How Effective is Monetary Policy at the Zero Lower Bound? Identification Through Industry Heterogeneity

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. How Effective is Monetary Policy at the Zero Lower Bound? Identification Through Industry Heterogeneity Arsenios Skaperdas 2017-073 Please cite this paper as: Skaperdas, Arsenios (2017). “How Effective is Monetary Policy at the Zero Lower Bound? Identification Through Industry Heterogeneity,” Finance and Economics Discussion Series 2017-073. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2017.073. 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.

How Effective is Monetary Policy at the Zero Lower Bound? Identification Through Industry Heterogeneity Arsenios Skaperdas†‡ July 3, 2017 First Draft: December 11, 2014 Abstract US monetary policy was constrained from 2008 to 2015 by the zero lower bound, during which the Federal Reserve would likely have lowered the federal funds rate further if it were able to. This paper uses industry-level data to examine how growth was affected. Despite the zero bound constraint, industries historically more sensitive to interest rates, such as construction, performed relatively well in comparison to industries not typically affected by monetary policy. Further evidence suggests that unconventional policy lowered the effective stance of policy below zero. JEL Classification: E32, E43, E47, E52 †Board of Governors of the Federal Reserve System. Email: arsenios.skaperdas@frb.gov. ‡I am thankful for the guidance of my dissertation advisors, Michael Hutchison and Carl Walsh, and themembersofmycommittee, DavidDraperandKenKletzer. IamalsogratefultoJohannaFrancisfor continued feedback and encouragement at an early stage. Michael Bergman, Valentina Colombo, Carlos Dobkin, Peter Dunne, Gabriel Fagan, Benjamin Garcia, Stefan Gerlach, Matija Lorenz, Jeffrey Huther, Fergal McCann, Laura Moretti, Gerard O’Reilly, Jonathan Robinson, Ajay Shenoy, Evan Smith, Eric T. Swanson, UCSC Macro, Central Bank of Ireland, Workshop on Empirical Monetary Economics, and Fed Boardseminarparticipantsprovidedhelpfulcomments. Finally, IthanktheMonetaryPolicyDivisionat the Central Bank of Ireland for an excellent environment to work on revising the paper during April to September 2015. The views expressed in this paper are solely my own and should not be interpreted as reflectingtheviewsoftheBoardofGovernorsoftheFederalReserveSystemorofanyoneelseassociated with the Federal Reserve System. 1

1 Introduction The federal funds rate reached essentially zero in December 2008, at which point the Federal Reserve could no longer provide further monetary accommodation through its traditional policy instrument. Following that time, the unemployment rate continued to rise, while inflation fell below target. The Federal Reserve would thus likely have lowered the federal funds rate further if it were able to. In order to continue influencing economic activity, the Federal Reserve turned to nontraditionalpolicies. Amongthese,Large-ScaleAssetPurchaseprograms(LSAPs)involved the purchases of assets while those purchases could no longer affect short-term interest rates. Instead, these programs were meant to lower longer-term interest rates, and influence important yields affecting corporate bond and household mortgage rates. However, these programs were motivated by theoretical channels not believed to be particularly strong before the crisis (Woodford, 2012; Walsh, 2014). This paper examines growth during this unforeseen period of low interest rates and unconventional monetary policy. Given the scale of the Great Recession, estimates of previous behavior of the Federal Reserve indicate that the federal funds rate would likely have been decreased to -5 percent (Rudebusch, 2009). This suggests that a large amount of monetary accommodation- which had been provided in previous recoveries- did not occur, unless unconventional policy had significant affects. Accordingly, one would expect the recovery from the Great Recession to be weaker than previous recoveries. In that case, it would also necessarily be true that industry growth rates would be negatively affected. The contribution of this paper is thus to ask the following question: did industries typically affected by monetary policy suffer during the recovery from the Great Recession? It is commonly thought that some industries are much more interest rate sensitive than others. Using monetary vector autoregressions, I confirm that industries involved in longer-term projects, or that require large consumer loans, such as construction and durable goods manufacturing, respond more strongly to interest rate shocks (Dedola and Lippi, 2005). I validate the industry ranking through an alternative measure of monetary shocksfromRomerandRomer(2004). Ifunconventionalpolicyhadlittleeffect, weshould expect to see that these industries, such as construction, performed poorly in comparison to industries not thought to be typically affected by monetary policy. In contrast, I find that interest rate sensitive industries performed just as well- if not more so- than in previous US recoveries. As firms also grew about as well on average as in previous recoveries, it must be the case that either unconventional policy had significant 2

real effects, or that some other shocks on net caused heterogeneous industry effects as an expansionary monetary shock would have. However, I find no evidence for any other such shocks, including plausible alternative explanations such as expansionary fiscal policy or industry-specific productivity growth. In order to further answer whether unconventional monetary policy had real effects, I present a new time series method for measuring monetary accommodation. Departing from previous literature (Krippner, 2013; Wu and Xia, 2014; Lombardi and Zhu, 2014; Christensen and Rudebusch, 2013), I develop an approach that infers an implied monetary stance directly from observed industry growth rates and standard variables used in monetary vector autoregressions. I find that that the zero lower bound period is characterized by “below zero” properties, in that observed outcomes occurred as if further federal funds rate decreases had taken place. This provides evidence that unconventional monetary policy was effective, and accounts for the observed positive relative growth rates of interest rate sensitive industries. Overall, the findings in this paper mirror other results in the literature (Swanson and Williams, 2014; Gilchrist et al., 2014) which suggest that the Federal Reserve had leewaytoinfluencetheeconomythroughunconventionalpolicy. Giventhepositiverelative growth rates of interest rate sensitive industries, and measured accommodative monetary shocks following the incidence of the zero bound, I infer suggestive evidence that the combination of zero interest rates and unconventional policy increased economic activity in a comparable fashion to previous US recoveries for the public firms examined. The layout of the paper is as follows: Section 2 establishes how this paper relates to theliteratureonthezerolowerbound. Sections3, 4, and5explainthedata, methodology, and results of the growth rates of interest rate sensitive industries through time. Section 6 discussesalternativeexplanationsforthefindings. Section7presentsexploreswhetherthe results are attributable to unconventional monetary policy. Finally, Section 8 concludes. 2 Related Literature The identification strategies of this paper rely in part on industry heterogeneity of the effects of monetary policy. A variety of strands of literature have shown heterogeneous effects. Barth and Ramey (2002) find that heterogeneous industry effects can arise from differing degrees of working capital usage through the cost channel, whereby firms’ marginal costsaredirectlyaffectedbychangesinnominalinterestrates(RavennaandWalsh,2006). Gertler and Gilchrist (1994) find that small manufacturing firms are more responsive to 3

monetary policy shocks than large firms, and find evidence that this is at least partly due to the credit channel. Ehrmann and Fratzscher (2004) and Maio (2013) examine heterogeneous stock market movements in response to monetary shocks. Finally, structural VAR evidence for heterogeneous industry effects can be found in Dedola and Lippi (2005) and Peersman and Smets (2005). The approach in this paper builds on that of Dedola and Lippi (2005) in the basic specification of a VAR to identify how sectors respond to interest rate shocks. Thispaper’sresultsareparticularlyrelevantinlightofmacroresearchthathasimplied that the effective short-term interest rate following 2008 was negative. Krippner (2013) and Wu and Xia (2014) use term structure models to calculate “shadow rates” of the stanceofmonetarypolicy, byevaluatingwhatnominalinterestratewouldnormallycreate observed yield curves. These papers attempt to measure the short-term interest rate implied by financial markets. In contrast, Section 7 estimates the effective federal funds rate as implied by real macroeconomic variables and industry growth rates. Finally, this paper follows Claessens et al. (2011) and in the use of firm-level data to answer macro questions. Claessens et al. (2011) examine the diffusion of the Global Financial Crisis on firms from a large sample of countries using differences-in-differences from the start of the crisis to the trough. In contrast, this paper examines economic recoveries, and uses previous recoveries to benchmark what we should quantitatively expect for firm growth. For the sectoral (industry) interest rate sensitivity index, I draw inspiration from Rajan and Zingales (1998) in capturing time-invariant intrinsic sector characteristics. Theanalysisinthispaperhasbegunfromthepremisethatinterestratedecreaseshave non-negligible effects on output. If this is not the case, this paper’s results provide no meaningful evidence that monetary policy is effective at the zero lower bound. However, evidence for nominal rigidities is unambiguous (Steinsson and Nakamura, 2014; Gorodnichenko and Weber, 2014). I thus proceed under the presumption that these rigidities do result in real effects. 3 Data 3.1 Data Sources I use Compustat data for access to a firm-level database with relevant firm covariates. This database contains quarterly and yearly information on the universe of publicly listed 4

companies in the US. As such, all results are local to public firms. In the conclusion, I discuss how representative is this sample of firms as a whole. For ranking industry interest rate sensitivity, I use the Compustat Quarterly North America database and the 1970:Q1-2008:Q3 time period. This sample is chosen such that it precedes the crisis period and the incidence of the recovery at the zero lower bound. I then use the resulting ranking to predict revenue growth using Compustat annual data. All US macro data is from FRED. Fiscal variables are taken from the Bureau of Economic Analysis. 3.2 Industry Time Series Construction To my knowledge, no consistent time series of industry growth exists.1 I thus construct time series using Compustat data. The timing of firm entry and exit into the Compustat database is an endogenous outcome of many variables. Regulatory changes, investor sentiment, and other factors could make it more appealing for firms to go public or merge during certain time periods. When examining revenues, this would erroneously make it appear during times of increased corporate entry that an industry had grown (in aggregates) or shrunk (in averages, assuming entering firms are small). In fact, mean industry revenues do not follow any clear growth trend. I thus examine growth rates of revenues rather than aggregates or averages.2 I follow the micro employment literature (Davis et al., 1998) in specifying firm-level growth rates as y −y i,t i,t−k g = (1) i,j,t−k,t 0.5[y +y ] i,t i,t−k Where y is the variable of interest of firm i, in industry j, in quarter t. This bounds the i,t measure to a [−2,2] interval and approximates the effect of log differences. This growth rate measure accounts for firm entry and exit symmetrically, corrects for the effects of outliers, and helps control for mergers. Finally, I use year-over-year growth rates to avoid deseasonalizing. This ignores firms that exist for less than four quarters, provides a less volatile metric of firm growth than quarter over quarter measures, and allows for firms that only report revenues once a year over parts of the sample. The growth rate in quarter t of industry j is thus constructed as the industry average of (1) 1TheBureauofEconomicAnalysisdoescollectindustryvalueadded, amongothervariables,buthas no industry data spanning a time period long enough for a VAR analysis due to reclassification. 2I cannot calculate industry value added, as the Compustat database has only sparse data on wages and salaries. 5

1 nj,t g = X gi (2) j,t n t−4,t j,t i=1 Where n is the number of firms in industry j at time t with year-over-year revenue j,t statistics. 4 Methodology This section discusses the methodology of the first approach I use to investigate monetary policy at the zero lower bound. The approach involves two steps: estimating an industry interest rate sensitivity index, then using the index in a second-stage regression which evaluates industry revenue growth at the zero lower bound. This section explains the methodology, while Section 5 presents the results from the two steps. 4.1 Industry Interest Rate Sensitivity Index Construction The interest rate sensitivity index is created using industry-level impulse responses from structural vector autoregressions. This approach makes a minimum number of assumptions on the data, identifying the effect of interest rate changes that cannot be predicted by contemporaneous and lagged values of other variables in the model.3 I use the SVARs to measure to what extent an unpredictable interest rate increase is associated with future decreases in revenues for each industry. The ranking is highly correlated across SVAR lag lengths, recursive orderings, and a variety of other specification choices, including subsample analysis where the VARs are estimated beginning in the 1980s.4 In order to measure the interest rate sensitivity index, I employ the following SVAR model for each two digit NAIC industry: B Y = C +B Y +u (3) 0 t 1 t−1 t The SVAR is identified recursively, with the exception that industries are small relative to the macroeconomic aggregates, and cannot affect them. I impose constraints on parame- 3An alternative approach to isolate industry level interest rate sensitivity, which requires market efficiency, would be to examine the industry heterogeneity of stock returns due to unanticipated federal funds rate changes. This would require differentiation of stock movements into real interest rate news, dividend news, and future excess return news. Unfortunately, this approach yields statistically imprecise estimates of industry heterogeneity, as found in Bernanke and Kuttner (2005). 4These robustness checks are presented in the online appendix. 6

