The Decline in Corporate Liquidity During the Japanese Lost Decade, 1991-2003
Teun Brinkman
Master Thesis - M.Sc Econometrics Final version
December 31, 2010
Abstract
We examine the decline in corporate liquidity in Japan from 1991 to 2003, a period referred to as a "lost decade", due the low growth of real GDP per capita. Cash to assets in 2003 was about half its size of 1991. Using the …rst-di¤erences IV estimator for panel
…xed e¤ects, we examine the determinants of liquidity demand for Japanese …rms between 1991 and 2003, and try to …nd an explanation for the decline in cash holdings. Controlling for other factors, we …nd that the decline is mainly accounted for by a decline in growth opportunities and long-term debt holdings before 1997, where cash holdings were declining the most. After 1997, credit factors seem to play a role: we provide evidence that …rms with poor access to external …nance, so-called …nancially constrained …rms, were saving more cash after 1997. We …nd that this is likely due to an increase in cash ‡ow uncertainty, measured by industry cash ‡ow risk. Using several speci…cations, we do not …nd evidence for a higher ‘cash ‡ow sensitivity of cash’for …nancially constrained Japanese corporations.
This thesis is partially written during an internship at the Institute of Social and Economic Research at Osaka
University. I would like to thank Prof.Dr. Kazuo Ogawa from Osaka University for his very useful comments and
Prof.Dr. Ichiro Tokutsu from Kobe University for his help with collecting the data.
1 Introduction
In this paper we examine the decline in corporate liquidity in Japan from 1991 to 2003, a period often referred to as a "lost decade" (see e.g. Hayashi and Prescott (2002)). 1 The average real GDP per capita growth rate from 1981 to 1990 was 3.4 percent, compared to 1.1 percent from 1991 to 2003. In the same period, cash holdings of Japanese …rms plummeted. Where Japanese corporations were known for their high amount of cash reserves compared to other industrial nations (e.g. Rajan and Zingales (1995) and Pinkowitz and Williamson (2001)), this picture changed during the 1990s. In 2003 the median cash and deposits of our balanced sample of 843 non…nancial …rms was half its size in 1991.
Why did this decline in Japanese corporate liquidity take place? In this paper, we estimate a liquidity demand function for Japanese corporations, and try to answer this question. The demand function is based on a wide range of theoretical hypotheses. As an example, …rms hoard liquidity when they expect pro…table investment opportunities in the near future. We use the …rst-di¤erences IV estimator for …xed e¤ects as a consistent estimator, and estimate the determinants of the corporate cash holdings in Japan. Using this model we try to …nd an explanation for the decline in corporate liquidity. We also pay attention to macroeconomic determinants of corporate liquidity and the role of …nancing constraints.
Our analysis is interesting, since it reveals some of the motives behind the …rms’ …nancial policies during the lost decade. Subjecting our liquidity demand equation to di¤erent speci…ca- tions, variables and samples, we …nd the following main results. If we control for balance sheet substitutes as net working capital and long-term debt, we …nd that …rms save in the prospect of growth opportunities, measured by the market-to-book ratio of assets or two-year sales growth.
Industry cash ‡ow risk is a signi…cant positive determinant, when we exclude non-manufacturing
…rms, or when we remove …rms with an average negative cash ‡ow during the lost decade.
To …nd an explanation for the decline in corporate liquidity, we split the sample around March 1998. In March 1998, median cash and deposits was 35% lower than in 1991. We …nd that the decline in growth opportunities and the decline in long-term debt holdings, were the main factors in this decline up to 1998. After 1997, we …nd a di¤erence in saving behaviour between …nancially constrained and unconstrained …rms. Doing a regression analysis, we …nd that mainly constrained …rms started saving after March 1998, as a response to industry cash
‡ow risk.
Looking at various macroeconomic factors, we …nd in particular that credit, proxied by a bank stock index or a corporate credit spread, plays a role: when credit conditions are poor, …rms hoard liquidity. Almeida, Campello and Weisbach (2004), have hypothesized that …nancially constrained …rms have a higher cash ‡ow sensitivity of cash. Using several speci…cations, we do not …nd evidence for this hypothesis for Japanese …rms from 1991 to 2003.
Our research contains two main novelties. First, we pay more attention to consistent estima-
1
The year 2003 is chosen, since the Japanese economy starting growing again in 2004 and 2005.
tion than in any previous paper on this topic. We estimate a liquidity demand function for …rms, by taking into account endogeneity of the regressors and weak exogeneity of the instrument set.
Where most papers on this topic (e.g. Opler, Pinkowitz, Stulz and Williamson (1999) and Bates, Kahle and Stulz (2009)), implicitly assume strong exogeneity when estimating …xed e¤ects, we argue that this is a wrong assumption. As …rms’ …nancial policies are dynamic by nature, it is unlikely that unexpected time-t shocks to cash holdings, do not a¤ect regressors - or more precisely: instruments - at time t+1.
The second novelty, is that we are the …rst to research the causes of the decline in corporate liquidity throughout the lost decade. Some studies have focused explicitly on the liquidity demand of Japanese corporations. Hoshi, Kashyap and Scharfstein (1991) show that …rms which are part of some sort of conglomerate (a keiretsu) hold more liquidity. Pinkowitz and Williamson (2001) look at corporate liquidity in a di¤erent time period, namely from 1974 to 1995. Luo and Hachiya (2005) examine corporate liquidity in Japan between 1989 and 2002, but they are concerned with the role of corporate governance in cash holdings, and the e¤ect of cash holdings on agency problems. The decline in corporate liquidity is not a subject of their study.
This study relates to three broader categories of literature. First, it relates to the empirical literature on corporate liquidity demand, such as Kim, Mauer and Sherman (1998), Opler et al.
(1999) and more recent Bates et al. (2009). Bates et al. (2009) try to explain the rise in U.S. cash holdings during the 1990s and 2000s. Our empirical strategy is mostly related to the strategy employed in Bates et al. (2009), although we pay more attention to consistent estimation.
Second, this study relates to the literature on the role of …nancing constraints in corporate
…nancial policies. In their seminal but often critized work, Fazzari, Hubbard and Peterson (1988) hypothesize and …nd, that …rms who have more di¢ culty to obtain external …nance, have a higher cash ‡ow sensitivity of investment. More recently, Almeida et al. (2004) describe and test a theoretical model that predicts that …rms with …nancial constraints should have a positive cash
‡ow sensitivity of cash, where this is zero for unconstrained …rms.
Finally, this paper relates to the literature on Japanese …rms during the lost decade. Nagahata and Sekine (2005) and Ogawa (2007b) study the e¤ect of the …rms’ high leverage on …rms’
investment policies during the lost decade. Gan (2007) looks at the e¤ect of collateral values on
…rm investment. We add to this literature, with our e¤ort to explain the decline in corporate cash holdings during the lost decade.
