Labour Reallocation and Productivity Dynamics: Financial Causes, Real Consequences

Transcription

1 Labour Reallocation and Productivity Dynamics: Financial Causes, Real Consequences Claudio Borio, Enisse Kharroubi, Christian Upper, Fabrizio Zampolli March 2016 Abstract This paper investigates causes and consequences of labour reallocation across sectors in advanced economies over the last 35 years. First, it shows that the contribution of labour reallocation across sectors to aggregate productivity growth systematically weakens during episodes of high credit growth: When credit outgrows GDP, workers move into low productivity gains sectors. Second we build a model where credit growth hurts productivity growth because entrepreneurs reallocate during credit booms into projects with relatively low return but high borrowing capacity. Moreover the model predicts that ressource allocation across sectors tends to be endogenously more persistent during episodes of tight credit conditions. We finally take this prediction to the data and show that the labour reallocation contribution to productivity growth measured prior to a recession affects significantly the path of productivity following the recession, and much more so when the recession is associated with a financial crisis. These results shed light on the recent secular stagnation debate and provide an alternative interpretation to the hysteresis hypothesis. Keywords: Labour reallocation, productivity, financial booms, hysteresis JEL codes: E24; E51; O47 We thank, Martin Brown, Andy Filardo, Paloma Garcia, Jonathan Kearns, Gianni Lombardo, Jose Luis Pedro, Thijs Van Rens, and seminar participants at Banque de France, the BoC-ECB conference on "The underwhelming global post-crisis growth performance determinants, effects and policy implications", the CEPR Conference on Persistent Output Gaps: Causes and Policy Remedies; the St-Gallen conference on "Finance, Capital Reallocation and Growth". The views expressed here are those of the authors and do not necessarily represent the views of the BIS. Monetary and Economic Department - Bank for International Settlements. 1

2 1 Introduction [... ] growth was adequate perhaps even good during the period. [But] It would not be right to say either that growth was spectacular or that the economy was overheating [... ]. And yet this was the time of vast erosion of credit standards, the biggest housing bubble in a century, the emergence of substantial budget deficits and what many criticize as lax monetary and regulatory policies.[...] US economic growth has averaged only 2% over the last 5 years [ ] despite having started from highly depressed state. In a similar vein, credit spreads in Europe have come way down and fears of the dissolution of the Eurozone have been sidelined, yet growth has been glacial over the past several years and is not expected to rapidly accelerate. Lawrence Summers, World Economic Forum, October 31, Five years after the worst financial crisis the world economy has experienced in the last 80 years, growth is still very low, particularly in Europe and in many countries, either output is still below pre-crisis levels or output growth is significantly below pre-crisis trends. Yet, if history shows that recovering from financial crises can be particularly difficult, it is striking to note that in many advanced economies, sub-par growth is now a pretty old story, at least one which pre-dates the financial crisis. As is suggested in the first quote above, growth was decent in advanced economies prior to the crisis, but that owed to wide policy stimuli, rapid credit expansion and asset price bubbles -notably housing-, all of which had a sizable positive, albeit temporary, contribution to growth. In other words, absent such factors, it is an open question as to whether the pre-crisis growth performance would have been any better than the post-crisis one. These two observations -modest growth prior to the crisis despite significant tailwinds and virtually no growth since the financial crisis- have been at the core of the Secular Stagnation hypothesis. Proponents of this view suggests that low growth is actually here to stay as Alvin Hansen thought would be the case for the US in There have been two dominant views up to now to think about secular stagnation The first states that current developments reflect slow-down in technological progress (Gordon 2012). According to this first view, recent innovations have had less of a positive effect on productivity and hence advanced economies are bound to grow more slowly in the future. The second view is based on a demand-side approach. 2