ters such that industries do not affect any of the macroeconomic variables.5 Empirically, quarterly corporate industry revenues are a relatively small portion of gross revenues, justifying this structure.6 Structural shocks are recovered through the following contemporaneous restrictions:      1 0 0 0 0 (cid:15) u Output,t Output,t a 1 0 0 0(cid:15)  u   2,1  IndustryRev,t  IndustryRev,t a 0 1 0 0(cid:15)  = u   3,1  Commodities,t  Commodities,t a 0 a 1 0 (cid:15)   u   4,1 4,3  Inflation,t   Inflation,t  a 0 a a 1 (cid:15) u 5,1 5,3 5,4 Fedfunds,t Fedfunds,t As shown, the Y vector is composed of and ordered in the following manner: real t GDP growth, industry revenue growth, commodity price inflation,7 personal consumption expendituresinflation,8 andthefederalfundsrate.9 TheC vectorcontainsonlyconstants; no other exogenous terms are needed as all variables other than the federal funds rate are year-over-year growth rates, in order to match with time series of industry revenues. A lag length of 4 is chosen as suggested by the AIC. Three industries, Agriculture, Forestry, Fishing and Hunting (11), Transportation and Warehousing (49), and Educational Services (61), have significantly fewer quarterly firm observations in the sample.10 Since these industries each have such a small number of firms, VAR results using these industry aggregates are more uncertain due to the larger effects of idiosyncratic firm-level shocks. The series for these industries are characterized by much higher variability than the other industries, and as Table 1 shows, NAICS (11), (49), and (61) have among the largest standard deviations of industry revenues. I deal with this problem by dropping these industries from the sample. This results in dropping all industries with fewer than 24 firms per quarter on average, which comprise 0.09% of firm observations used in the second-stage analysis.11 I quantify the SVAR revenue effects of monetary policy by recording the minimum 5Thus, the parameters characterizing macroeconomic aggregates are by construction equal to the parameters resulting from a standard recursive VAR. Relaxing VAR specifications to allow industry outcomestoaffectmacroeconomicaggregatesyieldsverysimilarindustryimpulseresponsestoamonetary shock, as the Federal Reserve is not measured to react to individual industry oucomes. 6This figure is shown in the online appendix. 7Series PPIACO from FRED. 8TheistheprimarydesignatedinflationrateusedbytheFederalReserve,seriesPCEPIfromFRED. 9The identification scheme used is standard in the literature for VARs analyzing monetary policy with quarterly data (Christiano et al., 1999). 10A chart of quarterly firm observations by industry is shown in the online appendix, as is a plot of the constructed industry time series. 11Intheonlineappendix,itisshownthatincludingtheseindustriesdoesnotchangeanyoftheresults. 7

point estimate12 of the orthogonalized industry revenue response, over a four year horizon, to a positive one standard deviation interest rate shock from each SVAR in system (3). To ease interpretation of the index in predicting firm recovery, I standardize the orthogonalized impulse responses with respect to each other and multiply by −1. In second-stage regressions used to investigate the zero lower bound, this makes it such that a positive value on the index, denoted by φ , is associated with positive growth. Each j firm i’s index sensitivity is based on its industry’s standardized orthogonalized impulse response. A one unit increase in the index thus signifies a one standard deviation increase in interest rate sensitivity.13 (cid:16) (cid:17) Formally, let χ = min16 ∂gj,t , where g is the revenue growth rate at quarterly j t=0 ∂u5t j,t horizon t for industry j, and u is a one standard deviation orthogonalized shock to the 5t federal funds rate. Then each firm in industry j has its interest sensitivity assigned to be (cid:18)χ −χ¯ (cid:19) j j φ = −1· (4) j σ where σ denotes the standard deviation of χ . j 4.2 Second-Stage Analysis I first use the industry ranking to trace out the time-varying evolution of the relative growth rates of industries identified as interest rate sensitive. This will reveal that interest rate sensitive industries have not suffered at the zero bound relative to historical norms. I then use the standardized sensitivity index to predict real revenue growth, as defined in equation (1), in the following more rigorous second-stage specification: g = ψ ·(φ I )+δφ +γ I +βX ·I +(cid:15) (5) i,j,t−k,t j 2009,2012 j t−k,t t−k,t i,j,t−k,t t−k,t i,j,t−k,t Whereg istherevenuegrowthratebetweenyearst−k andtoffirmi, inindustry i,j,t−k,t j. I isanindicatorforeachofthethreerecoveries, whileγ captureseachrecovery’s t−k,t t−k,t trend growth. The coefficient ψ on the interaction term φ I captures the difference j 2009,2012 in real revenue growth for firms in sensitive industries relative to insensitive industries, 12Industries with positive responses to monetary shocks are thus set to zero before standardization. Alternative specifications of the index are presented in the online appendix and yield similar results. 13The literature on monetary policy has noted that the effects of monetary shocks are asymmetric depending on the phase of the business cycle (Weise, 1999; Chen, 2007; Tenreyro and Thwaites, 2013). Forexample,acontractionaryshockduringanexpansionmayhavedifferenteffectsthanacontractionary shock during a recession. For the purposes of this paper, I examine the symmetric case, as is the case in the majority of the VAR literature. 8

at the zero lower bound, in comparison to the previous two recoveries, captured by δ. In baseline specifications, I examine recoveries using the timing of 1991-1994, 2001-2004, and 2009-2012 for the recovery years, as these correspond to periods after which the Fed eased policy following a recession.14 The vector X consists of a number of control i,j,t−k,t variables, to be explained, allowed to have different effects in each recovery.15 Equation 5 is thus similar to running three separate OLS regressions over the periods 1991-1994. 2001-2004, and 2009-2012, with the difference being that the correlation between industry revenue growth and industry interest rate sensitivity is constrained to be the same over 1991-1994 and 2001-2004 recoveries (δ), while I measure the difference in this correlation over the 2009-2012 ZLB period (ψ). In all specifications, ψ captures the difference in revenues growth for firms in highly interest rate sensitive industries relative to firms in low sensitivity industries. If the zero bound has been a severely binding constraint on monetary policy, and hence aggregate growth, it must be the case that industry growth rates have been negatively affected. We should then expect to see that ψ is significantly negative. Table 11 presents summary statistics on relevant firm variables used in this paper. It is of interest to note that the 2009-2012 recovery is not characterized by lower revenue or employment growth than previous recoveries. In fact, the 2009-2012 period has greater employment and revenue growth than the 2001-2004 recovery. 5 Empirical Results 5.1 SVAR Index Results I first investigate the industry interest rate sensitivity classification. Figure 1 graphs the orthogonalized industry impulse responses to an interest rate shock upon which the standardized index is created.16 Table 2 presents the minimum response values to a one 14The year of each recovery start is set according to NBER dates. These most recent recoveries are examined as they provide a closer comparison to the zero lower bound recovery rather than pre-Great Moderation. Examining other year choices for the recoveries give similar results. 15Note that the control variables are thus allowed to be time-varying. This is important because, for example, industry GDP elasticity will have different effects depending on the strength of GDP growth, which varies across the three recoveries examined. Likewise, log firm employment at the start of each recovery may have different effects in 1991 than 2009, as firms on average are larger in 2009. 16In all figures, 90% standard error bands are calculated using a circular (moving block) bootstrap. Steps are as follows: an artificial time series is created, equal in length to the sample, using random samples of block length 4 drawn with replacement from a circle of the SVAR residuals, initiated with 1971:Q1-1971:Q4values. TheSVARmodelsareestimatedovertheartificialtimeseries,yieldingdistribu- 9

standard deviation shock, along with the standardized reciprocal explained previously. The ordering of industries aligns with common priors. Broadly speaking, industries classified as more interest rate sensitive are capital intensive and produce goods of greater durability, as found in previous literature (Dedola and Lippi, 2005).17 According to the responses, Construction (23)18 and Mining, Quarrying, and Oil and Gas Extraction (21)19 are the most sensitive sectors. These are industries that are highly affected by investment funding levels. The most responsive manufacturing sector, NAIC 33, is composed mostly of durable goods (as in Erceg and Levin (2006)), which are intertemporally substitutable. Overall, the most sensitive sectors consist of firms that either undertakelong-termprojectssensitivetothecostofcapital, orproducegoodsthatrequire large consumer loans. Service industries including Health Care and Social Assistance (62), Professional, Scientific, and Technical Services (54), and Other Services (81), are less affected by interest shocks. These industries do not rely on inventory or production chains for revenues, making them less affected by the credit channel. Arts, Entertainment, and Recreation (71), though a service industry, may be more affected by decreases in disposable income. The finding that the Retail Trade sectors (44 and 45) are not more affected may be surprising. However, it is important to note that uncertainty over the impulses responses implies that many of the industries with intermediate responses to interest rate shocks are not statistically differentiable from each other. In addition, it is possible that exchange rate changes following interest rate changes mitigate the net effect on retail. That utilities are somewhat affected by interest rates may be counterintuitive. In fact, utility companies have high debt loads,20 and many utility companies are subject to regulation which can cause decreases in their ability to adjust to interest rate changes. It is important to mention that the ranking of industries by magnitude to an interest rate shock: construction, durable goods, non-durable goods, and the services sector, is mirroredbyimpulseresponseresultsfromaVARusingthesedisaggregatedcomponentsof tions of impulse responses. This process is repeated 200 times. The standard error bands are calculated using the 5th and 95th empirical quantiles of the impulse responses at each horizon. 17These authors also find evidence that sensitive industries have higher degrees of working capital usage and external financial dependence 18EvidencefromtheBureauofEconomicAnalysisnationalaccountsalsoshowsthatresidentialinvestment is the most affected of all components of GDP by VAR interest rate shocks; as found by Bernanke and Gertler (1995). 19The oil and natural gas industries have similar rates of growth to other industries in NAIC 21, such as coal mining, during the zero lower bound period. 20Of all industries, utility companies have the largest firm-year average of long term debt expense as a percent of total assets in sample. 10

GDP. Thus, the industry impulse responses using Compustat firms are similar to impulse responses of aggregates from the national accounts. The VAR impulse responses to an interest rate shock are shown in Figure 2. Although prices do rise initially in response to an interest rate shock, this increase is small and statistically insignificant.21 5.2 Second-stage Correlations It is valuable to present the data in a graphical manner. Figure 3 charts coefficients on interest rate sensitivity from three year industry-level growth rates regressed on year indicators and year indicator-interest rate sensitivity index interactions. This charts the equivalent of the coefficient estimates on interest rate sensitivity (φ ) in an OLS regression j over each three year window of revenue growth in the following regression.22 g = α+βφ +(cid:15) (6) j,t−k,t j j,t−k,t The coefficients trace the time-varying association between interest rate sensitivity and relative industry growth rates. As shown, the post-2009 period has a similar trajectory to the 2001 recovery. This is a major finding of this paper. If the zero lower bound has severely constrained monetary policy, we should see that interest rate sensitive industries have been detrimentally affected. This is not the case. Another interesting aspect of Figure 3 is that the pre-crisis years in the 2000s are characterized by large increases in revenues of interest rate sensitive industries. Expost, we now know that this time period was characterized by over-exuberant growth in industries sensitive to loosening credit standards. These are also the industries, such as construction, which would be typically affected by monetary policy.23 Figure 3 is reassuring in the sense that it confirms the relevance of the interest rate sensitivity index with priors about how industries affected by monetary policy evolved before the crisis, though it does not prove that monetary policy was the cause of the buildup. We would expect to see growth in interest rate sensitive industries in the post-1991 21Thefactthatpricesinitiallyriseasaresultofacontractionaryshock, thepricepuzzle, isacommon findingintheliterature. Onetheoreticalreasonproposedisthecostchannelofmonetarypolicy(Ravenna and Walsh, 2006). In a meta-analysis, Rusn´ak et al. (2013) find that the only consistent way to solve the price puzzle is to implement sign restrictions which impose prices to decrease. 22Two year industry growth rates result in similar dynamics. 23The finding that interest rate sensitive industries grew abnormally fast in the buildup to the crisis is consistent with a vast literature examining the effects of the housing bubble (Mian and Sufi, 2009). 11

period. The fact that no such growth occurred is worrying given the substantial easing that occurred during that time. However, the late 80s and early 90s were characterized by a very high number of FDIC insured bank failures, as shown in Figures 4.24 Furthermore, the savings and loans crisis contributed to a larger number of failures of financial intermediaries. If one takes the findings of this paper at face value, the Federal Reserve’s proactive actions in stabilizing the financial sector following the 2008 crisis were more effective at increasing the growth of interest rate sensitive industries than the easing that occurred in the 1990s. However, it is likely that other factors were also at play than monetary policy, which is not normally the predominant factor in business cycles, especially outside of the zero bound. Thus, it may be the case that the post-2001 recovery provides a better counterfactual for the zero lower bound period, with which the post-2009 period still compares favorably. Regardless, the important aspect of Figure 3 is that the post-crisis recovery, relative to historical norms, does not seem to be characterized by the monetary contraction implied by a severe liquidity trap.25 As a second exercise, I chart scatter plots of industry revenue growth rates and interest rate sensitivity in Figure 5. As mirrored in the preceding figure, the 1991-1994 recovery is characterized by relatively little growth in interest rate sensitive industries. The post- 2001 recovery is characterized by 5% growth for each standard deviation increase in index sensitivity. Finally, the 2009 recovery is characterized by similar industry growth patterns to the post-2001 period. If the zero lower bound period resulted in roughly a 5% contractionary monetary shock, as implied by Taylor rule type reaction functions (Rudebusch, 2009), we would not expect industry revenue differentials due to interest rate sensitivity to be positive. There is no evidence here that interest rate sensitive industries have been detrimentally affected by monetary policy during the zero lower bound period. Econometric Analysis Table 3 presents the basic test of this paper, estimating equation (5) at the firm-level. Firm-level data is useful, as it will allow firm-level controls. The coefficient on the index interacted with 2009-2012 reveals that a one standard deviation increase in industry sensitivity is associated with positive increases in relative revenues over that period in comparison to previous recoveries, though this is statistically insignificant.26 24Ofcourse,theTreasuryandFederalReservealsotookactionsthatresultedinlargebailoutsfollowing the 2008 financial crisis, while the FDIC was not as involved. 25Although Figure 3 appears to have a trend, this must be incidental, as it would imply that interest rate sensitive industries increasingly become a larger portion of the economy. 26The index coefficient itself equals zero, as it is a combination of the 2001-2004 recovery, which is characterized by negative growth, and 1991-1994, which is characterized by positive growth. 12

