ABSTRACT

Straightforwardapplicationofusualinferenceprocedures(testsandconfidenceregions)ineconometricsisoftenimpossible.Theproblemusuallycomesfromaninsufficientspecificationoftheprobabilitydistributiongeneratingtheasoccurs forexamplewhenonemakesassumptionsonlyonthefirstfewmomentsofanerror distribution.However,theproblemalsoarisesinparametricmodelsthatspecifythe datageneratingprocessuptoafinitevectorofparameters.Thisistypicallythecase whentheassumptionsmadeonthedistributiondepartfromthosemadeinthestandardlinearregressionframework,suchastheabsenceofserialindependenceand homoskedasticity.