The effects of private schooling on pupil achievement: a global systemic analysis

ABSTRACT

Globally, the private school share of enrollment increased from about 14 percent in 2000 to about 18 percent in 2019. We estimate the systemic effect of private enrollment share on learning outcomes. Estimates of the systemic effect of private school enrollment capture any competitive effects as well as any differences between public and private schools in raising student outcomes. We use new data from the World Bank on harmonised learning outcomes for mathematics, reading, and science to produce an unbalanced sample of 120 countries from 2000 to 2017. We find that, all else equal, on average, a one percentage point increase in private enrollment is associated with null to at most weakly positive effects on country-level learning outcomes. Countries that increased private school enrollment experience as much or slightly more learning than countries with no change in private school enrollment.

摘要

在全球范围内，私立学校的入学份额从2000年的14%左右增至2019年的18%左右。我们估算了私立学校入学份额对学习成绩的系统性影响。这一估算反映公立学校和私立学校在提高学生成绩方面的竞争效应及差异。我们使用世界银行关于数学、阅读和科学三个科目学习成绩经过统一后的新数据，得出 2000年至2017年间120个国家的非平衡样本。我们发现，在其他条件相同的情况下，平均而言，私立学校入学份额提高一个百分点，对国家层面学习成绩有零到至多微弱的积极影响。与私立学校入学份额没有变化的国家相比，私立学校入学份额增加的国家的学习成绩与之相当或略有提高。

KEYWORDS:

• Private school
• harmonised learning outcomes
• test scores
• country

关键词:

• 私立学校
• 统一学习成果
• 考试分数
• 国家

Introduction

Over the last half century governments around the globe have invested in improving the quality of education in public schools. For example, the United States tripled inflation-adjusted expenditures per pupil over forty years (Hanushek 2006), and per-pupil expenditures continue to grow (U. S. Department of Education 2021). Internationally, enrollment in the public-school sector significantly increased during the 19th and 20th centuries due to government efforts and state-subsidised education (Nishimura and Ogawa 2009; Ramirez and Boli 1987). Schooling around the globe, however, is privatising with increased enrollment in private schools, with varying degrees of government oversight and public funding support (Lewis and Patrinos 2011). Globally, the share of school children enrolled in private schools has increased by 28% since 2000 (Baum et al. 2014; Glewwe and Muralidharan 2016; World Bank 2021). Most of the increase in private school enrollment has come from developing countries in Asia and Africa where public schools are free and parents pay tuition to send their children to private schools (Dixon et al. 2019; Tooley, 2009). In some countries, such as India (Sarkar and Cravens 2022) private schools receive public funding for disadvantaged students; in India, about half of urban enrollment and a fifth of rural enrollment were in private schools in 2014 and 2015 (Kingdon 2020). UNESCO data from 2020 show private school share of about a fifth in low- and middle-income countries (Crawfurd, Hares, and Todd 2023).

Proponents of private schooling claim positive effects on learning due to competition among public and private schools, a better match between schools and family’s needs, or both (Chubb and Moe 1990; DeAngelis and Erickson 2018). Opponents claim that any positive effects are due to student selection (Akmal, Crawfurd, and Hares 2019; Somers, McEwan, and Willms 2004). Research finds conflicting results from large scale publicly-funded vouchers for private schools. In Colombia, children winning a voucher to attend private school scored higher on achievement tests and were more likely to finish eighth grade than those not winning a voucher (Angrist et al. 2002). In New Orleans, students winning a voucher to attend private school scored lower on mathematics assessments and had similar rates of college-going compared to those who did not win a voucher (Erickson, Mills, and Wolf 2021).

Few studies have estimated the association between private school share and student achievement at the country-level. A handful of studies use cross-sectional data to estimate this systemic effect. Woessmann (2003) found a positive correlation on average between the share of private enrollment and mathematics and science performance in a sample of 39 countries¹ on the 1995 Trends in International Mathematics and Science Study (TIMSS). Schuetz (2009) found positive association between the share of private enrollment and mathematics outcomes for 28 countries (22 of these countries are from the OECD)² in the 2003 Program for International Student Assessment (PISA). The effect may differ for students with different socioeconomic backgrounds. In a sample of 14 mostly high-income countries using PISA and Progress in International Reading Literacy Study (PIRLS) reading outcomes, Ammermueller (2013) found that larger shares of private school enrollment are associated with larger SES test score gaps.³

Countries may have more private schools for reasons correlated with overall student outcomes such as when the public schools are low quality or when parents are willing to invest more in their children’s human capital. West and Woessmann (2010), D’Agostino (2016), and Heller-Sahlgren (2018) use instrumental variables to address the bias concern caused by students selecting into private schools due to public school quality, parental investment in human capital, or both. These studies use the historical share of Catholics in 1900 as an instrument for private school share.⁴ West and Woessmann (2010) and Heller-Sahlgren (2018) found positive effects of the share of private schools on PISA test scores in 2003 and 2012 respectively. Using PISA 2012 scores, D’Agostino (2016) found no statistically significant differences in scores between private and public schools in 35 countries, 29 of which are from the OECD.

