STATISTICS COLLOQUIUM

THURSDAY, May 14th, 2009 at 4.10pm, MSB 1147 (Colloquium Room)
Refreshments: 3.30pm, MSB 4110 (Statistics Lounge)

Speaker: Guido Kuersteiner (Economics, UC Davis)

Title: Bandwidth Choice for Bias Estimators in Dynamic Nonlinear Panel Models

Abstract: This paper considers bandwidth selection for spectral density estimators based on panel data sets. The spectral densities of greatest interest in this paper are the ones that appear in the bias expression for fixed effects estimators in nonlinear dynamic panel models obtained by Hahn and Kuersteiner. The bias estimation problem is different from the usual zero frequency spectral density estimation problem because the need for positive definiteness does not arise. As a consequence, the usual justification for kernel smoothing of spectral estimators does not apply to this case.

However, without kernel smoothing the bandwidth selection problem is significantly more difficult because in this case not only the usual proportionality constants are data-dependent but also the optimal rate at which the bandwidth parameter grows with sample size. In this paper an infinite order VAR model is used to obtain an estimate of the approximate mean squared error of the spectral estimator. It is shown that selecting the bandwidth parameter based on the estimated mean squared error criterion is asymptotically equivalent to the optimal infeasible bandwidth choice. Monte Carlo simulations show that truncated spectral estimates significantly outperform kernel weighted estimates in terms of their effectiveness in reducing bias in the panel application. In addition to the specific results in this current paper I will also talk a bit more broadly about the work I have done on estimating panel data models with fixed individual effects.

Joint work with Jinyong Hahn.