, because the sample excludes inter-subdistrict movers. school attendance raises a student’s test score 0.19 to 0.26 standard deviations. Using district-level access instruments generally results in similar estimated effects, as the estimated effect ranges from 0.17 to 0.31. The similarity of the magnitudes of the OLS and the instrumental variable estimates suggest that in total, the endogeneity bias resulting from parent’s choice of school type does not invalidate the qualitative con- clusions drawn from the OLS and fixed-effect estimates. Furthermore, the consistent finding of a positive public school premium across all estimation strategies is strong evidence that public junior secondary schools, on average, provide superior prepara- tion for the national exam. Although the focus has been on test scores, the effect of public school attendance could be extended to additional affects of attending public junior secondary school. Moreover, these outcomes may provide some insight into the channels through which public school attendance affects test scores. In 1997, the IFLS records the self- reported amount of hours a person spent studying at home per week during attendance of junior secondary school. Because information on study hours was only collected in 1997, the sample size is limited and results should be interpreted with caution. In an OLS regression, public school attendance is associated with a minor, but statistically significant, increase of 0.7 hours per week in study time. The estimated magnitude is similar for the fixed effect model, but rises to an unrealistic 9.3 extra study hours in the IV model, and neither the fixed effect nor the IV coefficient is statistically signif- icant. The results are consistent with public school attendance slightly raising study hours, but are only suggestive, as the OLS results may reflect the tendency for more motivated students to select public schools. We also examine the effect of school type on completion of junior secondary school, though there is less variation in this variable, as more than 95 percent of all students who start junior secondary school complete the third grade in that school level. The effect of public school attendance on probability of completion is not sig- nificant in any of the samples, and the sign of the effect varies across samples.

VIII. Different Types of Private Schools