One study used panel-data to estimate systemic effects of private school enrollment. DeAngelis (2019) pooled PISA mathematics, reading, and science scores between 2000 and 2012 for 63 countries and used country-fixed effects and IV methods. DeAngelis uses demographic changes as an instrument for percent private school enrollment. He found positive effects of the private share of schooling on test scores, where only the reading result remained statistically significant in the IV analysis.

We expand on this literature using new, globally representative data on student learning outcomes from the Harmonized Learning Outcomes (HLO) database. We estimate a general equilibrium, systemic effect of private school enrollment on student learning on country-level data with country fixed effects. One advantage of estimating this systemic effect is that these estimates include any effects of private schools on enrolled students and any competitive effects of private schools on public school students while avoiding concerns about selection into different school sectors. Our data allows for country-fixed effects while expanding the countries analysed in DeAngelis (2019). Our sample comprises 98% of the global population, which is an improvement upon the aforementioned studies which largely comprise of middle- and high-income countries, and are restricted to certain regions, and to a few years based on a few assessments. We are also able to assess whether effects differ by region, assessments, gender, and income levels of nations. With this comprehensive data our study can provide a benchmark for future studies for average systemic effects from the global expansion of private schooling.

Our estimates range from null to at most weakly positive effects. We conclude that, all else equal, private schools are as good as or, at most, weakly better than public schools around the globe in improving learning. We generally do not observe large differences between boys and girls or between types of countries and regions. Parental selection of schools and switching between school sectors may equalise learning outcomes across school sectors.

The remainder of the paper appears in the following order. We first review the importance of test scores in country-level well-being. We then review the literature on private schooling and student achievement. We present our hypotheses; describe the data, methodology, and results; and then conclude.

Importance of test scores in country-level well-being

United Nations development goal #4 targets both enrollment in schooling at all levels as well as student achievement of minimum proficiency levels on assessments (United Nations 2023). Some scholars suggest that standardised tests provide imperfectly reliable or valid information about student abilities (Jacob and Levitt 2003; Koretz 2017). Other scholars show that student test scores predict individual economic outcomes such as college attainment, future earnings, teenage pregnancy rates, physical and mental health, and voting behavior (Borghans et al. 2016; Chetty, Friedman, and Rockoff 2014) as well as similar outcomes at the societal level (Hanushek and Woessmann 2012). Nations with higher student test scores progress faster than nations with lower scores (Hanushek and Woessmann 2008, 2012) although note that scholars disagree (Komatsu and Rappleye 2017).

Since the 1960s, there has been a rise in international assessments which assess student achievement across nations such as PISA and TIMSS. Such assessments test students on a common framework of questions, most of which comprise of mathematics, reading, and science items. Verger, Parcerisa, and Fontdevila (2019) compile a database of countries’ adoption and analysis of international large-scale assessments and suggest that national cross-country assessments are a key ingredient of the global education reform movement based on accountability (enforced through administrative incentives and sanctions, and school choice), standards (enforced through quality standards and curriculum), and decentralisation (enforced though national, state, local, and school-level governance). The outcomes from these tests have drawn the interest of governments, international aid agencies, educators, researchers, and the public. These assessments may capture imperfect information on learning in a country if schools can choose whether to participate in the assessments; to the extent that these assessments reflect average learning outcomes in a country, test scores provide useful information to compare how private school enrollment over time correlates with student learning outcomes.

International evidence on private schooling and student outcomes

A broad literature uses international data to estimate performance differences between public and private schools, with mixed results. Toma (1996) analysed the 1981 mathematics study conducted by the International Association for the Evaluation of Educational Achievement (IEA) for Belgium, France, New Zealand, Ontario (Canada), and the United States. Fuchs and Woessmann (2007) examined mathematics, reading, and science assessments in PISA 2000. Both studies estimated positive effects on average for private schools. Vandenberghe and Robin (2004) analysed PISA 2000 and found positive effects in Dutch- and French-speaking Belgium and Brazil, but not in Austria, France, Ireland, Mexico, Spain, and Denmark. Sakellariou (2017) analysed 2012 PISA mathematics for 40 countries, but found positive effects only in Argentina, Belgium, Canada, Greece, Indonesia, Jordan, and Netherlands. Corten and Dronkers (2006) found positive effects for students from low socioeconomic status families in PISA 2000 data for 19 countries.⁵ Using similar PISA 2000 data, Dronkers and Robert (2003, 2008) found private independent schools to be less effective than public schools on average at improving test scores. Most of these studies focus on students in OECD countries.

When estimating differences between school sectors, we may be concerned that students from well-off socio-economic backgrounds select into private schools thereby leading to an overestimation of the private school effect on learning (French and Kingdon 2010). Using 1997 Primer Estudio Internacional Comparativo (PEIC) data for ten Latin American countries, Somers, McEwan, and Willms (2004) found that, on average, the achievement advantage of private schools was accounted for by peer group characteristics.⁶

The experimental evidence of private school vouchers on the effects of private schooling on student achievement also shows modestly positive effects (Dixon et al. 2019; Shakeel, Anderson, and Wolf 2021), with null to negative effects observed in some contexts (Erickson, Mills, and Wolf 2021).

