《IMF:陷入财政乘数的愤怒:小细节大影响(英文版)(49页).pdf》由会员分享,可在线阅读,更多相关《IMF:陷入财政乘数的愤怒:小细节大影响(英文版)(49页).pdf(49页珍藏版)》请在三个皮匠报告上搜索。
1、IMF Working Papers describe research in progress by the author(s)and are published to elicit comments and to encourage debate.The views expressed in IMF Working Papers are those of the author(s)and do not necessarily represent the views of the IMF,its Executive Board,or IMF management.2023 FEB Getti
2、ng into the Nitty-Gritty of Fiscal Multipliers:Small Details,Big Impacts Jos Federico Geli and Afonso S.Moura WP/23/WP/23/29 *We are grateful to Ali Alichi,Marie-Pierre Aquino Acoste,Andrew Berg,Francesca Caselli,Antoine Cornevin,Nikolay Gueorguiev,Zeina Hasna,Ryota Nakatani,Martin Schlinder,Koon Te
3、e,Anna Ter-Martirosyan and participants at the ICD DepartmentalSeminar for very helpful comments and suggestions.2022 International Monetary Fund WP/23/29 IMF Working Paper Institute for Capacity Development Getting into the Nitty-Gritty of Fiscal Multipliers:Small Details,Big Impacts Prepared by Jo
4、s Federico Geli,Afonso S.Moura*Authorized for distribution by Ali Alichi February 2023 IMF Working Papers describe research in progress by the author(s)and are published to elicit comments and to encourage debate.The views expressed in IMF Working Papers are those of the author(s)and do not necessar
5、ily represent the views of the IMF,its Executive Board,or IMF management.ABSTRACT:Despite the remarkable progress the literature has made throughout the past years in studying fiscal multipliers,estimates still vary considerably across studies.Partly,estimates differ because of context-specific vari
6、ables that affect multipliers,but also because of the lack of a standardized framework to calculate and report them,making comparisons among studies hard to make.In this paper,we use a large panel of countries to study how some important methodological details affect the empirical estimates.Focusing
7、 on emerging economies,we show how slight changes in the filtering approach of fiscal forecast errors or the accumulation procedure of responses can significantly impact estimates.We emphasize that one of the most important features of estimating multipliers is the endogenous dynamic responses of fi
8、scal variables to fiscal shocks,and therefore we argue against reporting multipliers as simply the output response to exogenous fiscal innovations.Although our baseline results are in line with the previous studies,our standardized framework allow us to make fairer comparisons of multiplier estimate
9、s across budgetary items and country income groups.RECOMMENDED CITATION:Geli,Jos Federico and Afonso S.Moura,2023,“Getting into the Nitty-Gritty of Fiscal Multipliers:Small Details,Big Impacts”,IMF Working Papers 23/29.JEL Classification Numbers:E60,H30,H50 Keywords:Fiscal policy;Fiscal multipliers.
10、Authors E-Mail Address:Jos Federico Geli:jfgeliimf.org;Afonso S.Moura:afonso.mouranovasbe.pt WORKING PAPERS Getting into the Nitty-Gritty of Fiscal Multipliers:Small Details,Big Impacts Prepared by Jos Federico Geli and Afonso S.Moura1 1 We are grateful to Ali Alichi,Marie-Pierre Aquino Acoste,Andre
11、w Berg,Francesca Caselli,Antoine Cornevin,Nikolay Gueorguiev,Zeina Hasna,Ryota Nakatani,Martin Schlinder,Koon Tee,Anna Ter-Martirosyan and participants at the ICD Departmental Seminar for very helpful comments and suggestions.Contents1Introduction62Data93Analytical framework93.1Baseline specificatio
12、n.93.2Identification of fiscal shocks.114Results134.1Baseline results.134.2Alternative cleaning procedures of forecast errors.184.3One-stage vs Two-stage Estimation.204.4Different approach to accumulate responses.284.5Alternative data sources.305Conclusion and Policy Implications316Appendix376.1Desc
13、riptive Data.376.2Summary of Main Estimations.446.3Different Specifications.446.4Uncertainty surrounding Forecast Errors.466.5Instrument Relevance.48List of Figures1Multiplier heterogeneity across different groups of countries.162Multiplier heterogeneity across fiscal tools.183Multipliers using diff
14、erent ways of cleaning forecast errors.2044One vs Two-stage estimation.235Dynamic response of fiscal variable to the fiscal shock:Local projections offirst stage regressions.246Cumulative response of government budget balance to public investment andpublic consumption shocks.267SNA vs GFM data.318Hi
15、stogram of fiscal shocks for all sample,including AEs,EMEs and LICs.379Histogram of fiscal shocks for AEs.3810Histogram of fiscal shocks for EMEs.3811Histogram of fiscal shocks for LICs.3912GDP growth and Fiscal Shocks.40List of Tables1Multipliers when using FE of growth rates.282Different approache
16、s to calculate cumulative multipliers.293Income classification of countries.414Descriptive Statistics.425Sample size:number of countries per year.436Summary of Results.447Local projections including Covid years in the estimation.458Different fixed-effect specifications.459Different lag specification
17、s.4610Different clustering levels of Standard Errors.4611Estimated vs not estimated fiscal shock.4712F-statistic of first stage regression for instrument relevance.4851IntroductionIn the past years,especially after the financial crisis,there has been a remarkable interestin studying fiscal multiplie
18、rs.This has led to considerable innovations in both the empir-ical methodologies as well as in identifying theoretical determinants of their size1.One ofthe most important takeaways from this bulk of new research is how much multipliers arecontext-dependent.They can vary given the amount of slack in
19、 the economy(Auerbach andGorodnichenko,2012,Nakamura and Steinsson(2014),Hernndez de Cos and Moral-Benito(2016);the stance of monetary policy(Miyamoto,Nguyen and Sergeyev,2018);the direc-tion of the fiscal intervention(Barnichon,Debortoli and Matthes,2022);the financing source(Kraay,2012),among many
20、 other factors.However,estimates also diverge due to methodological differences,limiting comparisonacross studies.Results can vary due to the econometric approach,the identification strategy2and the multiplier definition.For this reason,comparing estimates across studies is not astraightforward task
21、.Without a standardized framework,this comparison exercise may leadto wrong conclusions and ill-informed policy advice.This raises the question of what methodsshould policymakers rely on to assess the impact of fiscal decisions,a question that is speciallyrelevant for Emerging Market economies(EMs)a
22、nd Low Income Countries(LICs)for whichthe literature on the topic is still relatively scarce.In this paper,we address this sources of heterogeneity among estimates in a consistent,uniform,and integrated framework,digging into the nitty-gritty of multipliers estimation andreporting.Although our basel
23、ine results are in line with the range of estimates Ramey(2019)reports,we show how some subtle methodological details can have a significant impact onthe results.We do so by empirically3estimating multipliers across different horizons,country1See,for example,Ramey(2019)for a survey on the topic.2Cal
24、dara and Kamps(2017)derive an uniform framework and compare how different commonly usedidentification schemes implicitly or explicitly determine the size of the estimated fiscal multiplier.3We focus on empirical methodologies instead of multipliers calculated from estimated and calibrateddynamic sto
25、chastic general equilibrium(DSGE)models.For a survey on the differences in the estimatedmultipliers across the two different approaches,see Ramey(2019).6groups and budgetary items,using the Jord(2005)local projection methodology.We followRamey and Zubairy(2018)and estimate the local projections with
26、in a two-stage instrumentvariable framework using exogenously identified fiscal shocks as instrumental variables,andemphasize some of the most important advantages of this methodology.We identify fiscal shocks by calculating forecast errors(FE)of fiscal variables using IMFsWorld Economic Outlook dat
27、a.4Our analysis shows that obtaining valid and relevant ex-ogenous fiscal shocks using forecast errors requires a careful filtering procedure.Otherwise,apparently good candidates for instruments could bias the results,usually increasing theestimated multiplier.Another dimension we take into account
28、is how fiscal variables endogenously respondto fiscal shocks throughout the horizon and how this impacts multipliers.We do this bycomparing the results of the two-stage instrument variable(IV)approach with a(commonlyused in the literature)one-step estimation that calculates the impact of exogenous s
29、hocksdirectly on output,without using them as IVs.Our results imply that neglecting this dynamiceffect that shocks have on fiscal variables leads to incorrect fiscal multiplier estimates,bothin terms of magnitude and persistence.In addition,we highlight how following the two-stagemethodology can att
30、enuate differences in results from using forecast errors calculated underalternative approaches.Finally,we also show how the estimation framework can easily beextended to assess the impact of the shocks on the governments financing needs.Different data sources and reporting methodologies can also be
31、 a source of differencesacross estimates.Most expenditure-based estimates of fiscal multipliers use as input govern-ment expenditure data from the System of National Accounts(SNA).However,fiscal policymeasures are usually designed using Government Finance Statistics(GFS)data.In manycountries,in part
32、icular in EMs and LICs,one of the main differences between these two typesof data is that SNA data is an accrued estimate of the cash-based fiscal accounts.Our resultssignal that using these two sources of data can also lead to substantial differences across re-4We follow this identification approac
33、h throughout the paper because we aim to compare results in auniform framework.7sults.Furthermore,different approaches to accumulate responses yield different multipliersestimates,even if the data source,identification strategy and methodological approach arethe same.An additional contribution to th
34、e literature is that we focus our analysis mainly on EMsand compare estimates of multipliers for both expenditure and revenue categories.Due todata availability or statistically quality reasons,most studies on fiscal multipliers have focusedon advanced economies(AEs).However,studies that do focus on
35、 EMs can reach differentconclusions not because of methodological differences,but due to different sample size andlength considered5.A uniform framework is again important to overcome this comparisonlimitation and extend our understanding of multipliers in EMs and LICs.In addition,mostauthors who re
36、ly on forecast errors to identify fiscal shocks focus on expenditure multipliers6.To the best of our knowledge,we are the first ones to extend this identification approachto the revenue side.By reporting fiscal multiplier estimates within the same analyticalframework,we can better single out differe
37、nces among public investment,public consumptionand personal income tax multipliers.Finally,by relying on a uniform framework,our estimates can also be used to enhancetools such as the bucket approach(Batini,Eyraud and Weber,2014),which is based onestimates from studies using different methodologies.
