Real Estate Collateral and Labor Demand Thomas Chaney David Sraer David Thesmar August 5, 2013 Abstract This paper shows that the availability of real estate collateral a ects labor demand. We regress firm-level labor demand on the value of their real estate holdings. We use a large administrative dataset of French firms, which has the advantage of including small, unlisted firms, and of providing reliable information on employment and real estate holdings. We find that collateral shocks have a strong and statistically significant impact on the labor demand of firms. Aggregate e ects are sizable. During the 2002-2006 real estate price run-up, we find that some 10% of aggregate job growth in France was due to increased availability of real estate collateral to firms. We find, however, that the response of employment to collateral shocks is smaller than what would be expected given the e ect of collateral on investment and stable labor-capital complementarity. This suggests either (1) large adjustment costs on labor or (2) the adoption by firms of labor-saving technologies during the period. TSE and NBER (e-mail: thomas.chaney@gmail.com) Princeton University and NBER and CEPR (e-mail: dsraer@princeton.edu) HEC and CEPR (e-mail: thesmar@hec.fr) 1
1. Introduction This paper estimates the e ect of real estate collateral value on labor demand fluctuations. When firms are financially constrained, a positive shock to the value of their collateral makes it easier to borrow, and therefore to invest (Bernanke and Gertler (1986), Kiyotaki and Moore (1997)). Given that labor and capital are complement in the production process, collateral fluctuations should therefore a ect employment dynamics. Besides, even in the absence of capital, the financing of labor may itself be subject to credit constraints, since firms need to pay their workers before they can sell their production (Benmelech, Bergman, and Seru, 2010). For these two reasons, we expect labor demand to respond to changes in the value of real estate collateral. This paper empirically evaluates this mechanism. Our focus on real estate collateral helps us to focus on exogenous shocks to collateral value. But real estate is interesting in itself: Because of its redeployability, it is a large supplier of collateral in the economy. Booms and busts of the housing market thus have the power to be important drivers of employment fluctuations. The empirical challenge is to attribute the response of employment fluctuations to movement in real estate prices, to the collateral channel. For instance, house prices are comoving with global economic conditions, and may therefore be correlated with business opportunities. In this paper, we use firm-level data and two sources of identification. First, we focus on reactions of employment to local house prices. This allows us to abstract from aggregate fluctuations that may be correlated with, say, aggregate demand. Second, we compare firms that own their real estate to firms that rent it. This technique allows us to control for shocks to local economic conditions that may be correlated with house prices: Owners benefit from the collateral channel, while renters do not. To implement these tests, we use a large administrative dataset of French firms over 1998-2007. A key advantage of our data is that they span the universe of French firms, both public and private, large and small. They also provide us with reliable information on employment, firm location, and real estate holdings. Our methodology which borrows from Chaney, Sraer, and Thesmar (2012) allows to 2
recover the firm-level response of labor demand to 1 Euro of collateral gain. The richness of the data allows to control for local industry-specific demand shocks and permit a host of robustness checks. Our findings are twofold. First, we find sizable e ects of collateral on employment. 1m Euro of collateral value leads to the creation of 1.5 additional job. During the 2002-2006 run-up in real estate prices, the collateral channel alone may have been responsible for the creation of 40,000 new jobs, approximately 10% of the aggregate employment growth during this period. Secondly, we find that labor demand reacts little, compared to what we would expect given the response of investment to collateral shocks. Indeed, according to our estimates, a 1m Euro increase in collateral value leads to about 190,000 Euro of additional investment. Given an average capital to labor ratio of 45,000 Euros per worker, we would expect firms to hire on average more than 4 new workers, rather than just 1.5. There can be two credible and non-exclusive explanations for this. Either the production function is changing: Firms use the additional collateral to invest in labor-saving technologies. Or there are large adjustment costs to labor, which limit the propensity of firms to adapt their workforce to the new capital. Our paper contributes to two strands of literature. The first one is the nascent literature about the e ects of credit constraints on labor demand. Using individual data from the US Current Population Survey, Duygan-Bump, Levkov, and Montoriol-Garriga (2010) show that employees of small firms that rely heavily on external finance, were more likely to lose their jobs during the great recession. Looking at survey evidence from US managers during the same period, Campello, Gimabona, Graham, and Harvey (2011) show that firms reacted to financing constraints by reducing investment, R&D spending, and employment. Chodorow- Reich (2012) uses firm-level data to show that firms borrowing from banks most a ected the US banking crisis tended to reduce employment the most. Finally, Benmelech, Bergman, and Seru (2010) use three di erent credit supply / cash-flow shocks to identify the e ects of credit constraints on the labor demand of US listed firms. These papers have in common that 3
they focus on direct and rather extreme credit supply shocks (the recent US banking crisis in most cases), and use various measures of exposure to bank financing to implement di -in-di methodologies. We approach the question of credit constraints indirectly through the role of collateral. In our setting, firms di er, not by the extent to which they are credit-constrained, but by the extent to which they can overcome constraints via collateral pledging. In this sense, our study is more one of the joint dynamics of collateral value and labor demand in normal times. In doing so, our paper also contributes to the recent literature on the real e ect of collateral supply shocks. This literature like our paper focuses on real estate collateral. Mian, Rao, and Sufi (2011) and Mian and Sufi (2011) document the e ect of housing prices on household consumption both in the house price run-up of 2002-2006, and in the economic slump of 2007-2009. Adelino, Schoar, and Severino (2013) and Schmalz, Sraer, and Thesmar (2013) look at the impact of house prices on entrepreneurial activity. Gan (2007) and Chaney, Sraer, and Thesmar (2012) focus on the e ect of real estate prices on corporate investment. The present paper is the first one, in this literature, to document employment e ects. We describe our dataset in Section 2. In Section 3, we describe our empirical strategy, and outline a simple model to understand the results. Section 4 describes the results and details the many robustness checks that we make. Section 5 contains the discussion of our two findings: The size of aggregate e ects, and the di erence between the e ect of collateral on labor and investment. Section 6 is the conclusion. 2. Data This Section describes the data and the algorithm that we use to calculate the value of real estate holdings at the firm-level. 4
