Fiscal policy and collateral constraints in an estimated DSGE model:

Transcription

1 U N I V E R S I T Y O F C O P E N H AGEN D E P A R T M E N T O F E C O N O M I C S F A C U L T Y O F S O C I A L S C I E N C E S Master Thesis Rasmus Bisgaard Larsen & Goutham Jørgen Surendran Fiscal policy and collateral constraints in an estimated DSGE model: Do collateral constraints amplify or weaken fiscal policy? Supervisor: Assistant Professor Søren Hove Ravn, PhD ECTS credits: 3 Date of submission: 1/8/16 Key strokes: 77,733

2 Summary The development of the United States housing market during the s has lead to an increased academic interest in the housing market. Economists have studied how the housing market affects the propagation of macroeconomic shocks through the economy, whether fluctuations in the housing market are just a consequence of general macroeconomic fluctuations or not, and if the housing market itself is a driver of the business cycle. Another subject that has received renewed interest is fiscal policy. This thesis bridges the gap between these two research agendas by analyzing how collateral constraint tied to housing values influence the propagation of fiscal policy shocks. We analyze the effects of collateral constraints on fiscal policy by developing a New Keynesian dynamic stochastic general equilibrium (DSGE) model that includes a housing market with a collateral constraint. The model features a rich fiscal policy block with distortionary taxes on consumption, labor, capital and housing as well as lump-sum transfers to households and government spending. This allows us to investigate the effects of multiple fiscal policy actions. The fiscal instruments react endogenously to output and government debt. Moreover, the model features a measure of unemployment as we extend the standard formulation of staggered wage setting with a relationship between the wage markup and unemployment. The model is estimated on a sample of quarterly U.S. data covering the period of 1985Q1-7Q4 by using Bayesian techniques. The sample includes conventional macroeconomic aggregates as well as fiscal variables. We show that the model implies that some fiscal expansions are accelerated by the collateral constraint, while others are decelerated. Fiscal shocks that cause house prices to increase are accelerated, while shocks that cause house prices to decrease are decelerated. Hence, the collateral constraint weakens the expansive effect on output from government spending shocks and cuts in the capital or consumption tax rates, while increased lump-sum transfers and cuts in labor or housing tax rates have a larger effect on output because of the collateral constraint. These effects are especially large when collateral constraints are loose and households can borrow against a large share of their housing wealth. In our estimated model, however, the quantitative effects of the collateral constraint are relatively small for the shocks to government spending and the consumption tax rate compared to the constraint s effects on the other fiscal instrument. The effect of the collateral constraint on the present value multipliers for these two fiscal instruments is also small, and the effect on short-run and long-run multipliers differs. We wish to thank our supervisor Assistant Professor Søren Hove Ravn for providing excellent guidance during the process of writing this thesis. His comments have been constructive and good-natured throughout the last six months. We also thank Danmarks Nationalbank for providing an office and the economists at the Department of Economics and Monetary Policy for useful comments on our thesis. We especially thank Head of Economic Research Kim Abildgren. In addition, we are grateful to Lars Sparresø Merklin and Bjørn Bjørnsson Meyer for comments on our drafts. The usual disclaimer applies. i

3 The stimulative effects of fiscal policy are evaluated by calculating present value multipliers for each fiscal instrument based on the posterior distribution of the parameters. We show that a shock to government spending raises output almost one-to-one on impact and that it is the most stimulative fiscal instrument on impact. The output effect of a government spending shock decreases across the horizon, while the effects of lump-sum transfers and cuts in the tax rates take time to build: lumpsum transfers and cuts in the labor, consumption and housing tax rates have their largest effect on output after about 1 year, while a capital tax rate cut becomes more stimulative over a longer horizon. While collateral constraints in the housing market affect the transmission of fiscal policy shocks, the housing market does not contribute much to fluctuations in either output or fiscal policy in our estimated model. Instead, output fluctuations are mostly driven by shocks to the wage markup, monetary policy and productivity. We show this by performing a forecast error variance decomposition of the model. We conduct a series of counterfactual experiments to analyze the sensitivity of the government spending multiplier to 1) the stance of monetary policy, ) whether the government adjusts distortionary tax rates or not following a government spending shock, and 3) how much the distortionary tax rates react to government debt. When monetary policy reacts less to output or inflation, the present value government spending multiplier increases at all horizons. Similarly, a higher degree of interest rate smoothing increases the government spending multiplier. The financing decisions of the government also play an important role in determining the size of the government spending multiplier as the expected path of the tax rates shape the response of rational agents following a government spending shock to the economy. The estimated model allows us to quantify the contribution of fluctuations in output and government debt to changes in tax rates. We argue how major tax reforms during our sample period were driven by either output stabilization motives, to stabilize government debt or exogenous shocks. While the estimated model suggests that the tax reforms during the administrations of President George H.W. Bush (Sr.) and President Bill Clinton were largely driven by either output or debt stabilization motives, the model attributes the tax cuts during the presidency of President George W. Bush (Jr.) to exogenous shocks. These findings are related to the narrative analysis by Romer and Romer (1). Finally, we discuss various theoretical aspects of our model. First, we discuss how some of the theoretical explanations for the comovement between consumption and government spending often found in the data alternative utility functions and rule-of-thumb households have trouble generating an increase in both house prices and consumption following a boost to government spending. A central bank that accommodates government spending shocks, however, can generate an increase in both house prices and consumption when government spending is increased. Second, we discuss how assuming that the collateral constraint always binds might affect the ii