38、The rest of the paper is organized as follows.Section 2 describes the data used in thestudy.Section 3 explains the analytical framework.We show the baseline econometricspecification and carefully explain the process to identify exogenous fiscal shocks.Section 4explores the results.We show the baseli
39、ne results and all the important changes that lead to5A few examples on how much the sample size can vary are:Ilzetzki,Mendoza and Vgh(2013)who usea panel of 44 countries,of which only 24 are developing economies;Carrire-Swallow,David and Leigh(2021)who present evidence for 14 Latin American and Car
40、ibbean countries;Restrepo(2020)who focus only onLatin America economies;and Honda,Miyamoto and Taniguchi(2020)who use data for 42 LICs which arenot resource rich economies6They do so either by looking at government consumption or government expenditure.Moreover,eventhough the evidence shows that mul
41、tipliers for public consumption and investment can differ,a usual practicein the literature is to sum their shocks as a measure of total government expenditure innovations(IMF,2018,Colombo et al.,2022).8significant differences in the estimates.We explore either different specifications comparedto th
42、e baseline results,or use different data(either by slightly changing the fiscal shocks usedas instruments or the source of the data being used).Section 5 concludes and analyzes thepolicy implications of the paper.2DataWe use an unbalanced panel of countries with annual data taken from the IMFs World
43、Economic Outlook(WEO)database archives.For public consumption and public investmentmultipliers,the exercises include data spanning from 1995 until 2019.Personal income taxesseries,however,is shorter.Because we need real-time vintage data when computing thefiscal forecast errors,for this variable ava
44、ilable data starts only in 2010.All variables areexpressed in real terms,deflated by the GDP deflator.Since forecast errors of fiscal variablesor variations in output can be particularly large for the Covid-19 years,only data until 2019was included in the exercises.7The baseline exercises include 38
45、 AEs,85 EMs,and 54 LICs.See appendix 6.1 for descriptive statistics of the data used.3Analytical framework3.1Baseline specificationTo calculate8multipliers over a certain horizon after the fiscal shock takes place,we usethe Jord(2005)local projection methodology9.We specify the regressions such that
46、 theimpulse response functions that the method yields can be directly interpreted as cumulativemultipliers for each horizon h.As in Auerbach and Gorodnichenko(2013),we apply thismethod in a context of a panel data framework.Additionally,we follow Ramey and Zubairy(2018)and extend the local projectio
47、ns to an instrumental variable framework.7Appendix 6.3 shows how results change when including the year of 2020 and 2021 in the sample.8All codes were written using R and can be made available upon request.9Local projection methods do not impose any dynamic restrictions to the model,as opposed to a
48、VARstructure.Additionally,the local projection model can be more parsimonious than the VAR specification.9For the baseline model,we estimate the following regressions for each horizon h=0,1,.,H:yi,t+h=i,h+t,h+hfi,t+h+hXi,t1+i,t+h(1)fi,t+h=i,h+t,h+hshockfi,t+hXi,t1+ui,t+h,(2)where equation 1 is the s
49、econd stage regression and equation 2 is the first stage regression inwhich an exogenous fiscal shockfvariable is used as an instrument for the respective fiscalvariable.We focus on three different types of fiscal tools,so f can be either governmentconsumption,government investment or personal incom
50、e taxes.Both regressions includecountry and time fixed effects.In the baseline specification,Xi,t1includes one lag of yi,t,fi,tand shocki,t10.The dependent variable yi,t+his the cumulative change in GDP from t tot+h and fi,t+his the cumulative change in the fiscal variable of interest from t to t+h.
51、Wefollow Canova and Pappa(2021)(henceforth,C&P)and express the cumulative variables,inorder to calculate cumulative multipliers at each horizon h,as yi,t+h=hP1Yi,t+h1hYi,t1Yi,t1andfi,t+h=hP1Fi,t+h1hFi,t1Yi,t1.All variables are normalized by the same period GDP(at t 1),and therefore can be interprete
52、d as unit multipliers directly,instead of elasticities11.We define cumulative multipliers as in Ramey and Zubairy(2018),which slightly changesthe Mountford and Uhlig(2009)definition of present value multipliers.That is,we acknowl-edge that in a dynamic environment,multipliers should be reported as t
53、he integral of theoutput response divided by the integral government spending response after a fiscal shock tak-ing place12.Because of the IV approach and the way we accumulate both output and the10Appendix 6.3 shows results using different econometric specifications.Especially for the case of themu
54、ltiplier for personal income taxation,since the number of the years included in the exercise is smaller,theuse of a within estimator when the model has a dynamic component(the lagged term yi,t)can introducesome bias in the estimations.In the appendix,we report the results(for EMEs)when we drop the d
55、ynamiccomponent of the model and results do not significantly change.11See Ramey and Zubairy(2018)for a detailed explanation on the drawbacks of estimating elasticitiesand then convert the estimates into unit multipliers.12Mountford and Uhlig(2009)calculate the present value multiplier,discounted by
56、 the average interestrate of their sample.However,as our estimation is done in a panel data framework,discounting the multiplierswith the average of the interest rates across country and across time would not be a good exercise,given the10fiscal variable across each horizon,the estimates directly gi
57、ve dynamic cumulative multi-pliers at each horizon h.In section 4.3 we explore how results change when this dynamicenvironment is neglected.3.2Identification of fiscal shocksThe literature mostly relies on three different approaches to identify exogenous fiscalshocks:the narrative approach(Romer and
58、 Romer,2010),recursive approach(Blanchardand Perotti,2002)or the use of forecast errors(Auerbach and Gorodnichenko,2013).AsColombo et al.,2022 points out,the economic or institutional heterogeneity that may bepresent in a large panel of countries can make the identification of shocks difficult using
59、 thefirst two approaches due to their information requirements.For EMs and LICs,in whichpublicly available official documents are scarce,the narrative identification becomes unfea-sible.Additionally,the lack of macroeconomic statistics at a quarterly frequency also doesnot allow for a time restricti
60、on identification of fiscal shocks.To overcome this constraint,we construct forecast errors of fiscal variables and use it to derive a measure of fiscal shocks.Forecast errors for a given variable are the difference between the projected and therealized value.By construction,they should be unexpecte
61、d shocks to the economy13.Intheory,forecast errors avoid the problem of fiscal foresight:if a shock happens at t but waspreviously announced at t-1,an econometrician might not be able to find any significanteconomic response because economic agents may have reacted before the shock actually tookplac
62、e.By capturing only purely unanticipated changes,forecast errors reduce this fiscalforesight bias.We construct fiscal forecast errors using World Economic Outlook real-time vintage data,in the following way:FEfi,t=fi,t fi,t|t1,(3)heterogeneity among the all the countries in our panel.13As An et al.(
63、2018)demonstrate,WEO fiscal projections are relatively precise and unbiased in compar-ison with other private forecasters.11where f is the fiscal variable of interest,fi,trepresents its realized value,for a given countryand period of time,and fi,t|t1is the forecast for that same fiscal variable14.Bo
64、th fi,tandfi,t|t1are expressed as a share of its corresponding vintage contemporaneous GDP15.Thatis,fi,t=Fi,tYi,tand fi,t|t1=Fi,t|t1Yi,t|t1.However,even if forecast errors are unexpected shocks in theory,in practice they mightnot be completely orthogonal to past macroeconomic trends.If that would be
65、 the case,byexploiting information on past economic developments,agents could,at least partially,adjusttheir actions before the shock takes place.To avoid this effect,we regress the fiscal forecasterrors on a set of lagged macroeconomic variables,such as GDP growth rate,total revenueand total expend
66、iture as a percentage of GDP,exchange rate growth and inflation16.Another potential issue related to the use of fiscal forecast errors is that fiscal forecastsare usually based on a set of macroeconomic assumptions.If these assumptions change,evenif no new fiscal measures are implemented,fiscal fore
67、casts are probably going to diverge fromfinal outcomes.For example,a lower-than-expected VAT collection can be simply the resultof a lower-than-expected private consumption expenditure and not due to any additionaltax reduction or exemption measure.Hence,it is important to guarantee that identifiedf
68、iscal shocks are not correlated with contemporaneous economic conditions or other typesof shocks affecting the economy.To tackle this issue,we also regress fiscal forecast errorson contemporaneous forecast errors of GDP growth,exchange rates growth and inflation(asmeasured by the percent change of t
