Attention-Dependent Monetary Transmission to Household Beliefs

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Attention-Dependent Monetary Transmission to Household Beliefs Jaemin Jeong, Eunseong Ma, Choongryul Yang 2025-084 Please cite this paper as: Jeong, Jaemin, Eunseong Ma, and Choongryul Yang (2025). “Attention-Dependent Monetary Transmission to Household Beliefs,” Finance and Economics Discussion Series 2025-084. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2025.084. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Attention-Dependent Monetary Transmission to ∗ Household Beliefs JaeminJeong† EunseongMa‡ ChoongryulYang§ DukeUniversity YonseiUniversity FederalReserveBoard September2025 Abstract WhendohouseholdslistentotheFed?Weshowtheanswerliesinasimplebutpowerful force:householdattentiontomacroeconomicconditions.Wedevelopamodelwhereattentionactsasacrucialgatekeeperforthepass-throughofpolicynewstobeliefs,andconfirm itspredictionsusinghouseholdsurveydata.Wefindthatbeliefrevisionstomonetarypolicy surprisesareconcentratedamongattentiveindividuals—particularlythosewithhighfinancialstakes—andthiseffectstrengthensdramaticallyduringuncertaintimes.Thisimpliesthe expectationschannelismostpotentwhenitmattersmost,suggestingpolicymakersshould accountforthetime-varyingandheterogeneousnatureofpublicattention. Keywords:Inflationexpectations,Monetarypolicy,Rationalinattention,Behavioralmacroeconomics JELCodes:D83,D84,E31,E52 ∗ WethankChaeWonBaek,AndrewFigura,ByoungchanLee,EkaterinaPeneva,JaneRyngaert,andparticipants at2023SNDEYoungEconomistWorkshopandthe16thJointEconomicsSymposiumofSixLeadingEastAsian Universitiesfortheirvaluablecommentsandsuggestions. Wegratefullyacknowledgefinancialsupportsfrom R.K.ChoResearchClusterProgram. Theviewsexpressedherearethoseoftheauthorsanddonotnecessarily reflectthoseoftheFederalReserveBoardortheFederalReserveSystem.Firstversion:January2024.Thisversion: September2025 †419ChapelDrive,Box90097,Durham,NC27708,U.S.A.Email:jaemin.jeong@duke.edu. ‡50Yonsei-ro,Seodaemun-gu,Seoul03722,SouthKorea.Email:masilver@yonsei.ac.kr. §20thStreet&ConstitutionAvenueNW,Washington,DC20551,U.S.A.Email:cryang1224@gmail.com. 1

1 Introduction Central banks emphasize expectations as an important channel of monetary transmission. Yetwhenhouseholdsactuallyupdatetheirinflationbeliefsinresponsetopolicynews—and which households do so—has been hard to pin down empirically. This paper studies when households“listen”totheFed. Ourcentralclaimisthatattentiontomacroeconomicconditions isakey,heterogeneous,andtime-varyingdeterminantofthepass-throughfromconventional monetarypolicy(MP)surprisestohouseholdinflationexpectations.Wecombineasimplemodel ofendogenousattentionwithnewmicroandtime-seriesevidencefromalong-runningU.S. householdsurveyandexternallyidentifiedpolicyshocks.Fourkeyresultsemerge:attentiongates theindividual-levelimpactofMPonbeliefs;aggregatepass-throughscaleswiththeeconomy’s averageattentiveness;theeffectstrengthensinperiodsofelevateduncertainty;andtheresponse islargestforhouseholdswithhigherpayoffstobeinginformed. We begin with a minimal behavioral framework, following Gabaix (2020), in which each householdchoosesanattentionlevelpriortothearrivalofshocksandformsexpectationsasan attention-weightedcombinationofalong-runanchorandthefullyinformedforecast. Attention balancesforecast-lossreductionsagainstmentalcostsandisincreasinginthepayoff-relevant newsvariance—thevolatilityofmonetaryandnon-monetarydisturbancesthatwouldmove thefullyinformedforecast. Themodeldeliversfourtestableimplications: (i)onlytheattentive componentofbeliefsloadsonpolicynews(attentiongatespass-through);(ii)aggregatepassthroughintimeseriesisproportionaltoaverageattentiveness;(iii)higheruncertaintyraises attentionandthereforeamplifiesbeliefresponsestopolicy;and(iv)pass-throughislargerfor householdswithhigherpayoffstoinformation(e.g.,stockholdersandhomeowners),consistent withahigherbenefitparameterinthemodel. WethentakethesepredictionstothedatausingtheMichiganSurveyofConsumers(MSC). Exploitingitsrotatingpanel,weconstructapredeterminedattentivenessindicatorbycontrasting respondents’assessmentsofrecentbusinessconditionswithanexternalbenchmark. Monetary policysurprisesareidentifiedwithhigh-frequencymethods. Ourempiricalstrategytestseach ofthemodel’spredictions: webeginwithamicroevent-studyoftheeffectofconventionalMP surprisesonrevisionsinone-year-aheadinflationexpectations,followedbyatime-seriesregressionthatteststhescalingwithaggregateattentiveness. Wethenanalyzestatedependenceby interactingshockswithmacrouncertaintyand,finally,testthepayoff-heterogeneitypredictions usinghouseholdcharacteristicsincludingstockholding,homeownership,age,andincome. Foursetsoffindingsaligncloselywiththemodel’spredictions. First,inthemicrodata,a contractionaryshockreducesone-year-aheadinflationexpectationsonlyamongrespondents classifiedasattentive;theestimateforinattentiverespondentsissmallandstatisticallyindistin- 2

guishablefromzero. Thisindividual-levelpatternistheattention-gatedpass-throughpredicted bythemodelanddirectlylinkspolicysurprisestobeliefupdateswhenattentionishigh. Second, in a time-series design that splits months by ex ante economy-wide attentiveness, the passthroughofacontractionarymonetarypolicyshockislargeandnegativeinhigh-attentiveness monthsandnearzerootherwise,consistentwithaggregatepass-throughbeingproportionalto averageattention. Third,pass-throughstrengthensinmoreuncertainperiods—duringrecessionsandwhenrealorfinancialuncertaintyiselevated—andthisamplificationisconcentrated amongtheattentivehouseholds. Thesefactsmatchthecomparativestaticsthatoptimalattentionriseswithpayoff-relevantnewsvarianceandhelpreconcilewhymeasuredeffectsofMP on the economy could vary across environments (Vavra, 2014, Tenreyro and Thwaites, 2016, Alpanda, Granziera and Zubairy, 2021). Fourth, consistent with the model’s payoff logic, we findsystematicheterogeneityintheresponse. Amongattentiverespondents,stockholdersand homeownersexhibitanespeciallylargepass-through,whileyoungerandmiddle-agedindividualsreactmorestronglythanolderones. Thesepatternsconfirmthepredictionthatgroups withahigherstakeintheeconomyareendogenouslymoreresponsivetopolicynews. Theyalso complementagrowingbodyofevidenceonfirmattentionheterogeneityandtheefficacyof MP(e.g.,AfrouziandYang,2021,Yang,2022,Afrouzi,2024,Wu,2024). Ourfindingsprovidea household-levelanalogue: justasmorecomplexfirmspaycloserattention,householdswith greaterfinancialstakesaremoreattunedtopolicynews. Forbothfirmsandhouseholds,higher attention leads to expectations that align more tightly with fundamentals and react more to policynews. Ourcontributionistoshow,inasingleframeworkanddataset,thathouseholds’attention mediateshowconventionalmonetarypolicyshockspassthroughtoinflationexpectations,that averageattentivenessorganizesthestrengthoftheexpectationschannelovertime,andthatthe effectbecomesstrongerinmoreuncertainperiodsandforhouseholdswithhigherpayoffsto information. Conceptually,theresultsunderscorethattheexpectationschannelisattention sensitive: thesamepolicyactioncanhavesharplydifferenteffectsonbeliefsdependingonhow muchattentiontheaudienceendogenouslydevotestomacroeconomicnews. Inpractice,they suggestthatcommunicationstrategiesandpolicyevaluationsshouldaccountforvariationin attentivenessacrossgroupsandovertime. Thispaperbridgestheoriesofinattentiveexpectationswithempiricsonthemonetarytransmission of beliefs. On the theory side, our setup nests classic information frictions—sticky information and rational inattention (Mankiw and Reis, 2002, Sims, 2003, Mac´kowiak and Wiederholt,2009)—withinthebehavioralexpectationsoperatorofGabaix(2020),andrelates tobroaderbounded-rationalityapproaches(AngeletosandLian,2018,Bordalo,Gennaioliand Shleifer,2018). Ontheempiricalside,weconnecttoworkonlimitedinformationandlearning 3

amonghouseholdsandfirms(CoibionandGorodnichenko,2015a,Candia,CoibionandGorodnichenko,2024),theeffectsofcentral-bankcommunicationsonhouseholdbeliefs(Carvalho andNechio,2014,LamlaandVinogradov,2019,ClausandNguyen,2020,KryvtsovandPetersen, 2021, Coibion, Gorodnichenko and Weber, 2022, Bauer, Pflueger and Sunderam, 2024), and experience/salienceinexpectationformation(MalmendierandNagel,2016,Cavallo,Crucesand Perez-Truglia,2017,DAcunto,Malmendier,OspinaandWeber,2021). Ourcontributionistofuse thesestrandsbyembeddingclassicinformationfrictionswithinabehavioralexpectationsmodel thatdeliverssharp,state-contingentpredictionsforbeliefupdatingafterexternallyidentified MPshocks,andtestingthesepredictionsusingawidelyusedhouseholdsurveybymeasuring attentivenesspriortopolicynewsandshowingthatitgovernswhoupdates,byhowmuch,and when. Householdandfirmattentivenesstoinflationhasbeenmeasuredinseveralcomplementary ways. One strand uses “revealed attention” from search behavior and news supply, such as internet search for inflation-related queries and counts of inflation articles in major outlets (Kumar,Coibion,AfrouziandGorodnichenko,2015,MarcellinoandStevanovic,2022,Korenok, Munro andChen, 2023). Pfäuti (2024)infersattentionfromupdatingbehavior, estimatinga time-varyingattentionparameterfromhowstronglyshort-runinflationexpectationsloadon recentinflationandclassifying“high-attention”regimeswhenthisresponsivenessexceedsan estimatedthreshold. Kroner(2025)introducesacomplementarypre-announcementindexof investorattentionaroundCPIreleasesaggregatesnewscoverage,mainstreammediamentions, andGooglesearchintensityforinflationintoaCPI-attentionmeasureusedtopredictmarket reactions. Micro-basedapproachescomplementtheseaggregatesbyinferringattentiveness directlyfromsurveybehavior(e.g.,BraitschandMitchell,2022,SongandStern,2024). Inparticular, BrachaandTang(2024)proxyinattentionfromtheMSC’stwo-stepinflationmodule: among respondents who first say prices will “stay the same,” low attention is flagged if they answer“don’tknow”atthenumericfollow-upor,iftheygiveanumber,whenitdepartssubstantially from contemporaneous inflation. Relative to these papers, our contribution is to measureattentivenessattherespondentlevelbeforepolicynewsandconnectittoexternally identifiedmonetarypolicyshocks,showingthatattentiongovernswhoupdates,howmuch,and when—andthataggregatepass-throughscaleswithindependentlymeasuredattentivenessover time. Thisbridgesaggregatesearchandnews-basedindicatorsandmicroconsistency-based measuresbyprovidingadirect,policy-linkedmappingfromattentiontobeliefupdating. Recentevidenceindicatesthatinattentionitselfisendogenousandvarieswiththeenvironment: wheninflationormacroriskishigh,agentsacquiremoreinformationandalignbeliefs morecloselywithfundamentals(FlynnandSastry,2024,Weber,Candia,Afrouzi,Ropele,Lluberas,Frache,Meyer,Kumar,Gorodnichenko,Georgarakos,Coibion,KennyandPonce,2025). 4

Webuildontheseinsightstoprovideaunified,micro-foundedexplanationofhowattention shapestheMPexpectationschannelwhenpolicyshocksareidentifiedexternallyandattentivenessismeasuredbeforetheshockrealizes.Wealsospeaktostatedependenceinmonetarypolicy. Whilepriorexplanationsemphasizenon-linearpricing(Vavra,2014),andbroadernonlinear propagation(TenreyroandThwaites,2016), wehighlightaninformationalchannel: inmore volatileoruncertainenvironments,agentsendogenouslyraiseattention,whichamplifiesthe beliefsresponsetopolicy. Thismechanismcomplementsrecentevidenceontime-varyingfirm inattentionandMPefficacy(SongandStern,2024). Thepaperisorganizedasfollows. Section2presentsthebehavioralexpectationsmodeland testableimplications. Section3describesthedataandtheconstructionoftheattentiveness proxy. Section4reportsthemainempiricalresults,andSection5providesrobustnesschecks. Section6concludes. 2 Behavioral Expectations with Endogenous Attention Thissectiondevelopsaminimalbehavioralframeworkinwhichhouseholdschoosehowmuch attentiontodevotetoinflation-relevantnews. Buildingonthebounded-rationalexpectations operatorofGabaix(2020)andtheendogenous-attentionlogicusedinDietrich(2024),wederive fourtestableimplicationsthatguideourempiricalworkinSections3and4: (i)attentiongates thepass-throughofmonetarypolicy(MP)shockstohouseholdinflationexpectations;(ii)aggregateMPpass-throughintimeseriesscaleswiththeeconomy’saverageattentiveness;(iii)state dependenceisstrongerforalready-attentiveagents,ashigherpayoff-relevantuncertaintyraises attentionandamplifiesresponses;and(iv)payoffheterogeneity: groupswithahigherbenefit ofbeinginformed(largerω )orlowerattentioncosts(smallerκ )choosemoreattention,are i i morelikelytobeclassifiedasattentive,andexhibitlargerpass-through. Section3introduces ourempiricalproxyforattentiveness;Section4implementsthecorrespondingtests. 2.1 Setup Timing. Atthestartofmontht,householdi choosesattentionm ∈[0,1]. Thentheperiod-t i,t shocks are realized, and the household forms a one-year-ahead inflation expectation using abehavioraloperator. Westudytheimpact changeinexpectationsaroundtheshockarrival (holdingπ fixedandvaryingonlythenewsrealizedwithint). t Inflationfundamentals. Thefullyinformed(rational)forecastofnext-periodinflationis π∗ = π¯+ρ(π −π¯)+θεmp+Γ(cid:48)εo, (2.1) t+1 t t t 5

whereπ¯ isthesteady-stateanchor,ρ∈(0,1),εmp istheMPsurprise,andεo ∈(cid:82)K stacksother t t contemporaneousdisturbances(e.g.,markup,energy/importprices,wagegrowth,commodity, taxchanges). ThescalarθandvectorΓ=(γ ,...,γ ) (cid:48) aresemi-elasticitiesmappingstandardized 1 K innovationsintothefullyinformedforecast. Weadoptthesignconventionthatcontractionary monetarypolicyshockslowerthefullyinformedinflationforecast,implyingθ<0. Shocknormalizationandcovariance. Wenormalizetheshockstobemean-zeroGaussian: εmp ∼N (0,1), εo∼N (cid:161) 0, Σ (cid:162) , t t o,t whereΣ isaK ×K positivesemidefinitecovariancematrixwithonesonthediagonal. Unless o,t statedotherwise,weassumeCov (εmp ,εo)=0withintheidentificationwindow;off-diagonal t t t elementsofΣ allowcontemporaneouscorrelationamongnon-MPshocks.1 o,t Behavioralexpectationsandattentionchoice. Householdi formsabehavioralexpectation byblendingacoarseanchorwiththefullyinformedforecast: (cid:69)B i,t (cid:163)π t+1 (cid:164) = (1−m i,t )π¯ + m i,t (cid:69) t (cid:163)π∗ t+1 (cid:164) , (2.2) where(cid:69) [·]isthefull-informationconditionalexpectation.2 Giventhemarginalbenefitofbeing t informedω andattentioncostsκ ,theagentchoosesm tominimizeastandardquadraticloss i i i,t function—whichcanbeviewedasasecond-orderapproximationtoamoregeneralproblem— thatbalancesforecastinaccuracyagainstmentalcosts:3 1 κ m = arg min ω U (1−m)2+ i m2, (2.3) i,t i t m∈[0,1]2 2 withclosed-formsolution ω U m ∗ (U ) = i t ∈ [0,1]. (2.4) i,t t ω U +κ i t i 1Anyunconditionalvariancescanbeabsorbedinto(θ,Γ). TimevariationinΣ captureschangingmacro o,t uncertaintyacrossstatesoftheworld. 2Forsimplicity,wemodelthelong-runanchorπ¯asfixed.Thisassumptioncouldberelaxedtoatime-varying anchor,π¯ ,toaccountforpotentialshiftsintheinflationregime.Ourmodel’scoremechanismremainsunchanged, t asthehousehold’sbehavioralexpectationinEquation(2.2)wouldsimplybecome(cid:69)B i,t (cid:163)π t+1 (cid:164) = (1−m i,t )π¯ t + m (cid:69)(cid:163)π∗ (cid:164) . Thekeyprediction—thatthepass-throughofashockεmp ,whichrepresentsnewsrelativetothe i,t t t+1 t currentanchor,isscaledbyattentionm —isrobusttothisextension. i,t 3Thepayoffparameterω canbemicro-foundedbylinkingittohouseholdeconomicdecisions.Forinstance,ina i consumption-savingproblemwithutilitydependingontheperceivedrealinterestrate,thelossfrommis-forecasting inflationislargerforhouseholdswithnominallyexposedbalancesheets(e.g.,netnominalassetsormortgagedebt), makingω anendogenousfunctionofthoseexposures.Inaheterogeneous-agentrational-inattentionmodelwith i homeownersandrenters,Ahn,XieandYang(2024)showthatthepayoffparameteriscloselylinkedtosteady-state mortgagedebt.Forparsimony,wetreatω asareduced-formparameter,whichissufficientforourcomparative i staticsandtestableimplications. 6