4 transmission of fiscal policy in contrast to a model, wherein the collateral constraint is occasionally binding. We also discuss the importance of another occasionally binding constraint that has received much attention in monetary policy analysis: the zero lower bound on nominal interest rates. Third, we compare the formulation of the labor market in our model with the search and matching frictions in a related model constructed by Andrés et al. (15). iii

5 Contents 1 Introduction 1 What is the empirical evidence on fiscal policy shocks? 3.1 The SVAR approach The narrative approach Reconciling SVARs with narrative analyses Fiscal policy and the housing market Theoretical models of the propagation of government spending shocks Non-Ricardian households Deep habit formation Non-separable utility Government spending in the utility function Fiscal-monetary interactions Model Households The labor market Wholesale firms Retail firms Fiscal authority Monetary policy Market clearing conditions Estimation Data Measurement equations Calibrated parameters Prior distributions Posterior distributions Application Transmission of fiscal policy shocks Fiscal multipliers Fiscal multipliers and financial conditions Variance decomposition 73 8 Counterfactual experiments of financing government spending Lump-sum versus distortionary financing Debt versus tax financing Government spending and monetary policy 84 1 Analysis of tax reforms 86 iv

6 11 Discussion: The response of consumption and house prices to government spending shocks 89 1 Discussion: Occasionally binding collateral constraints and the zero lower bound Discussion: Comparison with Andrés et al. s (15) model Conclusion 1 References 14 Appendix 111 A Model derivations 111 A.1 Derivation of savers first-order conditions A. Derivation of the borrowers first-order conditions A.3 Derivation of steady state solution A.4 Log-linearization of the model A.5 Derivations of the New Keynesian Wage Phillips Curves B Data sources 131 C Posterior and prior distributions 13 D Figures and tables 134 D.1 Impulse response functions and tables D. Government spending with/without detailed fiscal block D.3 Government spending under different φ R s D.4 Government spending and monetary accommodation D.5 Government spending and nominal wage rigidities Individual contributions Rasmus Bisgaard Larsen: Sections, 4., 4.4, 4.5, 5.4, 6.3, 7, 1, 11 and 1 as well as text on Bayesian inference in section 5 and introduction to section 6 before impulse response functions. Goutham Jørgen Surendran: Sections 3, 4.1, 4.3, 4.6, 4.7, , 5.5, 6.1, 6., 8, 9, 13 as well as introduction to section 4. Collectively written: Summary and sections 1 and 14. v

7 1 Introduction Housing wealth constitutes a significant share of households wealth in the United States. According to table 1 below, the aggregate housing wealth of households was around 31.7 trillion dollars by the end of 15. This was about a third of the households net worth and substantially larger than the total GDP of about 17.9 trillion dollars in the same year. In addition, the United States experienced a prolonged and large boom in house prices prior to the Great Recession. These phenomena are not unique to the United States: the housing market is relatively large in many economies, while the house price boom during the s was exceptionally large and synchronized across countries (André, 1). Table 1: Household wealth in the United States, 15 Households balance sheet, 15 billion dollars A Assets 11,769.6 B Real estate (owner-occupied homes) 5,9.5 C Residential real estate of noncorporate businesses (rented homes) 6,368.4 D Other tangible assets 5,7.5 E Financial assets less residential real estate of noncorporate businesses 64,41. F Liabilities 14,5. G Household net worth (A-F) 87,49.6 H Housing wealth (B+C) 31,658.9 I Non-housing wealth (D+E-F) 55,59.7 Note: The source is the Z.1 Financial Accounts of the United States (downloaded on 14 July 16 from federalreserve.gov) and the balance items have been calculated by using the definitions by Iacoviello (1). The Z.1 table entries are: assets (B.11:1), real estate (B.11:3), residential real estate of noncorp. businesses (B.14:4), other tangible assets (B.11: less B.11:3), financial assets less residential real estate of noncorp. businesses (B.11:9 less B.14:4) and liabilities (B.11:31). The size of the housing market relative to the rest of the economy as well as the recent cycle in house prices has lead many to investigate how the housing market affects the business cycle. Notably, economists have studied how collateral constraints tied to housing values influence macroeconomic fluctuations. Prominent contributions to this research area, amongst others, include Iacoviello s (5) development of a model with collateral constraints for monetary policy analysis, the analysis of the sources of U.S. housing market fluctuations as well as its spillover effects on the wider economy by Iacoviello and Neri (1) and Gerali et al. s (1) model with an explicit formulation of the banking sector. In this thesis, we contribute to this strand of the macroeconomic literature by analyzing how collateral constraints tied to households housing wealth affect the transmission of fiscal policy. This is done by constructing and estimating a DSGE model featuring collateral constrained households and a rich fiscal policy block with several distortionary 1