69、he output deflator).A similar approach is followedby Abiad,Furceri and Topalova(2016),IMF(2017)and IMF(2018),for example,thateither regress fiscal forecast errors on output and inflation forecast errors or test if regressingon other forecast errors(such as private consumption or investment)influence
70、s the results.14fi,tis taken from the October WEO released at t+1,which contains the realized variable at t,andfi,t|t1is taken from the October WEO released at t.15This also reduces the risk of artificially having a high forecast error because statistical changes occurredbetween WEO vintages,for exa
71、mple,because of a currency re-denomination.If this is the case,then nor-malizing the level of each fiscal variable by its corresponding GDP using the same vintage should overcomethis issue.16See,for example,Colombo et al.(2022)and Cacciatore et al.(2021)for similar ways to filter out theexpected com
72、ponent of forecast errors.12Additionally,we include in the regression forecast errors of exchange rate movements to(partially)capture the presence of external shocks.Before regressing each fiscal forecast errors on all of the above mentioned variables(ina panel data framework,including also time and
73、 country fixed effects),we first trim thesample,ruling out 1%of each tail of the distribution to remove(most likely)data outliers17.Additionally,we drop from our sample any forecast error that is exactly equal to zero.Theresiduals of this regression are used as the fiscal shocks18.The former filteri
74、ng process warrants that identified shocks respect the properties Ramey(2016)points out as being necessary for a shock to be considered as such:uncorrelatedwith past conditions of the economy,with other shocks and,by construction,unanticipated.Appendix 6.1 plots the histograms of the shocks for the
75、different tools by income groups.Onsection 4.2 we explore the impact of using alternative definitions of forecast errors.4Results4.1Baseline resultsThe first source of heterogeneity in estimates that we want to explore is the difference inoutput responses by country groups(AEs,EMs and LICs).In theor
76、y,estimated multipliersacross different country groups can be heterogeneous.For example,they can vary due todifferences in the exchange rate regime(Ilzetzki,Mendoza and Vgh,2013),on the quality ofthe institutions(Honda,Miyamoto and Taniguchi,2020),the degree of informality(Colomboet al.,2022),among
77、many other reasons.Figure 1 plots the cumulative multipliers,fromimpact until 2 periods-ahead of the fiscal shock taking place,for different country groups.We start by showing the impulse response functions(IRFs)when we use all the countries17If outliers of forecast errors are not excluded before re
78、gressing them on current and past macroeconomicvariables,they would bias the coefficients of such regression and the residuals would became artificially big.18There is,however,one drawback of using the estimated residuals of this regression as our fiscal shocks.Because these variables will be estima
79、tes,instead of observed variables,there is some uncertainty associatedto them that can impact the final estimates of fiscal multipliers,specially the standard errors of the estimatedfinal coefficients.This concern has been overlooked,however,in the literature that follows a similar procedure.In the
80、appendix we explore this problem and show the results of one specification that tries to overcome thisproblem.13in the sample.Then,we show the same IRFs by income group and focus on two differentfiscal categories:personal income taxes on the revenue side and public investment on theexpenditure side.
81、For AEs and EMs economies,although our estimates for personal income tax multipliersare relatively high,they are in line with what the empirical literature reports,with esti-mates usually around-2 and-3(Ramey,2019,Restrepo,2020).In a similar fashion,publicinvestment multipliers for these two country
82、 groups are also in accordance with the litera-ture:impact multipliers bellow unit but higher afterwards.However,there is one importantdifference between AEs and EMs.Public investment multipliers are close to zero after twoyears for the first group,whereas the cumulative response of output is still
83、increasing aftertwo years in EMs.This suggests that,although the short-run multiplier for public invest-ment in these two groups may be similar,the longer-term effects may differ considerably.AsIzquierdo et al.(2019)point out,the reason for this may hinge on the fact that countrieswith an initial hi
84、gher stock of capital experience lower marginal productivity of additionalunits of capital,and may experience no crowding-in of private investment.That may explainwhy the positive impact in output from public investment in AEs is short-lived.Boehm(2020)finds similar results,estimating that,for OECD
85、economies,government investmentmultipliers are close to zero while government consumption ones are close to one two yearsafter the shock.Unlike the previous cases,the results for LICs are considerably different both in termsof sign and precision.Estimates for personal income tax and public investmen
86、t have anassociated multiplier with the opposite sign one would expect.Still,the relatively highstandard errors imply that estimates are not statistically different from zero for most ofthe horizons considered.In the case of public investment,a weaker institutional frameworkmight signal a less effic
87、ient implementation of public projects(IMF,2015)leading to lower(or even null)multipliers19.However,in the case of the personal income tax,the results are19See also Acemoglu,Johnson and Robinson(2001);Ederveen and Nahuis(2006);Rodrguez-Pose andGarcilazo(2015)or Avelln,Andrade and Len-Daz(2020)for a
88、role of institutions in the size of expenditure14somehow puzzling.Since the same methodology is being followed for all the country groups,this results can be indicative of data quality issues,not properly reflecting the underlyingeconomic developments,which bias the estimates.multipliers.15Figure 1:
89、Multiplier heterogeneity across different groups of countries.Note:C&P multipliers.One-standard deviation confidence bands using robust standard errors,clus-tered at the country level.Shocks to personal income taxes and public investment are expressed as thecumulative response of GDP to a unit incre
90、ase in the respective cumulative fiscal variable,at a givenhorizon.Number of observations included in each estimation(for the h=0),personal income taxes:all sample n=626;AEs n=111;EMs n=341;LICs n=174;public investment:all sample n=2202;AEsn=342;EMs n=1125;LICs n=721.16After analyzing the difference
91、s among country groups,we further explore the differencesin multipliers when focusing on the impact of shocks in different budgetary items20.Figure2 plots cumulative multipliers for the three variables analyzed,using the EMs sample.Be-sides differences on size(i.e larger multipliers,in absolute term
92、s,for personal income tax),the main difference across results seems to be the persistence of responses.While personalincome tax shocks have a considerable contemporaneous impact which becomes virtuallyzero after two years(i.e.:the cumulative effect remains stable),the marginal impact of bothgovernme
93、nt consumption and investment shocks remains positive along the estimated hori-zon.In addition,the estimates for public investment multipliers and government investmentmultipliers are very similar at the horizons considered,although confidence intervals for thefirst are wider.A technical explanation
94、 for these differences lies in the endogenous responseour shocks induce in the fiscal variables,as we explore in section 4.3.In that section weshow that the output response to investment shocks is stronger than the response to pub-lic consumption shocks.However,our public investment shocks are more
95、persistent-thatis,they have a higher impact on public investment also in the years following the shock21.Therefore,by using the integral multiplier definition that we follow,estimates for the twotypes of expenditure become similar.20For the remaining exercises of the paper we will focus only on EMs
96、for the following reasons:1)to ensurethat results are comparable across different exercises,2)because it is a less studied income group relativelyto AEs and 3)due to the statistical quality concerns around LICs data that might bias the estimates,asfigure 1 shows.21One reason why public investment ca
97、n show persistence years after a shock is implemented is due tothe multi-annual nature of many investment projects.In addition,in many cases public contractors fail tocomply with original deadlines and exceed the pre-agreed budget ceilings.17Figure 2:Multiplier heterogeneity across fiscal tools.Note
98、:C&P multipliers.One-standard deviation confidence bands using robust standard errors,clusteredat the country level.Shocks to personal income taxes,public investment and public consumption areexpressed as the cumulative response of GDP to a unit increase in the respective cumulative fiscal variable,
99、at a given horizon.4.2Alternative cleaning procedures of forecast errorsThe identification of fiscal shocks using forecast errors,apart from capturing unexpectedchanges in the fiscal variables,also relies on the assumption that there is a lag associatedto the implementation of fiscal policy measures
100、.In other words,it relies on the quarterlytime restriction(Blanchard and Perotti,2002).That is because,presumably,when using theforecast of a fiscal variable taken from the Octobers WEO of that same year,the realizedfiscal information until September is already incorporated in that forecast.Therefor
