
 Start

2193
 Prefix

American Economic Association Meetings in Atlanta, the 2001 Annual Conference of the Royal Economic Society in Durham, the Conference on Fixed Investment at the City University in London,
the 49th International Atlantic Conference in Munich and on seminars at the University of Mannheim and the Deutsche Bundesbank are gratefully acknowledged.
1
 Exact

KNIGHT (1921)
 Suffix

argues that the absence of objective probabilities is constitutional for genuinely
economic decisions, as opposed to situations of statistical "risk". The latter, he maintains, can be
diversified away by pooling ("homogeneous grouping"), whereas the readiness to assume the former type of uncertainty on the part of the entrepreneurs is the true foundation for
 (check this in PDF content)

 Start

2669
 Prefix

The latter, he maintains, can be
diversified away by pooling ("homogeneous grouping"), whereas the readiness to assume the former type of uncertainty on the part of the entrepreneurs is the true foundation for the existence
of profits in a market economy; see
 Exact

KNIGHT (1921),
 Suffix

Chapters VIII and IX. It is not clear, however, whether KNIGHT himself really believed in "Knightian uncertainty" as defined above and
elsewhere  the use of subjective probabilities seems perfectly compatible with his account.
 (check this in PDF content)

 Start

4860
 Prefix

The modern literature distinguishes four
main channels by which uncertainty might affect firms' investment outlays: risk aversion,
financial constraints, irreversibility and convexity of marginal returns.
The oldest and most intuitive account focuses on firms' attitude towards risk; see eg
 Exact

HARTMAN (1976),
 Suffix

the textbook exposition by NICKELL (1978) or, most recently, Bo
(2001). Riskaverse owners and their managers will systematically trade expected returns for certainty. The standard capital asset pricing model (CAPM) shows how this
risk aversion is translated into the equilibrium framework.
 (check this in PDF content)

 Start

4944
 Prefix

distinguishes four
main channels by which uncertainty might affect firms' investment outlays: risk aversion,
financial constraints, irreversibility and convexity of marginal returns.
The oldest and most intuitive account focuses on firms' attitude towards risk; see eg
HARTMAN (1976), the textbook exposition by NICKELL (1978) or, most recently,
 Exact

Bo (2001).
 Suffix

Riskaverse owners and their managers will systematically trade expected returns for certainty. The standard capital asset pricing model (CAPM) shows how this
risk aversion is translated into the equilibrium framework.
 (check this in PDF content)

 Start

6751
 Prefix

An exception is the study of NISHIMURA and OZAKI (2002), comparing the effect of risk and
Knightian uncertainty within the framework of the same model.
3. SCOTT (1976), p. 13.
4. For a discussion of decisions under uncertainty with lexicographic preferences, see
 Exact

SINN (1983),
 Suffix

p. 59.
THE ROLE OF UNCERTAINTY FOR CORPORATE INVESTMENT DECISIONS 175
try might be graver  and the resulting constraints severer  if the prospects of the firm
look more uncertain from the outside.
 (check this in PDF content)

 Start

7468
 Prefix

GHOSAL and LOUNGANI (2000) argue that the impact of uncertainty on investment might differ across firms, depending on their access to
the capital markets. Conversely, STERKEN, LENSINK and
 Exact

Bo (2002)
 Suffix

expect the sensitivity
of investment with respect to internal finance to depend on the degree of uncertainty.
But even financially unconstrained investors who maximise the expected value of
their companies given an exogenous discount rate will not be indifferent towards risk.
 (check this in PDF content)

 Start

10197
 Prefix

CHIRINKO and
SCHALLER (2002) attempt to test for the existence of an irreversibility premium using a
structural model.
5. See, for example, CHIRINKO, FAZZARI and MEYER (1999). However, HARHOFF and RAMB (2001),
as well as VON
 Exact

KALCKREUTH (2001)
 Suffix

find a rather solid interest rate channel in Germany using the
Bundesbank data set. For French firms, CHÂTELAIN and TIOMO (2001) find rather small and fragile user cost effects, whereas GAIOTTI and GENERALE (2001) pictures Italian firms as somewhere
in between.
 (check this in PDF content)