We proceed as follows. In the next section, we present theoretical literature on the deter-
minants of corporate liquidity demand. After this, we document the subject of this study: the
decline in corporate cash holdings during the Japanese lost decade. In the fourth section, we
describe our consistent estimator. Using this estimator, we estimate a liquidity demand equation
and check its robustness to di¤erent speci…cations in section …ve. In section six, we try to explain
the decline in corporate liquidity and look at the role of macroeconomic factors. In section seven,
before we conclude our paper, we look at the role of …nancial constraints in liquidity demand.
2 The corporate demand for liquidity
What is the rationale for saving and borrowing at the same time? There is a very wide range of explanations and models in the literature. We will use the most important determinants in our empirical model, to be explained in section 4. In his General Theory of Employment, Interest and Money, Keynes (1936) gives an analytical explanation for the motives to demand and hold liquidity. His analysis is still often referred to in the literature, and gives us a good understanding of the determinants of corporate liquidity demand. His analysis starts with the question "[W]hy should anyone prefer to hold his wealth in a form which yields little or no interest to holding it in a form which yields interest [?]" (p. 168).
He identi…es three motives. Two of these, the transactions motive and the speculative motive are exempli…ed in the liquidity demand models of Baumol (1952) and Tobin (1956). Following the transactions motive, agents need cash for everyday transactions, but raising cash has a …xed cost C. On the other hand, according to the speculative motive, cash holdings have an opportunity cost equal to the interest rate or return r on other assets A. 2 With N withdrawals at regular intervals, the total opportunity cost resulting from foregone returns is 2N rA . Minimizing the sum of transaction costs N C and foregone interest 2N rA , with respect to N , we get an optimal number N of withdrawals q
rA
2C . With an average size of liquid funds of 2N A , the optimal average amount of liquid funds L is
L = r AC
2r
The result of this intuitive model is that the optimal ratio of liquid to illiquid funds L A , is declining in the size of illiquid funds A, but increasing with the …xed transaction cost. Therefore, bigger …rms are expected to hold less cash as a share of assets. This can as be interpreted as economies to scale in cash management, something found empirically by Mulligan (1997). Also, a decline in …xed transactions costs over the years with the advancements in risk management techniques and internet technology, should lead to less transactions-based holdings of cash.
Where the transactions motive is about regular and certain expenses, the precautionary motive induces …rms to hold cash as an insurance mechanism against uncertain and irregular expenses. Firms want to avoid ‘liquidity shortages’, since these are more costly to …rms than having an excess of liquidity. Holmstrom and Tirole (2000) present a model with very intuitive explanation why it is expensive for …rms to be short of liquid funds. In their model, …rms and investors design a contract for a two-period investment project at date 0. At date 1, the
…rm experiences an unexpected exogenous liquidity shock, e.g. a cost overrun, requiring a cash infusion. Firms can either choose to hoard liquidity before date 1, or ‘wait-and-see’. A ‘wait- and-see’ policy is only optimal, if the reinvestment need resulting from the liquidity shock, is lower than the continuation value or pledgeable income for investors. In that case, investors
2
More precisely, Keynes (1936) asserts that agents manage their cash holdings in order to "[secure] pro…t from
knowing better than the market what the future will bring forth." (p.170)
are willing to lend more money. However, when there is a probability that this will not occur,
‘wait-and-see’is not optimal and …rms will hoard liquidity at date 0 to insure against liquidity shock, thereby creating a corporate demand for liquidity.
From this precautionary motive, Kim et al. (1998) and Opler et al. (1999) hypothesize in their empirical studies of cash holdings in the United States, that …rms with better growth opportunities and higher cash ‡ow risk hold more cash. Or, in the words of Tirole (2006):
"As ongoing entities, …rms are concerned that they may in the future be deprived of the funds that would enable them to take advantage of exciting growth prospects, strengthen existing investments or simply stay alive." (p. 199)
The precautionary motive also implies that, those …rms for which external …nance is more costly, will hoard more liquidity. In the literature, …rms with more costly external …nance, are said to be ‘…nancially constrained’ (e.g. Fazzari et al. (1988)). In this context, cash holdings contribute to the …rms’‘…nancial ‡exibility’. Using a theoretical model, Almeida et al. (2004) hypothesize and …nd empirically that …rms with costly external …nance have a higher incentive to save cash out of cash ‡ow. Hoshi et al. (1991) …nd that Japanese …rms with closer bank ties have less di¢ culty in obtaining funds and investment is less sensitive to liquidity. We will discuss the role of …nancing constraints in corporate liquidity more explicitly in section 7.
The literature on corporate cash holdings has also identi…ed other motives to hold cash than the classic motives from Keynes (1936). First, there is a literature that stresses the interaction between debt and cash policies. In the model of Holmstrom and Tirole (2000), …rms that choose to borrow more long-term debt, have more trouble to obtain funds at the intermediate stage. If they expect a high liquidity shock at the intermediate date, they will hold less long-term debt and more cash. This suggests that in a cross-section, …rms with more long-term debt are likely to have a lower liquidity need. Acharya, Almeida and Campello (2007) research substitutability between cash and long-term debt in a simultaneous equations model. They identify a hedging motive: …rms with a low correlation between cash ‡ow and investment needs save cash by holding debt, while …rms with a high correlation try to reduce debt by using internal funds.
Second, Jensen (1986) argues that entrenched managers have an incentive to keep money in the …rm, as this gives them power vis-a-vis shareholders and other investors. Also, having internal funds has the advantage of avoiding monitoring.
Third, in their study of the question why cash holdings in Japan where so much more higher than in the United States and Germany between 1974 and 1995, Pinkowitz and Williamson (2001) …nd that …rms with powerful main banks hold more cash, "predominantly to bene…t the main bank". When …rms started to use other sources of …nance, such as public debt, their cash holdings declined.
In section 4, we translate these hypotheses in a regression model, used for tests in the rest
of this paper. Before presenting the model, we describe the decline in corporate liquidity during
the Japanese lost decade.
3 The decline in corporate liquidity in Japan
In this section, we document the subject of our paper: the decline in cash holdings of Japanese companies between 1991 and 2003. We use data from the Firm-level Data Base of the Develop- ment Bank of Japan. 3 Our sample is an unbalanced panel of 1,202 publicly traded non-…nancial
…rms for …scal years 1991 to 2003, with 13,565 …rm-year observations. A …scal year in Japan runs from April this year to March next year. 4 All data is de‡ated to 2005 yens. For descriptions of all data used in this paper, see appendix A. In this section we analyze 843 …rms with balanced data between 1991 and 2003. The results are similar for the unbalanced sample.