Here U ≡ Var (cid:161)π∗ (cid:162) = θ2Var (εmp ) + Γ(cid:48)Σ Γ + 2θCov (cid:161)εmp ,Γ(cid:48)εo(cid:162) , (2.5) t t t+1 t t o,t t t t isthepayoff-relevantnewsvarianceatthetimeattentionischosen. Underthebaselinenormalizationandorthogonality, Var (εmp )=1, Cov (cid:161)εmp ,εo(cid:162)=0 ⇒ U = θ2+Γ(cid:48)Σ Γ. t t t t t t o,t ∗ Intuition. Optimal attention m rises when the incoming news that would move the fully i,t informedforecastismorevolatile(largerU ,∂m ∗ /∂U >0),whenattentionismorevaluable t i,t t forthehousehold(higherω ,∂m ∗ /∂ω >0),andfallswhenattentionismorecostly(higherκ , i i,t i i ∂m ∗ /∂κ <0).4 i,t i 2.2 TestableImplications WenowcharacterizeindividualandaggregateresponsestoacontractionaryMPsurprise(εmp >0 t withθ<0). ProofsaredeferredtoAppendixA. Proposition2.1(AttentiongatesMPpass-through).Forhouseholdi,theimpactchangeininflationexpectationsinresponsetoacontractionaryMPsurprise(εmp >0withθ<0)is t ∆π i,t+1 ≡ (cid:69)B i,t (cid:163)π t+1 (cid:164)−π i,t = θm i ∗ ,t (U t )εm t p . Proof. SeeAppendixA.1. Proposition2.1showsthatthepass-throughofpolicynewsisscaledbythehousehold’slevel ofattention,amechanismwhereattentionmediatestheresponse. InSection4.1,wewilltestthis mechanismbyinteractingMPsurpriseswithanattentivenessproxytoshowthattheresponseis concentratedamongagentsweclassifyasattentive. Proposition2.2(Aggregateattentivenessraisestime-seriespass-through).Let∆πe denotethe t+1 aggregate(e.g.,meanormedian)revisionininflationexpectations. AggregatingEquation(2.2) acrosshouseholdsyields (cid:90) ∆πe = Λ θεmp + υ , Λ ≡ m ∗ di, (2.6) t+1 t t t t i,t (cid:124)(cid:123)(cid:122)(cid:125) ∈[0,1] 4ω scales the marginal loss from forecast errors (“benefit of being informed”) while κ captures cognii i tive/opportunitycosts.Heterogeneityin(ω ,κ )willmapintocross-sectionaldifferencesinpass-through. i i 7

whereΛ istheaverageattentivenessintheeconomyandυ collectsaggregationresidualsorthogt t onaltoεmp . t Proof. SeeAppendixA.2. The time-series impact of conventional MP on aggregate belief revisions scales with the economy’s average attention. In Section 4.2, we will sort months by aggregate attentiveness andshowthattheMPslopeislargeandnegativeinhigh-attentiveregimesandnegligiblein low-attentiveregimes. Proposition 2.3 (State dependence is stronger for more attentive households).LetU be the t payoff-relevantnewsvarianceinEquation(2.5). ForacontractionaryMPsurprise(θ<0), ∂∆π m ∗ (U ) (cid:161) 1−m ∗ (U ) (cid:162) i,t+1 =θεmp i,t t i,t t < 0, ∂U t U t t sohigherU makestheexpectationdecline more. Ifgroup A ismoreattentivethangroup I at t eachU (i.e.,m (U )>m (U )),then t A t I t (cid:175) (cid:175) (cid:175) (cid:175) (cid:175)∂(cid:161) m (U )θ(cid:162) /∂U (cid:175) > (cid:175)∂(cid:161) m (U )θ(cid:162) /∂U (cid:175) (cid:175) A t t(cid:175) (cid:175) I t t(cid:175) whenever m (U ) (cid:161) 1−m (U ) (cid:162) > m (U ) (cid:161) 1−m (U ) (cid:162) . A simple sufficient condition is if both A t A t I t I t groups’attentionisbelowthispeak,i.e.,0≤m <m ≤ 1. I A 2 Proof. SeeAppendixA.3. Endogenousattentioncreatesstatedependence: whentheenvironmentismoreuncertain (largerU ),attentiveagentsreducetheirinflationexpectationsbymoreafteracontractionary t MPshock,andthesensitivitytoU isitselfstrongerforthealready-attentivegroup. t Proposition2.4(Payoffheterogeneityandcross-sectionalpass-through).FixU >0. Lethouset holdsdifferonlyin(ω ,κ )inEquation(2.3)–Equation(2.4). Then: i i 1. Attentionordering. m ∗ (U )isstrictlyincreasinginω andstrictlydecreasinginκ (i.e., i,t t i i ∂m ∗ /∂ω >0and∂m ∗ /∂κ <0). i,t i i,t i 2. Pass-throughordering. TheindividualMPpass-throughmagnitude, (cid:175) (cid:175) ∂∆π i,t+1 (cid:175) (cid:175)=|θ|m ∗ (U ), (cid:175) ∂εmp (cid:175) i,t t t isstrictlyincreasinginω andstrictlydecreasinginκ . i i 8

3. Selection into “attentive/accurate”. For any threshold τ∈(0,1), the probability of being classifiedasattentive(accurate) A =1{m ∗ ≥τ}isweaklyincreasinginω andweakly i,t i,t i decreasinginκ . i 4. Conditionalorderingwithintheattentivegroup. Amongagentswith A =1,thecondii,t tionalpass-through|θ|(cid:69)[m ∗ | A =1]islargerforgroupswithhigherωand/orlowerκ i,t i,t (wheneverthesupportofm ∗ haspositivemeasureaboveτ). i,t Proof. SeeAppendixA.4. Groupsforwhomreducingforecasterrorsismorevaluable(higherω )orlesscostly(lower i κ )choosehigherattention,aremorelikelytobeclassifiedasattentiveunderanyfixedthreshold, i and,crucially,displaylargerMPpass-throughandstrongerstatedependence. InSection4.4,we willtreathomeowners,stockholders,prime-age,andhigher-incomehouseholdsasempirical counterpartsofhigher-ω(and/orlower-κ)groups,andtestthecorrespondingcross-sectional predictions. In sum, the simple behavior expectations model delivers four testable implications: (i) attention gates the impact of MP shocks on individual expectations; (ii) aggregate MP passthroughscaleswiththeeconomy’saverageattentiveness;(iii)higherpayoff-relevantuncertainty strengthenspass-through—especiallyforalready-attentiveagents;and(iv)groupswithhigher payofffrominformation(largerω )orlowerattentioncosts(smallerκ )choosemoreattention i i andexhibitlargerpass-through.5 InSection3,wedefinetheempiricalattentivenessproxyand construct the aggregate attentiveness index used to verify these predictions. Section 4 then implementsthecorrespondingmicroandtime-seriestests. 3 Data Thissectiondescribesthedatasetsandtheconstructionofourempiricalattentivenessproxy, whichwewilltaketothetestsimpliedbySection2. Wefirstoutlinesourcesandsampledefinitions,thenconstructanindividual-levelaccuracyindicator(ourproxyforattentioninthemodel), 5Theseresultsdonotrelyonalinearattentionweightorquadraticattentioncosts.Moregenerally,if(i)higher attentionplacesmoreweightonthefullyinformedforecastand(ii)thementalcostofattentionisconvex,thenthe optimalattentionchoiceriseswiththevolatilityofpayoff-relevantnewsandwiththebenefit/stakesω ,andfalls i withthecostparameterκ .Theconclusionsalsosurviveacommon,noisypublicsignalaboutfutureinflationthat i arrivesbeforeattentionischosen:amoreprecisesignalreducesresidualforecastuncertaintyandmaycompress averageattention, butconditionalonthechosenattentionthepolicypass-throughtermisstillscaledbythe attentionweight,sotheindividualgating,aggregatescaling,andpayoff-heterogeneityimplicationsareunchanged. Becausethesignaliscommon,itsleveleffectisabsorbedbytimecontrolsanddoesnotaffecttheestimatedslopes inourempiricalsetting.SeeAppendixBfordetails. 9

andfinallydefineanaggregateattentivenessindexusedinourtime-seriesexercises. Section4 willbringthesemeasurestothemicroandaggregateregressionsimpliedbyPropositions2.1–2.4. 3.1 SourcesandSamples Micro survey and demographics. Our micro data come from the Michigan Survey of Consumers(MSC),whichinterviewsanationallyrepresentativesamplemonthlyandre-interviews arotatingpanelofrespondentsroughlysixmonthslater. Weusetherotating-panelstructure toconstructrevisionsinexpectationsattheindividuallevelandtocontrolforobservedheterogeneity(age,income,education,homeownership,stockownership,gender,region). TheMSC providesone-year-aheadinflationexpectationsandarichsetofqualitativequestionsonrecent business conditions. We focus on the one-year horizon because it is standard for near-term transmission,alignswithoursix-monthpanelandidentificationwindow,andisthemeasure mostresponsivetocontemporaneousmacroandpolicynewsinhouseholddata(e.g.,Cavallo etal.,2017,Coibionetal.,2022,DAcunto,MalmendierandWeber,2023). Ourbaselinemicro samplespansSeptember1998toMarch2020,whichistheintersectionofMSCavailabilityfor thenecessaryitemsandtheavailabilityofourhigh-frequencymonetarypolicyshocks.6 Monetarypolicyshocks. Ourbaselinemeasureofmonetarypolicy(MP)surprisesusesthe high-frequencyseriesfromNakamuraandSteinsson(2018),asextendedbyBauer,Lakdawala andMueller(2022). Thesesurprisesareidentifiedfromchangesinfederalfundsfuturespricesin anarrowwindowaroundFOMCannouncementsandarestandardintheliterature. Apotential concernwiththisapproachisthatitmaycapturenotjustpurepolicyactionsbutalsoaFed “information effect.” We retain this series as our baseline because its narrow identification windowiscrucialforpreciselytimingpolicynewsrelativetooursurvey’sinterviewdates. To ensureourresultsarenotdrivenbyinformationeffects,weconfirmourfindingsusingalternative shocksfromBu,RogersandWu(2021)thataredesignedtopurgesucheffects.Forourtime-series analysisoftheGreatModeration(Section4.2)), wealsousethenarrative-basedshocksfrom RomerandRomer(2004). Toensureconsistentinterpretationacrossallspecifications,wenormalizetheshockseries. First, we set the sign so that a positive value always represents a contractionary surprise (an unexpectedpolicytightening).Second,wescaletheseriessothataone-unitchangecorresponds toaone-percentage-point(100basispoint)tightening. Thisnormalizationallowsourreported regressioncoefficientstobeinterpreteddirectlyasthepercentage-pointresponseofinflation 6WedropNovember2002andMay2003duetomissingstock-ownershipinformation. FollowingBachmann, BergandSims(2015),wetrimobservationswithabsoluteone-year(orfive-year)inflationexpectationsabove20% tomitigateoutliers. 10

expectationstoaone-percentage-pointpolicyshock.7 Ouranalysisusesallidentifiedsurprises, both contractionary and expansionary. For expositional clarity, we discuss the effects of a “contractionary”shockinthetext,asthemodel’spredictionsaresymmetric. Othermacroseries. WeobtainourmacroeconomicdatafromtheSt. LouisFederalReserve’s FREDdatabase. Weusetheunemploymentrate,IndustrialProduction(IP),inflation,andthe NationalFinancialConditionsIndex(NFCI)aseitherbenchmarksforourattentivenessproxy orascontemporaneouscontrols. Forourstate-dependenceanalysis,weusetheNBER-dated recessionindicatorandtheCBOEVolatilityIndex(VIX). 3.2 MeasuringAttentiveness: AnAccuracyProxy Section2formalizesattentionasalatentweightm ∗ ∈[0,1]. Inthedata,weproxyattentiveness i,t withapre-determinedindicatorbasedoneachrespondent’squalitativeassessmentofrecent businessconditions,recordedatthefirstinterview. Step1: Perceivedbusinessconditions(favorable/unfavorable/nonews). Atthefirstinterviewinmonth t,eachrespondentreportswhethertheyhaveheardfavorableorunfavorable changesinbusinessconditionsin“thelastfewmonths,”orhavenotheardofchanges. Wecode athree-waycategoricalvariable News ∈{Fav, Unfav, Haven’theard}, i,t whichrecordsthesignoftherespondent’sperceivedbusinessnewsattimet (orlackofexposure). Step2:Benchmarkforbusinessconditions. Toconstructouraccuracybenchmark,weseeka macroeconomicindicatorthatiscanonical,widelyreported,andmapscloselytothesurvey’s phrasing of “changes in business conditions.” The unemployment rate is arguably the most salientandeasilyunderstoodmeasureofrealeconomichealthforthegeneralpublic.Specifically, we compare perceived favorability with the three-month change in the unemployment rate tosmoothouthigh-frequencynoisewhilestillcapturingtherecenteconomicdevelopments respondentswereaskedabout: ∆Unrate t ≡Unrate t −Unrate t−3 . 7Inpanelspecificationswecumulatetheannouncement-windowshocksfromt tot+5tomatchthesix-month interviewhorizon. 11

Whileweviewthisasthemostnaturalbenchmark,weconfirmtherobustnessofourfindings usingalternativerealandfinancialindicatorsinSection5. Accuracyclassification. Wedefinethreemutuallyexclusivegroupsatthefirstinterviewdatet:  Accurate ifFav&∆Unrate <0, orUnfav&∆Unrate ≥0,   t t   Accuracy = Inaccurate ifUnfav&∆Unrate <0, orFav&∆Unrate ≥0, i,t t t     Haven’theard otherwise. Forestimation,weencodeattentivenessusingathree-waysetofmutuallyexclusiveindicators, A =(cid:161) 1{Accurate }, 1{Inaccurate }, 1{Haven’theard } (cid:162) , i,t i,t i,t i,t andusethecorrespondinggroupdummiesinourspecifications(withonecategoryomittedas thereferencegroup). Timingandidentification. Crucially, the attentiveness indicators, A , are measured at the i,t first interviewinmontht,priortotheFOMCannouncementwindowthatdefinesthemonetary policy surprise εmp . Hence they are pre-determined with respect to the shock. Under our t high-frequencyidentification, (cid:69)(cid:163)εm t p(cid:175) (cid:175)A i,t ,X i,t ,α t (cid:164)=0, whereX collectsobservedcovariates(agebins,education,income,homeownership,stocki,t holding,gender,region,maritalstatus,andsurvey-modecontrols)andα aremonth-yearfixed t effectsthatabsorbcommonmacro/newsvariation. Thistiming,combinedwiththeexogeneity ofhigh-frequencysurprises,formsourkeyidentifyingassumption,allowingustointerpretthe coefficientsontheinteractiontermsasthedifferentialpass-throughofpolicynews,rulingout reversecausalityorwithin-monthinformationacquisition. Descriptivestatisticsbyaccuracygroup. Table1reportsrespondentcharacteristicsacrossthe threegroups.Thegroupsarebalancedinthesample(Accurate:37.2%,Inaccurate:29.7%,Haven’t heard: 33.1%). Accurate and Inaccurate respondents look strikingly similar on observables: homeownership(83.2%vs.81.9%), stockholding(76.4%vs.75.6%), education(about56%vs. 55%withacollegedegree),age(35-64: 63.4%vs.61.7%;65+: 22.0%vs.23.3%),gender,region, maritalstatus,andaverageincome(both∼$94k). Bycontrast,theHaven’theardgroupdiffers systematically: lowerhomeownership(76.9%),lowerstockholding(60.8%),lowereducational attainment(36.4%withacollegedegree;5.6%lessthanhighschool),youngeronaverage(18-34: 12

