G. Natvik - Quantitative and Qualitative Analysis in Social Sciences ...

qass.org.uk

G. Natvik - Quantitative and Qualitative Analysis in Social Sciences ...

We start by evaluating the policy inefficiency hypothesis on U.S. data. Moreover,due to international financial integration, it is likely that U.S. stockmarket volatility and U.S. corporate spreads at least partially capture movementsin global economic uncertainty. We therefore extend our analysis byestimating how these two uncertainty measures interact with the transmissionof monetary policy in six small open economies: Australia, Canada, NewZealand, Norway, Switzerland and the United Kingdom. Beyond serving as arobustness check of our results from the United States, this extension has thebenefit that the U.S.-based uncertainty measures are likely to be exogenouswith respect to monetary policy in each of these countries.Our main finding is that the policy ineffectiveness hypothesis has bite at themacro level. Across the countries we study, the effects of monetary policyshocks tend to be considerably weaker when uncertainty is high. The patternis particularly stark for GDP and investment, and the effects are sizeable.In the United States, a monetary tightening has almost no effect when stockmarket volatility is in its upper decile, while the same impulse makes investmentand GDP drop by approximately one percent when volatility is in itslower decile. The effect when volatility is in its lower decile is more than fivetimes larger than when it is in its upper decile, and we show these differencesto be statistically significant. When uncertainty is measured by the corporatespread, forecaster disagreement or the Google-index, the results are qualitativelysimilar, but quantitatively somewhat smaller. For these measures, thepolicy effect is approximately halved when uncertainty is in its upper ratherthan in its bottom decile. The differences in policy effects remain statisticallysignificant. It is only economic policy uncertainty that does not seemto dampen the influence of monetary policy, which might reflect that policyactions directly influence policy uncertainty.For GDP and particularly investment, our findings from the six small openeconomies are qualitatively in line with the pattern for the United States.The investment responses to a monetary policy shock are approximatelyhalved when stock market volatility is in its bottom instead of its top decile,and the response differences are statistically significant in all countries. Whenuncertainty is measured by the spread, investment responses are also weakerif uncertainty is high, and the 84% probability bands for the response differencesexceed zero in 4 out of 6 countries. That the interaction effects arequantitatively smaller here than in the United States, is hardly surprisinggiven that the uncertainty measures are obtained from U.S. financial markets.Our study is closely related to the recent literature initiated by Bloom (2009),3


that uses SVARs to identify uncertainty shocks and how they affect themacroeconomy. Bloom (2009) showed that shocks to U.S. stock marketvolatility are followed by contractions in U.S. employment and investment.Bachmann et al. (2013) measure uncertainty by forecast disagreement anddispersion in forecast errors in business surveys from Germany and the U.S.,document that the two measures are correlated, and show how innovationsin these indexes are followed by economic contractions in the two countries.They also show that shocks to spreads, stock market volatility and theGoogle-index foreshadow similar contractions. Baker et al. (2012) constructindexes of economic policy uncertainty in the United States, Canada andEurope and find similar effects. Alexopoulos and Cohen (2009) construct anuncertainty index reflecting the frequency with which the news media refersto economic uncertainty, and also they find that increased uncertainty is followedby reduced economic activity. Other studies in this vein are Bachmannand Bayer (2011) and Knotek and Khan (2011). Notably, the recent strandof macroeconomic empirical research has focused exclusively on the questionof how movements in uncertainty affect economic activity, while no attentionhas been directed to the policy-effectiveness hypothesis which we address.The evidence that exists on policy-effectiveness and uncertainty is obtainedeither through micro-data (e.g. Bloom et al. (2007)), or through structuralmodels where the implication of policy ineffectiveness is largely imposed bytheory (e.g. Bloom (2009), Bloom et al. (2007)). A particularly relevantpaper is Bloom et al. (2012), who build and calibrate a dynamic stochasticgeneral equilibrium model with non-convex adjustment costs, and study howthe influence of expansive policy depends on uncertainty. Our contributionis to empirically evaluate with macroeconomic data if the effect of monetarypolicy interacts with the level of uncertainty in the economy. This relates ourpaper to the vast empirical literature on the transmission of monetary policyshocks, summarized at one stage by Christiano et al. (1999). 3 Within thisfield, several studies explore how the monetary transmission mechanism hasevolved over time, such as Boivin et al. (2010), Canova and Gambetti (2009),Boivin and Giannoni (2006) and Kuttner and Mosser (2002). However, toour knowledge, this literature has not previously addressed how uncertaintymight alter the effectiveness of monetary policy.The remainder of our paper is organized as follows. Section 2 gives a theoreticalmotivation for our empirical investigation, based on a stylized modelof an irreversible investment decision. Section 3 describes our data and the3 See, for example, Sims (1992), Bernanke and Mihov (1998), Romer and Romer (2004) andBernanke et al. (2005).4


