INVESTMENT, FINANCIAL FRICTIONS AND THE DYNAMIC EFFECTS OF MONETARY POLICY

Transcription

1 INVESTMENT, FINANCIAL FRICTIONS AND THE DYNAMIC EFFECTS OF MONETARY POLICY James Cloyne Clodomiro Ferreira Maren Froemel Paolo Surico March 2018 Abstract This paper assesses the role of financial frictions in business cycle fluctuations by looking at the response of firm-level investment to unanticipated interest rate changes. Using detailed demographic, balance sheet and income statement information, we find that age is a far stronger predictor (than liquidity, leverage and to a lesser extent size) of a large and significant capital expenditure response. Furthermore, the behaviour of young firms paying no dividends accounts for the bulk of the aggregate investment dynamics following a monetary policy shock. This group of firms has worse access to financial markets and face both an increase in leverage ratios and a decrease in Tobins Q after a rise in interest rates. Our evidence highlights the role of financial frictions as an important amplification mechanism of business cycle fluctuations. JEL classification: E21, E32, E52. Key words: monetary policy, investment, financial frictions, firm balance sheets. First Draft: March We thank staff in the Bank of England s Macroeconomic and Financial Analysis Division for the construction of the firm level dataset. Address for correspondence: James Cloyne (University of California Davis, NBER and CEPR) jcloyne@ucdavis.edu; Clodomiro Ferreira (Bank of Spain) clodomiro.ferreira@bde.es; Maren Froemel (London Business School) mfroemel@london.edu; Paolo Surico (London Business School and CEPR) psurico@london.edu. 1

2 1 Introduction Do credit conditions amplify business cycle fluctuations? What group of firms adjust their capital expenditure the most in response to changes in interest rates? Frictions in financial markets have played a prominent role in leading narratives of both the Great Depression of and the Great Recession of But, despite this, empirical evidence on how these frictions constrain firms decisions and whether they affect the transmission of monetary and other macroeconomic shocks is, at best, scant. This seems particularly surprising given that investment accounts for about 20% of GDP in advanced economies, is the most volatile component of GDP, and is likely to be one of the expenditure components most sensitive to changes in financial conditions and interest rates. This paper provides some novel macro evidence on the role of financial frictions in the monetary transmission mechanism by looking at the response of capital expenditure to unanticipated interest rate changes, using micro data on firms demographics, balance sheet positions and income statements. The goal of our analysis is twofold. On the academic side, it aims to revisit the empirical relevance on the financial accelerator, namely the notion that adverse shocks are amplified by worsening credit conditions, by providing direct evidence at the firm level. On the policy side, it seeks to inform future policy interventions by identifying which firm characteristics are most likely to predict a stronger response of investment to aggregate demand shocks and, more specifically, monetary policy. Empirical evidence based on structural VARs and aggregate time series from the national accounts reveals that investment and measures of financial frictions, such as interest rate spreads, are among the most responsive variables to monetary policy shocks (see for instance Christiano et al. (1999), and Gertler and Karadi (2015)). Still, by their very nature, aggregate analyses are silent on: (i) the specific channel(s) through which financial frictions distort the propagation of demand shocks onto capital expenditure and (ii) what group of firms drive the aggregate results. Our macro approach based on micro data at the firm level, in contrast, is well suited to explore heterogeneity along a plethora of demographic and balance sheet dimensions, so as to disentangle and quantify the theoretical mechanism(s) most relevant in the data; furthermore, it allows us to identify the individual characteristics that make the capital expenditure of a specific group of firms account for most of the aggregate investment dynamics following a demand shock. Our empirical strategy combines the benefits of micro data where we can exploit the heterogeneity in the cross section of firm-level investment with the benefits of macro data where we can 2

3 exploit the high frequency of monetary policy shock measured as surprise movements in short-term interest rate future within a 30 minutes window around a policy announcement (see Gurkaynak et al. (2005) and Gertler and Karadi (2015)). This compares favourably to a long-standing tradition of panel regression analyses which typically correlate firm-level characteristics, including capital expenditure, with aggregate variables such as short term interest rates. To the extent that monetary policy responds to aggregate conditions driven by a particular group of firms, our approach based on high frequency identification of policy shocks is not subject to this reverse causality concern which would otherwise be present in a micro data analysis using a short term interest rate as a regressor. Our main results can be summarized as follows. First, firms with less than 10 years since incorporation display the largest and most significant capital expenditure adjustment following a monetary policy shock. Older firms, in contrast, are characterized by small and insignificant changes. Second, this heterogeneity survives when we further split the age groups by liquidity and leverage but the converse is not true, revealing that age is a stronger predictor of larger investment adjustments, especially among small firms. Third, the capital expenditure response of young firms is concentrated among those paying no dividends. Fourth, only for this latter group, leverage and Tobin s Q move significantly after a monetary policy shock whereas changes in interest payments and net sales are far more muted and less heterogeneous. We conclude that the capital expenditure response of young firms especially small and paying no dividends drive the aggregate investment dynamics following an unanticipated interest rate change and that our balance sheet evidence is consistent with a quantitatively important role for the financial accelerator in amplifying business cycle fluctuations. Our work relates to several strands of literature. A long-standing tradition in macro econometrics has used structural VARs and data from national statistics to measure the impact of monetary policy shocks on investment. We show that the heterogeneity uncovered in our micro data provides sharper inference and a novel interpretation of both the specific channel(s) of monetary transmission and the driver(s) of the aggregate response. A long-standing tradition in micro econometrics has looked at the response of firm-level capital expenditure to aggregate short term interest rates but has typically abstracted from reverse causality and mostly reported insignificant estimates. In two very recent studies, independently developed, Ottonello and Winberry (2018) and Jeenas (2018) look at the heterogeneous responses of investment to high frequency identified monetary policy shock across groups of low leverage and high leverage firms. We show that age and paying no dividends 3