The fact that the interaction of the index and the 2009-2012 indicator is no less than previous recoveries is a major finding of this paper. Firms in industries typically sensitive to monetary policy have not had diminished revenue growth in the current recovery, providing evidence that these industries have not been affected by the zero lower bound. Note that if spreads have moved differently relative to the federal funds rate since 2009, and transmission from the federal funds rate to other interest rates has been weaker, this should also result in a decreased coefficient for 2009-2012. We do not observe that this is the case. In order to rule out other factors that could bias the coefficients, I include control variablesincolumns(2)through(4). Thecontrolvariablesdonotdiminishthemagnitude of the coefficient ψ at the zero lower bound, meaning that the index is capturing variation separate from traditional characteristics of firms affected by monetary policy. To show thattheresultsarenotdrivenbybusinesscyclevariations, Iincludeasectorlevelmeasure of GDP growth elasticity. Specifically, I regress revenue growth on annual GDP growth for each sector using yearly data from 1985 to 2012. I then save each elasticity, and interact it with GDP growth from each recovery in second-stage specification (3). All results are unchanged when controlling for sector-specific business cycle elasticities. This isreassuring,asitsignifiesthattheVARapproachiscapturingashockwhichisorthogonal to industry-level GDP dependencies. I include lagged log firm employment in the regression in column (3). Firm employment at the start of each recovery could be correlated with firm revenue growth due to financing friction differences between small and large firms, as documented by Gertler and Gilchrist (1994) in the manufacturing industry. In addition, important differences in employment growth dynamics have been found across small and large firms, depending on the aggregate unemployment rate (Moscarini and Postel-Vinay, 2012). The use of firm-level data provides the benefit of being able to adequately control for these factors. Reassuringly, the inclusion of firm employment does not change the relationship between the index and firm revenue growth. One possible confounder is that industry trends could be correlated with the interest rate sensitivity index, thus biasing the results. If firms ranked as sensitive have secular growth over the sample period, becoming a relatively larger share of output, this would erroneously make it appear that sensitive firms grew more over the recovery periods due to monetary policy. In order to test for this, I take quarterly firm averages of revenues of each NAIC sector from 1984-2008. I then calculate trend growth of revenues by regressing revenues on a time trend. I use the predicted trend coefficient as a measure of secular 13

industry growth in the second-stage regressions. In all cases, this secular growth measure does not affect the results in any way, implying that secular trends are not driving the index coefficient estimates.27 Although column (4) of Table 3 controls for secular industry trends over the sample period, it is still possible that shorter growth trends could exist in industries sensitive to monetary policy. If interest rate sensitive industries had relatively faster growth before the recoveries, and this differential continued, it would erroneously make it appear that monetary policy was causing the recovery. Table 4 estimates equation (5), using the time periods preceding the recoveries (1988-1991, 1998-2001, and 2006-2009). The index coefficients are negative, meaning that there were industry pre-trends correlated with interest rate sensitivity. However, the 2006-2009 coefficient estimate is not statistically significant. It is worrying that the index coefficients are significantly negative in Table 4, as it means that firms in interest rate sensitive industries suffered greater decreases in revenues during those recessions. It is often thought that recessions are followed by above-average growth, returning economies to pre-recession trajectories of potential output. This would likely be mirrored by above-average growth in industries that were hit hardest in the recession. If this is the case, we would expect the index coefficient to be of a stronger magnitude than in previous recoveries.28 In order to control for this potential mean reversion, I use three year lagged growth ratesascontrolvariables.29 Table5presentstheresults,showingthattheindexcoefficients from 2009-2012 are unchanged from Table 3, meaning that mean reversion cannot account for the growth in industries sensitive to monetary policy.30 ˆ The index φ is itself actually a generated variable, φ, whose uncertainty has not been 27ThiscontrolvariableisnotinteractedwiththerecoveryindicatorsI ,constrainingittohavethe t−k,t same value across recoveries. I do this since the purpose of the control variable is to show that growth rates during the recoveries are not driven by a single factor common across the time sample. 28Martinetal.(2014)examineHPfilteredoutputbeforeandafter149recessionsinadvancedeconomies from 1970, and find that, in fact, recessions are not followed by above average growth. Recessions are followed by decreased revisions of potential output. The authors state that it is likely researchers using two-endeddatafindthattheeconomyreturnstopotential,butwithone-sideddata,thetruetrendisnot apparent. 29Specifically, I include the growth rates from 1988-1991, 1998-2001, and 2006-2009 interacted with the recovery indicators in equation (5). 30Using the growth rates from 1988-1990, 1998-2000, and 2006-2008 as controls interacted with the recovery indicators yields the same results. Note that these estimates could be biased due to the correlation of firm fixed effects with the error term, though using previous years means that the firm fixed effect is not mechanically correlated. Using lagged dependent variables as instruments is difficult in this context given the small number of industry observations. In addition, controlling for firm employment helps to control for differential growth trends related to firm size. 14

taken into account thus far. Table 6 bootstraps over all possible indices to adjust for this.31 The standard errors correcting for generated regressors are comparable to the industry-clustered standard errors, though slightly larger on the index interactions. The bootstrapped corrected standard errors do not meaningfully change the value of inference from the results. 5.3 Romer and Romer’s Alternative Shock Measure In this section, I provide external validation of the industry ordering using Romer and Romer (2004)’s alternative method to identify monetary shocks.32 The authors use internal Fed forecasts to construct a series of intended federal funds rate changes that are purged of information of future output and inflation. This leaves variation that should be orthogonal to developments of these variables, allowing the series to be used as an explanatory variable while being relatively free of endogeneity. Following Coibion (2012), I cumulate R&R’s shock measures and use it in hybrid VARs, analogous to (3), with the cumulative new shock measure in place of the federal funds rate.33 I construct the index using the SVAR impulse responses in the same fashion as explained in Section 4.1. The resulting industry ordering using R&R has a correlation of 0.85 with the baseline SVAR index (Table 7). Thus, using the federal funds rate or the R&R shock measure results in a very similar industry ordering. Likewise, controlling for fiscal variables, and using non-financial output in the VARs with the R&R shocks results in a highly correlated measure.34 Table 8 reveals that second-stage estimates using the R&R shock series result in zero lower bound estimates that are quite similar to the results using the federal funds rate, though the third column does have increased standard errors. 31Pagan(1984)hasshownthatequation(5)willyieldinconsistentstandarderrorsifuncertaintyfrom indexconstructionisnottakenintoaccount. Iemployabootstrappingtechniquetoshowthatcorrecting for generated regressors does not change the statistical significance of coefficients. For each sequence of impulse responses, 200 possible indices are created from each iteration of industry SVARs described in Footnote 16. An index is drawn with replacement from this distribution, then 19 industry clusters are drawn with replacement for the second-stage, and Equation (5) is run. The standard deviations of the point estimates are then the bootstrapped standard errors of the coefficients. Industry-level data is used for this calculation so as to decrease computational burden. 32The shock series, updated and used from 1971:Q1-2007:Q4, can be obtained from Silvia Tenreyro’s website. 33The hybrid VAR methodology involves summing up the series of shocks up to time t for each t in the sample. This creates a level series comparable to the federal funds rate for the VARs, except that it contains only exogenous movements of the interest rate as identified by R&R. I use hybrid VARs since thishelpstocontrolformanyfactorsthatcancauseommittedvariablesrelativetoaunivariateregression with volatile industry data. 34The specification of the SVARs for this index is expanded upon in Section 7. 15

6 Possible Alternative Explanations 6.1 Productivity Increases In a recovery generated by monetary policy, we would expect revenues in industries positively affected by monetary policy to have proportional increases in hours. That is, we would expect sectors to have movements along their production functions, rather than shifts in their production functions due to technological advances. Recent literature (Fernald, 2014) has found diminished productivity growth leading up to and through the Great Recession. Nonetheless, a formal test of productivity growth against the monetary industry index is necessary. Table 9 estimates equation (5) with real revenues over employment growth, for a measure of labor productivity, as the independent variable.35 In none of the the recoveries is the index significant, meaning that firm labor productivity growth cannot account for the growth of interest rate sensitive industries, including at the zero bound. I additionally test for the interaction between multi-factor productivity growth and the sensitivity index by using data from Rosenthal et al. (2014). The authors combine industry-level output and intermediate inputs from the Bureau of Economic analysis with capital and labor statistics from the BLS Productivity Program to calculate contributions toaggregategrowthinvalue-addedbyindustry. Thedataisaggregatedslightlydifferently thanNAICscode. Specifically,theauthorshaveadurableandnon-durablecategorization. I thus assign durable goods to be NAIC 33, and 31 and 32 to be non-durables, as the 3-digit NAIC categorizations fall mostly along this line.36 Figure 6 presents scatter plots and an OLS regression of the sensitivity index and multi-factor productivity growth from 2009-201237 by 2-digit NAIC industry. As shown in the figure, there is little relationship. Thus, productivity increases cannot account for the positive coefficient on the index sensitivity index, and are not biasing the results. 38 35Bureau of Labor Statstics data reveals that correlations between employment and hours worked in 1991,1994,2009,and2012areallgreaterthan0.98. Thecorrelationsin2001and2004are0.895and0.921 respectively. Hence, the BLS data suggests that using employment as a proxy for hours from 2009-2012 is acceptable. 36ThemappingfromRosenthaletal.(2014)’sstudytoNAIC31,32,and33,doesnotaffecttheresults. For retail trade, I assign both NAIC groups (44 and 45) to the same value that the authors provide. 37These are the only years for which multi-factor productivity growth data is available, so I cannot test the relationship for 1991-1994 and 2001-2004. 38Includingmulti-factorproductivitygrowthinEquation1leavescoefficientestimatesintheregression unchanged. 16

6.2 Fiscal Policy One observationally equivalent explanation for the positive coefficient on interest rate sensitivity from 2009-2012 is that fiscal policy was able to counteract the effects of a contractionary monetary shock. Faced with the greatest financial crisis since the Great Depression, Congress passed the American Recovery and Reinvestment Act of 2009 (ARRA) in February of that year. The economic stimulus package resulted in an estimated $831 billion increase in spending between 2009 and 2019, including increases in unemployment benefits, and direct spending on infrastructure. The literature has found that the net effect of ARRA was relatively small. In comparison to a panel of countries, Aizenman and Pasricha (2013) find that the US was ranked in the bottom third in growth of public spending. Likewise, Aizenman and Pasricha (2010) find that after factoring in decreases in state and local government spending, consolidated government spending was insufficient to counteract decreases in private demand. These evaluations of ARRA are corroborated by Cogan et al. (2010). The authors estimate the output effects of government purchases using the Smets-Wouters model, and find maximum effects of 0.3% of GDP in 2009:Q3. Following the end of the Great Recession, fiscal policy was unprecedentedly contractionary (Cashin et al., 2016).39 I thus control for fiscal policy by constructing an industry-level measure of the effects of government spending in a similar to manner to the approach for measuring interest rate sensitivity. I run VARs for each industry while adding the growth rates of government spending40 and tax receipts as the first and second ordered variables in the VAR. The VAR specification is described at greater length in Section 7. I quantify the resulting industry responses to government spending by summing the cumulative response over a 16 quarter horizon to an orthogonalized government spending shock.41 The cumulative industry responses are then standardized with respect to each other. The resulting measure of industry sensitivity to fiscal policy is negatively correlated with the baseline interest rate sensitivity index (ρ = −0.37).42 This negative correlation is intuitive, as interest rate sensitive industries would also be sensitive to increases in interest rates due to crowding out, as long as they are not disproportionately awarded government funds. Mining, Quarrying, and Natural Gas, for example, was found to be 39Quarterly year-over-year growth rates of government consumption and investment spending were negative from mid-2010 to mid-2014. 40The series used is government consumption and gross investment from the National Accounts. 41Note that since government spending shocks can have both positive and negative industry revenue effects due to crowding out, the cumulative sum of the revenue impulse response is a more adequate measure than the minimum used for constructing industry interest rate sensitivity 42The resulting fiscal policy sensitivity index is presented in the online appendix. 17

highly sensitive to interest rates. Similarly, a positive government spending shock results in a decrease in revenue for this industry.43 Including industries’ sensitivity to fiscal policy to the baseline regressions of this paper does not change the finding that the zero bound recovery is no different than previous recoveries. Column (2) of Table 10 controls for industry sensitivity to fiscal policy. The second-stage estimates of industry interest rate sensitivity during the zero bound, though diminished, are not significantly different from other recoveries. Column (3), which includes controls, yields similar results. Interestingly, the coefficient on fiscal policy sensitivity interacted with the zero bound period shows that those industries fared significantly worse in revenue growth than during other times. This is consistent with unprecedented fiscal contraction during the zero bound decreasing growth. 7 Identifying the Policy Stance 7.1 Methodology Thus far, evidence on monetary policy effectiveness has been presented through an inferential approach comparing recoveries. In this section, I evaluate how much the results are attributable to unconventional policy, and provide time series rather than cross-sectional evidence that the zero lower bound period was characterized by accommodative monetary policy. I find evidence that unconventional policy succeeded in stimulating growth in excess of what a federal funds rate of zero would suggest. I infer the implied policy stance based on the estimated SVARs, growth rates of industry revenues, and macroeconomic aggregates used throughout the paper. The method uses the Kalman filter to extract a measure of the interest rate shocks on a quarterly basis from the structural forms of the models in (3). For expositional purposes, consider a bivariate VAR(1) system in structural form B Y = B Y +u (7) 0 t 1 t−1 t Rewriting as explicit equations for the two variables while assuming a recursive ordering, and letting (i,j) denote the ith row and jth column of the preceding matrix, x = B (1,1)x +B (1,2)z +u (8) t 1 t−1 1 t−1 1t z = −B (2,1)x +B (2,2)x +B (2,1)z +u t 0 t 1 t−1 1 t−1 2t 43These impulse responses are presented in the online appendix. 18

