Swedish Institute for Social Research (SOFI)
Stockholm University
WORKING PAPER 3/2020
THE IMPACT OF VOUCHER SCHOOLS:
EVIDENCE FROM SWEDISH UPPER SECONDARY SCHOOLS
by
Karin Edmark, Iftikhar Hussain & Carla Haelermans
The Impact of Voucher Schools:
Evidence from Swedish Upper Secondary Schools ⃰
Karin Edmark
♣Iftikhar Hussain
♦Carla Haelermans
♠24 March 2020
ABSTRACT
Empirical studies investigating the impact of private voucher schools on student outcomes have focused on a number of mechanisms, including productivity and competitive effects. Arguably, the possibility that these voucher schools may provide greater variety, in terms of education options or tracks remains an understudied area. This paper exploits the rapid expansion of private academic and vocational track schools in Sweden, to address this question. We uncover new evidence that the introduction of private voucher schools induced greater vocational education participation, and not simply a substitution of public for private vocational schools. In effect, private school penetration lead to a switch away from academic tracks, including both science and social science, in favour of vocational options. We then ask what impact inducing greater participation in vocational education had on short- and medium-term outcomes, including GPA, on-time graduation from high school, university participation and field of study at university. We discuss other possible mechanisms, including changes in peer and teacher quality.
JEL-Classification: H44, I21, I26, I28.
Key words: Private provision, independent schools, voucher school reform, vocational education, upper secondary education.
⃰We acknowledge funding from the Swedish Research Council, Project Nr: 2014-01783. We are grateful for comments from seminar and conference participants at the Fourth Lisbon Research Workshop on Economics, Statistics and Econometrics of Education; The 73rd Annual Congress of the IIPF in Tokyo; the Swedish
Institute for Social Research at Stockholm University (SOFI); the Uppsala Centre for Labour Studies Workshop;
and the Research Institute of Industrial Economics (IFN).
♣ SOFI, Stockholm University, SE-106 91 Stockholm, Sweden, IFAU, IFN and CESIfo, karin.edmark@sofi.su.se, http://orcid.org/0000-0002-5629-5499.
♦ Department of Economics, University of Sussex. iftikhar.hussein@sussex.ac.uk
♠ Research Centre for Education and the Labour Market (ROA), School of Business and Economics, Maastricht University. http://orcid.org/0000-0002-9202-8427; carla.haelermans@maastrichtuniversity.nl
1. Introduction
The question of how to improve the educational quality is of great importance for policy- makers around the world. One policy that has gained popularity in many countries is to increase the role of private provision. Typically, this has been done by introducing school vouchers that allow students to choose the school of their liking, be it publicly or privately provided, while the funding follows the student to the chosen school.1 Notably, this system allows funding to be kept strictly public, while adding market based elements of (more or less) free entry for private providers, and competition between schools. These market based elements, goes the argument, will improve efficiency and innovation in the education sector. Opponents to these reforms however claim that the effects will be increased segregation and quality-reducing cost- cutting2.
Sweden was one of the first countries to implement such policies at a large scale. A series of reforms in the early 1990s introduced practically free entry for privately operated schools, with full public funding granted through a voucher system. The reforms applied both to the lower, compulsory, education system (grades 1–9), and to the voluntary upper secondary education system (grades 10–12). These privately operated but fully publicly funded schools are sometimes referred to as “independent schools”, or “voucher schools”, but we will in this paper mainly use the term private schools.
This paper focuses on the expansion of the independent upper secondary education that took place during the first decade of the 2000s. Our research design takes advantage of the rapid increase in the number of private schools following the voucher school reform by estimating how the opening of new private schools nearby a student’s residence affected students’ school and track choice, and via that, educational outcomes.3 More precisely, we measure the private school share nearby a student’s home (that is, the share of all schools nearby a students’ home that are private), at the time when the student is expected to choose an upper secondary school. We then base the analysis on the variation in private school access, caused by school openings, among students who reside within the same 250m-square grid cell but who start upper secondary school in different years. Alternatively, we limit the sample to siblings who reside within the same grid cell, and use only the within-sibling variation.
How do we expect the private school expansion to affect student outcomes? In theory, the effects of introducing private provision of education and school choice can go either way.
On the one hand, school choice can be argued to give rise to better matching between students and schools/tracks, and also to give rise to a positive competition for students which leads to improved quality in all schools. These processes are reinforced by allowing private school entry, as this provides both more alternatives for students and increased competition. Private schools additionally have more autonomy than the public counterparts, and are in Sweden allowed to run as for-profit firms, which can be argued to make them more efficient providers of education. On the other hand, the theoretical arguments can be turned up-side down: there is no guarantee that the competitive pressure among schools will improve quality – it may even be damaging if schools start competing with measures that could be harmful to learning, such as lax discipline, short school days or overly generous grading. There is also no guarantee that the autonomy that the private schools enjoy make them better providers of education than their
1 See for example Epple (2017), for a review of voucher systems, or Peterson et al. (2003), Chingos and Peterson (2012) and Krueger and Zhu (2004).
2 See e.g. the overview study of Rouse and Barrow (2008) for effects of increased competition.
3 Our research strategy builds on the literature using local variation in the access to education, e.g. Currie and Moretti (2003).
public counterparts. Free private entry may also make local planning more difficult, by adding uncertainty to what the future supply will look like as private agents enter and exit the market.4
It shall be noted that private provision and school choice may work differently for the upper secondary level compared to lower levels of education. For example, students entering upper secondary education are older and thus likely to have more say in the choice of school than younger students, where it is probably the parents who make the choice. In addition, upper secondary education in Sweden is organized in around 20 educational tracks, which means that students simultaneously choose a school and a track. Private schools that enter the market can choose which of these educational tracks they offer, and are not required to offer any certain mix of tracks. The effects of private provision and the ensuing school choice and competition may hence differ between the lower and upper levels of education. Furthermore, males and females tend to choose different tracks – so are likely differently affected depending on what tracks are offered by the entering private schools. More research on the effects of private upper secondary schools is hence needed, and this paper aims to help fill this gap.