4 are stronger predictors of a larger investment response and that the balance sheet position of firms evolves over time in response to changes in monetary policy. Finally, our empirical analysis provides some of the first direct evidence on the financial accelerator mechanism put forward by the theories of Bernanke and Gertler (1989), Bernanke et al. (1999) and Kiyotaki and Moore (1997). The paper is structured as follows. In Section 2, we present the macro data and the high frequency identification strategy that we use to extract a series of monetary policy shocks based on a monthly proxy-structural VAR. In Section 3, we report firm-level statistics from the micro data and the empirical strategy to match the annual frequency of the firm-level data with the monthly frequency of the policy shock series. Here, we also verify that the time series of investment aggregated from the micro data as well as the average response to the monetary policy shock in the panel regressions is consistent with the macro evidence based on national statistics and time series models. The heterogeneous adjustment in capital expenditure following an unanticipated interest rate change is the focus of Section 4, where we emphasize that our newly constructed measure of age is a stronger predictor than size, liquidity or leverage of a larger firms response. Finally, in Section 5, we provide some evidence on the specific channels of monetary transmission, and the financial accelerator in particular, by reporting the responses of several balance sheet and income statement variables. The Appendix conducts further, complementary analyses and robustness checks. 2 Macro data and monetary policy identification Identifying the dynamic causal effects of changes in monetary policy requires isolating exogenous variation in the monetary policy interest rate. This issue is well-known in the empirical macro literature and, although we are interested in the heterogeneous responses across firms using micro data, this is still a key challenge that we need to tackle. Our micro-econometric regressions will be based on firm-level panel data. But common macro shocks will still drive firm-level decisions and aggregate monetary policy changes. Furthermore, this issue cannot be tackled using time fixed effects these would be collinear with the aggregate interest rate. Our identification strategy is based on the proxy-var/external instruments approach of Mertens and Ravn (2013), applied to monetary policy in the United States by Gertler and Karadi (2015). The idea is to isolate surprise movements in interest rates using the movements in financial markets data in a short window around central bank policy announcements. Building on Gurkaynak et al. (2005), Gertler and Karadi (2015) measure financial market surprises from Fed Funds Futures 4

5 contracts using a small window around Fed policy announcements. The, very plausible, identifying assumption is that nothing else occurs within the window (that would drive both private sector behavior and Fed policy decisions). The technical innovation in Gertler and Karadi (2015) is to use these high frequency surprises as proxies for the true structural monetary policy shocks in a Vector Autoregression. This approach has recently been applied to the United Kingdom by Gerko and Rey (2017) and we make use of their high frequency data. 1 The Gerko and Rey (2017) shocks are based on surprise movements in Short-Sterling Futures, derived using the same approach as in Gurkaynak et al. (2005). More details can be found in Gerko and Rey (2017). These data are available for the period but, ideally, we would like a series of monetary policy shocks that span the full sample covered by our firm-level micro data ( ). One advantage of the Gertler and Karadi (2015) proxy-var (as noted by Stock and Watson (2018)) is that the VAR can be estimated on a longer sample, and the identification of the contemporaneous causal relationships can be conducted using the sample for which the proxy/instrument is available. Our innovation is to estimate a Gertler and Karadi (2015) type-var for the United Kingdom and extract a series of monetary policy shocks that can be used with the full sample spanned by our micro data. To extract the monetary policy shocks, we first estimate a proxy-var for the UK following Gerko and Rey (2017). In particular, we estimate a monthly reduced-form VAR including 5 year gilt yields (as the measure of interest rates, as in Gerko and Rey (2017)), log industrial production, the employment rate, the unemployment rate, the log of the retail prices index (excluding mortgage interest payments), the Gerko and Rey (2017) measure of corporate spreads (mortgage rates minus the 3-month bill rate) and the dollar-sterling exchange rate. The structural VAR is given by: AX t = B(L)X t 1 + ɛ t (1) where X is the vector of endogenous variables and ɛ t is a vector containing the structural shocks, one of which is the true, but latent, monetary policy shock. Using the high frequency surprises as proxies/instruments, the method in Mertens and Ravn (2013) allows us to identify the elements of the structural impact matrix A that are relevant for measuring the effects of monetary policy. 1 These data are available from the journal s website. 5

6 Using the mapping between the reduced-form residuals u t and ɛ t : u t = A 1 ɛ t we are able to extract the implied series of monetary policy shocks. These are shown in Figure 1. We then use these monetary policy shocks in our micro panel regressions. Since these are exogenous disturbances, there is no need to include further time fixed effects or additional macro controls. Furthermore, since firms are typically small relative to the size of the economy, the is no reverse causality problem, an issue that typically afflicts the time series literature. To estimate the dynamic causal effects from the micro data we use a panel local projection instrumental variable technique, following Jorda et al. (2017). This is extremely flexible and allows us to estimate impulse response functions on firm level panel data using monetary shocks as instruments for interest rate changes. With this technique, it will then be straightforward to conduct multivariate heterogeneity analysis in a dynamic setting. 2 Before turning to the precise micro-econometric specification, it is first useful to study whether monetary policy has an effect on business investment in the aggregate. Official business investment data are available at quarterly frequency. As is common in the macro literature, we therefore sum our monetary policy shocks to quarterly frequency. A simple regression will then allow us to uncover the impulse response functions for business investment and, to keep the specification close to the micro regressions below, we estimate the following sequence of local projections: y t+h y t 1 = α h + β h R t + ν t+h (2) where R t is the end-of-quarter 5 year gilt yield (as in the VAR) and this is instrumented using our extracted series of monetary policy shocks summed to quarterly frequency. 3 y t+h is log real quarterly business investment and the β h refers to the impulse response function at period h. 4 Figure 2 shows that a initial 25bp rise in the 5 year interest rate leads to a fall in business 2 Ottonello and Winberry (2018) run panel regressions assuming that the high frequency shock is the true monetary policy shock. Our methodology is therefore more general, and allows us to estimate the effects of monetary policy in panel data where the instrument is only available at the macro level for a shorter time period. 3 It is possible to estimate these macro responses in one step using the proxy-var, although this requires summing the instruments to quarterly frequency. We prefer to instrument interest rates at a monthly frequency and estimate the quarterly IRFs ex-post. This two-step approach is also more consistent with the way we estimate the effects on investment using the micro data. 4 Generally, the error terms exhibit serial correlation, so these are corrected using the Newey-West method. 6

