When Does Inter-School Competition Matter? Evidence from the Chilean “Voucher” System

In Gallego (2006), I present a model to rationalize the effects of competition on student outcomes. The main idea is that there are two types of schools in a market: public schools with no direct incentives to produce quality beyond meeting a minimum enrollment level and voucher schools that face explicit competitive incentives. In this context, the model predicts (i) positive effects of voucher-school entry on the quality offered by both voucher and public schools and (ii) that the extent of the response of public schools to voucher-school entry depends on the minimum enrollment level needed by a public school to operate and on the size of the school-age population. McMillan (2005), Ferreyra and Liang (2012), and Tapia (2010) present models with the same prediction regarding the ways in which competition affects public schools’ rents and through this channel thereby affects student outcomes.

Notice that by the Cannon Law of the Catholic Church, a school is formally recognized as Catholic when either (i) the Church directly appoints the school principal or (ii) the Church approves the appointment of the school principal. Therefore, many schools that are related to the Catholic Church are not considered formally Catholic.

I focus on the results using 2002 test scores, because the source of identification I use for the presence of soft-budget constraints in public schools is not available outside the 1999–2003 period.

A small group of subsidized private schools did operate before the 1981 reform. These schools enrolled about 7% of the school-age population (estimates using data from the 2002 Social Protection Survey) and were financed through small public subsidies and private donations (Aedo-Richmond 2000).

Elacqua (2009) estimates that about 66% (59%) of voucher-school students were enrolled in for-profit schools in 2008 (1990). Note that some for-profit schools are formally Catholic.

Using a structural model of school choice, Gallego and Hernando (2009) present evidence of a negative effect of school market power on the level and growth rate of test scores.

Another important difference between the 1990s and the 1980s is that test scores for the complete population are only available for the 1990s.

As mentioned above, I also ran regressions for 2006 and 2010, also using the SIMCE dataset, but I have not presented descriptive statistics to save space.

I use five categories to measure mother’s education (having attained, at most, primary education, secondary general education, secondary technical education, post-secondary technical education, and college or postgraduate education).

It may be possible that there is within-municipality variation in competition levels that is different across municipalities. Given that the main estimates in this article use instrumental variables that vary at a more aggregate level, my main estimates have to be interpreted as the average effect of inter-school competition at the municipality level.

In particular, I use data on priests per capita for 1950 from the Annuario Pontificio. I estimate the number of priests in the different dioceses in 1950 by considering the territorial division existing in the 1990s (which includes 26 dioceses) and the number of priests in 1950. I make this adjustment, because some dioceses (namely Santiago) included disproportionately large areas of the population in 1950. Between the 1960s and the 1990s, new dioceses appeared when some dioceses were split up – so the number of dioceses increased from 19 to 26. I assume that the distribution of priests within the split dioceses is given by the distribution when the new dioceses were created (the distribution of priests within dioceses in the following periods is quite stable). In all the empirical applications, I cluster standard errors at the 1950 dioceses level.

There may also be bias caused by miss-measurement of inter-school competition. See, for instance, my discussion in footnote 10.

By controlling for the share of the Catholic population, I take potential direct effects of this variable on educational results into account (as suggested by recent research on the effects of religious affiliation on income, education, and other social and economic variables, e.g., Barro and McCleary 2003 and Gruber 2005).

These estimates are computed as follows: the Social Protection Survey implies that 7% of those in pre-voucher cohorts attended subsidized private schools. Espinola (1993, quoted in Hsieh and Urquiola 2006) reports that 53% of pre-reform subsidized private schools were Catholic. If Catholic and non-Catholic schools are of the same size, then I estimate that 3.5% of the school-aged population attended Catholic subsidized schools before the reform. However, I know that today Catholic schools tend to be bigger than non-Catholic schools. Therefore, using the ratio of size between Catholic and non-Catholic schools in 2002, I estimate that around 4.7% of the school-aged population before the reform was in Catholic subsidized schools.