8 taxes that react endogenously to output and government debt. The housing market is based on the model by Iacoviello and Neri (1), while the fiscal policy block is based on Zubairy s (14) model. In addition, we embed the theory of unemployment by Galí (11a) into the model to study fiscal policy s effect on unemployment. We estimate the model using quarterly U.S. data on standard macroeconomic variables covering the period 1985Q1-7Q4, while also using fiscal variables such as government spending, tax rates, transfers and government debt as observables. In the context of a model with collateral constraints, relatively few authors have studied fiscal policy (see e.g. Andrés et al. (15), Andrés et al. (16), Bermperoglou (15), Khan and Reza (14) and Callegari (7)), fewer have studied distortionary taxes and as far as we know none have estimated a DSGE model that focuses on both fiscal policy and collateral constraints with Bayesian techniques. We find that collateral constraints do not have an uniform effect on the stimulative effects of a fiscal expansion. The effect on output can either be amplified or weakened depending on which fiscal instrument is used to stimulate the economy. Whether a fiscal shock is amplified or weakened depends on how house prices react in response to the shock. This is because higher house prices will increase households housing wealth and enhance the borrowing capability of collateral constrained households, which stimulates their consumption. Vice versa, fiscal shocks that cause house prices to fall will force collateral constrained households to deleverage and decrease consumption. Thus, expansionary fiscal policy that boosts house prices higher transfers to households and cuts in housing or labor taxes will be accelerated by the collateral constraint, while expansionary fiscal policy that decrease house prices higher government spending and lower capital or consumption taxes is decelerated. Collateral constraints especially have a large effect on the transmission of fiscal policy when households can borrow against a large share of their housing wealth. The quantitative effects in our estimated model, however, of the collateral constraint are relatively small for the shocks to government spending and the consumption tax rate compared to the constraint s effects on the other fiscal instrument. Moreover, the effect of the collateral constraint on the present value multipliers for these two fiscal instruments is small, and the effect on short-run and long-run multipliers differs. In order to quantify the stimulative effects of fiscal policy, we calculate present value fiscal multipliers based on the posterior distribution of the parameters. Especially government spending is stimulative in the short run and the posterior mean multiplier is.94 on impact. A tax cut that decreases total tax revenues by 1 % has a posterior mean impact multiplier of.6 if it is driven by a cut in the consumption tax rate, while the posterior mean impact multiplier is.5 if the tax cut is driven by either cuts to the labor tax rate or the capital tax rate. The posterior mean of the housing tax multiplier is.1 on impact, while the transfers multiplier is.9. However, while the government spending multiplier is largest on impact, a capital tax cut is more stimulative over a longer horizon

9 with a posterior mean present value multiplier of 3 over 5 years. The remaining fiscal instruments have a maximum effect on output after about 1 year. This highlights that it is important to evaluate multipliers both in the short run and the long run. We show how the stance of monetary policy affects the transmission of government spending shocks and that more accommodative monetary policy increases the stimulative effects of government spending. This is something that has gained significant attention after the Great Recession as central banks in many Western countries lowered interest rates towards the zero lower bound (see e.g. Christiano et al. (11), Woodford (11), Davig and Leeper (11), Mertens and Ravn (14b) and Zubairy (14)). We also show how the financing decisions of the government have consequences for the size of the government spending multiplier, something which has also been analyzed by Leeper et al. (1) and Zubairy (14) amongst others. The thesis is structured as follows. In the next two sections, we review the empirical literature on fiscal policy as well as the theoretical literature on the propagation of government spending shocks. The model is presented in section 4. Section 5 presents the data, details on the Bayesian estimation procedure as well as the prior and posterior distributions. Section 6 shows impulse response functions, fiscal multipliers and an analysis of how financial conditions affect multipliers. A forecast error variance decomposition is shown in section 7. Sections 8 and 9 explore alternative specifications for financing government spending and monetary-fiscal interactions respectively. Section 1 contains an analysis of historical tax reforms in the sample period. Discussions of various model features are in sections 11 to 13. Finally, section 14 concludes. What is the empirical evidence on fiscal policy shocks? The empirical analyses of the impact of fiscal policy and the size of fiscal multipliers mainly follow two approaches to identify fiscal policy shocks. One approach relies on restrictions on structural vector autoregressions (SVARs), while the other employs a narrative method to identify exogenous changes in fiscal policy. As discussed by Ramey (11a) in her review of the literature on fiscal output multipliers, estimates of the multiplier span the range of for the government spending multiplier, while it is between -.5 and -5. for the tax multiplier. 1 The meta-analysis of Gechert (15) indicates that the government spending multiplier is about 1 on average, while the average tax multiplier is about -.6 to -.7. Thus, there is a consensus on the sign of the output multiplier with regards to 1 The interpretation of the government spending multiplier is relatively straightforward: the effect on the level of output of an increase in the level of government spending. The interpretation of the tax multiplier is a bit more complicated but the tax multipliers in this section can generally be interpreted as the effect on the level of output of an increase in the level of total tax revenues. Note that the model-based multipliers we report in section 6 are defined as the increase in the level of output of a decrease in the level of total revenue from distortionary taxes. Hence, the model-based multipliers will have the opposite sign of the multipliers reported in this section. 3