2.1. Data Sources The sample is a large dataset of French firms from 1998 to 2007. We use administrative data sources only, made available by INSEE, the French statistical o ce. Accounting data come from tax files. To tax authorities, all French firms which pay the corporate tax under the standard regime have to report detailed balance sheet and income statements. We exclude industries directly related to the real estate market: construction, finance and real estate development: This ensures that real estate prices hence collateral value do not obviously proxy for investment opportunities. Employment data come from plant-level data. We merge plant-level data on employment with firm-level accounting data using the firm identifying number. One firm may have several plants. Most of our regressions restrict the sample to single plant firms, as it makes the determination of the location of a firm s real estate easier. We do, however, run robustness checks including multi-plant firms. The variables are constructed to follow, as much as possible, those used in our previous study on US firms (Chaney, Sraer, and Thesmar, 2012). As our normalization variable, we used Immobilisations Productives, whichmeasuresphysicalcapital,andistheequiva- lent to the PPE variable used in all US-based investment papers ( Plant, Property, and Equipment ). We label it henceforth PPE. Investment is the ratio of capital expenditures to lagged PPE. Cash flows are the standard measure of Net cash flows used in the literature: Net Income - Exceptional Items + Depreciation. Investment is Capital Expenditure. Labor demand is measured with two variables. We discuss in Section?? the meaning of our two measures of labor demand. The first measure is employment change divided by lagged PPE ( L/K). The second one is the symmetrized employment growth (2 (L t L t 1 )/(L t + L t 1 )). This measure of employment growth is taken from Davis, Haltiwanger, and Schuh (1996). It is bounded above by 2 (from 0 to x jobs) and below by -2 (from x to 0 jobs), and can accommodate both entry and exit. Employment growth is less thicktailed than employment to physical capital ratio (s.d. equal to 0.23 instead of 1.98). Results 5
are therefore likely less sensitive to extreme values for our second measure although all our variables are windsorized. We perform robustness checks to assess the importance of extreme values. Our main sample is unbalanced, firms may enter and exit the sample as they start and cease to pay the regular corporate tax. It contains slightly more than 1m observations - about 100,000 mono-plant firms per year. Entry and exit may be related to creation and bankruptcy, but it may also be due to the fact that firms cross the size threshold below which it is optimal to opt for the simplified tax regime (a few hundred thousands Euro of annual turnover, depending on the firm). Firms may exit the sample because they are terminated more likely for smaller firms or acquired by a larger entity more likely for larger firms. In order to control for observable determinants of real estate holdings, we also need to run anumberofregressionsonfirmsthatarepresentinoursamplein1998. Thisallowsus ton control for such determinants firm size, industry etc. as of 1998. This alternative sample is also unbalanced, but there is by construction no entry after 1998. It has 750,000 observations. All variables are trimmed: We take out observations whose value is above the 99 th and below the 1 st percentile. For most variables, this procedure leads to the suppression of all outliers. One exception is the ratio of employment change to physical assets L/K, which has a bit more outliers than the other variables. We test that our estimates are robust to the trimming/windsorizing technique in Table 5. 2.2. Estimating Real Estate Value at the Firm-Level Key to our estimates is the calculation of the ratio of real estate value to physical assets, RE Value. To maximize the precision of our real estate value variable, we first focus on single plant firms. Because our real estate measure relies on several assumption, we also provide, in robustness checks, regressions using an Owner dummy interacted with house price levels. All our results go through using this alternative measure, but the advantage of RE Value is 6
that it makes quantification much more intuitive. For each firm-year observation, we first estimate the date at which the real estate was purchased. To do this, we use the fact that French tax files provide us with the gross (of depreciation) and the net (of depreciation) book value of buildings. Assuming linear amortization over 25 years, we calculate one minus the ratio of net building assets divided by gross building assets, and multiply it by 25. This gives us the implicit average age of buildings held by the firm, which allows us to deduct the date at which the average Euro of building has been purchased. Our methodology thus implicitly assumes that real estate has not been purchased prior to 25 years before the current date of the observation. There are very few observations with building age above 20 years, which suggests that this leftcensorship of our data is unlikely to be a concern. We then use local house prices to impute the market value of real estate holdings of each firm. To do this, we focus on single establishment firms, and assume that all real estate holdings are in the same département as the establishment there are 95 départments in mainland France. We need to make this assumption because our data does not report the exact location of real estate holdings. We then multiply the book value of real estate at the time of purchase (measured by the gross book value) by one plus the cumulative price growth of house prices in the département where the establishment is located. As commercial real estate price series are unavailable in France, we use house prices. Our previous study on investment in the US showed that using commercial or residential real estate prices led to similar estimates. Due to data limitations, we need to make additional assumptions, however. House prices are obtained from INSEE, but are unfortunately not available back to 1973 (1998-25). They are available at the département-level starting in 1998 which is why we start our study in 1998. Between 1986 and 1998, we can backfill the series using region-level price data also available from INSEE there are 20 regions in France, each of them including on average 5 départements. Prior to 1986, we use the nationwide inflation index to backfill house price data. 7
After this procedure, we normalize the value of real estate holdings by physical capital. This number is available for more than 1m observations. The median RE value is zero: 56% of the observations correspond to firms, which own zero real estate. The average RE value is.28: For the average firm, real estate holdings are worth nearly a third of their physical capital. Figure 1 calculates the year-by-year average of RE value. Clearly, fortheaverage firm, the value of of real estate holdings increases by some 10 percentage points. This is because the period that we study is one of rising house prices: in Figure 2), we report the house price levels in all départements since 1998. 2.3. Comparison with US Data In this Section, we check that our French data have properties similar to COMPUSTAT, the standard firm-level data in US studies. To do this, we regress employment growth and investment divided by PPE, on a measure of cash-flows. We check that estimates have the same order of magnitude. In both samples, we use the same accounting items. Cash-flows are Net Income before Extraordinary Items plus Depreciation, both in the French data and in COMPUSTAT. Investment is measured through CAPEX, normalized by lagged PPE. The employment variable is notoriously unreliable in COMPUSTAT, but we retrieve it nonetheless, and calculate as we do for the French sample its growth. For firm j, inindustrys, atdatet, wethenrunthefollowinginvestment-to-cashregressions: Inv t = a j + b st + cash jt PPE jt 1 + jst (1) where we cluster the residual jst at the firm level. We allow for firm-specific fixed-e ets a j and industry shocks b st. Industry shocks are here to capture time-varying investment opportunities we have no Tobin s Q for unlisted firms. We also run the same regression 8
