Title: Outlier testing in robust two-stage least squares models
Authors: Jonas Kai Kurle - University of Oxford (United Kingdom) [presenting]
Xiyu Jiao - University of Oxford (United Kingdom)
Abstract: A frequent concern in applied economics is that key empirical findings may be driven by a tiny set of outliers. To perform outlier robustness checks in instrumental variables regressions, the common practice is first to run ordinary two-stage least squares (2SLS) and classify observations with residuals beyond a chosen cut-off value as outliers. Subsequently, 2SLS is re-calculated based on the non-outlying observations, and this procedure may be iterated until robust results are obtained. However, the above trimmed 2SLS has a positive probability of finding outliers even when the data generating process contains none. To answer the question of whether observations are correctly classified as outliers, the false outlier detection rate (gauge) is studied asymptotically using an empirical processes argument. The established asymptotic theory of the gauge forms a basis for tests for the overall presence of outliers. Simulation studies lend further support to the theory, and an empirical illustration shows its utility.