 Start

10784
 Prefix

For an overview, see CHÂTELAIN, GENERALE, HERNANDO, VON KALCKREUTH and
VERMEULEN (2003).
176 ULF VON KALCKREUTH
A different conclusion is reached by the literature stressing the convexity of the marginal product of capital, as in ABEL (1983) and
 Exact

HARTMAN (1972, 1976).
 Suffix

If variable factors, such as labour, energy or raw materials can be optimally adjusted after demand uncertainty is resolved, marginal returns of capital are not linear in product prices any
more  the functional relationship will be convex, and Jensen's inequality makes expected profits an increasing function of risk, given optimal adjustment.
 (check this in PDF content)

 Start

12672
 Prefix

In general, uncertainty can
act as a deterrent from investment, be neutral or even create new incentives; see DIXIT
and PINDYCK (1994, Chs. 6 and 11), DARBY, HUGHESHALLET, IRELAND and PISCI
 Exact

TELLI (1999),
 Suffix

or BÖHM, FUNKE and SIEGFRIED (2001).
Empirically, it is not easy to test an isolated hypothesis on the effect of uncertainty on
investment expenditure. In general, for a given firm or sector, several mechanisms will
be at work simultaneously.
 (check this in PDF content)

 Start

14969
 Prefix

low adjustment rates.6
There is a relatively recent but growing literature that investigates the significance of
uncertainty by means of firm level data, of a very diverse nature. LEAHY and WHITED
(1996), MINTON and SCHRAND (1999), and DRIVER, YIP and DAKHIL (1996) use data on
US firms. Guiso and PARIGI (1999) work on a panel of Italian companies,
 Exact

PATILLO (1998)
 Suffix

utilises a panel of Ghanaian firms and Bo (2002) as well as STERKEN, LENSINK
and Bo (2002) work on panels of listed Dutch firms. PEETERS (2001) investigates and
compares panels of Belgian and Spanish firms, and BLOOM, BOND and VAN REENEN
(2001) use data on companies quoted on the UK stock market.
 (check this in PDF content)

 Start

15029
 Prefix

LEAHY and WHITED
(1996), MINTON and SCHRAND (1999), and DRIVER, YIP and DAKHIL (1996) use data on
US firms. Guiso and PARIGI (1999) work on a panel of Italian companies, PATILLO (1998) utilises a panel of Ghanaian firms and
 Exact

Bo (2002)
 Suffix

as well as STERKEN, LENSINK
and Bo (2002) work on panels of listed Dutch firms. PEETERS (2001) investigates and
compares panels of Belgian and Spanish firms, and BLOOM, BOND and VAN REENEN
(2001) use data on companies quoted on the UK stock market.
 (check this in PDF content)

 Start

15073
 Prefix

LEAHY and WHITED
(1996), MINTON and SCHRAND (1999), and DRIVER, YIP and DAKHIL (1996) use data on
US firms. Guiso and PARIGI (1999) work on a panel of Italian companies, PATILLO (1998) utilises a panel of Ghanaian firms and Bo (2002) as well as STERKEN, LENSINK
and
 Exact

Bo (2002)
 Suffix

work on panels of listed Dutch firms. PEETERS (2001) investigates and
compares panels of Belgian and Spanish firms, and BLOOM, BOND and VAN REENEN
(2001) use data on companies quoted on the UK stock market.
 (check this in PDF content)

 Start

15127
 Prefix

Guiso and PARIGI (1999) work on a panel of Italian companies, PATILLO (1998) utilises a panel of Ghanaian firms and Bo (2002) as well as STERKEN, LENSINK
and Bo (2002) work on panels of listed Dutch firms.
 Exact

PEETERS (2001)
 Suffix

investigates and
compares panels of Belgian and Spanish firms, and BLOOM, BOND and VAN REENEN
(2001) use data on companies quoted on the UK stock market. BUTZEN, FUSS and VERMEIDEN (2002) work on a large panel of Belgian firms.
 (check this in PDF content)

 Start

15925
 Prefix

Using a measure of market turbulence, DRIVER, YIP and DAKHIL (1996) find that uncertainty on its own has a significant negative effect in only one out of twelve industries,
which could be accounted for by a Type I error. Interacting uncertainty with a proxy for
vertical market integration results in significant negative effects in four industries.
 Exact