While the effects of the Swedish privatization reform on students in the lower level grade 1–9 education system has been the subject of a handful of previous economics research studies5, the research on the upper secondary level is scarce. To our knowledge, it is limited to Hinnerich Tyrefors and Vlachos (2017), who study the impact of private school attendance for a sample of roughly 10 percent of Swedish upper secondary students.6 Using a conditional-on- observables approach, they find that students in private schools on average achieve better on the teacher-assessed standardized tests. However, when the Swedish School Inspectorate re- graded the tests, the independent school students actually scored worse than students in public schools. Systematic differences in grading standards between private and public schools is thus a concern, and we shall, as a result, be careful when interpreting teacher-assessed performance measures.7
The international evidence on voucher systems is mixed. Epple et al. (2017) survey the literature on voucher schools and report evidence of positive competition effects on public schools, but also of increased sorting of students. The effect of voucher school attendance differs across settings: positive and sometimes large effects are estimated for some programs and groups of students, whereas a negative impact is found for others. A mixed picture also emerges from the U.S. literature on charter schools, see Epple et al. (2015) for an overview.8 Finally, Kortelainen and Manninen (2019) find no impact on student exit exams of attending a private upper secondary school, compared to a public ditto. Kortelainen and Manninen study only academic track students, and is limited to schools Helsinki. The Finnish private school sector is smaller than the Swedish one, and allows only non-profit schools.
The varying empirical results from the previous literature suggest that it is useful to consider the specific contexts that prevail in different voucher systems, and how they shape the
4 How to improve the local planning capabilities for upper secondary education is the topic of an inquiry commissioned by the Swedish government (Komittédirektiv 2018:18, Planering och dimensionering av gymnasial utbildning), which is due in June 2020.
5 Böhlmark and Lindahl (2015), Hennerdahl et al. (2018), Sandström and Bergström (2005), Ahlin (2003), Björklund et al. (2005).
6 Private provision in the Swedish upper secondary education system is also studied a forthcoming working paper, which uses matching and RD methods to study the impact of independent school attendance, see Edmark and Persson (2020).
7 Vlachos (2018) and Skolverket (2019) are examples of additional studies that point to the prevalence of more lenient grading standards among independent schools. Diamond och Persson (2016) furthermore suggest that grade inflation has positive long term spill over-effects on students’ educational attainment and earnings.
8 See also e.g. Dobbie and Fryer (2019), Abdulkadiroglu et al. (2018), Berends and Waddington (2018), Hahn et al. 2018, and Chabrier et al. (2016) for more recent studies on voucher and charter schools.
way the system works. This project adds to this notion by investigating not only how increased private school supply affects student outcomes, but also the impact on what educational tracks they attend.
The results of this paper suggest that an increased presence of private schools near a student’s home raises the likelihood to attend a private school. It also affects the type of educational track the student attends in upper secondary school: a rise in the private school share leads to an increase in the probability of enrolling into a vocational track. This is a key finding and demonstrates that private school penetration is not neutral in its impact on the mix of academic and vocational tracks chosen.
We also investigate the impact of private school expansion on the probability of graduating on time, students’ GPA, and university attendance. We find some indication that a higher vocational private school share increases grades and graduation rates, whereas a higher academic private school share gives the opposite pattern. However, these results are sensitive to changes to the regression model and private school share variables, and should therefore be treated with caution.
2. Swedish Context: Track Choice and the Voucher School Expansion
9Students enter upper secondary education at age 16, after 9 years of compulsory education.10 They are free to choose among all voucher schools in the country, and among all municipal schools in their home region.11 Upper secondary school is divided into a number of academic and vocational tracks.12 There is also a shorter preparatory track for students with insufficient grades to qualify for a regular track.13 Admission to a school and track is determined solely by the students’ final grade from lower secondary education, in a deferred acceptance system.14
The current voucher school regulation has its roots in the privatization reforms of the early 1990s, which greatly improved the economic situation for privately provided schools through the introduction of government-funded vouchers. The vouchers are paid by students’
home municipalities and provide full funding to approved schools. Additional tuition fees are not allowed. Apart from being privately operated, the voucher schools are by large subject to the same regulations as the municipally operated schools15: They are monitored by the Swedish School Inspectorate, follow the same curriculum and educational goals, and need to hire certified teachers (although exceptions are allowed in case of teacher scarcity). For-profit organizations may run schools, and the vast majority of upper secondary private schools belong to larger corporate groups. Conversions of public schools to voucher schools have occurred, but have been rare.16 There is no cap on the total number of voucher schools, but approval can
9 For a more detailed overview of the institutional setting for the Swedish voucher funded schools, see e.g.
Edmark and Persson (2020).
10 There is also a preparatory pre-school year, which is mandatory since 2018.
11 They may apply to municipal schools outside their home region, but home student applicants are given priority.
12 See Table A1 in the Appendix for a full list of the tracks in each category. We will in the following often analyse these two groups of tracks separately, and will then broadly refer to them as the Academic and Vocational tracks.
13 The preparatory track is rarely provided by the private schools. We will in this paper therefore focus on private provision of academic and vocational tracks, and limit the sample to students who qualify for these.
14 Ability tests may be used for admission to the arts track and special profile tracks.
15 In the early days of the reform the voucher schools were less regulated, but over time, the regulation governing the municipal school has been applied also to the voucher schools.