using Emp t Emp t 1 on the LHS. The results of these regressions are reported in Appendix Table A.1, using both LHS variables (capital and labor growth), and with and without firm fixed e ects. The bottom line is that we obtain similar orders of magnitude. For investment to cash sensitivity, we obtain.05 on French data, vs.02 on US data. Both estimates are strongly significant. The French estimate is larger, consistent with the idea that French firms may be more financially constrained. Note however that cash-flows are also a measure of productivity, in which case this might mean that investment of French firms is more sensitive to productivity, and hence that French credit markets are more e cient. Given the massive endogeneity concerns in these regressions, it is hard to draw any inference from these comparisons. The following analysis, on French data only, deals with such concerns. 3. Empirical Strategy and Economic Framework This Section describes our estimation strategy and outlines a simplistic model to understand our results. 3.1. Empirical Strategy For firm j, indépartementd, atdatet, weestimatethefollowingregression: Y jdt = a j + b dt + RE Value jdt +controls jdt + jdt (2) where a j is a firm fixed e ect, and b dt a fixed e ect designed to control for département-level shocks. RE Value jdt is our measure of the value of real estate. We cluster error terms at the département level (95 clusters). This is conservative but designed to capture that our source of variation of RE Value is in partly but not only driven by département level house prices. Y jdt is a measure of investment in our first regression table and of labor demand in all our subsequent tables. 9
Equation (2) is identified of o the interaction of two main sources of variation: real estate ownership by firms, and département-level house price movements. The first source is the age of real estate holdings, that varies across observation, both across and within firms firms may acquire or sell real estate. In 1998, 44% of the firms own some real estate. To measure the role of the extensive (own vs rent), as opposed to the intensive (firm-level sequence and size of real estate purchases) margin, we run, in robustness checks, regressions including the Owner dummy interacted with département-level house price level. This amount to assuming that the age of buildings is constant over time and across owning firms. As we will see, this approach already leads to strong identification. The second source of identification is the variation of house prices, which di ers across départements. 1998-2007 is a period during which house prices increase dramatically in the whole country, but they increases di erentially across regions. We report house price indices by département in Figure 2. Panel A plots the trends for all départements taken together. Panel B plots trends for the top 20 départements in terms of number of firms in 1998: Altogether, these départements account for half of the firms in 1998. These two panels show that real estate inflation varies considerably across départements, from about 50% to more than 150%. This is our second source of variation. 3.2. A Set-Up to Help Interpretation We will seek to interpret the magnitude of the coe cient in equation (2). We will compare the coe cients obtained in investment and labor regressions. To understand how we can make this comparison, let us start from the following benchmark: there is no adjustment cost for labor, and the production process can be modeled as a CRS Cobb-Douglas production function: Q t = A.L t Kt 1. 1 In this model, K t is given by a dynamic optimization problem taking into account adjustment costs in capital and financial constraints. Given the absence of labor adjustment costs, labor demand is simply given by: 1 The analysis presented here holds for all production functions that are homogeneous of degree 1. 10
A 1 L t = K t (3) w Assuming productivity and wage constant, labor demand thus has to be proportional to capital. Let us now assume that some shock to financing constraints leads to a change in investment K along the optimal path, and investigate the impact on labor demand. In our regressions, we will take two specifications. In our first set of specifications, we use change in labor force normalized by physical capital L/K. If our benchmark model applies, then equation (3) imposes that: L K = K K A w 1 K L = K K (4) Thus, when we regress L and K on our collateral shock, we expect the ratio K K L/ K to be equal to L K. If it is smaller than this, then this suggests that labor demand is less responsive than under the no-adjustment cost model. In our second set of specifications, we use employment growth as the LHS variable of equation (2). Equation (3) shows that, in this case, L/L = K/K. Thus, when we regress L/L on the collateral shock, we expect to find the same coe cient as in the investment regression. If it is smaller, this suggests that adjustment costs on labor prevent the firm from fully adjusting labor demand to the increase in capital stock. 4. Results 4.1. Investment We first replicate the results of Chaney, Sraer, and Thesmar (2012), which look at the impact of collateral shock on investment, using a sample of publicly listed firms in the US. We thus need to verify first that the baseline investment regression still works (1) in a di erent country and (2) on a sample of smaller, unlisted firms. 11
The results are reported in Table 2, columns 1-4. Column 1 only includes firm and year fixed e ects. Column 2 includes further time-varying controls. These controls are designed to capture observable determinants of real estate holdings. This is because real estate value is partly endogenous for instance through the decision to own vs rent, but also through the decision of when to buy (market timing). For instance, firms who are able to do better real estate market timing may also be those who invest more in upswings. This should be partly captured by firm fixed e ects, and the fact that we exclude construction and real estate industries, but we can do better: We include the price index interacted with firm observables which are known to predict the value of real estate holdings, such as firm size, industry, and profitability. We take these observables in 1998 (the first year of our sample) and therefore require firms to be present since the beginning: This is why the number of observations falls by some 250,000. We find that these controls do not change the point estimate at all. In column 3, we add one additional control: the cash-flow to capital ratio, which is a standard determinant of investment because it captures both financial constraints and firm productivity. Our estimate decreases a little but remains very strongly significant. We finally seek to control for investment opportunities. Because most of our firms are not listed, we cannot calculate the market-to-book ratio, that the investment literature typically uses as a proxy for marginal Q. Instead, we include industry region year dummies in order to capture industry-level, local, time variation in investment opportunities. In spite of this demanding set of controls (11 years 95 regions 20 industries = 20,900 di erent dummies, in addition to firm-fixed e ects), our baseline estimate is una ected. We find much bigger e ects in the sample of French firms than in our previous US study. The fully saturated specification in column 4 estimates that 1 Euro of real estate wealth leads to an increase in investment by some.19 Euro. In our sample of publicly listed US firms, we found that a 1$ increase in real estate value led to a 0.06 $ increase in capital expenditures. Hence, the e ect on French firms that we estimate here is about three times larger. The exact specification di ers slightly because we do not have Tobin s Q for French 12
