Working Paper: CEPR ID: DP17941
Authors: Martin O'Connell; Howard Smith; Øyvind Thomassen
Abstract: In GMM estimators moment conditions with additive error terms involve an observed component and a predicted component. If the predicted component is computationally costly to evaluate, it may not be feasible to estimate the model with all the available data. We propose an estimator that uses the full data set for the computationally cheap observed component, but a reduced sample size for the predicted component. We show consistency, asymptotic normality, and derive standard errors and a practical criterion for when our estimator is variance-reducing. We demonstrate the estimator's properties on a range of models through Monte Carlo studies and an empirical application to alcohol demand.
Keywords: GMM estimation; micro data
JEL Codes: C20; C51; C55
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Full sample for observed component (C29) | More efficient estimator (C51) |
| Smaller sample for predicted component (C29) | More efficient estimator (C51) |
| Covariance between observed and predicted components small relative to variance of observed component (C29) | Proposed estimator outperforms small-sample GMM estimators (C51) |
| Proposed estimator exhibits consistency and asymptotic normality (C51) | Lower variance under certain conditions (C29) |
| Proposed estimator maintains precision close to full-sample GMM estimators (C51) | More accurate and efficient estimation in large datasets (C51) |