Are Private Schools More Effective than Public Schools? Mohamad Fahmi * Department of Economics, Universitas Padjadjaran, Indonesia July 27, 2010 Abstract I attempt to replicate the research carried out in the paper entitled "The Effectiveness of Private Versus Public Schools: Case of Indonesia" the work of Bedi and Garg (2000). Bedi and Garg (2000) find that selectivity bias in the earnings estimation reverses the superiority of public schools over private schools. To confirm the findings, I use the same sample data used by Bedi and Garg (2000) and a sample data that created from the Indonesia Family Life Survey 1. I also discuss the use of school quality proxies by Bedi and Garg (2000) that may bias their estimates of the earnings differential. My findings show that the surprising findings of Bedi and Garg (2000) are not robust and suggest that public school graduates earn significantly higher than private non-religious school graduates and imply that the quality of public schools are better than private non-religious schools. JEL classification: J31 Keywords: School effectiveness; Earnings; Indonesia 1. Introduction While it is generally accepted that public secondary schools are at the highest quality in In- donesia (Strauss et al., 2004; Newhouse and Beegle, 2006), Bedi and Garg (2000) find that workers who attended private non religious schools experience a 75 per cent earnings advan- tage over graduates of public schools. Without correcting for selectivity bias, public school * e-mail: [email protected]. 1
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Are Private Schools More Effective than PublicSchools?
Mohamad Fahmi∗
Department of Economics, Universitas Padjadjaran, Indonesia
July 27, 2010
Abstract
I attempt to replicate the research carried out in the paper entitled "The Effectiveness ofPrivate Versus Public Schools: Case of Indonesia" the work of Bedi and Garg (2000). Bediand Garg (2000) find that selectivity bias in the earnings estimation reverses the superiorityof public schools over private schools. To confirm the findings, I use the same sample dataused by Bedi and Garg (2000) and a sample data that created from the Indonesia FamilyLife Survey 1. I also discuss the use of school quality proxies by Bedi and Garg (2000) thatmay bias their estimates of the earnings differential. My findings show that the surprisingfindings of Bedi and Garg (2000) are not robust and suggest that public school graduatesearn significantly higher than private non-religious school graduates and imply that thequality of public schools are better than private non-religious schools.
JEL classification: J31Keywords: School effectiveness; Earnings; Indonesia
1. Introduction
While it is generally accepted that public secondary schools are at the highest quality in In-
donesia (Strauss et al., 2004; Newhouse and Beegle, 2006), Bedi and Garg (2000) find that
workers who attended private non religious schools experience a 75 per cent earnings advan-
tage over graduates of public schools. Without correcting for selectivity bias, public school
graduates are found to earn 31 per cent more over private non religious graduates. However,
a negative selection effect for private non religious schools is identified and correcting for this
selection effect results in a large earnings premium of private non religious schools over public
schools. Despite private non religious school graduates generally having lower academic qual-
ifications than public school graduates, it is argued that the private non-religious schools are
more effective than the public schools. Bedi and Garg’s (2000) finding supports Hannaway’s
(1991) claim, “that private schools perform better due to greater school level autonomy and
their responsiveness to the needs of students and parents”. The policy implication of these
findings is to encourage a greater private sector role in Indonesian education since the results
suggest the private sector is more efficient and effective in delivering education.
Along with Bedi and Garg (2000), Newhouse and Beegle (2006) investigate the effective-
ness of private and public lower secondary schools in the Indonesian context, focus on the
relationship between school choice and academic performance rather than school choice and
future earnings. Newhouse and Beegle (2006) found the academic performance of public lower
secondary school students was superior to private school students as measured by national fi-
nal test exam scores (UN1) upon completion of lower secondary school. An important aspect
of the findings by Newhouse and Beegle (2006) is that they are difficult to reconcile with the
earlier findings by Bedi and Garg (2000). In particular, Newhouse and Beegle (2006) suggest
that the resource advantages of public schools are unlikely to be outweighed by any efficiency
benefits of private schools.
