A Sieve Approach to Nonparametric Estimation of Exchangeable Networks: Examples from International Trade. Abstract This paper considers the estimation and inference of nonparametric parameters of interest in a network model using the sieve extremum estimator. The main result is a central limit theorem for sieve extremum estimates that enables standard methods for inference of nonparametric parameters. The results rely on an assumption of exchangeability on the data generating process which in turn can be used to establish a conditional dependency graph to model the network interdependence. This varies from previous work on sieve extremum estimation that has only considered i.i.d. and weakly dependent data. The results in this paper justify the application of sieve estimation procedures to a wide range of network models. To illustrate this flexibility, I show how this method can be used to estimate infinite and finite-dimensional parameters in two models of international trade. I also show the equivalence of this setting to a generalization of psi-mixing developed for use in network models.
Robust Propensity Score Estimation and the Overlap Condition. Abstract A large body of research in the social and physical sciences is concerned with the estimation of causal effects in observational studies. These analyses often employ propensity score matching methods to recover these parameters. Unfortunately, this methodology relies on strong assumptions on the support of the propensity score that are not always satisfied. This paper attempts to provide a rigorous analysis and assessment of the propensity score overlap condition. To this end, I introduce the support vector machine (SVM) as a powerful statistical approach to estimate propensity scores and assess the overlap assumption. I use simulated data to show the comparative advantage of this methodology relative to previous methods especially in the case where there is a nonlinear determination of treatment. Finally, I illustrate how to apply this methodology using the commonly studied LaLonde (1986) experimental and non-experimental data sets.