Class in the classroom: the relationship between school resources and math performance among low socioeconomic status students in 19 rich countries

Abstract

This paper investigates achievement gaps between low and high socioeconomic students in 19 high-income countries. On average, math scores of students with indicators of high socioeconomic status (SES) are over one standard deviation above those with low SES indicators. The paper estimates the extent to which these achievement gaps can be attributed to differences in classroom- and school-level resources available to students from different SES backgrounds. In some countries, achievement gaps can be largely explained by differences in the characteristics of schools attended. However, in many other countries, the gap appears more closely related to differences in the characteristics of the students. The results point to the importance of institutional difference among countries in explaining international differences in the quality of education received by different groups within a nation.

Keywords:

• math performance
• educational policies
• disadvantaged students
• international comparison

JEL classifications:

• I21
• I28
• I39

Acknowledgements

I would like to thank the US Department of Education’s National Center for Education Statistics for training support, Sara de la Rica, Mark Long and Ricardo Mora for helpful suggestions, two anonymous reviewers, and participants in a seminar at the Departamento de Fundamentos del Análisis Económico II, Universidad del País Vasco, and at the University of Washington’s Center for the Study of Demography and Ecology, for useful comments. A draft of this paper was presented at the International Association for the Evaluation of Educational Achievement (IEA) Third Annual Research Conference, Taipei, Taiwan, September 2008.

Notes

The Economist. 15 December 2007. Chile: Playground harmony, p. 54.

See http://web.worldbank.org/WBSITE/EXTERNAL/DATASTATISTICS/.

The regions are Indiana, the Basque Country, Flemish Belgium, Ontario, Quebec, and Hong Kong. The countries in the study are Australia, Italy, Japan, Korea, Netherlands, New Zealand, Norway, Singapore, Sweden, England, Scotland, Taiwan, and the USA.

Further on we address why we do not pool the data.

In a few instances, the unmeasured characteristics are negative (the model overestimates the test-score gap). In these cases, we leave PU out of the denominator, thus estimating the percent of the total test score gap as: Equation (2) (PS)/(PS+PB). As in Equation (3), the purpose of this is to bias downward our estimate of the importance of school characteristics.

This approach is similar to a fixed effect approach, but assumes some underlying distribution to these school level effects. As discussed below, in this study, we substitute the classroom level for the school level.

Using the average of the five plausible values does lead to a greater likelihood that the explanatory variables will be found to be significantly related to test scores; however it also generates unbiased estimates of the size of this relationship (the coefficient). For this paper, it is the size and not the statistical significance of the variables that we are most interested in, since the focus of the study is on the cumulative effect, rather than the effect of individual variables. Moreover, a comparison of HLM results using all five plausible values with the average value found no real difference in the coefficients or their significance.

Only 1.5% of students in this study did not answer the question concerning number of books in the home, whereas over 30% either did not answer or answered ‘I don’t know’ to questions concerning their parents’ educational attainment. Moreover, analyses indicate that these 30% missing responses are strongly correlated with ‘books in the home’, suggesting a strong non-random component to who did and did not answer questions about parent’s educational level. The overall correlation between books in the household and parents’ educational attainment is 0.36, and the correlation between these alternative ways of measuring the achievement gap in this study is 0.7.

As mentioned earlier, where on average the model underestimates the gap, the remaining unexplained gap we attribute to unobserved student characteristics, thereby giving a lower bound on the estimated importance of school characteristics.

The model used in this and the next section is identical to that used above, with two differences. First, estimates are derived based on OLS rather than HLM. This is because limiting the analysis to low SES student results in too few observations per classroom to use HLM. The second difference is that the dependent variable is the average of the five plausible values, in contrast with HLM which performs five separate regressions and reports the average estimated value and standard deviation of the coefficients. While using the average value does lead to a lower standard error, the estimated βs are unbiased, which is all that is used in estimating the counterfactuals in this study.

Wöβmann (2004), for instance, found that school characteristics are of greater importance for high scoring students.

In this manner, a student will of course be labeled as an H student if it is the unmeasured characteristics of the student’s school, rather than the students’ unmeasured individual or background characteristics, that leads to a higher test score. If true, then the H captures a more highly effective school than a more able or motivated or otherwise advantaged student. While we cannot identify the reason for the students’ higher-than-predicted performance, we believe that selecting for students in this way does reduce the heterogeneity among the students, and thus reduces the potential problem of self selection and omitted variables bias in this study, which is the objective of this additional estimation.

Evidence available from author.