Notice that by construction, the structural shocks u and u are uncorrelated with 1t 2t each other, and are assumed to be normal with sufficient lag length. This implies that given data and parameters, (8) can be reformulated as a state-space model that can be solved using the Kalman filter out-of-sample. In brief, the Kalman filter minimizes one step ahead prediction errors, optimally weighting the uncertainty of the measurement equation and the uncertainty of the state. AssumeVARparametershavebeensolvedforusingasampleofobservedvalues. Then in state-space form the measurement equation is as follows: ˆ ˆ x = B (1,1)x +(B )(1,2)z +u (9) t 0 t−1 1 t−1 1t | {z } deterministic while the state equation is defined as: ˆ ˆ ˆ z = −B (2,1)x +(B )(2,2)x +B (2,1)z +u (10) t 0 t 1 t−1 1 t−1 2t | {z } deterministic Usingtheobservedseriesx , theunobservedseriesz canbesolvedforusingKalmanfilter, t t ˆ both in or out-of-sample, using the estimated VAR parameters in the B matrices. Outof-sample, the Kalman filter is essentially decomposing the forecast errors of structural forecasts of the SVAR equations into an interest rate component and other disturbances. I make three main changes from the baseline SVARs used throughout the paper.44 As shown in the industry impulse responses in Figure 1, the financial sector (NAIC 52) has a starkly different response to a monetary contraction than any other industry, with a large initial increase in revenues, followed by a subsequent decrease to negative growth. In the shortrun,financialfirmscanbenefitfromincreasedinterestrates,andhencelargerspreads between short and long-term rates, before the macroeconomic effects of monetary policy depress output. As a result of the strong response of the financial sector, VARs examining the post-1980 subsample often find indeterminate output responses to monetary shocks.45 I thus place non-financial corporate output into the SVARs in place of GDP growth,46 as the Kalman filter cannot accurately extract monetary shocks following the 1980s if an important measurement equation capturing aggregate demand ceases being informative. 44The resulting industry interest sensitivity ranking is highly correlated (ρ > 0.8) regardless of these changes, as shown in Section 5.3. 45Asecondreasonhas been proposedin the monetaryVARliteratureforthe weakresponse ofoutput to monetary shocks for post-1980 subsamples. Boivin and Giannoni (2006) find that by responding more strongly to inflation, monetary policy has been more effective in counteracting shocks since the 1980s. Since interest rate movements are more systematic and endogenous than in earlier periods, VARs are unable to identify the effects of unsystematic interest rate shocks. 46This follows Lozej (2012). 19

Second, I incorporate the growth rates of government spending and taxation into the SVARs.47 Past literature has found that fiscal shocks are an important driver of medium-run macroeconomic dynamics (Rossi and Zubairy, 2011), and controlling for fiscal policy thus results in much more accurate estimates when cross-validating the SVAR model Kalman filtered state estimates of monetary shocks out-of-sample. Furthermore, controllingforfiscalpolicyisimportantduringthezeroboundperiod,asitischaracterized by an unprecedented number of years of fiscal contraction. Finally, industries can be heterogenously affected by fiscal policy, as found in Section 6.2. Third, I use cumulative Romer and Romer shocks in place of the federal funds rate.48 Since the R&R shocks are constructed so as to be relatively free of endogeneity and unpredictable, this allows the state equation to be set up as a random walk, whereby increases or decreases in the monetary stance are agnostically treated as equally likely. ThisthereforedoesnotimposeanimplicitFedreactionfunctionuponthestateequation.49 This is useful at the zero lower bound, as it shows that state estimates during that period are not driven by previous Federal Reserve behavior imposed in the state equation. In matrix form, the actual state-space system used can be written in the following manner: z = z +u (11) t t−1 6t ˆ ˆ ˆ X = C +F(L)W +Λ(L)z +ζ t t t−1 t where z , the state, is the cumulative sum of R&R shocks up to quarter t, and X is a t t vector of the variables used in the SVARs besides fiscal variables: non-financial corporate output, industry revenue growth rates, and PCE inflation, all in year-over-year growth rates.50 The vector W includes all variables in X , and also includes fiscal variables. t t The error term u and the vector of errors ζ are normally distributed. The term c and 6t t vector C are the constants from the right hand side variables of each line in the system. Contemporaneous restrictions are imposed in the following manner: 47InordertomaintainaparsimoniousVARmodelwhileaddingfiscalvariablesIalsoomitcommodity prices. 48This shock series ends 2007:Q4. 49In practice, the Cholesky-identified monetary shocks are not predicted well in placebo tests with either a random walk state equation or imposing the VAR federal funds rate equation as the state equation. 50Note that on the right hand side side, the parameter in Fˆ for each x ’s contemporaneous effect on 0 t x is set to zero. t 20

 1 0 0 0 0 0 (cid:15)   u  GovtCons,t GovtCons,t a 2,1 1 0 0 0 0 (cid:15) Tax,t   u Tax,t       a a 1 0 0 0(cid:15)  u   3,1 3,2  NFCoutput,t =  NFCoutput,t a a a 1 0 0(cid:15)  u   4,1 4,2 4,3  IndustryRev,t  IndustryRev,t      a a a 0 1 0 (cid:15) u  5,1 5,2 5,3  Inflation,t   Inflation,t  0 0 a 0 a 1 (cid:15) u 6,3 6,5 R&Rshocks,t R&Rshocks,t The model is essentially Cholesky identified with fiscal variables ordered first, with a few differences. As before, industry revenues are constrained to not affect any other variables either contemporaneously or through lagged parameters. I also impose that the Romer and Romer shocks are not affected by fiscal policy contemporaenously or through lagged parameters.51 The SVARs are run, yielding structural forms for the federal funds rate, macro variables, and industry revenues.52 I then input the parameters and data from both the ˆ 1971:Q1-2007:Q4 and post-2008 period into system (11), treating F(L)W as exogenous t disturbances to the measurement equations. Finally, I run the Kalman filter on this system, treating z as the unobserved state. t It is important to mention an aspect of the estimated state series z . This method does t not explicitly capture a market “shadow” interest rate as other papers in the literature have done (Wu and Xia, 2014). The state series after 2008:Q3 is yielding an estimate of cumulative Romer and Romer shocks that would imply the realizations of X observed t post-zero lower bound. Thus, z reveals the amount of monetary accommodation in excess t of the Federal Reserve’s typical reaction function as identified by Romer and Romer, which during the zero bound would imply a negative federal funds rate. It is important to note that if the monetary transmission mechanism has been impaired- say by financial disruption- this will result in the state estimate measuring contractionary policy, in spite of the Fed’s actions. The state estimate is conditional on the pre-zero lower bound VAR dynamics. If the entirestructureoftheeconomyhaschanged-includingtherelationshipsbetweeninflation, output, and the federal funds rate- this estimate will not reflect the true monetary stance. In order for the state estimates to be valid, the VAR parameters from before the zero lower bound need to accurately capture dynamics post-2008:Q1.53 This implies that the 51Feedback effects between fiscal and monetary policy are ignored, as they are not important for the question of interest. This is also consistent with the exclusion of fiscal variables in state-space system (11). These constraints do not affect the results. 52Impulse responses of the macro and industry responses to Romer and Romer shocks are available in the online appendix. 53Section 7.3 provides evidence that this is the case. 21

distribution of structural shocks since 2008 need to be estimated correctly from before the zerolowerbound. ItislikelythatnegativeshockstoGDPwereunderestimatedintheprezlbsample. Intuitively,however,thiswouldlikelybiasthestateestimatespositivelyrather than negatively. We would expect the Kalman filter to erroneously attribute negative shocks to GDP and industry revenue growth rates to restrictive monetary policy. Due to the direction of expected bias, we can thus infer that, if anything, the state estimate of the R&R shocks likely understates the true monetary stance created by unconventional policy. It is of note that other papers in this literature (Krippner, 2013; Wu and Xia, 2014; Lombardi and Zhu, 2014) also make strong assumptions regarding the relationships between observed variables and unobserved measures of monetary policy during the zero lower bound period, including linearity. 7.2 Results Figure 7 plots the filtered state equation from System 11.54 The estimated state is characterized by negative (expansionary) shocks soon following the incidence of the zero lower bound, which is denoted by a red line. This implies that unconventional policy was able to stimulate corporate industries in excess of the Federal Reserve’s previous reaction function, meaning that policy after this point was even more accommodative than what a Taylor-type rule would suggest. This fact helps to explain why it does not seem that interest rate sensitive industries had diminished revenue growth at the zero bound, as it implies that unconventional policy was able to provide monetary accommodation. This result is especially interesting bearing in mind the setup of the state equation: the Romer and Romer shock series was imposed to follow a random walk, meaning that no structure in the state equation is responsible for the measurement of accommodative shocks. 7.3 Placebo Test The length of the sample allows cross-validation of the results. If the approach accurately captures the stance of policy during the zero bound, it should also accurately capture policy during a subsample in which the VARs have not been estimated. In Figure 8, I 54In all charts, standard errors are calculated so as to correct for VAR parameter uncertainty. Steps are as follows: first, 100 new datasets are created using a four quarter moving block bootstrap from the VAR, initialized at 1971:Q1-1971:Q4 values. The VAR is reestimated over each of these new datasets, yielding a distribution of the VAR parameters. The Kalman filter is then run using each estimated VAR in state-space form, with the actual 1971:Q1-2014:Q2 data. The resulting 5th and 95th percentile values oftheKalmanrootmeansquarederrorsarethenpresented,thustakingintoaccountboththeuncertainty of the VAR parameters and measurement of the unobserved variable. 22

estimate the SVARs from 1971:Q1-1999:Q3,55 then run the Kalman filter over 1971:Q1- 2014:Q2 to extract the R&R shocks. A line is drawn at 1999:Q3 to show that after this point, the Kalman estimates are all using VARs out-of-sample, while another reference line is drawn for the incidence of the zero bound. In support of the credibility of the approach, the 1999:Q4-2007:Q4 estimates of the state are very close to the actual R&R shocks over this period, and mirror the results in Figure 7. The VARs thus are capturing time-invariant relationships and do not suffer from overfitting. 7.4 Great Moderation Subsample Analysis Although the state-space model performs well out-of-sample, one may still wish to see that the method presented finds accommodative shocks using only data from the Great Moderation. Figure 9 plots a similar estimation process as in Figure 7, except that VARs are run on data from 1988:Q1-2007:Q4. The zero bound period is still characterized by accommodative monetary policy. 8 Conclusion The use of industry data presents a new avenue for examining growth at the zero lower bound. I measure an industry ranking of interest rate sensitivity, and cross-sectionally examine its relationship with revenue growth while Federal Reserve policy was constrained. The approach helps inform current knowledge of monetary policy’s real effects by placing into context how industries have grown with respect to the historical record, and shows that interest rate sensitive industries have performed well in the latest recovery. There is no evidence that interest rate sensitive industries have been detrimentally affected by monetary policy, in spite of the large difference between zero and the negative federal funds rate characterized by the Federal Reserve’s pre-zero lower bound reaction function. As firms have grown equally well as in previous recoveries,56 it must be the case that either unconventional monetary policy had real effects, or other shocks on net caused differential growth in interest rate sensitive industries. However, a thorough evaluation of other plausible shocks leaves monetary policy as the explanation for this differential growth. Theevidenceforunconventionalmonetarypolicy’sbeneficialeffectsisvalidatedthrough 55A variety of breakpoints were experimented with yielding similar results. 56This is in aggregate, not cross-sectional terms, as shown in Table 11. 23

a new time series method measuring monetary accommodation at the zero bound using observed macro and industry data. This quantifies the portion of observed growth attributable to monetary policy. I use estimated structural VAR parameters from before the zero lower bound to predict the most likely monetary shocks given observed data after 2008:Q3. The estimated shocks closely follow those observed in-sample, and are crossvalidated with out-of-sample testing. During the zero bound, the measure turns negative significantly. This implies that the zero bound period is characterized by expansionary monetary shocks in excess of typical Federal Reserve behavior, and that unconventional policy stimulated growth as further decreases in the federal funds rate would have. This approach merits further use and could be extended to create an implied stance measured from the macroeconomy, rather than from the public firms examined in this paper. The results presented in this paper have important ramifications for the conduct of monetary policy at zero interest rates. Despite the inability of the Federal Reserve to further decrease the federal funds rate, there is no evidence that monetary policy has inhibited growth. Standard theory and empirical evidence would have led researchers to believe that announcements of future policies, alteration in relative asset supplies, and quantitative easing would have had limited ability to influence real output (Walsh, 2014). In contrast, the results show that the US recovery since 2009 compares favorably to past recoveries, and that this is in part due to unconventional monetary policy. An important caveat to these results is that all data is derived from publicly-listed firms. Public firms are larger than non-public firms, and hence are thought to be less affectedbycreditconstraintsandthecreditchannel(GertlerandGilchrist,1994). Fromthis perspective, we would expect non-public firms to be more affected by restrictive monetary policy. On the other hand, programs such as quantitative easing may disproportionately assist public firms, as they directly benefit from increases in stock and bond valuation. Thus, it is likely that the firms examined in this paper would be the least likely to be affected by the zero bound constraint on policy. 24