101、e,anyforecast error must come from unexpected changes in the fiscal variable happening betweenOctober to December.However,we argue that,although this may be a reasonable assump-tion for the majority of the AEs,macroeconomic statistics for the third quarter of the yearare most likely not yet availabl
102、e for the majority of the EMs and LICs at the time that the18October projections for the year are constructed.In practice,the October WEO projectionsonly incorporate actual data from the first two quarters(or sometimes even one)of the con-temporaneous year,which leads the time restriction assumption
103、 to be unreliable.Therefore,there is still the need to guarantee that the shocks are not responding to contemporaneouseconomic conditions,as we explain in section 3.2.This section shows the importance of guaranteeing that forecast errors are purely ex-ogenous innovations.For this exercise we focus o
104、n fiscal multipliers for personal income inEMs,but using different shocks as the shockfvariable in equation 2.We plot the IRFs whenusing as IVs:i)uncleaned FE,ii)FE filtered from past conditions only,and iii)additionallymaking sure FE are orthogonal to several contemporaneous shocks impacting the ec
105、onomy.Using uncleaned FEs as the IVs leads to multipliers which are almost twice as big as theestimates we get when using our cleaning approach,for every horizon considered,as figure 3shows.In addition,unlike the case of multipliers obtained with FEs filtered only from laggedmacro variables,the marg
106、inal multiplier(i.e.the period difference in cumulative multipliers)is always decreasing if estimated with FEs filtered from contemporaneous shocks.Theseresults suggest that assuming that the forecast errors are unpredictable by past conditionsof the economy,as some studies do,may invalidate the exo
107、geneity assumption and thereforebias the results.19Figure 3:Multipliers using different ways of cleaning forecast errors.Note:C&P multipliers.One-standard deviation confidence bands using robust standard errors,clusteredat the country level.Unfiltered shocks are forecast errors which were not cleane
108、d from any predictablecomponent or any other contemporaneous shock to the economy.Filtered with lagged macro variablesonly cleans the shocks from predictable past variables.Filtered shocks make sure the shocks are orthog-onal to other contemporaneous shocks,as explained in section 3.2.4.3One-stage v
109、s Two-stage EstimationThe dynamic impact of fiscal shocksFiscal policy shocks have a dynamic impact both on GDP and on the fiscal variableitself22.Therefore,estimates that report fiscal multipliers as simply the impact on outputfrom an initial fiscal shock ignore the dynamic endogenous response that
110、 fiscal variables mayalso have to that initial fiscal shock.Ramey and Zubairy(2018)emphasize the importanceof considering this effect by showing that the higher response of output to fiscal shocks22In the long run,authorities have control on the evolution of public consumption and investment.How-eve
111、r,due to implementation lags or political feasibility,for example,some government expenditure items arenot fully controlled by the authorities in the short term.The quintessential example being the quantity ofgovernment employees,which typically shows a substantial degree of inertia.20in periods of
112、economic slack is accompanied also with a stronger response of governmentspending to fiscal shocks in those periods.Therefore,although output response is higher,these findings suggest that the implied fiscal multiplier does not significantly differ accordingto the state of the economy.This is at odd
113、s with the results of Auerbach and Gorodnichenko(2012),which do not take into consideration the dynamic behavior of the fiscal variables andtherefore conclude that multipliers are higher during recessions.This section analyzes how important is to account for this dimension even in a linearframework(
114、that is,without considering state-dependencies).Therefore,we show how thiscan account for considerable differences in the multipliers reported across studies.Figure 4plots the estimated multipliers calculated using the following one-stage regression,for eachhorizon h=0,1,.,H:yi,t+h=i,h+t,h+hshockfi,
115、t+hXi,t1+i,t+h(4)and compares the IRFs to the multipliers obtained using a two-stage IV approach reportedpreviously23.IRFs obtained by equation 4 focus only on the output response to a fiscalshock at time t and are commonly reported by the literature as the estimates for the fiscalmultiplier.For per
116、sonal income taxation,the estimate for the impact multiplier is almost twicehigher when using the two-stage estimation than the results obtained when using the one-stage approach.Similar differences are also reported for the public investment multiplier,with the second-stage approach yielding a 50%h
117、igher multiplier 2 periods after the shockthan the one found when using the one-stage approach.The lower short-term impact onoutput we find when we use the one-stage approach is in line with,for example,estimatesobtained by Furceri and Li(2017)that use a similar methodology to equation 4.Perhaps the
118、 most striking difference in estimates corresponds to public consumption,with23Table 11 in the Appendix reports point estimates and standard errors as a result of directly includingthe vector of variables used to clean FEs in the two-step IV equations 1 and 2.21the multipliers from the one-stage app
119、roach being much lower than the two-stage estimation.This is in line with,for example,lower multiplier estimates obtained by Alichi et al.(2019),Honda,Miyamoto and Taniguchi(2020)or Cacciatore et al.(2021),that estimate multipliersusing similar one-stage procedures24.This shows that results in the l
120、iterature are in lineto what we find if we were using a similar methodology.However,as we argue that theendogenous fiscal movements following a shock should not be ignored,this indicates thatmany studies are missing the important dynamic component that our approach addresses.Ramey and Zubairy(2018)a
121、lso point out that another advantage of using the two-stage IV regression is that it also stresses the importance of analyzing the relevance of theinstruments.Using the one-stage approach usually overlooks this issue.Appendix 6.5 reportsthe F-Statistics of the first-stage regressions.24A similar one
122、-stage approach is follow by IMF(2018).However,their specification accumulates theshocks over the horizon.Although this tries to control for any future shock that may be influencing futureoutput,it still does not capture the endogenous response of fiscal variables.22Figure 4:One vs Two-stage estimat
123、ion.Note:C&P multipliers.One-standard deviation confidence bands using robust standard errors,clusteredat the country level.Left column reports the IRFs when using directly the exogenous shockfon thesecond-stage regression(regression 4),right column reports the multipliers when using the two-stages
124、IVapproach(regression 1).The reason why the multipliers obtained using the one-stage are different from the multi-pliers obtained using the two-stage IV approach is that some fiscal shocks create endogenousmovements on the fiscal variables across the considered horizon and some others do not.Fig-23u
125、re 5 plots the local projections for equation 2,that is the IRF across the considered horizonsof the fiscal variable to the fiscal shock.If multipliers are defined as the integral of the outputresponse divided by the integral of the fiscal variable response to a fiscal shock,the one-stageapproach of
126、 regression 4 would wrongly imply a one-to-one impact of the shock in the fiscalvariable across all horizons.This difference is key,especially for government consumption.As already explained before,this results also highlight why we obtain similar multipliersfor government consumption and government
127、 investment:the output response to investmentshocks is stronger than the response to public consumption shocks,which is a common findingof the literature for EMs.However,our public investment shocks are more persistent-thatis,they have a higher impact on public investment also in years following the
128、 shock,whencompared to public consumption.Hence,when the multiplier definition used is as describedin the previous paragraph,multiplier estimates for the two instruments become similar.Figure 5:Dynamic response of fiscal variable to the fiscal shock:Local projections of firststage regressions.Note:C
129、&P cumulative response of fiscal tools to a shock in the same tool(regression 2).One-standarddeviation confidence bands using robust standard errors,clustered at the country level.A Glance at the Source of FinancingThe financing source of a fiscal policy expansion has been recognized as another impo
130、rtantfactor that may influence the size of multipliers.The neoclassical mechanism suggests that afiscal expansion may have low multipliers because they decrease the intertemporal wealth ofprivate agents,given the increase in future taxes to finance current increase in expenditure(Kraay,2012).However
131、,specially at lower horizons,private agents may not internalize into24their decisions this decrease in their present value wealth given the higher future taxes,whichmakes Ricardian equivalence to fail and multipliers to be higher.Ricardian equivalence maynot hold for other different reasons as well:
132、Liquidity constrained agents that make theirconsumption decisions based on their current income are not impacted by changes in theirpermanent wealth.Alternativelly,myopic agents,which ignore the intertemporal dimensionof their decisions,may also make Ricardian equivalence not to hold(Chodorow-Reich,
133、2019).For this reason,it may or may not matter if the expansion is financed by an increase incontemporaneous revenues or by future revenues(that is,deficit-financed).The literaturealso highlights the higher multipliers that outside-financed fiscal expansions have relative todeficit-financed shocks.I
134、n those cases,it is not that Ricardian equivalence does not hold,but that the local government does not have to raise(present or future)taxes to finance theexpansion25.Another advantage of estimating fiscal multipliers using the two-stage IV approach isthat it allows the researcher to partially addr
135、ess this issue.The local projections for the firststage regression,shown in figure 5,show the dynamic response that a fiscal variable has to ashock in itself.But we can further explore the response of the primary budget balance andanalyze the impact that an expenditure shock has on the dynamics of g
136、overnment finances.By doing so,one can understand how,a posteriori,an expenditure shock at time t is beingfinanced in the following years.To do so,we compute the local projections for the followingregression,for each horizon h=0,1,.,H:bbi,t+h=i,h+t,h+hshockfi,t+hXi,t1+i,t+h(5)where bbi,t+his the cum
137、ulative change on primary budget balance at a given horizon h.Public investment shocks have a negative impact on government budget,as figure 6 shows.This result might suggest that the shocks we are considering are mostly deficit-financed.The reduced-form estimates do not allow us to fully disentangl
138、e the response of the budget25See,for example,Coelho(2019)for the multipliers of federal transfers in eurozone countries.25balance into what is driven by automatic stabilizers or discretionary measures in responseto the expansionary expenditure shock.Still,a deteriorating budget balance indicates th
139、attaxes are not increasing enough to compensate for the increase in expenditure.The samecannot be said about our public consumption shocks,as point estimates are close to zeroand not statistically significant.Therefore,we cannot exclude the hypothesis that our publicconsumption shocks are compensate
140、d by a decrease in some expenditure component or anincrease in revenues(or both).Figure 6:Cumulative response of government budget balance to public investment and publicconsumption shocks.Note:C&P cumulative response of the budget balance to a fiscal shock.One-standard deviation confi-dence bands u
141、sing robust standard errors,clustered at the country level.Forecast errors in different unitsInstead of calculating the forecast errors as a share of GDP(or,more precisely,as thedifference between the realized fiscal variable as a share of GDP and the forecasted value forthe fiscal variables over th
142、e forecasted GDP),some papers calculate the forecast errors ofthe fiscal variable growth rates(Colombo et al.,2022).The drawback of doing so is that theforecast errors are calculated as a percentage of the fiscal variable itself,not as a percentageof GDP26.If the procedure used for estimating the mu
143、ltipliers is the one-stage procedure,26For example,IMF(2017)calculates the forecast errors of growth rates of the fiscal variables.Then,theyconvert these forecast errors into levels using a base year.This new level series is then divided by laggedGDP,so that the variables are expressed as percentage
144、 of the same variable in the right and left hand sideof equation(1).26in which shocks are included directly on the final regression(as explored throughout thissection),then this causes variables on the right and lefts hand side of equation(4)to bein different units and therefore the IRFs cannot be i
145、nterpreted as multipliers.However,ifinstead an IV two-stage procedure is used as we propose,as long as the variables in regression1 are in the same unit,this does not change the final interpretation of the coefficients.We proceed by showing how using the two-stage IV estimation attenuates this probl
146、em.We assess if changing the initial unit in which forecast errors are calculated changes theresults.Instead of calculating forecast errors as in equation 3,we calculate them as thedifference in realized and projected growth rates:FEfi,t=fgri,t fgri,t|t1,(6)where fgri,tis the realized growth rate,an
147、d fgri,t|t1is the projected growth rate.After thisinitial difference in their calculation,we follow the same filtering procedure explained insection 3.2.Table 1 reports the results of using this alternative procedure with the baseline results.We can see some differences in results,especially with es
148、timates for personal income taxes notbeing now statistically significant even when we consider a one standard deviation confidenceinterval.However,the point estimates for this fiscal tool and the other two expendituretools are very close to the estimates obtained using the baseline specification.The
149、 maintakeaway from this exercise is,therefore,that using the two-stage IV approach can limitthe differences in results when identified shocks are not even in the same unit.Using theseshocks directly on the the second stage regression,however,would yield IRFs that should beinterpreted as elasticities
150、 instead of unit multipliers.This is particularly important becausethe methodology used to calculate forecast errors in different papers is not always the same.27Table 1:Multipliers when using FE of growth rates.BaselineFE growth ratesImpactt+2Impactt+2Personal Income Taxes-1.71*-2.47*-1.32-2.68(1.2
151、4)(2.02)(1.61)(4.16)Public Investment0.65*1.68*0.49*1.39*(0.35)(0.73)(0.23)(0.47)Public Consumption0.88*1.91*0.91*1.74*(0.72)(1.19)(0.58)(0.88)Note:C&P multipliers for EME economies.Robust standard errors,clustered at the country level,inparentheses.Stars indicate significance at different confidenc
152、e intervals:*68%,*90%,*95%.4.4Different approach to accumulate responsesEven if one agrees on the definition of cumulative multipliers we use in this paper,thereare still different approaches to achieve it.Variables can be accumulated across each horizonusing the C&P approach,as we have been reporti
153、ng the results.However,Ramey andZubairy(2018)(R&Z)suggest a slightly different approach.They accumulate variables inthe following way27:yi,t+h=hP0Yi,t+hYi,t1and fi,t+h=hP0Fi,t+hYi,t1.On the other hand,Auerbachand Gorodnichenko(2013)(A&G)accumulate the output response as yi,t+h=Yi,t+hYi,t1Yi,t128.How
154、ever,in their paper they use their exogenous fiscal shocks directly in the regression(sincethey calculate fiscal spillovers from one country to the other,and not exactly multipliers ofone country).Since we are using a two-stage IV approach,we also accumulate the fiscalvariable in the same way for th
155、e A&G case,fi,t+h=Fi,t+hFi,t1Yi,t129.Table 2 shows the impact multipliers,as well as one-and two-periods ahead estimates,forthe three different accumulating approaches.We report the estimates for the three variableswe have being using:personal income tax,public investment and public consumption.27In
156、 their paper,the authors normalize the US variables by trend GDP.Since we are working in the contextof a large panel of countries,and focusing on EMEs,trend GDP estimates are less reliable and therefore wenormalize our variables by GDP at t 1.28This is also the most common approach followed in the l
157、iterature.29Notice that in every considered methodology,variables are normalized by output at t 1.As both thecumulative change in output on the right hand side of the equation,and cumulative change on the fiscalvariable on the left hand side are normalized by the same variable,IRFs can be directly i
158、nterpreted as unitmultipliers.28Results for public investment and public consumption multipliers using both the C&Papproach and the R&Z approach are very similar.However,the R&Z methodology yieldshigher cumulative multipliers for personal income taxes.For the multipliers of these twotypes of shocks,
159、we can also see that the A&G approach inflates estimates,specially for theestimates of the multipliers two-periods after the shock occurring.These results reinforce the importance of comparing estimates in a standardized frame-work.The three different approaches,because of the two-stage IV estimatio
160、n,are reportingcumulative dynamic multipliers.Still,minor changes in the way variables are accumulatedacross the horizons leads to different multiplier estimates.Table 2:Different approaches to calculate cumulative multipliers.Impactt+1t+2Personal Income TaxesC&P-1.71*-2.48*-2.47*(1.24)(1.92)(2.02)R
161、&Z-1.83*-3.27*-4.61*(1.39)(2.90)(4.39)A&G-1.71*-4.31*-4.18(1.24)(4.04)(4.92)Public InvestmentC&P0.65*1.09*1.68*(0.35)(0.52)(0.73)R&Z0.71*1.26*1.97*(0.33)(0.49)(0.72)A&G0.65*1.69*3.62*(0.35)(0.74)(1.58)Public ConsumptionC&P0.88*1.48*1.91*(0.72)(0.76)(1.19)R&Z0.77*1.35*1.89*(0.65)(0.71)(1.19)A&G0.88*1
162、.84*3.84(0.72)(0.91)(4.49)Note:Reported results in the table are for EME economies.Robust standard errors,clustered atthe country level,in parentheses.Stars indicate significance at different confidence intervals:*68%,*90%,*95%.C&P accumulates the variables using the Canova and Pappa(2021)approach,R
163、&Zaccumulates the variables using the Ramey and Zubairy(2018)approach,and A&G uses the(slightlymodified)Auerbach and Gorodnichenko(2013)approach.294.5Alternative data sourcesAnother source of noisy estimates may come from the use of alternative data sources,which specially for EME and LIC economies