In this paper, I re-examine the earnings differential between public and private lower sec-
ondary school students, originally studied by Bedi and Garg (2000). Despite careful attempt to
replicate these earlier results, I am unable to do so. While Bedi and Garg (2000) use the first
wave of the Indonesia Family Life Survey (IFLS1) that issued in 1996 (DRU-1195-CD), my
sample data are obtained from the re-release version of IFLS1 (IFLS1-RR).
In order to explain the differences, I discuss the use of school quality proxies by Bedi
1UN or Ujian Nasional is the newest system of national centralized final examination. In 2008, IndonesianGovernment increased the standard average passing grade from minimum 5.25 to 5.50 for 6 subjects.
2
and Garg (2000). Several variables which identify the condition of the last school attended
confound lower and upper secondary schools. The use of such proxies may bias their estimates
of the earnings differential.
Using the re-release sample data of IFLS1 and with an absence of school quality indicators,
my findings suggest that the surprising findings of Bedi and Garg (2000) are not robust. My
findings suggest that public school graduates earn significantly higher than private non reli-
gious school graduates and imply that the quality of public schools are better than private non
religious schools.
In the next section, I try to replicate Bedi and Garg (2000) data sample. The following
section introduces the empirical strategy to re-estimate the effect of lower secondary schools
quality on earnings differential. Section 4. provides the results of school choice and earnings
decomposition estimates, while Section 5. concludes the paper.
2. Sample Replication
In order to replicate the result of Bedi and Garg (2000), I first try to create an identical dataset,
using the Indonesia Family Life Survey 1 (IFLS1) 1993. The IFLS1 is a large-scale lon-
gitudinal survey of socio economic and health status of the individual and household level.
The IFLS1 sampling scheme was based on provinces, then the sample was randomly selected
within these provinces. For cost-effectiveness reasons the survey focused on only 13 out of
26 provinces on the Island of Java, Sumatra, Bali, West Nusa Tenggara, Kalimantan, and Su-
lawesi. These were selected to represent approximately 83 per cent of the Indonesian popu-
lation. While waves 2, 3, and 4 of the IFLS have been released between 1997 and 2009, the
analysis in Bedi and Garg (2000) focuses on IFLS1. I assume that their study commenced
when only IFLS1 was available, and while IFLS2 was released in 1997, it did not contain
employment data required to extend the analysis.
I created a sample data based on Bedi and Garg’s (2000) guidance (pages 467-468). Fol-
3
lowing the guidance in pages 467-468 of Bedi and Garg (2000), I attempt to replicate their
sample. However, I was unable to exactly reproduce their sample. My initial sample data set
consisted of 7220 respondents who have earnings and are no longer students. The size of the
initial data was almost twice the size of Bedi and Garg’s (2000) initial sample data of 4900
observations. Missing and miscoded data and also sample restrictions then reduced the data set
by 6170 (more than 85 per cent) to 1050 observations. Most of the observations, 5448, were
dropped as respondents had not proceeded beyond primary school, while 274 observations
were dropped since respondents had more than 12 years of education. Moreover, I dropped 13
respondents due to missing information on school type and 9 observations as they had either
99997 or 999997 on total monthly earnings. Finally, I exclude a further 389 observations as
they had some missing information, miscoded class size (41 observations), number of months
in school period per year (45), failed in primary school (1), parents’ education (294), province
where school is located (6) and religion (2). Table 1 presents the full comparison of the exclu-
sion process with results of Bedi and Garg (2000).