10 both government spending and tax shocks but their sizes are contested. The multiplier can also be calculated in different ways as we explain below, which by itself can lead to different multipliers. The literature is more divided on not only the size but also the sign of the multiplier for other central macroeconomic variables such as consumption, real wages and investment, which typically depends on which identification strategy is used (Ramey, 16). We summarize the main results from important contributions to the empirical literature on fiscal policy shocks and discuss methodological issues below. Unless otherwise stated, all results are from analyses of fiscal multipliers in the United States, which also constitute the bulk of the literature and are relevant for our model, which is estimated by using U.S. data. We have chosen to ignore the growing literature that use microeconometric methods on cross-regional data sets since these papers mostly estimate regional multipliers that are difficult to translate into aggregate multipliers and therefore of limited relevance for our model in section 4 (Ramey, 11a). Comparing estimated multipliers across papers should be done with care as different papers report different types of multipliers. Some report impact multipliers (the withinperiod response of output to a shock), others report peak multipliers (the response of output in the period with the largest response) and more recently some have started reporting present value multipliers (the present value of changes in output divided by the present value of changes in government spending or taxes in response to an initial shock). Cumulative multipliers, which are similar to the present value multiplier but where the changes in the variables are not discounted, are also reported (i.e. a change in output in the distant future has the same weight as the change in output on impact). Finally, some authors distinguish between multipliers depending on the fiscal shocks effect on the government s budget (e.g. an increase in government spending can either by financed by issuing debt or increasing taxes to balance the budget)..1 The SVAR approach The SVAR approach utilizes a VAR model and identifies fiscal policy shocks by imposing restrictions on the structure of the reduced form VAR. Consider, as an example, the baseline VAR model used in the seminal contribution by Blanchard and Perotti (): (.1) q Z t = αd t + Φ s Z t s + U t s=1 For example, Nakamura and Steinsson (14) exploit regional variations in military spending in the United States to estimate what effect an increase in government spending in one region relative to other regions has on the regional output relative to the rest of country. 4

11 Z t = [Y t, G t, T t ] is a vector of the logarithms of output, government spending and net taxes, while Φ s for s = {1,,..., q} are three-dimensional matrices of parameters that allow for the current endogenous variables to respond to lagged, endogenous variables and d t is a deterministic term to account for deterministic trends. The reduced form residuals, U t = [y t, g t, t t ] = Bɛ t, are related to the three structural shocks to output, government spending and taxes, ɛ t = [ɛ Y,t, ɛ G,t, ɛ T,t ], through the matrix B. The structural shocks are typically assumed to be orthogonal to each other and serially uncorrelated, while their distribution is normalized to a standard normal distribution: E[ɛ t ] = E[ɛ t ɛ t] = I E[ɛ t ɛ s] = for s t This decomposition of the reduced form residuals into structural shocks allows for an economic interpretation of the model. Since the structural shocks are independent of the other shocks and independent of the endogenous variables in the model, they can be interpreted as primitive and exogenous shocks to the system (Ramey, 16). The reduced form residuals, instead, carry little economic meaning by themselves since they are functions of all of the structural shocks. While the parameters α and Φ s and the residuals U t are easily estimated, we cannot estimate B without imposing identifying restrictions. The identity E[U t U t] = E[Bɛ t ɛ tb ] = BB imposes 6 restrictions by itself since the matrix is symmetric and has ones in the three diagonal elements but we need to impose 3 further restrictions in order to identify B. Blanchard and Perotti () were some of the first authors to analyze the effects of fiscal policy by using the SVAR approach. They do so by decomposing the reduced form residuals as y t = c 1 t t + c g t + ɛ Y,t g t = b 1 y t + b ɛ T,t + ɛ G,t t t = a 1 y t + a ɛ G,t + ɛ T,t B is identified by using institutional information on a 1 and b 1 and assumptions about the interaction between taxes and spending to identify a and b. Specifically, they assume that there is a fiscal policy lag such that spending shocks do not react to current output shocks (b 1 = ), while an estimate of the elasticity of tax revenues to GDP is used to construct a 1. They consider two types of restrictions on a and b by assuming that either spending shocks react to tax shocks but tax shocks do not react to spending shocks (b and a = ) or the other way around (a and b = ). The SVAR is estimated with U.S. data and they find a peak government spending multiplier of 1.9 and a peak tax 5