firms most of them are not listed. This is consistent with the idea that our sample of small, bank-depend firms, are more financially constrained. We then check that capital gains also lead to debt issue. The underlying mechanism of collateral e ects is that firms pledge real estate holdings to increase their debt capacity. Thus, we expect firms to issue enough new debt to finance their investment. We look at debt issues in Table 2, columns 5-8. The LHS variable in these columns is the change in total debt (bank-financed and trade payables) normalized by lagged capital. These four columns progressively include the same controls as in columns 1-4. We take comfort in the fact that the progressive inclusion of these controls does not a ect our baseline estimate too much. Using the last column of Table 2, we find that each additional Euro of real estate collateral leads to 17 cents of additional debt. This is consistent with the idea that most, if not all, of the additional investment observed in columns 1-4 is financed through debt issues, lending further credence to the collateral interpretation. 4.2. Labor Demand We explore two distinct specifications using labor. First, we use, as a LHS variable, the ratio of employment change to capital. These results are reported in Table 3, columns 1-4. This variable is multiplied by 1000 to make estimates readable. Hence, a coe cient of 1.5 means that, for about 1,000,000 Euro increase in real estate value, labor demand increases by 1.5 job. This estimate is robust to the inclusion of the controls mentioned in the previous Section. Second, we use, as an alternative LHS variable, employment growth. Estimates are reported in Table 3, columns 5-8. We find that an increase in real estate value by 10 percentage point of capital leads to job growth of 0.2 basis points ( 0.018 0.10). This is a small e ect, in line with estimates from columns 1-4. We will discuss in detail in Section 5whatwemeanby small. 13
4.3. Robustness Checks We provide numerous robustness checks in Tables 4-8. 4.3.1. Sample Selection Table 4 checks the impact of our sample selection. In all columns, we use the fully saturated specification (columns 4 and 8, Table 3) with firm- and region-industry-time fixed e ects, as well as 1998 firm characteristics interacted with price level. Columns 1,3,5,7 use employment / capital as the LHS variable; Columns 2,4,6,8 use employment growth. First, we find that potentially endogenous attrition is unlikely to be a concern. Columns 1-2 look at the balanced sample of firms continuously present between 1998 and 2007. This is a way to evaluate the the extent to which our estimate depends on endogenous attrition rentingfirmsenteringthesamplewithlowinvestmentandlowrealestateholdingsforinstance. We find that our estimates are una ected. In columns 3-4, we further investigate the issue of endogenous attrition by including entering firms that start with at least 5 employees. Again, estimates are unchanged which suggests that our regressions are not very sensitive to the entry of tiny firms. Second, we seek to look at the impact of firm structure. In columns 5-6, we restrict the sample to firms that do not belong to a business group. These firms are likely to be more financially constrained, but on the other hand their financial statements are likely to be more precise. Indeed, subsidiaries in business groups may benefit from collateral held by other firms of the group, as well as internal capital markets; they may also lease equipment from other subsidiaries instead of investing. This may make their accounts less reliable and our estimates noisier. Overall, we find similar estimates as in Table 3, suggesting that group structure does not interfere with our estimates. In columns 7-8, we expand the sample to firms with several plants. So far, all of our regressions were run on a sample of singleestablishment firms: This was done to maximize the precision of our real estate variables. Things may become slightly more complex when a firm has plants in several departments. 14
Since we only know the aggregate book value of real estate holdings, we do not know which plant the firm owns, and how much each plant is worth. We thus make the arbitrary choice of assigning to these aggregate holdings the weighted average of price indices in all départements where the firm has plant, and use as weights the number of employees in the plant (the only plant-level information available in the data). Using this alternative measure, and the larger sample of multi-plant firms, we find similar estimates as in our baseline specification. 4.3.2. Estimate Precision In both investment and employment regressions, t-stats are very high, in spite of the fact that we cluster error terms at the département level, a very conservative approach. This is probably due to the huge number of observations that we use in our sample (between 750k and more than 1 million). As an additional test, we run simulations to check for possible spuriousness. We seek to obtain the distribution of parameter estimates ˆ under the null hypothesis that = 0. We use the following procedure. First, our data provide us with the empirical distribution of RE V alue. Then,foreachofourfirm-yearobservations,wedrawonevalue of RE V alue from the empirical distribution, with replacement, and assign this random draw to the observation. Each observation thus has a random value of RE V alue and should equal zero. On this fake sample, we then re-run the regression of column 1, Table 4 fullysaturatedspecificationonthebalancedsample,andobtainoneestimate ˆ. We then repeat this procedure 50 times: we thus obtain a simulated distribution of possible estimates ˆ under the null that real estate holdings are unrelated to investment. We report this simulated distribution in Figure 3. Clearly, the distribution of obtained parameters is tightly located around zero, always much smaller than our baseline estimate. Our estimates, although very precise, are therefore not spurious. 15
4.3.3. Variable Definition In Table 5, we test alternative variable definitions. Columns 1-2 show that a large fraction of the identification comes from the interaction between an Ownership Dummy (equal to 1ifthefirmhaspositiverealestateholdings)andpricechangeinthedépartementwhere the firm is located. Columns 3-6 revert to the RE Value. Columns 3 uses the standard (ie not symmetrized) employment growth measure ( Emp/Emp t 1 ). We find the same e ect as with the symmetrized job growth measure. Column 4-6 check how sensitive our results with Emp/ as a LHS variable are to the trimming/windsorizing procedure. We do this because Emp/ still has extreme values in spite of our trimming procedure that removes observations above the 99 th and below the 1 st percentile. In column 4, we remove all observations that are more than 5 times the interquartile range away from the median. In column 5, we windsorize at the 5 th and 95 th percentiles. In column 6, we trim at this level. The estimate is a bit sensitive to the treatment of outliers, but always remains strongly significant and in the vicinity of.5-1.5. 4.3.4. Large vs Small Firms Tables 6 and 7 show that the estimated collateral channel on labor is not a small firm e ect. Table 6 focuses on firms already present in 1998, and splits the sample into three terciles of firm size, as measure through total assets in 1998. The estimated e ect is, if anything, slightly larger for larger firms. This suggests that larger firms are not less likely to be financially constrained, a result in line with our earlier study on US corporations. Table 7 checks that the e ect is not driven by local booms. Indeed our estimates remain vulnerable to one criticism: house prices may capture local booms, while real estate ownership may capture exposure to local booms. For instance, retailers may tend to own their building, and are more likely to hire when local consumption is picking up. To account for this, we split industries into local, intermediate and global categories. To construct these categories, we calculate, at the 2-digit industry level, the average fraction of exports in total sales 16