4
Table 1: Comparison of Exclusion Process
Item Bedi and Fahmi*
Garg (2000)
Initial income information 4900 7220
Had not proceeded beyond primary education 3391 5448
Had more than 12 years of education 291 274
Lack of information on hours of work 33 37
Missing information on school type 10 13
Reported incomes seemed implausibly high 3 9
Missing information on class size - 41
Attend(ed) school more than 12 months (miscoded) - 45
Missing information on failed in primary school - 1
Missing information on father’s education - 214
Missing information on mother’s education - 80
Missing information on school location - 6
Missing information on religion - 2
Number of remaining observation 1194 1050
5
In order to understand the differences between my sample and the sample extracted by Bedi
and Garg (2000) it is important to note that Bedi and Garg (2000) used the IFLS1 issued by
RAND in 1996 (DRU-1195-CD), while I used the IFLS1 data set called IFLS1-RR (re-release
in 2000) that updates the original IFLS1. Peterson (2000) explains that IFLS1-RR revises and
restructures the original IFLS1 to accompany with IFLS2. The different between IFLS1 DRU-
1195-CD and IFLS1-RR maybe the source of the discrepancy between my sample and that
of Bedi and Garg (2000). Arjun Bedi kindly sent the sample data set, PUBPRIV.DTA2. Bedi
and Garg (2000)create the file on 7 February 1998 which consists of 1527 observations and
231 variables. However, they were unable to provide the code used to construct the sample,
making it impossible to clearly identify the sources of the discrepancies.
2The file PUBPRIV.dta originally consists of 1,527 observations, 231 variables. It is created on 7 February1998. The exclusion of some observations of missing data on earnings information drops the final sample data setto 1194.
6
Table 2: Tracking Process of Mismatch Sample Data
No. Note
Obs.
745 Identical
17 Unidentified
152 Had more than 12 years’ education.
34 - Missing information on period in school in months.
- Bedi and Garg (2000) substitute the missing data by sample mean.
32 - Missing information on class size.
- Bedi and Garg (2000) substitute the missing data by sample mean.
154 - Missing information on father’s education.
- Bedi and Garg (2000) put "0" instead of missing values in three dummy
variables of father’s of education.
- Three variables of father’s education are FATH_PRI, FATH_JH and
FATH_SH.
60 - Missing information on mother’s education.
- Bedi and Garg (2000) put "0" instead of missing values in two dummy vari-
ables of mother’s education.
- Two variables of mother’s education are MOTH_PRI and MOTH_SEC.
I was able to compare my sample (1050) with Bedi and Garg’s (2000) sample (1194) using
survey identification code. It is possible to match 745 respondents precisely. Of the remaining
449 observations, 17 observations are unidentified and 432 are considered as missing informa-
tion. Conversely, my sample contained 305 observations that were not in the sample provided
by Bedi and Garg (2000).
Of the 305 observations with missing data, 34 observations have no information on the
number of months per year of attending school and 32 observations have no information on
7
class size. Bedi and Garg (2000) substitute for the missing data on those observations by using a
sample mean instead of dropping the number of observations. The remaining 214 observations
have no information on either father’s or mother’s education. Bedi and Garg (2000) put "0"
value on those observations rather than dropping them. While the exclusion process is clearly
documented, the substitution process on the 305 observations is not explained in the paper. I
provide the details of my comparison in Table 2. I also present a full complete comparison of
summary statistics between the two sample in Table 3 and the description of all variables in
Table 9 in Appendix 5.. Given the differences in the sample data, I proceed to replicate their
methodology using the sample that I have extracted.
Table 3: Comparison of Descriptive Statistics
Variable Bedi and Garg (2000) Fahmi
Mean Std. Dev Mean Std. Dev
LOGEARN -0.202 1.079 -0.290 1.063
EARN 1.492 2.567 2.030 17.655
AGE 34.66 7.502 34.264 7.321
JUNIOR 0.307 0.462 0.415 0.493
SENIOR 0.521 0.499 0.527 0.500
MALE 0.672 0.469 0.689 0.463
BAHASA 0.404 0.491 0.370 0.483
HIN_BUD 0.066 0.248 0.074 0.262
CHRIST 0.091 0.289 0.092 0.290
PRI_FAIL 0.204 0.403 0.208 0.406
SCHOLAR 0.048 0.215 0.040 0.196
FATH_PRI 0.422 0.494 0.521 0.500
FATH_JH 0.101 0.302 0.113 0.317
FATH_SH 0.085 0.279 0.084 0.277
MOTH_PRI 0.380 0.485 0.470 0.499
Continued on Next Page. . .