12 multiplier of In addition, government spending shocks raise consumption, hours and real wages, while private investment decreases. Their estimates of the multipliers are sensitive, however, to whether the trend is assumed to be deterministic or stochastic, and the inclusion of a stochastic trend yields multipliers of.9 and for government spending and taxes respectively. Other authors have applied the Blanchard-Perotti identification scheme of using shortrun point restrictions. Perotti (5) extends the three-variable Blanchard-Perotti SVAR with inflation and the nominal interest rate and estimates the model on Australian, Canadian, German, UK and U.S. data. He finds that the peak government spending multiplier is only above 1 in the U.S. and Germany, while there are signs of subsample instability because the response of output is more muted after 198 for all countries. Galí et al. (7) finds government spending multipliers of a similar magnitude as Blanchard and Perotti () as well as similar impulse responses. Another approach does not impose point restrictions on the error term structure but only relies on sign restrictions to identify structural shocks. Mountford and Uhlig (9) analyze the effects of government spending and tax shocks in a 1 variable SVAR with 4 different structural shocks: a business cycle shock, a monetary policy shock, a government spending shock and a tax shock. They identify the fiscal policy shocks by imposing 7 sign restrictions on the response of the variables for 4 quarters after a shock (e.g. the interest rate increases following a monetary policy shock) and assuming that the monetary policy, business cycle and fiscal policy shocks are orthogonal. 3 Mountford and Uhlig find a peak government spending multiplier of.65 in response to a deficit-financed spending shock, substantially lower than the estimate by Blanchard and Perotti (). They also find that output increases most on impact and reverts towards its trend whereas Blanchard and Perotti () find that the output response is hump-shaped with output reaching its maximum value after 15 quarters. Like Blanchard and Perotti (), they find that investment falls, while consumption only rises on impact and the response of real wages is insignificant on impact but negative over longer horizons. With regards to tax shocks, they estimate a peak multiplier of -3.6 for a deficit-financed tax shock at the 13th quarter, which is considerably larger than the estimate by Blanchard and Perotti (). Some authors have recently analyzed whether fiscal multipliers are state-dependent or not. Auerbach and Gorodnichenko (1) use the identification method of Blanchard and Perotti () to analyze the government spending multiplier in a smooth-transition VAR (STVAR), wherein the coefficients in the SVAR can switch between two regimes (a recession regime and an expansion regime). The coefficients switch smoothly such that they are weighted averages of the two regimes coefficients. Specifically, the weight depends 3 While the fiscal policy shocks are assumed to be orthogonal to the monetary policy and business cycle shocks, the two fiscal policy shocks government spending shocks and tax shocks are not orthogonal to each other. 6

13 on a seven-quarter moving average of the growth rate of output as an index for the business cycle (the weight s response to this measure is calibrated rather than estimated). The authors find that the government spending multiplier is state-dependent: it is considerably larger during recessions (the cumulative government spending multiplier is.-.5 during expansions and during recessions over a quarter period). Similar results are found by Fazzari et al. (15) who estimate a SVAR, wherein the coefficients switch discretely once a measure of slack in the economy is above a threshold (they estimated this threshold). They estimate a cumulative government spending multiplier of 1.6 in the slack regime, while the multiplier is less than half of that in the no-slack regime. Unlike Auerbach and Gorodnichenko (1), however, Fazzari et al. (15) also analyze the effect of government spending on consumption and investment. They find a positive response of consumption in both regimes although the response is larger when there is slack in the economy. Investment only decreases in the no-slack regime regime, while it increases albeit not significantly in the slack regime.. The narrative approach The narrative approach exploits additional, historical information other than the standard aggregate data used in traditional SVARs to identify exogenous changes in fiscal policy. Thus, this identification scheme is in some ways similar to microeconometric methods such as IV estimators and natural experiments. Ramey and Shapiro s (1998) paper is one of the earliest contributions to this strand of the literature on fiscal policy shocks. They use large military buildups (the Korean War, the Vietnam War and the Reagan-Carter military buildup) to identify anticipated changes in fiscal policy that are exogenous to macroeconomics variables. In addition, military spending is theoretically appealing since it is unlikely to enter households utility functions, substitute for private consumption or impact private productivity. Ramey and Shapiro read Business Week to pinpoint the moments, where the private sector expected future military spending to increase and construct a dummy time series for these dates. Similar to Blanchard and Perotti (), they find that the output response to military spending is hump-shaped: it increases on impact but reaches its maximum value after 4-6 quarters. In contrast to the typical findings in SVAR analyses, however, consumption and real wages fall, while non-residential investment increases and residential investment decreases. 4 The Ramey-Shapiro war dates have been used in a number of other papers. For example, Burnside et al. (4) analyze the response of hours and real wages to the military spending shocks, and Cavallo (5) study the effects of military spending on government output and employment. 4 While non-durables consumption falls, durables consumption actually rises on impact but quickly falls again. Ramey and Shapiro contribute this to households hoarding durables in anticipation of the Korean War. 7