References Aizenman, J. and Pasricha, G. K. (2010). The net fiscal expenditure stimulus in the us 2008–2009: Less than what you might think. VoxEU. org, 3. Aizenman, J. and Pasricha, G. K. (2013). Net fiscal stimulus during the great recession. Review of Development Economics, 17(3):397–413. Barth, M. J. and Ramey, V. A. (2002). The Cost Channel of Monetary Transmission. In NBER Macroeconomics Annual 2001, Volume 16, NBER Chapters, pages 199–256. National Bureau of Economic Research, Inc. Bernanke, B. S. and Gertler, M. (1995). Inside the black box: the credit channel of monetary policy transmission. Technical report, National bureau of economic research. Bernanke, B. S. and Kuttner, K. N. (2005). What Explains the Stock Market’s Reaction to Federal Reserve Policy? Journal of Finance, 60(3):1221–1257. Boivin, J. and Giannoni, M. P. (2006). Has monetary policy become more effective? The Review of Economics and Statistics, 88(3):445–462. Cashin, D., Lenney, J., Lutz, B., and Peterman, W. (2016). Fiscal policy changes and aggregate demand in the u.s. during and following the great recession. Chen, S.-S. (2007). Does monetary policy have asymmetric effects on stock returns? Journal of Money, Credit and Banking, 39(2-3). Christensen, J. H. and Rudebusch, G. D. (2013). Modeling yields at the zero lower bound: Are shadow rates the solution? Citeseer. Christiano, L. J., Eichenbaum, M., and Evans, C. L. (1999). Monetary policy shocks: What have we learned and to what end? Handbook of macroeconomics, 1:65–148. Claessens, S., Tong, H., andWei, S.-J.(2011). FromtheFinancialCrisistotheRealEconomy: Using Firm-level Data to Identify Transmission Channels. In Global Financial Crisis, NBER Chapters. National Bureau of Economic Research, Inc. Cogan, J. F., Cwik, T., Taylor, J. B., and Wieland, V. (2010). New keynesian versus old keynesian government spending multipliers. Journal of Economic dynamics and control, 34(3):281–295. Coibion, O. (2012). Are the effects of monetary policy shocks big or small? American Economic Journal. Macroeconomics, 4(2):1. Davis, S. J., Haltiwanger, J., and Schuh, S. (1998). Job Creation and Destruction, volume 1. The MIT Press, 1 edition. Dedola, L. and Lippi, F. (2005). The monetary transmission mechanism: Evidence from the industries of five OECD countries. European Economic Review, 49(6):1543–1569. 25

Ehrmann, M. and Fratzscher, M. (2004). Taking stock: Monetary policy transmission to equity markets. Journal of Money, Credit, and Banking, 36(4):719–737. Erceg, C. and Levin, A. (2006). Optimal monetary policy with durable consumption goods. Journal of Monetary Economics, 53(7):1341–1359. Fernald, J. (2014). Productivity and potential output before, during, and after the great recession. Technical report, National Bureau of Economic Research. Gertler, M. and Gilchrist, S. (1994). Monetary policy, business cycles, and the behavior of small manufacturing firms. The Quarterly Journal of Economics, 109(2):pp. 309–340. Gilchrist, S., L´opez-Salido, D., and Zakrajˇsek, E. (2014). Monetary policy and real borrowing costs at the zero lower bound. Technical report, National Bureau of Economic Research. Gorodnichenko, Y. and Weber, M. (2014). Are sticky prices costly? evidence from the stock market. Technical report, American Economic Review. Krippner, L. (2013). Measuring the stance of monetary policy in zero lower bound environments. Economics Letters, 118(1):135 – 138. Lombardi, M. J. and Zhu, F. (2014). A shadow policy rate to calibrate us monetary policy at the zero lower bound. Lozej, M. (2012). Financial innovation and firm debt over the business cycle. Maio, P. (2013). Another look at the stock return response to monetary policy actions*. Review of Finance. Martin, R., Munyan, T., and Wilson, B. A. (2014). Potential output and recessions: Are we fooling ourselves? Technical report, IFDP Notes. Mian, A. and Sufi, A. (2009). The consequences of mortgage credit expansion: Evidence fromtheu.s.mortgagedefaultcrisis. The Quarterly Journal of Economics,124(4):1449– 1496. Moscarini, G. and Postel-Vinay, F. (2012). The contribution of large and small employers to job creation in times of high and low unemployment. The American Economic Review, 102(6):2509–2539. Pagan, A. (1984). Econometric issues in the analysis of regressions with generated regressors. International Economic Review, pages 221–247. Peersman, G. and Smets, F. (2005). The Industry Effects of Monetary Policy in the Euro Area. Economic Journal, 115(503):319–342. Rajan, R. G. and Zingales, L. (1998). Financial Dependence and Growth. American Economic Review, 88(3):559–86. 26

Ramey, V. A. (2015). Macroeconomic shocks and their propagation. Ravenna, F. and Walsh, C. E. (2006). Optimal monetary policy with the cost channel. Journal of Monetary Economics, 53(2):199–216. Romer, C. D. and Romer, D. H. (2004). A new measure of monetary shocks: Derivation and implications. The American Economic Review, 94(4):1055–1084. Rosenthal, S., Russell, M., Samuels, J. D., Strassner, E. H., and Usher, L. (2014). Integrated industry-level production account for the united states: Intellectual property products and the 2007 naics. May, 15:2014. Rossi, B. and Zubairy, S. (2011). What is the importance of monetary and fiscal shocks in explaining us macroeconomic fluctuations? Journal of Money, Credit and Banking, 43(6):1247–1270. Rudebusch, G.D.(2009). Thefed’smonetarypolicyresponsetothecurrentcrisis. FRBSF economic Letter. Rusna´k, M., Havranek, T., and Horv´ath, R. (2013). How to solve the price puzzle? a meta-analysis. Journal of Money, Credit and Banking, 45(1):37–70. Steinsson, J. and Nakamura, E. (2014). High frequency identification of monetary nonneutrality. In 2014 Meeting Papers, number 96. Society for Economic Dynamics. Swanson,E.T.andWilliams, J.C.(2014). Measuringtheeffectofthezerolowerboundon medium- and longer-term interest rates. American Economic Review, 104(10):3154–85. Tenreyro, S. and Thwaites, G. (2013). Pushing on a string: Us monetary policy is less powerful in recessions. Walsh, C. E. (2014). Monetary policy transmission channels and policy instruments. Weise, C. L. (1999). The asymmetric effects of monetary policy: A nonlinear vector autoregression approach. Journal of Money, Credit and Banking, 31(1):pp. 85–108. Woodford, M. (2012). Methods of policy accommodation at the interest-rate lower bound. In The Changing Policy Landscape: 2012 Jackson Hole Symposium. Federal Reserve Bank of Kansas City. Wu, J.C.andXia, F.D.(2014). Measuringthemacroeconomicimpactofmonetarypolicy at the zero lower bound. Technical report, National Bureau of Economic Research. 27

sexa-x ehT .3Q:8002-1Q:1791 ,kcohs etar sdnuf laredef noitaived dradnats eno ot sesnopser dezilanogohtro yrtsudnI :1 erugiF .eulb ni deniltuo partstoob partstoob kcolb gnivom ralucric gnisu slavretni ecnedfinoc %09 .sretrauq etoned 2 1 0 1- 2- 3- )12( .cte ,gniyrrauQ ,gniniM 51 01 5 0 5. 0 5.- 1- )22( seitilitU 51 01 5 0 1 0 1- 2- 3- )32( noitcurtsnoC 51 01 5 0 1 5. 0 5.- 1- )13( gnirutcafunaM 51 01 5 0 1 5. 0 5.- 1- )23( gnirutcafunaM 51 01 5 0 1 5. 0 5.- 1- 5.1- )33( gnirutcafunaM 51 01 5 0 1 5. 0 5.- 1- 5.1- )24( edarT elaselohW 51 01 5 0 1 5. 0 5.- 1- )44( edarT liateR 51 01 5 0 1 5. 0 5.- 1- 5.1- )54( edarT liateR 51 01 5 0 1 5. 0 5.- 1- 5.1- )84( gnisuoheraW dna noitatropsnarT 51 01 5 0 1 5. 0 5.- 1- )15( noitamrofnI 51 01 5 0 3 2 1 0 1- 2- )25( ecnarusnI dna ecnaniF 51 01 5 0 1 5. 0 5.- 1- )35( .cte ,latneR ,etatsE laeR 51 01 5 0 1 5. 0 5.- )45( .cte secivreS lanoisseforP 51 01 5 0 1 5. 0 5.- 1- )65( .cte evitartsinimdA 51 01 5 0 5.1 1 5. 0 5.- )26( ecnatsissA laicoS dna eraC htlaeH 51 01 5 0 1 0 1- 2- )17( .ceR dna ,tnemniatretnE ,strA 51 01 5 0 1 5. 0 5.- 1- )27( .cte noitadommoccA 51 01 5 0 5.1 1 5. 0 5.- 1- )18( secivreS rehtO 51 01 5 0 28

ehT .elbairav esnopser ,elbairav eslupmi yb skcohs noitaived dradnats eno ot sesnopser dezilanogohtro etagerggA :2 erugiF .eulb ni deniltuo partstoob ralucric gnisu slavretni ecnedfinoc %09 .sretrauq etoned sexa-x 1 5. 0 5.htworg PDG ,htworg PDG 51 01 5 0 1 5. 0 5.htworgmoC ,htworg PDG 51 01 5 0 6. 4. 2. 0 2.htworg ECP ,htworg PDG 51 01 5 0 8. 6. 4. 2. 0 sdnuf def ,htworg PDG 51 01 5 0 4. 2. 0 2.- 4.- 6.htworg PDG ,htworgmoC 51 01 5 0 2 1 0 1htworgmoC ,htworgmoC 51 01 5 0 6. 4. 2. 0 2.htworg ECP ,htworgmoC 51 01 5 0 6. 4. 2. 0 2.- 4.sdnuf def ,htworgmoC 51 01 5 0 2. 1. 0 1.- 2.- 3.htworg PDG ,htworg ECP 51 01 5 0 5.1 1 5. 0 5.htworgmoC ,htworg ECP 51 01 5 0 6. 4. 2. 0 htworg ECP ,htworg ECP 51 01 5 0 6. 4. 2. 0 2.sdnuf def ,htworg ECP 51 01 5 0 4. 2. 0 2.- 4.- 6.htworg PDG ,sdnuf def 51 01 5 0 1 5. 0 5.- 1htworgmoC ,sdnuf def 51 01 5 0 2. 0 2.- 4.- 6.htworg ECP ,sdnuf def 51 01 5 0 1 5. 0 5.sdnuf def ,sdnuf def 51 01 5 0 29

Figure 3: Coefficient estimates of a one standard deviation increase in index sensitivity. Dependent variable: rolling three year revenue growth. htworG euneveR raeY eerhT 02 01 0 01- 02- 1990 1995 2000 2005 2010 2014 Year Coefficient on Interest Rate Sensitivity (90% confidence interval) Figure 4: Estimated FDIC Losses Due to Failures and Acquisitions of Banks and Savings Institutions )PDG fo %( sessoL detamitsE 1 8. 6. 4. 2. 0 1986 1989 1992 1995 1998 2001 2004 2007 2010 2013 2016 YEAR Source: FDIC, Bureau of Economic Analysis, and author’s calculations. 30