7 investment of around 0.5-1% after 2 years. Qualitatively and quantitatively these findings are consistent with the Gertler and Karadi (2015) results for industrial production and reflect the contractionary effects of interest rate hikes on activity. In the micro analysis below we will also find a similar relationship and this will be our point of departure for studying various dimensions of heterogeneity. 3 Micro data and empirical framework In this section we describe the firm-level data and the empirical framework employed in the estimation of heterogeneous effects of monetary policy shocks. 3.1 Firm-level data Detailed balance sheet and income statement data comes from Thomson Reuters WorldScope database. It provides detailed / high quality information on all balance sheet and income statement components as reported by publicly traded firms in several countries, although we restrict to firms physically operating in the U.K.. Firms report on a fiscal year basis 5 ; crucial for our identification strategy, though, they do so in different months. Thomson Reuters started collecting data since 1980, but consistent information for a sufficiently large number of firms only starts in We then restrict attention to the period January December At the highest level of sample selection, we discard firms with SIC codes in the finance, insurance, real estate or public administration sectors. All these characteristics allow us to construct a panel of firms with high quality data, for a relatively long period of time compared to other available data-sets. The main disadvantage is that we do not have information for non-publicly traded firms, and the fact that going public is an endogenous decision. We come back to this issues later in the text. Main variables. Our benchmark investment measure is the investment rate i j,t k j,t, where i j,t capital expenditures net of disposals / sales carried out by firm j throughout a fiscal year ending in t, and k j,t 1 is the level of physical capital, as captured by net plant, property and equipment, at the beginning of the reporting year t. In addition to being a standard measure in the investment 5 WorldScope provides some variables at the interim / quarterly frequency within a fiscal year, as well as information for some private firms; however, such information is scarce, so we restrict to publicly traded firms annual reports. 7

8 literature (see for example Chaney et al. (2012)), it is the relevant investment measure in standard theories. As Cooper and Haltiwanger (2006) have shown for the US, the distribution of the investment rate exhibits fat tails and has a significant mass around 0. This reflects the nature of investment and adjustment costs at the firm (or plant) level. Our main contribution is to shed light on the role of a firm s age as a source of heterogeneity in the effects of monetary policy, over and above the more traditional firm characteristics such as size, leverage and liquidity. We define a firm s age as the number of years since incorporation. Though this measure is different from founding date, it is reported by a larger number of firms and measured with more precision. Following standard practice, we define the size of a firm in terms of the number of employees at the beginning of the fiscal year. Leverage is defined as the ratio of total debt to total assets. Total debt includes both short and long term obligations, while the measure for total assets is the book value as reported at the beginning of the fiscal year. Liquidity is defined as the ratio of cash and short-term investments over total assets. Tobins Q is defined as the ratio of market value to book value of total assets. 6 The Appendix presents further details on the data-set, variables and definitions. Table 1 presents some descriptive statistics on firm characteristics for two age groups: firms with age up to 10 years ( younger ) and firms with age above 10 ( older ). There are a few features worth noting. First, the average age for younger firms is around 6 years, while it is 47 years for older firms. Second, older firms are far more likely to pay dividends, pay higher dividends, and have much higher employment investment levels; in contrast, the investment rate, investment and employment growth, and the change in the investment ratio are significantly larger for younger firms. Third, the leverage ratio distribution is more skewed for younger firms: the average leverage ratio and the upper quartile for the two groups is virtually identical, whereas the lower quartile and median are significantly lower for younger firms. Comparison with aggregate investment from national accounts. As a first step, we are interested in understanding how much of the total aggregate investment level and dynamics is captured by publicly traded firms in the U.K. 7 To do so, we need to aggregate the investment 6 We calculate the market value as the sum of the market value of its outstanding stock at the end of the accounting year (market capitalization) plus the value of its outstanding debt. This is an upper bound on Tobins Q, as we do not subtract deferred taxes from the value of debt, as this variable has many missing values in our dataset. 7 Note that, differently from a survey where we would use sampling weights in order to corroborate that the sample is representative for the population, here we are not dealing with a sampling issue. 8

9 reported by each firm for a given fiscal year into a measure of investment at the calendar quarter / year. We proceed in the following way: (i) for each firm we first distribute the amount of capital expenditures reported at the end of the fiscal period uniformly throughout the corresponding months; (ii) we then aggregate investment at the monthly frequency by summing up all the firmmonth investment amounts. In this way, we can construct a calendar quarter aggregate investment level. The comparison with the aggregate series from the Office of National Statistics (ONS), which includes investment by both public and private firms, is presented in Figure 3. 8 The left panel compares the levels, while the right panel compares the quarter-on-quarter growth rates as computed from the level series. Similarly to the well-known facts in the US (see for example Covas and Den Haan (2011)), both the level of investment, as well as its unconditional dynamics, of total aggregate investment is explained by publicly traded firms. Compared to private (non-traded) firms, public firms tend to be relatively big. We will show, however, that the behavior of investment after a monetary policy shock presents a different degree of heterogeneity, over and above the small-big dichotomy. 9 Baseline sample selection for the empirical analysis. We discard outliers in the following way: (i) the investment ratio is trimmed by year; if more than 1% of observations exceed 10, we discard the top 1%, else all observations above 10 are discarded. (ii) we discard negative values of capital expenditures and firms with zero employees; (iii) we discard the top 1% of the leverage ratio; (iv) we discard the top and bottom 1% of the equity ratio; (v) we discard observations where growth of sales and growth of cost of goods sold is in the top / bottom one percent; (vi) We discard the top 1% of Tobin s Q; (vii) discard observations where the accounting identity is violated by more than 5% of total assets; (viii) liquidity ratio higher than In the estimations below, we keep firms with at least 5 consecutive years of data such that each impulse response function contains complete data for all firms. 8 The divergence at the end of the sample of our series is due to the behaviour of firms in the Basic material and Oil and Gas sector, which showed a much more pronounced increase/decrease in capital expenditure abroad than other firms in their total expenditure between Capital expenditure abroad by domestic firms is not included in the ONS investment series. See figure 18 in the appendix, which was kindly provided to us by the BoE. The main results are robust to including firms from these sectors. 9? present similar heterogeneity for employment decisions of private firms in the U.K. 10 We also drop all observations that miss an incorporation date, and current and previous accounting year observations for the number of employees, cash dividend payments, leverage liquidity, number of employees, and the book value of equity. 9

