A general approach to Lagrange multiplier model diagnostics (Q801625)

A diagnostic test can be formulated by taking the current model as the null hypothesis and a particular form of misspecification as the alternative. The general approach of this paper is to calculate the score for some alternative and base a test on its distribution under the null. This distribution is derived in two theorems corresponding to different models. Such a test is a Lagrange multiplier test and is asymptotically optimal. These tests can be formulated as ways of looking at the residuals for particular types of non-randomness. The tests are simple to calculate even when the alternative may be very complicated. Examples are given of tests for omitted variables and block-recursive structures.