methodology we use. Section 4 and 5 presents our results for the UnitedStates and the six small open economies, respectively. Finally, we makesome concluding remarks in Section 6.2 Theoretical MotivationIn this section we use a simple theoretical model to show how uncertaintyinfluences a decision that can be postponed. The model is highly stylized,in order to allow an analytical expression of the hypothesis we want to test.The same qualitative prediction is obtained numerically in the more detailedmodels with non-convex adjustment costs used by for instance Bloom et al.(2012), Bloom et al. (2007) and Bloom (2007). The classic textbook expositionof these effects is given in Dixit and Pindyck (1994).There are 3 periods, 0, 1 and 2. In period 0, a continuum of entrepreneursindexed by i face the opportunity to invest in a project. The cost of undertakingthis investment is γ i , which is uniformly distributed across investors withdensity 1/α. If invested in, the project pays off y in periods 1 and 2, wherey is stochastic. With probability p, y = y h , while y = y l with probability1 − p. The distance between y h and y l captures the degree of uncertainty inthis economy, and is denoted σ. Uncertainty about y is realized in period 1,and after observing this level, an entrepreneur who did not invest previouslymay choose whether or not to invest for period 2. We assume that after y isrealized, the resale price of capital does not exceed y, and hence projects arenever terminated in period 1. The alternative to investing in the project isto spend γ i on a risk-free asset yielding the gross interest rate R.To make the investment decision interesting, we assume that the project isunprofitable if the state with low productivity materializes: γ i > y l /R +y l /R 2 for all investors i. On the other hand, if the high-productivity statematerializes, the project is profitable even if it operates for one period only:γ i ≤ y h /R.The net present value from investing the amount γ i in period 0 is givenby E ( )πi,0inv = E (y) /R + E (y) /R 2 − γ i , while the net present valuefrom postponing the investment decision until period 1 is E ( )π no−invi,0 =(1 − p) γ i +p/R [ ] 2 y h + (R − 1) γ i −γi . The latter expression reflects that bydelaying the investment decision, the investor gains the option to invest in therisk-free asset rather than a project that has turned out to be unprofitable.Naturally, the investment will be made in period 0 if and only if E ( π invi,05)−


E ( )π no−invi,0 ≥ 0. From the two profit expressions, we can therefore expressthe individual investment decision in terms of the fixed investment cost.Entrepreneur i will choose to invest if γ i ≤ γ, whereγ =RE (y) + (1 − p) ylR 2 (1 − p) + (R − 1)pAggregate investment in period zero, I 0 , is then given by the mass of investorswith γ i ≤ γ. Hence, I 0 = 1/2 + [γ − E (γ)] /α.We can now answer how uncertainty affects investment. 4 First, the effect ofa mean-preserving increase in uncertainty is: 5∂I 0∂σ | E(y) =− (1 − p) p[R 2 (1 − p) + (R − 1) p] α < 0.We see that higher uncertainty leads to lower investment. This is the “delay”effect of higher uncertainty. The effect follows from equation (1), as themean preserving increase in uncertainty by definition reduces y l , and therebytightens the condition for investors to take action. Intuitively, a wider distributionof potential payoffs increases the cost of making a wrong decision andtherefore raises the value of postponing the investment decision in order togain further information. As consequence, the investment cost that triggersdelay falls, and fewer invest. This delay effect has been scrutinized in thegrowing literature on uncertainty shocks referred to in the introduction.Second, higher uncertainty influences how strongly movements in the interestrate R, affects the motive to invest:(1)∂ 2 I 0∂R∂σ | E(y) =(1 − p) p [2R (1 − p) + p][R 2 (1 − p) + (R − 1) p] 2 α > 0.As the partial effect of a higher interest rate on investment, ∂I 0 /∂R, is negative,it follows that higher uncertainty reduces the influence of monetarypolicy. This reflects the “caution effect” that uncertainty creates. Whenuncertainty goes up, there is more at stake when deciding whether to investor not, and hence a marginal change in investment incentives, as induced bya change in R, has a smaller impact. In terms of equation (1), the effect can4 Notably, the key assumption behind the uncertainty effects here is that projects cannot beliquidated at a price that exceeds their productivity. This assumption makes investmentirreversible in this model. If projects could be terminated and sold at prices that exceededy, the option value of postponing investments would not be increased by higher uncertainty.5 Here we have utilized that dyldσ | E(y) = −p.6


e observed by noting that an interest rate hike reduces investment incentivesby raising the discounting of E (y) and y l . The higher is uncertainty,the lower is y l , and hence the smaller is the effect of R on the individualinvestment decision as captured by γ. This “caution effect” is what we willexplore in this paper.3 Data and Empirical StrategyWe aim to estimate how economic uncertainty influences the transmission ofmonetary policy shocks. We perform our analysis first for the United States,and thereafter for six small economies. These are Canada, Australia, NewZealand, Switzerland, Norway and the United Kingdom. 6We use quarterly data throughout. For the US, we cover the period 1970Q1to 2011Q3. 7 For all the other countries we use data from 1978Q4 to 2011Q3,since this is the longest period over which we have consistent macroeconomicseries.3.1 Measuring Economic UncertaintyAs our main measure of uncertainty, we will use the series for volatility in theUS stock market constructed by Bloom (2009), extended to cover the periodfrom 2008 to 2011. From 1986 and onward, these data are taken from theChicago Board Options Exchange VXO index of implied volatility. For theperiod before 1986, the implied volatility index is unavailable, and realizedvolatility, calculated as quarterly standard deviation of the daily S&P500,is used instead. We will refer to the stock market volatility index as theVXO-index throughout.6 We do not include Euro area countries in the analysis as it complicates the identificationof monetary policy shocks and requires us to tackle the issue of the introduction of a singlecurrency.7 We start in 1970 because our longest series for uncertainty is the US stock market volatilityindex from Bloom (2009), which trends upward from a low level until 1970. The volatilitylevel is therefore unlikely to treat periods of uncertainty in a symmetric way before andafter 1970. For instance, the Cuban missile crisis in 1962 leads to a spike in the volatilityindex that lies below its average over the entire period. Using dummies to capture uncertaintyspikes, as in Bloom (2009), does not suffer from this problem, but this approach isnot compatible with the interaction VAR strategy that we will apply.7