10 3.2 Empirical specification In order to capture the heterogeneous effects of monetary policy, in our benchmark empirical specification we estimate impulse response functions using an instrumental variable (IV) variation of Jorda (2005)s local projection (LP) approach (a LP-IV) ( ) ( ) i i 11 G G = α j + δ m,h mth j,m + α g 11 k j,t+h k {grpj =g} + β g,h 11 {grpj =g} r m,t +ɛ j,t+h j,t 1 m=1 g=1 g=1 (3) mth j,m is a dummy that takes value 1 if firm j reports in month m, r m,t is the 5-yr gilt as recorded at the end of the reporting month m in year t, and 11 {grpj =g} is a dummy that takes value 1 if firm j belongs to group g in year t. We estimate (3) first separately for groupings by age, size, liquidity and leverage, and use a double-sorting along age and each of the other variables. We allow for firm fixed-effects (FE), and standard errors are clustered at the firm level. 11 We instrument the policy rate r m,t in the following way: for each firm j reporting in month m of year t, we use the monetary policy shock in the accounting year as instrument, extracted and aggregated to annual frequency as discussed in the previous section. The advantage of the LP approach is that the IRF for horizon h is simply given by the coefficient β g,h. The results below present the IRF over a 4 year horizon (h=0,...,4) for a 25 basis points contractionary monetary policy shock. 3.3 The average effect We first estimate equation (3) without any heterogeneity. This can be interpreted as estimating the average effect of interest rates on the investment ratio in the micro data. Figure 4 reports the impulse response function (β h 0 for each horizon h). The investment ratio declines significantly following a 25bp rise in the 5 year interest rate. The effect is felt immediately in the first year and peaks in the second year. These dynamics are consistent with the impulse response functions using official aggregate data presented in Figure 2. This result is reassuring: the average impulse response for the investment ratio estimated from the micro data is consistent with the macro response of 11 Given that we are restricting to firms for which we observe at least 5 consecutive years of data, the firms that we keep might present an auto-correlation across time both in their errors and regressors which are not fully representative of the population of public firms. In addition, it is sensible to assume that the treatment assignment (i.e. the interaction of the interest rate with firm characteristics) presents correlation within firm and across time. For an up-to-date discussion of these issues, see for example Abadie et al. (2017). 10

11 investment using official aggregate time series. This gives us a meaningful benchmark against which to study the heterogeneous effects across different groups. 4 The heterogeneous response of capital expenditure In this section, we focus on heterogeneity in the capital expenditure responses to interest rate changes according to a firm age, namely the number of years since incorporation. More specifically, we show that grouping firms by age not only produces a more pronounced heterogeneity than grouping firms by size, liquidity or leverage (all reported in the Appendices C to E) but also that the differences in the investment responses along the demographic dimension are robust to controlling for other firm-specific characteristics using a double sorting strategy. 4.1 Results based on age Based on the firm-specific variable age constructed in Section 3, we group firms into three categories depending on whether at the time of the monetary shock they were operating below ten years (younger firms), between ten and fifty years (middle-aged firms) or above fifty years (older firms) since incorporation. The reason to focus on age is twofold. First, unlike other firms characteristics such as leverage, liquidity and (to a lesser extent) size, it is fully predictable and thus represents an ideal dimension along which splitting our sample. The other variables, in contrast, are far more likely to be endogenous. In Section 5, we will show that leverage and liquidity do respond to monetary policy shocks, thereby posing issues of selection and compositional changes across firms grouped along these dimensions. 12 Second, in a parallel with the consumption literature, age is likely to be correlated with the unobserved characteristics shaping a firms access to credit markets and therefore may represent a proxy for the degree of financial frictions that each group of firms face. In the next section, we will present evidence consistent with this hypothesis. The heterogeneous capital expenditure responses by age are reported in Figure 5: the top row refers to younger firms, the middle row to middle-aged firms and the bottom row to older firms. Three results clearly emerge from this exercise. First, there is pervasive evidence of considerable differences across groups. Second, the investment of younger firms is by far the most responsive: 12 Bahaj et al. (2018) show that a firms number of employees (a commonly used measure of size) respond to monetary policy shocks and that this result is significantly more marked for younger firms. 11