PEETERS (2001)
 Suffix

finds negative effects of price uncertainty, but no effects of sales uncertainty. BLOOM, BOND and VAN REENEN (2001) find no longrun effect on the capital,
but uncertainty will slow down the reaction of firms to sales shocks.
 (check this in PDF content)

 Start

16618
 Prefix

firms, by BÖHM, FUNKE and SIEGFRIED (2001), identifies a positive uncertaintyinvestment link in a sample of 70 large listed German corporations, which turns
negative for firms in very concentrated industries.7 A positive effect is also obtained by
LENSINK and STERKEN (2000) in their study on Czech firms. See CARRUTH, DICKERSON
and HENLEY (2000) and
 Exact

Bo (2001)
 Suffix

for detailed recent surveys of the empirical literature.
The Bundesbank's financial statement database {Unternehmensbilanzstatistik, UBS)
constitutes the largest source of accounting data for nonfinancial firms in Germany.8
About 70'000 annual accounts were collected per year on a strictly confidential basis by
6.
 (check this in PDF content)

 Start

17689
 Prefix

Nevertheless, aggregation will be much
less of a problem using firmlevel data, and our data set contains many very small firms.
7. Using aggregate data and time series methodology,
 Exact

MAILAND (1998) and WERNER (2001)
 Suffix

find negative effects of uncertainty on investment in Germany.
8. This discussion draws on DEUTSCHE BUNDESBANK (1998), FRIDERICHS and SAUVÉ (1999), and
STÖSS (2001), which contain more detailed descriptions of the UBS data.
 (check this in PDF content)

 Start

17836
 Prefix

Using aggregate data and time series methodology, MAILAND (1998) and WERNER (2001) find negative effects of uncertainty on investment in Germany.
8. This discussion draws on
 Exact

DEUTSCHE BUNDESBANK (1998),
 Suffix

FRIDERICHS and SAUVÉ (1999), and
STÖSS (2001), which contain more detailed descriptions of the UBS data. With respect to investment demand, the UBS has been utilized by HARHOFF and RAMB (2001) in a user cost study, by
VON KALCKREUTH (2001) in a study on the monetary transmission process, by CHIRINKO and VON
KALCKREUTH (2002) on financial constraints, by B
 (check this in PDF content)

 Start

18117
 Prefix

This discussion draws on DEUTSCHE BUNDESBANK (1998), FRIDERICHS and SAUVÉ (1999), and
STÖSS (2001), which contain more detailed descriptions of the UBS data. With respect to investment demand, the UBS has been utilized by HARHOFF and RAMB (2001) in a user cost study, by
VON
 Exact

KALCKREUTH (2001)
 Suffix

in a study on the monetary transmission process, by CHIRINKO and VON
KALCKREUTH (2002) on financial constraints, by BEHR and BELLGARDT (2002) on a dynamic Q
equation and by BREITUNG, CHIRINKO and VON KALCKREUTH (2003) in a panel study on the interaction of investment and finance.
178 ULF VON KALCKREUTH
the Bundesbank's branch offices, in the context of t
 (check this in PDF content)

 Start

18214
 Prefix

With respect to investment demand, the UBS has been utilized by HARHOFF and RAMB (2001) in a user cost study, by
VON KALCKREUTH (2001) in a study on the monetary transmission process, by CHIRINKO and VON
 Exact

KALCKREUTH (2002)
 Suffix

on financial constraints, by BEHR and BELLGARDT (2002) on a dynamic Q
equation and by BREITUNG, CHIRINKO and VON KALCKREUTH (2003) in a panel study on the interaction of investment and finance.
178 ULF VON KALCKREUTH
the Bundesbank's branch offices, in the context of the rediscount and lending operations.
 (check this in PDF content)

 Start

18357
 Prefix

With respect to investment demand, the UBS has been utilized by HARHOFF and RAMB (2001) in a user cost study, by
VON KALCKREUTH (2001) in a study on the monetary transmission process, by CHIRINKO and VON
KALCKREUTH (2002) on financial constraints, by BEHR and BELLGARDT (2002) on a dynamic Q
equation and by BREITUNG, CHIRINKO and VON
 Exact