3 It posits that under the influence of some structural factors (eg. globalisation, drop in capital goods price, rise in inequality), the natural rate of interest has become negative (Eggertson et al. (2014)). Yet, because of institutional constraints, policy makers are not able to take the real interest rate down to such negative levels. Policy hence ends up being permanently too tight which constrains growth at sub-par levels. In this paper, we provide a different albeit complementary approach to think about low growth and stagnation in advanced economies. For that, the paper builds on a set of stylized facts which suggest that the demand-side approach to secular stagnation described above could be usefully complemented with a supply-side view focusing on the real effect of credit booms and busts. We first document that credit booms can have significant negative effects on productivity growth as they occur. 1 Specifically, we provide empirical evidence that during periods of rapid credit expansion, employment grows more quickly in low productivity gain sectors and such labour reallocation pattern acts as a drag on aggregate productivity growth. 2 To uncover why credit booms hurt productivity growth through labour reallocation, we build a simple model with two sectors, the higher productivity sector having a lower borrowing capacity. When the economy faces a positive shock on future credit supply, then entrepreneurs invest more in the sector with a higher borrowing capacity but with a lower return. This way, they can take advantage of large future credit expansion. Higher credit growth is therefore systematically associated with reallocation in favor of the low productivity sector. 3 Next, we extend the analytical framework to investigate how current labour reallocations affect future productivity growth. We highlight two different channels. On the one hand, when the high productivity sector exhibits positive learning externalities, an increase in the current size of this sector raises the future return to this sector which raises future aggregate growth, and the more so when credit is more available. 1 From that point view, the observation that growth was only modest in the pre-crisis period is perfectly consistent: even if the credit boom contributes to increase aggregate demand in the short-run, we argue that it does also weaken growth over the medium term through the supply side. The exceptionally large US trade deficitatthattimeinthepre-crisisperiodsomewhat confirmsthisview. 2 To take a concret example, at the beginning of a credit boom, a large numbe of people work in manufacturing, a sector with large productivity gains. But by the end of the credit boom, many have shifted to construction and real estate services, both sectors featuring much low productivity gains than manufacturing. 3 Interestingly, the model delivers the same prediction whether we assume a negative correlation between productivity and borrowing capacity or a negative correlation between productivity and optimal investment size. In the latter case, a credit boom also shifts entrepreneurs to the low productivity sector as it allows to close a larger part of the distance to the optimal production plan for the less productive sectors. On the contrary for the higher productivity sectors which are very close to their optimal size, the benefits to derive from the credit boom, in terms of attaining the optimal size are much smaller. 3

4 This is because higher leverage magnifies the positive effect on growth of a higher return. On the other hand, an increase in the current size of the high productivity sector raises the future size of the high productivity sector raises but less so when credit is more available. The intuition is pretty straightforward: When credit is more easily available, changes in the current allocation generate larger changes in current profits. As a result, the current allocation needs to respond less to changes in the past allocation if credit is more plentiful. 4 Which of these two effect eventually dominates is a matter of parameters. Specifically, the second effect dominates when the high productivity sector is relatively credit insensitive, i.e. its borrowing capacity does not respond to credit supply shocks. In this case, the tighter the credit conditions, the stronger is the dependence of subsequent growth on past labour allocation. We finally test for this prediction focusing on a cross-section of recessions where we identify the occurence of a financial crisis to tight credit conditions. Consistent with the model s prediction, we find that labour reallocation prior to a recession has significant, long-lasting, effects on productivity after the recession, but only when the economy is hit with a financial crisis. This evidence hence suggests that weak productivity after the recession is the joint outcome of the financial crisis and the previous credit boom which led employment to be severely skewed towards low productivity gain sectors. This paper relates to three different strands of literature. It first relates to the literature which quantifies job reallocation and look at its causes and consequences. On quantification, Davis and Haltiwanger (1992) estimate in their seminal paper that job creation and destruction accounts roughly for 20% of jobs while Campbell and Kuttner (1996) find that reallocation shocks account for roughly half of the variance in total employment growth. A large literature also explores the relationship between reallocation and the business cycle given that causality can go in both ways. More recently, a burgeoning literature has started looking at the effect of credit booms on reallocations: Acharya et al. (2010) looks at the effect of cross-state banking deregulation in the US on the allocation of output and employment across sectors at the state level. Gorton and Ordonnez (2014) show that credit booms can have negative implications for aggregate TFP, which can be accounted for in a model where agents do not produce information about the quality of collateral during 4 Credit therefore acts in the model to increase the economy s convergence speed -to the steady state allocation- while with scarce credit, the starting point matters more. 4

5 the boom. A second strand of literature deals with the macroeconomic implications of microeconomic distortions. The seminal paper by Hsieh and Klenow (2009) measures within-industry dispersion in productivity with the idea that such dispersion indicates resources misallocation, which has negative effects at the aggregate level. The authors can therefore infers the increase in aggregate productivity that would result from eliminating such dispersion. Our approach is here is more based on econometric inference testing for the possibility that past labour reallocation across sectors may actually be driving future aggregate labour productivity. Last, the paper relates to the literature on financial factors and the dynamics of the real economy. Yet contrary to most of this literature which looks at how financial frictions can directly affect output and productivity, our findings highlight that developments in the financial sector can have both direct and indirect real effects. For example when credit outgrows GDP, this comes with stronger labour reallocation towards low productivity gains sectors which represents a direct drag on aggregate labour productivity. Moreover, there is also an indirect effect as credit booms can have weaken the path for subsequent productivity through the type of labour reallocation they can generate. The paper is organized as follows. Section 2 lays down the methodology and details the data used for the empirical analysis. Then section looks at the relationship between credit booms and labour reallocation across sectors. Building on these results, section 4 develops a simple analytical model of how credit booms affect factor allocation. It also looks at how credit conditions affect the dependence of growth on past allocation. Section 5 then tests the prediction of the model looking at the implications of labour reallocation across sectors for the dynamics of productivity depending on the occurence of a financial crisis. Section 6 concludes. Finally, proofs and estimation details are located in the annex. 5