Throughout this paper we measure cash holdings as the size of cash and deposits. For between-…rm comparisons, we normalize this, using some benchmark, most often assets at book- value. In the literature on cash holdings in the United States, cash holdings also include mar- ketable securities. We do not include this in our measure of liquidity. As mentioned by Pinkowitz and Williamson (2001), many …rms in Japan have signi…cant cross-holdings of shares, because they are part of some conglomerate (a keiretsu). These cross-holdings cannot be easily sold and therefore they are not very liquid. 5
The cash holdings of Japanese …rms used to be relatively high compared to …rms in other developed countries. Rajan and Zingales (1995) show that industrial …rms in Japan in 1991 had twice as much cash to assets on average than …rms in any other G7 country. Pinkowitz and Williamson (2001) show that Japanese …rms hold more cash than U.S. and German …rms between 1974 and 1995, also conditionally on …rm charactertics. For 1991 in the United States, Bates et al. (2009) report a median cash ratio of 0:072, where the median ratio in Japan in our sample was 0:113. This all changed during the lost decade. In 2003, the …gure was reversed.
The median cash ratio in the United States rose to 0:133, where this …gure in Japan declined to 0:067. 6
In …gure 1, we look at the development of median cash holdings, debt and total assets in real terms since 1991. We see that the median size of cash and deposit holdings in 2003 was about half the size of the cash and deposit holdings in 1991. This is an enormous dive. Where did this decline in Japanese cash ratio’s come from?
Comparing …gure 1 to real GDP per capita growth (e.g. reported in …gure 3 on p.21), we see that short-term debt, and as a consequence net debt (total debt minus cash) and assets at bookvalue, are pro-cyclical, with an upward spike in 1996 and 2000. 7 The median …rm from
3
We would like to thank Prof.dr. Ichiro Tokutso from Kobe University for his help with collecting the data.
4
We exclude …rms who do not have March as its reporting month. This led us to take out about a …fth of all
…rm-years.
5
In their study of cash holdings in Japan, Pinkowitz and Williamson (2001) identify that about 49% of the about 1,300 companies from 1974-1995 are members of some sort of keiretsu. Unfortunately, these data are hard to obtain, so we do take into account keiretsu -membership in this paper. Their e¤ect should be captured by the inclusion of …xed …rm e¤ects.
6
In the unbalanced sample, the median cash ratio declined from 0:113 in 1991 to 0:076 in 2003.
7
Real GDP per capita growth spikes in 2001, but note that the calender year 2001 and the Japanese …scal
40 50 60 70 80 90 100 110
1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 Fiscal Year
In d exed M ed ian ( 1991= 100)
Cash and deposits Assets (bookvalue) Net debt Long-term debt Short-term debt
Figure 1. Real Cash Holdings, Assets, and Debt of Japanese Corporations From 1991 to 2003. For each …scal year, median real values are calculated, and are indexed at 100 in 1991.
Fiscal years run from April this year to March next year. We use a balanced sample of 843 …rms. Net debt is the sum of long- and short-term debt minus cash and deposits.
1991 to 2003 was a net saver, but mainly during economically bad times. 8
We see median long-term debt decline by 25% from 1993 to 1997. Looking at net debt, this decline seems to o¤set the decline in cash and deposits. According to Hayashi and Prescott (2002), …rms relied on cash to deal with the decline in bank loans, due to problems in the banking sector. This is consistent with what we see in …gure 1, but only up to 1997. After 1997 cash and deposit holdings still decline, but long-term debt is rising.
It is also interesting to look at some cross-sectional regularities. In …gure 2, we look at median cash to asset ratio for di¤erent size quartiles. We see that the largest 25% of …rms hold less cash in cross-section. The decline in cash ratio’s occured for all size quartiles.
Also interesting is the occurence of a rise in the holdings of cash during 1997/1998 and after 2000. Both these periods are marked by the …rst and second credit crunch in the Japanese banking sector. Also, after the drop, we see cash to assets rise for all size quartiles. This e¤ect is larger for smaller …rms. Later in this paper, we will show, that adverse credit conditions seem to contribute positively to cash holdings. We also show that …nancially constrained …rms were more a¤ected in the second half of the lost decade.
year 2000 overlap by three months.
8
The average …rm’s net debt was 12% lower in 2003 than in 1991, implying that in the aggregate, saving of
…rms was more serious than the …gure suggests.
0.03 0.05 0.07 0.09 0.11 0.13
1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 Fiscal Year
Cash/ Asset s
Q1: Smallest firm size quartile Q2 Q3 Q4: Largest firm size quartile
Figure 2. Median Cash/Assets of Japanese Corporations for Di¤erent Quartiles of Size. For each …scal year, median values are calculated for di¤erent quartiles of size for a balanced sample of 843 …rms. Fiscal years run from April this year to March next year. Size is measured as sales.
In appendix B, we present a correlation matrix of the main data we employ throughout this paper. We note the main cross-sectional regularities in the data. First of all, the negative correlation between …rm size and cash holdings is con…rmed. For example, the logarithm of sales and the cash to asset ratio have a correlation of 0:18. Second, bigger …rms carry more debt, implying also that …rms with more debt carry less cash. The correlation between the logarithm of sales and total debt to assets is 0:31. The correlation between cash to assets and total debt to assets is -0:33.
However, this does not tell anything about causality. In the next section, we present our empirical model to explore the determinants of corporate cash holdings. In the three sections after that, we will use this model to see which factors determine cash holdings and are likely contributors to the decline in corporate cash holdings in Japan during the lost decade.
4 Empirical model speci…cation
In this section, we …rst describe our empirical model and the measurement of the main variables.
We choose our explanatory variables, following the theoretical decription of the determinants
of cash holdings in section 2. In the second subsection, we present several speci…cation tests,
to come to a consistent estimator for our empirical model. These include tests for …xed e¤ects,
and several tests suggesting the use of instrumental variables, such as endogeneity tests, weak instrument tests and tests for overidenti…cation.
4.1 Model and hypotheses
We test for the determinants of corporate cash holdings by estimating the parameters of the following corporate liquidity demand equation for all …rms i and years t = 1991; ::::; 2003.
CashHoldings i;t = 1 CashHoldings i;t 1 + 2 CashFlow i;t + 3 Size i;t + 4 Market-to-book i;t
+ 5 Industry i;t + 6 Long-termDebt i;t + 7 CapitalExpenditures i;t + 8 NetWorkingCapital i;t + i + t + " i;t (1)
The ’s are coe¢ cients to be estimated. Parameters i and t are respectively …rm and year
…xed e¤ects. The term " it is an error term which is weakly exogenous with respect to the instru- ment set, to be described below. In our estimations, we control for between-…rm dependence in error terms and heteroskedasticity using the cluster-robust estimate of the variance-covariance matrix. All variables, except size, the market-to-book ratio and industry are normalized by assets at bookvalue, unless noted otherwise.
This equation is the most commonly tested equation in this paper, which we also use for speci…cation tests in the next section. As we will explain in the next section, there are some drawbacks of this model, so we will also present results using other dependent and explanatory variables.