KALCKREUTH (2003)
 Suffix

in a panel study on the interaction of investment and finance.
178 ULF VON KALCKREUTH
the Bundesbank's branch offices, in the context of the rediscount and lending operations.
 (check this in PDF content)

 Start

22591
 Prefix

As with all surveys, one has to
ensure that the questionnaire is answered in the first place, that it is answered by the
right person and that it is answered correctly. Guiso and PARIGI (1999) exploit data on
the subjective probability distribution of investors contained in a survey conducted by
the Banca dTtalia, and
 Exact

PATILLO (1999)
 Suffix

uses a similar data set for 200 entrepreneurs in
Ghana constructed with administrative help from the World Bank and the UK government. A cheaper alternative is to make use of regular industry survey data.
 (check this in PDF content)

 Start

23556
 Prefix

A second approach is to rely on highfrequency financial market data and to use volatilities, either of commodity prices or exchange rates, or else of stock prices. The first line
of research, exemplified in the paper of DARBY, HUGHESHALLET, IRELAND, and PISCI
 Exact

TELLi (1999),
 Suffix

directly quantifies the degree of uncertainty with respect to some crucial
economic variables; however, it cannot differentiate between firms. The use of stockmarket data, as in BLOOM, BOND and VAN REENEN (2001), or BÖHM, FUNKE and SIEGFRIED (2001), assumes a strong form of market efficiency and implicitly equates firms'
information on future profits to the
 (check this in PDF content)

 Start

24908
 Prefix

Finally, one can generate uncertainty indicators from annual or quarterly financial
statements of individual firms, measuring the volatility of operating profits, cash flow
and other variables. This is the route that will be taken here. GHOSAL and LOUNGANI
(1996, 2000), MINTON and SCHRAND (1999),
 Exact

PEETERS (2001), Bo (2002) and
 Suffix

BUTZEN,
Fuss and VERMEULEN (2002) proceed in the same fashion. Balance sheet or income
statement data naturally yield firmspecific indicators and thus can exploit the individual
variability of a large panel data set.
180 ULF VON KALCKREUTH
3.2.
 (check this in PDF content)

 Start

27656
 Prefix

This equation describes a random walk with a firmspecific drift; the
shift parameter a, will indicate the long run deviation of firm i's growth rate from industry average. This equation (1) can be regarded as a version of the famous law of proportional effect, or G IB RAT'S law.
 Exact

GIBRAT (1931)
 Suffix

explains the approximately lognormal size
distribution of income, firm sizes and other economic variables by positing a stochastic
process in which growth rates develop as a series of independent shocks, unrelated to
the level (size).
 (check this in PDF content)

 Start

28232
 Prefix

Thus the absolute increment to a firm's size in each period will be proportional to the current size of the firm. Since its inception in the early 1930s, GIBRAT'S law
has received a lot of attention among industrial economists; see the survey article by
 Exact

SUTTON (1997).
 Suffix

More recent studies however, most notably the work by EVANS
(1987a,b) and HALL (1987), suggest that the law does not hold. As one of two stylised
10. Of course, sales uncertainty is not the only type of uncertainty relevant for a firm.
 (check this in PDF content)

 Start

28345
 Prefix

Since its inception in the early 1930s, GIBRAT'S law
has received a lot of attention among industrial economists; see the survey article by
SUTTON (1997). More recent studies however, most notably the work by EVANS
(1987a,b) and
 Exact

HALL (1987),
 Suffix

suggest that the law does not hold. As one of two stylised
10. Of course, sales uncertainty is not the only type of uncertainty relevant for a firm. A prior version
of this paper, VON KALCKREUTH (2000), also considers cost uncertainty.
 (check this in PDF content)

 Start

28574
 Prefix

More recent studies however, most notably the work by EVANS
(1987a,b) and HALL (1987), suggest that the law does not hold. As one of two stylised
10. Of course, sales uncertainty is not the only type of uncertainty relevant for a firm. A prior version
of this paper, VON
 Exact

KALCKREUTH (2000),
 Suffix

also considers cost uncertainty. The cost uncertainty indicators were discarded as the results lacked robustness with respect to GMM estimation.
THE ROLE OF UNCERTAINTY FOR CORPORATE INVESTMENT DECISIONS 181
facts for the relationship between size and growth, SUTTON (1997, p. 46) proposes instead:
(a) The probability of survival increases with f
 (check this in PDF content)