6 2 Labor reallocations, productivity growth and credit booms 2.1 Methodology In this section we lay out the methodology used to isolate the effect of labor reallocations on productivity growth. For this sake, we write aggregate output (aggregate employment ) as the sum of individual sectors nominal output (individual sectors employment ): = X and = X (1) Assuming the economy is made of different sectors and denoting the unweighted average for variable across all sectors ( = ; = ), aggregate productivity can be written as the sum of two terms: = 1 X µ µ µ = + ; (2) The first term represents unweighted average productivity computed across all sectors in the economy while the second term measures whether sectors with high productivity also account for a large share in total employment. When this is the case, i.e. when sectors with higher productivity account for a large share of total employment, the covariance is positive and aggregate productivity is higher than the unconditional average sector-level productivity. Building on the decomposition in expression (2) and denoting sector relative output size as =, it is possible to write the growth rate of aggregate labour productivity as follows: 1+ ( ) = 1 X µ 1+ ( µ ) 1+ ( ) ( ) Then using the property =1, the growth rate of aggregate real labour productivity writes as the sum of 6

7 two different terms: 1+ ( ) " #" # = 1+ ( ) 1+ ( ) {z } common component µ µ ( ) + ; 1+ ( ) {z } allocation component (3) We will call the first term of the right hand side in expression (3) the common component of real labour productivity growth (henceforth, ( )) and the second term of the right hand side, the allocation component of real labour productivity growth (henceforth, ( )). We now turn to quantifying each of these two terms. In a second step, we will be investigating what drives their fluctuations and what their implications can be, in particular for the dynamics of productivity. 2.2 The data We rely on three different sources of industry-level data: the OECD-STAN database, the EU-KLEMS database and the GGDC 10-sector database. These three datasets provide information on value added and employment at the sector level following the ISIC 3 rev.1 classification at the 1-digit level. Overall, we consider 9 different sectors: Agriculture (A and B), Mining (C), Manufacturing (D), Utilities (E), Construction (F), Trade services (G and H), Transport services (I), Finance, Insurance and Real Estate services (J and K) and Government and Personal services (L to Q). To build our dataset, we require for each data point that industry-level output and employment sum up to the economy-wide aggregates. This limits the number of countries and years that can be included in the analysis. 5 Weendupwithanunbalancedsamplecovering 21 countries starting in 1979 and ending in Following the previous notation, using decomposition (3), aggregate real labour productivity growth in 5 In this paper we focus on net changes in sector-level employment, without separating employment destruction from employment creation. Another difference with the literature is that we focus on employment or persons employed as opposed to jobs. As a result, we are probably under-estimating the extent of labour reallocation in the economy. For example, Davis and Haltiwanger (1992) estimate that each year around 20 per cent of jobs are either created or destroyed in US manufacturing. By contrast, our net employment change barely represents a few percentage points of total employment in our sample. 6 The countries included in the sample are Australia, Austria, Belgium, Canada, Switzerland, Germany, Denmark, Spain, Finland, France, Greece, Ireland, Italy, Japan, Korea, Netherlands, Norway, Portgual, Sweden, the United Kingdom and the United States. The start and end dates (1979 and 2009) were chosen mainly because of constraints on industry data consistent availability. 7

8 country between year and year + can be written as + + =( ) + +( ) + (4) On the right- hand side, ( ) represents the common component of productivity growth and ( ) represents the allocation component as defined in decomposition (3). To compute the various growth measures we consider non-overlapping periods of either three or five years. This is because reallocations must surely take considerable time, especially across industries as widely defined as those considered here. 7 Shorter periods, of, say, one or two years, could mask the "true" extent of the reallocations. Using 5-year windows yields 120 observations and 3-year windows 182 observations. 2.3 A first glance Table 1 provides summary statistics -polling all the data together- for aggregate real labour productivity growth, i.e. the left-hand side in expression (3), and for its common and allocation components, i.e. respectively the first and second terms on the right-hand side in expression (3). The first three columns of Table 1 provide summary statistics using five-year windows and the last three using three-year windows. Insert Table 1 here Over a 5-year interval, real labour productivity grows on average 8.6 percent, i.e. 1.6 percent per year. On average, the common component represents around 5.4 percentage points (or just under two thirds) and the allocation component the remaining 3.2 percentage points. The figures based on 3-year windows are similar: aggregate real labour productivity grows on average by 1.7 percent per year, with the common component representing two-thirds of the total. 7 Blanchard and Katz (1992) consider the effect of state-specific shocks to labour demand across US states. According to their estimates, it can take up to 7 years for their effects on state unemployment and participation to disappear. More recently, based on longitudinal data, Walker (2013) estimates the transitional costs associated with reallocating workers from newly regulated industries to other sectors in the wake of new environmental regulations. His results suggest that these costs are significant: the average worker in a regulated sector experienced a total earnings loss equivalent to 20% of their pre-regulatory earnings, with almost all of the estimated earnings losses driven by workers who separate from their firm. 8

