Running Gauss code in Ox

RunningGAUSSprogramsUnderOx3PhilipA.VitonOctober9,2003Contents1Introduction22Updates23PrinciplesofConversion34OxEdit44.1ObtainandinstallOxEdit...........................44.2Con…gureOxEditforOx............................54.3Con…gureOxEditforOxGauss........................54.4SetuptheRunbuttons............................54.5One-linecomments..............................65Anextendedexample65.1GetandinstalltheGausscode........................65.2Preliminarysteps................................75.3Round1—relocatingthecode........................75.4Round2—initializingglobals........................85.5Round3—disambiguatingthecode.....................95.6Round4—maxlik...............................95.7Round5—load...............................105.8Round6—indirection.............................115.9The…nalcode..................................115.10Usabilitynotes.................................125.10.1Optimizationfailures..........................125.10.2Robuststandarderrors........................125.10.3Initialpoints..............................135.10.4Dataconversion............................136CaveatandConclusion1511IntroductionThisnotedescribeshowtogetGausscodetorununderJurgenDoornik’sOxsystem,version3.x.ItismeantasasupplementtoJurgen’sowndiscussioninOxAppendix.pdf,distributedwithOx.Frommyperspectiveasacompletebeginner,theproblemwithJurgen’swrite-upwasthattheexamplesturnedouttobe“tooeasy”toconvert,anddidn’tprovideguidanceonwhattodowhenthingsdidn’tgoquitethatsmoothly.Fortheacademicuser,theadvantageofOx+Gauss(OxGauss)isthatitisine¤ectafreesubstituteforGaussitself.(Notethatthisappliesonlyforacademicresearchandteaching:Oxisotherwisenotfree.)Asalong-timeuserofGaussremarkedtomerecently,Aptechseemstobemakingitmuchmoredi¢cultandexpensivetouse(forexample,moredi¢culttohaveacopyofthesoftwareforhomeusewhenyouhavealicensedo¢ceversion):indeedhewentsofarastoguessthatifthiskeepsup,“Gauss’daysarenumbered”.Idon’tknowwhetherthisisanover-reaction,butifitistrue,OxGaussprovidesasolution.ThisnoteisaimedattheGaussuserwhojustwantstogetprogramsrunningunderOxasquicklyaspossible.Forthisreason,IincludedetailsonsettingupOxEdit,thefreeOxtexteditor,forusewithOxGauss.Ifyouhaveanothereditorwhichallowsyoutorunprogramsandcapturetheiroutput,youshouldbeabletogetitworkingwithOxGaussquiteeasily,thoughsyntaxhighlightingmaytakeabitmoree¤orttosetup.IfyoucametoOxafterreadingFranciscoCribari-Neto’sreviewintheJournalofAppliedEconometrics(Vol12,1997,pp.77–89),there’soneveryimportantpointtobemade.FranciscomentionstheavailabilityofaGauss-to-Oxconverter(g2ox.exe).Inmyview,whileusingthisconvertermighthavebeennecessaryinpreviousversionsofOx,youshouldnowavoiditcompletely.Asofversion3,OxcanrunGaussprogramsonitsown,withverylittle…xuponyourpart(aswe’llseebelow),andwithoutconvertingthemtotheOxlanguage.Ifyoudodecidetogiveg2oxatry,bepreparedforlotsof(fairlyroutine)work;youcouldeasilyspend…vetotentimesaslongontheconversion(expertocrede)thanifyouuseOx’s“native”abilities,asdescribedbelow.2UpdatesThissectionlistsupdatestothediscussion,foreasyreference.October7,2003:ItappearsthatOx3.3hasimprovedtheOxGausssubsystemtothepointthatverylittleofthediscussionofhowtoconvertGausscodeforuseinOxappearsnecessary.Inparticular,Train’smixed-logitcoderunsout-of-thebox—noreorderingnecessary—ifyoujustchangethedata-loadingstatementasdescribedinSection5.7.Also,2version3.3issigni…cantlyfasteratcomputingtherobuststandarderrorsduetoachangeinthewayGauss’gradpprocedureisimplementedinOx.Obviously,everyoneshouldupgrade.Inmyoriginalwrite-upInotedthatonecouldn’trunGaussroutineswhichinvokeitsextra-costmaxlikmodule,andIsuggestedthatitmightbepossibletomapcallstomaxlikintocallstoOx’sownoptimizers.ThishasnowbeendonebySébastienLaurentandJean-PierreUrbain.Usingtheirworkitispossibletorunthemixed-logitcodeevenwiththemaxlikoptionspeci…ed(acoupleofotheradd-onoptimizersarealsosupported,butI’veexperimentedwiththem).Onereasonthisisimportantisthattheincludeddomaxoptimizerishighlysensitivetoanumberorparameters,forexampletheinitialpointandthenumberofHaltondraws.TheimplementationofmaxlikforOxGaussisavailablehere.March5,2003:VERYIMPORTANT:Train’soriginalcodeappearstohavecontainedanerror,whosesymptomwasthatyougotidenticalresultswhenestimatingamixedlogitmodelwithuniformdraws,nomatterhowmanyrepetitionsyouused.Theerrorhasbeensilentlycorrectedinthelatestversionsfromhiswebsite.Youshouldcheckthatyouranswerschangewiththenumberofrepetitionsbeforeyoubelievetheresults.Itisnotclearwhatwasbeingestimatedinthiscase.Theerrordoesnotseemtoa¤ectestimationwithHaltonnumbers.Asofnow,Trainappearstohavewithdrawnthetwozip…lesoriginallyreferredtobelow.Insteadthereisasinglezip…le,train0401.zip,containingavarietyof…lesassociatedwithacoursehetaughtonsimulationmethodsindiscretechoice.Themixedlogit…lesareinthesubfolderps05.mxl,andthemain…leismxlp.g.Thisisapanel-dataformulationofthemodel;youcanobviouslyadaptittoanon-panel-datacontextbyspecifyingTIMES=1.Becausehistwooriginalarchivesarenolongeravailable,Ihavewithdrawnreferencestomyconvertedco