 Start

28897
 Prefix

The cost uncertainty indicators were discarded as the results lacked robustness with respect to GMM estimation.
THE ROLE OF UNCERTAINTY FOR CORPORATE INVESTMENT DECISIONS 181
facts for the relationship between size and growth,
 Exact

SUTTON (1997,
 Suffix

p. 46) proposes instead:
(a) The probability of survival increases with firm (or plant) size.
(b) The proportional rate of growth of a firm (or plant) conditional of survival is decreasing in size.
 (check this in PDF content)

 Start

29256
 Prefix

between size and growth, SUTTON (1997, p. 46) proposes instead:
(a) The probability of survival increases with firm (or plant) size.
(b) The proportional rate of growth of a firm (or plant) conditional of survival is decreasing in size.
For our data set, GIBRAT'S law is clearly rejected by two tests on unit roots for panel data
proposed by
 Exact

BREITUNG (1997);
 Suffix

see Appendix B for a detailed account. Both tests are
based on estimations using the GMM procedure for dynamic panel data models discussed in Section 5. The first is a ttest; the second employs the SarganHansen Jstatistic
as a comprehensive specification test.
 (check this in PDF content)

 Start

36368
 Prefix

The model platform corresponds to that used recently by, among others, BOND, EL
STON, MAIRESSE and MULKAY (2003), MAIRESSE, HALL and MULKAY (1999) CHIRINKO
FAZZARI and MEYER (1999), HARHOFF and RAMB (2001), as well as BLOOM, BOND anc
VAN REENEN (2001). CHIRINKO and VON
 Exact

KALCKREUTH (2002)
 Suffix

present a detailed discussion. The point of departure is a static neoclassical equation for capital demand. Using a
generalised CES production function, EISNER and NADIRI (1968) derive the following
linear equation from the firstorder conditions of profit maximisation:
log Ku = 9 log Su  a log UCC:t + log h]A, (6)
 (check this in PDF content)

 Start

40968
 Prefix

estimation
Equation (8) corresponds to the common twoway error component regression model,13
which can be written in shorthand as
yu = ß'xi.t + £/./, with Çu = Vi + vu (10)
12. On the interpretation of a cash flow variable in the context of a neoclassical investment function
see CHIRINKO and VON
 Exact

KALCKREUTH (2002),
 Suffix

Appendix A.
13. See the textbook expositions by BALTAGI (2001), Ch. 2 and 3, or HSIAO (2003), Ch 3.
186 ULF VON KALCKREUTH
where yiìt is the endogenous variable for firm i, x,., the vector of explanatory variables,
including time dummies, and ß the vector of coefficients including the time effects A,.
 (check this in PDF content)

 Start

41081
 Prefix

On the interpretation of a cash flow variable in the context of a neoclassical investment function
see CHIRINKO and VON KALCKREUTH (2002), Appendix A.
13. See the textbook expositions by BALTAGI (2001), Ch. 2 and 3, or
 Exact

HSIAO (2003),
 Suffix

Ch 3.
186 ULF VON KALCKREUTH
where yiìt is the endogenous variable for firm i, x,., the vector of explanatory variables,
including time dummies, and ß the vector of coefficients including the time effects A,.
 (check this in PDF content)

 Start

46738
 Prefix

with &,, = <# + vu (14)
Using the standard mean difference transformation (13) will introduce correlation between the (transformed) error terms and the (transformed) lagged dependent variable.
One solution to both problems is to use some sort of instrumental variable (IV) estimation. ANDERSON and
 Exact

HSIAO (1981)
 Suffix

propose transforming the regression equation by
first differencing in order to get rid of the individual specific effects. In the case of equation (14), this leads to
Ayu = aAyui + ß'Axu + Avu (15)
Now a set of instruments has to be chosen.
 (check this in PDF content)

 Start

62977
 Prefix

We report the second step results on the
basis of an asymptotically efficient weighting matrix for the orthogonality conditions.
The standard deviations are corrected for the small sample bias reported in ARELLANO
and BOND (1991); see
 Exact