our accounting data report, at the firm level, total sales and total exports. We then split industries by tercile of average export rate. We find that hiring in global industries tends to be more sensitive to the real estate value, than in local sectors. This suggests that the sensitivity to local booms interpretation of our result does not seem to be validated by the data. 4.3.5. Dynamics As a last robustness check, Table 8, Panel A, explores the dynamics of the collateral channel on labor. Our main specification assumes very simplistic dynamics: It implicitly assumes that a 1 Euro increase in collateral value leads to a given increase in labor demand during the same period. The true dynamic is probably more complicated, even though it is hard to address properly in a panel regression. To describe the dynamics a little bit better, for each k between 0 and 5, we regress Emp t+k /P P E t+k 1 on RE Value t.wereporttheresultsin Panel A. We find that the maximum e ect of 1 Euro increase in collateral is reached after 1 year (2.7 jobs created per million Euro of collateral), and it decreases afterwards. It becomes insignificant and small after 3 years. One possible concern with this approach is that these lead-lag regressions change de sample composition drastically. This is because our sample spans a relatively short period of time: 1998-2007. Hence, making sure that Emp t+k /P P E t+k 1 is not missing reduces the sample period by k years. For instance, the last regression of Table 8, Panel A, is estimated over 1998-2002 only. If the parameter of interest is di erent in the early part of the sample, then the lead-lag regressions may just capture the fact that sensitivity of collateral is higher in the early part of the sample. To check whether this is the case, we re-run the baseline regression on each of the subsamples on which the lead-lag regressions are estimated. We report these results in Table 8, Panel B, and find that, if anything, what happens is the opposite: the instantaneous e ect is larger at the beginning of our sample period, as shown in Panel B, column 6. Hence, the dynamic that we observe in Panel A is not mechanically 17
driven by right censoring of the data. 5. Interpretation & Discussion We discuss here our two main findings. First, the aggregate e ects implied by our microestimates are quite large. Second, at the micro-level, labor is much less a ected by credit constraints than what one would expect given their measured impact on capital. 5.1. Aggregate E ects The first finding of this paper is that the aggregate e ect of the collateral channel is quite large. While our model like most micro-econometric models of firms only captures a small fraction of the cross-firm variation, it has non-negligible macro implications on investment. As documented in Figures 1 and 2, France has experienced a sharp increase in real estate prices during the 2000s. Between 2002 and 2006, Figure 1 shows that average real estate value to PPE goes up from 25 to 35%. 2 Given our estimates from Table 2, this suggests an increase in aggregate investment by approximately 2 percentage points higher than without the collateral e ect (.19.1). This is very large compared to the cumulative growth in aggregate non-financial corporate investment over the same period, which is equal to 16.5 percent. Hence, our estimates suggest that the collateral channel alone may have been responsible for about 12.5 percent of French aggregate investment growth over 2002-2006 ( 2/16.5). When looking at employment, we find the e ect to be smaller, yet quite sizable. Taking the estimate from Table 3, column 8, we find that aggregate job growth due to the collateral channel to be equal to 0.2 percent over 2002-2006, since the average RE V alue goes up by 10 percentage points ( 0.017 0.10). Given that total employment in France is in the 2 Aggregate numbers show the same trend: total real estate value divided by total capital goes up from 40 to 50%. The increase has the same size, but levels are higher because, in the data, larger firms are more likely to own valuable real estate collateral. 18
vicinity of 20m, this suggests that the collateral channel may have led to the creation of approximately 40,000 jobs between 2002 and 2007, as much as 10% of aggregate job growth during the same period (source: INSEE, national accounts). Another way to get at the same aggregate number is to directly look at the increase in the value of real estate held by firms between 2002 and 2006. To be conservative, we focus only on firms already present in 1998, and aggregate our measure of market values of real estate holdings, over all firms in the sample. We find an aggregate capital gain on real estate of some 26 Bn Euro during the period. Using the estimate from Table 3, column 4, this leads to slightly less than 40,000 jobs created over the period (12.5 26, 000 1.5). 5.2. Collateral E ect on Labor vs Capital The second finding of this paper is that, when collateral value goes up, firms seem to be hiring much less that the typical labor-capital complementarity seems to suggest. Recall from Section 3.2 that K = L K,where K K K L L is the average capital to labor ratio that we observe in the data. Start from the fact that 1m Euro capital gain leads to a capital stock increase by about 190,000 Euro ( K =.19 1m). We have seen in Table 2, that a 1m Euro capital gain typically leads to the creation of 1.5 new job ( L =1.5). This suggests a marginal capital to labor ratio ( K )of130,000europerjob. Thisismuchhigherthan L the average firm-level capital to labor ratio in the existing stock ( K L ), which is equal to 45,000 Euro per worker (median: 19,000): 130,000 is in fact at the 95 th percentile of the distribution of capital-to-labor ratios. Hence, the assumption of stable production function and zero labor adjustment cost made in Section 3.2 are violated. Job growth estimates from columns 5-8 deliver the same message. Recall, from section 3.2, that with a Cobb Douglas technology, we expect L L = K K. As we have seen from Table 2, columns 1-4, capital growth is equal to.2 percentage point for a 1 percentage point increase in real estate value to capital. The same increase in real estate value leads to a.02 percentage point increase in labor, from Table 3, columns 5-8. Hence, labor demand 19
reacts 10 times less strongly than investment to a given capital gains on real estate. This contradicts the Cobb-Douglas assumption without labor cost adjustment of Section 3.2. All this suggests that either: Labor may have much larger adjustment costs than capital, which explains why it responds less to decrease in financing contraints, or Firms use improved access to capital to adopt labor-saving technologies, or Labor requires upfront financing, before it actually generates profits. Such need for working capital finance can be subject to financing constraints itself. The last point is less likely to be relevant, given the orders of magnitude we found in our regressions. This is because the working capital need coming from hiring is small compared to the investment made by firms. Recall that a 1m Euro gain in collateral leads to an increase in the capital stock by about 190,000 Euro, while only 1.5 job is created. Assuming the average job costs some 40,000 Euro (the average annual labor cost in our data), and that it takes, say, 6 months, before cash flows start flowing in a conservative estimate. We obtain that the extra working capital generated by 1.5 job is some 30,000 Euro (= 1.5 40, 000 0.5). Hence, given these assumptions, 1m capital gain leads to an extra investment of 190,000 Euro, and labor-related working capital need of only 30,000. The need to finance working capital looks small compared to how much is spent on capital goods, and is thus not likely to matter. Put di erently, had the firm wanted to preserve a constant labor to capital ratio, it could have invested slightly less, and hired much more. 6. Conclusion TO BE COMPLETED 20