Our data allow us to estimate separate effects for boys and girls. Some studies reveal gender gaps in private school choice contexts. For example, in Pakistan, boys are more likely to attend low-cost private schools, but girls outperform boys in learning assessments (Bizenjo 2020). In an experimental evaluation of a targeted school voucher programme in India, Dixon et al. (2019) did not find statistically significant differences in learning between boys and girls. In contrast, Angrist et al. (2002) found larger learning impacts for girls than boys in their experimental evaluation of Colombia’s PACES voucher programme.

As described above, these studies consider specific sub-samples of countries and a variety of empirical strategies, making it difficult for a lay reader to adjudicate the overall takeaways. Given the growing international focus on large-scale assessments, our study takes advantage of globally-representative, country-level data to estimate global average effects of changes in private school enrollment rates. One way countries may meet UN development goal #4, for example, is by expanding financial support for private schools instead of providing and funding elementary and secondary schooling. Our estimates provide some information as to the likely results of these efforts.

Hypotheses

Overall, the literature reviewed above shows modest benefits of the private share of schooling on student achievement in some, but not all, countries. We accordingly hypothesise that an increase in the private share of schooling at the country-level would be modestly and positively associated with student learning. However, we expect effects may differ across regions and countries. We do not expect large gender differences in private schooling’s impact.

Data

Harmonised learning outcomes (HLO)

Our analysis examines student achievement data from the Harmonized Learning Outcomes (HLO) database developed by Angrist et al. (2021). This data represents 98% of the global population, with two-thirds of the sample comprised of developing countries. It includes country-level achievement scores for 164 countries between 2000 and 2017 and it is constructed from seven assessments which are psychometrically linked and administered across multiple countries. HLO is created to be comparable across countries and years.

The seven assessments comprise the TIMSS, the PIRLS, the PISA, the Southern and Eastern Africa Consortium for Monitoring Educational Quality (SACMEQ), the Programme d’Analyse des Systèmes Éducatifs (PASEC, or Programme of Analysis of Education Systems), the Latin American Laboratory for Assessment of the Quality of Education (LLECE), and the Early Grade Reading Assessment (EGRA). TIMSS and PIRLS are administered by the International Association for the Evaluation of Educational Achievement (IEA). TIMSS assesses students in mathematics and science in grades 4 and 8, whereas PIRLS covers reading assessments in grade 4. PISA tests students in mathematics, reading and science assessments at age 15 and it is administered by the OECD, a 37-member international organisation of industrialised countries that collects statistics and information on the economic and social well-being of member countries. SACMEQ assesses students in mathematics, reading, and English in grade 6 and it is administered by a consortium of 16 countries in the Southern and Eastern African region. The Conference of Ministers of Education of French-Speaking Countries (CONFEMEN) administers PASEC, and students are assessed in mathematics and French in grades 2 and 5. LLECE is administered by the UNESCO Regional Bureau for Education in Latin America and the Caribbean. The assessments include reading and mathematics in primary school (grades 3-6). EGRA is implemented by the United States Agency for International Development (USAID), Research Triangle Institute (RTI) and their local partners. Students are generally assessed in basic literacy assessment in early grades (usually grades 2-4). The seven surveys differ in their administration, scope, purpose, sampling frames, test content, grades/ages and subjects assessed. Together they provide reliable and globally representative information on student learning as measured by mathematics, reading, and science assessments.

The cross-country data are linked, and the country-level scores are harmonised on a common scale. The database provides mean scores and standard errors for each measure, and the achievement scores are disaggregated by schooling level (primary and secondary), subject (mathematics, reading, and science) and gender (male and female). We examine HLO at the primary level, chiefly because the private school information is more available for primary enrollment. However, in Appendix Table A1, we examine the HLO data with the smaller secondary school sample. The estimates suggest a statistically null relationship between the percent enrolled in private school at the secondary level and learning outcomes. HLO is provided for three subjects: science, mathematics, and reading. We present each subject separately in the analysis below. The sample sizes differ depending on the subject examined.

World bank and UNESCO indicators

The World Bank (WB) and UNESCO provide similar data on private school enrollment. We use the World Bank’s World Development Index (WDI) measure of private school enrollment and fill in missing observations, when possible, with private school enrollment data from UNESCO. Private school enrollment is measured by the enrollment in private primary schools divided by the total enrollment in primary school. The World Bank compiles all of the economic, demographic, and social variables in its WDI data from ‘officially-recognized international sources’ and ‘presents the most current and accurate global development data available.’⁷ Regions differ significantly in their private school penetration. For example, between 2000 and 2017, Europe & Central Asia average 5 percent, Latin America & the Caribbean 25 percent, and Sub-Saharan Africa 15 percent.

For our sample with the largest number of observations (N = 291), the sample comprises of 120 countries with at least one year of data. This sample includes 30 countries with only one year of data; 37 with two years of data; 28 with three years of data; 23 with four years of data; one with five years of data; and one with six years of data. By region, the sample includes 26 observations in East Asia & Pacific; 89 in Europe and Central Asia; 44 in Latin America and Caribbean; 31 in Middle East and North Africa; 7 in North America; 7 in South Asia; and 87 in Sub-Saharan Africa.