In the rest of this subsection, we describe the hypotheses behind the inclusion of the ex- planatory variables in equation (1), based on our theoretical overview in section 2. In the next subsection, we describe the requirements for consistent estimation of this equation, based on speci…cation test results.
Cash holdings are most often measured as cash and deposits divided by total assets at book- value. We will also include other measures of cash holdings such as cash to sales or a more normally distributed measure: the logarithm of cash to net assets (assets minus cash). Thereby we check the robustness of our results to the dependent variable. We include the lagged value of cash holdings in the equation for two reasons. First, cash holdings can be ‘transitory’in the words of Opler et al. (1999), implying that liquidity is hoarded now for the next period. Sec- ond, …rms are likely to become more cautious in using their liquidity when it is low. Therefore, controlling for the size of existing cash holdings could pick up this e¤ect.
Everything else equal, we expect that the demand for liquidity depends positively on the size
of cash ‡ow to assets. Cash ‡ow is measured as net income minus interest and dividend expenses
plus depreciation. We expect that cash ‡ow a¤ects cash holdings positively. On the other hand,
cash ‡ow could have a negative e¤ect on cash holdings, since a high cash ‡ow usually implies
high growth opportunities. We circumvent this, by controlling for a …rms’ growth opporunties (see below).
As described in section 2, following the transactions motive, we theorized that bigger …rms can manage their cash holdings more e¢ ciently, so size will have a negative e¤ect on cash holdings.
Size is often measured as the logarithm of assets. Since the denominator of our measure for cash holdings is assets at bookvalue, we will measure it as the logarithm of sales to avoid automatic negative correlation. When we use cash to sales as dependent variable, we measure size by the logarithm of assets.
Firms hold precautionary cash in expectation of growth opportunities and cash ‡ow risk. We proxy growth opportunities by the ratio between the marketvalue of assets and the bookvalue of assets, following the q-theory of investment. The market-to-book ratio should proxy for Tobin’s marginal q, which measures the market value of the company’s asset vs. its replacement value.
As an alternative measure, we use past two-years sales growth. Cash ‡ow risk is measured as industry cash ‡ow risk, computed as in Bates et al. (2009). We compute the standard deviation of cash ‡ow to assets for this year and previous four years (where we use a minimum of two years) and average this over industry. 9 We hypothesize that …rms with more risky cash ‡ows hold more cash, although higher cash ‡ow risk could also a¤ect pro…tability and therefore reduce cash holdings.
Financial decisions of …rms, such as the decision to change the capital structure, to invest or to build up liqudity, are made simultaneously, as also stressed in section 2. Therefore, we control for net working capital (net of cash), long-term debt to assets and capital expenditures to assets. Net working capital (net of cash) is measured as current assets minus short-term debt and cash holdings. We expect that these explanatory variables are endogenous, as …rms can use cash to increase or decrease working capital, leverage or investments. We account for endogeneity by using instrumental variables. As we use a …rst-di¤erences estimator, we will simply use lagged values as instruments, to be discussed below. We expect a negative e¤ect of net working capital and capital expenditures, since they substitute for cash. Even though we …nd cash and long-term debt to be negatively related in cross-section, we expect a positive substitution e¤ect of long-term debt on cash holdings. Since within-…rms, an increase in debt will a¤ect cash holdings positively. On the other hand, Bates et al. (2009) mention that if debt is su¢ ciently constraining, …rms with high debt should have low cash holdings.
In section 2, we also discussed the in‡uence of some factors that can arguably attributed to non-time varying factors, such as …nancial constraints, keiretsu-membership (see Hoshi et al.
(1991)) or agency problems. 10 Another reason why cash holdings could be higher for speci…c
…rms are regulatory issues, for example in industries with public utilities (public transport, water, etc.). By including …rm …xed e¤ects, we controls for these non-time varying e¤ects and other
9
In section 5, we will also discuss results with slightly alternative risk measures.
1 0
Almeida et al. (2004) relate …nancial constriants to …rms and not to …rm-years. Agency problems are for
example attributed to corporations with entrenched managers, something which has high persistence.
unobserved heterogeneity among …rms. We also mentioned the role of technological advancement in cash management and macroeconomic uncertainty. To control for these time factors we include year …xed e¤ects. In the next subsection we discuss Hausman test results, con…rming that both
…rm and year …xed e¤ects are important for liquidity demand.
Including …rm …xed e¤ects makes the lagged dependent variable endogenous in panel data with a small time-dimension (see Cameron and Trivedi (2005), p.764). 11 This requires the use of instrumental variables. Anderson and Hsiao (1982) and Arellano and Bond (1991) have suggested the use of second and higher-order lagged dependent variables as instruments. To avoid overidenti…cation, we use the second-order lagged dependent variable as instrument, as in Anderson and Hsiao (1982).
To deal with outliers, we winsorize our data as follows. 12 Cash ‡ow to assets is winsorized at the 1% level. All leverage measures are winsorized between 0 and 1. The top tails of Tobin’s q, capital expenditures to assets, depreciation to assets and industry cash ‡ow risk are winsorized at the 1% level. After this, our free cash ‡ow measures and net working capital are computed.
4.2 Speci…cation issues
Consistent estimation of the coe¢ cients of equation (1) requires us to take some measures. In this subsection we give a motivation on theoretical and empirical grounds for our esimator: the
…rst-di¤erences IV estimator for …xed e¤ects. This estimator is consistent under the presence of endogenous variables and weakly exogenous instruments. 13 We present the results of alternative (but inconsistent) estimators in a table.
With our estimator choice we account more seriously for consistent estimation than other empirical studies. For example, Kim et al. (1998), Opler et al. (1999), Pinkowitz and Williamson (2001) and Bates et al. (2009) do not include instrumental variables for endogenous regressors, such as debt to assets, capital expenditures or net working capital. 14 Also, if these studies present
…xed e¤ects regressions (e.g. in Opler et al. (1999) or Bates et al. (2009)), they use the within- estimator for …rm …xed e¤ects, an estimator which is inconsistent under weak exogeneity. 15 We
1 1
With the …rst-di¤erences estimator for …xed e¤ects, the moment condition for the lagged dependent variable is E [("
i;t"
i;t 1) (C
i;t 1C
i;t 2)] = 0, where C stands for cash holdings. Obviously following (1), "
i;t 1is correlated with C
i;t 1, implying that this condition is not satis…ed. Something similar holds for the within estimator for …xed e¤ects.
1 2
Winsorization is a method to deal with outliers without losing data, also employed in Bates et al. (2009).
Winsorizing a variable at the 1% level, implies taking the 0.5 and 99.5 percentiles. All values below the 0.5 percentile take the value of the 0.5 percentile. All values above the 99.5 percentile, take the value of the 99.5 percentile.
1 3
The assumption of weak exogeneity can be written as E ("
i;tjZ
i;s) = 0 for s t, where Z
i;sis a vector of instruments.
