NONPARAMETRIC INFERENCE BY QUASI-LIKELIHOOD METHODS

Richard Spady

Nuffield College, Oxford University

 

March 1996

 

Abstract

Two techniques for nonparametric inference concerning statistical functionals are examined as quasi-likelihood methods. Empirical likelihood, introduced by Owen (1988), and the least-favorable family construction of DiCiccio and Romano (1990) are considered as defining test statistics through maximization problems. Using first-order theory only, a statistic derived from the latter approach outperforms the second-order (Bartlett) corrected version of the former in all of the cases considered in DiCiccio, Hall, and Romano (1991), and the fully iterated bootstrap results for variances reported in Lee and Young (1995). This suggests that coverage accuracy comparable to the computationally intensive iterated bootstrap can be achieved without recourse to simulation or higher-order asymptotic calculations.