14 Ramey (11b) expands her original analysis with Shapiro by constructing a series of changes in the expected present value of government spending by using Business Week to gauge the public s expectations of military spending. She estimates a peak government spending multiplier of 1.1, a little lower than the multiplier estimated by Blanchard and Perotti (). Contrary to the results by Ramey and Shapiro (1998), non-residential investment actually falls after an anticipated increase in military spending when this new data series is used, while consumption still falls. To account for the relatively little informational content in the Ramey-Shapiro war dates, Ramey (11b) also uses an additional measure of news about defense spending by including the forecast errors of defense spending based on a survey of professional forecasters instead of the war dates. In this case, the peak multiplier is.8 but the present value multiplier is actually negative since output quickly falls. Owyang et al. (13) extend the military spending news series of Ramey (11b) back to 189 and analyze whether the government spending multiplier is state-dependent or not by letting the coefficients in the impulse response function switch between two regimes depending on the unemployment rate (the unemployment rate is used as a measure of slack in the economy). Contrary to the results of Auerbach and Gorodnichenko (1), they find that the government spending multiplier is not statedependent but lies between.7 and.9 in both states depending on how the multiplier is calculated. Owyang et al. (13) also estimate their model on Canadian data and find evidence of a government spending multiplier that is larger when there is slack in the economy in contrast to what they found in the U.S. data. Romer and Romer (1) analyze major post-war tax changes in the United States by relying on the narrative record such as congressional reports and presidential speeches. This approach allows them to classify tax changes as either endogenous or exogenous by using the political motivation for the tax changes. While endogenous tax changes are done for countercyclical reasons or to counteract a change in government spending, exogenous tax changes are not done to return growth to trend or to offset government spending initiatives; instead, exogenous tax changes can be done for ideological reasons or to spur long-run growth. In addition, the narrative record contains the timing and size of the tax changes. The Romers estimate a peak tax multiplier of -3.1 after 1 quarters a large and persistent effect on output close to Mountford and Uhlig s (9) estimate while a tax increase gives rise to a drop in consumption and a large decrease in investment (investment has a peak multiplier of -11. after 1 quarters). Their approach differs from the Ramey-Shapiro dates, however, in that all tax changes are dated at implementation such that anticipation effects are ignored. The Romers data are used by Mertens and Ravn (14a) who propose a new method the proxy SVAR for identifying shocks by using external instruments in the same vein as traditional IV methods are used in microeconometrics. Specifically, they use the narrative measure of tax changes by Romer and Romer (1) as a proxy for the structural shocks 8

15 in the Blanchard-Perotti SVAR. This method allows them to estimate the coefficients in the relationship between the reduced form residuals and the structural shocks in contrast to the Blanchard-Perotti method, where the coefficients are calibrated. The estimated peak tax multiplier is -3., which is close to the estimate by Romer and Romer (1) but considerably larger than the estimated multiplier by Blanchard and Perotti (). This indicates that the output elasticity of tax revenue used by Blanchard and Perotti () is too low, which attenuates the tax multiplier towards zero (Blanchard and Perotti () did highlight that the size of the response to a tax shock is sensitive to this elasticity). Mertens and Ravn (13) have also used the Romers data and the proxy SVAR in a previously published article, wherein they study the effects of changes in the average personal income tax rate and the average capital income tax rate. By contrast, most authors only look at changes in total tax revenue and not at changes in different tax instruments. Mertens and Ravn (13) find that a one percentage point cut in the personal income tax rate increases GDP by 1.4 per cent on impact, which is equal to a multiplier of -. A one percentage point cut in the capital income tax rate increases GDP by.4 per cent on impact, while its multiplier is not well defined since the effect on tax revenues is very small. Both tax cuts increase investment but only the cut in the personal income tax rate increases consumption and hours, while also lowering unemployment..3 Reconciling SVARs with narrative analyses Both the SVAR approach and the narrative approach yield a positive response of output to a government spending shock and a negative response after a tax hike. However, the two approaches typically estimate different responses of some macroeconomic variables to government spending shocks: consumption and real wages rise in the SVAR approach, while the opposite is the case for the narrative approach. It should be stressed, however, that although positive shocks to government spending in SVARs typically yield an increase in consumption following a government spending shock, this does not mean that the government spending multiplier from this approach is necessarily larger than in the narrative approach since the size of the multiplier ultimately depends on the response of total private spending (Ramey, 16). How can we reconcile that the two approaches yield opposite results with regards to some central macroeconomic variables? The discussion above illustrates that shocks are typically treated differently in the two approaches: shocks are unanticipated in the SVAR approach, while they are not in the narrative approach (Ramey, 16). 5 Hence, 5 The SVAR approach does not ignore anticipation effects completely. For example, Blanchard and Perotti () analyze a large tax cut in 1975 by including dummies in the quarter up to the cut but find no anticipation effects. Mountford and Uhlig (9) account for anticipated tax and government spending shocks by restricting these variables to only move a year after a shock. They find that an anticipated increase in taxes reduces output, consumption and the interest rate immediately, while an anticipated increase in government spending causes output and interest rates to increase immediately. 9