Figure 5: Industry growth rates and interest rate sensitivity, 95% confidence intervals in grey. Each industry labeled with North American Industry Classification Code. 62 42 51 72 71 81 44 32 45 56 48 54 33 23 31 21 52 22 53 htworG euneveR 06 04 02 0 02- 1991-1994 21 23 62 45 48 32 42 54 51 3 5 1 6 4 5 4 3 33 71 52 81 72 22 -1 0 1 2 3 Interest Rate Sensitivity htworG euneveR 06 04 02 0 02- 2001-2004 -1 0 1 2 3 Interest Rate Sensitivity 21 51 81 4 3 4 1 32 33 42 48 23 62 54 72 56 53 45 71 52 22 htworG euneveR 06 04 02 0 02- 2009-2012 -1 0 1 2 3 Interest Rate Sensitivity Source: Compustat and author’s calculations. 31

Figure 6: Industry Multi-Factor Productivity Growth and Interest Rate Sensitivity Index 33 53 54 21 22 51 72 42 48 56 71 23 44 45 52 81 62 31 32 htworG ytivitcudorP rotcaF-itluM 4. 2. 0 2.- -2 -1 0 1 2 Index Sources: Compustat, Rosenthal et al. (2014), and author’s calculations. Figure 7: Implied Stance of Monetary Policy, Benchmark Estimate 5 0 5- 01- 1972q1 1982q3 1993q1 2003q3 2008q3 2014q1 Quarter Romer and Romer Shocks Filtered State Estimate (+/- 2 RMSE) 32

Figure 8: Implied Stance of Monetary Policy: Out-of-Sample Cross-Validation 5 0 5- 01- 1972q1 1982q3 1993q1 1999q3 2008q3 2014q1 Quarter Romer and Romer Shocks Filtered State Estimate (Benchmark +/- 2 RMSE) Estimates following 1999:Q3 are using out-of-sample VAR parameters from before that date. Figure 9: Implied Stance of Monetary Policy: Great Moderation Data 5 0 5- 01- 1990q1 1995q1 2000q1 2005q1 2008q3 2015q1 Quarter Romer and Romer Shocks Filtered State Estimate (Benchmark +/- 2 RMSE) 33

3Q:8002-1Q:1791 :htworG euneveR yrtsudnI naeM :1 elbaT xaM niM DS naideM naeM sbO sciaN emaN 71.24 18.82- 90.41 05.4 86.4 151 11 gnitnuH dna gnihsiF ,yrtseroF ,erutlucirgA 55.83 15.64- 83.51 48.8 01.8 151 12 noitcartxE saG dna liO dna ,gniyrrauQ ,gniniM 58.92 48.72- 3.7 70.5 92.4 151 22 seitilitU 89.92 45.42- 23.11 39.4 43.4 151 32 noitcurtsnoC 38.41 25.41- 48.4 34.3 54.3 151 13 )selbarud-non yltsom( gnirutcafunaM 41.71 71.11- 04.5 04.7 67.6 151 23 )selbarud-non yltsom( gnirutcafunaM 21.91 39.61- 22.7 11.7 92.6 151 33 )selbarud yltsom( gnirutcafunaM 67.71 39.61- 54.7 10.8 00.6 151 24 edarT elaselohW 03.61 70.5- 64.4 90.7 88.6 151 44 edarT liateR 75.53 09.5- 70.7 98.8 48.8 151 54 edarT liateR 73.92 71.51- 42.7 94.8 74.7 151 84 gnisuoheraW dna noitatropsnarT 94.93 48.91- 62.21 82.8 04.8 151 94 gnisuoheraW dna noitatropsnarT 44.93 12.4- 23.8 92.01 21.11 151 15 noitamrofnI 67.42 53.03- 01.8 89.6 05.6 151 25 ecnarusnI dna ecnaniF 81.32 67.31- 61.8 85.5 09.4 151 35 gnisaeL dna latneR dna etatsE laeR 23.72 86.8- 52.7 64.9 55.9 151 45 secivreS lacinhceT dna ,cfiitneicS ,lanoisseforP 61.72 44.11- 32.8 62.01 77.9 151 65 secivreS noitaidemeR/tnemeganaM etsaW ,troppuS ,evitartsinimdA 65.56 3.72- 09.41 27.21 04.41 151 16 secivreS lanoitacudE 17.23 39.9- 42.8 00.51 11.51 151 26 ecnatsissA laicoS dna eraC htlaeH 06.23 05.81- 28.01 88.01 83.01 151 17 noitaerceR dna ,tnemniatretnE ,strA 86.52 99.11- 68.6 93.7 39.6 151 27 secivreS dooF dna noitadommoccA 81.92 73.71- 17.9 47.4 83.5 151 18 )noitartsinimdA cilbuP tpecxe( secivreS rehtO 65.56 15.64- 87.9 87.7 17.7 2233 latoT .snoitaluclacs’rohtuadnatatsupmoC :ecruoS 34

xednI ytivitisneS etaR tseretnI RAVS CIAN tigiD owT :2 elbaT rotceS xednI )%( esnopseR edoC tigiD owT noitcurtsnoC 35.2 68.1- 32 noitcartxE saG dna liO dna ,gniyrrauQ ,gniniM 68.1 45.1- 12 gnisuoheraW dna noitatropsnarT 38.0 40.1- 84 ecnarusnI dna ecnaniF 96.0 79.0- 25 noitaerceR dna ,tnemniatretnE ,strA 34.0 58.0- 17 edarT liateR 83.0 28.0- 54 edarT elaselohW 83.0 28.0- 24 )selbarud yltsom( gnirutcafunaM 92.0 87.0- 33 )selbarud-non yltsom( gnirutcafunaM 12.0- 35.0- 23 )selbarud-non yltsom( gnirutcafunaM 53.0- 74.0- 13 gnisaeL dna latneR dna etatsE laeR 53.0- 74.0- 35 edarT liateR 93.0- 54.0- 44 seitilitU 25.0- 83.0- 22 secivreS noitaidemeR dna tnemeganaM etsaW dna ,troppuS ,evitartsinimdA 45.0- 73.0- 65 secivreS dooF dna noitadommoccA 55.0- 73.0- 27 )noitartsinimdA cilbuP tpecxe( secivreS rehtO 19.0- 91.0- 18 noitamrofnI 80.1- 11.0- 15 secivreS lacinhceT dna ,cfiitneicS ,lanoisseforP 91.1- 60.0- 45 ecnatsissA laicoS dna eraC htlaeH 13.1- 0 26 35

Table 3: Real Revenue Growth, Baseline Specification (1) (2) (3) (4) Index -0.028 0.696 0.531 0.885 (1.779) (2.545) (2.361) (2.653) Index·2012 2.442 2.346 2.611 2.617 (3.692) (2.970) (3.087) (3.096) Elasticity·94 4.489** 3.943** 4.311** (1.842) (1.718) (1.767) Elasticity·09 -0.586 -0.050 0.121 (3.628) (3.503) (3.349) Elasticity·12 1.088 0.788 1.091 (4.260) (4.149) (3.893) Emp·94 -5.874*** -5.830*** (0.954) (0.960) Emp·09 0.531 0.667 (0.675) (0.701) Emp·12 -0.027 0.080 (0.685) (0.671) Secular Trend -3.490 (5.083) 1991-1994 19.222*** 12.532*** 17.787*** 20.181*** (2.146) (3.862) (3.907) (3.904) 2001-2004 12.107*** 12.992* 10.295 12.699** (1.494) (6.831) (6.557) (5.717) 2009-2012 14.887*** 13.263 13.276 15.496* (3.183) (7.984) (8.152) (8.211) Observations 17,723 17,723 16,480 16,480 Adjusted R-squared 0.068 0.069 0.067 0.068 *** p < 0.01, ** p < 0.05, * p < 0.1. Standard errors clustered at NAIC 2 digit level in parentheses. 36

Table 4: Real Revenue Growth: Pre-trends (1) (2) (3) (4) Index -4.589 -4.932 -4.318* -3.729* (4.145) (2.847) (2.340) (2.075) Index·2009-2009 -0.961 -2.729 -3.677 -3.693 (3.080) (2.973) (2.592) (2.548) Observations 18,333 18,333 16,313 16,313 Adjusted R-squared 0.063 0.064 0.041 0.042 **** p < 0.01, ** p < 0.05, * p < 0.1. Standard errors clustered at NAIC 2 digitlevelinparentheses. Timeindicatorsnotshown. Column(2)controlsfor industry-level GDP elasticity, column (3) controls for initial log employment, andColumn(4)controlsforsecularindustrytrendsasinTable3. Table 5: Real Revenue Growth: Controlling for Mean Reversion (1) (2) (3) (4) Index 1.833 2.134 2.582 3.043 (1.340) (1.855) (1.822) (1.855) Index·2012 0.913 1.497 1.532 1.526 (3.139) (2.692) (2.679) (2.678) Lagged Growth 0.046*** 0.047*** 0.048*** 0.047*** (0.016) (0.016) (0.014) (0.014) Observations 14,734 14,734 14,157 14,157 Adjusted R-squared 0.041 0.042 0.053 0.054 ****p<0.01,**p<0.05,*p<0.1. StandarderrorsclusteredatNAIC2digitlevelin parentheses. Time indicators not shown. Column (2) controls for industry-level GDP elasticity, column (3) controls for initial log employment, and Column (4) controls for secularindustrytrendsasinTable3. 37

Table 6: Real Revenue Growth: Bootstrapped Standard Errors Accounting for Generated Regressors Clustered Bootstrapped Index 1.087 (1.701) (2.248) Index·2012 2.762 (2.947) (3.439) 1991-1994 20.554*** (2.342) (2.637) 2001-2004 13.827*** (2.342) (1.931) 2009-2012 15.671*** (2.342) (2.208) Observations 57 57 Adjusted R-squared 0.731 *** p < 0.01, ** p < 0.05, * p < 0.1. Bootstrap procedure describedintext. Table 7: Romer and Romer Alternative Shock Measures Baseline R&R Shocks W/ NFC and Fiscal Variables Baseline 1.00 R&R 0.85 1.00 R&R NFC 0.80 0.73 1.00 Column(2)usesanindexmeasuredwithcumulativeRomerandRomershocksinplaceofthe federalfundsrate. Column(3)usesthecumulativeRomerandRomershocksinplaceofthe federalfundsrate,non-financialcorporateoutputinplaceofGDPgrowth,includesgovernment consumptionandinvestmentorderedfirst, includesgovernmenttaxrevenuesorderedsecond, andexcludescommodityprices. 38

Table 8: Real Revenue Growth, Romer and Romer Shock Measures (1) (2) (3) Index 0.885 0.662 2.481 (2.653) (1.962) (9.835) Index·2012 2.617 2.741 4.937 (3.096) (2.652) (12.467) 1991-1994 20.181*** 19.771*** 18.461** (3.904) (4.096) (7.380) 2001-2004 12.699** 12.671** 10.991 (5.717) (5.552) (9.181) 2009-2012 15.496* 20.098** 11.716 (8.211) (9.471) (9.860) Observations 16,480 16,480 16,480 Adjusted R-squared 0.068 0.068 0.067 *** p<0.01, ** p<0.05, * p<0.1. Industry clustered standard errors in parentheses. Column (1) uses the baseline sensitivity index. Column (2) usesanindexmeasuredwithcumulativeRomerandRomershocksinplace ofthefederalfundsrate. Column(3)usesthecumulativeRomerandRomer shocksinplaceofthefederalfundsrate,non-financialcorporateoutputin placeofGDPgrowth,includesgovernmentconsumptionandinvestmentorderedfirst,includesgovernmenttaxrevenuesorderedsecond,andexcludes commodityprices. Table 9: Labor Productivity Growth (1) (2) (3) Index 0.771 0.391 0.466 (2.085) (2.892) (3.190) Index·2012 -0.705 0.451 0.454 (3.019) (2.759) (2.757) 1991-1994 13.643*** 8.393 8.933** (2.403) (5.007) (3.125) 2001-2004 4.091 9.805* 10.368** (2.494) (5.334) (3.865) 2009-2012 11.251*** 9.226** 9.742*** (1.139) (3.215) (3.359) Observations 15,912 15,912 15,912 Adjusted R-squared 0.038 0.040 0.040 *** p<0.01, ** p<0.05, * p<0.1. Industry clustered standard errors in parentheses. Control variable coefficients omitted from table. Column (2) controls for sectoral GDP elasticity. Column (3) controls for industry-specificgrowthtrends. 39