WINDMEIJER (2000).
 Suffix

The SarganHansen statistics are insignificant both with and without uncertainty, and so is the Lagrange Multiplier statistic that
tests for autocorrelation of second order in the residuals of the transformed equation.
 (check this in PDF content)

 Start

63410
 Prefix

statistics are insignificant both with and without uncertainty, and so is the Lagrange Multiplier statistic that
tests for autocorrelation of second order in the residuals of the transformed equation.
For the ADL model, the long run elasticity of capital with respect to real sales is calculated as 7/A.5 = B(l)/A(l); see CHIRINKO and VON
 Exact

KALCKREUTH (2002),
 Suffix

Appendix A.
For the models excluding and including the uncertainty variables, this elasticity is calculated as 0.40 and 0.41, respectively. The measured uncertainty coefficient in column (2)
is insignificant and close to zero.
 (check this in PDF content)

 Start

67757
 Prefix

, the instruments are all available lags of length 1 and higher for the first differences of 7^)/A',,/i, A log 5',,/,
CFu/I<ui, apart from a constant and time dummies; additionally all available lags of length 0 and
higher of U(3) for the estimation in column (4). Windmeijercorrected, robust standard errors are in
parentheses; see
 Exact

WINDMEIJER (2000).
 Suffix

SH is the pvalue for the SarganHansen statistic testing overidentifying restrictions. LM is the pvalue for the Lagrange Multiplier statistic testing for qthorder autocorrelation of the error terms: columns (1) and (2), q = 2; columns (3) and (4), q = 1.
 (check this in PDF content)

 Start

73941
 Prefix

The instruments, apart from a constant and time dummies, are lags of length 1 and higher for
the first differences of IU)/KU\, A log Sut, CF^/A^i, and all available lags of length 0 and higher for
the relevant uncertainty variable. Windmeijercorrected, robust standard errors are in parentheses, see
 Exact

WINDMEIJER (2000).
 Suffix

SH is the pvalue for the SarganHansen statistic testing overidentifying restrictions. LM(1) is the pvalue for the Lagrange Multiplier statistic testing first order autocorrelation of
the error terms.
 (check this in PDF content)

 Start

82460
 Prefix

As is well known, in the time series context such a test is compii200 ULF VON KALCKREUTH
cated by the fact that under the unitroot hypothesis, the OLS estimates of the autoregressive coefficients have a nonstandard limiting distribution for T —> oo, such that the
critical values for the normal distribution cannot be used.20
 Exact

BREITUNG (1997)
 Suffix

shows that
this feature is absent in the panel data context, as long as the limiting distributions for
fixed T and N —+ oo are considered. Thus, in a panel with many individuals, the standard
inference procedures will be applicable irrespective of whether the coefficients are on
the unit circle or within.
 (check this in PDF content)

 Start

82815
 Prefix

distribution cannot be used.20 BREITUNG (1997) shows that
this feature is absent in the panel data context, as long as the limiting distributions for
fixed T and N —+ oo are considered. Thus, in a panel with many individuals, the standard
inference procedures will be applicable irrespective of whether the coefficients are on
the unit circle or within.
 Exact

BREITUNG (1997)
 Suffix

proposes two different tests for unit roots in panel equations. The
first one is a simple £test on the autoregressive coefficient, estimated using GMM. Subtracting log Sij on both sides, we can write equation (Bl) as:
A log Su = o, + (6  1) log Sui +qt+ eu (B2)
and test the hypothesis 6 = 1 by looki
 (check this in PDF content)

 Start

84416
 Prefix

At the second step, the procedure yielded a coefficient estimate of
0.5367, corresponding to a value of 6 = 0.4633, with a Windmeijercorrected standard
error of 0.01975. The null hypothesis of 6 = 1 can thus be rejected on the basis of a ^statistic of 28.9.
On the basis of a simulation study,
 Exact

BREITUNG (1997)
 Suffix

prefers yet another test on unit
roots. If we assume a secondorder autoregressive process with firmspecific fixed effects,
log Su = dj f 6i log Sui + b2Sjj2 + qt + elA (Bl)
this will include the AR(1) in equation (1) as a special case.
 (check this in PDF content)