1 4
Opler et al. (1999) note the potential endogeneity, but they respond to it by excluding these potentially endogenous variables. We will also present results doing this.
1 5
Bates et al. (2009) present results using the …rst-di¤erences estimator, but they do not take into account
endogeneity of the lagged dependent variable.
will argue below that weak exogeneity is a very reasonable assumption, given the forward-looking nature of cash policies.
We use this subsection to discuss the details of our speci…cation choice, whose results are reported in column (1) of table 1. We discuss respectively, the econometric rationale behind the inclusion of …rm and year …xed e¤ects, the estimation of the …rm …xed e¤ects, and last but not least, endogeneity issues and the instrumental variables. The results in columns (2)-(6) in table 1, each di¤er in one or two respects from our baseline speci…cation (1). Throughout this subsection, we discuss these results along with the divergent speci…cation issues.
[Insert table 1.]
In column (1), we present results from our consistent estimator. We see that cash holdings are positively a¤ected by existing cash holdings and cash ‡ow. Firms’growth opportunities and industry cash ‡ow risk have positive e¤ects, although the latter is statistically insigni…cant at the 10% level. Cash holdings substitute positively for long-term debt and negatively for capital expenditures and net working capital (net of cash). Unexpectedly, we see that size enters the equation positively and signi…cant at the 5% level. We discuss the implications of these results in the next section. For comparison with the other columns, we note that following our assumptions, only the estimation results in columns (1) and (6) are consistent.
In the previous subsection, we already discussed the rationale behind using …rm and year
…xed e¤ects. We also used econometric tests for the inclusion of these …xed e¤ects using the robust Hausman test statistic. The robust Hausman test, tests whether there is a systematic di¤erence in coe¢ cients of two regressions. In appendix C, we can see that this test especially suggests the use of both …rm and year …xed e¤ects, but there is less evidence for the inclusion of year …xed e¤ects. In table 1, we present results with …rm and year …xed e¤ects (1), no …xed e¤ects (2) and only …rm …xed e¤ects (3). Comparing (2) with (1) and (3), we see that the coe¢ cient estimates are highly di¤erent when excluding …rm …xed e¤ects, as already suggested by the robust Hausman test results. When …rm e¤ects are excluded, the results follow the correlations in the data: bigger …rms hold less cash, and more indebted …rms hold less cash.
The inclusion of year …xed e¤ects in (1) versus (3) alters mainly the sign and signi…cance of industry cash ‡ow risk. One explanation could be that exogenous shocks to all …rms such as macroeconomic shocks a¤ect cash holdings negatively, while a¤ecting industry cash ‡ow volatility positively, giving a negative relation when we do not control for year e¤ects. We research this issue more thoroughly in section 6.
To estimate our liquidity demand equation consistently with …xed e¤ects, we can choose between two estimators. These are the within or mean-di¤erenced estimator and the …rst- di¤erences estimator. 16 The estimator that is most often used in the literature, is the within
1 6
The …rst-di¤erences estimator estimates …xed e¤ects by subtracting the …rst-order lag of (1), and estimating
estimator. However, this estimator is only consistent under strong exogeneity of the regressors.
The …rst-di¤erences estimator, however, is consistent under a much weaker assumption: weak exogeneity of the regressors. Especially with forward-looking motives to hold cash, such as the precautionary motive, it is highly likely that unexpected time-t shocks to cash holdings " i;t , in‡uence cash policies at time-t + 1, making strong exogeneity an invalid assumption and within estimation for …xed e¤ects inconsistent (see Cameron and Trivedi (2005), p.758). In appendix D, we discuss robust Hausman test results, suggesting that the instruments are weakly exogenous.
In table 1, we can also see that the results from within estimation in (4) di¤ers from …rst di¤erence estimation in (1). These results suggest that the within estimator for …xed e¤ects is inconsistent. This point is completely ignored in the literature.
As mentioned, …rm policies to invest, hold cash, save or borrow are determined simultaneouly.
Also noted above, using …xed …rm e¤ects in a dynamic panel model, implies that the lagged dependent value is endogenous. Therefore, we test and control for endogeneity of the following four variables: long-term debt, capital expenditures, non-cash net working capital and lagged cash holdings. With our …rst-di¤erence estimator for …xed e¤ects, the choice of instruments is relatively easy. We use lagged values of these variables as instruments, due to their natural correlation with the …rst-di¤erences.
Using these instruments, we tested for endogeneity using auxilary regressions following Haus- man (1978) and the robust Hausman test. Comparing column (1) with (5) in table 1, we see that the use of instruments alters the size of the coe¢ cients. In appendix E.1, we give our detailed endogeneity test results. In appendix E.2, we present tests showing that the instruments are not
‘weak’.
The choice of our set of instruments makes our equation exactly identi…ed. In table 1, col- umn (6), we report the results of the same instrument set augmented with third-order lagged dependent variable as instrument. This model is not overidenti…ed, and we see that these results are comparable to the results of the exactly identi…ed equation in column (1). 17 This suggests that the results of the exactly identi…ed equation are consistent. In appendix E.3, we show that increasing the amount of instruments by additional lags or using the instrument set by Arellano and Bond (1991), quickly leads to overidenti…cation. Another reason to choose for this parsimonious instrument set, is that we are able to include 1991 as a year in the regression.
In the next three sections we present our regression results, using the described speci…cation in this section. In section 5 we look at the …rm-speci…c determinants of corporate cash holdings.
In section 6, we look at the causes of the decline in corporate liquidity in Japan.
the coe¢ cients. The within estimator estimates …xed e¤ects by subtracting the averaged LHS and RHS of (1) from (1). This estimator is equivalent in terms of coe¢ cients to the least square dummy variable estimator, which is an impractical estimator in our case with around 1,200 cross-sections.
1 7
As with the …rst-di¤erences model, we expect that the result of within estimation in (4) is not overidenti…ed.
Using one more lagged dependent variable as instrument leads to a Hansen’s J statistic of 1:765 with a p-value
of 0:1840, suggesting that the fully identi…ed within model is not overidenti…ed.
5 The determinants of corporate liquidity during the Japanese lost decade
We use our consistent estimator to estimate the liquidity demand equation (1) for Japan between 1991 and 2003. First, we critically asses our liquidity demand equation, based on the included dependent and explanatory variables. In the second subsection, we build on our main model in equation (1) and take a deeper look at some of the main results. We use alternative explanatory variables and look at di¤erent subsets of …rms.
5.1 Main results
In the …rst 7 columns of table 2, we present alternative ways to estimate the liquidity demand equation (1). We present these results to deal with the following disadvantages of this equation.
First, the test equation as speci…ed in (1) has some similarities with an accounting identity.