Table 1 presents summary statistics. In our sample, private school enrollment ranges from zero (Kenya 2000) to 88 percent (Zimbabwe 2013). Field work by some researchers suggests that private schools are undercounted in developing nations (Dixon et al. 2019; Härmä 2019; Tooley, 2009). Thus, the WB and UNESCO data likely represent a lower-bound for our independent variable of interest.

Table 1. Summary statistics.

Variable	Obs	Mean	Std. Dev.	Min	Max
HLO
science	167	478.19	70.32	263.65	590.48
mathematics	215	459.62	77.44	295.97	617.67
reading	291	432.49	95.77	229.18	580.77
percent private	291	13.94	17.47	0.00	87.77
ln(real GDP pc)	291	8.60	1.55	5.68	11.57
polity2	291	5.11	5.79	−10	10
mortality rate 5–9 y.o. per 100,000	291	4.49	5.79	0.30	30.70

Note: Table displays summary statistics for the variables analysed in this paper.

In the analysis, we control for GDP per capita, the mortality rate of children, and a measure of political regime stability. Countries with higher GDP per capita may invest more resources in their children, including in their human capital; shorter lives mean fewer years to recoup any investment in human capital. Political stability may affect families’ willingness or ability to send their children to school and their trust in any public education system. The WDI provides data on real, purchasing-power-adjusted GDP per capita and the mortality rate per 100,000 for 5- to 9-year-olds. Real GDP per capita averages $14,619 in our sample with a range from $294 (Malawi 2000) to $105,572 (Luxembourg 2006). This range reflects the wide variety of countries in the sample. Child mortality averages 4.5 per 100,000 and ranges from 0.3 per 100,000 (Denmark 2016) to 30.7 per 100,000 (Niger 2002). The Center for Systemic Peace quantifies political regime characteristics with a measure named polity2. The polity2 variable ranges from hereditary monarchy (−10; for example, Qatar 2006) to consolidated democracy (+10; for example, Canada 2016).

Methodology

This study estimates the general equilibrium effects of private school penetration on student learning. By using country-level data on student learning and private school enrollment, the estimates capture the average effect of within country changes in private schooling enrollment including any differences in value-added across sectors and competitive effects. Our focus on cross-country analysis and estimating the general equilibrium effect, avoids any concerns about individual selection into school sector.

We estimate for country c at year t the following: $\mathrm{HL}{O}_{\mathrm{ct}}={\beta }_1\mathrm{pctprivat}{e}_{\mathrm{ct}-1}+X^{\mathrm{\prime }}\gamma +{\theta }_c+{\tau }_t+{\epsilon }_{\mathrm{ct}}$Vector X controls for real, purchasing power adjusted GDP per capita, the polity2 measure of democracy, and the mortality rate per 100,000 of 5- to 9-year-olds for the reasons discussed above.

Country fixed effects, ${\theta }_c$, account for time-invariant cultural differences across countries; these include time-invariant characteristics of educational institutions (such as school size and government regulations), cultural differences, differences in family structure or parenting style, and student characteristics (such as race, gender, and parental education). The country fixed effects also control for country-wide educational policies that do not change during our sample period, 2000–2017.

The specification includes year fixed effects, ${\tau }_t$, which account for any global trends in test scores over time. Standard errors are clustered by country to account for heteroskedasticity and serial correlation within country.

Our coefficient of interest is ${\beta }_1$, the within-country effect of lagged private school enrollment rates on HLO. We lag private school enrollment by one year to allow some time for different enrollment in school sector to affect student learning. The coefficient captures the net effect of increasing enrollment in private school including any change in the effectiveness of schooling as well as any competitive effects of increased enrollment in the private sector.

Results

Table 2 presents estimates of the above regression equation by subject. Panel A presents estimates for science, panel B for mathematics, and panel C for reading. Sample sizes differ across subjects.

Table 2. HLO and percent private.