References Adelino, M., A. Schoar, and F. Severino (2013): House Prices, Collateral and Self- Employment, Discussion paper, MIT. Benmelech, E., N. Bergman, and A. Seru (2010): Financing Labor, Discussion paper, Northwestern University. Bernanke, B. S., and M. Gertler (1986): Agency costs, collateral, and business fluctuations, Discussion paper, National Bureau of Economic Research Cambridge, Mass., USA. Campello, M., E. Gimabona, J. Graham, and C. Harvey (2011): Liquidity Management and Corporate Investment During a Financial Crisis, Review of Financial Studies, 24(6). Chaney, T., D. Sraer, and D. Thesmar (2012): The Collateral Channel: How Real Estate Shocks a ect Corporate Investment, American Economic Review, 102(6),2381 2409. Chodorow-Reich, G. (2012): The Employment E ects of Credit Market Disruptions: Firm-level Evidence from the 2008-09 Financial Crisis, Discussion paper, Harvard University. Davis, S., J. Haltiwanger, and S. Schuh (1996): Job Creation and Destruction. MIT Press. Duygan-Bump, B., A. Levkov, and J. Montoriol-Garriga (2010): Financing Constraints and Unemployment: Evidence from the Great Recession, Discussion paper, Boston Fed. Gan, J. (2007): Collateral, debt capacity, and corporate investment: Evidence from a natural experiment, Journal of Financial Economics, 85(3),709 734. 21
Kiyotaki, N., and J. Moore (1997): Credit Cycles, The Journal of Political Economy, 105(2), 211 248. Mian, A., K. Rao, and A. Sufi (2011): Household Balance Sheets, Consumption, and the Economic Slump, Working Paper. Mian, A., and A. Sufi (2011): House Prices, Home Equity-based Borrowing and the U.S. Household Leverage Crisis, American Economic Review, 101(5). Schmalz, M., D. Sraer, and D. Thesmar (2013): Housing Collateral and Entrepreneurship, Working Paper. 22
Table 1: Descriptive Statistics mean median min max sd p25 p75 N Panel A: Sample of all firms from 1998 to 2007 23 CAPX t 0.19 0.08-0.93 1.09 0.28 0.02 0.23 1,078,211 (Debt) 0.11 0.01-2.66 2.67 1.05-0.19 0.35 1,078,211 (Employment) 0.60 0.00-28.55 28.55 13.01-2.21 3.64 1,078,211 (Emp) 2 Emp t+emp t 1 0.00 0.00-0.72 0.72 0.23-0.06 0.08 1,062,435 (Emp) Emp t 1 0.02 0.00-0.73 0.73 0.23-0.06 0.08 1,059,840 RE Value 0.28 0.00 0.00 1.82 0.49 0.00 0.36 1,078,211 Dep. Index 1.38 1.24 0.75 3.25 0.46 1.05 1.64 1,078,211 Cash t 0.85 0.28-2.67 4.79 1.29 0.05 0.97 1,078,211 Panel B: Sample of firms in 1998 Owner Dummy 0.44 0.00 0.00 1.00 0.50 0.00 1.00 125,065 Assets 9.28 9.15 3.74 19.03 1.19 8.49 9.93 86,228 ROA 0.06 0.05-0.43 0.53 0.12 0.01 0.11 86,228 Leverage 0.16 0.11 0.00 1.08 0.17 0.04 0.23 86,228 Note: CAPX t is Capital Expenditure normalized by previous year PPE. (Employment) (Debt) is the change in total debt normalized by previous year PPE. is the change in total employment normalized by previous year PPE (in million Euro). 2 Emp t+emp t 1 is the symmetric growth rate (Emp) of employment. Emp t 1 is the standard growth rate of employment. RE Value t is our proxy for the market value of firms real estate assets. Dep. Cash Index is the real estate price index defined at the département level. t is the ratio of current cash flows to previous year PPE. Owner dummy is a dummy equal to 1 if the firm owns some real estate assets in 1998. Assets is total assets. ROA is a return on asset computed as EBIT normalized by total assets. Leverage is total debt normalized by total assets. (Emp)
Table 2: Investment, Capital Structure and Collateral Value 24 CAPX t (Debt) (1) (2) (3) (4) (5) (6) (7) (8) RE Value t.22***.22***.19***.19***.29***.3***.17***.17*** (46) (48) (46) (45) (29) (26) (13) (13) Dep. Prices -.01**.013 -.0059 -.09*** -.054 -.13*** (-2.1) (.99) (-.46) (-3.8) (-1.5) (-3.7) Log(assets) 98 Dep. Prices -.0013.001.0014 -.0038.0058*.0066* (-.95) (.85) (1.2) (-.95) (1.8) (1.9) ROA 98 Dep. Prices -.099*** -.068*** -.071***.022.15***.15*** (-10) (-6.4) (-6.1) (.7) (5.1) (4.9) Leverage 98 Dep. Prices -.059*** -.065*** -.067***.0066 -.02 -.023 (-10) (-11) (-10) (.26) (-.72) (-.83) Cash t /.046***.046***.19***.19*** (30) (30) (44) (42) Observations 1078211 745680 745680 745680 1078211 745680 745680 745680 Adj. R-Square.26.29.3.019.03.047 Year FE Yes Yes Yes Yes Yes Yes Yes Yes Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Dep. Industry Year FE No No No Yes No No No Yes Note: The dependent variable is Capital Expenditure normalized by previous year PPE (Column 1 to 4) and Changes in total debt normalized by previous year PPE (Column 5 to 8). RE Value t is our proxy for the market value of firms real estate assets. See Section 2 for the construction of this proxy. Dep. Index is a département-level residential price index. Log(Assets 98 ) is the log of the firm s total assets in 1998. ROA 98 is the firm s Return on Asset in 1998. Leverage 98 is the firm s leverage in 1998. Cash t is operating cash flows. Column 2 to 4 and 6 to 8 include industry and département fixed e ects interacted with the département-level real estate price index. All specifications include year and firm fixed e ects. Column 4 and 8 include département-industry-year fixed e ects. Our industry classification has 20 groups. Standard errors are clustered at the département level. T-statistics are reporter in parenthesis.***, ** and * means significant at the 1, 5 and 10% significance level.
Table 3: Employment Changes, Employment Growth and Collateral Value 25 (Employment) (Emp) 2 Emp t+emp t 1 (1) (2) (3) (4) (5) (6) (7) (8) RE Value t 2.5*** 2.2*** 1.5*** 1.5***.027***.029***.016***.018*** (28) (23) (15) (16) (12) (12) (6.9) (7.1) Dep. Prices -1.2*** -3.7*** -4.1*** -.019*** -.004 -.011 (-3.8) (-10) (-12) (-2.8) (-.48) (-1.3) Log(assets) 98 Dep. Prices.32***.37***.38*** -.0013* -.0004 5.9e-06 (5.7) (7.3) (7.6) (-2) (-.67) (.0096) ROA 98 Dep. Prices -5*** -4.2*** -4.3*** -.052*** -.04*** -.04*** (-14) (-13) (-12) (-8.8) (-7) (-6.5) Leverage 98 Dep. Prices -.55*** -.7*** -.74*** -.031*** -.034*** -.034*** (-3.4) (-4.2) (-4.4) (-7.1) (-7.3) (-7.4) Cash t / 1.1*** 1.1***.019***.018*** (28) (29) (22) (22) Observations 1078211 745680 745680 745680 1062435 737684 737684 737684 Adj. R-Square.023.043.047.027.043.046 Industry FE Dep. Index No Yes Yes Yes No Yes Yes Yes Dep. FE Dep. Index No Yes Yes Yes No Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Dep. Industry Year FE No No No Yes No No No Yes Note: The dependent variable is Change in employment normalized by previous year PPE (Column 1 to 4) and symmetric growth rate of employment (Column 5 to 8). RE Value t is our proxy for the market value of firms real estate assets. See Section 2 for the construction of this proxy. Dep. Index is a département-level residential price index. Log(Assets 98 ) is the log of the firm s total assets in 1998. ROA 98 is the firm s Return on Asset in 1998. Leverage 98 is the firm s leverage in 1998. Cash t is operating cash flows. Column 2 to 4 and 6 to 8 include industry and département fixed e ects interacted with the département-level real estate price index. All specifications include year and firm fixed e ects. Column 4 and 8 include département-industry-year fixed e ects. Our industry classification has 20 groups. Standard errors are clustered at the département level. T-statistics are reporter in parenthesis.***, ** and * means significant at the 1, 5 and 10% significance level.
