Joint work with Santiago Pereda Fernández. Published in: Econometric Reviews, 44(9), September 2025 (pp. 1321-1360).
Abstract. Conventional methods for the estimation of peer, social or network effects are invalid if individual unobservables and covariates correlate across observations. In this paper we characterize the identification conditions for consistently estimating all the parameters of a spatially autoregressive or linear-in-means model when the structure of social or peer effects is exogenous, but the observed and unobserved characteristics of agents are cross-correlated over some given metric space. We show that identification is possible if the network of social interactions is non-overlapping up to enough degrees of separation, and the spatial matrix that characterizes the co-dependence of individual unobservables and covariates is known up to a multiplicative constant. We propose a GMM approach for the estimation of the model’s parameters, and we evaluate its performance through Monte Carlo simulations. Finally, we revisit an empirical application about classmates in college. Contrasting with conventional methods, our methodology can estimate zero, non-significant peer effects on both academic performance and major choice.
Abstract. This paper develops an estimator for quadratic forms in semiparametric models where covariates enter through an unknown function. The leading application is a generalized AKM wage decomposition: we estimate worker and firm variance components and sorting while flexibly partialling out observed worker and firm characteristics. The estimator combines leave-one-out bias correction with series approximation, is robust to heteroskedasticity and many fixed effects, and is consistent when both fixed effects and basis terms grow with sample size under a strengthened smoothness condition. Simulations show large bias reductions relative to plug-in estimators in nonlinear settings. In linked Portuguese employer-employee data, adding rich controls to linear AKM generates large discontinuities, whereas the generalized estimator yields coherent movements across control sets and basis richness, with sorting attenuating toward zero as flexibility increases.