While financial market volatility is a natural starting point for measuringuncertainty, in the recent literature on uncertainty shocks several alternativemeasures have been proposed. We therefore consider a host of these.First, we use the alternative measures provided in Bachmann et al. (2013),namely the corporate bond spread, forecast disagreement, and the Googleindex. The corporate bond spread is the spread of the 30-year Baa-ratedcorporate bond yield index over the 30-year treasury bond yield, where the20-year treasury bond has been used when the 30-year bond was missing. Theforecast disagreement measure stems from the the Federal Reserve Bank ofPhiladelphia’s Business Outlook Survey, where large manufacturing firms inthe Third Fed district are asked to give their own “evaluation of the generalbusiness activity six months from now.” 8 Disagreement is measured as thecross-sectional dispersion in these point forecasts. The Google index is thenumber of articles in a given month that refer to “uncertainty” and phrasesrelated to the economy, divided by the number of articles containing the word“today” in order to control for the overall increasing news volume. Note thatthis variable is only available from 1985Q1.Second, we use the economic policy uncertainty (EPU) index constructed byBaker et al. (2012). The EPU index is built on the following components:the frequency of newspaper references to economic policy uncertainty, thenumber of federal tax code provisions set to expire, and the extent of forecasterdisagreement over future inflation and government purchases. As forthe Google index, this variable is only available from 1985.The uncertainty measures are plotted in Figures 1 and 2.3.2 Macroeconomic variablesFor each country we include the following macroeconomic variables: the CPIinflation rate, GDP, investments, private consumption, the short term interestrate. In addition we include the real effective exchange rate for the sixsmall open economies. Inflation is computed as the annual growth rate of theCPI index and is expressed in percentage terms. GDP, investment and privateconsumption are instead expressed in real terms and transformed usingnatural logarithms. 98 More precisely, according to Bachmann et al. (2013), the survey is sent to firms inDelaware, the southern half of New Jersey, and the eastern two-thirds of Pennsylvania,with voluntary participation, and monthly responses from executives in 100-125 firms.9 For Norway we exclude the oil sector and use GDP and investments for the mainland-Norway only. This is important because the oil sector is relatively large in Norway, and8


To construct investment series which are both comparable across countriesand consistent with the theoretical model, we use “gross fixed capital formation”of the private sector. These series are taken from Datastream wherethe original sources are the national statistical offices. For all countries, theshort term interest rate is the interest rate on 3 months government bonds.It is expressed in percentage terms and is considered as an indicator of themonetary policy stance. Our measure of the real exchange rate is the “CPIreal effective exchange rate” published in the International Financial Statisticsof the IMF. It is a trade weighted measure of real exchange rate, and itis constructed so that an increase means an appreciation of the currency.3.3 Empirical modelTo study how time-varying uncertainty affects the transmission of monetarypolicy, we estimate an interacted structural VAR model. By interacting themacroeconomic variables with the uncertainty index, we allow the impact ofmonetary policy to change with the degree of uncertainty. The model buildson the interacted panel VAR model developed by Towbin and Weber (2013)and Sa et al. (2013). However, it differs in two aspects. First, while the twomentioned papers use a panel approach, we apply a separate model to eachof the economies we study. Second, the interaction term we use differs fromthe two mentioned papers. Notably, with this approach time-variation inthe effect of policy is directly linked to a specific determinant, uncertainty,in contrast to the studies that use VARs with stochastically time-varyingcoefficients, such as Canova and Gambetti (2009) and Primiceri (2005).We consider the following interacted VAR model:Y t = A 0 + B 0 X t +L∑l=1(Al Y t−1 + B l Y SMt−1 X t)+ CZt + E t (2)where E t is a vector of reduced form residuals at time t. The vector Y tcontains CPI inflation, GDP, private consumption, private investments, andthe short term interest rate. For the small open economies we also includethe real exchange rate. Furthermore, the model allows the variables in Y t tointeract with X t . X t is also included as an additional regressor. 10largely driven by other forces than the rest of the Norwegian economy.10 This is the multivariate analog to a standard interaction term, in which the effects of Y t−lon Y t is a linear function of X t .9


The interaction term X t is the uncertainty measure, assumed to be exogenousin our model. We use a four quarter moving average to account for a laggedreaction of consumers and investors to the level of uncertainty. Note thatin our exercise we want to quantify the extent to which the response of theendogenous variables to the interest rate changes with the level of uncertainty.Therefore, in each equation we interact X t only with the lags of the interestrate.The vector Z t contains additional exogenous variables. For all countrieswe include a linear trend in Z t , and for the small open economies we alsoinclude the U.S. interest rate. We have also considered specifications wherewe include commodity prices or oil prices in Z t , both because their inclusionhas been proposed as important to identify monetary policy shocks withoutgenerating a “price puzzle”, and because several of the small open economieswe consider are commodity producers. Our results are robust to includingcommodity prices or oil prices in the model.A 0 is a vector of constant terms, while B 0 , A l , B l and C, are parametervectors for the interacted variable (X t ), the endogenous variables (Y t ), theinteraction term (Y t−1 X t ) and the exogenous variables (Z t ), respectively.Finally, several of the countries we study have changed monetary policyregime and adopted inflation targeting over our sample period. In Z t wetherefore include a dummy variable which takes the value 1 for the periodafter the regime change. 11To study the effect of policy effectiveness in our econometric model, we needto identify monetary policy shocks. Our identification strategy is explainedbelow.3.4 Estimation and IdentificationFor the U.S. analysis, we identify monetary policy shocks by imposing recursiveshort-run restrictions, allowing the interest rate to respond within thesame quarter to all the macroeconomic variables, but not vice versa. Hence,macroeconomic variables react with a lag to monetary policy shocks. Thisrecursive structure in the identification of monetary policy is the most con-11 For Norway we also follow Bjørnland and Jacobsen (2010) and include three dummyvariables for 1992Q3, 1992Q4 and 1993Q1 to account for episodes of extreme turbulencein the interest rate and exchange rate due to the breakdown of the fixed exchange rateregime.10