9 The volatility (standard deviation) of the allocation component accounts for 45 to 55% of the volatility of aggregate productivity growth, depending on the window length. The common component is roughly as volatile as aggregate productivity growth, implying a negative covariance with the allocation component. This means that changes in the common component are systematically associated with opposite, but smaller, changes in the allocation component. For example, an economy-wide shock that raises productivity growth uniformly across all sectors tends to be partly offset by labour reallocations towards those with lower productivity growth. Table 2 provides the correlation matrix for aggregate productivity growth and the two components, focusing on within-country correlations. Correlations in the upper left matrix are computed using 5-year windows; those in lower right matrix using 3-year ones. Insert Table 2 here The matrix shows that labour reallocations towards high productivity sectors tend to boost aggregate productivity growth. Aggregate productivity and its allocation component co-move positively within countries and the relationship is statistically significant. 3 Credit booms, labour reallocations and productivity growth 3.1 Relative impact on the allocation and common components of productivity What is the impact of credit booms on the two components of productivity growth as the credit booms occur? Simply put, we find that credit booms depress productivity growth and that their impact works through the allocation component the only one for which a statistically significant link is apparent. This result survives increasingly demanding tests. The basic result emerges already quite clearly in simple bivariate tests. Graph 1 plots the allocation components (left panel) and the common component (right panel), respectively, against the growth in credit to the private sector to GDP (shown on the x-axes) our benchmark measure of credit expansion, the latter 9

10 being drawn from the BIS database on credit to the non-financial private sector. We use 5-year windows and focus on deviations from country and time averages. The graph traces a negative and statistically significant relationship between credit growth and the allocation component. By contrast, no such relationship emerges with the common component. Insert Graph 1 To test whether these bilateral correlations survive a more rigorous econometric analysis, we estimate the following three regressions: + + = ( ) + = (5) ( ) + = Here, + + stands for the growth rate of labour productivity in country between year and +, and ( ) + and ( ) + for the allocation and the common components, respectively. The independent variables include a set of country and time dummies ( ; ) as well as a vector of (pre-determined) control variables. 8 The growth rate of employment in country between year and + is denoted + and + isthevariablemeasuringtheintensityofthecreditboomincountry between year and +. Finally, 0 are residuals. 9 We estimate regressions (5) using the two different window-lengths, 3 and 5 years, and two different measures of credit booms, the rate of growth in the private credit-to-gdp ratio (our benchmark measure) and the deviation of the same ratio from its long-term trend (the credit gap ). 10 The vector of controls includes the following variables: (i) the ratio of credit to GDP, (ii) government size, measured as the ratio of government consumption to GDP, (iii) CPI inflation, (iv) openness to trade, 8 Note that including country fixed effect ensures we focus on within-country credit booms while including time fixed effects ensures we focus on country-specific creditboomsandfilter for global ones. 9 With decreasing marginal returns to labour, productivity and employment growth tend to be negatively correlated. Moreover, aggregate employment growth controls for the cyclical position of the economy and hence ensures that we tease out credit booms from economic expansions as the former is likely to bring the latter. 10 Data on the credit gap is obtained from Borio et al (2011). Moreover, for the sake of brevity, we only report estimations using five-year windows. Estimations using three-year windows are available in a separate annex. 10

11 measured as the ratio of imports plus exports to GDP; (v) a dummy for the occurrence of a financial crisis; and (vi) the log of the initial level of output per worker. These data are all drawn from the OECD economic outlook database, except the data on financial crises drawn from Laeven and Valencia (2012). The choice of control variables deserves some explanation. Credit in relation to GDP can help avoid confusing the effect of credit levels and growth. If, say, the growth rate in the credit-to-gdp ratio is lower when the credit-to-gdp ratio is higher, then a negative correlation between our measure of credit booms and productivity growth could simply reflect the positive effect of a higher credit-to-gdp ratio. We include government expenditures because credit booms boost tax revenues, allowing the government to increase its spending and employment. Since, by construction, the government sector exhibits low productivity growth, the negative correlation identified above might just be capturing changes in its size. The addition of inflation reflects the well-known view that inflation can lead to misallocations by introducing noise in the signals agents receive about relative prices (Lucas (1975)). If credit booms coincide with higher inflation then we may just be picking up this effect. Trade openness should be expected to boost productivity gains across sectors, including through reallocations towards sectors enjoying some comparative advantage. And, as already discussed, financial crises, which tend to be preceded by credit booms, are also known to generate output losses: the previous results may simply reflect their costs. Insert Table 3 here The regression results using the growth rate in private credit to GDP as a measure of credit booms fully confirm the preliminary bivariate tests (Table 3). Based on 5-year windows, private credit to GDP growth is negatively correlated with aggregate productivity growth and the result appears to be entirely driven by a strong and highly statistically significant relationship with the allocation component (column (iii.a)): there is no significant relationship between credit growth and the common component Results using a 3-year window are very similar, except that now there is some evidence of a statistically significant link also with the common component, albeit only at the 10%-level. Possibly, over the shorter window, credit booms boost demand across all sectors, leading to a generalised increase in employment which leads to a productivity slow-down. But as the credit boom proceeds, the incidence across sectors becomes more differentiated so that the average effect fades out while labour reallocations keep taking place. 11