12 following a 25 basis points unanticipated changes in the interest rate, the peak occurs within the second year after the shock, just above 2%, before returning to zero, statistically, within the fourth year. Third, middle-aged firms investment dynamics share a similar time profile although the peak response is significantly weaker around 0.7%. In contrast, older firms adjust their capital expenditure by a small and insignificant amount. Combined with the descriptive statistics in Table 1, which shows that firms incorporated over the ten years before the shock hits are characterized by the largest (and most volatile) growth rates of capital expenditure, the estimates in Figure 5 reveals that younger firms are the ones driving the cumulated response of investment in the aggregate data (as well as the average effect estimated using the micro data) reported in Figure Controlling for size The evidence in Figure 5 suggests that age is a strong predictor for heterogeneity in the investment response to monetary policy shocks. In Section 5, we will show that age is also correlated with the presence of financial frictions at the firm-level. Here, we ask whether age is correlated with other characteristics which may be in fact the genuine driver of the heterogeneous responses in Figure 5. In particular, we will consider grouping firms also according to (i) their number of employees (size), (ii) cash and short term investment as a share of total assets (liquidity) and (iii) total debt as share of total assets (leverage). In the Appendices, we show that the evidence of heterogeneity by size, liquidity and leverage is weaker (and often non-monotonic) than when firms are grouped by age. In the rest of this section, we want to assess whether the heterogeneous responses by age groups are robust to netting out any possible heterogeneous effects due to size, liquidity and leverage. It is worth noting that unlike traditional panel regression analyses in which the identification exploits exogenous variation in the cross section, our identification is based on exogenous changes (in monetary policy) that vary over time but are common across firms. Accordingly, the notion of controlling for other characteristics requires a different approach than simply adding further regressors to our baseline empirical specification. More specifically, we further split the age groups according to the second dimension of interest, a strategy that we refer to as double sorting. By looking at the differences between small-young (old) firms and large-young (old) firms, for instance, we will be able to assess the marginal contribution of size for a given age. Similarly, by 12

13 comparing the response of small (large) young firms to small (large) old firms, one can infer the marginal contribution of age for a given size. As double sorting is very demanding on the data, we maximize the number of observations per sub-groups by using only two categories of age and two categories of size. Firms are counted as younger (older) if they have less (more) than ten years since incorporation at the time of the shock whereas are deemed as smaller (larger) if they have less (more) than 250 employees in the year before the shock hits. Figure 5 and Figure 15 in the Appendix confirm that this specific 2x2 choice for this double sorting is representative of the heterogeneity we uncovered based on a large number of groups for each dimension in isolation. The results based on age controlling for size are presented in Figure 6 for firms that are smalleryounger (top-left panel), larger-younger (bottom-left panel), smaller-older (top-right panel) and larger-older. The marginal impact of size can be read across rows whereas the marginal impact of age can be read across columns. Smaller-younger firms in the top-left corner exhibit the largest capital expenditure response following a monetary policy shock with a peak around 4% during the second year, reverting to insignificant values after four years. In contrast, smaller-older firms, which are recorded in the same row of Figure 6, displays a far smaller and statistically negligible response, which is significantly different from the investment dynamics of smaller-younger firms in the top-left panel but is very similar to the response of larger-older firms in the bottom-right panel. Finally, a comparison of the two charts on the left column suggests that size does play a marginal role in further amplifying the capital expenditure adjustment of younger firms but it has no marginal effect on the older firms, as can be seen by comparing the two charts on the right coolumn of Figure 6. We conclude that age is a strong predictor of the heterogeneous responses of investment, over and above any possible marginal effect of size, and that the effects of monetary policy is particularly pronounced among firms who are both young and small. 4.3 Controlling for liquidity In Figure 7, we apply our double sorting strategy to liquidity. We keep the same two categories of age as in the previous sub-section but further split these groups depending on whether their liquidity position is below (low liquidity) or above (high liquidity) the median of the distribution of cash and short term investment over total asset during the year before the shock. 13 A comparison of the two panels in the top row reveal that young firms always adjust their capital expenditure 13 Similarly resulted are obtained here and for leverage in the next sub-section if the median is computed over the previous two or three years distribution. 13

14 after a monetary policy shock by a large and significant amount, which is only slightly (but not significantly) larger for younger-low liquidity firms. Similarly, the response of older firms in the bottom row is statistically indistinguishable from zero independently from the level of liquidity held across the two sub-groups in each column. In summary, the heterogeneity in liquidity positions reported across age groups in Table 1 does not seem to be responsible for the heterogeneity in capital expenditure dynamics recorded in Figure 5. A possible explanation for this result, further corroborated by the evidence in Appendix D, is that firms adjust their liquidity position in response to business cycle fluctuations, possibly reflecting current and expectations of future access to financial markets. Accordingly, grouping by liquidity may pool together two very different set of firms: for instance, one group may hold high liquidity because is unconstrained whereas another group may hold high liquidity because anticipate hard time to obtain credit in the future. 4.4 Controlling for leverage Another significant dimension of balance sheet heterogeneity across age groups in Table 1 is leverage, which in our sample is typically lower for many younger firms but also for few older firms. Applying the same logic used above for liquidity, a low leverage ratio could be consistent not only with firms that have little access to credit markets and thus are unable to lever on their (possibly few) assets but also with a very different set of firms that have sufficient assets/cash flows to be less reliant on credit markets. Looking at the response of investment using a double sorting strategy based on age and leverage could help shedding some lights on the potential ambiguity of interpretation by looking across leverage groups only. In Figure 8, we assess the marginal contribution of leverage for a given age (and of age for a given leverage) by dividing the younger and older categorization into two further groups with leverage ratios below (low leverage) or above (high leverage) the median of the firms distribution in the year before the shock. A comparison of the capital expenditure response of younger-lower levered firms in the top left panel with the adjustment of older-lower levered firms in the bottom-left panel suggest that being younger predict a far larger investment response conditional on being in the low leverage group. The response heterogeneity in age is only slightly less pronounced among high levered firms in the left column of Figure 8. In contrast, conditional on being old (bottom row), having lower leverage does not seem to trigger a stronger capital expenditure response, though it 14