This could lead to an automatic …t. Therefore we present results where we exclude long-term debt and net working capital. Second, the …rst-di¤erences of long-term debt and net working capital have a relative high correlation of 0:50 (see appendix B). We replace these two by total debt and current assets (net of cash), which a correlation in …rst-di¤erences of 0:06. Third, most explanatory variables are divided by total assets at bookvalue. This could lead to an automatic correlation with the dependent variable, cash to assets. We present results using a di¤erent dependent variable, cash to sales. Fourth, cash to assets can be seen as a limited dependent variable, since it can only take on values between 0 and 1. We also present results, where we use the logarithm of cash to net assets as dependent variable, which is more normally distributed.
If …rms with low cash holdings are indeed more cautious in using their liquidity, as hypothesized in the previous section, we expect that this e¤ect is not picked up in the log-model. If there still is a signi…cant e¤ect of lagged cash holdings, it could be due to transitory cash holdings.
[Insert table 2.]
In columns (1)-(3) we present results from three speci…cations with cash to assets as de- pendent variable. In columns (4)-(5) and (6)-(7), we have respectively cash to sales and the logarithm of cash to net assets, as the dependent variable. In regressions (2), (5) and (7) we excluded long-term debt and net working capital. In (3) we replace long-term debt and net working capital with total debt and non-cash current assets.
What can we conclude from the results in columns (1) to (7) in table 2, with respect to the
four identi…ed speci…cation problems above? Replacing the dependent variable, by either cash
to sales in column (4) or the logarithm of cash to net assets in (6), the results do not di¤er
much. With the log of cash to net assets, lagged cash holdings turn insigni…cant with a very
low t-statistic of 0:254, suggesting that cash holdings in column (1) are only signi…cant since they control for low cash holdings. 18 In the case of cash to sales, we see that the t-statistics are generally lower.
The results are a¤ected more drastically when we remove long-term debt and net working capital from our regression. In columns (2), (5) and (7), we see that excluding long-term debt and net working capital decreases the e¤ect of cash ‡ow and increases the e¤ect of cash holdings and capital expenditures. This is quite logical, given, for example, that the inclusion of net working capital, controls for other destinations of previous period cash holdings or cash ‡ow.
Of the other three explanatory variables, we see that the e¤ect of size is most robust across the di¤erent regressions. Interestingly, the e¤ect is positive, rejecting the existence of scale economies in cash management. Below we will show that this positive e¤ect can be contributed to shrinking
…rms.
Also, excluding long term debt and net working capital, leads to an insigni…cant e¤ect of the market-to-book ratio. It could be that these variables are important control variables for the e¤ect of the market-to-book ratio. For example, when growth opportunities arise, …rms might reduce long-term debt or invest more in net working capital. On the other hand, excluding long- term debt and net working capital, increases the coe¢ cient of industry cash ‡ow risk giving it statistical signi…cance. We do not have a direct explanation for this result. Below we discuss the e¤ect of industry cash ‡ow risk in more detail.
The inclusion of long-term debt and net working capital could also a¤ect the estimates due to a multicollinearity problem, as described above. In (3), we replace these explanatory variables by total debt to assets and current assets (net of cash). We …nd similar results, showing that multicollinearity is not the main problem. We also …nd that total debt to assets has a much lower t-statistic and current assets has a much higher t-statistic. This shows the important role of non-cash current assets as a substitute for cash.
The results in columns (1)-(7) reveal that the inclusion of long-term debt and net working capital seem to have a large e¤ect on our estimates. As we expect that these variables are very important control variables, we include these in our model. 19
5.2 Other results
We take a deeper look at the liquidity demand equation in (1) in this subsection. We replace explanatory variables by alternative measures, add explanatory variables and look at di¤erent subsets of …rms. We report four of these regressions in columns (8)-(11) in table 2. We also discuss some unreported results.
As we see in most columns, the e¤ect of size on cash holdings seems to be positive. This
1 8
Excluding the lagged value, does change the results however: the market-to-book ratio becomes insigni…cant.
1 9
We also tested a range of other measures which do include some measure of net working capital or long-term
debt, but remove the similarity of equation (2) with an accounting identity. An example is dividing these control
variables by net assets (assets minus cash). These methods reveal similar information, as in table 2.
seems to reject our hypothesis of scale economies in cash management. A disadvantage of our size measure as the logarithm of sales, is that it could measure the e¤ect of year t sales on end of the year t cash holdings, something likely to be positive. An alternative size measure is the logarithm of assets, which is most common in the literature, see e.g. Opler et al. (1999) and Bates et al. (2009). However, the e¤ect is likely to be negative, without any theoretically founded hypothesis. Opler et al. (1999) and Bates et al. (2009) …nd a negative e¤ect, and contribute this to the economies to scale in cash management as found in Mulligan (1997). In unreported results, we use the logarithm of assets as our size measure, and we indeed …nd a negative e¤ect.
The same occurs, when we use cash-to-sales as dependent variable and the logarithm of sales as our size measure.
We found that the positive size e¤ect is likely to be contributed to a small fraction of …rms, which we label ‘cash poor’…rms. We identify ‘cash poor’…rms as …rms with a negative average cash ‡ow between 1991 and 2003. Surprisingly, in the unbalanced sample these are still 176
…rms with 2,145 …rm-year observations (in the balanced sample 154 …rms). 20 We expect these cash-poor …rms to have other priorities and behave di¤erently from …rms that are ‘cash rich’.
We assign a dummy value 1 to cash-poor …rms and add interaction terms of this dummy with some of the explanatory variables. The estimated coe¢ cient of each of these interaction terms, measures the di¤erence in slope between cash-poor and cash-rich …rms. Results are reported in (6). 21
We see in column (8) of table 2, a t-statistic for the size e¤ect for all …rms of 0:72. The slope however is much higher and signi…cant for cash-poor …rms. The positive size e¤ect is likely to be contributed to cash-poor …rms. Why? Looking at assets at bookvalue of both cash-poor and cash-rich …rms, we …nd that the median cash-poor …rm declined by 35 between 1991 and 2003, while the median cash-rich …rm increased in real asset size by 17 percent. The cash-poor …rms we identi…ed, were shrinking, likely to be a¤ecting cash holdings.
What about the other results in (8)? First, we see that coe¢ cient of lagged cash holdings is about four times higher for cash-poor …rms. Cash-poor …rms have less ‡exibility in changing their cash holdings, and have therefore higher persistence in cash holdings. This adds to the suggested role of a necessary minimum in cash holdings for Japanese …rms.
Also in (8), we …nd a positive e¤ect of industry cash ‡ow risk for all …rms. Cash-poor …rms seem to have a much lower slope - the sum is negative - although the di¤erence is not signi…cant with a p-value of 0:115. Industry cash ‡ow risk could play a di¤erent role for cash-poor …rms.
They could be ‘gambling for resurrection’ and disregard risk, or they simply don’t have the
2 0
This could be related to the often-heard comment that Japanese banks kept lending to technically bankrupt
…rms, so-called "zombie companies" (see The New York Times, October 29, 2002, "They’re Alive! They’re Alive!