16 http://www.statskontoret.se/globalassets/publikationer/2017/201724.pdf
be denied if entry is determined to have substantial negative financial consequences for the municipality’s ability to provide education.
Figure 1 charts the expansion of upper secondary private schools over the period 2001 to 2010.17 It clearly demonstrates the dramatic expansion in the number of private schools, from around 150 at the beginning of the period to close to 500 by the end of the decade. The share of upper secondary students enrolled in voucher schools also exhibits a large increase, though less dramatic in proportionate terms than for the total number of schools, a consequence of the fact that private schools are on average substantially smaller than their public counterparts.
Figure 1 Private and public upper secondary schools, and share of entry-level (grade 10) students enrolled in private schools
Figure 2 shows trends in the number of entry-level (grade 10) students enrolled in public and private schools by track type. The total number of students in public schools in both tracks has declined in the most recent years of the figure, whilst for private schools there has been a steady rise across the two tracks. Appendix Table A4 shows the total number of public and private schools, by the track offered, as well as mean enrolment size of the grade-10 intake.
Overall, in terms of enrolment, the average private school is around a third as large as the average public school.
Figure 3 breaks down the number of schools by track into the following categories:
those offering both academic and vocational tracks, as well as those offering only one of these two types of tracks. Private schools exhibit large increases across the board over this period.
There are no strong trends in the public sector. A notable feature of the supply of private schools is the preponderance of single track schools: over two thirds of them offer either academic or vocational tracks only.
17 A private school is any private school offering both vocational or academic tracks, or only one of these two tracks.
Figure 2 Number of entry-level (grade 10) students enrolled in public and private schools by track type.
Table notes: Students in the preparatory track and students with missing or erroneous tracks information are omitted from this figure.
Figure 3 Number of upper secondary schools offering either only Academic or Vocational tracks, or both, among private and public schools.
Figure notes: the figure omits a very small number of schools for which we were unable to identify the tracks offered.
Finally, Figure 4 shows the share of grade-10 students enrolling in the different tracks types. It shows an upward trend in the share enrolled in the Vocational tracks over most of the period under investigation, while the Academic track share was declining.18 The figure also shows an increase in the preparatory track share. The preparatory track was however very rarely offered by the private schools over time period, and we will therefore focus on the Vocational and Academic tracks in this paper.
18 It can be noted that this trend was reversed in the second decade of the 00s: after the period studied here, the vocational track shares decreased. It is believed that this was due to a reform, outlined in the government proposition 2008/9:199 and implemented in 2011, which reduced the theoretical content of the Vocational tracks, which meant that they were no longer not automatically qualifying students to basic university education.
Figure 4: Share of entry-level students enrolled in Vocational, Academic and Preparatory tracks
3.1 Descriptive Evidence: Selection into Private School, Student Outcomes and School Quality
Our data are based on various registers held by Statistics Sweden, and mainly cover the first decade of the 00s.19 We observe students’ educational careers starting from the last year of lower secondary education and throughout upper secondary school. We also observe if students attend university at age 22, although this information is not available for the last of our cohorts.20 The data furthermore include a large set of demographic and family level background characteristics, such as age, gender, parental education levels, country of birth (aggregated to larger regions), household disposable income, etcetera. We are able to link parents and children, which means that we can identify siblings. The school level data include information on number of and qualifications of the staff, and whether or not the school is privately operated.
For both schools and students, we observe geographical location at the level of 250m- square grid cell. The grid cells are, for integrity reasons, not available for very rural regions, which means that these are omitted from the analysis. Grid cells are missing for around 15%
of all schools; and this is partly due to the omission of very rural regions, and partly to failure to link school addresses.21 Appendix Figure A1 shows the locations of the private schools in our data in 2000 and 2010. It illustrates that the rapid expansion of the private schools over this period by large correlates with the population density – schools did open up in all parts of the country but more so in the more populous areas.
Table 1 presents descriptive statistics for students, including family characteristics as well as end of upper secondary and post-schooling outcomes, by public and private school tracks. Panel A of Table 1 demonstrates that there is evidence of strong positive selection into private vocational schools: relative to those attending public vocational schools, students attending private vocational schools have higher prior grades (grade 9) and come from households with higher income and where at least one parent is more likely to have attended
19 More detailed information on the data is available in the supplementary Appendix to this paper.
20 University outcomes are in our data available until 2014, and does hence not over our last cohorts of data. We also estimated the impact on having any work-income, from employment or self-employment, at age 22, based on data available until 2013. We however recognize that age 22 is a very early age to measure earnings, and therefore present the results for this outcome in the Appendix.
21 See the supplementary Appendix for details.
university. There is little evidence of selection into private academic versus public academic schools. Finally, private vocational track schools appear to be more segregated by gender:
males represent 66 percent of students in vocational private track school, whilst they are 55 percent of the total in public vocational track schools.
