Edgeworth Expansions for FisherConsistent Estimators and Second Order Efficiency
Using a recent result of Bhattacharya, R. N. and Ghosh, J. K. ( Annals of Statistics, 1978, 434451), Edgeworth expansions of distributions of Fisherconsistent estimators for curved exponential family of parent distributions (dominated by the Lebesgue measure) can be obtained. We compare directly the first four cumulants of an arbitrary Fisherconsistent estimator with those of the MLE after correcting them for their bias. This leads to a key probability inequality which immediately implies the second order efficiency of the MLE wrt any bounded bowl shaped loss function. If the assumption of a dominating Lebesgue measure is dropped the formal Edgeworth expansions are no longer valid. However it turns out that if the loss function satisfies certain additional conditions the second order efficiency of the MLE holds. This modification takes care of the curved multinomial.