Unfortunately, there are no systematic data on enrollment in Catholic schools before and after the reform. My estimates, using data from Pasalacqua (2004), are that 25% of enrollment in Catholic voucher schools in 2002 corresponds to brand new Catholic schools and between 15 and 25% of the enrollment corresponds to schools that became voucher schools after the reform.

Anecdotal evidence from one of the most important Catholic groups in Chile (the Marist brothers) may help to understand the increase in enrollment in Catholic voucher schools after the reform. Enrollment in Marist schools was 5,000 students in 1980, with about 10% of these students in subsidized schools. In contrast, enrollment was 14,800 in 2002, with 40% corresponding to voucher schools. About two thirds of enrollment in voucher schools corresponds to schools established after the reform (Personal communication with Br. Aldo Pasalcqua).

This value is computed as follows. The initial value of the voucher was 30% higher than expenditure per student in public schools before the reform. Before 1981, private schools received an average of 50% of public schools expenditure per-student (Hsieh and Urquiola 2006). Therefore, the nominal value of the voucher increased by 160%.

Again, I have no systematic data to quantify this hypothesis, but two examples may help us to understand the magnitude of this phenomenon. About 40% of enrollment in Jesuit-connected voucher schools (through the Red Educacional Ignaciana and Fe y Alegria Chile) is not counted as enrollment in Catholic schools. Most of these schools were established after the reform. Similarly, enrollment in the (two large) Opus Dei voucher schools in Santiago is not counted as in Catholic schools, because Opus Dei schools are not connected to the local Catholic bishop.

In Gallego (2006), I also show that priests do not affect the propensity to attend Catholic vis-à-vis non-Catholic voucher schools at the individual level.

In the Catholic Church, a diocese is an administrative territorial unit, composed of many parishes and governed by a bishop. Technically, each diocese is independent of the others, and the bishop is answerable only to the Pope.

These numbers were computed using information from the 2002 directory of the Catholic Church in Chile.

Municipalities in Chile belonged to 13 different regions in 2002.

In addition, it is worth mentioning that both and are statistically different from 0 (with p-values of 0.07 and 0.04, respectively).

In addition to these exercises, using individual data for a small sample of individuals from the 2002 Social Protection Survey, I studied whether the effect of priests on attendance at voucher schools changes for members of cohorts that attended school after and before the voucher reform, controlling for a dummy for urban areas, and region and age dummies. Results, available in Gallego (2006), suggest that the number of priests is more connected with the decision to attend a voucher school after the reform, as expected.

Results for other variables included in these regressions are similar to other articles in the literature: mean (standard deviation of) education and income have positive (negative) effects on the availability of voucher schools, and more populated areas have more voucher schools.

I also carried out an exercise in which I estimated the effect of the number of priests on the change in public school enrollment from the pre-reform period. I was able to compute public enrollment rate for 52 geographical areas for 1975 from MINEDUC (1975) — the geographical classification corresponds to the division of 26 national Chilean provinces into urban and rural areas. I define the change in enrollment in public schools as enrollment today minus enrollment in 1975 in the area where the school is located. Results, available in Gallego (2006), imply that public school enrollment decreased more in areas with more priests, confirming my argument in Section 4.

I thank a referee for suggesting this exercise.

One example may help us understand the working of this correlation. Let us consider the city of Los Angeles (a big city located in the South of Chile) in which the supply of Catholic schools is high (about 44% of voucher schools belong to Catholic orders) and more than half of the for-profit schools have religious names. Interestingly, more than half of these for-profit schools with religious names were funded in the 1990s (in contrast, just around 1/3 of the other for-profit schools started operating in the 1990s) suggesting that entry of these schools to the market was related to a latent demand for “religious” education.

Since I include multiple observations of variables in the same area, I use the White/Huber estimator of the variance-covariance matrix to compute corrected standard errors that are robust to arbitrary heteroskedasticity and clustered standard errors.

