Can Schools Reduce the Indigenous Test Score Gap? Evidence from Chile

Abstract

In Chile, indigenous students obtain lower test scores, on average, than non-indigenous students. Between two cohorts of eighth-graders in the late 1990s, the test score gap declined by 0.1 to 0.2 standard deviations. An Oaxaca decomposition and related descriptive evidence suggest that the most plausible explanation is related to Chile's large-scale school reforms that were targeted at low-achieving schools and students. The paper evaluates and rules out alternate explanations such as relative improvements in indigenous socioeconomic status, and sorting of indigenous students between schools. The results highlight a potential lever for reducing earnings gaps between indigenous and nonindigenous adults.

Acknowledgements

I am grateful to Gregory Elacqua and many officials of Chile's Ministry of Education for providing data. Gillette Hall, Joseph Shapiro, Miguel Urquiola, Emiliana Vegas and two anonymous referees provided helpful comments. The research received financial support from the World Bank, although the views expressed are only attributable to the author.

Notes

1. The countries include Bolivia (Vera, 1998; McEwan, 2004), Canada (Ma and Klinger, 2000), Ecuador (Garcia Aracil and Winkler, 2004), Guatemala (Wu et al., 2003; Hernandez-Zavala et al., 2006; McEwan and Trowbridge, 2007), Mexico (Hernandez-Zavala et al., 2006), Peru (World Bank, 2001; Hernandez-Zavala et al., 2006), and the US (Freeman and Fox, 2005). See McEwan and Trowbridge (2007) for a summary.

2. Ethnographic evidence from region nine suggests that gaps within schools might be traced to low teacher expectations for indigenous students, and use of instructional materials more suited to non-indigenous, Spanish-speaking students (Herrera Lara, 1999).

3. The 1980s were characterised by public school decentralisation, privatisation and few direct interventions in school quality by the national Ministry of Education (McEwan and Carnoy, 2000; Hsieh and Urquiola, 2006).

4. This paper's task is similar to US literature that explores the causes for the decline of the black-white test score gap in the 1980s and its later stagnation or increase in the 1990s (Grissmer et al., 1998; Hedges and Nowell, 1998, 1999; Cook and Evans, 2000; Hanushek, 2001; Neal, 2005).

5. The 1980s reforms may have also played a role, albeit more indirect, in affecting test score gaps by removing constraints on school choice and perhaps allowing indigenous students to choose higher-quality schools. Nonetheless, both cohorts of students in this paper's data were exposed to a similar regime of school choice, and the evidence presented below does not suggest that re-sorting of indigenous students across schools can explain test score gaps.

6. The 2002 census asked: ‘Do you belong to one of the following first or indigenous nations?’ (author's translation). The mutually-exclusive options included Alacalufe (Kawashkar), Atacameño, Aimara, Colla, Mapuche, Qechhua, Rapa Nui, Yámana (Yagán), and none of the above.

7. Among Chile's smaller indigenous groups, the Aymara are concentrated in the far north (region one), near Bolivia, and the Atacameño in region two.

8. Chilean household surveys began collecting data on indigenous status in 1996. Using similar questions, the CASEN household surveys in 1996 and 2000 identified 4.5 per cent and 4.4 per cent of the population, respectively, as indigenous (MIDEPLAN, 2002). The CASEN household survey in 2003 found a slightly higher proportion indigenous (5.4%; author's calculations with CASEN 2003 data). The discrepancy could be due to sampling error or the use of imperfect weights derived from population projections based upon the 1992 census.

9. The 1992 census tabulations only report the indigenous population aged 14 or more (INE, 1993). Thus, the absolute number of the indigenous in 1992 would be even higher. In calculating the indigenous proportion, Gacitúa-Marió (2002) used the total population, ages zero and above, as the denominator, thus reporting a lower proportion.

10. The 1992 census asked: ‘Do you feel that you belong to one of the following cultures?’ (author's translation). Only three mutually exclusive options were provided (Mapuche, Aymara, and Rapanui).

11. The 2000 CASEN survey yields similar proportions (MIDEPLAN, 2002).

12. It asks: ‘In Chile, the law acknowledges the existence of eight first nations or indigenous peoples; does the student's mother belong to one of them?’ (author's translation). The mutually exclusive options are Atacameño, Aymara, Rapa-Nui, Quechua, Mapuche, Coya, Káwaskar and Yagán.

13. That is generally the case in countries without universal primary enrolments such as Guatemala (McEwan and Trowbridge, 2007).

14. The estimates are obtained from the 2003 CASEN household survey.

15. In 1980, the Ministry of Education decentralised the management of public schools to more than 300 municipalities (McEwan and Carnoy, 2000; Hsieh and Urquiola, 2006). It also began financing schools via a per-student subsidy based on monthly attendance. Private schools were eligible to accept such subsidies if they did not charge tuition. Parents were able to attend either public or private schools, presuming they were admitted. The reform precipitated a large-scale expansion of private schooling throughout the 1980s.

16. The 1997 data have also been used to analyse the effectiveness of private schools (McEwan, 2001) and the existence of peer-group effects (McEwan, 2003)

17. For an analysis of indigenous test score gaps in the 1999 cross-section, see McEwan (2004).

18. The original income categories in each year are more extensive and do not always overlap. They were collapsed to facilitate comparisons between the two survey years.

