Separating gender composition effects from peer effects in education

ABSTRACT

This paper aims to demonstrate the importance of controlling for endogenous peer effects in estimating the influence of gender peer effects on educational outcomes. Using Manski's linear-in-means model, this paper illustrates that the estimation of gender peer effects is potentially biased in the presence of endogenous peer effect in education. The appropriate gender peer effect is estimated after identifying and controlling for the endogenous effect through the use of Graham's variance-restriction method.

KEYWORDS:

• Social interaction
• social multiplier
• gender peer effect
• INVALSI
• Project STAR

JEL CLASSIFICATION:

• I21
• J16

Acknowledgments

I thank Giulio Zanella for excellent research assistance and the editor and two anonymous referees and participants of the 5th International Workshop on Applied Economics of Education for their very helpful comments.

Disclosure statement

No potential conflict of interest was reported by the author.

ORCiD

Babak Jahanshahi http://orcid.org/0000-0002-8641-5543

Notes

1. In the case where group-level heterogeneity () is amplified by the social multiplier I can prove that the procedure to identify γ and the underlying assumptions are identical to the one of Graham. In order to avoid confusion the details of the computation are not provided here and will be available for interested readers upon request.

2. Following Graham closely the social multiplier is calculated for the US. However, the application of Graham's method for the Italian school is the main contribution of this part of the paper which requires more explanation.

3. This is the major requirement in identification of social multiplier as is explained in Section 2 and the appendix of the paper.

4. The National Institute for the Evaluation of the Education System.

5. The fact that the first stage of the estimation shown in Table 11 is highly significant confirms this assumption.

6. This test was first used by Ammermueller and Pischke (2006).

7. This test is suggested by Graham (2008) to ensure the stochastic separability assumption holds in US Project STAR.

8. This is a dichotomous variable which takes value 1 if the classes are on the right of the threshold (where large classes expected) and zero otherwise. This variable is a class size indicator to instrument within-school variation.

9. Between and within statistics are computed by Galbiati and Zanella (2012) in the case where group-level heterogeneity (α_c) is amplified by the social multiplier. I prove that the underlying assumptions one needs to identify γ are identical to the one of Graham. For simplicity the details of the computation is not written here and will be provided for interested readers upon request.