Table 4: Employment Changes, Employment Growth and Collateral Value: Robustness Checks I 26 (Emp) (Emp) (Emp) (Emp) (Emp) (Emp) (Emp) (Emp) 2 Emp t+emp t 1 2 Emp t+emp t 1 2 Emp t+emp t 1 2 Emp t+emp t 1 (1) (2) (3) (4) (5) (6) (7) (8) RE Value t 1.7***.023*** 1.4***.017*** 1.5***.015*** 1.1***.015*** (15) (9.2) (13) (6.2) (14) (5.9) (11) (5.8) Cash t 1***.016*** 1.2***.019*** 1.1***.018***.91***.017*** (23) (24) (24) (16) (26) (18) (30) (32) Log(assets) 98 Dep. Prices.31*** -.0015***.071 -.0041***.41*** -.00033 -.0026 -.005*** (7.5) (-2.7) (1) (-4.1) (6.2) (-.36) (-.058) (-6.6) ROA 98 Dep. Prices -3.9*** -.043*** -3.2*** -.018*** -4.7*** -.051*** -3.9*** -.039*** (-8.9) (-5.2) (-9.2) (-2.7) (-12) (-7.9) (-14) (-6.6) Leverage 98 Dep. Prices -.49*** -.025*** -.39** -.024*** -.76*** -.034*** -.51** -.026*** (-2.9) (-6.3) (-2.1) (-6.4) (-4) (-7.7) (-2.5) (-6.2) Observations 540740 536465 853931 844234 648515 642034 786392 776745 Adj. R-Square Year FE Yes Yes Yes Yes Yes Yes Yes Yes Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Dep. Industry Year FE Yes Yes Yes Yes Yes Yes Yes Yes Note: The dependent variable is Change in employment normalized by previous year PPE (Column 1, 3, 5 and 7) and symmetric growth rate of employment (Column 2, 4, 6 and 8). Column 1 and 2 uses a balanced sample of firms from 1998 to 2007. Column 3 and 4 allows for entry of new firms in the sample provided these firms have more than 5 employees. Column 5 and 6 use a sample of stand-alone firms, i.e. not a lliated to a business group. Column 7 and 8 add multi-establishment firms and assign to these firms a real estate price index equal to a weighted-average of the real estate price index for each plant location, where the weights are the number of employees in the plant. RE Value t is our proxy for the market value of firms real estate assets. See Section 2 for the construction of this proxy. Dep. Index is a département-level residential price index. Log(Assets 98 )isthe log of the firm s total assets in 1998. ROA 98 is the firm s Return on Asset in 1998. Leverage 98 is the firm s leverage in 1998. Cash t is operating cash flows. All columns include industry and département fixed e ects interacted with the département-level real estate price indexes well as year, firm and département-industry-year fixed e ects. Our industry classification has 20 groups. Standard errors are clustered at the département level. T-statistics are reporter in parenthesis.***, ** and * means significant at the 1, 5 and 10% significance level.
Table 5: Employment Changes, Employment Growth and Collateral Value: Robustness Checks II 27 (Emp) (Emp) (Emp) (Emp) 2 Emp t+emp t 1 Emp t 1 (1) (2) (3) (4) (5) (6) Owner dummy Dep. Prices.69***.0035*** (7.4) (2.6) Cash t 1.1***.019***.018*** 1.5***.4***.67*** (33) (24) (20) (30) (19) (19) Log(assets) 98 Dep. Prices.35***.00011.0015***.65***.044**.21*** (6.4) (.18) (2.6) (11) (2.2) (5.8) ROA 98 Dep. Prices -4.4*** -.042*** -.047*** -5.3*** -2.8*** -3.4*** (-13) (-6.9) (-8.4) (-12) (-10) (-10) Leverage 98 Dep. Prices -.73*** -.034*** -.038*** -.92*** -.63*** -.82*** (-4.4) (-7.3) (-7.8) (-4) (-5.5) (-5.3) RE Value t.014*** 2.1*** 1.2*** 1.8*** (5.1) (16) (15) (15) Observations 745680 737684 736595 745680 655687 679251 Adj. R-Square Year FE Yes Yes Yes Yes Yes Yes Firm FE Yes Yes Yes Yes Yes Yes Dep. Industry Year FE Yes Yes Yes Yes Yes Yes Note: The dependent variable is Change in employment normalized by previous year PPE (Column 1, 4, 5 and 6), the symmetric growth rate of employment (Column 2) and employment growth (Column 3). Owner dummy is a dummy variable equal to 1 if the firm reports some real estate (Emp) assets in its 1998 balance sheet. Column 4 trims the dependent variable ( ) at the median +/- 5 times the interquartile range, Column 5 windsorizes at the 5th and 95th percentiles while Column 6 trims at the 5 and 95th percentiles. RE Value t is our proxy for the market value of firms real estate assets. See Section 2 for the construction of this proxy. Dep. Index is a département-level residential price index. Log(Assets 98 )isthe log of the firm s total assets in 1998. ROA 98 is the firm s Return on Asset in 1998. Leverage 98 is the firm s leverage in 1998. Cash t is operating cash flows. All columns include industry and département fixed e ects interacted with the département-level real estate price index, as well as year, firm and département-industry-year fixed e ects. Our industry classification has 20 groups. Standard errors are clustered at the département level. T-statistics are reporter in parenthesis.***, ** and * means significant at the 1, 5 and 10% significance level.
Table 6: Employment Changes, Employment Growth and Collateral Value: Variations by Size 28 (Emp) (Emp) 2 Emp t+emp t 1 Small Medium Large Small Medium Large (1) (2) (3) (4) (5) (6) RE Value t 1.4*** 1.5*** 1.5***.012**.014***.025*** (5.7) (7.7) (9) (2.2) (2.7) (5.8) Cash t 1.1*** 1*** 1***.019***.017***.018*** (19) (12) (22) (18) (12) (19) Log(assets) 98 Dep. Prices.52***.64***.53***.00061.0016.00074 (5.1) (5.9) (7.9) (.37) (.94) (.64) ROA 98 Dep. Prices -5.5*** -5.2*** -2.5*** -.07*** -.054***.0046 (-6.8) (-6.6) (-3.5) (-4) (-4.2) (.34) Leverage 98 Dep. Prices -.58-1*** -.77** -.033*** -.032*** -.037*** (-1.4) (-3.5) (-2.3) (-4.1) (-4.1) (-5.1) Observations 252182 227663 238049 249569 225096 235545 Adj. R-Square Year FE Yes Yes Yes Yes Yes Yes Firm FE Yes Yes Yes Yes Yes Yes Dep. Industry Year FE Yes Yes Yes Yes Yes Yes Note: The dependent variable is Change in employment normalized by previous year PPE (Column 1, 2 and 3) and the symmetric growth rate of employment (Column 4, 5 and 6). Column 1 and 4 (resp. 2 and 5, 3 and 6) use the sample of firms in the first (resp. second and third) tercile of firm size as measured by total assets in 1998. RE Value t is our proxy for the market value of firms real estate assets. See Section 2 for the construction of this proxy. Dep. Index is a département-level residential price index. Log(Assets 98 ) is the log of the firm s total assets in 1998. ROA 98 is the firm s Return on Asset in 1998. Leverage 98 is the firm s leverage in 1998. Cash t is operating cash flows. All columns include industry and département fixed e ects interacted with the département-level real estate price index, as well as year, firm and département-industry-year fixed e ects. Our industry classification has 20 groups. Standard errors are clustered at the département level. T-statistics are reporter in parenthesis.***, ** and * means significant at the 1, 5 and 10% significance level.