Table1: DemographicandSocioeconomicCharacteristicsbyAttentivenessGroup Accurate Inaccurate Haven’tHeard PanelA:Homeownership (1)Homeowner(%) 83.2 81.9 76.9 (2)Renter(%) 16.8 18.1 23.1 PanelB:Stockownership (3)Stockholder(%) 76.4 75.6 60.8 (4)Non-stockholder(%) 23.6 24.4 39.2 PanelC:Educationlevel (4)Grade0-8nohsdiploma(%) 0.5 0.6 1.5 (5)Grade9-12nohsdiploma(%) 1.6 1.3 4.1 (6)Grade0-12w/hsdiploma(%) 16.1 16.3 28.0 (7)Grade13-17nocoldegree(%) 25.7 26.7 29.8 (8)Grade13-16w/coldegree(%) 30.7 29.3 22.8 (9)Grade17w/coldegree(%) 25.2 25.5 13.6 PanelD:Age (10)18-34(%) 14.4 14.8 22.8 (11)35-64(%) 63.4 61.7 53.3 (12)65+(%) 22.0 23.3 23.7 PanelE:Gender (13)Male(%) 56.3 56.2 53.1 (14)Female(%) 43.6 43.7 46.8 PanelF:Region (15)West(%) 22.3 22.2 20.5 (16)NorthCentral(%) 27.0 27.0 27.8 (17)Northeast(%) 17.4 17.4 16.3 (18)South(%) 33.2 33.2 35.2 PanelG:Maritalstatus (19)Married/partner(%) 67.2 67.0 60.0 (20)Divorced(%) 13.5 13.5 13.9 (21)Widowed(%) 6.3 6.3 8.4 (22)Nevermarried(%) 12.8 13.0 17.4 PanelH:Averageincome (23)Averageincome 93,911.9 93,886.0 71177.6 Total(%) 37.2 29.7 33.1 Notes:Table1reportsrespondentcharacteristicsbyattentivenessgroup(Accurate,Inaccurate,Haven’tHeard). Allentriesarecolumnpercentagesunlessotherwisenoted;“Averageincome”ismeannominalhouseholdincome (USD).Demographiccategoriesincludehousingtenure,stockholding,education,age,gender,region,marital status,andincome.Samplecoversfirstinterviewsfrom1998m09–2020m03.SeeSection3fortheconstructionof theattentivenessmeasureandvariabledefinitions. 22.8%),lesslikelytobemarried/partnered(60.0%),andloweraverageincome($71.2k). These patternsareconsistentwithinterpretingouraccuracyindicatorA asanattentivenessproxy i,t 13

ratherthanaproxyforfixedtraits;observablecompositiondifferencesareconcentratedinthe Haven’theardcategory,whileAccurateandInaccuraterespondentsaresimilaronobservables. Section4willcontrolforthefullsetofdemographicsinallspecifications. Discussion of the accuracy proxy. Our accuracy-based indicator is a noisy measure of the latentattentionvariable,m ,inourmodel. Tostrengthenthetheoreticaljustificationforthis i,t proxy,wegrounditdirectlythroughthelensofourmodel. Proposition2.4(part3)providesa directmotivationtoourempiricalclassification: A =1{m ∗ ≥τ},whereanagentisclassified i,t i,t as “attentive” if their chosen attention level m ∗ surpasses a certain threshold τ required for i,t accurateperception. Fromthisperspective,our“Accuracy”indicatorisnotjustaproxyforthecontinuouslatent ∗ variablem ,butanempiricalimplementationofthistheoreticalclassificationrule. Anagent i,t isclassifiedas“Accurate”becausetheirattentionlevelwassufficientlyhightocorrectlyparse the direction of recent economic news. This approach still allows for misclassification—an attentiveagent(m ∗ ≥τ)mightmisreadaspecificsignal,oraninattentiveone(m ∗ <τ)might i,t i,t guesscorrectly—whichwouldinduceattenuationbiasandworkagainstusfindingsignificant results. Nonetheless,therobustalignmentofourempiricalresultswithallfourofthemodel’s predictions, asshowninSection4, suggeststhatthisproxysuccessfullycapturesthissalient theoreticaldimensionofhouseholdattentiveness. Onemightbeconcernedthatour“Accuracy”proxycapturesfactorsotherthanattention, suchascognitiveabilityorpoliticalbias. Whilewecannotrulethesechannelsoutentirely,three piecesofevidencepointtowardanattentioninterpretation. First,ourframeworkprovidesa unifiedexplanationforthefullsetofourfindings: thescalingofaggregatepass-through,the amplificationduringuncertaintimes,andthestrongerresponseamonghigh-payoffgroupslike stockholdersandhomeowners,anditislessclearhowtheseotherfactorswouldjointlyexplain thisspecificconstellationofresults. Second,therobustnessofourmainempiricalfindingsto usingdifferentmacroindicators,asshowninSection5,mitigatesconcernsthattheresultsare drivenbyaspecificpoliticalnarrativetiedtounemployment. Third,politicalcompositionis balancedacrossattentivenessgroupssuggestingthatpartisanbiasisnottheprimarydriverof theclassification.8 Takentogether,thesefactsmakeitdifficultfornon-attentionexplanationsto jointlyaccountfortheconstellationofpatternswedocument. 8WeclassifypoliticalstancerelativetothesittingU.S.presidentatthetimeofthefirstinterview. Supporters arerespondentswhoself-identifywiththepresident’sparty;Opponentsidentifywiththeout-party;Independents includeself-reportedindependents,otherparties,andnopreference. AmongAccuraterespondents,32.5%are supporters,30.3%opponents,and36.7%independents,withsimilarsharesfortheInaccurategroup. 14

3.3 AggregateAttentivenessIndex TotestProposition2.2intimeseries,weconstructanaggregateattentivenessmeasureasthe cross-sectionalshareofattentiverespondentsatthefirstinterviewdatet: A agg ≡ 1 (cid:88) Nt A ∈[0,1], t N i,t t i=1 agg where N isthenumberofrespondentswithnon-missingA . WeuseA directlyasacont i,t t tinuous index and, for regime analyses, define high-attentive months as those in the upper agg quantileofA (e.g.,top30%intheGreatModerationsubsample)andlow-attentivemonths t agg asthecomplement. Byconstruction,A istheempiricalcounterparttothemodel’saverage t attentionΛ =(cid:69) [m ∗ ]inProposition2.2. t i i,t 3.4 VariableAlignmentandConstructionNotes Expectationrevisions. Forindividuali,wecomputetherevisioninone-year-aheadinflation expectationsoverthesix-monthpanelwindow,aligningthetimingsothatthefirstinterview (where A ismeasured)precedestheMPshockandthesecondinterviewfallsatt+h (typically i,t h = 6 months). Aggregate revisions ∆πe (e.g., median) are computed analogously across t+h individualsinterviewedinmontht andre-interviewedint+h. Controlsandscaling. Whenused,contemporaneousmacrocontrolsaremeasuredbetween thefirstandsecondinterviews(e.g.,∆IPand∆πfromt tot+h). Monetaryshocksarecumulated fromt throught+h−1tomatchthesurveyhorizonwhenappropriate;Section4reportsthe exacthorizonchoiceandrobustnesstoalternatives. 4 Empirical Results Thissectionbringsthemodel’spredictionstothedata. WetestfourimplicationsfromSection2 usingthemeasuresdefinedinSection3.First,atthemicrolevel,attentiongatesthepass-through of contractionary monetary policy surprises to one-year-ahead inflation expectations: only attentive(accurate)respondentsrevisedownonimpact. Second,intimeseries,theaggregate pass-through scales with the economy’s average attentiveness. Third, pass-through is state dependent andstrengthenswhenuncertaintyishigh,especiallyamongtheattentive. Fourth, cross-sectionalheterogeneitylinesupwithpayoffdifferences: groupsforwhominformation is more valuable—homeowners, stockholders, prime-age, and higher-income households— displaylargerresponseswhentheyareaccurate.Throughout,identificationexploitsexogenously 15

identified monetary policy shocks and the fact that accuracy is measured before the shock window;wereportrobustnesstoalternativeshockmeasures,controls,andsamples. 4.1 AttentionGatesMonetaryPolicyPass-Through We begin by testingProposition 2.1 inthe micro data: only attentive (accurate) respondents shouldloadoncontractionarymonetarypolicy(MP)newsonimpact. Identificationrestsontwo timingfeatures. First,attentivenessismeasuredatthefirst interviewinmontht andistherefore predeterminedwithrespecttotheFOMCannouncementwindowthatgeneratestheMPsurprise inmontht. Second,theMPshockismeasuredinhighfrequencyaroundtheannouncementand thencumulatedfromt tot+5sothattheinformationsetbetweenthetwointerviews(typicallysix monthsapart)alignswiththesurveyhorizon. Underthistiming,andconditionalonobservables, thesurprisecomponentof MPS isorthogonaltorespondents’pre-shockattentivenessand t demographics, sotheinteractioncoefficientsbelowidentifydifferentialpass-throughrather thanreversecausalityorwithin-monthinformationacquisition. Our baseline specification, adapted from Coibion and Gorodnichenko (2015b), tests this attention-gatingmechanismbyinteractingthepolicyshockwithourattentivenessindicators: ∆πe =α+β(cid:48) (cid:161) MPS ×A (cid:162)+β(cid:48) (cid:161) Z ×A (cid:162)+Γ(cid:48) X +ε , (4.1) i,t+6 M,A t i,t Z,A t i,t i,t i,t where∆πe isthechangeinahousehold’sone-year-aheadinflationexpectationbetweenthe i,t+6 twosurveyinterviews,MPS isthenormalizedcumulativeMPshockfromt tot+5,A isour t i,t three-wayvectorofattentivenessindicators(Accurate/Inaccurate/Haven’theard),Z contains t concurrent macro changes between interviews (IP growth and inflation), and X includes i,t standarddemographiccontrolsincludingageandage2,incomequartiles,education,gender, homeownership,stockholding,maritalstatus,region,andsurvey-modecontrols.9 Coefficients inβ arethegroup-specificpass-throughslopesimpliedbyProposition2.1. M,A Table2reportstheestimates. Theresultslineupcloselywiththegatingprediction. Forthe Accurate group, a1pptighteningintheshadowpolicyratelowersone-year-aheadexpected inflationby−0.359percentagepoints(t =−4.56). FortheInaccurategroup,theslopeissmall andstatisticallyindistinguishablefromzero(0.088,t =0.81). TheHaven’theardgroupshowsa modestnegativeandonlymarginallysignificantcoefficient(−0.155,t =−1.66),aneffectmuch smallerinmagnitudethanthatoftheAccurategroup.10 Quantitatively,theAccurate–Inaccurate 9Ourattentivenessmeasureisrecordedatthefirstinterviewinmontht,priortothenarrowFOMCannouncement windowusedtoformMPS ;accuracyisthereforepredeterminedwithrespecttotheidentifiedsurprise.Cumulating t theshocksoversixmonthsalignstheinformationsetwiththeinterviewhorizonandhelpsensuretheestimated interactionisnotdrivenbywithin-monthlearning. 10Onepossibleinterpretationisthatthisgroup—which,asshowninTable2,isobservationallydistinct—may 16

Table2: AttentionShapesMonetaryPolicyEffectsonInflationExpectations (1) (2) (3) Accurate Inaccurate Haven’tHeard ∗∗∗ ∗ (1)MPS -0.360 0.088 -0.155 t (-4.56) (0.81) (-1.66) (2)∆IP 0.060 ∗∗∗ -0.008 0.013 t (3.60) (-0.49) (0.70) (3)∆π 0.370 ∗∗∗ 0.272 ∗∗∗ 0.325 ∗∗∗ t (9.81) (6.41) (7.21) Controls Yes Observations 37,445 R2 0.0138 Notes:ThistableshowsthebaselineregressionresultsofEquation(4.1).Dependentvariableistherevisionin one-year-aheadinflationexpectationsbetweenthefirstandsecondMSCinterviews(t tot+6). MPS isthe t high-frequencymonetarypolicysurprisecumulatedfromt tot+5andnormalizedsothatoneunitcorresponds toa1ppchangeintheshadowpolicyrateoverthatwindow. ∆IP isthelogchangeinindustrialproduction t and∆π isthechangeininflation.Columnsreportcoefficientsfrominteractionswiththethreeattentiveness t groups(Accurate,Inaccurate,Haven’tHeard)definedatthefirstinterviewinmontht.Allspecificationsinclude individualcontrols(ageandage2,incomequartiles,education,gender,homeownership,stockholding,marital status,region,andsentiment). Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. wedgeislarge: accuraterespondentsrevisedownbyroughlyonethirdofapercentagepointper 1pptightening,whileinaccuraterespondentsdonotreactonimpact. Thispatternisexactly whatProposition2.1implieswhenattentiveagentshavem ∗ >0andinattentiveagentshave i,t m ∗ ≈0. i,t Beyond statistical significance, our estimates imply that attention has an economically meaningfulimpactonthemonetarytransmissionmechanism.Ourbaselinemicro-levelestimate indicatesthatforattentive(“Accurate”)individuals,astandard25-basis-pointcontractionary policysurpriselowersone-year-aheadinflationexpectationsbyapproximately9basispoints. Foragivenpathofthenominalinterestrate,thisrevisiondirectlyamplifiestheintendedpolicy tighteningbyraisingtheperceivedshort-termrealinterestrateforthisgroup. The controls behave sensibly. IP growth between interviews is positively associated with revisionsonlyfortheAccurategroup(consistentwithreal-sidenewsbeingprocessedbyattentive respondents),whilecontemporaneousinflationchangesloadpositivelyandsignificantlyfor all groups, reflecting the salience of price changes in household belief formation. Crucially, theprimaryempiricalsupportforourmechanismcomesfromthesharpcontrastbetweenthe engageinindirectorpassivebeliefupdating.Forexample,theymightreacttohighlysalientsignalslikechangesin gasolinepricesorabsorbbroadeconomicsentimentfrommediaheadlines,eveniftheydonotfollowspecificnews aboutbusinessconditions. 17

Regime.pdf .8 .6 .4 .2 0 1985m1 1988m1 1991m1 1994m1 1997m1 2000m1 2003m1 2006m1 High Attention Regime Share of Accurate Respondents Figure1: AggregateAttentiveness: ShareofAccurateRespondents(1985–2007) Notes:Thisfigurerepresentsthemonthlyaggregateattentiveness(accuracy)ratefromJanuary1985toDecember 2007,definedastheshareofrespondentsatthefirstinterviewinmontht whoseassessmentofrecentbusiness conditionsalignswiththesignofthethree-monthchangeintheunemploymentrate(seeSection3forconstruction). Weusedatathrough2007m6todefinethe“high-attentive”regimeasthetop30%ofthedistributionemployedin thetime-seriesanalysis. “Accurate”and“Inaccurate”groups. AsshowninTable1, thesetwogroupsarenearlyidentical across a wide range of demographic and socioeconomic characteristics. Their divergent responsestomonetarypolicyshocksthereforecannotbeeasilyattributedtoobservableheterogeneity,lendingstrongsupporttoourinterpretationthatpre-shockaccuracy—ourproxyfor attention—isthekeymediatingfactor. Takentogether,thespecification,timing,andmagnitudes supporta“attentiongatespass-through”interpretationatthemicrolevel: contractionaryMP newslowersexpectedinflationprimarilyamongrespondentswhoaccuratelyperceivedrecent businessconditionsbeforethepolicynewsarrived. 4.2 AggregatePass-ThroughScaleswithAttentiveness WenowtestProposition2.2inaggregatetimeseries: theimpactofaconventionalmonetarypolicy(MP)surpriseonrevisionsininflationexpectationsshouldbeproportionaltotheeconomy’s averageattentivenessΛ . Toleveragealongertimeseriesandfocussquarelyonconventionalpolt 18

Table3: AggregatePass-ThroughScaleswithAttentiveness (1) (2) AccuracyRegime High Low ∗∗∗ (1)RRshock -0.620 -0.009 t (-3.17) (-0.09) (2)∆IP 0.183 ∗∗∗ -0.032 t (2.82) (-1.54) (3)∆π 0.333 ∗∗∗ 0.223 ∗∗∗ t (3.01) (4.66) Observations 269 R2 0.394 Notes: ThistableshowstheregressionresultsofEquation(4.2). Dependentvariableisthemedianrevision in1-year-aheadinflationexpectations. RRshock isthethecumulativeRomerandRomer(2004)monetary t policyshocksfromperiodt tot+5. ∆IP isthelogchangeinindustrialproductionand∆π isthechangein t t inflation.Columnsreportregime-specificcoefficientswherehigh-attentivemonthsarethosewiththeaggregate agg attentivenessindexA inthetop30%ofits1985m1–2007m6distribution(Figure1)andlow-attentivemonths t−1 arethecomplement.Newey-weststandarderrorswith6lagsareusedfortheinference;t-statisticsinparentheses. ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. icyactionspriortothezerolowerbound,wefocusontheGreatModeration(1985m1–2007m12) andusethenarrative-basedshocksRomer-Romershockseries,RRshock (RomerandRomer, t agg 2004). WeconstructanaggregateattentivenessindexA asthecross-sectionalshareclassified t Accurateatthefirstinterview(Section3.3). Todefineourpolicyregimes,weclassifymonthsas “high-attentive”iftheaggregateattentionindexfallsinthetop30%ofitshistoricaldistribution. WemeasurethisdistributionusingdataonlythroughJune2007toensuretheclassificationis pre-determinedrelativetoourfullsample(Figure1). Thistop-30%cutoffisastandardapproach forregimeanalysis,andourqualitativefindingsarerobusttousingalternativethresholds,such asthetopquartileortercile. Proposition2.2impliesalarger(morenegative)policyslopein thesemonths: βH =θ(cid:69)[Λ |High]vs.βL =θ(cid:69)[Λ |Low]with|βH|>|βL|forcontractionaryMP t t shocks(θ<0). Ourtime-seriesregressionmirrorsthemicrodesignbutaggregatesthedependentvariable tothemonthlymedianrevision,andsplitsmonthsbyIA =1{A agg intop30%}: t−1 t−1 ∆πe =α+β (cid:161) RRshock ×IA (cid:162)+β(cid:48) (cid:161) Z ×IA (cid:162)+ε , (4.2) t+6 M t t−1 Z,A t t−1 t whereZ containscontemporaneousIPgrowthandinflationchangesbetweenthetwosurvey t interviews. Newey-Weststandarderrors(6lags)accountforserialcorrelationatthesix-month horizon. Table3showsthattheresultsaligntightlywithProposition2.2. Inhigh-attentivemonths,a1 19