8
Table 3 – Continued
Variable Bedi and Garg (2000) Fahmi
Mean Std. Dev Mean Std. Dev
MOTH_SEC 0.109 0.312 0.094 0.292
DIRT FLOOR 0.067 0.251 0.044 0.205
CLASS SIZE 36.47 9.301 36.651 8.884
MONTHS 9.459 1.849 9.638 1.710
OTH_PR 0.023 0.148 0.031 0.175
SKALI_ED 0.043 0.204 0.036 0.187
NSUMA_ED 0.106 0.308 0.097 0.296
WSUMA_ED 0.068 0.253 0.049 0.215
SSUMA_ED 0.051 0.220 0.052 0.223
LAMP_ED 0.023 0.151 0.027 0.161
EJAVA_ED 0.120 0.325 0.135 0.342
WJAVA_ED 0.139 0.346 0.131 0.338
CJAVA_ED 0.141 0.348 0.155 0.362
BALI_ED 0.048 0.215 0.058 0.234
NTB_ED 0.042 0.200 0.056 0.230
YOGYA_ED 0.067 0.251 0.065 0.246
SSULA_ED 0.042 0.202 0.038 0.192
JAKAR_ED 0.079 0.270 0.069 0.253
URBAN 0.708 0.455 0.670 0.470
SKALMNT 0.043 0.204 0.050 0.219
NSUMATRA 0.098 0.297 0.084 0.277
WSUMATRA 0.066 0.250 0.045 0.207
SSUMATRA 0.053 0.225 0.057 0.232
EJAVA 0.103 0.304 0.117 0.322
WJAVA 0.131 0.338 0.125 0.331
CJAVA 0.088 0.284 0.098 0.298
BALI 0.054 0.226 0.068 0.251
Continued on Next Page. . .
9
Table 3 – Continued
Variable Bedi and Garg (2000) Fahmi
Mean Std. Dev Mean Std. Dev
NTB 0.042 0.202 0.057 0.232
LAMPUNG 0.029 0.168 0.034 0.182
YOGKARTA 0.067 0.251 0.065 0.246
SSULAWES 0.042 0.202 0.040 0.196
JAKARTA 0.176 0.381 0.160 0.367
Number of Sample 1194 1050
3. Estimation Procedure and Sample
I use a two-step earnings estimate with selection bias correction and Blinder-Oaxaca Decom-
positions to determine the earnings differential between public school group and private school
groups. The two step earnings estimates that is corrected for selection bias problem is based
on the technique that developed by Lee (1983). I follow Lee (1983) to employ an unordered
multinomial logit (MNL) model in obtaining the selection correction terms and estimating the
lower secondary school choice.
To determine the effect of school quality on earnings differential, I follow Bedi and Garg
(2000) to estimate four separate earnings estimates: public, private non religious, private Is-
lamic and private Christian. According to Kingdon (1996) this method is able to avoid endo-
geneity problem. The earnings determination of individual i who attended school type j may
be written as
10
Yij = βjXij + eij (1)
where Y is earnings, β is parameters of exogenous variable X that consists of personal and
family characteristics, and u is the error terms.
The inclusion of selection correction terms to overcome the selection bias problem modifies
the equation 1 to
Yij = βjXij + θjλij + ηij (2)
where
λij =φ(Hij)
Φ(Hij
(3)
and
Hij = Φ−1(Pij) (4)
θ is the coefficient on inverse Mills ratio λij , ηij is error terms, φ(Hij is the standard normal
density function, Φ(Hij is the normal distribution and Pij is probability of individual i chooses
the type school j.
I follow Bedi and Garg (2000) using the Blinder-Oaxaca decomposition to estimate earn-
ings differential between public school and private school graduates. Bedi and Garg (2000)
use the two-fold decomposition that included some non-discriminatory coefficient vectors to
determine the contribution of the gap in the predictors. According to Reimers (1983), the two