Table 1: Descriptive Statistics
(Main Regression Sample of First Year Upper Secondary Students 2001–10)
(1) (2) (3) (4) (5) (6)
Academic
Private
Academic Public
Difference Vocational Private
Vocational Public
Difference
Panel A: Student background characteristics Final grade lower
secondary education
237.44 237.85 -0.41 200.14 188.60 11.55
(49.52) (44.46) (0.17) (44.73) (38.31) (0.19)
Male 0.44 0.48 -0.03 0.66 0.55 0.11
(0.50) (0.50) (0.00) (0.47) (0.50) (0.00)
Log of household disposable income
13.01 12.98 0.03 12.89 12.82 0.07
(0.51) (0.46) (0.00) (0.46) (0.43) (0.00)
Parent high education 0.64 0.62 0.01 0.43 0.32 -0.02
(0.48) (0.48) (0.00) (0.49) (0.47) (0.00)
Parent Swedish born 0.86 0.86 0.01 0.89 0.90 0.11
(0.34) (0.35) (0.00) (0.32) (0.30) (0.00)
Panel B: End of upper secondary and post-secondary schooling outcomes
Graduate on time 0.77 0.80 -0.03 0.73 0.74 -0.01
(0.42) (0.40) (0.00) (0.45) (0.44) (0.00)
Percentile of final GPA 0.60 0.57 0.02 0.44 0.40 0.04
(0.30) (0.28) (0.00) (0.27) (0.24) (0.00)
Enrolled University (age 22)
0.45 0.49 -0.04 0.22 0.11 0.11
(0.50) (0.50) (0.00) (0.41) (0.31) (0.00)
Enrolled in STEM field (if enrolled age 22)
0.23 0.31 -0.08 0.47 0.19 0.28
(0.42) (0.46) (0.00) (0.50) (0.39) (0.01)
Number of students 92 835 409 393 56 098 215 529
Table notes: The table is based on the regression sample, and this excludes students with insufficient grades to qualify for the Academic or Vocational tracks. The summary statistics may thus differ slightly from those presented in figures and tables based on the full population. Household disposable income is given in 2010 monetary value. Share variables indicate that at least one parent is High Education/Swedish-born. The final grade from lower secondary education is defined as the sum of the grade credits for the 16 best subjects, and ranges from 0 to 320. Percentile of final GPA is defined as the year-wise percentile rank (from 0 to 0.9999) of students' final GPA from upper secondary school. Standard deviation in parenthesis under means, and standard error in parenthesis under differences.
In panel B of Table 1, we report end of upper secondary school and post-schooling outcomes. These show that private school students (across both tracks) are less likely to graduate on time than their public school peers (in the respective tracks). However, private school students have higher GPA scores. Students graduating from private academic tracks are slightly less likely to be enrolled in university by age 22 compared to those graduating from public academic track schools, whilst private vocational track students are twice as likely as public vocational track students to be enrolled in university.
Conditional on attending university, the statistics on studying a STEM subject are mixed, with private academic track students less likely to do so than public academic track students, whilst private vocational track students are more likely to do so than students from the public vocational track. Finally, private schools tend to have a lower proportion of qualified teachers and have higher student-teacher ratio.22 One reading of these data might be that private schools offer better peer quality peers (via sorting) but arguably worse school characteristics, at least on conventional measures.
As explained in greater detail in the next section, the treatment variable is the proportion of private schools within a 20km radius of the centroid of the 250m-square grid cell. As an example, Figure 5 shows a heat map of growth in private schools over the period 2001 to 2010 for the municipalities in Stockholm county. This shows that growth was more pronounced in the more urban areas.
Figure 5: Heat map Stockholm county
Figure notes: The figure depicts the change in the private school share between 2001 and 2010, measured within a 20km radius around each grid cell, in the municipalities of Stockholm county. Each dot represents a grid cell, and a darker shading signifies a larger increase in private school share. The white areas denote grid cells where there was no student residing in the period under study, and that are as a result omitted from our regression sample.
22 Our measure on qualified teachers is defined as the share of teachers with a pedagogical degree relevant for the level and subjects taught, and is given as a share of the total school staff.
2.2 Private School Expansion
Private school expansion is unlikely to proceed in an idiosyncratic or random fashion. In order to evaluate this proposition, we investigate the relationship between the change in private school availability in a neighbourhood and characteristics of the neighbourhood at the start of the rapid expansion process, i.e. 2001. Specifically, we run the following regression:
∆𝑆ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝑔𝑡 = 𝛼0 + 𝛼1𝐶𝑔2001+ 𝜎𝑟+ 𝜀𝑔𝑡 , (1)
where 𝑆ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝑔𝑡 is the proportion of private schools in a radius of 20km from the centre of the grid cell 𝑔, and ∆ signifies the long difference between 2001 and 2010. This difference is regressed on 𝐶𝑔2001, the mean of characteristics of 16-year old students and their families in the 20km circle, also centred at 𝑔, in 2001 (the ‘baseline’). These characteristics comprise log of disposable family income, proportion of students with at least one parent with a post- secondary degree, proportion of students with at least one Swedish born parent, log of the number of 16-year olds, and the log of students’ grade sum at grade 9 (i.e. prior to upper secondary school entry). These variables are summarised in Appendix Table A5, which shows that there are just over 66,000 grid cells and associated 20km circles, with a mean change in private school share over the 2001 to 2010 period of 0.24 (s.d. = 0.21).23 Labour market region fixed effects, 𝜎𝑟 are also included in this regression.24
We estimate equations of this form separately for (i) all private schools; (ii) private academic track schools; and (iii) private vocational track schools.25 We also repeat this analysis at the municipality-level, with change in private school share and 2001 characteristics measured at the municipality level. The results from the municipality level exercise are reported in the Appendix Table A6. They are often not statistically significantly different from zero, which likely reflects that the municipality averages mask relevant local variation.
Table 2 presents estimates for the grid cell level models based on equation (1). Columns 1 to 3 do not include labour market region fixed effects, whilst columns 4 to 6 do include them.
The large change in many coefficients when we include labour market region fixed effects is clear, and we focus on these results, i.e. we will investigate correlates of private school expansion within labour market regions. Column 4 shows that for all private schools, expansion is related negatively to 2001 income in the local neighbourhood but positively to parental education and the proportion of parents who are Swedish born. Private schools also expand more where initial student performance is lower.26
23 The minimum number of students residing within the 20km radius from the grid cells, in year 2001, is 36. In other words, very low-populated areas are excluded from the regression sample. This reflects that our sample only includes i) grid cells where there is at least one school within the 20km radius, and ii) grid cells for which there is at least one student resident in at least one year of the data panel.