ppconventionaltighteningreducesone-year-aheadexpectedinflationbyabout−0.62pp(significant),whereasinlow-attentivemonthstheslopeissmallandstatisticallyindistinguishablefrom zero. Controlsalsobehavesensibly:realactivityandinflationchangesloadpositivelyinthehighattentiveregimeandaremutedotherwise. Thedifferenceinslopesisconsistentwithahigher averageattentivenessΛ t inhigh-attentivemonths: β (cid:98) H ≈θΛ (cid:98) H vs.β (cid:98) L≈θΛ (cid:98) L≈0. Quantitatively, inhigh-attentivemonths,a25-basis-pointtighteningreducesmedianinflationexpectationsbya substantial16basispoints. Thissuggeststhatduringsuchperiods,theexpectationschannelcan amplifytheeffectofapolicysurpriseonex-anterealratesbymorethan60%. Conversely,the absenceofthiseffectinlow-attentiveperiodsdemonstrateshowacrucialchannelofmonetary transmissioncanbecomedormant,highlightingthatthestateofhouseholdattentivenessisa keydeterminantoftheoverallpotencyofmonetarypolicy. Ourresultsimplythatbeliefpass-throughisstate-dependentandscaleswithanindependentlymeasuredattentivenessindex.Thiscomplementsmicroevidenceoninformationfrictions inexpectationsformation(e.g.,CoibionandGorodnichenko,2015a;Gabaix,2020)byproviding acleantime-seriescounterpart: whenmorehouseholdsareattentive,aggregateexpectations respondstronglytopolicynews;whenfewerareattentive,pass-throughisweak. 4.3 StateDependence: UncertaintyRaisesAttentionandAmplifiesExpectationResponses Proposition2.3predictsthatwhenpayoff-relevantuncertaintyU ishigher,optimalattention t ∗ m risesandtheimpactofacontractionaryMPshockonexpectationsbecomesmorenegative, i,t withastrongersensitivityamongalready-attentiveagents.Webringthistothedatabyinteracting MPsurpriseswith(i)ouraccuracy indicatorsand(ii)proxiesforU measuredat t−1: NBER t recessions,theLudvigson,MaandNg(2021)real-uncertaintyindex(LMN),andfinancial-market volatility (VIX). We select these three measures to span canonical business cycle, real, and financialuncertainty,ensuringourfindingsarenotspecifictoonedomain. FortheLMNandVIX indices,ourdefinitionofahigh-uncertaintystateisbasedontheircyclicalcomponenttoisolate deviationsfromtherecenttrendinuncertainty,whichmaybemoresalienttohouseholdsthan theabsolutelevel. TheestimatingequationextendsEquation(4.1)withatripleinteraction, ∆πe i,t+6 =α+β(cid:48) M,A,C (cid:161) MPS t ×A i,t ×State t−1 (cid:162)+β(cid:48) Z,A,C (cid:161) Z t ×A i,t ×State t−1 (cid:162)+Γ(cid:48) X i,t +ε i,t , (4.3) whereState t−1 ∈{Recession,HighLMN,HighVIX};coefficientsonMPS t ×A i,t ×State t−1 recover howthepolicyslopevarieswithuncertaintyfortheattentivegroup,whilethecorresponding “Inaccurate”termsbenchmarktheinattentioncase. 20

Table4: UncertaintyRaisesAttentionandAmplifiesExpectationResponses (1) (2) (3) (4) (5) (6) Accurate Inaccurate Accurate Inaccurate Accurate Inaccurate PanelA:NBER (1)Recession×MPS -1.730 ∗∗∗ -1.125 t (-4.01) (-1.00) (2)Normal×MPS -0.039 0.115 t (-0.49) (1.12) PanelB:LMNRealUncertainty (3)High×MPS -0.539 ∗∗∗ 0.048 t (-5.51) (0.35) (4)Low×MPS -0.269 ∗ 0.250 t (-1.77) (1.33) PanelC:VIX (5)High×MPS -0.456 ∗∗∗ 0.040 t (-4.06) (0.22) (6)Low×MPS -0.007 0.100 t (-0.07) (0.79) Controls Yes Yes Yes Observations 37,445 37,445 37,445 R2 0.0170 0.0146 0.0182 Notes:Thistableshowsregime-andgroup-specificpolicycoefficientsfromthetriple-interactionregressionin Equation(4.3).Thedependentvariableistherevisionin1-year-aheadinflationexpectationsbetweeninterviews, ∆πe .MPS isthenormalizedcumulativemonetarypolicyshockfromt tot+5.A isthethree-wayaccuracy i,t+6 t i,t indicator(Accurate/Inaccurate/Haven’theard)measuredatthefirstinterviewinmontht.State t−1 is(i)the NBERrecessiondummy(PanelA);(ii)HighLMNreal-uncertainty(PanelB)and(iii)HighVIXfinancialvolatility (PanelC),eachdefinedatt−1;“Normal/Low”arethecomplementaryregimes(seeSection4.3forconstruction). WeincludeconcurrentIPgrowthandinflationchangesbetweent andt+6.Weuseindividualinformationabout age,income,homeownership,stockownership,gender,educationlevel,region,maritalstatusandsentimentas controls.Robuststandarderrorsareusedfortheinference;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. The results, reported in Table 4, closely match the theory.11 During recessions, Accurate respondentsrevisedownstronglyonimpact(−1.73ppper1pptightening;significant),while Inaccuraterespondentsdonotrespond.12 InHigh-LMNandHigh-VIXmonths,thesamequalitativepatternholds: Accuratehouseholdsreduceexpectedinflationby≈−0.5pp;Inaccurate 11AllregressioncoefficientsarereportedinAppendixTablesC.1–C.3inAppendixC.1. 12TheestimatedeffectforaccuraterespondentsduringNBER-datedrecessionsiseconomicallyverylarge.This substantialmagnitudemayreflectthenatureofrecessionsasperiodsofheightenedmacro-financialriskand policyscrutiny.Duringsuchcriticalperiods,attentivehouseholdsmaybecomehyper-responsivetoFedactions, perceivingthemascrucialsignalsaboutthefuturestateoftheeconomy.Thispointestimateisconsistentwithour model’scorepredictionthatuncertaintyandriskdramaticallyamplifytheexpectationschannelforthosewhoare payingattention. 21

householdsagainshownosignificantreaction.13 InLow-uncertaintyorNormalstates,policy slopesaresmallandstatisticallyindistinguishablefromzeroorweaklysignificantforallgroups. Thiscross-statecontrastistheempiricalcounterpartof ∂ m ∗ (1−m ∗ ) (cid:161) m ∗ (U )θ(cid:162)=θ i,t i,t <0 (forcontractionaryMPshocks), ∂U i,t t U t t and the Accurate-Inaccurate wedge in high-uncertainty states is exactly the “stronger state dependenceforattentiveagents”inProposition2.3. Inshort,uncertaintyraisesattention,and higherattentionscalestheexpectationsresponsetopolicynews. Thesefindingscomplementexistingstate-dependenceevidenceobtainedfrompricesand quantities. Vavra(2014)showsthattime-varyingvolatilitychangesfirms’adjustmentbehavior and thereby alters aggregate inflation dynamics; our mechanism works on the expectations margin,withuncertaintyinducinggreaterhouseholdattentionandsharperbeliefupdatesto policynews. Relatedly,ourfindingscanbereconciledwithmacrostudiesdocumentingweaker ultimateeffectsofpolicyonrealactivityincertainstates(e.g.,deeprecessionsorhighvolatility). Our evidence points to a stronger initial impact through the expectations channel: in highuncertaintystates, attentivehouseholdsaligntheirinflationexpectationsmoresharplywith policynews. Thisleadstoalargeradjustmentintheirperceivedex-anterealinterestrates. This veryalignment,however,canexplainwhytheultimaterealeffectsonspendingmightbemuted. If expectations adjust swiftly, there is less scope for policy surprises to generate real effects through informational frictions or misperceptions. In this view, a more potent expectations channel could lead to a more muted response in real activity, as well-informed agents have alreadyincorporatedthepolicystanceintotheirdecisions. Twoadditionalpatternsareworthnoting. First,thestatedependenceweuncoverdoesnot requiretimevariationinthevolatilityoftheMPshockitself;increasesinΓ(cid:48)Σ Γ(e.g.,energyor o,t markupvolatility)sufficetoraiseU and,therefore,attention. Second,controlsbehavesensibly t acrossstates: real-sidechanges(IP)loadmoreinhigh-uncertaintystatesfortheAccurategroup, whilecontemporaneousinflationchangesremainsalientacrossgroups. Together,themicro evidencesupportsasimplemessage: theexpectationspass-throughofmonetarypolicyshocks isattentionweightedandthereforestatedependent. 4.4 PayoffHeterogeneityandAccuracy: WhoReactstoPolicyNews? GuidedbyProposition2.4,inthissection,weaskwhethergroupsforwhombeinginformedis morevaluable(higherω)orlesscostly(lowerκ)displaylargermonetarypolicypass-through 13Thiscorefinding—thatamplificationisconcentratedamongtheattentive—alsoholdswhenusingabroad, text-basedmeasureofEconomicPolicyUncertainty,asshowninSection5. 22

whentheyareaccurate. Weproxythesehigherpayoffgroupswithrthreecharecteristics. First,we useassetexposure(stockholdingandhomeownership),aspolicymovesdirectlyaffectportfolio valuesandmortgagefinancing.Second,weexamineage,wheredifferentlife-cyclestagespresent distinctincentives: youngerhouseholds’lifetimeearningsarehighlysensitivetothebusiness cycle, while prime-age households (35-64) typically have the largest balance sheet exposure throughassetsandmortgages. Third,weusehigherincome,whichcorrelateswithbothasset ownershipandinformationuse. Empirically,weextendEquation(4.1)byinteractingMPS withtheaccuracyindicatorsand t each demographic partition, controlling for group means and the full set of covariates. Let D beamutuallyexclusivedemographicpartition(e.g.,Stockholder/Non-stockholder;Homei,t owner/Renter;Young/Middle/Old;Incomequartiles),withonecategoryomittedinestimation. Ourgeneralspecificationreplacesthedemographicblockasneeded: ∆πe =α+ β(cid:48) (cid:161) MPS ×A ×D (cid:162) +β(cid:48) (cid:161) Z ×A ×D (cid:162)+Γ(cid:48) X +ε , (4.4) i,t+6 M,A,D t i,t i,t Z,A,D t i,t i,t i,t i,t (cid:124) (cid:123)(cid:122) (cid:125) group-andaccuracy-specificMPpass-through whereMPS isthenormalizedcumulativeMPsurprisebetweeninterviews,Z collectsconcurt t rentmacrochanges(IPgrowth,inflation)betweenthetwointerviews,andX includesthefull i,t setofdemographicsandsurveycontrols;alllower-ordertermsandfixedeffectsareincluded. Thecoefficientsinβ delivertheimpactslopesbyaccuracy×demographiccell. Forcontrac- M,A,D tionaryshocks,themodelpredictslargenegativeslopesforAccurate×(high-ω/low-κ)groups (e.g.,stockholders,homeowners,prime-age,higher-income)andslopesnearzeroforInaccurate cells. WeestimateEquation(4.4)separatelyforeachpartitionD andTable5reporttheβ i,t M,A,D blocks.14 Stockholding Proposition2.4predictsstrongermonetary-policy(MP)pass-throughamong householdsforwhomthepayofftopayingattentionishigher(largerω ). Stockholdersarea i naturalcandidate: thevalueoftheirportfoliosismoreexposedtomacroandpolicynews,which raisesthemarginalbenefitoftrackingandinterpretingsuchnews. PanelAofTable5estimatesEquation(4.4)withinteractionsbetweenMPshocksand(i)our pre-determinedattentivenessproxyand(ii)stockholdingstatus. Wefindalargeandstatistically significantresponseonlyforaccuratestockholders:a1ppcontractionaryMPsurpriselowerstheir one-year-aheadinflationexpectationsonimpactbyabout−0.41pp(t =−4.57). Incontrast,the coefficientissmallerandstatisticallyindistinguishablefromzeroforaccuratenon-stockholders, andallcoefficientsarenearzerofortheinaccurategroups. Theabsenceofanyresponseamong 14AllregressioncoefficientsarereportedinAppendixTablesC.5–C.6inAppendixC.2. 23

Table5: AttentionandDemographicHeterogeneityinMonetaryPolicyPass-Through (1) (2) (3) (4) (5) (6) Accurate Inaccurate Accurate Inaccurate Accurate Inaccurate PanelA:Stockholding (1)Stock×MPS -0.410 ∗∗∗ 0.150 t (-4.57) (1.20) (2)NonStock×MPS -0.228 -0.047 t (-1.42) (-0.21) PanelB:Homeownership (1)Homeowner×MPS -0.436 ∗∗∗ 0.063 t (-5.08) (0.54) (2)Renter×MPS 0.026 0.214 t (0.13) (0.76) PanelC:Age (1)Young×MPS -0.613 ∗∗∗ 0.260 t (-3.22) (0.82) (2)Middle×MPS -0.350 ∗∗∗ 0.140 t (-3.68) (1.12) (3)Old×MPS -0.234 -0.264 t (-1.22) (-0.98) Interaction Stockownership Homeownership AgeGroup Controls Yes Yes Yes Observations 37,445 37,445 37,445 R2 0.0142 0.0144 0.0150 Notes: Thistablereportgroup-andaccuracy-specificpolicycoefficientsfromtheinteractedspecificationin Equation(4.4).Thedependentvariableistherevisionin1-year-aheadinflationexpectationsbetweeninterviews, ∆πe .MPS isthenormalizedcumulativemonetarypolicysurprisefromt tot+5(mappedtoa1ppchange i,t+6 t intheshadowrate).A isthethree-wayaccuracyindicator(Accurate/Inaccurate/Haven’theard)measuredat i,t thefirstinterviewinmontht.D denotesthedemographicpartitionusedineachpanel:(A)Stockholdervs. i,t Non-stockholder;(B)Homeownervs.Renter;(C)Agegroups(Young18-34,Middle35-64,Old65+).Weinclude concurrentmacrochangesbetweeninterviews(IPgrowthandinflation)aswellasthefullsetofdemographics andsurveycontrols.Alllower-ordertermsandgroupmeansareincluded.ReportedcoefficientsareonMPS × t A ×D .Robuststandarderrorsareusedfortheinference;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, i,t i,t ∗∗∗ p<0.01. inaccurate stockholders, alongside the strong effect for accurate ones, points to attention— rather than simple selection on unobservable traits—as the operative channel. This pattern mapstightlytoProposition2.1(attentiongatespass-through)andProposition2.4(higher-ω typesexhibitstrongerpass-through). AhnandXie(2024)independentlydocumentthatstock-marketparticipationisassociated withgreaterhouseholdattentivenessandmoreaccurateinflationbeliefs. UsingMSCmicrodata, 24

theyshowthatstockholdersaremoreattentiveandholdmoreaccurateinflationbeliefs;they updatemoretomacronewsthannon-holders,andtheattentiongapwidenswhenuncertaintyis high(consistentwitharisk-hedgingmotive).OurfindingiscomplementaryalongtheMPmargin: conditioningonapre-determinedattentivenessproxy,theimpact pass-throughofconventional MPsurprisesisconcentratedamongaccuratestockholders,whereasinaccuratestockholdersdo notreact—exactlytheattention-gatinglogicofProposition2.1. Quantitatively,thisdeliversa larger(morenegative)slopeforstockholderswithintheAccurategroup,anempiricalcounterpart toProposition2.4(higherω). Homeownership Forhomeowners,interest-ratemovementsaredirectlysalientviamortgage payments, refinancing options, and housing wealth, raising the marginal benefit of tracking ∗ policy news and plausibly increasing optimal attention m . This mechanism complements i evidencethathomeownersareespeciallysensitivetoratechangesthroughrefinancing/payment channels(e.g.,Ahnetal.,2024). Estimating Equation (4.4) with interactions between MP surprises, our pre-determined accuracyindicators,andhomeownershipstatussupportsthesepredictions. PanelBofTable5 showsthat,amongaccuraterespondents,a1ppcontractionaryMPsurprisereducesone-yearaheadexpectedinflationbyabout−0.434ppforhomeowners(t =−5.08),whereasrentersexhibit nodetectableimpactresponse;fortheinaccurategroups,coefficientsaresmallandstatistically indistinguishablefromzero. Thissharpcontrastprovidesevidencethattheresultsaredriven bytheproposedattentionchannel, ratherthanbyselectiononunobservablecharacteristics correlatedwithhomeownership. ThepatternmirrorsProposition2.1—attentiondrivespassthrough—andalignswithProposition2.4: conditionalonbeingattentive,thehomeownergroup (ahigh-payoff-to-informationmargin)transmitspolicynewsmorestronglyintoexpectations. In magnitude, the homeowner effect is comparable to the stockholder effect in Panel A, suggestingtwocomplementarymargins—portfolioexposureandmortgage-linkedexposure— throughwhichhigherω amplifiesexpectationresponseswhenattentionispresent. Crucially, i theprerequisiteofaccuracyremainscentral: absentpre-shockattentiveness,neitherhomeownersnorrenterstransmitpolicynewsintoexpectedinflationonimpact. Age Group Age offers another natural partition for the attention sensitivity. Younger and prime-agehouseholdshavegreaterlabor-marketexposureandmorehigh-frequencyeconomic decisions, which plausibly raises ω ; they may also face lower information costs (lower κ ). i i Moreover, the personal-experience framework of Malmendier and Nagel (2016) implies that youngerindividualsplacemoreweightonrecentmacroinformationandthusupdatebeliefs morestrongly,whereasolderindividualsrelymoreonlonger-horizonexperienceandupdate 25