	(1)	(2)	(3)	(4)	(5)	(6)
Panel A: Science
% privatet-1			1.106	1.033	1.008	0.476*
			(0.957)	(0.984)	(1.025)	(0.279)
ln(real gdp pc)	31.87***	42.78**		38.10**	37.06*	82.55***
	(9.312)	(16.77)		(18.41)	(19.41)	(26.54)
polity2					0.146	−1.578
					(1.819)	(1.562)
Mortality rate 5–9 y.o.					−2.649	6.706
					(10.85)	(9.877)
Observations	162	162	162	162	162	162
Adjusted R-squared	0.308	0.937	0.939	0.940	0.939	0.973
Panel B: Mathematics
% privatet-1			1.283	1.305	1.312	0.875**
			(0.843)	(0.841)	(0.832)	(0.382)
ln(real gdp pc)	27.09***	−8.396		−13.12	−5.144	−2.316
	(5.743)	(20.70)		(20.06)	(18.51)	(18.46)
polity2					−1.761	−2.235***
					(1.214)	(0.668)
Mortality rate 5–9 y.o.					6.152*	4.544*
					(3.115)	(2.395)
Observations	204	204	204	204	204	204
Adjusted R-squared	0.410	0.946	0.950	0.950	0.953	0.977
Panel C: Reading
% privatet-1			0.387	0.382	0.350	−0.427
			(0.408)	(0.433)	(0.438)	(0.846)
ln(real gdp pc)	40.19***	4.449		2.907	4.803	72.34**
	(3.809)	(23.50)		(23.73)	(22.74)	(32.87)
polity2					−0.643	−0.304
					(1.235)	(0.988)
Mortality rate 5–9 y.o.					4.302	−1.399
					(2.635)	(3.388)
Observations	291	291	291	291	291	291
Adjusted R-squared	0.634	0.956	0.956	0.956	0.957	0.975
Country fixed effects	No	Yes	Yes	Yes	Yes	Yes
Weighted by population	No	No	No	No	No	Yes

Note: All regressions also include year fixed effects. Standard errors clustered by country in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

We first regress the HLO on only logged real GDP and year fixed effects. The country-fixed effects account for important time-invariant differences across countries but also consume many degrees of freedom. We present column (1) to illustrate the role of country fixed effects. This specification displays the expected correlation that HLO is higher in countries with higher GDP per capita. Column (2) onwards adds country fixed effects. We observe that as a country’s GDP per capita increases, science HLO increases; the relationships for mathematics and reading are not statistically significant. In column (3), the specification includes the percent private, year fixed effects, and country fixed effects. For all three subjects, percent private is positively and not statistically significantly related to HLO.

We add real GDP per capita in column (4). In column (5), we add polity2 and the mortality rate of 5- to 9-year-olds as control variables. We continue to find a positive and not statistically significant relationship between percent private and HLO. Adding polity2 and the mortality rate of elementary school aged children does not impact the estimated coefficients on percent private.

In the complete specification in column (6), we weight the regressions by total country population. Weighting the observations by population accounts for the greater variance in test scores in smaller countries. Further, it generates effects that are an average over the population instead of an average over countries. The weighted estimate on percent private is statistically significant for mathematics and science, but null for reading; as expected, weighting reduces the standard errors, increasing precision. To interpret these effects, we divide the estimates by the standard deviation of the subjects in Table 1. These estimates imply that, all else equal, on average, a one percentage point increase in the percent enrolled in private school at the primary level is associated with an increase in science HLO by about 0.476/70.32 = 0.007 standard deviations, an increase in mathematics HLO by about 0.875/77.44 = 0.011 standard deviations, and a decrease in reading HLO by about −0.427/95.77 = −0.004 standard deviations at the country-level; only the former two are statistically significant. To convert these numbers to weeks of learning, we use 0.31 standard deviation as the average annual learning between grade 4 and 8 on the 2005–2017 NAEP test (see Shakeel and Peterson 2020, 612–613). HLO outcomes are on the TIMSS scale, and TIMSS items are fairly aligned with NAEP items, as both tests are tied to the curriculum. We also assume that learning takes place only 40 weeks per year with a 12-week break. Under these assumptions, a percentage point increase in percent private enrolled at the primary level is associated with an increase in science HLO by (0.007/.31) $\times$ 40 = 0.90 weeks, and an increase in mathematics HLO by (0.011/0.31) $\times$ 40 = 1.42 weeks at the country-level.

Table 3 presents separate estimates for boys and girls for each subject using the specification from column (6) of Table 2 (weighted with all controls). The coefficient on percent private exhibits at most a very modest advantage for boys in science, and for girls in mathematics. Reading estimates are similar between boys and girls.

Table 3. HLO and percent private by gender.

	(1)	(2)	(3)	(4)	(5)	(6)
	Boys			Girls
	science	mathematics	reading	science	mathematics	reading
% privatet-1	0.531*	0.679**	−0.655	0.404	0.901**	−0.413
	(0.273)	(0.291)	(0.954)	(0.294)	(0.352)	(0.858)
ln(real gdp pc)	76.15**	−5.440	78.26***	90.83***	17.96	86.64***
	(30.87)	(18.85)	(27.69)	(22.77)	(22.17)	(32.99)
polity2	−2.088	−2.535***	−0.552	−0.811	−1.906***	−0.689
	(1.684)	(0.657)	(0.808)	(1.383)	(0.670)	(1.070)
Mortality rate 5–9 y.o.	6.840	6.023***	0.242	7.007	5.140*	0.0113
	(11.22)	(1.639)	(2.990)	(8.696)	(2.787)	(3.333)
Observations	162	194	278	162	194	278
Adjusted R-squared	0.972	0.979	0.981	0.973	0.980	0.979

Note: Regressions also include year and country fixed effects. Observations weighted by population. Standard errors clustered by country in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

Limitations

Although linking data on different tests on a harmonised scale and using the same variable for public and private schooling across countries allows us to study their net systemic effects with global panel-data, some limitations should be considered.