24 We could alternatively include municipality fixed effects. However, given the potentially large size of school catchment areas for this age group of students, we investigate variation within the larger geography of labour market regions. This should yield a reasonably large geographical area whilst at the same time being relatively homogeneous.
25 Note that a school offering both tracks is counted in both the voucher academic track schools’ share and the voucher vocational track schools’ share.
26 The positive correlation with parental education is in line with the findings in Edmark (2019), which studies the location patterns of lower primary private schools. Edmark however estimates a negative correlation with the share of students who have a Swedish background.
Table 2: Correlates of Private School Expansion
(Dependent variable: change in private school share, 2001 to 2010, RHS-variables averages within 20km radius from grid cell)
(1) (2) (3) (4) (5) (6)
No FE’s Labour Market Region FE's
∆ Share Private
∆ Share Private Academic
∆ Share Priv Vocational
∆ Share Private
∆ Share Private Academic
∆ Share Private Vocational
Log of household disposable income
-0.6075*** -0.4365*** -0.6498*** -0.2603*** 0.1684*** -0.4639***
(0.0091) (0.0085) (0.0096) (0.0139) (0.0126) (0.0152)
Parent high
education 1.1036*** 1.0692*** 1.4958*** 1.6394*** 1.4207*** 1.9768***
(0.0157) (0.0133) (0.0167) (0.0217) (0.0202) (0.0214)
Parent Swedish born
0.2058*** 0.4235*** -0.4065*** 0.1292*** 0.3898*** -0.5373***
(0.0161) (0.0171) (0.0205) (0.0226) (0.0199) (0.0265)
Log of average grade sum
-0.7491*** -1.0946*** -0.6393*** -1.3455*** -1.8994*** -1.1465***
(0.0283) (0.0246) (0.0320) (0.0513) (0.0462) (0.0553)
Log of population density
0.0258*** 0.0456*** 0.0080*** 0.0164*** 0.0401*** 0.0111***
(0.0015) (0.0014) (0.0018) (0.0017) (0.0016) (0.0021)
Observations 66,225 66,225 66,225 66,196 66,196 66,196
R-squared 0.1376 0.1852 0.2417 0.5818 0.6179 0.6112
Table notes: The number of grid cells in this table is slightly lower than the number of grid cells in the tables for the regressions on the full panel. This is due to the fact that the private school share is missing for grid cells in years when there was no upper secondary school within a 20km radius. For the grid cells that are missing in this table, compared to the (unbalanced) regression panel data, this was the case in either 2001 or 2010. Standard errors (in brackets) are clustered at the grid cell level. *** p<0.01, ** p<0.05, * p<0.1.
Turning to the separate measures for academic and vocational track schools, there are similarities but also some clear differences across these two tracks. Both academic as well as vocational track private schools expand more where initial grade performance is low: a 10 percent decline in the average grade sum is associated with an 18 percentage point increase in the share of private academic track schools and an increase of 11 percentage points in the share of private vocational track schools.
However, turning to some of the other correlates shows up some stark differences in the patterns of penetration for the two types of private school tracks. The results for academic track private schools show clearly that these schools expanded more in more prosperous areas (as measured by higher parental income and education) and in areas where a greater proportion of parents are Swedish born. Vocational private schools, on the other hand, expand more in areas where income is lower and a lower proportion of parents are Swedish born. For example, a decline in the proportion of Swedish born parents of 10 percent is associated with a rise in the Vocational private school share of 5 percentage points.
In short, we see that private school expansion is not random, and in order to identify its causal impact, we need to choose an identification strategy that takes this into account.
3. Empirical Strategy
Our empirical strategy exploits variation in the availability of private schools, whilst controlling for a number of potentially confounding fixed and time-varying factors. In particular, we control for grid cell fixed effects, year fixed effects, municipality linear trends, labour market region-by-year fixed effects, municipality-by-year covariates, detailed student characteristics, including student’s prior grades, as well as family background characteristics.27 Our main estimation model is as follows:
𝑦𝑖𝑔𝑡= 𝛼 + 𝛿0𝑆ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝑔𝑡+ 𝛽1𝑋𝑖𝑡+ 𝛽2𝑊𝑚𝑡+ 𝛽3𝐷𝑔𝑡+ 𝜃𝑔+ 𝜆𝑚𝑡 + 𝛾𝑟𝑡+ 𝜇𝑡 + 𝑢𝑖𝑔𝑡 (2)
where 𝑦𝑖𝑔𝑡 is an outcome (for example, school track choice, GPA in grade 12, university attendance, or university field) for student 𝑖, residing in grid cell 𝑔 in year 𝑡. 𝑆ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝑔𝑡 is the proportion of private schools in a radius of 20km from the centre of grid cell 𝑔.28 𝜃𝑔 are grid cell fixed effects, 𝛾𝑟𝑡 are labour market region-year effects and 𝜆𝑚 are municipality- specific linear time trends. 𝑊𝑚𝑡 are municipality-year covariates, which include the share of students attending a private school in grade 9, measures of per-student municipality funding for students in public schools29, and an indicator for whether the municipality is run by a left- wing political majority. Also included in the regression are predetermined family and student covariates, 𝑋𝑖𝑡. These include student’s prior (grade 9) grade sum, whether the student finished 9th grade in a private school, student gender and age, household disposable income, parental education and employment status, and dummies for parental and student’s country of birth.
