欧洲央行-IV回归中的贝叶斯推理-2026.pdfVIP

WorkingPaperSeries

DomenicoGiannone,MicheleLenza,BayesianinferenceinIVregressions

GiorgioE.Primiceri

No3189

Disclaimer:ThispapershouldnotbereportedasrepresentingtheviewsoftheEuropeanCentralBank

(ECB).TheviewsexpressedarethoseoftheauthorsanddonotnecessarilyreflectthoseoftheECB.

Abstract

ItiswellknownthatstandardfrequentistinferencebreaksdowninIVregressionswith

weakinstruments.Bayesianinferencewithdiffusepriorssuffersfromthesameproblem.

Weshowthattheissuearisesbecauseflatpriorsonthefirst-stagecoefficientsoverstatein-

strumentstrength.Incontrast,inferenceimprovesdrasticallywhenanuninformativeprior

isspecifieddirectlyontheconcentrationparameter—thekeynuisanceparametercapturing

instrumentrelevance.TheresultingBayesiancredibleintervalsareasymptoticallyequiva-

lenttothefrequentistconfidenceintervalsbasedonconditioningapproaches,andremain

robusttoweakinstruments.

JELCodes:C11,C26,C36,C55

Keywords:bayesiananalysis,concentrationparameter,weakinstruments

ECBWorkingPaperSeriesNo31891

Non-TechnicalSummary

Estimatingcausaleffectsineconomicsisdifficultwhenkeyexplanatoryvariablesareen-

dogenous.Inthesesituations,researchersoftenrelyoninstrumentalvariablestorecovercausal

relationships.Akeyeconometricchallengearises,however,whentheseinstrumentsareweak—

thatis,onlyweaklyrelatedtotheendogenousvariable.Inthatcase,manystandardstatistical

methods,bothc