some recent developments in econometric test methodology

abstract (introduction): a crucial element in the development of econometric methodology over the past decade or so has been the concern with testing, as opposed to estimating, econometric models. this has been brought about partly as the result of the subject reaching a new level of maturity and partly because the lack of success of econometric modelling in the seventies, particularly at the macro level, made clear the need for a deeper concern with model evaluation. modern econometric practice advocates that a given specification should be subject to a rigourous testing procedure and it is now becoming routine to test for misspecifications such as omitted variables, serially correlated disturbances and structural change and, in addition,to test for heteroscedasticity and incorrect functional form. this kind of intensive misspecification testing leads to problems of distortions in the inference procedures but leading econometricians believe that the importance of carrying out such tests overrides these problems. ... hendry advocated that this should be done notwithstanding the difficulties involved in calculating and controlling type 1 and type 2 errors. while it is important to test econometric models rigourously, it is also important to seek to structure the testing procedure in such a way that problems of data mining are minimised. in particular we seek test procedures to test for the presence of, possibly, several misspecifications simultaneously in such a way that: (a) the overall type 1 error probability is controlled within acceptable limits, and (b) the test procedure while having good power properties provides some opportunity for detecting individual types of misspecification. in this paper i consider some recent advances in test methodology which contribute to the development of such procedures. there are two general approaches to conducting tests for misspecification in econometric models. in the first approach, we obtain some sample statistic whose distribution is known under the null hypothesis, i.e. when the model is assumed correct, and if the statistic assumes a significant value this is taken as evidence that the model is misspecified in some unknown way. the d.w. statistic is sometimes used for such a test because it is relatively sensitive to various departures from the null hypothesis. in this case we do not necessarily regard a significant test result as implying that the disturbances are first order serially correlated and proceed with a cochrane orcutt type estimation procedure. the significant test result is simply taken as evidence that something is wrong. tests of this type are known as pure significance tests and do not require the specification of a particular alternative model.;