Table 10: Real Revenue Growth: Controlling for Fiscal Policy (1) (2) (3) Index -0.028 0.553 0.814 (1.779) (2.309) (2.553) Index·2012 2.442 0.639 3.228 (3.692) (2.671) (2.641) FiscalIndex 2.021 1.976 (2.963) (3.561) FiscalIndex·2012 -5.808* -9.499*** (2.815) (2.697) 1991-1994 19.222*** 19.068*** 22.195*** (2.146) (2.128) (7.546) 2001-2004 12.107*** 11.833*** 14.105** (1.494) (1.769) (6.374) 2009-2012 14.887*** 15.191*** 8.006 (3.183) (3.061) (7.970) Observations 17,723 17,723 16,480 Adjusted R-squared 0.068 0.069 0.070 *** p<0.01, ** p<0.05, * p<0.1. Standard errors clustered at NAIC 2 digit level in parentheses. Column (3) controls for sectoral GDP elasticity, initiallogemployment,andindustry-specificgrowthtrends. Table 11: Summary Statistics: Key Dependent Variables Real Revenue Growth Years Obs Mean Median Std. Dev. Min Max 1991-1994 5987 19.27 15.00 59.89 -199.68 199.58 2001-2004 6778 12.12 11.45 60.12 -199.81 199.88 2009-2012 5124 15.12 13.22 52.11 -199.91 199.87 Employment Growth 1991-1994 5249 13.65 8.08 56.55 -200 200 2001-2004 6144 4.09 4.20 52.11 -200 200 2009-2012 4668 11.26 9.01 44.21 -200 200 40

Appendix to “How Effective is Monetary Policy at the Zero Lower Bound? Identification Through Industry Heterogeneity” Arsenios Skaperdas For Online Publication Robustness Alternative Interest Rate Sensitivity Indices As shown in the paper, industry orthogonalized impulse responses exhibit variation at varied quarters. I use alternative measure of index construction to test for the importance of this variation. Table 1 reveals that constructing the index using 7 quarter structural impulse response point estimates, or 16 quarter cumulative responses, result in highly correlated industry orderings to the baseline using the minimum point estimate responses. Table 2 shows that the 2012 interest rate sensitivity coefficient interactions are in fact positive and significant in the second-stage regressions. With these definitions of interest ratesensitivity, thezerolowerboundrecoveryisactuallystrongerthanpreviousrecoveries in interest rate sensitive industries.1 The ordering of the SVARs was chosen in accordance with the literature. However, the index and the second-stage results are unaffected by ordering. Specifying the SVARs as normally recursively identified VARs, where industry revenues affect the macro variables, results in a highly correlated index. Furthermore, rearranging the ordering of the VARs leaves all results unchanged. Table 3 presents the index correlations resulting from different orderings, while Table 4 presents the second stage estimates. Table 5 presents correlations of the indices with a number of other specifications of the SVARs, while Table 6 presents the second-stage results. Column (2) includes indicators for each quarter to control for any additional seasonality in the year over year growth rates. Column (3) omits commodity prices from the SVARs. Column (4) runs the SVARs and second-stage estimates including the three industries that are omitted in the paper due to the fact that they have few firm-observations per quarter (Figure 2) and resulting 1Part of this is due to the fact that NAICS 52, Finance and Insurance, is ranked as less interest rate sensitive using the alternative measures, as this industry has an initial increase in revenues in response to a contractionary monetary shock. Since this industry did poorly following the financial crisis, ranking it as less sensitive makes the 2012 index interaction coefficient more positive. 1

noisy time series (Figures 3). Column (5) constrains industry revenues to not affect the macroeconomic aggregates, and with industry revenues ordered last. In all cases, the industry indices are highly correlated, and the second-stage estimates are largely unchanged. It is thought the transmission mechanism of monetary policy has changed through time, particularly since the early 1980s. In evaluating the transmission mechanism, Barth and Ramey (2002) find evidence that the cost channel of monetary policy decreased after the Volcker period. During the 1970s, the financial structure of credit markets was affected by deregulation and and the regional de-segmentation of banking markets Friedman (1986). The US exchange rate regime also moved from fixed to floating. At times before the 1980s, the Federal Reserve sought to directly limit the amount of bank lending (Romer and Romer, 1989). Finally, the Fed also officially targeted monetary aggregates rather than interest rates from 1979-1982. In spite of evidence for changes in the monetary transmission mechanism, Table 5 shows that the index is still correlated (ρ = 0.66) when using only the post-1988:Q1 subsample for VAR index construction. Column (6) of Table 6 also shows the second-stage results are unaffected. In fact, using only more recent data for the SVAR index suggests that interest rate sensitive industries had greater revenue growth during the recovery at the zero lower bound.2 As a last check on the VAR results, the 2012 period compares favorable regardless of lag length selections, though the magnitude and significance of the index coefficient do fluctuate (Tables 7 and 8). In conclusion, the VARs are remarkably robust across all specifications in the outcome of the industry ordering, and the second-stage results. Further Second-Stage Robustness The results hold using a balanced panel of firms, that is, using firms that exist in all recovery years. This alleviates concerns associated with bias due to firm entry and exit. As shown in Table 9, the coefficient on the index from 2009-2012 is not statistically different from zero. One possible criticism of all regressions presented thus far is that firm-level data is examined while the index varies at only the industry level. Since only 19 industries are used, the effective sample size of the differences-in-differences estimates is more closely approximated as 19 for each recovery, rather than thousands of firms. Although standard errorsareclusteredattheindustrylevel, thesmallnumberofclustersmeansthatclustered 2This is partly due to the fact that Finance and Insurance is ranked much less sensitive in the post-1988 subsample, and this industry, as expected, fared poorly following the financial crisis. 2

standard errors may overstate significance. One way to evaluate whether this problem has merit is to take industry averages, and completely ignore the firm dimension in the differences-in-differences specification. Table 10 presents the results from this process. The coefficients on the index recovery interactions give similar estimates. In addition, the R2 increases to 70 − 73% from the 7−8% observed in baseline firm-level regressions. The 2009-2012 results compare favorably to previous recoveries when dropping industries one at a time (Table 11), though dropping Mining, Quarrying, and Oil and Gas Extraction (21) does decrease the zero lower bound coefficient. However, it is still not statistically different from previous recoveries. In addition, it is of note that when dropping the financial industry (NAIC 52), the 2009-2012 period is characterized by significantly greater revenue growth in interest rate sensitive industries. This is notable, as we would expect the financial industry to recover slowly after the financial crisis; one could make the case that dropping the financial industry would be more sensible for the entire paper. In that case, most regressions would show increased growth of interest rate sensitive industries at the zero lower bound. VAR Results for Identifying the Policy Stance and Industry Effects of Fiscal Policy Figure 4 presents the industry responses to a monetary shock from VARs with cumulative Romer and Romer shocks in place of the federal funds rate, as described in Section 7 of the paper. These VARs are used for identifying the policy stance and for identifying industry senstivity to fiscal policy. Macro aggregate responses, shown in Figure 5, are as expected, with the exception of counterintuitive responses to tax shocks.3 However, the estimated Kalman filtered R&R shocks are similar when omitting tax revenues from the VAR. 4 Figure 6 presents the industry responses to a government spending shock, while Table 12 presents the resulting index based on Figure 6’s standardized cumulative impulse responses. 3VARs with government spending, tax receipts, and GDP growth over this time period result in similar impulse responses. 4Ilzetzki (2011) also finds similar results whether controlling or not for tax revenues in VARs estimating effects of fiscal spending. 3

Table 1: Cross-correlation Table: Alternative Index Construction Baseline 1.00 Sirf 7 quarters 0.93 1.00 Csirf 16 quarters 0.88 0.96 1.00 Sirf 7 quarters refers to the 7 quarter structural impulse response point estimate. Csirf 16 quartersreferstothe16quartercumulativestructural impulseresponse. Table 2: Alternative Index Construction (1) (2) (3) Baseline Sirf7 Csirf16 Index 0.885 -0.880 -1.456 (2.653) (2.898) (3.008) Index·2012 2.617 5.446* 6.867** (3.096) (2.890) (2.914) 1991-1994 20.181*** 20.832*** 21.306*** (3.904) (3.827) (4.123) 2001-2004 12.699** 13.571** 13.979** (5.717) (5.942) (5.741) 2009-2012 15.496* 14.700* 14.180 (8.211) (8.391) (9.166) Observations 16,480 16,480 16,480 Adjusted R-squared 0.068 0.068 0.068 ***p<0.01,**p<0.05,*p<0.1. Industryclusteredstandarderrorsinparentheses. Baselinerefersto16quarterminimumresponseasdescribedintext. Csirfandsirfindicatecumulativestructuralandstructuralimpulseresponses respectively. Table 3: Cross-correlation Table: Different VAR Orderings Baseline GDP FF P FF GDP P P FF GDP Baseline 1.00 GDP FF P 0.91 1.00 FF GDP P 0.93 0.90 1.00 P FF GDP 0.93 0.82 0.98 1.00 AllVARscontainfourlagsandaconstant. GDPreferstorealgrossdomesticproduct growth and industry revenue block, P refers to price block, and FF refers to federal fundsrate. 4

Table 4: Real Revenue Growth, Different VAR Orderings (1) (2) (3) (4) Baseline GDP FF P FF GDP P P FF GDP Index 0.885 0.348 0.155 0.030 (2.653) (2.759) (2.468) (2.420) Index·2012 2.617 3.437 6.451** 6.840*** (3.096) (3.136) (2.267) (2.370) 1991-1994 20.181*** 20.501*** 20.354*** 20.476*** (3.904) (4.131) (3.978) (4.092) 2001-2004 12.699** 13.115** 12.999** 13.125** (5.717) (5.722) (5.775) (5.875) 2009-2012 15.496* 16.908* 13.001* 11.095 (8.211) (8.906) (7.361) (7.189) Observations 16,480 16,480 16,480 16,480 Adjusted R-squared 0.068 0.068 0.069 0.069 *** p<0.01, ** p<0.05, * p<0.1 ***p<0.01,**p<0.05,*p<0.1. Industrystandarderrorsclusteredinparentheses. Table 5: Cross-correlation Table: Alternative Specifications Baseline Qrtr Ind No Com Post-1988 Baseline 1.00 Quarter Indicators 1.00 1.00 No Commodity Prices 0.98 0.97 1.00 Post-1988 0.66 0.64 0.65 1.00 5

snoitacfiicepS evitanretlA :6 elbaT )5( )4( )3( )2( )1( 1Q8891 tsoP srotceS llA secirP ytidommoC oN seimmuD retrauQ enilesaB 600.0 091.1- 953.1 407.0 588.0 xednIi )364.2( )359.2( )395.2( )586.2( )356.2( ***101.6 746.2- 050.5 135.1 716.2 2102·xednI )010.2( )212.3( )069.2( )091.3( )690.3( ***523.02 ***504.02 ***992.02 ***891.02 ***181.02 4991-1991 )901.4( )800.4( )908.3( )319.3( )409.3( **179.21 **496.21 **209.21 **417.21 **996.21 4002-1002 )628.5( )236.5( )104.5( )067.5( )717.5( **232.71 *865.61 583.31 *783.61 *694.51 2102-9002 )250.8( )258.8( )249.7( )675.8( )112.8( 084,61 436,61 084,61 084,61 084,61 snoitavresbO 960.0 860.0 960.0 760.0 860.0 derauqs-R detsujdA .sesehtnerapnideretsulcsrorredradnatsyrtsudnI .1.0<p*,50.0<p**,10.0<p*** 6

Table 7: Cross-correlation Table: Alternative Lag Lengths Baseline 2 Lags 3 Lags 5 Lags Baseline 1.00 2 Lags 0.94 1.00 3 Lags 0.96 0.96 1.00 5 Lags 0.93 0.90 0.93 1.00 7

Table 8: Real Revenue Growth, Alternative Lag Lengths Baseline (4 Lags) 2 Lags 3 Lags 5 Lags Index 0.885 0.337 0.906 -0.431 (2.653) (2.369) (2.792) (2.347) Index·2012 2.617 5.450** 3.513 0.958 (3.096) (2.369) (2.917) (2.802) 1991-1994 20.181*** 20.500*** 20.249*** 20.323*** (3.904) (3.845) (3.739) (4.192) 2001-2004 12.699** 13.134** 12.806** 13.090** (5.717) (5.688) (5.601) (5.707) 2009-2012 15.496* 14.167* 15.366 17.470* (8.211) (8.051) (8.925) (9.813) Observations 16,480 16,480 16,480 16,480 Adjusted R-squared 0.068 0.069 0.068 0.067 *** p < 0.01, ** p < 0.05, * p < 0.1. Standard errors clustered at NAIC 2 digit level in parentheses. Controlvariablecoefficientsomittedfromtable. Table 9: Real Revenue Growth, Balanced Subsample (1) (2) (3) (4) Index 1.927 2.985 2.689 3.058 (3.025) (3.362) (3.111) (2.674) Index·2012 1.450 1.129 1.299 1.181 (2.170) (2.116) (2.042) (2.082) 1991-1994 22.126*** 13.623** 20.162*** 27.410*** (2.604) (4.952) (4.572) (4.169) 2001-2004 10.778*** 7.537 7.480 14.600* (2.983) (9.212) (9.472) (7.028) 2009-2012 8.862*** 5.900 5.592 12.620** (2.707) (5.660) (6.067) (5.398) Observations 5,185 5,185 4,965 4,965 Adjusted R-squared 0.107 0.111 0.113 0.118 *** p < 0.01, ** p < 0.05, * p < 0.1. Standard errors clustered at NAIC 2 digit level in parentheses. Control variable coefficients omitted from table. Column (2) controls for sectoral GDP elasticity. Column (3) controls for initial firm employment. Column (4) controlsforindustry-specificgrowthtrends. 8