16 anticipation effects might be the source of the dispute. It can be argued that fiscal shocks are usually anticipated by the public. Government spending initiatives and tax cuts are announced (either by formal announcement or through politicians campaign pledges), undergo negotiations between lawmakers, are enacted or rejected, and finally taken into effect. Hence, it might not make a lot of sense to model fiscal shocks as unanticipated. This view is supported by the analysis by Ramey (11b): she finds that government spending shocks from a SVAR are actually Granger-caused by the Ramey-Shapiro war dates. The consequences of ignoring fiscal foresight in the context of anticipated tax changes have been shown by Leeper et al. (13): all dynamics associated with an anticipated tax change is attributed to the unanticipated component in a VAR analysis. Depending on the structure of the information flows, the estimated multipliers can be severely biased in any direction. This criticism might seem to underline the need for using the narrative approach instead of traditional SVARs but authors using the narrative approaches rarely treat information flows rigorously. For example, Ramey and Shapiro (1998) simply use dummies to capture anticipated military buildups, Romer and Romer (1) ignore anticipation effects, and few studies distinguish between different kinds of informational flows such as formally announced fiscal policies and more uncertain campaign pledges. Furthermore, analyses using the narrative approach are often plagued by problems of weak explanatory power and confounding effects. For example, the Ramey-Shapiro war dates only contain three military spending events over the sample period of , while patriotism might affect labor supply, and uncertainty related to wars can generally affect the economy negatively. Zubairy (9) also shows that adding the news measure by Ramey (11b) to a Blanchard-Perotti type SVAR as anticipated shocks orthogonal to unanticipated shocks yield similar results to those of Blanchard and Perotti (), while the impulse response functions from a SVAR with the news measure are economically identical to those from a SVAR without it. This indicates that Ramey s (11b) government spending series do not capture any significant anticipation effects. Luckily, recent papers suggest how anticipation effects can be treated with more rigor. We briefly summarize five of these paper below. Karel Mertens and Morten O. Ravn have written a number of articles about anticipation effects in fiscal policy. We will review two of them. First, Mertens and Ravn (1) study anticipated government spending shocks in an augmented SVAR, which is robust to the presence of anticipation effects, and they show that the estimates from a standard SVAR that does not account for anticipated shocks can be severely biased depending on the relative importance of anticipated shocks and the rate at which news are discounted Thus, both Blanchard and Perotti () and Mountford and Uhlig (9) analyze anticipation effects but in contrast to narrative studies they do not use external instruments but instead rely on dummy variables and restrictions on the SVAR. 1

17 by forward-looking agents. They do, however, find that consumption increases in response to both unanticipated and anticipated increases in government spending once they apply their model to U.S. data. Thus, their results do not invalidate the findings of Blanchard and Perotti (). Second, Mertens and Ravn (1) split the tax change measure by Romer and Romer (1) into unanticipated and anticipated tax changes by defining a tax change as unanticipated if its announcement and implementation dates are less than 9 days apart. They find that unanticipated tax cuts give rise to an increase in output, consumption and investment that peak after.5 years, while real wages rise persistently. This is largely in line with the results of Romer and Romer (1). Before the implementation date, an anticipated tax cut results in a drop in output, hours and investment with no response of consumption and a rise in real wages. The drop in investment and hours is largely consistent with forward-looking behavior, while the lacking response of consumption is not. Once the anticipated cut is implemented, it has a stimulating effect on the economy similar to the unanticipated cut. Fisher and Peters (1) use excess returns on military contractor stocks to identify anticipated changes in military spending. Their approach has an advantage over the Ramey-Shapiro dates in that public uncertainty about military spending is included in stock returns, while the timing of anticipated shocks are derived from the returns rather than determined by the econometrician. They find a cumulative government spending multiplier of 1.5 over a 5 year horizon. However, as Ramey (16) argues, the excess return on military contractor stocks can only explain a small part of the variation in government spending, which makes it a weak instrument. Zeev and Pappa (forthcoming) use a medium-run identification strategy by identifying a defense news shock as the shock that best explains variations in the next five years of defense spending, while being orthogonal to current defense spending. They find that consumption, output, investment and hours increase in response to a positive news shock, while real wages fall (the cumulative multiplier for output over 6 quarters is.14). The increase in consumption runs counter to the findings of both Ramey (11b) and Ramey and Shapiro (1998). In addition, the shock explains a larger share of macroeconomic fluctuations than Ramey s (11b) news shocks and a positive shock increases the excess return stock series by Fisher and Peters (1) contrary to the shocks by Ramey (11b), which suggests that the shocks by Zeev and Pappa (forthcoming) are more informative about future defense spending. Finally, Leeper et al. (1) exploit the differential tax treatment of municipal and federal bonds to construct a measure of anticipated tax changes. Since municipal bonds are exempt from federal taxes, the spread between similar municipal and federal bonds should reflect anticipated changes in future tax rates if asset markets are efficient. They do not use this measure in a VAR analysis but instead use it to provide information about the degree of fiscal foresight in a DSGE model. 11