164、can make a considerable difference.In thissection,we show how the use of data coming from the system of national accounting(SNA)or coming from the government finance statistics(GFS)may influence the estimates.Apart from being compiled from different statistical authorities,the main difference be-twe
165、en the two systems lies on when the transactions are recorded.For most of the EME andLIC economies,the GFS still uses a cash value approach:that is,transactions are recordedat the moment the cash payment related to a certain event is materialized.The SNA,onthe other hand,uses an accrual approach,rec
166、ording the economic transaction at the timethe economic event occurs.Since most of the studies for EME and LIC economies rely onannual data,a difference of a couple of months between the economic event and the cashpayment may imply that the same transaction is recorded in different years when using
167、thetwo different systems.However,the timing of the transaction is not the only source of differences between thetwo approaches.Using cash data also means that some transactions that do not have a cashflow counterpart might not be recorded.For example,donations in kind and debt forgivenessare two exa
168、mples of transactions which would not be recorded in a cash data framework.To show how these small details may affect the results,we estimate cumulative multipliersfor public investment and public consumption,using both SNA and GFS data30.Figure 7plots the different estimates found when using the di
169、fferent data sources.What we show is that the IRFs for the different approaches differ significantly.The impacton output found when using GFS data from a public investment shock is lower compared tothe impact found when using SNA data.Multipliers estimated are not statistically significantacross any
170、 considered horizon when using GFS data,which sharply contrasts whith the results30For SNA data,we use the following variables from the WEO database:NFIG for public investment,NCG for public consumption.For GFS data,for public investment we use GGAANT,which stands forgeneral government net acquisiti
171、on of nonfinancial assets,and for public consumption we sum GGECE andGGEGS,the compensation of employees and purchase of goods and services,respectively.30from SNA data.The same conclusion can be taken for public consumption.When using GFSdata,considerably higher standard errors make estimates not s
172、ignificantly different than zero.Figure 7:SNA vs GFM data.Note:C&P multipliers.One-standard deviation confidence bands using robust standard errors,clusteredat the country level.Shocks to public investment and public consumption are expressed as the cumulativeresponse of GDP to a unit increase in th
173、e respective cumulative fiscal variable,at a given horizon.5Conclusion and Policy ImplicationsThe literature has shown that multipliers are substantially context-specific.They dependon the characteristics of the countries,as well as on state-contingencies.This characteristicof fiscal multipliers mak
174、es different estimates particularly hard to compare.On top of that,most papers that study fiscal multipliers still do not use identical methodologies to calculatethem and consequently report estimates in a non-standardized way.As Ramey(2019)shows,however,standardizing the methodologies to calculate
175、multipliers considerably narrows therange of estimates reported.In this paper,we compare estimates in a uniform framework,across a different set of exercises.We show that small methodological details may lead toconsiderable differences in estimates.Focusing on EMEs,we show that small differences in
176、identifying exogenous fiscal shockscan considerably change results.Even under the same identification strategy that relies oncalculating forecast errors of fiscal variables,different cleaning procedures of these forecasterrors can have an impact in final estimates.Additionally,we show how different
177、reporting31approaches or different data sources also change the results.We emphasize different reasons why the two-stage IV procedure is important for theestimation of fiscal multipliers,and argue against estimating them as simply the response ofoutput to an exogenous fiscal shock.This is particular
178、ly important for policy design.Forpolicymakers,the question that matters is not simply what is the impact on output froma unit increase in fiscal expenditure or revenue.However,a considerable amount of paperswith policy implications are only answering that question,ignoring another dimension offisca
179、l shocks.Instead,policymakers should be interested in all the dynamic endogenousmovements that a fiscal shock triggers in the economy,especially future movements on thefiscal variables themselves.We show that,especially for public consumption multipliers,ignoring this dynamic environment can signifi
180、cantly affect estimates.Accounting for theseendogenous movements is therefore of the upmost importance for proper policy projections.Our paper focuses on the changes in results that different methodological frameworksmay cause.We show that there are crucial details that influence estimates,and asses
181、singthem in an uniform framework should be a prerequisite for proper comparison among studiesand therefore proper policy advise.Naturally,an implication of our argument is that hetero-geneity among estimates caused by state-dependencies or particular country characteristicshould also be assessed in
182、an uniform framework.We leave that for future work.32ReferencesAbiad,Abdul,Davide Furceri,and Petia Topalova.2016.“The macroeconomic effectsof public investment:Evidence from advanced economies.”Journal of Macroeconomics,50:224240.Acemoglu,Daron,Simon Johnson,and James A.Robinson.2001.“The ColonialO
183、rigins of Comparative Development:An Empirical Investigation.”American EconomicReview,91(5):13691401.Alichi,Ali,Mr Ippei Shibata,Kadir Tanyeri,et al.2019.“Fiscal Policy Multipliersin Small States.”IMF Working Paper.An,Zidong,Joao Tovar Jalles,Prakash Loungani,and Ricardo M Sousa.2018.“Do IMF fiscal
184、forecasts add value?”Journal of Forecasting,37(6):650665.Auerbach,Alan J,and Yuriy Gorodnichenko.2012.“Measuring the output responsesto fiscal policy.”American Economic Journal:Economic Policy,4(2):127.Auerbach,Alan J,and Yuriy Gorodnichenko.2013.“Output spillovers from fiscalpolicy.”American Econom
185、ic Review,103(3):14146.Avelln,Leopoldo,Arturo Jos Galindo Andrade,and John Len-Daz.2020.“Therole of institutional quality on the effects of fiscal stimulus.”IDB Working Paper SeriesIDB-WP-01113,Washington,DC:Inter-American Development Bank(IDB).Barnichon,Regis,Davide Debortoli,and Christian Matthes.
186、2022.“Understandingthe size of the government spending multiplier:Its in the sign.”The Review of EconomicStudies,89(1):87117.Batini,Nicoletta,Luc Eyraud,and Anke Weber.2014.A simple method to computefiscal multipliers.International Monetary Fund.33Blanchard,Olivier,and Roberto Perotti.2002.“An empir
187、ical characterization of thedynamic effects of changes in government spending and taxes on output.”the QuarterlyJournal of economics,117(4):13291368.Boehm,Christoph E.2020.“Government consumption and investment:Does the compo-sition of purchases affect the multiplier?”Journal of Monetary Economics,1
188、15:8093.Cacciatore,Matteo,Romain Duval,Davide Furceri,and Aleksandra Zdzienicka.2021.“Fiscal multipliers and job-protection regulation.”European Economic Review,132:103616.Caldara,Dario,andChristopheKamps.2017.“The analytics of SVARs:aunified framework to measure fiscal multipliers.”The Review of Ec
189、onomic Studies,84(3):10151040.Canova,Fabio,and Evi Pappa.2021.“What are the likely macroeconomic effects of theEU Recovery plan?”CEPR Discussion Paper.Carrire-Swallow,Yan,Antonio C David,and Daniel Leigh.2021.“Macroeco-nomic Effects of Fiscal Consolidation in Emerging Economies:New Narrative Evi-den
190、ce from Latin America and the Caribbean.”Journal of Money,Credit and Banking,53(6):13131335.Chodorow-Reich,Gabriel.2019.“Geographic cross-sectional fiscal spending multipliers:What have we learned?”American Economic Journal:Economic Policy,11(2):134.Coelho,Maria.2019.“Fiscal stimulus in a monetary u
191、nion:Evidence from eurozone re-gions.”IMF Economic Review,67(3):573617.Colombo,Emilio,Davide Furceri,Pietro Pizzuto,and Patrizio Tirelli.2022.“FiscalMultipliers and Informality.”IMF Working Paper.34Ederveen,Sjef,and Richard Nahuis.2006.“Fertile Soil for Structural Funds?A PanelData Analysis of the C
192、onditional Effectiveness of European Cohesion Policy.”Kyklos,59(1):17+.17.Furceri,Davide,and Bin Grace Li.2017.The macroeconomic(and distributional)effectsof public investment in developing economies.International Monetary Fund.Hernndez de Cos,Pablo,and Enrique Moral-Benito.2016.“Fiscal multipliers
193、inturbulent times:the case of Spain.”Empirical Economics,50(4):15891625.Honda,Mr Jiro,Hiroaki Miyamoto,and Mina Taniguchi.2020.Exploring the outputeffect of fiscal policy shocks in low income countries.International Monetary Fund.Ilzetzki,Ethan,Enrique G Mendoza,and Carlos A Vgh.2013.“How big(small?
194、)are fiscal multipliers?”Journal of monetary economics,60(2):239254.IMF.2015.“Making Public Investment More Efficient.”IMF Policy Papers,June 2015.IMF.2017.“WEO October 2017:Seeking Sustainable Growth.Chapter 4:Cross-borderimpacts of fiscal policy:Still Relevant?”World Economic Outlook,October 2017.