Table6: AttentionandIncomeQuartileinMonetaryPolicyPass-Through (1) (2) Accurate Inaccurate (1)YTL1×MPS 0.048 -0.192 t (0.18) (-0.58) (2)YTL2×MPS -0.669 ∗∗∗ 0.046 t (-3.90) (0.19) (3)YTL3×MPS -0.361 ∗∗∗ 0.201 t (-2.58) (1.04) (4)YTL4×MPS -0.298 ∗∗ 0.132 t (-2.47) (0.77) Interaction IncomeQuartile Controls Yes Observations 37,445 R2 0.0153 Notes:Thistablereportgroup-andaccuracy-specificpolicycoefficientsfromtheinteractedspecificationinEquation(4.4).Thedependent variableistherevisionin1-year-aheadinflationexpectationsbetweeninterviews,∆πe i,t+6 .MPStisthenormalizedcumulativemonetary policysurprisefromttot+5(mappedtoa1ppchangeintheshadowrate).Ai,tisthethree-wayaccuracyindicator(Accurate/Inaccurate /Haven’theard)measuredatthefirstinterviewinmontht.Thedemographicpartitionusedinthistalbeisincomelevel.WeuseYTL4variablefromMSCtodefineconsumers’incomequartile.Weincludeconcurrentmacrochangesbetweeninterviews(IPgrowthandinflation) aswellasthefullsetofdemographicsandsurveycontrols.Alllower-ordertermsandgroupmeansareincluded.Reportedcoefficientsare onMPSt ×Ai,t ×Di,t.Robuststandarderrorsareusedfortheinference;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. lessonimpact. EstimatingEquation(4.4)withinteractionsbetweenmonetarypolicy(MP)surprises,our pre-determinedaccuracyindicators,andage-groupstatus(Young18-34,Middle35-64,Old65+) yieldsacleargradientwithintheaccurategroup(PanelCofTable5).Accurateyoungrespondents reviseone-year-aheadinflationexpectationsthemostaftera1ppcontractionaryMPsurprise (−0.611,t =−3.23),accuratemiddle-agedrespondentsrespondlessbutstillsignificantly(−0.349, t =−3.68),andaccurateolderrespondentsshowasmaller,statisticallyinsignificantcoefficient (−0.234, t =−1.22). Forinaccurate respondents,coefficientsaresmallandindistinguishable fromzeroacrossallagegroups. ThispatternmirrorsProposition2.1: attentiongatespass-through,withvirtuallynoimpact amongtheinaccurate. Conditionalonbeingattentive,themagnitudeordering(Young>Middle > Old) is consistent with higher ω and/or lower κ for younger/prime-age households, and i i withtheexperience-basedupdatingofMalmendierandNagel(2016),wherebyyoungerindividualsplacegreaterweightonrecentpolicy-relevantinformation. Insum,theagegradientin impactresponsesprovidesanadditionalcross-sectionalvalidationofthemodel’spayoff-based heterogeneity. 26

IncomeQuartile Lastly,incomeoffersanothernaturalpartition:relativetothebottomquartile, middle-andhigher-incomehouseholdstypicallyhavemorepolicy-exposedstakes(labor-market risk,assetportfolios,mortgage/creditmargins),whichraisesω and,inturn,theattention-scaled i response|θm ∗|. i UsingtheMSCincomequartiles,weestimateEquation(4.4)withthedemographicpartition D ={YTL1,...,YTL4}andreportresultsinTable6.15 Theaccuracyprerequisiteremainsfirsti,t order: acrossallquartiles,inaccuraterespondentsdonotreactonimpact. Withintheaccurate group,wefindacleargradient: middle-incomehouseholds(YTL2,YTL3)displaythelargestand mostpreciselyestimateddeclinesin1-year-aheadexpectationsaftera1ppcontractionaryMP surprise(-0.667and-0.360,respectively),high-incomehouseholds(YTL4)reactmoderately(- 0.297),andthelowest-incomequartile(YTL1)showsnodetectableimpactresponse.Thispattern isconsistentwithourpayoff-basedmechanism(higherω outsidethebottomquartile)andwith i theideathatgroupswhoseexpenditurebasketsloadmoreonenergyandotherpolicy-sensitive categoriesanticipatelargernear-termdisinflationfollowingatightening.16 5 Robustness We assess the robustness of our findings along five dimensions and report full details and tablesinAppendixSectionC.First,weaddressconcernsthathigh-frequency(HF)monetary policy surprises may bundle a Fed information-effect component. We therefore re-estimate ourbaselinespecificationsusingtheBuetal.(2021)“BRW”shocks(PanelAofTable7). The signs,cross-groupordering,andsignificancemirrortheHFresults;magnitudesaresomewhat larger under BRW, which is consistent with differences in the mapping from shocks to rates (BRWinnovationsmove2-yearyieldsnearlyone-for-one, whereasHFfactorsneednot). We alsoverifiedrobustnesstothereassessedHFseriesinBauerandSwanson(2023);becausethat sampleendsin2019:M7,wedonottabulateit,butthecorepatternspersist(PanelBofTable7). Second,wevarytheconstructionofAccuracy. Ourbenchmarkmeasureusesrecentchanges in the unemployment rate; to check that results do not hinge on this choice, we reclassify Accuracy usingtwoalternativeaggregatesignalsthatproxytherealandfinancialsidesofthe macroenvironment: IndustrialProduction(IP)andtheNationalFinancialConditionsIndex (NFCI).ThebaselinegatingandheterogeneitypatternsareunchangedwhenweuseIP(Appendix 15AllregressioncoefficientsarereportedinAppendixTableC.7inAppendixC.2. 16Jaravel(2019)andMangianteandLauper(Forthcoming)investigatethelinkbetweenmonetarypolicyshocks andinflationinequality.Theyfindthattheinflationratesfacedbyhouseholdsresponddifferentlytopolicy,arguing thatmiddle-incomegroupsaremostaffectedbycontractionaryshocks.Thisphenomenonisprimarilydrivenby heterogeneousconsumptionbundles;sectorslikegasolineandenergyaremoreresponsivetopolicy,andthese goodsmakeupalargershareoftheconsumptionbasketforlow-andmiddle-incomehouseholds. 27

Table7: AlternativeMonetaryShockMeasure (1) (2) (3) (4) (5) (6) Accurate Inaccurate Haven’t Accurate Inaccurate Haven’t Heard Heard PanelA:Buetal.(2021) ∗∗∗ ∗∗ (1)MPS -1.411 -0.256 -0.505 t (-6.66) (-1.19) (-2.14) (2)∆IP 0.043 ∗∗∗ -0.001 0.004 t (2.77) (-0.10) (0.24) (3)∆π 0.349 ∗∗∗ 0.268 ∗∗∗ 0.319 ∗∗∗ t (9.29) (6.33) (7.04) PanelB:BauerandSwanson(2023) ∗∗∗ ∗∗∗ (1)MPS -1.343 -0.535 -0.898 t (-4.46) (-1.57) (-2.79) (2)∆IP 0.079 ∗∗∗ 0.056 ∗∗ 0.056 ∗∗∗ t (3.97) (1.97) (2.16) (3)∆π 0.275 ∗∗∗ 0.268 ∗∗∗ 0.275 ∗∗∗ t (7.45) (6.32) (6.11) Controls Yes Yes Observations 37,445 35,592 R2 0.0148 0.0168 Notes:Thistablereplacesthehigh-frequencyMPS serieswiththeBuetal.(2021)monetarypolicyshocks(Panel t A)andBauerandSwanson(2023)(PanelB)andre-estimatesthebaselinemicrospecificationEquation(4.1). Thedependentvariableistherevisioninone-year-aheadinflationexpectations.Shocksarecumulatedfrom t tot+5toalignwiththesix-monthsurveyhorizon. Accuracyismeasuredatthefirstinterview. Weinclude contemporaneousIPgrowthandinflationchangesbetweeninterviews;demographicsandsurveycontrolsare included.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. TablesinAppendixC.3)orNFCI(AppendixTablesinAppendixC.4)insteadofunemploymentto defineAccuracy.17 Third,weevaluaterepresentativenessbyreweightingthemicroregressionswithhouseholdheadweights(wt). BecausetherecontactedMSCpanelinagivenmonthcontainsatmostabout 250respondents,weightingisanaturalcorrection. Weightedregressions(AppendixTableC.18) yieldcoefficientsthatarestatisticallyandeconomicallyindistinguishablefromourbaseline, suggestingthatsmall-samplecompositiondoesnotdriveourresults. Fourth, we augment the macro controls to account for the salience of gasoline prices in householdbeliefformation. WeaddthelogchangeinU.S.RegularAllFormulationsGasPrice between the two interviews (from FRED) to the baseline controls (replacing crude oil prices used elsewhere). The gating and heterogeneity results are robust to this addition (Appendix 17WealsoreplaceIPwithitsyear-over-yeargrowthratetoremovetrend;resultsareessentiallyidentical. 28

TableC.19),indicatingthatourfindingsarenotanartifactofomittedgasoline-pricemovements. Finally,werevisitthestate-dependenceanalysisusinganalternativeuncertaintyproxy. We constructthevolatilitystatefromtheEconomicPolicyUncertainty(EPU)index(Baker,Bloom andDavis,2016a)—atext-basedmeasurethatcapturespolicy-relevantuncertaintyspanning bothrealandfinancialsources. Defininghigh-uncertaintymonthsbythecyclicalcomponentof EPUandre-estimatingthetriple-interactiondesignreproducesourbaselinepattern: Accurate respondentsloadmorestronglyoncontractionarypolicynewsinhigh-uncertaintystates,while Inaccuraterespondentsdonot(AppendixTableC.20). Acrossallchecks—alternativeAccuracydefinitions(IP,NFCI),alternativeshockmeasures (BRW,reassessedHF),populationweighting,richerpricecontrols,andalternativeuncertainty splits(EPU)—thecoreresultsremain: attention(Accuracy)mediatespass-throughonimpact, aggregatepass-throughscaleswithattentiveness,statedependenceisstrongerfortheattentive, andhigh-payoffgroups(stock-holders,homeowners,prime-age,higher-income)displaylarger effectswhenaccurate. 6 Conclusion Wedevelopaminimalbehavioralframeworkinwhichhouseholdsoptimallychooseattention to inflation-relevant news and derive four predictions: attention drives the pass-through of monetarypolicytoinflationexpectations;aggregatepass-throughscaleswiththeeconomy’s averageattentiveness;pass-throughisstatedependentandriseswithpayoff-relevantuncertainty; and,conditionalonbeingattentive,groupswithahigherpayofffrombeinginformeddisplay strongereffects. Usingpre-determinedAccuracy,high-frequencyidentifiedMPsurprises,and both micro and aggregate designs, the data align closely with these predictions. On impact, attentivehouseholdsrevisedownexpectedinflationaftercontractionaryshocks,theaggregate responseislargerinhigh-attentivemonths,statedependenceisconcentratedamongtheattentive,andstockholders,homeowners,prime-age,andhigher-incomehouseholdsreactmore whenaccurate. Thesefindingshaveclearpolicyandmacroimplications. Attentionactsasanexpectations multiplier: when attention is low, policy news barely reaches household beliefs; when high, thesamenewsmovesexpectationsstrongly. Thisprovidesamicrofoundationforwhybroadbased communications can have limited effects, as a large share of the audience may be in alow-attentionstate. Ourresultssuggestthattheexpectationschannelismostpotentwhen communications are timed to coincide with periods of high uncertainty or targeted toward high-payoffgroups—likehomeownersandstockholders—whoareendogenouslymoreattentive. Theeffectivenessoftoolslikeforwardguidanceisthereforenotconstantbutislikelyamplified 29

duringturbulenteconomictimes. Thisuneventransmission,whileusefulforfast-actingpolicy, meanscentralbanksmayconfrontdistributionalasymmetriesinhowexpectationsareupdated. Fromamacrolens,strongerbeliefpass-throughamplifiestheshort-runreal-rateeffectofagiven nominaltightening,potentiallymakingconventionalMPmorepowerfulindisinflatingwhile sharpeningnear-termtrade-offs. Ouranalysisfocusesonimpactrevisionsandleaveslonger-horizondynamicsandgeneralequilibrium propagation to future work. Natural next steps include causal manipulation of attention(e.g.,informationtreatments),linkingbeliefupdatestospending/refinancing/portfolio behavior, and integrating household and firm attention in a structural model, and studying optimalcommunicationunderattentionconstraints.Whileourworkfocusesonmonetarypolicy, themodelimpliesthatattentiongatesresponsestoanyinflation-relevantnews;investigating this mechanism for other disturbances, like fiscal or energy shocks, is a fruitful avenue for futureresearch. Acompanionagendaistoconnecttimevariationinattentioninequalitytothe changingeffectivenessofpolicyoverthebusinesscycle. References Afrouzi,Hassan,“StrategicInattention,InflationDynamics,andtheNonneutralityofMoney,” JournalofPoliticalEconomy,2024,132(10),3378–3420. andChoongryulYang,“DynamicrationalinattentionandthePhillipscurve,”WorkingPaper 8840,CESifo2021. Ahn,HieJooandShihanXie,“Stock-drivenHouseholdAttention,”2024. , ,andChoongryulYang,“Effectsofmonetarypolicyonhouseholdexpectations: Therole ofhomeownership,”JournalofMonetaryEconomics,2024,147,103599. Alpanda,Sami,EleonoraGranziera,andSarahZubairy,“Statedependenceofmonetarypolicy acrossbusiness,creditandinterestratecycles,”EuropeanEconomicReview,2021,140,103936. Angeletos,George-MariosandChenLian,“ForwardGuidancewithoutCommonKnowledge,” AmericanEconomicReview,2018,108(9),2477–2512. Bachmann, Rudiger, Tim O Berg, and Eric R Sims, “Inflation expectations and readiness to spend: Cross-sectionalevidence,”AmericanEconomicJournal: EconomicPolicy,2015,7(1), 1–35. Baker,ScottR.,NicholasBloom,andStevenJ.Davis,“MeasuringEconomicPolicyUncertainty,” QuarterlyJournalofEconomics,2016,131(4),1593–1636. Baker,ScottR,NicholasBloom,andStevenJDavis,“Measuringeconomicpolicyuncertainty,” Thequarterlyjournalofeconomics,2016,131(4),1593–1636. 30

Bauer,MichaelD,AeimitLakdawala,andPhilippeMueller, “Market-basedmonetary policy uncertainty,”TheEconomicJournal,2022,132(644),1290–1308. andEricTSwanson,“Areassessmentofmonetarypolicysurprisesandhigh-frequencyidentification,”NBERMacroeconomicsAnnual,2023,37(1),87–155. ,CarolinEPflueger,andAdiSunderam,“PerceptionsAboutMonetaryPolicy,”TheQuarterly JournalofEconomics,062024,139(4),2227–2278. Bordalo, Pedro, Nicola Gennaioli, and Andrei Shleifer, “Diagnostic Expectations and Credit Cycles,”JournalofFinance,2018,73(1),199–227. Bracha, Anat and Jenny Tang, “Inflation Levels and (In)Attention,” The Review of Economic Studies,062024,92(3),1564–1594. Braitsch, Hana and James Mitchell, “A New Measure of Consumersￜￜ (In)Attention to Inflation,”EconomicCommentary,October2022,2022(14),1–7. Bu,Chunya,JohnRogers,andWenbinWu,“AunifiedmeasureofFedmonetarypolicyshocks,” JournalofMonetaryEconomics,2021,118,331–349. Candia, Bernardo, Olivier Coibion, and Yuriy Gorodnichenko, “The Inflation Expectations of U.S. Firms: Evidence from a New Survey,” Journal of Monetary Economics, 2024, 145. Supplementissue. Carvalho,CarlosandFernandaNechio,“Dopeopleunderstandmonetarypolicy?,”Journalof MonetaryEconomics,2014,66,108–123. Cavallo,Alberto,GuillermoCruces,andRicardoPerez-Truglia,“Inflationexpectations,learning,andsupermarketprices: Evidencefromsurveyexperiments,”AmericanEconomicJournal: Macroeconomics,2017,9(3),1–35. Claus,EddaandVietHoangNguyen,“Monetarypolicyshocksfromtheconsumerperspective,” JournalofMonetaryEconomics,2020,114,159–173. Coibion,OlivierandYuriyGorodnichenko,“InformationRigidityandtheExpectationsFormationProcess: ASimpleFrameworkandNewFacts,”AmericanEconomicReview,2015,105(8), 2644–2678. and ,“IsthePhillipscurvealiveandwellafterall? Inflationexpectationsandthemissing disinflation,”AmericanEconomicJournal: Macroeconomics,2015,7(1),197–232. , ,andMichaelWeber,“Monetarypolicycommunicationsandtheireffectsonhousehold inflationexpectations,”JournalofPoliticalEconomy,2022,130(6),1537–1584. DAcunto,Francesco,UlrikeMalmendier,andMichaelWeber,“Whatdothedatatellusabout inflationexpectations?,”in“Handbookofeconomicexpectations,”Elsevier,2023,pp.133–161. , ,JuanOspina,andMichaelWeber,“Exposuretogrocerypricesandinflationexpectations,” JournalofPoliticalEconomy,2021,129(5),1615–1639. 31