18 .4 Fiscal policy and the housing market There are few authors who have included housing variables such as house prices, mortgage debt and the housing stock in empirical analyses of fiscal policy. Thus, the understanding of the responses of housing market variables to fiscal shocks is still rather limited. Andrés et al. (15) include household debt and house prices in a SVAR. Their results with a Blanchard-Perotti type identification scheme largely match the results of Blanchard and Perotti () with regards to output, consumption and real wages, while they find that house prices fall and private debt increases in response to a positive government spending shock. By contrast, while Khan and Reza (14) also find that consumption rises in response to government spending, they find that house prices increase following a positive government spending shock (the SVAR analysis by Khan and Reza (14) does not include mortgage debt and labor market variables, which the analysis by Andrés et al. (15) does). Khan and Reza s (14) results are unchanged when they control for expectations by including private sector agents forecast error of government spending although the responses of output and consumption become more hump-shaped. Among other things, Afonso and Sousa (1) analyze the response of house prices to government spending and taxes in a Bayesian SVAR with a Blanchard-Perotti type identification scheme, where they explicitly impose a feedback mechanism from government debt to inflation, interest rates, GDP growth, government spending and taxes through the government s intertemporal budget constraint. They find that house prices increase following a positive government spending shock, while they decrease after a positive shock to taxes irrespective of whether the government debt feedback mechanism is included in their model or not. Bermperoglou (15) analyzes state-dependent effects of fiscal policy in a threshold SVAR model with a Blanchard-Perotti type identification scheme. He uses real house prices as a proxy for housing wealth as the threshold variable: once the real house prices reach a certain level, the coefficients in the SVAR switch to another regime. He finds that the effects of government spending shocks are highly state-dependent. When house prices are above the threshold (i.e. housing wealth is high), then output and consumption increase persistently. The opposite is the case when house prices are below the threshold but house prices will decrease following a positive government spending shock in both regimes. Bermperoglou (15) also includes a narrative measure of the average personal income tax rate. The response to a tax cut is also state-dependent. When house prices are below the threshold, output increases significantly, while its response in the regime with high house prices is positive but insignificant. Consumption increases slightly and house prices rise persistently in both regimes. Thus, the state-dependent effects on output of fiscal expansions depends on the instrument: when house prices are high, then government spending shocks are accelerated, while the opposite is the case for cuts to the personal 1

19 income tax rate. While not focusing exclusively on the housing market, Berger and Vavra (1) analyze the response of durables consumption in which housing investment is included in the STVAR of Auerbach and Gorodnichenko (1). They find that the government spending multiplier on durables consumption is substantially larger during expansions than in recessions. In an expansion, the multiplier is hump-shaped and reaches a maximum value of.8 after 3 years, while it is negative during recessions. This procyclical impulse response is largely in accordance with a model, wherein households face fixed adjustment costs when purchasing durables. When the authors decompose durables into housing investment and consumer durables, they find similar results for both variables. The response of housing investment, however, is more state-dependent than consumer durables (this could be explained by larger fixed adjustment costs in comparison to consumer durables). 3 Theoretical models of the propagation of government spending shocks How do theoretical models of fiscal policy stack up with the empirical evidence discussed in the previous section? As we discuss below, a model with forward-looking households can have difficulty generating positive responses of consumption and real wages when government spending increases as many empirical studies find. This is primarily due to the presence of a negative wealth effect on households. Baxter and King (1993) highlight the role of wealth effects of government spending shocks in a simple RBC model, where government spending is financed with lump-sum taxes. Consider a temporary increase in government spending. Since households will inevitably face higher taxes after a positive shock to government spending, their permanent income decreases, which causes them to lower consumption and leisure if these are normal goods. The increase in leisure is equivalent to an outwards shift in the labor supply curve, which causes real wages to decrease and hours to increase. Thus, the standard RBC model creates comovements of some variables that are contrary to the results of Blanchard and Perotti (): consumption and real wages decrease when government spending increases. Hours do increase but this is due to supply effects, not demand effects. Investment can either increase or decrease depending on how much hours rise: if hours increase sufficiently then the marginal product of capital rises enough to induce an increase in investment (Fatás and Mihov, 1). The forward-looking behavior of households is essentially the reason for the responses in this framework. In contrast, in a Keynesian model such as the IS-LM model the households consumption is a function of their current income and not permanent income as in the RBC model, while prices are fixed 13