ventional one in the established SVAR literature, see for instance (see Stockand Watson (2001) and Christiano et al. (1999)).For the small open economies, we deal with the simultaneity between theexchange rate and the interest rate by employing a mixture of sign andcontemporaneous zero restrictions. We allow both the interest rate and theexchange rate to respond within the same quarter to all other macroeconomicvariables included in the VAR, but not vice versa. In addition, we allowthe interest rate and the exchange rate to react contemporaneously to eachother. In order to to separate a monetary policy shock and an exchange rateshock we impose two additional sign restrictions. We follow Jarocinski (2010)and impose that a positive monetary policy shock is associated with a jointincrease in the interest rate and an appreciation of the exchange rate, whilean exchange rate shock is associated with a decrease in the interest rate andan appreciation of the exchange rate. 12 The sign restrictions are imposed tohold for two quarters.As with standard interaction terms, the effect of the identified shock canthen be graphed for different levels of the variable X t . We choose to focuson the bottom decile and on the top decile of the uncertainty indicators.We estimate our model using Bayesian techniques. Following Sims and Zha(1999), Uhlig (2005) and Sa et al. (2013) we impose that the priors, andhence also the posterior density, of the regression coefficients and the covariancematrix belongs to the Normal-Wishart family. We use uninformativepriors, and draw all parameters jointly from the posterior (including the coefficientson the interaction terms). 13 The priors and the implementation ofthe Bayesian estimation are discussed in more detail in appendix A. Oncewe have estimated the parameters of the VAR in equation (2) we identifystructural shocks with the strategies explained above. When imposing signrestrictions for the small open economies we follow Rubio-Ramirez et al.(2010). For each posterior draw of the covariance matrix we compute thematrix R, obtained by orthogonalizing the variance covariance matrix 14 andmultiplying it by an orthonormal matrix Q. The matrix Q is constructedexactly as in Jarocinski (2010): it is an identity matrix with a lower blockobtained via a QR decomposition of a 2 × 2 random matrix drawn from an12 See Bjørnland (2009) for an alternative way to separate exchange rate shocks and monetarypolicy shocks using long-run restrictions.13 We start with 22,000 draws, discard the first 2,000 of them, and thereafter select onlyevery tenth draw to avoid correlation. Consequentially, we are left with 2,000 draws of theparameters.14 We here use a standard Choleski factorization.11


independent standard normal. We keep the posterior draw if the matrix Rgenerates impulse responses that satisfy the sign restrictions for both shocks.For a given parameter draw, we keep 100 impulse responses which satisfy ourrestrictions.To draw the impulse responses to a given shock we must specify values of theinteraction variable X t at which to evaluate the effect of the shock. This isagain the multivariate analog to a standard interaction term. As for standardlinear regressions like y t = β 0 + β 1 z t + β 2 x t + β 3 x t z t + ε t , the presence of∂ythe interaction term implies that: t∂x t= β 2 + β 3 z t . This means that theeffect of a change in x t is a linear function of the value of the variable z t .To quantitatively evaluate the importance of the interaction effect we mustchoose two values of the variable z t , which we denote “z hight ” and “ztlow ”, andcompute impulse responses under each of them. In our exercise we use thevalues of the uncertainty measures corresponding to the top 10% and to thebottom 10% of their distribution. From the Figures 1 and 2 it is evident thatthe episodes of highest uncertainty in our sample correspond to the periodsaround the “Black Monday” stock market crash, the 9/11 terror attack, thecollapse of the dot-com bubble, and finally the onset of the sub-prime crisis.After having assigned values to the variable X t the estimated VAR reducesto:Y hightY lowt= ̂Dhigh ∑ Lhigh0 + (̂D l Yt−l ) + ĈZ t + Êtl=1= ̂Dlow ∑ Llow0 + (̂D l Yt−l ) + ĈZ t + Êtl=1where ̂D 0high= Â 0 + ̂B 0 X high and ̂D 0low= Â 0 + ̂B 0 X low . Similarly, ̂D lhigh=Â l + ̂B l X high and ̂D llow= Â l + ̂B l X low . These are standard reduced formVAR-models, and there is therefore no further complication associated withusing the restrictions discussed above to identify a monetary policy shockand analyze its effects at high and low levels of uncertainty on the variablesincluded in Y t . Finally, for the small open economies where a sign restrictionis employed, we compute the impulse responses applying the “median targetmethod” proposed by Fry and Pagan (2011).12


3.5 A Test Statistic for the Interaction EffectBecause there exists correlation between impulse responses for low and highlevels of uncertainty, the confidence bands around each response alone yieldsa distorted impression about the statistical significance of their difference.Hence, to assess the difference more formally, we perform a test using theimpulse responses based on draws of the posterior parameters. For eachof the 2,000 saved draws we compute the differences between the responseof the variables to a monetary policy shock under high and low uncertainty,respectively. This provides us with an empirical distribution of the differencebetween the responses, which we can then use to compute a probability band.The two impulse responses can be considered statistically different from eachother if the interval between the probability bands lie above or below zero.We will report these probability bands in figures. Each figure will displaythe distribution of the the difference between the impulse response underhigh and low uncertainty. For example, a positive value for investment willmean that a monetary tightening causes a greater drop in investment whenuncertainty is low than when uncertainty is high. For each variable, we willreport 84% and the 95% probability bands, respectively.4 Uncertainty and the Effect of MonetaryPolicy Shocks in the USWe first estimate how uncertainty affects the transmission of monetary policyshocks in the United States.4.1 Interaction with Stock Market VolatilityOur starting point is to estimate the effects of monetary policy shocks usingstock market volatility as our measure of uncertainty. Figure 3 displays theeffects of a one percentage point unanticipated increase in the policy rate.The impulse responses with circles are estimated for the case where volatilityis in its upper decile, while the curves with marks give the estimated responseswhen volatility is in its lower decile. The figure plots responses for the first20 quarters after the shock.We see that for GDP, investment and consumption, responses are sizeable,and in line with conventional monetary theory, when volatility is low. In13