Finally, 𝐷𝑔𝑡, measures the size of the student population around a grid cell, defined as the log of the number of 16-years olds that reside within a 20km radius from the grid cell in year t. We estimate year-specific coefficients all covariates, in order to take into account that their correlation with the outcome variables may vary over the period. All regressions include year effects, 𝜇𝑡. We cluster standard errors at the grid cell level.
As discussed earlier, less than a third of private schools offer both academic and vocational tracks. The majority of private schools offer a single track only. We exploit this variation to estimate separate effects of the availability of vocational versus academic track private schools. In order to do this, we make a distinction between the proportion of private schools in the neighbourhoods offering vocational tracks and the proportion offering academic tracks. Thus we adapt model (2) as follows:
𝑦𝑖𝑔𝑡= 𝛼 + 𝛿1𝑠ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝐴𝑐𝑔𝑡+ 𝛿2𝑠ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝑉𝑜𝑐𝑔𝑡+ 𝛽1𝑋𝑖𝑔+ 𝛽2𝑊𝑚𝑡+ 𝛽3𝐷𝑔𝑡+ 𝛾𝑟𝑡 + 𝜆𝑚𝑡 + 𝜃𝑔+ 𝜇𝑡+ 𝑢𝑖𝑔𝑡 ,
(3)
27 For the full list of covariates, see the table notes of Table 4.
28 The choice of 20km as cutoff was chosen in order to take into account that a large portion of private school students travel relatively far: the distance to the school of attendance among private school students in our regression sample averages 19km and the median value is 8km. Results for smaller cutoffs of 5km and 10km are reported in the Appendix.
29 This measure is based on spending per pupil in compulsory education, and is intended to capture the
municipality’s general generosity (or needs) towards education. It is not based on spending for upper secondary education, because the per-pupil cost at this stage varies markedly across the educational tracks.
where 𝑠ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝐴𝑐𝑔𝑡 and 𝑠ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝑉𝑜𝑐𝑔𝑡 are the proportion of private academic and vocational schools, respectively, in a radius of 20km from the centre of the grid cell.30
The parameters of interest in (3) are 𝛿1 and 𝛿2, the separate impact of the availability of private academic and vocational tracks, respectively, on the outcome of interest.
Identification relies on the assumption that after accounting for grid cell fixed effects, municipality time trends, labour market region-year effects, as well as observable student and municipality covariates, the share of private schools is uncorrelated with the error term, 𝑢𝑖𝑔𝑡. The balancing test, discussed in 3.3, lends credibility to this assumption.
3.2 Sibling Fixed Effect Design
We employ a second research design, exploiting sibling comparisons. The idea here is to account for time-invariant family traits and assess whether changing supply of private schools has an impact on take-up of private schooling as well as track choice. We estimate models of the following form:
𝑦𝑖𝑓𝑔𝑡 = 𝛼 + 𝛿0𝑆ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝑔𝑡+ 𝛽1𝑋𝑖𝑡+ 𝛽2𝑊𝑚𝑡+ 𝛽3𝐷𝑔𝑡 + 𝜌𝑓+ 𝜆𝑚𝑡 + 𝛾𝑟𝑡+ 𝜇𝑡 + 𝑢𝑖𝑔𝑡 , (4)
where 𝜌𝑓 is the family (strictly speaking, mother) fixed effect and all other variables are as defined before. We estimate the model on the sample of non-moving families. Results for the full sample, including families who move between grid cells, are available upon request, and are overall similar to the non-moving households. Note that the family fixed effect absorbs the grid cell fixed effect for the sample of non-moving households.31
3.3 Balancing test
As we now show, changes in private school share are unrelated to changes in pre-determined student and family covariates at the grid cell level, once the set of fixed effects and linear trends outlined above are controlled for. This lends credibility to the assumption that the rapid expansion of private schools generates useful variation in the supply of private schools. For example, if the variation we exploit is confounded by selective sorting of families into neighbourhoods experiencing greater private school expansion, or, alternatively by selective location of private schools to particular neighbourhoods, then we might expect to detect a relationship between our treatment variable and family background characteristics.
In order to carry out this balancing test, we run the following regression where the outcome variable is now the share of private schools:
𝑆ℎ𝑎𝑟𝑒_𝑃𝑟𝑖𝑣𝑔𝑡 = 𝛼 + 𝛽1𝑋𝑖𝑔𝑡 + 𝛽3𝐷𝑔𝑡+ 𝜃𝑔 + 𝜆𝑚𝑡 + 𝛾𝑟𝑡+ 𝜇𝑡+ 𝑢𝑔𝑡, (5)
30 As before, a school offering both tracks is counted in each of these two categories.
31 When the model is estimated on the full sample (i.e. movers as well as non-movers) we also report results from models which include grid cell fixed effects. These results are available upon request.
where all variables are as defined in (2). We ask whether, conditional on the grid cell, year, and region-year fixed effects as well as linear municipality trends, the share of private schools (as before, in a radius of 20km from the centre of grid cell 𝑔) is correlated with pre-determined covariates such as student’s prior GPA, household disposable income and parental education and immigrant status, as well as the student population density around the grid cell. Column 1 of Table 3 reports the results from this exercise. Columns 2 and 3 repeat this exercise for the academic and vocational private share outcomes, respectively. This set of balancing results demonstrates that there is very little evidence of any correlation between the predetermined covariates and the treatment variables after we account for the fixed effects and municipality trends.