15 does seem to make a marginal contribution among younger firms (see difference between the two top panels). In summary, the evidence of this section support the notion that age is an important driver of the heterogeneity in the capital expenditure responses to monetary policy shocks across firms, over and above any possible heterogeneity across size, liquidity and leverage. The heterogeneity of the investment adjustments in the latter seem far less marked (see Appendices C to E) and become only marginal significant, if any, once we control for a firms age using our double sorting strategy. 5 On the channels of transmission In the previous section, we have shown that age is a robust predictor of a larger and significant response of capital expenditure to a monetary policy shock. In the Appendix, we report that the estimated heterogeneity is significantly weaker when firms are grouped along more traditional dimensions such as size, liquidity or leverage. In this section, we use other balance sheet and income statement variables to interpret the results by age. In particular, we double sort the sample by dividing the demographic groups into firms paying or not paying dividends, and show that the change in investment after a policy shock is far larger and significant only for the latter. To explore this finding further, we look at firms leverage and Tobins Q as left-hand variable in an otherwise identical double sorting specification by age and paying dividends. Finally, we look at the response of interest payments and net sales, which we interpret as proxies for a cash-flow effect on existing debt and for a demand effect. Following a contractionary monetary policy shock, leverage (Tobins Q) increases (decreases) significantly only for young firms paying no dividends, consistent with the predictions of a financial accelerator mechanism of monetary transmission. In contrast, the impact on interest payments and net sales appears muted and more homogeneous 5.1 Paying dividends A possible interpretation of the demographic groups estimates of the previous section is that age is a more accurate predictor (than size, liquidity or leverage) for the heterogeneity in the access to financial markets. To corroborate this interpretation, we divide firms in younger (if incorporated since less than 10 years at the time of the shock) and older (otherwise). Each of these groups is further decomposed into firms that pay no dividends and firms that pay dividends. To the extent 15

16 that age is correlated with the presence and severity of credit constraints, we would expect young firms paying no dividends to adjust their investment the most following a monetary policy shock. The response of capital expenditure for this four-way categorization is reported in Figure 9. Three results emerge from this exercise. First, young firms paying no dividends exhibit a very large and significant response. Second, young firms paying dividends displays a smaller, yet statistically significant adjustment. Third, the capital expenditure of old firms is negligible and statistically indistinguishable from zero, independently on whether dividends are paid out. We conclude that the response of young firms paying no dividends drives the aggregate investment dynamics following a monetary policy shock and conjecture that this group may face tighter access to credit markets after a hike in interest rates. We assess this hypothesis in the rest of this section. 5.2 The response of leverage and Tobin s Q The financial accelerator theory of Bernanke and Gertler (1989) and Bernanke et al. (1999) predicts that, following a contractionary monetary policy shock, the leverage of constrained firms should increase whereas their Tobins Q should decrease. The reason is that the value of the collateral used to access credit falls together with the equity value and it does so more sharply than debt because the returns on capital are now higher. This results in constrained firms being forced to disinvest as the decrease in asset prices worsens financial conditions in the aggregate and at the firm-level. To investigate these issues, in Figure 10, we double sort firms by age and dividends paid but the dependent variable is now leverage defined as total debt over total assets. Two main results are worth emphasizing. First, the leverage of younger firms paying no dividends increases significantly after a rise in interest rates. Second, the responses of all other sub-groups are small and often not statistically different from zero. A very similar picture emerges from Figure 11, whose specification is all alike the one behind Figure 10 but the left-hand side variable, which is now Tobins Q defined as the ratio of market value to book value of a firms assets. Consistent with the estimates using leverage, only younger firms paying no dividends experience a fall in Tobins Q after a contractionary monetary policy shock, consistent with the theoretical predictions of the financial accelerator in Bernanke et al. (1999). 16

17 5.3 Interest payments Another potentially significant change in the resources available to firms after a monetary policy shock may come from interest payments, both in the form of new originations for firms with shortterm debt and, likely more important, in the form of existing debt contracted at an adjustable rate. To explore this mechanism, in Figure 12 we resort to the double sorting by age and dividends paid used in the rest of this section but look at the log of interest payments as dependent variable. The main inference we draw from this analysis is that while there seems to be a slightly more pronounced response for younger firms who pays no dividends, this is neither large nor significantly different from the interest payment changes for the other groups, which are typically statistically indistinguishable from zero. We conclude that while this mechanical interest cash flow effect moves qualitatively in the right direction (and may have made some small contribution), its overall magnitude is quantitatively modest and thus unlikely to play a major role to explain the investment response to interest rate changes. 5.4 The demand effect Another channel for the transmission of interest rate changes to capital expenditure may come from aggregate demand through a standard, neo-classical intertemporal substitution effect. The idea if that a hike in interest rates encourages savings and depresses consumption with negative consequences on firms sales. This effect may be further amplified by a contraction in labour demand, which may provide an additional motive to cut back households spending (see Cloyne et al. (2017)). The lower sales imply, in turn, lower cash flows available to firms not only to run their operations but also to invest in new projects and machines. A way to assess the empirical relevance of this mechanism is to look at the response of sales (net of operating costs and inventories) to a monetary policy shock. This is recorded in Figure 13, where the bottom (top) row refers to firms paying (no) dividends while the left (right) column displays the estimates for younger (older) firms. The responses of net sales appear rather homogeneous across the four groups and rarely statistically different from zero. We interpret these relatively small and often insignificant estimates as evidence against a quantitatively important role for an aggregated demand channel. Furthermore, we note that the lack of heterogeneity in Figure 13 implies that, despite all groups facing a similar change in cash flows through sales, only younger firms paying no dividends adjust their capital expenditure following a monetary policy shock, consistent with the 17