In Gallego (2006), I also present parsimonious models without including controls and only including my measure of voucher-school competition. The results are very similar.

Measurement error in voucher-school competition may also explain why my OLS estimates are smaller than my IV estimates.

The comparison of my estimates with the estimates in that article is interesting, because Gallego and Hernando use a measure of inter-school competition that varies within municipalities. The fact that my estimates, using only between municipalities variation in inter-school competition, are similar to the estimates in that paper suggest that the measurement error of my variable is not important.

In additional exercises, I replace my measure of voucher-school competition with three alternative measures: the share of enrollment in voucher schools (Auguste and Valenzuela 2003; Gallego 2002; Hsieh and Urquiola 2006), the Herfindahl index (Contreras and Macias 2002), and the change in enrollment in public schools between the pre-reform (1975) and post-reform periods (2002). Results reported in Gallego (2006) imply a positive effect of the three alternative measures of voucher-school competition on test scores.

I only present the estimates using both instruments to save space and to be able to report over-identification tests. The fact that my data pass the over-identification tests implies that estimates using just one of the instrumental variables are not statistically different among them.

To implement this procedure, I need to find a variable that affects the selection of students in different schools and has no direct effect on test scores. My instrument in the selection equation is a dummy that takes a value of one if the teaching of values was among the top three criteria used by parents for choosing schools. Since the mention of “teaching of values” (i.e. la enseñanza de valores in Spanish) has a religious connotation in Chile, this variable may capture relative preferences for voucher vis-à-vis public schools, or Catholic vis-à-vis non-Catholic voucher schools. Results for the selection equation are reported in the Appendix, and discussed in detail in Gallego (2006). Using this selection equation, I include the inverse of the Mills ratio in the regression reported in the last column of Table 6.

I have implemented several robustness and specification checks to the results in Table 6 (reported in Gallego 2006): (i) I estimate the regression excluding parents who attended school after the 1981 to address the fact that my instruments could be correlated with their educational outcomes and the estimated effect of voucher-school competition is mainly unchanged; (ii) I estimate regressions excluding schools located in rural areas, because the high physical transportation costs of moving from one school to another in those areas should decrease the impact of competition: results confirm this idea as the point estimate of inter-school competition increases when excluding rural schools; (iii) I include controls for systematic differences in pre-reform educational outcomes using three different proxies for pre-reform outcomes (high-school graduation rate in public schools for cohorts that attended school before 1980 from the Social Protection Survey, high-school graduation rate at the municipality level for cohorts that attended school before 1980 from the CASEN survey, and the average 1991 SIMCE test scores at the municipality level) and find that my main estimates of the effect of the ratio of voucher-to-public schools change slightly in value, but remain statistically significant; (iv) I introduce controls for the composition of students at the school level: mean and standard deviation of mother’s education and per-capita income (following Hoxby (2000) in including these variables without giving a formal interpretation to the estimates) and find that the point estimate of the effect of voucher-school competition is basically unchanged with respect to most other estimates I present in Table 6; and (v) I estimate quantile regression estimates of voucher-school competition for students in different positions of the distribution and find that the estimated effects do not vary a lot across quantiles, but are slightly smaller for the students in the first and tenth quantiles than for the other students.

This is the reason why I do not implement the same exercise for 2006 and 2010. During this period, the mayoral election was different: there were candidates specific for the mayor post (in contrast to the 1999–2003 period in which mayors were elected among the city council candidates), and, therefore, I cannot use the same identification strategy.

A specific example may clarify the working of my identification strategy. Let us consider El Bosque and San Bernardo, two neighboring municipalities located in Santiago. The total support for all the candidates of the opposition coalition who were standing for the municipality council was very similar — 32.3% in El Bosque and 34.6% in San Bernardo — but in San Bernardo the mayor belonged to the opposition coalition. The explanation for this is simple: the main candidate to the municipal council belonging to the opposition coalition obtained higher support with 27.7% of the vote, while the main candidate from the pro-government coalition got only 18.5%.