19. Within each year's sample, mean differences between indigenous and non-indigenous students are estimated by regressing test scores on Indigenous. Robust standard errors are adjusted for the clustering of students within schools.

20. Difference-in-differences are estimated by pooling the 1997 and 2000 samples and regressing students test score on Indigenous, a dummy variable indicating the 2000 sample, and the interaction between the two (where the last coefficient measures the difference-in-differences). Robust standard errors are adjusted for the clustering of students within year-by-school cells.

21. The US literature on black-white test score gaps has usually focused on mean test score differences. For an exception, see Clotfelter et al. (2006).

22. According to Cox (2004), it is possible to equate the 1997 and 2000 tests given anchor items duplicated on the 1997 and 2000 tests. He summarises such an exercise, and it shows that equated mean test scores in Spanish and mathematics are stable across time (implying similar mean achievement on anchor items).

23. For applications of the Oaxaca decomposition to the wage differences between indigenous and non-indigenous adults in Latin America, see Psacharopoulos and Patrinos (1994) and Hall and Patrinos (2006).

24. The present decomposition, and similar literature cited in footnotes 1 and 4, focuses on explaining mean test score gaps, and changes in mean gaps. Given the relative lack of empirical evidence in Chile or Latin America, this paper emphasises the application of well-known techniques for describing and decomposing mean test score differences. Nonetheless, the evidence in Table 2 suggested, primarily for math scores, that declines were more substantial at select percentiles of the indigenous and non-indigenous test score distributions. In future work, the author will explore this further with quantile regressions, though neither fixed-effects nor a Oaxaca-type decomposition can be applied without added methodological considerations (for example, see Koenker, 2004; Machado and Mata, 2005).

25. Many control variables are missing values for some observations (see the sample sizes in Table A1). To avoid reducing sample size, the author arbitrarily recode them to their sample means, and further control for dummy variables equal to one for missing values, and zero otherwise (Krueger and Zhu, 2002). The coefficients on these additional variables, though not reported in the tables, are included in regressions and later decompositions. Conducting the analyses with a smaller sample that excludes all observations with missing data does not substantially alter the results.

26. Also see Cook and Evans (2000), McEwan (2004), and McEwan and Trowbridge (2007).

27. Note that the second term could also be weighted by the 2000 coefficients, and the third term by the 1997 mean differences, as in , a well-known feature of an Oaxaca-type decomposition. All subsequent estimates were repeated both ways, but it does not substantially affect the results.

28. A potential disadvantage of Oaxaca-type decompositions based on linear regressions is that observations in the indigenous and non-indigenous groups may not fall within a region of common support (see especially Ñopo, 2004 and its references). That is, some indigenous children may be observationally quite dissimilar from non-indigenous children, and vice-versa, and the decomposition implicitly relies on projections of outcomes outside the observed range for such students. To assess the sensitivity of results, probits within each cross-sectional sample were estimated, regressing the indigenous dummy variable on the full set of family and school characteristics listed in Table A1. Propensity scores for each student were then calculated. In 1997 (2000), the region of common support includes 98.0% (97.7%) of non-indigenous students and 97.9% (98.6%) of indigenous students. The author then dropped observations outside of the region of common support and re-estimated the fixed-effects regressions in Table 3. The estimates and resulting decomposition results were not substantively different from the full sample estimates.

29. McEwan (2004) carried out a similar exercise; the results are similar, but not identical because of small sample differences.

30. For a similar discussion in the context of black-white test score convergence in the US, see Hanushek (2001).

31. Due to grade repetition, some eighth grade students may have spent more than eight years in school and, therefore, entered earlier than 1990 or 1993. The issue is less severe in Chile than other Latin American systems because of generally low rates of grade repetition, less than 3 per cent in the first grade (McEwan and Shapiro, 2008).

32. Declines in the test score gap between schools could have occurred even if indigenous students in the 2000 cohort were not more likely to be targeted by P-900 than students in the 1997 cohort. Anecdotal and empirical evidence that the programme became more effective over time, probably because programme interventions were added in the 1992 school year (Chay et al., 2005). Thus, the 2000 cohort may have been exposed to a relatively more effective intervention.

33. As Chay et al. (2005) note, there is no evidence on the relative cost-effectiveness of P-900 and other reforms in improving student achievement.

34. Despite this evidence, there is a surprising lack of descriptive evidence on how knowledge of native languages has shifted over time, and how such shifts may be related to children's school performance.

35. It may also be the case that non-indigenous families with high-achieving children become more likely to identify as indigenous over time. This seems implausible, given that indigenous identify is stigmatised in Chile, revealed in attitude surveys (see Gacitúa-Marió, 2002 and the citations therein).

36. The indigenous are also more likely to fall below the extreme poverty line (11% of the indigenous population, compared with 5.4% of the non-indigenous population). For similar data on other Latin American countries, see Psacharopoulos and Patrinos (1994), de Ferranti et al. (2004), World Bank (2004), Hall and Patrinos (2006).

37. In particular, see de Ferranti et al. (2004: 94) and Psacharopoulos and Patrinos (1994: 214). In each case, the results are obtained from an application of the Oaxaca decomposition to estimates from Mincer earnings regressions estimated separately among indigenous and nonindigenous adults.

38. It is also plausible that the gap reflects differences in Spanish language skills – as might be captured by test scores – an interpretation forwarded in the Bolivian context by Chiswick et al. (2000).