First, lack of student-level data prohibits us from exploring whether the effect of private school share differs by student demographics such as socioeconomic status. Second, combining tests on a harmonised scale comes with challenges. Shakeel and Peterson (2022) use the principal-agent model to show that measured per-decade performance differs among five tests in the U.S. For example, in mathematics, the median estimate of per-decade student progress ranges between a decline of 0.10 standard deviation per decade on the PISA test to a rise of 0.27 standard deviation on the National Assessment of Educational Progress (NAEP) test: a range of 0.37 standard deviation difference per decade. Moreover, content-analysis showed that PISA mathematics is not comparable to mathematics in NAEP and TIMSS. Shakeel and Peterson (2022) conclude that combining data across tests on a common harmonised scale across grades/age, subjects, time, and testing agencies is not without problems. Kamens and Benavot (2011) point out that regional assessments cater to the socio-cultural, economic, and technical/cost-related context for the countries administering these tests. Consequently, they conclude that regional tests are unfit for comparison with other countries.

Third, the definition of private and public sectors may not be strictly comparable both within and across countries. Public and private schools and systems differ with respect to accountability, standards, decentralisation, government regulation and support, admission and exit criteria, and how public-private partnerships in schooling are defined across and within countries (Hogan, Thompson, and Mockler 2022; Lubienski et al. 2022). The framework of religious instruction and school choice also differs across contexts. Schools operating in choice programmes likely differ from those that don’t participate in such programmes. Due to lack of cross-country panel data on such variables, we are unable to account for potential differences in private schooling across and within countries, including private tutoring.

These limitations suggest that our estimated effects may appear more homogenous than in reality. Thus, we next test for potential heterogeneities in effects allowed by our data.

Heterogeneities

In Table 4, we test whether private school share effects differ by OECD; country-income quartile; geographic region; and administered test. We use the same specification as Table 3: weighted with all controls.

Table 4. HLO and percent private by country-types.

	(1)	(2)	(3)
	science	mathematics	reading
Panel A: By OECD and non-OECD
% privatet-1	1.646	0.752	−1.497
	(1.278)	(0.862)	(1.716)
% privatet-1*OECD	−1.333	0.169	1.585
	(1.367)	(1.055)	(1.711)
Observations	162	204	291
Adjusted R-squared	0.973	0.977	0.975
Panel B: By income category
% privatet-1	0.388	0.768**	−0.229
	(0.247)	(0.352)	(0.610)
% privatet-1*low or middle lower income	1.528	0.633	−1.027
	(1.528)	(0.862)	(2.597)
Observations	162	204	291
Adjusted R-squared	0.973	0.977	0.975
Panel C: By region
% privatet-1	1.651	2.108**	−2.698
	(1.122)	(0.856)	(1.633)
% privatet-1*Europe/Central Asia	−1.322	−1.477*	2.803*
	(1.111)	(0.869)	(1.586)
% privatet-1*East Asian & Pacific	−1.264	−1.289	4.334
	(2.425)	(1.527)	(3.160)
% privatet-1*Latin America & Caribbean	−1.173	−1.066	−0.282
	(1.727)	(2.291)	(2.482)
% privatet-1*Sub-Saharan Africa		−1.228	6.221**
		(1.955)	(2.382)
Observations	162	204	291
Adjusted R-squared	0.973	0.977	0.977
Panel D: Limited to TIMSS for mathematics & science; PIRLS for reading
% privatet-1	0.468*	0.753***	0.002
	(0.275)	(0.273)	(0.351)
Observations	139	141	144
Adjusted R-squared	0.972	0.980	0.976

Note: Regressions also include year fixed effects, country fixed effects, logged real GDP per capita, mortality rate of 5- to 9-year-olds, and polity2. Observations weighted by population. Standard errors clustered by country in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

Much of the literature estimating systemic effects focuses on OECD countries. In Panel A, we allow the effect of private school penetration to differ for OECD and non-OECD countries by including an interaction term of percent private and an indicator for whether the country is a member of the OECD. The general pattern is that the effect of private schooling is positive for non-OECD countries in mathematics and reading, but not in science. However, these relationships are not statistically significant. Summing the coefficient on percent private with the interaction term produces a net effect close to zero in science and reading, but not in mathematics for OECD countries.

One possibility is that any learning differences across countries may be income-driven. In Panel B, we allow the effect of private school penetration to differ by the country’s income quartile. The World Bank categorises countries into four categories; we combine the low and lower-middle income categories as well as the upper-middle- and upper-income categories. In these specifications, we allow the effect of private school penetration to differ for low and medium-income countries in comparison to their higher-income counterparts. The coefficients on the interaction variable suggests a modest advantage in science and mathematics, but not in reading for low and medium-income countries. Yet, these results are not statistically significant. Summing up the coefficient on the interaction variable along with percent private generates a net positive effect in science and mathematics, but a net negative effect in reading.