12 Turning to the control variables, some interesting patterns emerge. There is little evidence of a financial deepening effect: the level of private credit to GDP does not seem to account for either aggregate productivity growth or its two components. On average, employment growth tends to coincide with lower aggregate productivity growth even as it goes hand-in-hand with productivity-enhancing reallocations: the common component dominates. And we can discard the view that labour reallocations are driven by changes in government expenditures. Government consumption does appear to dampen productivity growth, although the relationship is only weakly statistically significant, but no link is apparent with the allocation component. 12 The role of CPI inflation is consistent with priors: inflation correlates negatively and significantly with the allocation component of productivity growth, although there is no statistical significant relationship with productivity growth as a whole. Also as expected, openness to trade does co-vary positively with the common component of productivity growth, even if, as in the case of inflation, there is no statistically significant link with labour productivity growth as a whole. Finally, financial crises do not appear to affect productivity growth or any of its components in a statistically significant way. 13 The regressions also shed light on the so-called catch-up effect, i.e. the tendency for productivity growth to converge across countries. They indicate that the effect reflects almost exclusively the operation of the common component, since the correlation between the initial productivity level and the allocation component is not statistically significant. This implies that the allocation component is relatively more important in economies with higher productivity, because overall productivity growth will generally be lower there. If so, credit booms are likely to be more costly in advanced economies. 14 The conclusions are very similar if we use the credit gap as a proxy for credit booms, with some qualifications. In particular, in this case, the correlation with the allocation component is still apparent, but is statistically weaker. This difference probably reflects measurement errors: given the slow-moving trend, for 12 This hypothesis can be formally tested by computing productivity growth and its components, excluding the government sector. 13 This last result may sound surprising but it is important to remember that the financial crisis variable is a dummy which equals one if a financial crisis hits within the corresponding five-year bracket. Hence the economy may be experiencing a financial crisis one year but doing relatively well for the rest of the period. More importantly, an economy facing a financial crisis may be able over a 5-year period to adjust, in particular by reducing employment levels in accordance with the drop in output so that productivity is left relatively unchanged. 14 This may be a reason why such results may actually not extend to emerging market economies where productivity growth is more linked to common developments across sectors. 12

13 current purposes the credit gap is a noisier measure of credit booms, particularly over short windows. 3.2 Decomposing the allocation component further The previous results highlight how labour allocations during credit booms dampen productivity growth, but they are silent about the nature of the reallocations. Specifically, when the allocation component declines over time, is this because, for a given distribution of sectoral productivity growth, employment grows more rapidly in low productivity growth sectors ( employment-driven )? Or is this because, for a given distribution of sectoral employment growth, productivity slows-down in sectors with rapidly expanding employment ( productivity-driven )? Put differently, do changes over time in the allocation component reflect changes in the distribution of labour across sectors or changes in the distribution of productivity gains across sectors? To isolate these channels we carry out a variance decomposition exercise on the allocation component itself. Specifically, let be the average over time of any variable and = the deviation from the average. We can then write the allocation component as the sum of four terms: ³ ³ ( ) ; 1+ ( ) + µ = ³ ( ) ; µ + 1+ ( ) µ ³ ( ) ; 1+ ( ) + µ ( ) ; ³ ( ) ; 1+ ( ) ³ 1+ ( ) The first term on the right-hand side measures the covariance across sectors between average growth in sectoral employment shares and average growth in sectoral size-weighted productivity growth. Since this (6) varies only across countries, it will be captured by the country fixed effects in the regressions (6). The second term measures the covariance across sectors between average growth in sectoral employment shares and deviations of sectoral size-weighted productivity growth from averages. This component reflects the impact of changes in sector-level size-weighted productivity growth rates, holding changes in employment shares constant. We will call this second term, the productivity-driven allocation component. The third term measures the covariance across sectors between deviations of growth in sectoral employment shares from averages and sectoral size-weighted productivity growth. This term captures the impact of changes in employment shares, holding size-weighted sectoral productivity growth constant. We will call this third term, 13

14 the employment-driven allocation component. Finally, the fourth term measures the covariance between deviations of sectoral growth in employment shares from their long-run averages and deviations of sectoral size-weighted productivity growth rates from their own long- run average. This term measures how the allocation component of productivity growth depends on both types of changes. We therefore call it the jointly driven allocation component. We can now run the same regressions (5) using each component of the variance decomposition as a dependent variable. Insert Table 4 here The decomposition indicates that the decline in the allocation component during credit booms overwhelmingly reflects shifts in employment towards low-productivity gain sectors (Table 4). Specifically, the negative correlation between credit to GDP growth and the allocation component is explained almost exclusively by changes in industry-level employment growth rather than changes in size-weighted productivity growth across sectors: only the employment effect is statistically significant. In other words, credit booms do not appear to affect the sectoral distribution of productivity gains, turning potentially high productivity gain sectors into low productivity gain ones. Rather, they induce labour shifts into lower productivity gain sectors. Productivity in industries with rapid long run productivity growth do not grow any slower during credit booms, but these industries attract relatively fewer workers. Table 4 shows that more than 90% of the effectofcreditboomsreflects these shifts in employment shares. Switching to the credit to GDP gap provides very similar results. The negative correlation between credit booms and the allocation component remains largely unchanged, and still pertains to changes in the sectoral distribution of employment creation. The relative size of the effect is also very similar. 3.3 Investigating causality As a last robustness check, we investigate whether the evidence produced so far is simply a correlation or represents causality. But before we turn to the estimations designed to address this issue, it is worth noting 14