18 notion that these firms face a liquidity constraint. 6 Conclusions It is well-known that aggregate investment is particularly sensitive to changes in monetary policy. And, investment is one of the key channels of monetary transmission in a range of standard macro models. But, despite this, there is still relatively little evidence on why monetary policy affects firm investment, and which firms are likely to be the most sensitive to changes in monetary policy. We combine high-frequency external instrument techniques from the macro literature with a rich plethora of firm-level panel micro data. Using a panel local projections-iv approach, our contribution is to provide a systematic evaluation of the relevant dimensions of heterogeneity in the response of firms to monetary policy. Furthermore, we examine what this evidence can tell us about the monetary transmission mechanism. Across all the dimensions we consider, firm age seems to be the most robust predictor of having the highest sensitivity to changes in interest rates. Furthermore, young firms who are small and do not pay dividends seem particularly sensitive. Our rich set of heterogeneity results is also informative about the underlying channels. Different dimensions of heterogeneity are likely to be correlated. But our approach allows us to consider particular variables conditional on all the others. Specifically, heterogeneity by firm age survives when we further split the age groups by size, liquidity and leverage. Furthermore, we show that balance sheet variables respond more significantly for young firms paying no dividends. More specifically, we observe sizable falls in Tobin s Q for these firms and a rise in leverage (possibly reflecting the fall in asset values) following a contractionary monetary policy shock. In contrast, conditioning on age and dividend status, we do not see much difference across firms in the response of interest cash flows (which might reflect a mechanical effect) or sales growth (which might reflect heterogeneity in demand). Taken together, our evidence is consistent with a quantitatively important role for the financial accelerator in amplifying business cycle fluctuations. 18

19 Figures Figure 1: Monthly monetary policy shock Series ExtractedShock Time Notes: Extracted monetary policy shock based on a proxy-var estimated over the sample period 1982m1-2014m12. The proxy variable is included for 2003m1-2014m12. The first stage F-statistic is

20 deviation deviation deviation deviation deviation deviation Figure 2: Response of investment to monetary policy 1 INV (ONS) 0.1 Emp. Rate 0.2 Unemp. Rate Quarters log GDP Quarters log RPIX3 (SA) Quarters 5-yr gilt Quarters Quarters Quarters Notes: This figure shows the impulse response functions for various macroeconomic variables to a 25bps rise in the 5 year gilt yield. Gilt yields are instrumented using the extracted shock discussed in the text. ONS data from 1983 to

21 Figure 3: Investment over time: ONS vs. Worldscope (a) Investment levels: ONS vs. Worldscope log investment q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 investment (ONS) investment (WS) (b) Investment growth: ONS vs. Worldscope Correlation.58 (pvalue = 0) -on-year growth rates q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 investment (ONS) investment (WS) Notes: the growth rates on the right are computed from the levels on the left panel. We aggregate WorldScope firm level data in the following way: (i) for each firm we first distribute the amount of capital expenditures reported at the end of the fiscal period uniformly throughout the corresponding months; (ii) we then aggregate investment at the monthly frequency by summing up all the firm-month investment amounts. 21

22 Figure 4: Response of the investment ratio to a monetary policy contraction Notes: The figure shows the (average) impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

23 Figure 5: Response of the investment ratio by firm age (a) Age < 10 years (b) Age: years (c) Age > 49 years Notes: The figure shows the impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield by firm age. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

24 Figure 6: Response of the investment ratio by firm size and firm age Young firms (< 10 years) (a) Small firms Old firms (> 10 years) (b) Small firms (c) Large firms (d) Large firms Notes: The figure shows the impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield by firm age and firm size. Large firms refer to those in the top quartile of the size distribution. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

25 Figure 7: Response of the investment ratio by firm age and liquidity Young firms (< 10 years) (a) Low liquidity Old firms (> 10 years) (b) Low liquidity (c) High liquidity (d) High liquidity Notes: The figure shows the impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield by firm age and liquidity. Low liquidity firms refer to those in the bottom quartile of the liquidity ratio distribution. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

26 Figure 8: Response of the investment ratio by firm age and leverage Young firms (< 10 years) (a) Low leverage Old firms (> 10 years) (b) Low leverage (c) High leverage (d) High leverage Notes: The figure shows the impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield by firm age and leverage. Highly levered firms refer to those in the top quartile of the leverage ratio distribution. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

27 Figure 9: Response of the investment ratio by firm age and dividends status Young firms (< 10 years) (a) No dividends paid Old firms (> 10 years) (b) No dividends paid (c) Dividends paid (d) Dividends paid Notes: The figure shows the impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield by firm age and whether the firm paid dividends. Dividends paid refers to whether a firm paid dividends in the previous year. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

28 Figure 10: Response of leverage by firm age and dividends status Young firms (< 10 years) (a) No dividends paid Old firms (> 10 years) (b) No dividends paid (d) Dividends paid (c) Notes: The figure shows the impulse response function for leverage to a 25bps increase in the 5 year gilt yield by firm age and whether the firm paid dividends. Dividends paid refers to whether a firm paid dividends in the previous year. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

29 Figure 11: Response of Tobin s Q by firm age and dividends status Young firms (< 10 years) (a) No dividends paid Old firms (> 10 years) (b) No dividends paid Change in ratio Change in ratio (c) Dividends paid (d) Dividends paid Change in ratio Change in ratio Notes: The figure shows the impulse response function for Tobin s Q to a 25bps increase in the 5 year gilt yield by firm age and whether the firm paid dividends. Dividends paid refers to whether a firm paid dividends in the previous year. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

30 Figure 12: Response of interest expenditures by firm age and dividends status Young firms (< 10 years) (a) No dividends paid Old firms (> 10 years) (b) No dividends paid (c) Dividends paid (d) Dividends paid Notes: The figure shows the impulse response function for interest expenditures to a 25bps increase in the 5 year gilt yield by firm age and whether the firm paid dividends. Dividends paid refers to whether a firm paid dividends in the previous year. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

31 Figure 13: Response of sales by firm age and dividends status Young firms (< 10 years) (a) No dividends paid Old firms (> 10 years) (b) No dividends paid (c) Dividends paid (d) Dividends paid Notes: The figure shows the impulse response function for sales growth to a 25bps increase in the 5 year gilt yield by firm age and whether the firm paid dividends. Dividends paid refers to whether a firm paid dividends in the previous year. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

32 Tables Table 1: Firm Heterogeneity Overview: Aggregates (1) Inv DeltaInv InvRatio DeltaRatio age Empl DeltaEmpl Dividends DebtAssets Liquidity mean p p p