195、IMF.2018.“Regional Economic Outlook April 2018:Seizing the momentum.Chapter 4:Fiscal Multipliers:how will consolidation affect Latin America and the Caribbean?”Re-gional Economic Outlook,Western Hemisphere,April 2018.Izquierdo,Alejandro,Ruy E Lama,Juan Pablo Medina,Jorge P Puig,DanielRiera-Crichton,
196、Carlos A Vegh,and Guillermo Vuletin.2019.“Is the public in-vestment multiplier higher in developing countries?An empirical investigation.”NationalBureau of Economic Research.Jord,scar.2005.“Estimation and inference of impulse responses by local projections.”American economic review,95(1):161182.35Kr
197、aay,Aart.2012.“How large is the government spending multiplier?Evidence from WorldBank lending.”The Quarterly Journal of Economics,127(2):829887.Miyamoto,Wataru,Thuy Lan Nguyen,and Dmitriy Sergeyev.2018.“Governmentspending multipliers under the zero lower bound:Evidence from Japan.”American Eco-nomi
198、c Journal:Macroeconomics,10(3):24777.Mountford,Andrew,and Harald Uhlig.2009.“What are the effects of fiscal policyshocks?”Journal of applied econometrics,24(6):960992.Nakamura,Emi,and Jon Steinsson.2014.“Fiscal stimulus in a monetary union:Evi-dence from US regions.”American Economic Review,104(3):7
199、5392.Ramey,Valerie A.2016.“Macroeconomic shocks and their propagation.”Handbook ofmacroeconomics,2:71162.Ramey,Valerie A.2019.“Ten years after the financial crisis:What have we learned fromthe renaissance in fiscal research?”Journal of Economic Perspectives,33(2):89114.Ramey,Valerie A,and Sarah Zuba
200、iry.2018.“Government spending multipliers ingood times and in bad:evidence from US historical data.”Journal of political economy,126(2):850901.Restrepo,Jorge.2020.How big are fiscal multipliers in latin america?International Mon-etary Fund.Rodrguez-Pose,Andrs,and Enrique Garcilazo.2015.“Quality of G
201、overnment andthe Returns of Investment:Examining the Impact of Cohesion Expenditure in EuropeanRegions.”Regional Studies,49(8):12741290.Romer,Christina D,and David H Romer.2010.“The macroeconomic effects of taxchanges:estimates based on a new measure of fiscal shocks.”American Economic Review,100(3)
202、:763801.366Appendix6.1Descriptive DataTable 3 shows the classification of each country by income group.On table 4,we showsome of the descriptive statistics for fiscal variables and our calculated fiscal shocks,fro thethree different income groups.Additionally,on table 5 we report how does the sample
203、 sizeevolve on our compiled data set.Figure 8 plots the histogram of the shocks for all sample,figure 9 plots the fiscal shocksfor AEs,figure 10 plots the hocks for EMEs,and figure 11 the shocks for LICs.Shocks arein percentage of GDP.Figure 8:Histogram of fiscal shocks for all sample,including AEs,
204、EMEs and LICs.Note:GGRTII stands for personal income taxes,NFIG for public investment and NCG for publicconsumption.All variables are in real terms.37Figure 9:Histogram of fiscal shocks for AEs.Note:GGRTII stands for personal income taxes,NFIG for public investment and NCG for publicconsumption.All
205、variables are in real terms.Figure 10:Histogram of fiscal shocks for EMEsNote:GGRTII stands for personal income taxes,NFIG for public investment and NCG for publicconsumption.All variables are in real terms.38Figure 11:Histogram of fiscal shocks for LICs.Note:GGRTII stands for personal income taxes,
206、NFIG for public investment and NCG for publicconsumption.All variables are in real terms.On panel 12 we plot the relationship between GDP growth and our estimated exogenousfiscal forecast errors.As it is clear,there seems to be no clear contemporaneous correlationbetween the shocks and GDP growth,in
207、 any of the three cases(AEs,EMEs and LICs).R-squared for each scatter plot,presented in the top left of each graph,is approximatelyzero.39Figure 12:GDP growth and Fiscal Shocks.Note:Variables in percentage points.Fiscal shocks are the calculated forecast errors,cleaned fromendogenous movements in ma
208、cro variables and other forecast errors.Variables trimmed at 2.5%in eachtail of the distribution for these scatter plots.First line for advanced economies,second for emergingeconomies,third for low income economies.40Table 3:Income classification of countries.CountryGroupCountryGroupCountryGroupAfgh
209、anistanLICGermanyAENorth MacedoniaEMEAlbaniaEMEGhanaLICNorwayAEAlgeriaEMEGreeceAEOmanEMEAngolaEMEGrenadaEMEPakistanEMEAnguillaEMEGuatemalaEMEPalauEMEAntigua and BarbudaEMEGuinea-BissauLICPanamaEMEArgentinaEMEGuineaLICPapua New GuineaLICArmeniaEMEGuyanaEMEParaguayEMEArubaEMEHaitiLICPeruEMEAustraliaAE
210、HondurasLICPhilippinesEMEAustriaAEHong Kong SARAEPolandEMEAzerbaijanEMEHungaryEMEPortugalAEBahamasEMEIcelandAEPuerto RicoAEBahrainEMEIndiaEMEQatarEMEBangladeshLICIndonesiaEMERomaniaEMEBarbadosEMEIranEMERussiaEMEBelarusEMEIraqEMERwandaLICBelgiumAEIrelandAESaint LuciaEMEBelizeEMEIsraelAESamoaEMEBeninL
211、ICItalyAESan MarinoAEBhutanLICJamaicaEMESaudi ArabiaEMEBoliviaEMEJapanAESenegalLICBosnia and HerzegovinaEMEJordanEMESerbiaEMEBotswanaEMEKazakhstanEMESeychellesEMEBrazilEMEKenyaLICSierra LeoneLICBrunei DarussalamEMEKiribatiLICSingaporeAEBulgariaEMEKoreaAESlovak RepublicAEBurkina FasoLICKosovoEMESlove
212、niaAEBurundiLICKuwaitEMESolomon IslandsLICCabo VerdeEMEKyrgyz RepublicLICSomaliaLICCambodiaLICLao P.D.R.LICSouth AfricaEMECameroonLICLatviaAESouth SudanLICCanadaAELebanonEMESpainAECentral African RepublicLICLesothoLICSri LankaEMEChadLICLiberiaLICSt.Kitts and NevisEMEChileEMELibyaEMESt.Vincent and th
213、e GrenadinesEMEChinaEMELiechtensteinAESudanLICColombiaEMELithuaniaAESurinameEMEComorosLICLuxembourgAESwedenAECongoLICMacao SARAESwitzerlandAECook IslandsEMEMadagascarLICSyriaEMECosta RicaEMEMalawiLICSo Tom and PrncipeLICCroatiaEMEMalaysiaEMETajikistanLICCubaEMEMaldivesEMETanzaniaLICCyprusAEMaliLICTh
214、ailandEMECzech RepublicAEMaltaAETimor-LesteLICCte dIvoireLICMarshall IslandsEMETogoLICDemocratic Republic of the CongoLICMauritaniaLICTokelauEMEDenmarkAEMauritiusEMETongaEMEDjiboutiLICMexicoEMETrinidad and TobagoEMEDominicaEMEMicronesiaEMETunisiaEMEDominican RepublicEMEMoldovaLICTurkeyEMEEcuadorEMEM
215、ongoliaEMETurkmenistanEMEEgyptEMEMontenegro,Rep.ofEMETuvaluEMEEl SalvadorEMEMontserratEMEUgandaLICEquatorial GuineaEMEMoroccoEMEUkraineEMEEritreaLICMozambiqueLICUnited Arab EmiratesEMEEstoniaAEMyanmarLICUnited KingdomAEEswatiniEMENamibiaEMEUnited StatesAEEthiopiaLICNauruEMEUruguayEMEFijiEMENepalLICU
216、zbekistanLICFinlandAENetherlandsAEVanuatuEMEFranceAENew ZealandAEVenezuelaEMEGabonEMENicaraguaLICVietnamLICGambia,TheLICNigerLICWest Bank and GazaLICGeorgiaEMENigeriaLICYemenLICNote:The panel of countries used in the local projections is as unbalanced panel.Whenever available,weuse the maximum numbe
217、r of countries to which we have data to,in a given year.41Table 4:Descriptive Statistics.AEsEMEsLICsGDP growthMean2.75%3.97%4.35%Std.Deviation3.64%7.43%5.27%PITMean0.20%0.11%0.11%Std.Deviation0.47%0.42%0.39%Public InvestmentMean0.09%0.26%0.31%Std.Deviation0.5%2.00%2.92%Public ConsumptionMean0.47%0.5
218、7%0.61%Std.Deviation0.67%2.14%3.38%Shock PITMean0.00%0.00%0.00%Std.Deviation0.32%0.24%0.31%Shock IgMean0.00%0.00%0.00%Std.Deviation0.67%1.69%2.01%Shock GMean0.00%0.00%0.00%Std.Deviation0.98%1.98%2.37%Note:Sample between 1995-2019(for PIT shocks,between 2010-2019).PIT,public investment andpublic cons