15 that, in fact, the evidence so far does point to causality running from credit booms to productivity growth rather than the other way round. Intuitively, reverse causality does not look plausible. It would imply that productivity slowdowns induce either financial intermediaries to supply more credit or firms and households to demand more of it. True, in the very short term one could imagine, say, households borrowing more to protect their consumption in the face of an unexpected slowdown in productivity and hence income. But such an effect should wash out over the relatively long windows we are considering. Moreover, credit tends to be pro-cyclical. And since we control for cyclical conditions through employment growth, the response of credit to the real economy is already largely filtered out. Finally, it is hard to envisage that credit would systematically react to productivity slowdowns driven by labour reallocations but not to those driven by the common component. A first statistical safeguard against reverse causality is that in regressions (5) all right-hand-side variables are pre-determined with respect to the dependent variable, i.e. are measured at the beginning of the period. The exceptions are employment and credit growth, which are both measured over the same period. 15 Still, in order to lay to rest any residual doubts about the direction of causality even for these two variables, we instrument them. We do so with beginning-of-period values for the nominal long-term interest rate, the trade balance-to-gdp ratio, the current account balance-to-gdp ratio as well as the level and change in the financial liberalisation index constructed by Abiad et al (2008). Insert Table 5 here Table 5 provides estimation results using IV. Credit to GDP growth is the proxy for credit booms in the first four columns, and average credit to GDP deviation from trend in the last four. As in previous tables, the common and the allocation component sum up to productivity growth and the dependent variable in estimations in (iv.a) and (iv.b) is the part of the allocation component due to shocks to the sectoral distribution of employment, following the variance decomposition presented in section It is also the case that the financial crisis dummy is measured over the same period as productivity growth. However, this variable proves in practise to have very little influence on the empirical results. We therefore take it out from the IV estimations to ensure all right hand side variables, except those we instrument, are pre-determined. 15

16 Estimation results confirm our previous findings. Indeed, the results become even sharper. The estimated coefficient becomes larger in absolute value, suggesting that the OLS estimates may underestimate the effect of credit booms on productivity growth. For example, according to the OLS estimates, a 10 percentage point increase in private credit to GDP growth over 5 years reduces productivity growth by 0.8 per cent over the same period, arguably a relatively small number. But according to the instrumental variable (IV) estimates, the slow-down in productivity is closer to 1.4 percentage point over five years, which amounts to dampening productivity growth by percentage points per year. In addition, consistent with OLS results, IV estimates confirms that roughly 60% of the effect of credit booms on productivity reflects labour reallocations across sectors. In other words, labour reallocation is quantitatively the main channel through which credit booms affect productivity. Moreover, these results hold pretty much unaltered if credit booms are measured with the credit gap. We now turn to the model to provide a possible mechanism of how credit booms affect ressource allocation and growth. 4 The model 4.1 Main assumptions Consider a small open economy with overlapping generations of entrepreneurs who live for two periods. Entrepreneurs born at date receive a bequest from the generation born at date 1. Generation- entrepreneurs combine this bequest with some borrowing obtained from financiers and invest in a project. For simplicity and without loss of generality, we normalize the cost of capital to one. At date +1,the project produces output that is then used for three purposes: (i) paying back the loan (ii) bequesting +1 to generation- +1 entrepreneurs (iii) and saving +1. Entrepreneurs born at date then combine savings +1 with some new borrowing +1 to invest in the same project at date +1. Finally at date +2, entrepreneurs reap the project s output and use it to: (i) pay back liabilities +1, and (ii) consume

17 Timing of the model Entrepreneurs can either invest in a type- or a type- project. Importantly, they are committed to a single project type for their entire productive life. Project types { ; } are defined by the total return 1+ and the pledgeable return. Realistically, output from higher-return projects is more difficult to pledge, but. In addition, newly born entrepreneurs investing in a type- project need to pay an installation cost per unit of investment, which increases with the number of such entrepreneurs. Denoting the measure of newly born entrepreneurs investing in type- projects, the return to investing in a type- project is 1+ ( ) for newly born entrepreneurs (with 0 ( ) 0) and1+ for old entrepreneurs. Similarly, the return to investing in a type- project is 1+ ( ) for newly born entrepreneurs (with 0 ( ) 0)and1+ for old entrepreneurs. Finally all projects have positive NPV but are credit constrained, ( ) 0 and 1, { ; }. Denoting a positive scalar, we write the utility function of an entrepreneur born at date as =log+1 + log +2 (7) 4.2 The dynamics of the economy We denote ( ) the profit per unit of internal funds for newly-born entrepreneurs starting a type- project at date, being the number of such entrepreneurs investing in type- projects at date. Similarly, +1 17