Table 3: Balancing Test
(Dependent variable: private school share)
(1) (2) (3)
All private schools Private Academic Track Private Vocational Track
Log of household disposable income
-0.0002* -0.0001 -0.0001
(0.0001) (0.0001) (0.0001)
Parent high education 0.0000 -0.0000 -0.0000
(0.0001) (0.0001) (0.0001)
Parent Swedish born -0.0002* -0.0001 -0.0002
(0.0001) (0.0001) (0.0002)
Log of final grade lower secondary education
-0.0002 0.0002 -0.0001
(0.0002) (0.0002) (0.0002)
Log of population density -0.0556*** -0.0129 -0.0856***
(0.0067) (0.0080) (0.0070)
Observations 781,469 781,469 781,469
R-squared 0.8765 0.8343 0.8827
Number of grid cells 66,459 66,459 66,459
Table notes: Population density is measured as the number of 16-year olds residing within 20 km from the grid cell midpoint. Missing observations in covariates were replaced with a constant value, and dummy variables were added to indicate these occurrences (not shown in table).
All regressions include year and grid cell fixed effects, labor market region by year fixed effects, and municipality specific linear trends. Standard errors (in brackets) are clustered on grid cell. *** p<0.01, ** p<0.05, * p<0.1.
For example, for the private school vocational track share outcome, there is no significant relationship with parental characteristics such as education, Swedish born status, family income, nor with the student’s prior ability as measured by grade 9 GPA. There is evidence of an economically small but statistically significant negative relationship between growth in the private vocational school share and the log of number of 16-year olds residing in the grid cell (a decline of 10 percent in the number of 16-year olds is associated with a rise in the share of private vocational schools of approximately one percentage point). For the private academic school share there is no evidence of any statistically significant relationship. For the all private school share outcome (column 1) there is evidence of a very small but marginally
significant negative relationship with household disposable income and the share with a Swedish born parent. There is no evidence of any such relationship when the two tracks are analysed separately.
We repeat this exercise for the sibling sample (results are reported in Appendix Tables A7 and A8). For the non-moving households sample, there is no evidence of correlation between the treatment variables and time-varying family covariates such as family income.
There is some evidence of a negative relationship with mother’s and father’s education for the sample which includes moving households, which may arise from changes in family structure (such as divorce and remarriage).
4. School and Track Choice Results
In the next subsection we ask how the availability of nearby private school options affects private school enrolment. In subsection 4.2 we then move on to our main task and assess whether private schools induce students to switch upper secondary school track, whether private or public.
4.1 Private School Attendance
In Table 4 we investigate the impact of a rise in the local private school share on the probability that a student attends a private school. Column 1 of Table 4 reports the estimates of the coefficient on the proportion of private schools within a 20km radius of the student’s home (𝛿0 in equation (2) above). This treatment does not distinguish between academic and vocational schools. The result suggests that a 10 percentage point rise in the proportion of private schools results in a 1.1 percentage point rise in the probability of attending a private school. This represents a change in private school enrolment rate of around 10 percent from the baseline levels in 2001.32
Exploiting variation in the vocational and academic track provision, columns 2 and 3 report the impact on the private vocational and private academic track choices, respectively.
These correspond to estimates of model (2) above. Column 2 demonstrates that the impact of a 10 percentage point rise in the share of private vocational schools results in a 0.6 percentage point rise in the probability of attending a private vocational school. This effect is statistically significant at the 1 percent level. The impact of a rise in the share of private academic schools on the probability of attending a private vocational school is small and statistically insignificant.
Next, column 3 investigates the impact on private academic track school attendance.
This shows that a 10 percentage point rise in the share of private academic schools results in a 0.6 percentage point rise in the probability of attending a private academic track school. This impact is statistically significant at the 1 percent level. There is no statistically significant impact of the share of private vocational schools on this outcome.
Columns 4 to 6 in Table 4 produce estimates using the sibling fixed effect models corresponding to equation (4) above. This set of results is broadly in line with those using the
32 Recall that private school are significantly smaller than public schools, and hence even a relatively large increase in the share of private schools may result in a relatively small change in the private school enrolment probability. For example, Figure 1 demonstrates that although the share of private schools rose from around 20 percent to nearly 50 percent, the share of students enrolled in private schools rose by just over 10 percentage points.
full sample in the first half of Table 4: the overall impact on private school choice in column 4 is virtually identical to that in column 1, whilst the results for private vocational and academic tracks in columns 5 and 6 are somewhat larger but nevertheless broadly in line with the earlier results reported in columns 2 and 3.
Table 4: Impact of Private School Availability on Private School Enrollment
(1) (2) (3) (4) (5) (6)
Full sample: Grid cell FE-specification Siblings sample: Mother FE-specification
Private Private
Vocational
Private Academic
Private Private Vocational
Private Academic
Share Private 0.1149*** 0.1135***
(0.0110) (0.0179)
Share Priv Academic 0.0107 0.0613*** -0.0199 0.0727***
(0.0078) (0.0082) (0.0127) (0.0130)
Share Private
Vocational 0.0577*** 0.0007 0.0790*** -0.0027
(0.0076) (0.0082) (0.0128) (0.0133)
Observations 781,469 781,469 781,469 370,805 370,805 370,805
R-squared 0.0430 0.0435 0.0354 0.0339 0.0357 0.0245
Number of grid cells 66,459 66,459 66,459
Number of mother FEs 171,061 171,061 171,061
Table notes: All regressions include year and labor market region by year fixed effects; municipality specific linear trends;
and the following covariates: Student level dummy variables for mother employed; father employed; mother having a completed post-secondary degree; father having a completed post-secondary degree; mother having at most a completed secondary degree; father having at most a completed secondary degree; mother being born in Sweden; mother being born in Europe (except Sweden), North America or Europe; father being born in Europe (except Sweden), North America or Europe; student being male; student being age 15; student being age 17; student having attended a private lower secondary school; student level continuous variables for student final grade sum from lower secondary school in level and square;
household disposable income in level and square, and municipality level variables for the cost per student in compulsory education (primary and lower secondary education); the municipality share of students attending private school in grade 9;
a dummy variable for the municipality having a left-wing political local majority; and, finally, at the grid cell level, the log of the number of age 16-individuals residing within 20km from the grid cell midpoint. All covariates are measured the year that the student enters upper secondary education. Year-specific coefficients are estimated for all covariates. Missing observations in covariates were replaced with a constant value, and dummy variables were added to indicate these occurrences.
