Supplement to “identifying Social Interactions through Conditional Variance Restrictions: Additional Empirical Results, Specification Tests, and Power Calculations

SINCE THE ESTIMATOR PROPOSED in the paper is based on a simple conditional moment restriction, standard approaches to specification testing are available (e.g., the Sargan–Hansen test of overidentifying restrictions). However, as in any application, it is helpful to think about specific directions of misspecification and to construct tests accordingly. As an example, note that an implication of Assumption 1.2 (as stated in the paper) is that class size/type and teacher characteristics are ‘stochastically separable’ in the production of academic achievement. This appendix outlines a test for this assumption. Assume that teachers have L latent attributes ac = (ac1 acL)′ Conditional random assignment ensures that