18 denotes the profit per unit of internal funds for entrepreneurs born at date startingatype- projectatdate +1. In what follows, we will call ( ) the short-run profit rateand +1 the long-run profit rate.we are now able to write the utility maximization problem for an entrepreneur born at date as max =log+1 + log ; = ( ) s.t. +2 = (8) Optimal date +1bequest +1 and optimal date +2consumption +2 then satisfy: +1 = 1 1+ ( ) +2 = (9) ( ) As is clear, the first expression in (9) governs the growth rate of net wealth for entrepreneurs choosing type- projects. Next, we turn to entrepreneurs project choice to determine the profit rates ( ) and Financial constraints Entrepreneurs can borrow, but they can default strategically. To preclude default, financiers impose borrowing limits that ensure entrepreneurs are better-off paying back. To determine this no-default level of borrowing, consider an entrepreneur starting with a unit of net wealth, borrowing and investing 1+ in atype- project. If the entrepreneur chooses to repay financiers, the profit isthen(1 + )(1+ ). Alternatively, if the entrepreneur chooses to default, the pledgeable output (1 + ) is lost and the financiers can still recover some claims in the case of default. If the default rate is (1 ), with 1, the profit under default is (1 + )(1+ ) (1 1 ). Comparing the profit with and without default, the no-default constraint satisfies: 1 (10) 18

19 Given that projects are positive NPV, entrepreneurs always borrow as much as possible and (10) therefore always binds. Date- and date- +1 optimal borrowing for an entrepreneur born at date choosing to invest in a type- project satisfy = 1 and +1 = (11) Using (11), the short-run profit rate ( ) and the long-run profit rate +1 write as ( )=1+ ( ) and +1 =1+ (12) Denoting = log ( ) (1 + )log(1+ ), the indirect utility for an entrepreneur with unitary initial endowment, investing in a type- projectatdate is: = +(1+ )log ( )+ log +1 (13) Entrepreneurs welfare therefore depends on current and future credit conditions and +1. For example, apositiveshockto+1 raises future credit supply and hence mimics an increase in the growth rate of credit. Conversely, a negative shock to acts a cut to the current credit supply and can therefore mimic the effects of a credit crunch. The next section is devoted to investigate how these parameters affect entrepreneurs allocation between type- and type- projects and thereyby the growth rate of the economy. 4.4 Credit booms and reallocation across sectors. The equilibrium number of newly born entrepreneurs investing at date in type- projects is such that entrepreneurs should be indifferent between investing in either sector. Using (13), the break-even condition writes as (1 + )log ( )+ log +1 =(1+ )log ( )+ log +1 (14) 19

20 Writing (+1 )=[ ] 1+ = (14) simplifies as h 1+ i ih , the break-even condition ( )= (+1 ) ( ) (15) In the break-even condition (15), the LHS increases with the number of entrepreneurs investing in type- projects at date, while the RHS decreases with the number of entrepreneurs investing in type- projects at date. As a result, there is a unique equilibrium which pins down each sector size. In the next proposition, we look at the comparative statics of the equilibrium allocation of entrepreneurs. Proposition 1 Assuming and ,thenattheequilibrium, (i) the equilibrium number of newly born entrepreneurs investing in type- projects at date decreases when credit grows faster between date and date +1: +1 0 (16) (ii) the equilibrium number of newly born entrepreneurs investing in type- projects at date is such that entrepreneurs investing in type- projects enjoy larger short-term profit: ( ) ( ) (17) Proof. c.f. appendix. The assumption 1 1 still have a higher "leveraged" return, i.e. means that even if type- projects have a lower return, i.e.,they 1 1. This means that the long-run profit rateisalways larger for type- projects, As a result, for the break-even condition to be satisfied, the short-term profit ratemustbelowerfortype- projects, ( ) ( ). Hence at the equilibrium, type- projects, yield higher profits in the short-run while type- projects yield higher profits in the long-run. Moreover the long-run profit rate for type- projects is more sensitive to credit conditions, i.e Thus, when credit expands more strongly in the future, i.e. +1 goes up, 20

21 the long-run profit ontype- projects increases relative to type- projects, which leads more entrepreneurs to invest in type- projects at date. The equilibrium number of entrepreneurs investing in type- projects at date is therefore a decreasing function of the growth rate of credit between date and date +1,as measured by the parameter +1. Raising the growth of credit reduces the size of the high productivity sector. 4.5 Growth Growth is summarized in this model with the dynamics of entrepreneurs initial endowment. Given optimal individual choices (9), the growth rate of entrepreneurs initial endowment writes as +1 = (1 ) ( )+ ( ) 1+ (18) As expected, growth depends on the allocation of entrepreneurs across sectors. Interestingly, it also depends on short term profits, but not on long-profits which affect only consumption. Hence credit growth affect the growth rate of the economy essentially through two channels: First is the reallocation channel, as changes inthegrowthrateofcreditdirectlyaffect the distribution of entrepreneurs across sectors. Second is the implied change in short-term profits which change with the size of the corresponding sector. Interestingly, 21