In Panel C, we allow the effect of private school to differ by region. We interact percent private with indicators for whether the country is in Sub-Saharan Africa, Europe/Central Asia, Latin America & Caribbean, or East Asia & Pacific. We separated these regions out because our sample includes more observations from Europe & Central Asia (n = 91), Sub-Saharan African (n = 93), East Asia & Pacific (n = 34), and Latin America & Caribbean (n = 45). There are no sub-Saharan African countries in the science sample. The omitted regions are Middle East & North Africa (n = 33), North America (n = 7), and South Asia (n = 7). The negative sign on the interaction coefficient in science and mathematics suggests that private school penetration has a larger effect on HLO in countries in the omitted regions. The converse is true in reading, except for Latin America & Caribbean. Summing the coefficient on percent private with the interacted term for each geographic area provides the effect for that geographic area. We generally observe effects that are roughly zero to very modestly positive for Europe/Central Asia, East Asian & Pacific, Latin America & Caribbean, and Sub-Saharan Africa.

Much of the previous cross-country research focuses on results from a single test. To compare our results with this literature, in Panel D, we estimate the effect of private school penetration using HLOs with the most common source test for our sample: TIMSS for mathematics and science and PIRLS for reading; both tests are administered by the IEA. The results are similar to the full-sample results in Table 2. We continue to observe statistically significant effect of percent private on HLO in science and mathematics, and null effects in reading.

Robustness checks

One concern with estimating the effect of private school penetration on test scores is reverse causality. If, for example, parents respond to low performing public schools by opening or enrolling their children in private schools, we may see more private school penetration in areas with low average HLO scores. One way that previous work has addressed this concern is with instrumental variables. West and Woessmann (2010) use the historical share of Catholics as an instrument for private school share; this variable, however, prevents any use of the panel nature of our data and somewhat restricts the countries potentially included in the sample. DeAngelis (2019) used the percent of the population aged 0–14 as an instrumental variable for the percent private. As the fraction of the population under 15 grows, private schools may be less able to quickly respond to the resulting increase in demand; we expect a growing share of the younger population will correlate with a smaller share of private school enrollment.

We experiment with using the percentage aged 0–14 as an instrumental variable. Table 5 presents these IV estimates. The first stage estimates appear in the top panel. A growing younger population reduces the percent enrolled in private school, consistent with the private sector being capacity constrained. The percent aged 0–14 is a weak instrument, with an F-test less than 10. Results should be interpreted cautiously. The second stage estimates appear in the bottom panel of Table 5. The estimated coefficients on percent private are statistically null.

Table 5. IV estimates using school-aged population.

	(1)	(2)	(3)
	science	mathematics	reading
First Stage (% privatet-1 as dependent variable)
% aged 0 to 14t-1	−0.318	−0.627**	−0.608**
	(0.225)	(0.306)	(0.264)
F-test	1.99	4.20	5.28
p-value	0.164	0.044	0.024
Second Stage
% privatet-1	−8.437	−1.932	0.474
	(7.243)	(3.029)	(2.015)
ln(real gdp pc)	105.8***	13.79	66.77***
	(22.85)	(19.57)	(25.49)
polity2	−0.086	−2.415***	−0.323
	(2.678)	(0.875)	(0.724)
Mortality rate 5–9 y.o.	−19.32	3.622	−1.482
	(25.44)	(2.597)	(2.460)
Observations	143	191	261
# of countries	51	69	90

Note: Regressions also include year and country fixed effects. Observations weighted by population. Standard errors clustered by country in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

Conclusion

In this paper, we analyse the effect of share of private school enrollment on country-level achievement using a new cross-country panel of 120 countries between 2000 and 2017. Using the Harmonised Learning Outcomes (Angrist et al. 2021) we find null to at most weakly positive effects of private enrollment share on achievement around the globe. We generally do not observe effect size differences by gender, OECD, country income, and region. Most estimates are weak in magnitude or lack statistical significance.

The effects show very modest gender differences in science and mathematics. Results seem more positive in the Middle East & Africa, North America, and South Asia in science and mathematics, but not in reading. We do not have information to parse out reading effects between native vs. foreign language administered tests in the HLO sample. The reading results may be affected by challenges students in Middle East & Africa, North America, and South Asia face in tests conducted in a non-native language. The null to at most modestly positive effects we observe imply that students in countries with more private school enrollment learn as much or more than when private school enrollment is lower. One possibility is that parents may move children to private schools when their public schools underperform (and vice versa), roughly equalising learning outcomes in both sectors.

Implications

A rich literature explores whether private schools are better than public schools. Some researchers contest the rise of private schooling, saying that universal basic education is better delivered through public schooling (Smith and Joshi 2016). Other scholars disagree (Tulloch, Kramer, and Overbey 2014). In this ongoing debate, the lack of globally representative panel data has complicated drawing a conclusion regarding any net advantage the private or public sector may have in improving pupil achievement levels.

Previous international comparisons have relied on a single test or relatively high-income countries like the OECD. Our results with globally representative panel data show that an increase in private school enrollment share neither comes at the expense of systemic deterioration in student achievement, nor does it produce positive and large systemic benefits on achievement. These estimates are very modest in size, although in many contexts they might have been obtained at a lower per-student cost compared to the public schools. In other contexts, elite private schools have a higher per-student cost than public schools.