Table 10: Real Revenue Growth, Industry-Level Data (1) (2) (3) (4) Index 1.087 1.398 1.615 1.643 (1.701) (1.774) (1.829) (1.857) Index·2012 2.762 2.344 2.233 2.260 (2.947) (3.073) (3.228) (3.267) 1991-1994 20.554*** 16.403*** 9.882 10.328 (2.342) (4.485) (7.427) (7.977) 2001-2004 13.827*** 13.630*** 14.007* 14.196* (2.342) (4.485) (7.812) (7.979) 2009-2012 15.671*** 16.424*** 15.503* 15.679* (2.342) (4.524) (7.859) (8.014) Observations 57 57 57 57 Adjusted R-squared 0.731 0.721 0.711 0.705 ***p<0.01,**p<0.05,*p<0.1. StandarderrorsclusteredatNAIC2digitlevelin parentheses. Control variable coefficients omitted from table. Column (2) controls for sectoral GDP elasticity. Column (3) controls for initial firm employment. Column (4) controlsforindustry-specificgrowthtrends. 9

emiT a ta enO depporD seirtsudnI :htworG euneveR laeR :11 elbaT 18 27 17 26 65 55 45 35 25 15 54 44 24 33 23 13 32 22 12 llA yrtsudnI 88.0 80.1 98.0 09.2 68.0 98.0 14.0 36.0 79.1 67.0 27.0 13.1 36.0 03.1 53.1 38.0 16.0 79.1- 01.0 98.0 xednI )237.2( )887.2( )756.2( )820.2( )166.2( )356.2( )377.2( )696.2( )326.2( )085.2( )756.2( )737.2( )136.2( )058.2( )155.2( )837.2( )924.3( )731.2( )355.2( )356.2( 17.2 25.2 16.2 51.1 36.2 26.2 13.3 27.2 *51.5 48.3 78.2 84.2 18.2 09.0 04.2 19.2 92.3 58.1 40.0- 26.2 2102·xednI )511.3( )801.3( )401.3( )928.2( )211.3( )690.3( )203.3( )190.3( )496.2( )933.3( )431.3( )201.3( )351.3( )681.3( )789.2( )070.3( )802.4( )769.2( )493.2( )690.3( 014,61 241,61 273,61 581,61 541,61 084,61 957,51 310,61 719,31 868,41 002,61 230,61 098,51 353,21 746,41 738,51 952,61 417,51 808,51 084,61 snoitavresbO 70.0 70.0 70.0 70.0 70.0 70.0 70.0 70.0 80.0 70.0 70.0 70.0 60.0 60.0 70.0 70.0 70.0 70.0 70.0 70.0 derauqs-R .noissergermorfdeppordyrtsudniCIANetonedsgnidaehnmuloC .sesehtnerapnisrorredradnatsderetsulcyrtsudnI .1.0<p*,50.0<p**,10.0<p*** xednI ytivitisneS gnidnepS tnemnrevoG RAVS CIAN tigiD owT :21 elbaT rotceS xednI lacsiF edoC tigiD owT noitcartxE saG dna liO dna ,gniyrrauQ ,gniniM 94.2- 12 seitilitU 18.1- 22 noitaerceR dna ,tnemniatretnE ,strA 33.1- 17 gnirutcafunaM 86.0- 13 edarT elaselohW 84.0- 24 )noitartsinimdA cilbuP tpecxe( secivreS rehtO 42.0- 18 gnisuoheraW dna noitatropsnarT 60.0- 84 gnirutcafunaM 21.0 23 gnisaeL dna latneR dna etatsE laeR 61.0 35 noitamrofnI 61.0 15 noitcurtsnoC 42.0 32 gnirutcafunaM 82.0 33 ecnarusnI dna ecnaniF 35.0 25 secivreS noitaidemeR dna tnemeganaM etsaW dna troppuS dna evitartsinimdA 46.0 65 secivreS lacinhceT dna ,cfiitneicS ,lanoisseforP 27.0 45 edarT liateR 87.0 54 secivreS dooF dna noitadommoccA 60.1 27 ecnatsissA laicoS dna eraC htlaeH 61.1 26 edarT liateR 32.1 44 evitisopevahevitisoptsomdeknarseirtsudniehT .skcohslacsfiyranoisnapxeotsesnopsereuneverevitagenevahevitagentsomdeknarseirtsudniehT .skcohslacsfiyranoisnapxeotsesnopsereunever 10

Figure 1: Quarterly Industry Revenues: 2005Q1-2008Q3. snoitavresbO ylretrauQ fo rebmuN 051 001 05 0 0 .02 .04 .06 .08 Industry Revenues as a Fraction of Gross Output This figure shows that individual industry revenues tend to be a small portion of aggregate revenues. Sources: Compustat, Bureau of Economic Analysis, and author’s calculations. 11

Figure 2: Average Number of Quarterly Firm Observations by North American Industrial Classification Code, 1970:Q1-2008:Q3 snoitavresbO mriF fo rebmuN 005,1 000,1 005 0 11 21 22 23 31 32 33 42 44 45 48 49 51 52 53 54 56 61 62 71 72 81 North American Industrial Classification Code Source: Compustat and author’s calculations. 12

seireS emiT yrtsudnI detcurtsnoC :3 erugiF 2. 0 2.- 4. 2. 0 2.- 2.1. 01.- 4. 2. 0 2.- 4. 2. 0 2.- 51.1.50.050.- 2. 0 2.-4.- 3.2.1.01.- 4. 2. 0 2.- 2.1. 01.-2.- 4.3.2.1.0 6.4.2.02.- 4.2.02.-4.- 2.1. 01.-2.- 4. 2. 0 2.- 3.2.1.01.- 4. 2. 0 2.- 4.2.02.-4.- 2. 1. 0 1.- 4. 2. 0 2.- 3.2.1. 01.- 3.2.1.01.- 13 32 22 12 11 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 54 44 24 33 23 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 35 25 15 94 84 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 17 26 16 65 45 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 18 27 3Q:8002 1Q:0991 1Q:1791 3Q:8002 1Q:0991 1Q:1791 htworG euneveR retrauQ sciaN tigiD owT yb shparG 13

Figure 4: Industry impulse responses to one standard deviation R&R shocks. The x-axes denote quarters. 90% confidence intervals using circular moving bootstrap outlined in blue. Full specification described in text. 2 1 0 1- 2- 3- Mining, Quarrying, etc. (21) 0 5 10 15 1 5. 0 5.- 1- Utilities (22) 0 5 10 15 2 1 0 1- 2- Construction (23) 0 5 10 15 1 5. 0 5.- 1- Manufacturing (31) 0 5 10 15 1 5. 0 5.- 1- Manufacturing (32) 0 5 10 15 2 1 0 1- 2- Manufacturing (33) 0 5 10 15 2 1 0 1- 2- Wholesale Trade (42) 0 5 10 15 1 5. 0 5.- 1- Retail Trade (44) 0 5 10 15 2 1 0 1- 2- Retail Trade (45) 0 5 10 15 1 5. 0 5.- 1- 5.1- Transportation and Warehousing (48) 0 5 10 15 5.1 1 5. 0 5.- 1- Information (51) 0 5 10 15 2 1 0 1- 2- Finance and Insurance (52) 0 5 10 15 5.1 1 5. 0 5.- 1- Real Estate, Rental, etc. (52) 0 5 10 15 1 5. 0 5.- 1- Professional Services etc. (54) 0 5 10 15 5.1 1 5. 0 5.- 1- Administrative etc. (56) 0 5 10 15 5.1 1 5. 0 5.- 1- Health and Social Assistance (62) 0 5 10 15 2 1 0 1- 2- Arts, Entertainment and Rec. (71) 0 5 10 15 1 0 1- 2- Accomodation etc. (72) 0 5 10 15 2 1 0 1- 2- Other Services (81) 0 5 10 15 14

Figure 5: Aggregate impulse responses to one standard deviation shocks using R&R Shocks. The x-axes denote quarters. 90% confidence intervals using circular moving bootstrap outlined in blue. Charts labeled by impulse, response. Full specification described in text. 1 5. 0 Govt Cons, Govt Cons 0 5 10 15 5. 0 5.- Govt Cons, Taxes 0 5 10 15 4. 2. 0 2.-4.- Govt Cons, NFC Output 0 5 10 15 2. 0 2.-4.-6.- Govt Cons, PCE Inflation 0 5 10 15 2.51.1.50.050.- Govt Cons, R&R Shocks 0 5 10 15 2. 0 2.- 4.- Taxes, Govt Cons 0 5 10 15 25.11 5. 0 5.- Taxes, Taxes 0 5 10 15 1 5. 0 5.- Taxes, NFC Output 0 5 10 15 4. 2. 0 2.- Taxes, PCE Inflation 0 5 10 15 1.50. 050.-1.- Taxes, R&R Shocks 0 5 10 15 4. 2. 0 2.- NFC Output, Govt Cons 0 5 10 15 1 5. 0 5.-1- NFC Output, Taxes 0 5 10 15 5.11 5. 05.-1- NFC Output, NFC Output 0 5 10 15 4. 2. 0 2.-4.- NFC Output, PCE Inflation 0 5 10 15 51.1.50.050.-1.- NFC Output, R&R Shocks 0 5 10 15 2. 1. 0 1.-2.- PCE Inflation, Govt Cons 0 5 10 15 5. 0 5.- 1- PCE Inflation, Taxes 0 5 10 15 5. 0 5.- 1- PCE Inflation, NFC Output 0 5 10 15 6. 4. 2. 02.- PCE Inflation, PCE Inflation 0 5 10 15 2. 1. 0 1.-2.- PCE Inflation, R&R Shocks 0 5 10 15 4. 2. 0 2.- R&R Shocks, Govt Cons 0 5 10 15 1 5. 0 5.- 1- R&R Shocks, Taxes 0 5 10 15 5. 0 5.- 1- R&R Shocks, NFC Output 0 5 10 15 0 2.- 4.- 6.- R&R Shocks, PCE Inflation 0 5 10 15 6. 4. 2. 0 R&R Shocks, R&R Shocks 0 5 10 15 15

ehT .3Q:8002-1Q:1791 ,kcohs gnidneps tnemnrevog noitaived dradnats eno ot sesnopser dezilanogohtro yrtsudnI :6 erugiF .eulb ni deniltuo slavretni ecnedfinoc %09 citotpmyssA .sretrauq etoned sexa-x 2 0 2- 4- 6- )12( .cte gniyrrauQ ,gniniM 51 01 5 0 5. 0 5.- 1- 5.1- )22( seitilitU 51 01 5 0 2 1 0 1- 2- )32( noitcurtsnoC 51 01 5 0 1 5. 0 5.- 1- )13( gnirutcafunaM 51 01 5 0 1 5. 0 5.- 1- )23( gnirutcafunaM 51 01 5 0 1 5. 0 5.- 1- )33( gnirutcafunaM 51 01 5 0 1 5. 0 5.- 1- 5.1- )24( edarT elaselohW 51 01 5 0 8. 6. 4. 2. 0 2.- )44( edarT liateR 51 01 5 0 5.1 1 5. 0 5.- )54( edarT liateR 51 01 5 0 1 5. 0 5.- 1- 5.1- )84( gnisuoheraW dna noitatropsnarT 51 01 5 0 5.1 1 5. 0 5.- 1- )15( noitamrofnI 51 01 5 0 1 5. 0 5.- 1- 5.1- )25( ecnarusnI dna ecnaniF 51 01 5 0 1 5. 0 5.- 1- )35( .cte ,latneR ,etatsE laeR 51 01 5 0 1 5. 0 5.- 1- )45( .cte secivreS lanoisseforP 51 01 5 0 2 1 0 1- )65( .cte evitartsinimdA 51 01 5 0 5.1 1 5. 0 5.- )26( ecnatsissA laicoS dna eraC htlaeH 51 01 5 0 2 1 0 1- 2- )17( .ceR dna ,tnemniatretnE ,strA 51 01 5 0 1 5. 0 5.- 1- 5.1- )27( .cte noitadommoccA 51 01 5 0 5.1 1 5. 0 5.- 1- )18( secivreS rehtO 51 01 5 0 16

References Barth, M. J. and Ramey, V. A. (2002). The Cost Channel of Monetary Transmission. In NBER Macroeconomics Annual 2001, Volume 16, NBER Chapters, pages 199–256. National Bureau of Economic Research, Inc. Friedman, B. M. (1986). Money, credit, and interest rates in the business cycle. In The American business cycle: Continuity and change, pages 395–458. University of Chicago Press. Ilzetzki, E. (2011). Fiscal policy and debt dynamics in developing countries. World Bank Policy Research Working Paper Series, Vol. Romer, C. D. and Romer, D. H. (1989). Does monetary policy matter? a new test in the spirit of friedman and schwartz. In NBER Macroeconomics Annual 1989, Volume 4, pages 121–184. MIT Press. 17