22 the growth rate of credit has no direct effect on the growth rate of the economy. This is because short term profits do not depend on the growth rate but on the level of credit. Proposition 2 Assuming ( ) is separable, i.e. ( ) = ( ) and ( ) = (1 ), and assuming is decreasing concave, then the growth rate of the economy decreases with the growth rate of credit. This negative relationship is due to both reallocation towards the low growth sector as well as lower average growth across sectors. Proof. c.f. appendix An increase in the growth rate of credit affects aggregate growth through the reallocation channel: When +1 increases and credit growth more quickly, this leads fewer entrepreneurs to invest in type- projects, while type- projects happen to provide higher profits in the short-run. Moreover, the overall response of the growth rate of the economy depends not only on entrepreneurs reallocation across projects but also on changes in the short run profit toeachtypeofproject. When credit booms and fewer entrepreneurs chooses type- projects, short-run profits on type- projects increase as this sector shrinks in size while short-run profits on type- projects decrease as this sector expands in size. The net effect is hence ambiguous. Yet, the assumptions that ( ) and ( ) are separable and concave imply that there are more entrepreneurs choosing type- projects than entrepreneurs choosing type- projects. Hence the drop in short-run profits on type- projects tends to dominate the increase in short-run profits on type- projects. When credit booms average short-run profits hence decrease and so does the growth rate of the aggregate economy. 4.6 Credit conditions and the dynamics of the economy. We now turn to the dynamics of the economy. For this end, we extend the previous framework assuming now that the number of newly born entrepreneurs 1 whoinvestedintype- projects at date 1 affects positively the return on type- projects to newly born entrepreneurs at time, i.e. wenowassume 0 ( 1) 0. This new assumption translates the idea that the high return project benefits a positive learning-by-doing externality. By contrast no such externality exists for the low productivity project. Moreover for simplicity 22

23 and without any implication for next results, we will also assume that the current number of newly born entrepreneurs investing in type- projectsdoesnotaffect the return to type- projects, i.e. 0 ()=0.We otherwise keep the framework unchanged. Then following previous notation, short-term profits write as ( )=1+ ( ) 1 and ( 1 )=1+ ( 1 ) 1 Denoting =[ ] 1+, the break-even condition which determines the number of newly-born entrepreneurs investing in type- projects at date satisfies the dynamic relationship: ( )= ( 1 ) (+1 ) (19) Given that 0 and 1 0 (19) defines a positive dynamic relationship for the size of the -sector: A larger -sector at date 1 raises the return on type- projects at date, 1 0. Hence, the return on type- projects at date also needs to go up to ensure entrepreneurs are indifferent between the two types of projects. This requires that fewer entrepreneurs choose type- projects and more of them choose type- projects at date. Using (9) and (19) the growth rate of the economy writes as +1 = +(1 ) ( 1 ) (20) 1+ Using the expression (20) for the growth rate of the economy and the expression (19) for the dynamics of entrepreneurs allocation across sectors, we can derive the following result. Proposition 3 When type- projects are relatively credit-insensitive, i.e. is sufficiently low, then the growth rate of the economy increases with the number of newly-born entrepreneurs who previously invested in type- projects and the less so when credit is more easily available: and

24 Proof. c.f. appendix The first property that the growth rate of the economy +1 increases with the number of newly-born entrepreneurs 1 whopreviouslyinvestedintype- projects, is relatively straightforward. This is because of the direct positive effect on short-term profits for type- projects and because of the indirect positive effect on the number of newly-born entrepreneurs who invest in type- projects at the next date. The second property that the growth rate sensitivity to past allocation decreases with the current credit supply deserves a more detailed explanation. When current credit conditions improve, i.e. goes up, two opposite effects materialize. On the one hand, entrepreneurs enjoy larger short-term profits. Hence, the positive effect on short-term profits of more newly-born entrepreneurs investing in type- projects at date 1 is magnified: 1 0 On the other hand however, when type- projects are relatively insensitive to credit conditions, i.e. is close to zero, a positive shock to current credit conditions reduces the time persistence in the number of entrepreneurs investing in type- projects: The intuition for this last result is relatively straightforward: When the number of newly born entrepreneurs investing at date 1 in type- projects goes up, this leads to an increase in short-term profits both for type- and type- projects. Yet, when is larger, type- projects can borrow disproportionately more because of a higher pledgeable return. Hence the number of newly born entrepreneurs investing at date in type- projects goes up, but less so, to ensure that the short-term profit ontype- projects does not increase beyond proportion. The sensitivity of the current entrepreneurs allocation to the past entrepreneurs allocation is therefore a decreasing function of current credit extension. 24