Not!; Japan Hesitates to Put an End to Its ‘Zombie’Businesses."). This process is named "evergreening" by Peek and Rosengren (2005).
2 1
Simply adding interaction terms with cash ‡ow leads to high multicollinearity, since some variables hardly
change ‘within-…rms’ throughout the sample period. E.g. the between-…rm standard deviation for the log of
assets is more than 6 times higher than the within-…rm standard deviation.
means to take risk into account.
We also use alternative measures for industry risk, which results are unreported. Instead of using the industry mean of the standard deviation of cash ‡ow over the last …ve years (with at least two years), we use a median measure. We also estimate the e¤ect of a mean and median measure over the last eight years (with a least three cash ‡ows to obtain a standard deviation).
The median measure for cash ‡ow risk over the last …ve years is signi…cant at the 5%-level. The mean and median measure over the last eight years are respectively, signi…cant at the 10%-level and insigni…cant. A …rm-speci…c measure over 5 years is not signi…cant at all, with a t-statistic of 1:26.
For years 1999 to 2004 we have data on R&D expenditures. 22 Firms with high R&D expen- ditures could be risky start-up …rms or …rms in highly innovative industries. If we divide R&D expenditures by sales, we put more emphasis on the …rst as start-up …rms are expected to have relatively low sales. In both cases R&D expenditures proxy for risk. In column (9) of table 2 we add R&D to sales as explanatory variable, which we instrument using one lag of R&D to sales.
We …nd a positive signi…cant e¤ect. In unreported results, we …nd that adding R&D to assets as explanatory variable has an insigni…cant but positive e¤ect with a t-statistic of 1:2.
In (10), we only look at manufacturing …rms. Manufacturing …rms are least likely to be active in regulated industries, and some studies on corporate cash holdings only look at manufacturing
…rms (e.g. Almeida et al. (2004) and Acharya et al. (2007)). Eyeballing the results, we see that in general t-statistics are lower. However, we see a much larger e¤ect of the market-to-book ratio and a signi…cant e¤ect of industry cash ‡ow risk. In unreported results, we also …nd a signi…cant e¤ect of lagged two-year sales growth for manufacturing …rms. Running the regression in (8) for manufacturing …rms, by including interaction terms for cash-poor …rms, the positive size e¤ect disappears for all …rms, and lagged cash holdings are only signi…cant for cash-poor manufacturing
…rms.
In (11) we add dummy variables for public debt, dividend payouts and the number of shares outstanding. The public debt dummy is 1 if a …rm has positive public debt holdings and 0 otherwise. The dividend dummy is 1 if a …rm pays a positive dividend and 0 otherwise. The numbers of shares outstanding dummy is 1, if the number of shares outstanding increased, proxying for equity issues. 23 Since we estimate …xed e¤ects with the …rst-di¤erences estimator, the e¤ect is only measured for …rms that did not always have access to public debt markets (669 out of 1,202 …rms) and …rms that had years with and without dividend payment (352 out of 1,202 …rms).
First of all, the inclusion of these dummy variables does not seem to a¤ect the main results.
2 2
Due to the use of the …rst-di¤erences estimator, we can only estimate the e¤ect for years 2000-2004. Also, by adding an additional year, we perform similar winsorizing procedures as described in section 4, but now for the 1991-2004 sample.
2 3
For which we do not have data. This dummy is mismeasured to the extent that share splits and share
repurchases took place.
Concerning the speci…c dummies we expect and …nd the following e¤ects. Firms that have access to public debt markets need to retain less internal funds. In (11) we see a negative signi…cant e¤ect as expected. Following Bates et al. (2009), we expect that …rms that are able to pay a dividend are less risky and have better access to capital markets, and hold therefore less internal funds. Due to the use of …rm …xed e¤ects, the e¤ect can only be measure for …rms that switch from dividend paying to non-dividend paying within the sample. We …nd a positive e¤ect.
Removing …rms that stopped paying dividends, the e¤ect becomes insigin…cant. The positive e¤ect is therefore likely to be attributed to …rms that stopped paying dividends. 24 An equity issue, should raise the holdings of cash. In column (11), we see a positive e¤ect of this dummy variable.
Finally and unreported, we also use di¤erent methods to measure the market-to-book ratio.
In unreported results, we use two-year sales-growth as a proxy for growth opportunities. This measure has a correlation with the market-to-book ratio of 0:23. Due to the high correlation is
…rst-di¤erences with the logarithm of size of 0:79, we use the log of assets as our size measure (with a correlation of 0:27). We …nd a positive but insigni…cant e¤ect of sales growth, with a t-statistic of 1:23. Running this regression for only manufacturing …rms, we do …nd a signi…cant e¤ect, with a p-value of 0:02.
The market-to-book ratio is often critized as a poor measure of Tobin’s q. 25 Measurement error in one variable can lead to biased estimates of all coe¢ cients. Erickson and Whited (2000) and Riddick and Whited (2009) have proposed ways to deal with this measurement error, with the use on instrumental variables. As is shown in Cameron and Trivedi (2005), p.909-910, under certain assumptions, we can use the products and cross-products of the dependent variable and the market-to-book ratio as instrumental variables to obtain consistent estimates. 26 In unreported results, we look at this as a robustness check. We see that the market-to-book ratio turns negative and insigni…cant. This measure does not a¤ect the size of the other coe¢ cients but it lowers the t-statistics.
Concluding this section, we …nd the following results. Although industry cash ‡ow risk is not very signi…cant in our benchmark model, it becomes signi…cant when we remove control variables long-term debt and net working capital. Also, for di¤erent speci…cations of industry risk (e.g. a median value), or when we only look at ‘cash-rich’ …rms or manufacturing …rms, we do …nd a
2 4
Also, these …rms are more often cash-poor …rms, than the …rms that went from no dividends to become dividend-payers.
2 5
The market-to-book ratio measures Tobin’s average q, where it should measure marginal Tobin’s q. These two measures are only equal under certain conditions such as constant returns to scale and free competition (see Erickson and Whited (2000) for more details).
2 6
First, we assume that the measurement error is additive and is independently distributed with zero mean and
constant variance. Second, we assume that Tobin’s marginal q has a nonnormal distribution. Both assumptions
are hard to test empirically. For the second assumption we can infer on the mismeasured Tobin’s q, the market-
to-book ratio. We …nd in our balanced sample, that the market-to-book ratio has an average yearly skewness of
3:0 and the average yearly kurtosis of 17:11.
signi…cant e¤ect, including control variables. The market-to-book ratio is only signi…cant when we include long-term debt and net working capital. Size only seems to play a role for small subset of …rms which had an average negative cash ‡ow between 1991 and 2003. Using a speci…cation in logarithms, we …nd that lagged cash holdings turns insigni…cant, reducing the likelihood that
…rms hold transitory cash. The signs of cash ‡ow, long-term debt, capital expenditures and net working capital are as expected, statistically signi…cant and are quite robust to di¤erent speci…cations.