Dietrich,AlexanderM.,“Consumptioncategories,householdattention,andinflationexpectations: Implicationsforoptimalmonetarypolicy,”JournalofMonetaryEconomics,2024,147, 103594. Flynn,JoelPandKarthikSastry,“AttentionCycles,”WorkingPaper32553,NationalBureauof EconomicResearchJune2024. Gabaix,Xavier,“ABehavioralNewKeynesianModel,”AmericanEconomicReview,2020,110(8), 2271–2327. Jaravel,Xavier,“Theunequalgainsfromproductinnovations:Evidencefromtheusretailsector,” TheQuarterlyJournalofEconomics,2019,134(2),715–783. Korenok,Oleg,DavidMunro,andJiayiChen,“InflationandAttentionThresholds,”TheReview ofEconomicsandStatistics,112023,pp.1–28. Kroner, Niklas, “How Markets Process Macro News: The Importance of Investor Attention,” FinanceandEconomicsDiscussionSeries2025-022,BoardofGovernorsoftheFederalReserve SystemMar2025. Kryvtsov,OleksiyandLubaPetersen,“Centralbankcommunicationthatworks: Lessonsfrom labexperiments,”JournalofMonetaryEconomics,2021,117,760–780. Kumar,Saten,OlivierCoibion,HassanAfrouzi,andYuriyGorodnichenko,“InflationTargeting DoesNotAnchorInflationExpectations: EvidencefromFirmsinNewZealand,”Brookings PapersonEconomicActivity,2015,pp.151–208. Lamla,MichaelJandDmitriVVinogradov,“Centralbankannouncements: Bignewsforlittle people?,”JournalofMonetaryEconomics,2019,108,21–38. Ludvigson, Sydney C, Sai Ma, and Serena Ng, “Uncertainty and business cycles: exogenous impulseorendogenousresponse?,”AmericanEconomicJournal: Macroeconomics,2021,13 (4),369–410. Mac´kowiak,BartoszandMirkoWiederholt,“OptimalStickyPricesunderRationalInattention,” AmericanEconomicReview,2009,99(3),769–803. Malmendier, Ulrike and Stefan Nagel, “Learning from inflation experiences,” The Quarterly JournalofEconomics,2016,131(1),53–87. Mangiante,GiacomoandChristophLauper,“Monetarypolicyshocksandinflationinequality,” InternationalJournalofCentralBanking,Forthcoming. Mankiw,N.GregoryandRicardoReis,“StickyInformationversusStickyPrices: AProposalto Replace the New Keynesian Phillips Curve,” Quarterly Journal of Economics, 2002, 117 (4), 1295–1328. 32

Marcellino, MassimilianoandDaliborStevanovic, “The demand and supply of information aboutinflation,”WorkingPapers22-06,UniversityofQuebecinMontreal’sSchoolofManagementAug2022. Nakamura,EmiandJonSteinsson,“High-frequencyidentificationofmonetarynon-neutrality: theinformationeffect,”TheQuarterlyJournalofEconomics,2018,133(3),1283–1330. Pfäuti,Oliver,“TheInflationAttentionThresholdandInflationSurgers,”WorkingPapers2024- 02,EMPCTWorkingPaperSeries2024. Romer,ChristinaDandDavidHRomer,“Anewmeasureofmonetaryshocks: Derivationand implications,”Americaneconomicreview,2004,94(4),1055–1084. Sims,ChristopherA.,“ImplicationsofRationalInattention,”JournalofMonetaryEconomics, 2003,50(3),665–690. Song, Wenting and Samuel Stern, “Firm Inattention and the Efficacy of Monetary Policy: A Text-BasedApproach,”TheReviewofEconomicStudies,102024,p.rdae102. Tenreyro, Silvana and Gregory Thwaites, “Pushing on a String: US Monetary Policy Is Less PowerfulinRecessions,”AmericanEconomicJournal: Macroeconomics,2016,8(4),43–74. Vavra,Joseph,“InflationDynamicsandTime-VaryingVolatility: NewEvidenceandanSsInterpretation,”TheQuarterlyJournalofEconomics,None2014,129(1),215–258. Weber,Michael,BernardoCandia,HassanAfrouzi,TizianoRopele,RodrigoLluberas,Serafin Frache, Brent Meyer, Saten Kumar, Yuriy Gorodnichenko, Dimitris Georgarakos, Olivier Coibion,GeoffKenny,andJorgePonce,“TellMeSomethingIDon’tAlreadyKnow: Learning inLow-andHigh-InflationSettings,”Econometrica,2025,93(1),229–264. Wu,Edson,“RationallyInattentiveSeller: TheGeographicDispersionofU.S.RetailChainsand itsAggregateImplications,”TechnicalReport,KiltsCenteratChicagoBoothMarketingData Center2024. Yang,Choongryul,“Rationalinattention,menucosts,andmulti-productfirms: Microevidence andaggregateimplications,”JournalofMonetaryEconomics,2022,128,105–123. 33

APPENDIX A Proofs This appendix provides the formal mathematical derivations for the four main propositions presented in the theoretical framework of Section 2. It details the steps for deriving the impactofmonetarypolicyonindividualexpectations(Proposition2.1),thescalingofaggregate pass-throughwithattention(Proposition2.2),thestate-dependentnatureoftheresponseto uncertainty(Proposition2.3),andthecross-sectionalpredictionsbasedonpayoffheterogeneity (Proposition2.4). A.1 ProofofProposition2.1. ∗ CombineEquation(2.1)–Equation(2.2)evaluatedattheoptimumm (U ): i,t t E i B ,t π t+1 =(1−m i ∗ ,t )π¯+m i ∗ ,t (cid:163)π¯+ρ(π t −π¯)+θεm t p+Γ(cid:48)εo t (cid:164) . Holding π t fixed at impact and differentiating w.r.t. εm t p yields ∂E i B ,t π t+1 /∂εm t p =m i ∗ ,t θ. The impactchangeis∆π i,t+1 =m i ∗ ,t θεm t p . A.2 ProofofProposition2.2. Startfromtheindividualimpactchange, (cid:179) (cid:180) ∆π i,t+1 =(cid:69)B i,t [π t+1 ]−π i,t =m i ∗ ,t (U t ) θεm t p+Γ(cid:48)εo t , whichfollowsbysubstitutingEquation(2.1)intoEquation(2.2)andevaluatingatimpact(holding π fixed). Lettheaggregaterevisionbethecross-sectionalaverage: t ∆πe t+1 ≡(cid:69) i [∆π i,t+1 ]=(cid:69) i [m i ∗ ,t (U t )]θεm t p+(cid:69) i [m i ∗ ,t (U t )]Γ(cid:48)εo t . (cid:124) (cid:123)(cid:122) (cid:125) (cid:124) (cid:123)(cid:122) (cid:125) Λ υ t t ByconstructionΛ ∈[0,1]. Underthebaselineorthogonalitywithintheidentificationwindow, t Cov (εmp ,εo)=0,andsincem ∗ (U )ispredeterminedatthetimetheshocksarerealized,we t t t i,t t have (cid:69)(cid:163)εmpυ (cid:164)=0, so the regression coefficient of ∆πe on εmp equals θΛ , yielding Equat t t+1 t t tion(2.6). 34

A.3 ProofofProposition2.3. Let S (U) ≡ ∂[m (U)θ]/∂U = θ· mg(U)[1−mg(U)] for group g. For any U > 0, if m (1−m ) > g g U A A m (1−m ), then |S (U)|>|S (U)| because |θ| andU cancel in the comparison. A sufficient I I A I conditionism ∈(1,1)andm ∈(0,1)since f(m)=m(1−m)isstrictlyincreasingon[0,1]and A 2 I 2 2 strictlydecreasingon[1,1]withmaximumatm= 1. 2 2 A.4 ProofofProposition2.4. FixU >0. FromEquation(2.4), t ω U m ∗ (U )= i t . i,t t ω U +κ i t i (i) Attention ordering. A direct calculation gives ∂m ∗ /∂ω = Ut κ i > 0 and ∂m ∗ /∂κ = − ω iUt <0,som ∗ isstrictlyincreasinginω ands i t , r t ictly i dec ( r ω e i a U s t i + n κ g i) i 2 nκ . i,t i (ω iUt +κ i)2 i,t i i (ii)Pass-throughordering. TheindividualMPpass-throughmagnitudeis (cid:175) (cid:175) ∂∆π i,t+1 /∂εm t p(cid:175) (cid:175) = |θ|m ∗ (U )byProposition2.1. Monotonicitythenfollowsfrompart(i). i,t t (iii)Selectioninto“attentive/accurate”. Foranythresholdτ∈(0,1), A =1{m ∗ ≥τ}isnondei,t i,t creasinginω andnonincreasinginκ becausem ∗ ismonotoneinthoseparameters. i i i,t (iv)Conditionalorderingwithintheattentivegroup. On{A =1}wehavem ∗ ≥τ. Since i,t i,t m ∗ is increasing in ω and decreasing in κ pointwise, any upward (first-order) shift in ω or i,t i i downwardshiftinκraisesm ∗ foreveryindividual,andthusraises(cid:69)[m ∗ |A =1]whenever i,t i,t i,t thesupportaboveτhaspositivemeasure. 35

B Model Extension Thisappendixshowsthatourfourtestableimplicationsdonotrelyonthebaselinechoiceofa linearattentionweightorquadraticattentioncosts. Thefirstsubsectionestablishesageneral comparative-staticsresult(LemmaB.1)foranarbitraryincreasingattentionmappingφ(·)and strictlyconvexcostψ (·):theoptimalattentionm ∗ isunique,increaseswithpayoff-relevantnews i i varianceU andstakesω ,anddecreaseswithcostsκ .Thelinear/quadraticspecificationfollows t i i asacorollary. Thesecondsubsectionintroducesacommonnoisypublicsignalobservedbefore attention is chosen and shows that it synchronizes attention choices—micro-founding time variationintheaggregateattentivenessindex—whileleavingtheindividualgating,aggregate scaling,uncertaintyamplification,andpayoff-heterogeneitypredictionsunchanged. B.1 Generalattentionmappingandconvexcosts Weshowthatthemaincomparativestaticsdonotrelyonalinearattentionweightorquadratic costs. AssumptionB.1(Informationandcosts).TheexpectationsoperatorisEB =π¯+φ(m )(E ∗−π¯) i i with φ:[0,1]→[0,1], φ(cid:48) (m)>0, and φ(cid:48)(cid:48) (m)≤0. The attention cost is ψ (m), where ψ isC1, i i strictlyconvexon[0,1]withψ(cid:48) (0)=0andψ(cid:48)(cid:48) (m)>0. Benefitsarescaledbyω >0,costsbyκ >0 i i i i (possiblyviaψ (m)=κ ψ˜(m)withψ˜ (cid:48) (m)>0).LetU ≡Var(Γ(cid:48)ε +θεmp )denotepayoff-relevant i i t o,t t newsvariance. Withmean-squaredforecastloss,theper-periodobjectivecanbewritten(uptoapositive multiplicativeconstant)as 1 L (m;U ,ω ,κ ) = ω U [1−φ(m)]2 + ψ (m), i t i i i t i 2 sotheuniqueoptimumm ∗∈(0,1)solvesthefirst-ordercondition i ω U (cid:161) 1−φ(m ∗ ) (cid:162)φ(cid:48) (m ∗ ) = ψ(cid:48) (m ∗ ). (B.1) i t i i i i PropositionB.1(Comparativestaticsundergeneralφandψ).UnderAssumptionB.1,thereisa uniqueminimizerm ∗ (U ,ω ,κ )∈(0,1)satisfyingEquation(B.1). Moreover, i t i i ∂m ∗ ∂m ∗ ∂m ∗ i >0, i >0, andif ψ (m)=κ ψ˜(m) with ψ˜ (cid:48) (m)>0, then i <0. ∂U ∂ω i i ∂κ t i i Proof. StrictconvexityofL impliesauniqueinteriorsolution. DefineF(m;U,ω,κ)≡ωU(1− i 36

φ(m))φ(cid:48) (m)−ψ(cid:48) (m). ThenF =ωU{−(φ(cid:48) (m))2+(1−φ(m))φ(cid:48)(cid:48) (m)}−ψ(cid:48)(cid:48) (m)<0becauseφ(cid:48)>0, m φ(cid:48)(cid:48)≤0,andψ(cid:48)(cid:48)>0. Bytheimplicitfunctiontheorem, ∂m ∗ =− F U = ω(1−φ)φ(cid:48) >0, ∂m ∗ =− Fω = U(1−φ)φ(cid:48) >0. ∂U F −F ∂ω F −F m m m m Ifψ(m)=κψ˜(m)withψ˜ (cid:48) (m)>0,thenFκ =−ψ˜ (cid:48) (m)<0,so∂m ∗ /∂κ=−Fκ/F m <0. CorollaryB.1(Linearweight/quadraticcost).Ifφ(m)=mandψ(m)= 1κm2,thenEquation(B.1) 2 reducestoωU(1−m ∗ )=κm ∗ ,hencetheclosedform ωU m ∗ (U,ω,κ) = , φ(m ∗ )=m ∗ . ωU+κ Allfourtestableimplicationsinthemaintextfollowimmediately. Implications. Replacingm ∗ byφ(m ∗ )intheimpactcoefficientdeliversthesamefourpredici i tions: (i)individualgating (onlyattentivetypesloadonpolicynews),(ii)aggregatescalingby E[φ(m ∗ )],(iii)amplificationwhenU ishigher,and(iv)largerpass-throughforhigh-ω /low-κ i t i i groups. B.2 PublicSignalExtension This appendix shows that introducing a common, noisy public signal s that arrives before t attentionchoicesdoesnotalterthefourtestableimplicationsinthemaintext: (i)individual gating;(ii)aggregatescaling;(iii)uncertaintyamplification;and(iv)payoffheterogeneity. The publicsignalprovidesasimplemicro-foundationfortime-variationinaggregateattentiveness bysynchronizingattentionchoicesacrosshouseholds. Supposethepublicsignalisaboutthefullyinformedforecastofnext-periodinflationπ∗ t+1 (e.g.,ahighlypublicizeddatareleaseorheadline),observedbeforeattentionchoice. Lets be t informativeaboutπ∗ sothattheposteriorvariance t+1 U post≡Var(π∗ |s ) t t+1 t is(weakly)smallerthanthepriorvarianceU and(weakly)decreasinginthesignal’sprecision. t post Underquadraticforecastloss,therelevantlosscomponentscaleswithU . t AssumptionB.2.ThepublicsignalabouttheinflationlevelyieldsaposteriorvarianceU post= t H(U t ,τ s ) with H U > 0 and Hτ s < 0, where τ s is the signal precision. (For Gaussian-normal conjugacy,U post=(U −1+τ ) −1,whichdoesnotdependontherealizationofs .) t t s t 37

Givens (andτ ),thehouseholdchoosesattentiontominimize t s L (m;U post ,ω ,κ )=ω U post G(m)+κ C(m), i t i i i t i withG andC asabove. LemmaB.1(Optimalattentionwithapubliclevelsignal).UnderAssumptionB.2andthepropertiesofG andC statedabove,theuniqueoptimalattentionm ∗=m ∗ (U post ,ω ,κ )is(weakly) i i t i i increasinginU post andinω ,and(weakly)decreasinginκ andinthesignalprecisionτ (via t i i s post U ). t post PropositionB.2(Robustnessofimplications: levelsignal).ReplacingU byU leavesallfour t t implicationsintact: 1. Individual gating: the impact coefficient remains proportional to φ(m ∗ )θ with m ∗ = i i m ∗ (U post ,ω ,κ ). i t i i 2. Aggregatescaling: βagg=θE[φ(m ∗ )]scaleswithaverageattention;amoreprecisepublic t i post signalreducesU andthuslowersaverageattention,butdoesnotalterthegatinglogic. t post 3. Uncertaintyamplification:whenresidualuncertaintyU ishigher(e.g.,thepublicsignal t isimpreciseorabsent),optimalattentionishigherandpass-throughisstronger. 4. Payoffheterogeneity: foranyU post ,higherω /lowerκ typeschoosemoreattentionand t i i exhibitlargerpass-through. Discussion. Alevelsignalreducesresidualuncertaintyandtherebylowersthemarginalvalue ofcostlyattention,butconditionalonthechosenattention,thepass-throughofmonetarypolicy newsisstillmultipliedbytheattentionweight. Sinces iscommon,itsleveleffectonbeliefsis t absorbedbytimevariation(e.g.,monthfixedeffects)inourempiricaldesigns;theestimated slopewithrespecttopolicysurprisesisthereforeunaffected. 38