Table 7: Employment Changes, Employment Growth and Collateral Value: Variations by Industry trade costs 29 (Emp) (Emp) 2 Emp t+emp t 1 Local Interm. Global Local Interm. Global (1) (2) (3) (4) (5) (6) RE Value t 1.5*** 1.8*** 2***.016***.026***.026*** (8) (8.4) (5.6) (3.5) (5.4) (3) Cash t.93*** 1.1*** 1.1***.017***.018***.018*** (16) (21) (8.4) (13) (20) (9.1) Log(assets) 98 Dep. Prices.37***.41***.45***.0012 -.0009 -.0012 (7.7) (6) (3.9) (1.2) (-1) (-.63) ROA 98 Dep. Prices -2.7*** -6.3*** -6.4*** -.02* -.071*** -.034 (-4.8) (-10) (-3.8) (-1.7) (-5.7) (-1.1) Leverage 98 Dep. Prices -.73*** -.9** -2.4* -.029*** -.039*** -.072*** (-3.7) (-2) (-1.7) (-4.6) (-5.2) (-3.4) Observations 363093 321440 61147 356977 319821 60886 Adj. R-Square Year FE Yes Yes Yes Yes Yes Yes Firm FE Yes Yes Yes Yes Yes Yes Dep. Industry Year FE Yes Yes Yes Yes Yes Yes Note: The dependent variable is Change in employment normalized by previous year PPE (Column 1, 2 and 3) and the symmetric growth rate of employment (Column 4, 5 and 6). Column 1 and 4 (resp. 2 and 5, 3 and 6) use the sample of firms in the first (resp. second and third) tercile of industry. Industry trade costs are defined as the average firm-level ratio of export sales to total sales in each 2-digit industry. RE Value t is our proxy for the market value of firms real estate assets. See Section 2 for the construction of this proxy. Dep. Index is a département-level residential price index. Log(Assets 98 ) is the log of the firm s total assets in 1998. ROA 98 is the firm s Return on Asset in 1998. Leverage 98 is the firm s leverage in 1998. Cash t is operating cash flows. All columns include industry and département fixed e ects interacted with the département-level real estate price index, as well as year, firm and département-industry-year fixed e ects. Our industry classification has 20 groups. Standard errors are clustered at the département level. T-statistics are reporter in parenthesis.***, ** and * means significant at the 1, 5 and 10% significance level.
Table 8: Employment Changes, Employment Growth and Collateral Value: Dynamics 30 (Emp ) (Emp ) (Emp ) (Emp ) (Emp ) (Emp t) t+1 PPE t t+2 PPE t+1 t+3 PPE t+2 t+4 PPE t+3 t+5 PPE t+4 (1) (2) (3) (4) (5) (6) Panel A: RE Value t 1.5*** 2.7*** 2.5*** 1.5***.54**.22 (16) (23) (15) (8.4) (2.4) (.81) Cash t 1.1***.58***.23*** -.033 -.089* -.1 (29) (10) (5.1) (-.76) (-1.8) (-1.6) Log(assets) 98 Dep. Prices.38*** -.044.27***.39**.34.37 (7.6) (-.48) (2.7) (2.4) (1) (.86) ROA 98 Dep. Prices -4.3*** -3.7*** -3.3*** -1.9** 1.3 5.8*** (-12) (-8.2) (-4.5) (-2) (.87) (2.6) Leverage 98 Dep. Prices -.74*** -.57** -.47 -.97 -.3 -.89 (-4.4) (-2.6) (-.96) (-1.6) (-.45) (-.72) Observations 745680 647107 558582 477668 401642 328702 Adj. R-Square Year FE Yes Yes Yes Yes Yes Yes Firm FE Yes Yes Yes Yes Yes Yes Dep. Industry Year FE Yes Yes Yes Yes Yes Yes Panel B: (Emp t) on common sample RE Value t 1.5*** 2.1*** 2.5*** 3.1*** 3.9*** 4.5*** (16) (17) (16) (15) (13) (13) Observations 745680 647107 558582 477668 401642 328702 Adj. R-Square Note: In panel A, the dependent variable is Change in employment normalized by previous year PPE from year t+k-2 to year t+k-1 in column k. In panel B, the dependent variable is the current Change in employment normalized by previous year PPE. In panel B, the regression in column k is run for the sample where Change in employment normalized by previous year PPE from year t+k-2 to year t+k-1 is defined. RE Value t is our proxy for the market value of firms real estate assets. See Section 2 for the construction of this proxy. Dep. Index is a département-level residential price index. Log(Assets 98 ) is the log of the firm s total assets in 1998. ROA 98 is the firm s Return on Asset in 1998. Leverage 98 is the firm s leverage in 1998. Cash t is operating cash flows. All columns include industry and département fixed e ects interacted with the département-level real estate price index, as well as year, firm and département-industry-year fixed e ects. Our industry classification has 20 groups. Standard errors are clustered at the département level. T-statistics are reporter in parenthesis.***, ** and * means significant at the 1, 5 and 10% significance level.
Figure 1: Average Real Estate Collateral Value to Capital Ratio Mean Real Estate Value to PPE.2.25.3.35 1998 2000 2002 2004 2006 2008 Year Source: Administrative data and authors calculations. Note: We start from RE V alue, which is the firm-level ratio of estimated market value of real estate holdings to total capital. The average is then taken across all firms in our sample, which excludes construction, finance and real estate industries. 31
Figure 2: House Price Inflation: 1998-2007 Panel A: All 95 Regions Panel B: 20 Largest Regions 32 1 1.5 2 2.5 3 1 1.5 2 2.5 3 1998 2000 2002 2004 2006 2008 Year 1998 2000 2002 2004 2006 2008 Year Source: INSEE. Note: In Panel A, each line represents the house price index for one of the 95 French départments. In Panel B, we restrict ourselves to the 20 départements which have more than 2,000 firms in 1998. In 1998, these départements altogether have 62,504 firms, about half of the 125,065 firms present in our sample that year.
Figure 3: Placebo Regressions Density 0 1 2 3 4-2 -1 0 1 2 Coefficient on RE Value Note: This figure displays the distribution of coe cient estimates of placebo regressions relative to our baseline regression (red line). For each regression, we randomly draw from the balanced sample with replacement the market value of real estate of firms (RE Value) from its empirical distribution and run the regression corresponding to column 1 in Table 4 fully saturated specification on the balanced sample. We repeat this procedure 50 times. 33
A. Appendix Table Table A.1: The E ect of Cash-Flows on Employment and Capital growth: France and the US Inv t (Emp t) Inv t (Emp t) Emp t 1 Emp t 1 (1) (2) (3) (4) Panel A: French Data cash t.067***.035***.05***.035*** (61) (72) (46) (42) log(asset 1998 ).014*** -.00072*** (20) (-3.1) Observations 745680 736595 745680 736595 Adj. R-Square.078.039 Industry Year FE Yes Yes Yes Yes Firm FE No No Yes Yes Panel B: COMPUSTAT cash t.01***.02***.02***.02*** (2.6) (12) (4.9) (6.1) log(asset 1998 ) -.03*** -.01*** - - (19) (7.4) - - Observations 22435 21158 25462 23586 Adj. R-Square.12.056.42.29 Industry Year FE Yes Yes Yes Yes Firm FE No No Yes Yes Source: French administrative data (Panel A), COMPUSTAT (Panel B). Sample period: 1998-2007. Note: The dependent variable is capex over last year PPE (Column (1) and (3)) and change in employment over last year employment (Column (2) and (4)). Cash is earnings before extraordinary items plus depreciation normalized by last year PPE. Log(Asset) is the log of total assets in 1998. Standard errors are clustered at the firm level. *, **, and *** mean statistically di erent from zero at 10, 5 and 1% levels of significance. 34