contrast, when stock market volatility is high, the same variables respondnegligibly. Investment falls by less than 0.2 % when volatility is in its upperdecile, while it falls by more than 1 % when volatility is low. This differenceis consistent with the cautiousness effect explained above. Consumptiondisplays a similar pattern, and it falls by a maximum of almost 3 percentagepoints when volatility is low, which is about ten times more than whenuncertainty is high. The GDP response is consistent with the investmentand consumption movements, and it falls by at most 1 percentage pointwhich is about five to ten times the response when volatility is high.For inflation, we see that in both the high and low volatility scenarios, thereis a “price puzzle” as inflation initially increases in response to the monetarytightening. This puzzle is a well-known by-product of using a structural VARwith recursive ordering to identify monetary policy shocks, see for instanceSims (1992) and Christiano et al. (1999). More interestingly, we see that thisinitially positive inflation response is practically insensitive to the level ofstock market volatility. Moreover, once the initial increase has tapered off,the response under high uncertainty settles at zero, while the low-uncertaintyresponse becomes negative after about 10 quarters.The test statistic for each difference in impulse response is reported in Figure4. As we would expect from the large effects discussed above, we see that forthe real variables, the 95% probability bands all lie above zero. For inflationthe difference between responses becomes significant at the 84% level after14 quarters, but the 95% bands always contain zero.4.2 Interaction with Alternative Uncertainty MeasuresThe first alternative uncertainty measure we consider is the corporate bondspread. Figures 5 and 6 report impulse responses and the test statistic,respectively. Qualitatively, the pattern is very similar to what we obtainedusing stock market volatility, but quantitatively the effects are weaker. GDP,investment and consumption all fall about twice as much when uncertainty islow. The test statistic shows that these differences are statistically significant.For inflation, there is now no difference in responses.Next, we use forecaster disagreement. Figure 7 shows that the effect of amonetary tightening on the three real variables is approximately halved whenforecaster disagreement is in its upper decile rather than its lower decile. Thetest statistics in Figure 8 show that these differences are significant at the14


84% level, but not at the 95% level. For inflation, the initial increase isnow more pronounced when uncertainty is high. After 5 quarters forecasterdisagreement is irrelevant.The results using the Google-index are given in Figures 9 and 10. Note thatthe sample behind these results is considerably shorter, as the Google-indexgoes back only to 1985. Once again we see lower effects of policy on real activitywhen uncertainty is high. Quantitatively, the influence of uncertaintyis similar to what we found using the credit spread and forecaster disagreement:the responses of real variables are roughly halved when uncertainty isin its upper rather than in its lower decile. The 84% bands consistently lieabove zero, whereas the 95 % bands at some point exceed zero for all threereal variables. For inflation, the influence of uncertainty is negligible.Finally, we consider the EPU-index, which also starts in 1985. Figures 11 and12 shows that this uncertainty measure paints a very different picture thanthe other four. The initial effect of the monetary policy shock on the realeconomy is greater when the EPU-index is high, the direct opposite effectof what we have seen before. Moreover, the interest rate dynamics nowdiffer a great deal with the uncertainty measure, with far higher persistencewhen uncertainty is low. Inflation responds positively and persistently whenuncertainty is low, and negligibly when uncertainty is high. A potentialreason why the EPU-index gives such different results could be that theevents it captures are related to uncertainty about future monetary policy.In a period of high policy uncertainty, expansive policy might stimulate theeconomy by reducing the policy uncertainty itself, which would explain thereal effects in Figure 11. In a nutshell, our practice of treating the uncertaintyindexes as exogenous is likely to be particularly severe when we consider theEPU-index.5 Uncertainty and the Effect of MonetaryPolicy Shocks in Small Open EconomiesIn this section we extend our analysis to a set of six small open economies.We will study the interaction effects between monetary policy shocks and thesame U.S.-based uncertainty measures as we utilized above. Partly, this canbe seen as a robustness check, to see if our insights from the United Stateshold more generally. More importantly though, by studying how small openeconomies are affected by uncertainty measured in the United States, our15


esults are less prone to any potential biases from treating uncertainty asexogenous. For some of these countries, such as New Zealand and Norway,it is highly unlikely that their macroeconomic development is influencing theUS-based uncertainty measures to any noteworthy extent. For others, suchas the United Kingdom, this might be more likely. We therefore displayresults for each country separately.Using U.S. stock market volatility to measure uncertainty, our results for thesix countries are displayed in Figures 13 to 24. In general, the responses to aone standard deviation increase in the interest rate are as expected: inflation,output, investments and private consumption decrease, while the exchangerate appreciates in all countries except Australia.The starkest pattern from this analysis is that for investment, the effect ofuncertainty is highly similar to what we found for the United States: Whenuncertainty is high, the initial investment responses are dampened in allcountries. In Australia, Switzerland and New Zealand the response of privateinvestments is negative and statistically significant only when uncertainty islow, while it is small and statistically insignificant otherwise. For the UK,Norway and Canada the response of investments is statistically significantboth with low and high levels of uncertainty, but the magnitude of the responseis about three times larger in a low-uncertainty environment. In fiveof six cases the 95% probability bands of the response differences exceed zeroat some horizon after the shock. Quantitatively, the investment responsesare about two (New Zealand and the UK) to five (Australia) times higherwhen stock market volatility is in the bottom decile than when it is in thetop decile.For GDP, the response to a monetary policy shock is weaker under highuncertainty in four out of six countries, while in the last two (Norway andSwitzerland) the difference is negligible. The pattern of consumption responsesis what differs the most from our previous findings from the UnitedStates, and the effects vary considerably across countries. For Norway, consumptiondeclines more when uncertainty is high, while in the UK, NewZealand and Switzerland, its response is generally more muted and it is negativeand statistically significant only when the level of uncertainty is low.In Australia and Canada the response of consumption is negative and statisticallysignificant irrespective of the level of uncertainty. In four out of thesix countries we study there is a price puzzle. It is smaller for these smallopen economies than what we observed for the US.As we did for the United States, we have also extended our analysis of thesmall open economies by considering alternative uncertainty measures. With16