In the next two sections, we employ our regression model to …nd the causes of the decline in the cash holdings of Japanese …rms. We also look at the macro-economic determinants of corporate liquidity. Second, in section 7, by examing the role of …nancial constraints in liquidity demand.
6 What explains the decline in corporate liquidity in Japan during the lost decade?
In this section we examine the decline in corporate liquidity in Japan. We do this by estimating our liquidity demand function for di¤erent time periods. Secondly, we look at the role of time factors in corporate saving and see if there were macroeconomic factors that a¤ected liquidity demand.
6.1 The decline in corporate liquidity during the Japanese lost decade
Our main results are presented in table 3. To increase the stability of the results, we add the year 2004 to the regression. 27 In column (1), we present the same model as presented in column (1) of tables 1 and 2, extended with 2004. We see that the results do not di¤er much, except that industry risk turns signi…cant in this regression.
[Insert table 3.]
In columns (2)-(3) and (4)-(5) we report results for respectively the samples 1991-1997 and 1998-2004. We use a balanced sample in (2)-(5) to avoid sample attrition. 28 We choose to break the sample at the 31th of March in 1998. As we saw in …gure 1, the decline in corporate savings seems to slow down after that point.
Comparing equations (2) and (3), we see that the the size and signi…cance of lagged cash holdings and cash ‡ow changes drastically when we remove long-term debt and net working
2 7
We winsorize our 1991-2004 sample, as we did with our "lost decade"-sample, as described in section 4.1.
2 8
Which in this case, includes 842 …rms with full data coverage between 1991 and 2004. The balanced sample
for 1991-2003 contained 843 …rms.
capital from the equation. Including long-term debt and net working capital seems to take away the high persistence in the change in corporate cash holdings. Looking at the time series of median long-term debt in …gure 1, we see that long-term debt declined a lot between 1993 and 1997. Unreported, median net working capital was relatively stable in this period. We also see that the coe¢ cients of cash ‡ow, long-term debt and net working capital were much higher prior to 1998 than after 1997. This suggests that balance sheet factors played a much bigger role in the years before 1998.
As we saw in the previous section, removing long-term debt and net working capital decreases the sign of the market-to-book ratio. In table 3, we see the same. If we compare the results in columns (2) and (4), we see a very high and signi…cant coe¢ cient of the market-to-book ratio for 1991-1997 and a small and insigni…cant coe¢ cient for 1998-2004. The median market-to- book ratio declined from 1:53 in 1991 to 1:01 in 1998 and stabilized in the period after 1998.
This pattern is de…nitely picked up by our analysis, which shows that the declining growth opportunities were a factor in the decline in corporate liquidity. Size seems to have a negative e¤ect on cash to assets in the years 1991-1997, but a positive e¤ect in the second half of the sample. We do not have an explanation for this result.
We conclude that a growth opportunties, long-term debt and net working capital played a big role in the decline in corporate liquidity, prior to 1998. After 1997, these factors play a smaller role. In the next section, we provide an explanation for what happened after 1997 based on …nancing constraints. Before that, we discuss the role of macroeconomic factors in corporate liquidity.
6.2 Year …xed e¤ects, marcoeconomic factors and corporate liquidity
Following the Hausman test for year …xed e¤ects, reported in appendix C and described in section 4.2, we know that year e¤ects a¤ect the coe¢ cients of the liquidity demand equation systematically. Comparing the coe¢ cients estimates for regressions with and without year e¤ects (see column (1) in table 3, versus column (6) in table 3), we …nd that mainly the coe¢ cient of industry cash ‡ow risk is a¤ected by year e¤ects. Also, in our test equation (1) we …nd signi…cant year e¤ects in 1991, 1992, 2000 and 2004 for an unbalanced panel. These year …xed e¤ects are presented for a balanced and unbalanced sample in the lower right panel of …gure 3. In this subsection, we try to …nd the …rm and macro-factors that could be responsible for these year e¤ects.
First we look at …rm-speci…c factors. The year e¤ects - which is the yearly contribution to the di¤erenced cash ratio - are signi…cantly di¤erent from zero at the 1%-level, in only four years:
1991, 1992, 2000 and 2004 (for the balanced sample also 1999). Especially at the start of the lost decade in 1991 and 1992 and in 1999 and 2000, time factors contribute negatively to liquidity holdings. What happened?
One way to interpret these year e¤ects, is from the perspective of the …rm-speci…c explana-
-2 -1 0 1 2 3 4
91 92 93 94 95 96 97 98 99 00 01 02 03 04 Real GDP per Capita Gro wth (%)
10-to-1-year spread Gov.Bonds (% points)
0 1 2 3 4 5 6
91 92 93 94 95 96 97 98 99 00 01 02 03 04 Treasury Bill Rate (%)
-60 -40 -20 0 20 40 60 80 100 120
91 92 93 94 95 96 97 98 99 00 01 02 03 04 BB-to-A AA-rated Nonfin. Corpo rate Credit Spread (bps) Real Tokyo bank stock index gro wth (%)
-0.04 -0.03 -0.02 -0.01 0 0.01
91 92 93 94 95 96 97 98 99 00 01 02 03 04 Year effects (unbalanced sample) Year effects (balanced sample)
Figure 3. Macroeconomic Data for Japan from 1991 to 2004. See data sources in appendix A. Real GDP per capita growth and the real bank stock index growth measured for calender years. All other variables are averaged over months and measured for …scal years, running from April this year to March next year. Since credit rating data is only available from January 1998 onwards, the spread for 1997 is calculated as the average from January 1998 to March 1998. The year e¤ects are obtained from regression (1) in table 3 with either a unbalanced sample (1,202 …rms) or a sample of 842 …rms which have complete data coverage between 1991 and 2004.
tory variables. Large changes in explanatory variables, could lead to non-linear e¤ects on the dependent variable, implying a signi…cant value for the year …xed e¤ect. We see a negative signi…cant year …xed e¤ect in 1991 and 1992, implying that the model expected the cash ratio to be higher in those years than the actual cash ratio. Looking at time-series of the explanatory variables around 1991 and 1992, we see a relatively large increase in (average) long-term debt to assets and a large decline in the market-to-book ratio. The higher long term debt to assets and lower market-to-book ratio could have led to cash holdings much lower than according to the model.
Looking at the years 1999 and 2000, we see a large drop in net working capital to assets, which is mainly explained by the persistent decline in average current assets (net of cash) from 1996 to 2003. In the years 1999 and 2000, this decline was not matched by a declined in short- term debt, leading to a sudden drop in net working capital. Also, these two years real GDP per capita growth was rising after the …rst credit crunch of 1996/1997, 29 suggesting that …rms
2 9