Siblings are defined as having the same mother. The sibling sample is restricted to non-movers; defined as siblings for which the mother resides in the same grid cell measured the year the students turn 16.
Standard errors, clustered on 250-square meter grid cell, in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
Given that students select from one of the four mutually exclusive alternatives33 (public academic, public vocational, private academic, private vocational) a natural complement to the above analysis is a multinomial logit model of track choice. Table 5 presents the results from such an exercise. As before, our focus is the impact of the share of private schools near the student’s home on his or her schooling choices. In order to facilitate the estimations, we choose a simpler model for this estimation: it controls for student covariates, municipality fixed effects, municipality time varying covariates and year fixed effects. 34
Table 5 reports the impact of the share of private vocational and private academic schools on each of the four track options. The coefficients represent the average marginal effects, from a single multinomial logit model. The results suggest that a 10 percentage point rise in the share of private academic schools on average leads to a statistically significant 0.9 percentage point fall in the probability of attending a public academic track school and a 1.2 percentage point rise in the probability of attending a private academic track school. There is a 0.3 percentage point fall of private vocational track outcomes in response to private academic school expansion, but no estimated effect on public vocational school attendance.
Table 5: Multinomial Logit Model Estimates
(1) (2) (3) (4)
Public Academic
Public Vocational
Private Academic
Private Vocational
Share Private Ac -0.0863*** 0.0004 0.1178*** -0.0319***
(0.0073) (0.0055) (0.0059) (0.0041)
Share Private Voc -0.0120* -0.0562*** -0.0001 0.0682***
(0.0070) (0.0053) (0.0054) (0.0040)
Observations 773,855
Log likelihood -710552.37
The lower number of observations compared to Table 4 is due to that a small number of the students in our regression sample attend the preparatory track. They are included in the linear probability models but are excluded from this multinomial logit regression.
The table shows the average marginal effects, computed using the margins-command in Stata.
The regression model includes municipal and year fixed effects, and the same set of region and student covariates as Table 4 (but in contrast to Table 4, the covariates are not estimates separately by year).
Standard errors, clustered on 250-square meter grid cell, in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
A similar pattern, but at a slightly smaller magnitude is found for the private vocational school expansion: A 10 percentage point rise in the share of private vocational schools on average leads to a rise of 0.7 percentage point in the probability of attending a private vocational track school and a 0.6 percentage point fall in the probability of attending a public vocational track school. Both these effects are statistically significant at the 1 percent level. In
33 There is also an option to enrol in a preparatory year, but, given that our regression sample contains only students qualify to enter directly into the regular Academic and Vocational tracks, this is very rare in our regression sample.
34 In addition, we do not estimate year-specific coefficients for the covariates in this model.
addition, there is also a fall of 0.1 percentage point in the probability of attending a public academic track school in response to a 10 percentage point rise in the share of private vocational schools, albeit significant only at the 10 percent level. There is no evidence of a statistically significant impact on private academic track choice.
In line with the results from the linear probability models presented above (Table 4), the results from the multinomial logit model demonstrate that a rise in the private school share leads students to substitute public for private schools, within the academic and vocational track, respectively. There is also some evidence that a higher private academic school share increased the likelihood to attend the academic track, by reducing the probability to attend the vocational track in a private school. A weakly significant, and small, impact on public school academic track attendance is also estimated from a rise in in the private vocational school share.
To sum up this section, our multinomial logit model was a simpler model in terms of having less detailed fixed effects and trend variables, but the fact that it gave similar conclusions as the linear probability models is reassuring. We return to using the linear probability models for the remaining analyses.
4.2 Track Choice
The results in Table 4 are informative about whether penetration of voucher schools from a particular track lead to greater enrolment into that type of private school. These results are not informative about the change in the overall mix of vocational versus academic choices arising from the expansion of voucher schools. For example, if there is a one-for-one substitution from public vocational to private vocational, then the net impact on take up of vocational courses would be zero. Table 5 indicated that such substitutions may take place.
In this section we ask whether the take-up of private schools simply represent substitution away from public schools of the same track, or whether the introduction of private schools in fact leads to a different set of track choices than would have been the case in the absence of private school expansion.
In order to undertake this analysis, we ask whether enrolling in a given track, whether public or private, is influenced by increased availability of private schools. Table 6 reports the results from this analysis. In column 1, for all three panels of Table 6, the outcome variable is attendance at a vocational school (private or public). Panel A reports the estimate of the coefficient on the proportion of private schools within a 20km radius of the student’s home (equation (2)). The result shows that a 10 percentage point rise in the private school share leads to a 0.7 percentage point rise in the probability of enrolling into a vocational track. 35
This is a key finding and demonstrates that private school penetration is not neutral in its impact on the mix of academic and vocational tracks chosen.
Estimates corresponding to model (2), reported in panel B, help in breaking down the overall impact reported in panel A. These show the separate effects of the share of academic and vocational private school tracks. The results in column 1, panel B shows that a 10 percentage point rise in the availability of private vocational schools leads to a 0.5 percentage
35 The estimated impact for academic track choice are close to equal in magnitude and opposite in sign relative to those for vocational track choice (results are available upon request). The reason for being “close to equal”
instead of perfectly equal in absolute size is that a low number of the students in our sample (fewer than 1%) attend the preparatory track even though they are qualified to enter a Vocational or Academic track.