the U.S. credit spread as interacting variable, the results are qualitativelysimilar to what we found before. In all countries the investment response issmaller when uncertainty is high, and in all countries except Australia, the84% probability bands lie above zero over an extended period. In contrast,when uncertainty is measured by forecaster disagreement, the Google indexor the EPU index, no clear pattern arises.Overall, it seems that the two uncertainty measures based on U.S. financialmarkets yield considerable interaction effects for investments, while the nonfinancialindicators do not. A likely reason for this dichotomy is that whereasfinancial markets are internationally integrated, and therefore tend to reflectaspects of the economic environment that are relevant across countries, thenon-financial measures are more specific to the United States, and of lessrelevance for the small open economies.6 ConclusionOver the last years economic uncertainty has been given much attention,both by policymakers and in the academic literature, as a potential driverof business cycle fluctuations. Much of the debate has been motivated bythe fear that heightened uncertainty will motivate firms and households todelay decisions that are costly to reverse. This paper contributes here. Ourempirical findings indicate that monetary policy is less effective when uncertaintyis high. In the U.S.-data, both consumption and investment responsesto a monetary policy shock are approximately halved when uncertainty ishigh rather than low, and for investment the pattern holds across a set of sixsmall open economies. Taken at face value, this implies that in the face ofhigh uncertainty, monetary policymakers may face a trade-off between actingdecisively and acting correctly, as policy must be more aggressive thanotherwise in order to stabilize economic activity.Our results are consistent with the “cautiousness” effects suggested by economictheory about individuals who face non-convex adjustment costs. However,two of our findings may imply that alternative mechanisms are also atplay. First, non-convexities should be more important for investment thanconsumption, which conflicts somewhat with the finding that in the UnitedStates uncertainty dampens consumption responses as much as investmentresponses. Second, the pattern of reduced policy effect is particularly stark,and the effects particulary large, when uncertainty measures from financialmarkets are utilized. Further research on the exact mechanism behind the17


policy ineffectiveness effects we find seem warranted.ReferencesAlexopoulos, M. and J. Cohen (2009). Uncertain times, uncertain measures.Working Papers tecipa-352, University of Toronto, Department of Economics.Bachmann, R. and C. Bayer (2011). Uncertainty business cycles - really?NBER Working Papers 16862, National Bureau of Economic Research,Inc.Bachmann, R., S. Elstner, and E. Sims (2013). Uncertainty and economic activity:Evidence from business survey data. American Economic Journal:Macroeconomics forthcoming.Baker, S. R., N. Bloom, and S. J. Davis (2012). Measuring economic policyuncertainty. Working paper.Bernanke, B., J. Boivin, and P. S. Eliasz (2005, January). Measuring the effectsof monetary policy: A factor-augmented vector autoregressive (favar)approach. The Quarterly Journal of Economics 120 (1), 387–422.Bernanke, B. S. (1983). Irreversibility, uncertainty, and cyclical investment.The Quarterly Journal of Economics 98 (1), 85–106.Bernanke, B. S. and I. Mihov (1998). Measuring monetary policy. TheQuarterly Journal of Economics 113 (3), pp. 869–902.Bjørnland, H. C. (2009). Monetary policy and exchange rate overshooting:Dornbusch was right after all. Journal of International Economics 79 (1),64–77.Bjørnland, H. C. and D. H. Jacobsen (2010). The role of house prices in themonetary policy transmission mechanism in small open economies. Journalof Financial Stability 6 (4), 218–229.Bloom, N. (2007). Uncertainty and the dynamics of r&d. American EconomicReview 97 (2), 250–255.Bloom, N. (2009). The impact of uncertainty shocks. Econometrica 77 (3),623–685.18


Bloom, N., S. Bond, and J. V. Reenen (2007). Uncertainty and investmentdynamics. Review of Economic Studies 74 (2), 391–415.Bloom, N., M. Floetotto, N. Jaimovich, I. Saporta-Eksten, and S. J. Terry(2012). Really uncertain business cycles. NBER Working Papers 18245,National Bureau of Economic Research, Inc.Boivin, J. and M. P. Giannoni (2006, August). Has monetary policy becomemore effective? The Review of Economics and Statistics 88 (3), 445–462.Boivin, J., M. T. Kiley, and F. S. Mishkin (2010, January). How has themonetary transmission mechanism evolved over time? In B. M. Friedmanand M. Woodford (Eds.), Handbook of Monetary Economics, Volume 3 ofHandbook of Monetary Economics, Chapter 8, pp. 369–422. Elsevier.Canova, F. and L. Gambetti (2009, February). Structural changes in theus economy: Is there a role for monetary policy? Journal of EconomicDynamics and Control 33 (2), 477–490.Christiano, L. J., M. Eichenbaum, and C. L. Evans (1999). Monetary policyshocks: What have we learned and to what end? In J. B. Taylor andM. Woodford (Eds.), Handbook of Macroeconomics, Volume 1 of Handbookof Macroeconomics, Chapter 2, pp. 65–148. Elsevier.Dixit, A. and R. Pindyck (1994). Investment under Uncertainty. PrincetonUniversity Press.Fry, R. and A. Pagan (2011, December). Sign restrictions in structural vectorautoregressions: A critical review. Journal of Economic Literature 49 (4),938–60.Jarocinski, M. (2010). Responses to monetary policy shocks in the east andthe west of europe: a comparison. Journal of Applied Econometrics 25 (5),833–868.Kadiyala, K. R. and S. Karlsson (1997). Numerical methods for estimationand inference in bayesian var-models. Journal of Applied Econometrics12 (2), 99–132.Knotek, E. S. and S. Khan (2011). How do households respond to uncertaintyshocks? Economic Review (Q II).Koop, G., D. J. Poirier, and J. L. Tobias (2007).Methods. Cambridge University Press.Bayesian Econometric19