C Robustness This appendix contains the complete regression tables that support the robustness analysis discussedinSection5. Itincludesdetailedoutputfromthestate-dependenceanddemographic heterogeneityanalyses,aswellasacomprehensivesetofchecksusingalternativedefinitionsfor theaccuracyproxy(basedonIndustrialProductionandtheNFCI),alternativemonetarypolicy shockmeasures,populationweighting,andadditionalcontrols. C.1 FullReports: StateDependentAnalysis This section provides the complete regression output for the state-dependence analysis presentedinSection4.3.Thetablesreportthefullsetofcoefficients,includingthoseforthe“Haven’t Heard”groupandcontemporaneousmacrocontrols,forspecificationsusingNBERrecessions, theLMNrealuncertaintyindex,andtheVIXtodefinehigh-andlow-uncertaintystates. AppendixTableC.1: AttentionwithNBERBusinessCycleIndicator NBERRecession (1)Accurate (2)Inaccurate (3)Haven’tHeard PanelA:NBERRecession (1)Recession×MPS -1.701 ∗∗∗ -1.125 -0.988 t (-4.01) (-1.00) (-1.29) (2)Recession×∆IP 0.200 ∗∗∗ 0.084 0.120 ∗ t (5.19) (0.87) (1.68) (3)Recession×∆π 0.303 ∗∗∗ 0.258 0.243 t (3.29) (1.32) (1.52) PanelB:Normal (4)Normal×MPS -0.039 0.115 -0.123 t (-0.49) (1.12) (-1.43) (5)Normal×∆IP -0.028 -0.018 -0.021 t (-1.41) (-1.00) (-1.09) (6)Normal×∆π 0.332 ∗∗∗ 0.269 ∗∗∗ 0.275 ∗∗∗ t (8.17) (6.17) (5.86) Interaction NBER Controls Yes Observations 37,445 R2 0.0170 Notes: Thistablereportsthestate-dependentregressioninEquation(4.3)usingtheNBERrecessionindicatorasStatet−1. Thedependentvariableistherevisioninone-year-aheadinflationexpectationsbetweeninterviews.MPStdenotesthenormalizedcumulativehighfrequencymonetarypolicyshocksfrommontht tot+5. Ai,t isthethree-wayaccuracyvector(Accurate/Inaccurate/Haven’theard) measuredatthefirstinterview.WeincludecontemporaneousIndustrialProductiongrowthandinflationchangesbetweeninterviews;all lower-ordertermsandthefullsetofdemographicsandsurveycontrolsareincluded. Robuststandarderrorsarereported;t-statisticsin parentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 39

AppendixTableC.2: AttentionwithRealUncertaintyIndicator LMNRealUncertainty (1)Accurate (2)Inaccurate (3)Haven’tHeard PanelA:HighUncertainty (1)High×MPS -0.539 ∗∗∗ 0.048 -0.137 t (-5.51) (0.35) (-1.18) (2)High×∆IP 0.130 ∗∗∗ -0.003 0.022 t (4.37) (-0.09) (0.66) (3)High×∆π 0.304 ∗∗∗ 0.137 ∗ 0.324 ∗∗∗ t (5.29) (1.95) (4.38) PanelB:LowUncertainty (4)Low×MPS -0.269 ∗ 0.250 -0.248 t (-1.77) (1.33) (-1.49) (5)Low×∆IP 0.034 -0.012 0.008 t (1.62) (-0.63) (0.37) (6)Low×∆π 0.381 ∗∗∗ 0.397 ∗∗∗ 0.310 ∗∗∗ t (7.54) (7.73) (5.76) Interaction LMNrealuncertainty Controls Yes Observations 37,445 R2 0.0146 Notes:Thistablereportsthestate-dependentregressioninEquation(4.3)usingtheLudvigsonetal.(2021)real uncertaintyindex(LMN)todefineState t−1 (“High”whentheHP-detrendedindexisabovetrendatt−1).The dependentvariableistherevisioninone-year-aheadinflationexpectations.MPS isthenormalizedcumulative t high-frequencymonetarypolicyshocksfromt tot+5.Accuracyismeasuredatthefirstinterview;Weinclude contemporaneousIndustrialProductiongrowthandinflationchangesbetweeninterviews. Alllower-order interactions,demographics,andsurveycontrolsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 40

AppendixTableC.3: AttentionwiththeVIX VIX (1)Accurate (2)Inaccurate (3)Haven’tHeard PanelA:HighVolatility (1)High×MPS -0.456 ∗∗∗ 0.040 -0.098 t (-4.06) (0.22) (-0.70) (2)High×∆IP 0.078 ∗∗∗ 0.074 ∗∗ -0.000 t (2.73) (1.96) (-0.01) (3)High×∆π -0.011 0.141 ∗ 0.137 ∗ t (-0.17) (1.92) (1.83) PanelB:LowVolatility (4)Low×MPS -0.007 0.100 -0.179 t (-0.07) (0.79) (-1.45) (5)Low×∆IP 0.020 -0.027 0.007 t (0.97) (-1.40) (0.35) (6)Low×∆π 0.538 ∗∗∗ 0.338 ∗∗∗ 0.413 ∗∗∗ t (11.82) (6.51) (7.37) Interaction VIX Controls Yes Observations 37,445 R2 0.0182 Notes:Thistablereportsthestate-dependentregressioninEquation(4.3)usingfinancial-marketvolatility(VIX) todefineState t−1 (“High”whentheHP-detrendedlogVIXisabovetrendatt−1).Thedependentvariableisthe revisioninone-year-aheadinflationexpectations.MPS denotesthenormalizedcumulativehigh-frequency t monetarypolicyshocksfromt tot+5. Accuracyismeasuredatthefirstinterview. WeincludecontemporaneousIndustrialProductiongrowthandinflationchangesbetweeninterviews. Alllower-orderinteractions, demographics,andsurveycontrolsareincluded.Robuststandarderrors. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 41

C.2 FullReports: DemographicHeterogeneity Thissectionpresentsthefullregressiontablescorrespondingtothedemographicheterogeneity analysis inSection4.4. Eachtable detailsthe complete setofinteractioncoefficients for the partitionsbasedonstockholding,homeownership,agegroup,andincomequartile,including resultsforallthreeaccuracygroupsandmacrocontrolvariables. AppendixTableC.4: FullreportsforStockholding (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)Stock×MPS -0.410 ∗∗∗ 0.150 -0.299 ∗∗ t (-4.57) (1.20) (-2.38) (2)NonStock×MPS -0.228 -0.047 0.034 t (-1.42) (-0.21) (0.24) (3)Stock×∆IP 0.053 ∗∗∗ 0.001 0.020 t (2.84) (0.09) (0.91) (4)NonStock×∆IP 0.090 ∗∗∗ -0.049 0.004 t (2.35) (-1.22) (0.11) (5)Stock×∆π 0.394 ∗∗∗ 0.258 ∗∗∗ 0.377 ∗∗∗ t (9.67) (5.92) (6.89) (6)NonStock×∆π 0.292 ∗∗∗ 0.318 ∗∗∗ 0.241 ∗∗∗ t (3.18) (2.86) (3.06) Interaction Stockholding Controls Yes Observations 37,445 R2 0.0142 Notes:ThistablereportsthefullsetofcoefficientsforthehomeownershipspecificationofEquation(4.4).The dependentvariableistherevisioninone-year-aheadinflationexpectations. MPS denotesthenormalized t cumulative high-frequency monetary policy shocks from t to t+5. We interact MPS with the three-way t accuracyvector(measuredatthefirstinterview)andhomeownershipstatus(Homeowner/Renter).Weinclude contemporaneousIndustrialProductiongrowthandinflationchangesbetweeninterviews; Weincludeage andage2,incomeandquartiles,education,gender,homeownership,stockholding,maritalstatus,region,and ∗ sentimentascontrols;alllower-ordertermsareincluded.Robuststandarderrors;t-statisticsinparentheses. p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 42

AppendixTableC.5: FullreportsforHomeownership (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)Homeowner×MPS -0.436 ∗∗∗ 0.063 -0.148 t (-5.08) (0.54) (-1.40) (2)Renter×MPS 0.026 0.214 -0.181 t (0.13) (0.76) (-0.91) (3)Homeowner×∆IP 0.057 ∗∗∗ -0.0000 0.027 t (3.11) (-0.00) (1.22) (4)Renter×∆IP 0.075 ∗ -0.032 -0.024 t (1.88) (-0.81) (-0.67) (5)Homeowner×∆π 0.386 ∗∗∗ 0.286 ∗∗∗ 0.334 ∗∗∗ t (9.61) (6.43) (6.61) (6)Renter×∆π 0.297 ∗∗∗ 0.166 0.276 ∗∗∗ t (2.67) (1.30) (2.73) Interaction Homeownership Controls Yes Observations 37,445 R2 0.0144 Notes:ThistablereportsthefullsetofcoefficientsforthehomeownershipspecificationofEquation(4.4).The dependentvariableistherevisioninone-year-aheadinflationexpectations. MPS denotesthenormalized t cumulative high-frequency monetary policy shocks from t to t+5. We interact MPS with the three-way t accuracyvector(measuredatthefirstinterview)andhomeownershipstatus(Homeowner/Renter).Weinclude contemporaneousIndustrialProductiongrowthandinflationchangesbetweeninterviews; Weincludeage andage2,incomeandquartiles,education,gender,homeownership,stockholding,maritalstatus,region,and ∗ sentimentascontrols;alllower-ordertermsareincluded.Robuststandarderrors;t-statisticsinparentheses. p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 43

AppendixTableC.6: FullreportsforAgeGroup (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)[18−34]×MPS -0.613 ∗∗∗ 0.260 -0.006 t (-3.22) (0.82) (-0.03) (2)[35−64]×MPS -0.350 ∗∗∗ 0.140 -0.145 t (-3.68) (1.12) (-1.16) (3)[65+]×MPS -0.234 -0.264 -0.338 ∗ t (-1.22) (-0.98) (-1.66) (4)[18−34]×∆IP 0.110 ∗∗ 0.031 0.004 t (2.43) (0.72) (0.10) (5)[35−64]×∆IP 0.059 ∗∗∗ -0.019 0.015 t (2.95) (-0.84) (0.57) (6)[65+]×∆IP 0.038 0.004 0.022 t (1.01) (0.13) (0.60) (7)[18−34]×∆π 0.381 ∗∗∗ 0.033 0.206 ∗∗ t (3.68) (0.27) (2.07) (8)[35−64]×∆π 0.435 ∗∗∗ 0.328 ∗∗∗ 0.373 ∗∗∗ t (9.41) (6.17) (5.94) (9)[65+]×∆π 0.179 ∗∗ 0.243 ∗∗∗ 0.311 ∗∗∗ t (2.20) (2.92) (3.66) Interaction AgeGroup Controls Yes Observations 37,445 R2 0.0150 Notes: This table reports the full set of coefficients for the age-group specification of Equation (4.4). The dependentvariableistherevisioninone-year-aheadinflationexpectations.MPS isthenormalizedcumulative t high-frequencymonetarypolicyshocksfromttot+5.WeinteractMPS withthethree-wayaccuracyvector t andagegroups(Young18-34,Middle35-64,Old65+). WeincludecontemporaneousIndustrialProduction growthandinflationchangesbetweeninterviews;Weincludeageandage2,incomeandquartiles,education, gender,homeownership,stockholding,maritalstatus,region,andsentimentascontrols;alllower-orderterms areincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 44

AppendixTableC.7: FullReportsforIncomeQuartile (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)YTL1×MPS 0.048 -0.192 0.149 t (0.18) (-0.58) (0.71) (2)YTL2×MPS -0.669 ∗∗∗ 0.046 -0.212 t (-3.90) (0.19) (-1.12) (3)YTL3×MPS -0.361 ∗∗∗ 0.201 -0.119 t (-2.58) (1.04) (-0.77) (4)YTL4×MPS -0.298 ∗∗ 0.132 -0.416 ∗∗ t (-2.47) (0.77) (-2.07) (5)YTL1×∆IP 0.024 -0.043 -0.059 t (0.48) (-0.93) (-1.37) (6)YTL2×∆IP 0.142 ∗∗∗ -0.062 0.017 t (3.43) (-1.49) (0.49) (7)YTL3×∆IP 0.024 -0.018 0.020 t (0.82) (-0.59) (0.56) (8)YTL4×∆IP 0.054 ∗∗ 0.047 ∗ 0.069 ∗ t (2.26) (1.80) (1.93) (9)YTL1×∆π 0.427 ∗∗∗ 0.220 0.282 ∗∗∗ t (3.30) (1.47) (2.68) (10)YTL2×∆π 0.265 ∗∗∗ 0.313 ∗∗∗ 0.271 ∗∗∗ t (3.16) (3.12) (3.12) (11)YTL3×∆π 0.438 ∗∗∗ 0.307 ∗∗∗ 0.414 ∗∗∗ t (6.51) (4.35) (5.37) (12)YTL4×∆π 0.352 ∗∗∗ 0.238 ∗∗∗ 0.329 ∗∗∗ t (6.16) (3.88) (3.51) Interaction IncomeQuartile Controls Yes Observations 37,445 R2 0.0153 Notes:Thistablereportsthefullsetofcoefficientsfortheincome-quartilespecificationofEquation(4.4).The dependentvariableistherevisioninone-year-aheadinflationexpectations. MPS denotesthenormalized t cumulative high-frequency monetary policy shocks from t to t+5. We interact MPS with the three-way t accuracyvectorandincomequartiles(YTL1–YTL4).WeincludecontemporaneousIndustrialProductiongrowth andinflationchangesbetweeninterviews;Weincludeageandage2,incomeandquartiles,education,gender, homeownership, stockholding, maritalstatus, region, andsentimentascontrols; alllower-ordertermsare included.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 45

C.3 AccuracyMeasurewithIP Thissectionteststherobustnessofourfindingstoanalternativedefinitionoftheaccuracyproxy. Here, we reconstruct the “Accurate” and “Inaccurate” classifications using the three-month changeinIndustrialProduction(IP)insteadoftheunemploymentrate. AppendixTableC.8: AccuracyMeasurewithIndustrialProduction (1)Accurate (2)Inaccurate (3)Haven’tHeard ∗∗∗ ∗ (1)MPS -0.334 -0.044 -0.156 t (-3.84) (-0.47) (-1.67) (2)∆IP 0.053 ∗∗∗ -0.003 0.013 t (3.32) (-0.16) (0.71) (3)∆π 0.336 ∗∗∗ 0.351 ∗∗∗ 0.325 ∗∗∗ t (8.26) (8.90) (7.20) Controls Yes Observations 37,445 R2 0.0135 Notes: ThistablereconstructstheaccuracymeasureusingIPastheobjectivecomparator. Arespondentis“Accurate”ifthesignoftheir reportedbusinessconditionnewsalignswiththesignofthethree-monthchangeinIPbetweenthetwointerviewmonths;“Inaccurate”ifit doesnot;“Haven’theard”otherwise.Were-estimatethebaselinespecificationEquation(4.1)usingthisIP-basedaccuracy.Thedependent variableistherevisioninone-year-aheadinflationexpectations. MPSt isthenormalizedcumulativehigh-frequencymonetarypolicy shocksfromttot+5.WeincludecontemporaneousIPgrowthandinflationchangesbetweeninterviews;demographicsandsurveycontrols areincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. AppendixTableC.9: IPSpecificationforStockholding (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)Stock×MPS -0.398 ∗∗∗ -0.002 -0.299 ∗∗ t (-4.02) (-0.02) (-2.38) (2)NonStock×MPS -0.158 -0.154 0.033 t (-0.89) (-0.80) (0.23) (3)Stock×∆IP 0.053 ∗∗∗ 0.0007 0.020 t (3.07) (0.03) (0.93) (4)Non-stock×∆IP 0.055 -0.018 0.004 t (1.47) (-0.45) (0.11) (5)Stock×∆π 0.352 ∗∗∗ 0.351 ∗∗∗ 0.377 ∗∗∗ t (8.10) (8.50) (6.89) (6)Non-stock×∆π 0.284 ∗∗∗ 0.351 ∗∗∗ 0.241 ∗∗∗ t (2.79) (3.54) (3.06) Interaction Stockownership Controls Yes Observations 37,445 R2 0.0138 Notes: This table re-estimates Equation (4.4) with the IP-based accuracy measure and the Stockholding partition (Stockholder/Nonstockholder).Thedependentvariableistherevisioninone-year-aheadinflationexpectations.MPStdenotesthenormalizedcumulative high-frequencymonetarypolicyshocksfromttot+5.WeincludecontemporaneousIPgrowthandinflationchangesbetweeninterviews; ∗ demographicsandsurveycontrolsareincluded;alllower-ordertermsareincluded.Robuststandarderrors;t-statisticsinparentheses. p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 46

AppendixTableC.10: IPSpecificationforHomeownership (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)Homeowner×MPS -0.361 ∗∗∗ -0.186 ∗ -0.149 t (-3.79) (-1.83) (-1.40) (2)Renter×MPS -0.185 0.660 ∗∗ -0.182 t (-0.89) (2.56) (-0.91) (3)Homeowner×∆IP 0.051 ∗∗∗ 0.009 0.027 t (2.95) (0.43) (1.23) (4)Renter×∆IP 0.061 -0.060 -0.024 t (1.64) (-1.47) (-0.67) (5)Homeowner×∆π 0.353 ∗∗∗ 0.347 ∗∗∗ 0.334 ∗∗∗ t (8.31) (8.12) (6.61) (6)Renter×∆π 0.246 ∗ 0.369 ∗∗∗ 0.275 ∗∗∗ t (1.89) (3.56) (2.73) Interaction Homeownership Controls Yes Observations 37,445 R2 0.0142 Notes: This table re-estimates Equation (4.4) with the IP-based accuracy measure (see Table B.7) and the Homeownershippartition(Homeowner/Renter). Thedependentvariableistherevisioninone-year-ahead inflationexpectations. MPS isthenormalizedcumulativehigh-frequencymonetarypolicyshocksfromt t to t+5. WeincludecontemporaneousIPgrowthandinflationchangesbetweeninterviews; thefullsetof demographicsandsurveycontrolsisincluded; alllower-ordertermsareincluded. Robuststandarderrors; t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 47