33 References Abadie, A., S. Athey, G. Imbens, and J. Wooldridge (2017). When should you adjust standard errors for clustering? Manuscript. Bahaj, S., A. Foulis, G. Pinter, and P. Surico (2018). The employment effects of monetary policy: Macro evidence from firm-level data. Manuscript. Bernanke, B. and M. Gertler (1989, March). Agency Costs, Net Worth, and Business Fluctuations. American Economic Review 79 (1), Bernanke, B., M. Gertler, and S. Gilchrist (1999). The financial accelerator in a quantitative business cycle framework. Handbook of Macroeconomics 1, Chaney, T., D. Sraer, and D. Thesmar (2012, May). The collateral channel: How real estate shocks affect corporate investment. American Economic Review 102 (6), Christiano, L. J., M. Eichenbaum, and C. L. Evans (1999). Monetary policy shocks: What have we learned and to what end? In J. B. Taylor and M. Woodford (Eds.), Handbook of Macroeconomics, Volume 1A, Chapter 2, pp North-Holland. Cloyne, J., P. Surico, and C. Ferreira (2017). Monetary policy when households have debt: new evidence on the transmission mechanism. Manusript. Cooper, R. W. and J. C. Haltiwanger (2006). On the nature of capital adjustment costs. The Review of Economic Studies 73 (3), Covas, F. and W. J. Den Haan (2011, April). The cyclical behavior of debt and equity finance. American Economic Review 101 (2), Gerko, E. and H. Rey (2017). Monetary policy in the capitals of capital. Journal of the European Economic Association 15 (4), Gertler, M. and P. Karadi (2015). Monetary policy surprises, credit costs, and economic activity. American Economic Journal: Macroeconomics 7(1), Gurkaynak, R., B. Sack, and E. Swanson (2005). Do actions speak louder than words? the response of asset prices to monetary policy actions and statements,. International Journal of Central Banking 1(1). 33

34 Jeenas, P. (2018). Monetary policy shocks, financial structure, and firm activity: A panel approach. Manuscript. Jorda, O. (2005, March). Estimation and inference of impulse responses by local projections. American Economic Review 95 (1), Jorda, O., M. Schularick, and A. Taylor (2017). The effects of quasi-random monetary experiments. Federal Reserve Bank of San Francisco Working Paper ( ). Kiyotaki, N. and J. Moore (1997). Credit cycles. Journal of Political Economy 105(2). Mertens, K. and M. O. Ravn (2013). The dynamic effects of personal and corporate income tax changes in the United States. American Economic Review 103(4), Ottonello, P. and T. Winberry (2018). Financial heterogeneity and the investment channel of monetary policy. NBER Working Paper (24221). Stock, J. H. and M. W. Watson (2018). Identification and estimation of dynamic causal effects in macroeconomics using external instruments. (24216). 34

35 Appendix A Evidence from a monthly proxy-svar Figure 14: Response of investment to monetary policy using a monthly proxy-svar 5-yr gilt log IP Emp. Rate Months Unemp. Rate #10-3 Months log(rpix3) SA Months Corporate Spread Months Xrate Months Months Months Notes: Impulse response functions for all the macroeconomic variables in the monthly VAR, following Gerko and Rey (2017). This model is used to extract the series of monthly monetary policy shocks. See the text for more details. 35

36 B Firm level data Table 2: Descriptive statistics for the full sample Table 3: Firm Heterogeneity Overview: Aggregates Inv DeltaInv InvRatio DeltaRatio age Empl DeltaEmpl Dividends DebtAssets Liquidity mean p p p (1) C Results based on size Table 4: Descriptive statistics by size groups Table 5: Firm Heterogeneity Overview by Quartiles of size (employees) Inv DeltaInv InvRatio DeltaRatio age Empl DeltaEmpl Dividends DebtAssets Liquidity Quartile 1 mean p p p Quartile 2 mean p p p Quartile 3 mean p p p Quartile 4 mean p p p (1) 36

37 Figure 15: Response of the investment ratio by firm size (a) Quartile 1: small firms (b) Quartile 2: small-medium firms (c) Quartile 3: medium-large firms (d) Quartile 4: large firms Notes: The figure shows the impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield by quartiles of the firm size distribution. Firm size is based on the number of employees. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

38 D Results based on liquidity Table 6: Descriptive statistics by liquidity ratio. Table 7: Firm Heterogeneity Overview by Quartiles of liquidity Inv DeltaInv InvRatio DeltaRatio age Empl DeltaEmpl Dividends DebtAssets Liquidity Quartile 1 mean p p p Quartile 2 mean p p p Quartile 3 mean p p p Quartile 4 mean p p p (1) 38

39 Figure 16: Response of the investment ratio by liquidity (a) Quartile 1: low liquidity (b) Quartile 2: low-medium liquidity (c) Quartile 3: medium-high liquidity (d) Quartile 4: high liquidity Notes: The figure shows the impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield by quartiles of the liquidity ratio distribution. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

40 E Results based on leverage Table 8: Descriptive statistics by leverage. Table 9: Firm Heterogeneity Overview by Quartiles of leverage Inv DeltaInv InvRatio DeltaRatio age Empl DeltaEmpl Dividends DebtAssets Liquidity Quartile 1 mean p p p Quartile 2 mean p p p Quartile 3 mean p p p Quartile 4 mean p p p (1) 40

41 Figure 17: Response of the investment ratio by leverage (a) Quartile 1: low leverage (b) Quartile 2: low-medium leverage (c) Quartile 3: medium-high leverage (d) Quartile 4: high leverage Notes: The figure shows the impulse response function for the investment ratio to a 25bps increase in the 5 year gilt yield by quartiles of the leverage distribution. These are estimated using Worldscope micro data. The panel LP-IV specification is discussed in Section

42 Figure 18: Aggregated capital expenditure by UK-listed PNFCs decomposed by industry. Source: Thomson Reuters and Bank calculations. 42