Kuttner, K. N. and P. C. Mosser (2002). The monetary transmission mechanism:some answers and further questions. Economic Policy Review (May),15–26.Primiceri, G. (2005). Time varying structural vector autoregressions andmonetary policy. Review of Economic Studies 72 (3), 821–852.Romer, C. D. and D. H. Romer (2004, September). A new measure ofmonetary shocks: Derivation and implications. American Economic Review94 (4), 1055–1084.Rubio-Ramirez, J. F., D. F. Waggoner, and T. Zha (2010, 04). Structuralvector autoregressions: Theory of identification and algorithms for inference.Review of Economic Studies 77 (2), 665–696.Sa, F., P. Towbin, and T. Wieladek (2013). Capital inflows, financial structure,and housing booms. Journal of the European Economic AssociationForthcoming.Sims, C. A. (1992). Interpreting the macroeconomic time series facts : Theeffects of monetary policy. European Economic Review 36 (5), 975–1000.Sims, C. A. and T. Zha (1999, September). Error bands for impulse responses.Econometrica 67 (5), 1113–1156.Stock, J. H. and M. W. Watson (2001). Vector autoregressions. Journal ofEconomic Perspectives 15 (4), 101–115.Stock, J. H. and M. W. Watson (2013). Disentangling the channels of the2007-2009 recession. Brookings Papers on Economic Activity Forthcoming.Towbin, P. and S. Weber (2013). Limits of floating exchange rates: The roleof foreign currency debt and import structure. Journal of DevelopmentEconomics 101 (1), 179–101.Uhlig, H. (2005, March). What are the effects of monetary policy on output?results from an agnostic identification procedure. Journal of MonetaryEconomics 52 (2), 381–419.20


AppendicesAppendix ABayesian estimation and priorsWe use an uninformative version of the natural conjugate priors described inKadiyala and Karlsson (1997) and Koop et al. (2007). For simplicity, assumewe can rewrite equation 2 in the following form:Y = Xβ + ɛ (3)where X now includes all regressors in equation 2, i.e. lagged endogenous,exogenous and interacted variables, and ɛ has a variance-covariance matrixΣ.We apply the general Matricvariate Normal - Wishart priors for β and Σ −1 .whereandp(β, Σ −1 ) = p(β)p(Σ −1 ), (4)(β|Σ) ∼ MN (β, V ) (5)(Σ −1 ) ∼ W(H, ν) (6)Noninformativeness is then achieved by imposing that ν = 0 and H −1 = 0×I.By using the prior, we can derive the following conditional posteriors forp(β|Y, Σ −1 )(beta|Y, Σ −1 ) ∼ MN (β, V ) (7)whereandand likewise for p(Σ −1 |Y, β)V = (V −1 +β = V (V −1 +T∑X ′ Σ −1 X) −1 (8)t=1T∑X ′ Σ −1 Y ) (9)t=1(Σ −1 |Y, β) ∼ W(H, ν) (10)whereν = T + ν (11)21


andH = [H −1 +T∑X ′ Σ −1 (Y − Xβ)(y − xβ) ′ ] −1 (12)t=1We then use the Gibbs sampler to sequentially draw from the normalp(β|Y, Σ −1 ) and the wishart p(Σ −1 |Y, β).22


Appendix BFiguresFigure 1: Uncertainty Measures Since 1970Note: The quarterly average of the stock market volatility index constructed as in Bloom(2009), a corporate bond spread and the forecast disagreement measure used in Bachmannet al. (2013) for the period 1970Q1-2011Q3. Each series has been demeaned and standardizedby its standard deviation.Figure 2: Uncertainty Measures Since 1985Note: The quarterly average of the policy uncertainty index in Baker et al. (2012) andthe Google index used in Bachmann et al. (2013). Each series has been demeaned andstandardized by its standard deviation.23


Figure 3: Monetary Policy Shock - USA - Stock Market VolatilityNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the stock market volatility index. Low uncertainty depicts responses whenstock market volatility is in its lower decile. High uncertainty depicts responses when stockmarket volatility is in its upper decile.Figure 4: Test - USA - VXO24Note: The distribution of the difference between the impulse responses under high and lowlevels of stock market volatility as described in section 3.5. The shaded areas represent the84% and the 95% probability bands, respectively


Figure 5: Monetary Policy Shock - USA - Credit SpreadNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the credit spread. Low uncertainty depicts responses when the creditspread is in its lower decile. High uncertainty depicts responses when the credit spread isin its upper decile.Figure 6: Test - USA - Credit Spread25Note: The distribution of the difference between the impulse responses under high and lowlevels of the credit spread as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively


Figure 7: Monetary Policy Shock - USA - Forecaster DisagreementNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the forecaster disagreement index. Low uncertainty depicts responseswhen the uncertainty measure is in its lower decile. High uncertainty depicts responseswhen the uncertainty measure is in its upper decile.Figure 8: Test - USA - Forecaster Disagreement26Note: The distribution of the difference between the impulse responses under high and lowlevels of forecaster disagreement, as described in section 3.5. The shaded areas representthe 84% and the 95% probability bands, respectively