AppendixTableC.11: IPSpecificationforAgeGroup (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)[18−34]×MPS -0.439 ∗∗ -0.236 -0.007 t (-1.99) (-0.95) (-0.04) (2)[35−64]×MPS -0.327 ∗∗∗ 0.003 -0.145 t (-3.14) (0.03) (-1.16) (3)[65+]×MPS -0.322 -0.032 -0.338 ∗ t (-1.53) (-0.14) (-1.66) (4)[18−34]×∆IP 0.106 ∗∗ 0.018 0.004 t (2.52) (0.42) (0.11) (5)[35−64]×∆IP 0.057 ∗∗∗ -0.020 0.015 t (2.84) (-0.87) (0.58) (6)[65+]×∆IP 0.015 0.022 0.023 t (0.48) (0.53) (0.61) (7)[18−34]×∆π 0.304 ∗∗∗ 0.201 ∗ 0.206 ∗∗ t (2.70) (1.74) (2.07) (8)[35−64]×∆π 0.386 ∗∗∗ 0.426 ∗∗∗ 0.373 ∗∗∗ t (7.68) (8.73) (5.94) (9)[65+]×∆π 0.212 ∗∗ 0.227 ∗∗∗ 0.311 ∗∗∗ t (2.49) (2.83) (3.65) Interaction AgeGroup Controls Yes Observations 37,445 R2 0.0144 Notes:Thistablere-estimatesEquation(4.4)withtheIP-basedaccuracymeasureandtheAgepartition(Young 18-34,Middle35-64,Old65+).Thedependentvariableistherevisioninone-year-aheadinflationexpectations. MPS denotesthenormalizedcumulativehigh-frequencymonetarypolicyshocksfromt tot+5.Weinclude t contemporaneousIPgrowthandinflationchangesbetweeninterviews;demographicsandsurveycontrolsare included;alllower-ordertermsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 48

AppendixTableC.12: IPSpecificationforIncomeQuartile (1) (2) (3) Accurate Inaccurate Haven’tHeard (1)YTL1×MPS 0.009 0.018 0.148 t (0.03) (0.06) (0.71) (2)YTL2×MPS -0.595 ∗∗∗ -0.227 -0.213 t (-3.18) (-1.03) (-1.13) (3)YTL3×MPS -0.248 -0.145 -0.119 t (-1.59) (-0.88) (-0.77) (4)YTL4×MPS -0.358 ∗∗∗ 0.131 -0.416 ∗∗ t (-2.69) (0.89) (-2.08) (5)YTL1×∆IP 0.008 -0.066 -0.059 t (0.19) (-1.27) (-1.37) (6)YTL2×∆IP 0.084 ∗∗ 0.014 0.017 t (2.25) (0.30) (0.50) (7)YTL3×∆IP 0.031 -0.023 0.020 t (1.08) (-0.70) (0.56) (8)YTL4×∆IP 0.068 ∗∗∗ 0.031 0.069 ∗ t (2.93) (1.10) (1.93) (9)YTL1×∆π 0.313 ∗∗ 0.424 ∗∗∗ 0.282 ∗∗∗ t (2.11) (3.33) (2.68) (10)YTL2×∆π 0.297 ∗∗∗ 0.345 ∗∗∗ 0.271 ∗∗∗ t (3.25) (3.75) (3.12) (11)YTL3×∆π 0.375 ∗∗∗ 0.389 ∗∗∗ 0.414 ∗∗∗ t (5.31) (5.60) (5.36) (12)YTL4×∆π 0.333 ∗∗∗ 0.295 ∗∗∗ 0.329 ∗∗∗ t (5.45) (5.15) (3.51) Interaction IncomeQuartile Controls Yes Observations 37,445 R2 0.0148 Notes: Thistablere-estimatesEquation(4.4)withtheIP-basedaccuracymeasureandtheIncomepartition (YTL1–YTL4). Thedependentvariableistherevisioninone-year-aheadinflationexpectations. MPS isthe t normalizedcumulativehigh-frequencymonetarypolicyshocksfromt tot+5.WeincludecontemporaneousIP growthandinflationchangesbetweeninterviews;demographicsandsurveycontrolsareincluded;alllower-order termsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 49

C.4 AccuracymeasurewithNFCI Thissectionprovidesafurtherrobustnesscheckontheconstructionofouraccuracyproxy. We redefineaccuracyusingthethree-monthchangeintheNationalFinancialConditionsIndex (NFCI)asthebenchmark,wherearisingindexsignalsunfavorableconditions. 50

AppendixTableC.13: AccuracyMeasurewithNFCI (1)Accurate (2)Inaccurate (3)Haven’tHeard ∗∗∗ ∗ (1)MPS -0.336 -0.002 -0.155 t (-3.88) (-0.03) (-1.66) (2)∆IP 0.039 ∗∗ 0.021 0.013 t (2.46) (1.14) (0.70) (3)∆π 0.315 ∗∗∗ 0.375 ∗∗∗ 0.325 ∗∗∗ t (7.98) (8.95) (7.21) Controls Yes Observations 37,445 R2 0.0135 Notes: ThistablereconstructstheaccuracymeasureusingtheNFCIastheobjectivecomparatorforbusinessconditions. Arespondent is“Accurate”ifthesignoftheirreportednewsalignswiththesignofthethree-monthchangeinNFCI(withhigherNFCIinterpretedas tighter,i.e.,unfavorable,financialconditions);“Inaccurate”ifnot;“Haven’theard”otherwise. Were-estimatethebaselinespecification Equation(4.1)withthisNFCI-basedaccuracy. Thedependentvariableistherevisioninone-year-aheadinflationexpectations. MPSt is thenormalizedcumulativehigh-frequencymonetarypolicyshocksfromttot+5.WeincludecontemporaneousIPgrowthandinflation changesbetweeninterviews;demographicsandsurveycontrolsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. AppendixTableC.14: NFCISpecificationforStockholding (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)Stock×MPS -0.340 ∗∗∗ -0.028 -0.299 ∗∗ t (-3.51) (-0.26) (-2.38) (2)NonStock×MPS -0.336 ∗ 0.069 0.033 t (-1.74) (0.40) (0.24) (3)Stock×∆IP 0.038 ∗∗ 0.026 0.020 t (2.20) (1.23) (0.92) (4)Non-stock×∆IP 0.046 0.003 0.004 t (1.17) (0.09) (0.11) (5)Stock×∆π 0.308 ∗∗∗ 0.390 ∗∗∗ 0.377 ∗∗∗ t (7.43) (8.62) (6.89) (6)Non-stock×∆π 0.341 ∗∗∗ 0.329 ∗∗∗ 0.241 ∗∗∗ t (3.28) (3.33) (3.06) Interaction Stockownership Controls Yes Observations 37,445 R2 0.0137 Notes:Thistablere-estimatesEquation(4.4)withtheNFCI-basedaccuracymeasureandtheStockholdingpartition.Thedependentvariableistherevisioninone-year-aheadinflationexpectations. MPSt denotesthenormalizedcumulativehigh-frequencymonetarypolicy shocksfromttot+5.WeincludecontemporaneousIPgrowthandinflationchangesbetweeninterviews;demographicsandsurveycontrolsareincluded;alllower-ordertermsareincluded. Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 51

AppendixTableC.15: NFCISpecificationforHomeownership (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)Homeowner×MPS -0.430 ∗∗∗ -0.020 -0.148 t (-4.65) (-0.20) (-1.40) (2)Renter×MPS 0.170 0.086 -0.181 t (0.71) (0.43) (-0.91) (3)Homeowner×∆IP 0.046 ∗∗∗ 0.021 0.027 t (2.71) (1.02) (1.22) (4)Renter×∆IP 0.005 0.024 -0.024 t (0.15) (0.64) (-0.67) (5)Homeowner×∆π 0.317 ∗∗∗ 0.395 ∗∗∗ 0.334 ∗∗∗ t (7.84) (8.73) (6.61) (6)Renter×∆π 0.296 ∗∗ 0.258 ∗∗ 0.275 ∗∗∗ t (2.19) (2.32) (2.73) Interaction Homeownership Controls Yes Observations 37,445 R2 0.0141 Notes:Thistablere-estimatesEquation(4.4)withtheNFCI-basedaccuracymeasureandtheHomeownership partition(Homeowner/Renter).Thedependentvariableistherevisioninone-year-aheadinflationexpectations. MPS isthenormalizedcumulativehigh-frequencymonetarypolicyshocksfromt tot+5.Weincludecontemt poraneousIPgrowthandinflationchangesbetweeninterviews;thefullsetofdemographicsandsurveycontrols isincluded;alllower-ordertermsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 52

AppendixTableC.16: NFCISpecificationforAgeGroup (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)[18−34]×MPS -0.167 -0.379 ∗ -0.006 t (-0.68) (-1.83) (-0.04) (2)[35−64]×MPS -0.363 ∗∗∗ 0.068 -0.145 t (-3.50) (0.34) (-1.16) (3)[65+]×MPS -0.372 ∗ 0.079 -0.338 ∗ t (-1.81) (0.34) (-1.66) (4)[18−34]×∆IP 0.059 0.084 ∗ 0.004 t (1.40) (1.86) (0.10) (5)[35−64]×∆IP 0.031 0.026 0.015 t (1.57) (1.14) (0.57) (6)[65+]×∆IP 0.056 ∗ -0.024 0.022 t (1.72) (-0.61) (0.60) (7)[18−34]×∆π 0.046 0.440 ∗∗∗ 0.206 ∗∗ t (0.41) (3.84) (2.07) (8)[35−64]×∆π 0.418 ∗∗∗ 0.391 ∗∗∗ 0.373 ∗∗∗ t (8.34) (7.72) (5.94) (9)[65+]×∆π 0.145 ∗∗ 0.287 ∗∗∗ 0.311 ∗∗∗ t (1.96) (3.12) (3.65) Interaction AgeGroup Controls Yes Observations 37,445 R2 0.0147 Notes:Thistablere-estimatesEquation(4.4)withtheNFCI-basedaccuracymeasureandtheAgepartition(Young 18-34,Middle35-64,Old65+).Thedependentvariableistherevisioninone-year-aheadinflationexpectations. MPS denotesthenormalizedcumulativehigh-frequencymonetarypolicyshocksfromt tot+5.Weinclude t contemporaneousIPgrowthandinflationchangesbetweeninterviews;demographicsandsurveycontrolsare included;alllower-ordertermsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 53

AppendixTableC.17: NFCISpecificationforIncomeQuartile (1)Accurate (2)Inaccurate (3)Haven’tHeard (1)YTL1×MPS 0.118 -0.155 0.148 t (0.40) (-0.54) (0.71) (2)YTL2×MPS -0.563 ∗∗∗ -0.271 -0.212 t (-2.81) (-1.40) (-1.13) (3)YTL3×MPS -0.418 ∗∗∗ 0.240 -0.119 t (-2.75) (1.44) (-0.77) (4)YTL4×MPS -0.271 ∗∗ 0.024 -0.416 ∗∗ t (-2.03) (0.17) (-2.07) (5)YTL1×∆IP -0.003 -0.025 -0.059 t (-0.08) (-0.49) (-1.37) (6)YTL2×∆IP 0.040 0.067 0.017 t (1.14) (1.36) (0.49) (7)YTL3×∆IP 0.013 0.003 0.020 t (0.43) (0.10) (0.56) (8)YTL4×∆IP 0.076 ∗∗∗ 0.026 0.069 ∗ t (3.11) (1.00) (1.92) (9)YTL1×∆π 0.490 ∗∗∗ 0.273 ∗∗ 0.282 ∗∗∗ t (3.14) (2.17) (2.68) (10)YTL2×∆π 0.268 ∗∗∗ 0.343 ∗∗∗ 0.271 ∗∗∗ t (3.16) (3.51) (3.11) (11)YTL3×∆π 0.319 ∗∗∗ 0.464 ∗∗∗ 0.414 ∗∗∗ t (4.94) (6.09) (5.37) (12)YTL4×∆π 0.277 ∗∗∗ 0.352 ∗∗∗ 0.329 ∗∗∗ t (4.64) (5.66) (3.51) Interaction IncomeQuartile Controls Yes Observations 37,445 R2 0.0151 Notes:Thistablere-estimatesEquation(4.4)withtheNFCI-basedaccuracymeasureandtheIncomepartition (YTL1–YTL4). Thedependentvariableistherevisioninone-year-aheadinflationexpectations. MPS isthe t normalizedcumulativehigh-frequencymonetarypolicyshocksfromt tot+5.WeincludecontemporaneousIP growthandinflationchangesbetweeninterviews;demographicsandsurveycontrolsareincluded;alllower-order termsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 54

C.5 Others Thissectioncontainsasetofrobustnesschecks. Were-estimateourbaselinemicro-levelspecificationusinghousehold-headweightstoensurepopulationrepresentativeness,addcontrolsfor gasolinepricechangestoaccountfortheirsalience,andusetheEconomicPolicyUncertainty (EPU)indexasanalternativemeasureforthestate-dependenceanalysis. AppendixTableC.18: HouseholdHeadWeight (1)Accurate (2)Inaccurate (3)Haven’tHeard ∗∗∗ (1)MPS -0.298 0.097 -0.154 t (-3.56) (0.81) (-1.52) (2)∆IP 0.061 ∗∗∗ -0.014 0.008 t (3.50) (-0.76) (0.40) (3)∆π 0.357 ∗∗∗ 0.266 ∗∗∗ 0.300 ∗∗∗ t (8.85) (5.73) (6.20) Controls Yes Observations 36,565 R2 0.0130 Notes: Thistablere-estimatesthebaselinemicrospecificationEquation(4.1)usinghousehold-headweightsprovidedbythesurveyto improvepopulationrepresentativenessoftherecontactsample.Thedependentvariableistherevisioninone-year-aheadinflationexpectations. MPSt denotesthenormalizedcumulativehigh-frequencymonetarypolicyshocksfromttot+5. Accuracyismeasuredatthe firstinterview;WeincludecontemporaneousIPgrowthandinflationchangesbetweeninterviews;thefullsetofdemographicsandsurvey controlsisincluded.Weightedleastsquareswithrobuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. AppendixTableC.19: IncludingGasolinePriceControls (1)Accurate (2)Inaccurate (3)Haven’tHeard ∗∗ (1)MPS -0.193 0.170 -0.088 t (-2.49) (1.55) (-0.94) (2)∆IP -0.001 -0.063 ∗∗∗ -0.030 t (-0.09) (-3.45) (-1.51) (3)∆π 0.035 0.061 0.112 ∗∗ t (0.83) (1.27) (2.21) (4)∆Gas 0.042 ∗∗∗ 0.033 ∗∗∗ 0.026 ∗∗∗ t (14.62) (10.84) (8.18) Controls Yes Observations 37,445 R2 0.0283 Notes: ThistableaugmentsthebaselinemicrospecificationEquation(4.1)byaddingthelogchangeinU.S.RegularAllFormulations GasolinePricebetweenthetwointerviewmonths(fromFRED)tocontrolforsalientpricemovements. Thedependentvariableisthe revisioninone-year-aheadinflationexpectations.MPStdenotesthenormalizedcumulativehigh-frequencymonetarypolicyshocksfrom ttot+5.Accuracyismeasuredatthefirstinterview;WeincludecontemporaneousIPgrowthandinflationchangesbetweeninterviews; demographicsandsurveycontrolsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 55

AppendixTableC.20: AttentionwithEconomicPolicyUncertainty EPU (1)Accurate (2)Inaccurate (3)Haven’tHeard HighUncertainty ∗∗∗ (1)MPS -0.651 -0.242 -0.267 t (-5.69) (-1.29) (-1.81) (2)∆IP 0.056 ∗∗ 0.011 0.010 t (2.53) (0.50) (0.38) (3)∆π 0.324 ∗∗∗ 0.383 ∗∗∗ 0.292 ∗∗∗ t (5.08) (5.91) (4.26) LowUncertainty ∗∗ (4)MPS 0.168 0.276 -0.007 t (1.58) (2.17) (-0.07) (5)∆IP 0.051 ∗∗ -0.025 0.009 t (1.98) (-0.98) (0.37) (6)∆π 0.373 ∗∗∗ 0.223 ∗∗∗ 0.342 ∗∗∗ t (7.79) (3.98) (5.75) Interaction EPU Controls Yes Observations 37,445 R2 0.0167 Notes:ThistableestimatesEquation(4.3)usingtheEconomicPolicyUncertainty(EPU)index(Bakeretal.,2016b) todefineState t−1 (“HighEPU”whentheHP-detrendedEPUisabovetrendatt−1). Thedependentvariable istherevisioninone-year-aheadinflationexpectations. MPS isthenormalizedcumulativehigh-frequency t monetarypolicyshocksfromttot+5.Accuracyismeasuredatthefirstinterview.Weincludecontemporaneous IPgrowthandinflationchangesbetweeninterviews;thefullsetofdemographicsandsurveycontrolsisincluded; alllower-ordertermsareincluded.Robuststandarderrors;t-statisticsinparentheses. ∗ p<0.10, ∗∗ p<0.05, ∗∗∗ p<0.01. 56