Some scholars suggest that increased public funding for private schools may come at the cost of increased regulations, which may thereby blur the cross-sector differences (Zancajo, Verger, and Fontdevila 2022). Others suggest that increased government regulations of choice programmes reduce private school specialisation (DeAngelis and Burke 2017). If anything, our results suggest that both the fears of a deterioration and large promises of improvement in achievement due to private schooling are overblown. We do not find that private school enrollment share creates disproportionately large achievement differences across gender or country type (OECD, income indicator, or regions).

Our results offer an initial look into the global systemic effects of private schooling. Future exploration of this topic with individual-level data may explore whether the effects of private school share differ by student characteristics. As private schools educate a growing fraction of children around the world, better understanding of the likely consequences for human capital improvement and therefore economic growth and reduction in poverty rates remains important.

Acknowledgments

Prior versions of this study were presented at the 2022 Southern Economic Association Conference and the 2023 International School Choice & Reform Conference. We thank Albert Cheng, Aniruddha Bagchi, Benjamin Scafidi, and Patrick Wolf for their comments. We thank Harry Patrinos for answering questions about the HLO dataset and providing feedback. We own any remaining errors.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Notes on contributors

M. Danish Shakeel

M. Danish Shakeel is a professor and director of the E. G. West Centre for Education Policy at the University of Buckingham, UK.

Angela K. Dills

Angela K. Dills is Professor of Economics and the Gimelstob-Landry Distinguished Professor of Regional Economic Development at Western Carolina University.

Notes

1 Australia, Austria, Canada, Colombia, Cyprus, Czech Republic, Denmark, England, Flemish Belgium, France,

French Belgium, Germany, Greece, Hong Kong, Hungary, Iceland, Iran, Ireland, Israel, Japan, Korea, Kuwait, Latvia, Lithuania, Netherlands, New Zealand, Norway, Portugal, Romania, Russian Federation, Scotland, Singapore, Slovak Republic, Slovenia, Spain, Sweden, Switzerland, Thailand and the US.

2 https://www.oecd.org/

3 Those countries are Canada, the Czech Republic, England, France, Germany, Greece, Hungary, Iceland, Italy, Latvia, New Zealand, Norway, Russia, and Sweden.

4 IV is a statistical method to address bias in OLS estimates. To produce consistent estimates of the coefficient of interest, the researcher needs a variable (the instrument, here it is historical percent Catholic) that is correlated with the variable of interest (here, it is percent private school enrollment) but does not directly affect the outcome (here, student test scores). The IV estimates rely on the variation in the variable of interest coming from the instrument to estimate its effect. That is, here the researchers estimate the effect of percent private school enrollment on test scores using only the variation in percent private correlated with historical percent Catholic.

5 Austria, Belgium, the Czech Republic, Denmark, Finland, France, Germany, Hungary, Ireland, Italy, Netherlands, New Zealand, Poland, Portugal, Spain, Sweden, Switzerland, United Kingdom, and the USA.

6 Argentina, Bolivia, Brazil, Colombia, Chile, Dominican Republic, Mexico, Paraguay, Peru, and Venezuela.

7 See the documentation for the WDI at the World Bank data catalog: https://datacatalog.worldbank.org/search/dataset/0037712/World-Development-Indicators

Appendix

Table A1. presents estimates using the specification from Table 2 and the HLO and percent private for secondary school students. Sample sizes are roughly half and there are insufficient HLO observations for reading to estimate results. The sample includes data on 61 countries. Of these, we observe 29 countries once, 13 countries twice, 7 countries in three years, and 12 countries in four years. Any effects on students are likely larger for younger students than for older students.

Table A1. HLO and percent private for secondary school students.

Table

	(1)	(2)	(3)	(4)	(5)	(6)
Panel A: Science
% privatet-1			0.975	1.092	1.061	−0.0323
			(1.942)	(1.980)	(2.063)	(0.997)
ln(real gdp pc)	26.06***	23.68		27.73	27.35	56.29
	(6.830)	(20.87)		(24.46)	(24.14)	(34.00)
polity2					−0.186	−1.430
					(2.018)	(1.755)
Mortality rate 5–9 y.o.					−1.050	4.209
					(13.11)	(8.425)
Observations	118	118	118	118	118	118
Adjusted R-squared	0.282	0.893	0.894	0.894	0.890	0.966
Panel B: Mathematics
% privatet-1			1.066	1.100	1.065	0.263
			(1.633)	(1.679)	(1.714)	(1.006)
ln(real gdp pc)	29.07***	4.742		8.771	6.211	9.682
	(7.878)	(21.26)		(21.43)	(21.58)	(32.55)
polity2					0.135	0.887
					(1.590)	(2.142)
Mortality rate 5–9 y.o.					−4.911	−8.500
					(11.49)	(9.089)
Observations	124	124	124	124	124	124
Adjusted R-squared	0.266	0.943	0.945	0.944	0.942	0.983
Country fixed effects	No	Yes	Yes	Yes	Yes	Yes
Weighted by population	No	No	No	No	No	Yes

Note: All regressions also include year fixed effects. Standard errors clustered by country in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.