Figure 9: Monetary Policy Shock - USA - GoogleNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the Google-index. Low uncertainty depicts responses when the the Googleindexis in its lower decile. High uncertainty depicts responses when the the Google-indexis in its upper decile.Figure 10: Test - USA - Google27Note: The distribution of the difference between the impulse responses under high and lowlevels of the Google-index, as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively


Figure 11: Monetary Policy Shock - USA - Economic Policy UncertaintyNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the EPU-index. Low uncertainty depicts responses when the EPU-indexis in its lower decile. High uncertainty depicts responses when the EPU-index is in itsupper decile.Figure 12: Test - USA - Economic Policy Uncertainty28Note: The distribution of the difference between the impulse responses under high and lowlevels of the EPU-index, as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively


Figure 13: Monetary Policy Shock - AUS - Stock Market VolatilityNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the stock market volatility index. Low uncertainty depicts responses whenstock market volatility is in its lower decile. High uncertainty depicts responses when stockmarket volatility is in its upper decile.Figure 14: Test - AUS - Stock Market Volatility29Note: The distribution of the difference between the impulse responses under high and lowlevels of stock market volatility as described in section 3.5. The shaded areas represent the84% and the 95% probability bands, respectively


Figure 15: Monetary Policy Shock - CAN - Stock Market VolatilityNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the stock market volatility index. Low uncertainty depicts responses whenstock market volatility is in its lower decile. High uncertainty depicts responses when stockmarket volatility is in its upper decile.Figure 16: Test - CAN - Stock Market Volatility30Note: The distribution of the difference between the impulse responses under high and lowlevels of stock market volatility as described in section 3.5. The shaded areas represent the84% and the 95% probability bands, respectively


Figure 17: Monetary Policy Shock - CHE - Stock Market VolatilityNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the stock market volatility index. Low uncertainty depicts responses whenstock market volatility is in its lower decile. High uncertainty depicts responses when stockmarket volatility is in its upper decile.Figure 18: Test - CHE - Stock Market Volatility31Note: The distribution of the difference between the impulse responses under high and lowlevels of stock market volatility as described in section 3.5. The shaded areas represent the84% and the 95% probability bands, respectively


Figure 19: Monetary Policy Shock - NOR - Stock Market VolatilityNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the stock market volatility index. Low uncertainty depicts responses whenstock market volatility is in its lower decile. High uncertainty depicts responses when stockmarket volatility is in its upper decile.Figure 20: Test - NOR - Stock Market Volatility32Note: The distribution of the difference between the impulse responses under high and lowlevels of stock market volatility as described in section 3.5. The shaded areas represent the84% and the 95% probability bands, respectively


Figure 21: Monetary Policy Shock - NZL - Stock Market VolatilityNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the stock market volatility index. Low uncertainty depicts responses whenstock market volatility is in its lower decile. High uncertainty depicts responses when stockmarket volatility is in its upper decile.Figure 22: Test - NZL - Stock Market Volatility33Note: The distribution of the difference between the impulse responses under high and lowlevels of stock market volatility as described in section 3.5. The shaded areas represent the84% and the 95% probability bands, respectively


Figure 23: Monetary Policy Shock - UK - Stock Market VolatilityNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the stock market volatility index. Low uncertainty depicts responses whenstock market volatility is in its lower decile. High uncertainty depicts responses when stockmarket volatility is in its upper decile.Figure 24: Test - UK - Stock Market Volatility34Note: The distribution of the difference between the impulse responses under high and lowlevels of stock market volatility as described in section 3.5. The shaded areas represent the84% and the 95% probability bands, respectively


Figure 25: Monetary Policy Shock - AUS - SpreadNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the credit spread. Low uncertainty depicts responses when the creditspread is in its lower decile. High uncertainty depicts responses when the credit spread isin its upper decile.Figure 26: Test - AUS - Spread35Note: The distribution of the difference between the impulse responses under high and lowlevels of the credit spread as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively


Figure 27: Monetary Policy Shock - CAN - SpreadNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the credit spread. Low uncertainty depicts responses when the creditspread is in its lower decile. High uncertainty depicts responses when the credit spread isin its upper decile.Figure 28: Test - CAN - Spread36Note: The distribution of the difference between the impulse responses under high and lowlevels of the credit spread as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively


Figure 29: Monetary Policy Shock - CHE - SpreadNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the credit spread. Low uncertainty depicts responses when the creditspread is in its lower decile. High uncertainty depicts responses when the credit spread isin its upper decile.Figure 30: Test - CHE - Spread37Note: The distribution of the difference between the impulse responses under high and lowlevels of the credit spread as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively


Figure 31: Monetary Policy Shock - NOR - SpreadNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the credit spread. Low uncertainty depicts responses when the creditspread is in its lower decile. High uncertainty depicts responses when the credit spread isin its upper decile.Figure 32: Test - NOR - Spread38Note: The distribution of the difference between the impulse responses under high and lowlevels of the credit spread as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively


Figure 33: Monetary Policy Shock - NZL - SpreadNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the credit spread. Low uncertainty depicts responses when the creditspread is in its lower decile. High uncertainty depicts responses when the credit spread isin its upper decile.Figure 34: Test - NZL - Spread39Note: The distribution of the difference between the impulse responses under high and lowlevels of the credit spread as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively


Figure 35: Monetary Policy Shock - UK - SpreadNote: Impulse responses to a one percentage point increase in the interest rate, for twodifferent levels of the credit spread. Low uncertainty depicts responses when the creditspread is in its lower decile. High uncertainty depicts responses when the credit spread isin its upper decile.Figure 36: Test - UK - Spread40Note: The distribution of the difference between the impulse responses under high and lowlevels of the credit spread as described in section 3.5. The shaded areas represent the 84%and the 95% probability bands, respectively

More magazines by this user
Similar magazines