219、umption are presented as growth in percentage of GDP.Shock variables are the cleanedforecast errors of the respective fiscal variable,calculated as explained in the text.42Table 5:Sample size:number of countries per year.Shock PITShock IgShock GYear1995-1996--971-11712
220、12001-2002--1181-1171-920000000001987153175Note:Shocks are already the cleaned forecast errors,as a percentage of previous year GDP.436
221、.2Summary of Main EstimationsTable 6:Summary of Results.Impactt+1t+2Personal Income TaxesAll-0.45-0.38-1.35(1.09)(1.59)(2.39)AE-1.70*-1.27-2.11(1.32)(1.41)(2.37)EME-1.71*-2.49*-2.47*(1.24)(1.93)(2.02)LIC1.814.55*2.48(1.90)(3.99)(7.59)Public InvestmentAll0.34*0.69*1.14*(0.28)(0.45)(0.66)AE0.71*1.28*0
222、.28(0.55)(0.74)(0.81)EME0.65*1.09*1.68*(0.35)(0.52)(0.73)LIC-0.44-0.280.00(0.49)(0.53)(0.68)Public ConsumptionAll0.96*1.46*1.66(0.74)(0.95)(1.82)AE-0.41-1.57-4.32(1.28)(2.02)(4.38)EME0.88*1.48*1.91*(0.72)(0.76)(1.19)LIC0.39-0.364.18(2.08)(6.18)(11.42)Note:C&P multipliers.Robust standard errors,clust
223、ered at the country level,in parentheses.Starsindicate significance at different confidence intervals:*68%,*90%,*95%.6.3Different SpecificationsTable 7 shows the multipliers when we include Covid years in the exercises.Some differ-ences in estimates are worth mentioning:perhaps because the personal
224、income taxes exerciseuses less years than the exercises that calculate expenditure side multipliers,including theseatypical years has a higher impact on tax multipliers and bias the results.Additionally,44the public consumption impact multiplier also increases when we include this years:thestrong fi
225、scal response in some countries,higher fiscal forecast errors together with abnormalmovements in output,as well as fast recoveries,may have influence the estimate.Table 7:Local projections including Covid years in the estimation.Impactt+1t+2Personal Income Taxes1.642.72-7.69(3.74)(4.11)(8.23)Public
226、Investment0.66*0.89*1.23*(0.36)(0.51)(0.74)Public Consumption1.53*1.82*1.19(0.60)(0.64)(1.34)Note:C&P multipliers for EME economies.Robust standard errors,clustered at the country level,inparentheses.Stars indicate significance at different confidence intervals:*68%,*90%,*95%.On table 8 we show the
227、change in results for including only country fixed effects or timefixed effects separately.And table 9 shows results for different lag structures.Table 8:Different fixed-effect specifications.BaselineCountryTimeImpactt+2Impactt+2Impactt+2Personal Income Taxes-1.71*-2.47*-1.98*-3.45*-1.84*-2.94(1.24)
228、(2.02)(1.29)(2.5)(1.39)(4.29)Public Investment0.65*1.68*0.58*1.48*0.52*1.11*(0.35)(0.73)(0.40)(0.89)(0.3)(0.74)Public Consumption0.88*1.91*0.96*2.15*0.85*1.83*(0.72)(1.19)(0.79)(1.26)(0.74)(1.58)Note:C&P multipliers for EME economies.Robust standard errors,clustered at the country level,inparenthese
229、s.Stars indicate significance at different confidence intervals:*68%,*90%,*95%.Baselineuses both country and time fixed effects.Country specification uses only country fixed effects,Timespecification uses only time fixed effects.Table 10 reports the standard errors when clustered at the year-level,t
230、o account for thepossibility of cross-sectional dependence,and compares with the standard errors clustered atthe country-level.We see that standard errors remain practically unchanged.45Table 9:Different lag specifications.BaselineNo dynamic panelNo lagsTwo lagsImpactt+2Impactt+2Impactt+2Impactt+2PI
231、T-1.71*-2.47*-1.61*-2.26*-1.62*-2.18*-1.59*-2.49*(1.24)(2.02)(1.26)(1.89)(1.27)(1.86)(1.22)(2.08)Ig0.65*1.68*0.68*1.74*0.84*2.25*0.60*1.45*(0.35)(0.73)(0.39)(0.75)(0.39)(0.71)(0.34)(0.59)G0.88*1.91*0.95*2.20*0.98*2.32*0.97*2.05*(0.72)(1.19)(0.82)(1.23)(0.85)(1.39)(0.73)(1.21)Note:C&P multipliers for
232、 EME economies.Robust standard errors,clustered at the country level,inparentheses.Stars indicate significance at different confidence intervals:*68%,*90%,*95%.Baselinespecification includes one lag of yi,tand one lag of fi,t.No dynamic panel specification excludes thelagged yi,tcomponent,so that th
233、e panel model is no longer a dynamic one,No lags specification usesno lags of these variables,Two lags specification uses two lags of these two variables.All specificationsinclude one lag shockfi,tas a control.Table 10:Different clustering levels of Standard Errors.PITPublic InvestmentPublic Consump
234、tionImpactt+2Impactt+2Impactt+2Estimate-1.71-2.470.651.680.881.91Clustered Country-level(1.24)(2.02)(0.35)(0.73)(0.72)(1.19)Clustered Year-level(1.32)(2.21)(0.36)(0.69)(0.78)(1.67)Note:C&P multipliers for EME economies.Robust standard errors,clustered at the country-level andat the year-level,in par
235、entheses.6.4Uncertainty surrounding Forecast ErrorsFocusing on EMEs,in table 11 we report the point estimates and standard errors if thevector of variables used to clean the FEs,as explain in section 3.2,is directly included inequations 1 and 2 as controls.In this case unfiltered forecast errors are
236、 used as fiscal shocks.This methodological change has,as expected,two main implications.First,it shows therobustness of baseline point estimates reported in table 2 from sub-section 4.4.Second,itincreases standard errors,as the uncertainty surrounding the cleaning procedure of FEs isnow directly inc
237、orporated.In any case,the final impact in terms of inference can only be46assessed when incorporating the main drivers of multiplier heterogeneity.Table 11:Estimated vs not estimated fiscal shock.Filtered FEUnfiltered FE w/controlsImpactt+2Impactt+2Personal Income Taxes-1.71*-2.47*-1.10-0.55(1.24)(2
238、.02)(1.32)(2.27)Public Investment0.65*1.68*0.49*1.66*(0.35)(0.73)(0.37)(0.87)Public Consumption0.88*1.91*0.201.08(0.72)(1.19)(0.79)(1.54)Note:C&P multipliers for EME economies.Robust standard errors,clustered at the country level,inparentheses.Stars indicate significance at different confidence inte
239、rvals:*68%,*90%,*95%.FilteredFE are the fiscal shocks used throughout the paper(see section 3.2).Unfiltered FE with controls rep-resents the regression in which unfiltered forecast errors are used as IVs,but we add as controls in theregression the same variables used to filter the FE in the baseline
240、 exercises:the FE of GDP growth,FEof exchange rate,and FE of inflation,as well as lagged GDP growth,lagged government expenditure andrevenues growth as percentage of GDP,lagged growth in the exchange rate and lagged inflation rate.476.5Instrument RelevanceTable 12 shows the F-statistic of the first-
241、stage regressions of our baseline exercises,foreach of the three different ways of accumulating responses.Table 12:F-statistic of first stage regression for instrument relevance.Impactt+1t+2Personal Income TaxesC&P5.425.124.36R&Z160.4582.1151.53A&G5.423.681.52Public InvestmentC&P20.8810.1226.33R&Z10
242、5.2535.0020.61A&G20.8831.3244.43Public ConsumptionC&P5.795.274.99R&Z61.7242.9337.46A&G5.795.094.25Note:First-stage F-statistics across different horizons and ways of accumulating multipliers.EMEeconomies sample.48Getting into the Nitty-Gritty of Fiscal Multipliers:Small Details,Big Impacts Working Paper No.WP/2023/029