Content uploaded by E. Philip Davis
Author content
All content in this area was uploaded by E. Philip Davis
Content may be subject to copyright.
INVESTMENT AND UNCERTAINTY IN THE G7
Joseph P. Byrnea and E. Philip Davisb,c,∗
Strathclyde Universitya
National Institute of Economic and Social Researchb
Brunel Universityc
11 June 2004
Contents
1. Introduction
2. Literature Survey
2.1 Theory of Investment and Effects of Uncertainty
2.2 Empirical work on Investment and Uncertainty
3. Specification and Estimation Issues
3.1 Measures of Uncertainty
3.2 Specification of Investment Functions
3.3 Econometric Approach
4. Data
5. Results
5.1 Estimation of Investment Functions without Uncertainty
5.2 Impact of Uncertainty on Investment
5.3 Sensitivity Analysis
6. Conclusions
∗ Davis (corresponding author): Brunel University, Uxbridge, Middlesex, UB8 3PH and NIESR. e-mail:
e_philip_davis@msn.com. Byrne: Strathclyde University 16 Richmond Street, Glasgow, G1 1XQ , e-mail:
joseph.byrne@strath.ac.uk. The project was undertaken while both authors were at the National Institute of Economic
and Social Research. The authors wish to thank Paul Ashworth, Ray Barrell, Andy Blake, Stephen Hall, DeAnne
Julius, Campbell Leith, Nigel Pain, Simon Price, Martin Weale and other seminar participants at NIESR and
Warwick University for helpful suggestions. This research was financed by ESRC project L138250122, Fluctuations
and Long-Term Prosperity: A study of UK and international economies.
2
1. Introduction
The benefits of macroeconomic stability are increasingly emphasised by policy makers.
This has followed but also underpinned a growing academic interest in economic uncertainty and
its effects on macroeconomic variables. Output growth in both the short and long run is crucially
dependent upon aggregate business investment and the focus of this paper is the relationship
between investment and uncertainty. The two main strands of theoretical literature on this
relationship come to different conclusions. The call option approach to investment determination
(Dixit and Pindyck 1994), implies that increased uncertainty reduces investment given the
irreversibility of investment projects and the consequent option value of delaying expenditures. In
contrast, earlier work (Hartman (1972), Abel (1983)) suggested that increased uncertainty raises
investment where the marginal product of capital is an increasing function of prices and increases
in the variance of prices will increase the expected return. Given this theoretical dispute there is a
need for empirical analysis.
While there is an extensive empirical literature on uncertainty and investment,1 it is
mainly undertaken on the basis of one country or one indicator of uncertainty. Carruth et al.
(2000b) are of the view that the broad consensus is that the relationship is negative and this
consensus emerges from a wide range of models and alternative methods of proxying uncertainty.
On the other hand, Huizinga (1993) suggests that effects vary depending on the source of
uncertainty. Furthermore, differing results for a single indicator such as exchange rate volatility
have been found by authors such as Goldberg (1993) and Darby, Hughes Hallett, Ireland and
Piscittelli (1999) depending on the countries studied, methods of deriving measures of uncertainty
and the data period used. In this paper we argue that there are good reasons for using conditional
volatility measures such as those derived from GARCH estimation as the best measures of
uncertainty in a forward looking sense.
Recently, non-stationary panel econometric methods have become popular in multi-
country macroeconomic studies. These methods, which have both a time series and cross
3
sectional dimension, are a means of increasing the efficiency of parameter estimates when testing
a particular long run hypothesis whilst taking account of potential non-stationarity. While some
differences across countries are to be expected, when adopting this approach it is important to test
for cross sectional heterogeneity to ensure that panel estimates are not biased due to
“unreasonable” pooling of countries, see Pesaran and Smith (1995). Taking account of these
factors, we conduct a multi-country study of investment and a comprehensive range of
macroeconomic indicators of uncertainty using Pesaran, Shin and Smith’s (1999) Pooled Mean
Group Estimation (PMGE), and GARCH methods to derive uncertainty measures. We assess long
term effects on investment of measures of conditional volatility, and short term dynamic effects of
our uncertainty proxies on investment, while paying particular attention to whether long run
country homogeneity is accepted by the data. The indicators used include both financial variables
(equity prices, interest rates and exchange rates), and macroeconomic variables, (inflation and
industrial production).
We come to the conclusion that the source of volatility typically does matter, with
exchange rate volatility being most crucial. Besides the general interest in such work, the effect of
uncertainty on investment is an important aspect of the benefits of EMU membership, since EMU
is likely to remove part of effective exchange rate volatility and, possibly, part of long rate
volatility also. The paper is structured as follows. In Section 2 we provide a brief overview of the
theoretical and empirical literature on uncertainty and investment. In Section 3 we address
specification and estimation issues. In Section 4 we set out the data, Section 5 considers results
whilst Section 6 draws conclusions.
2. Literature Survey
2.1 Theory of Investment and Effects of Uncertainty
According to Dixit and Pindyck (1994), the effect of uncertainty on investment stems from the
option characteristics of an investment project, given the option of delaying the project and its
1 See the excellent review by Carruth, Dickerson and Henley (2000b).
4
irreversibility once begun, together with the uncertainty over future prices that will determine its
profitability. The value of the option stems from the fact that delaying the project may give a
more accurate view of market conditions. The call option implies a difference between the net
present value (NPV) of an investment and its current worth to the investor. To lead to
expenditure, the NPV has to exceed zero so as to cover the option value of waiting. The
expectation is that heightened uncertainty, by leading to delay in projects, would lead to a fall in
aggregate investment. There may also be threshold effects i.e. rates of return below which
investment is not undertaken, depending on investors’ risk aversion. Abel et al. (1996) extended
this theory of irreversibility to show that there could be both a call and put option feature in
investment, in terms of options to expand or contract the capital stock in the future.
The literature is not unanimous in suggesting a negative effect of uncertainty on investment.
Hartman (1972) and Abel (1983) show counter to the above that where there is perfect
competition and constant returns to scale as well as symmetric adjustment costs, an increase in
uncertainty may also raise the value of a marginal unit of capital and hence the incentive to invest.
Lee and Shin (2000) argue that the balance between the positive and negative effects of
uncertainty may depend on the labour share of firms’ costs.
Given these contrasting theoretical results, as well as ambiguity as to where uncertainty
originates at a macro level, empirical work is vital. We now go on to review empirical methods
and studies in terms of sources of uncertainty and their use in investment functions, before
undertaking our own empirical work.
2.2 Empirical work on Investment and Uncertainty
An extensive survey of the empirical literature on investment and uncertainty is provided
in Carruth et al. (2000b). Overall, they suggest there is a broad consensus that the effect from
proxy measures of uncertainty on aggregate investment is negative. This is for a wide range of
models and various methods of proxying uncertainty. Our intention is not to repeat that survey,
5
but rather to provide a thematic overview giving sufficient background for understanding our
results and possible contrasts with earlier studies, as well as to reference more recent work.
First, there are issues in choosing the source of uncertainty where there are often
conflicting theoretical arguments. Taking the example of share price volatility, it is argued in
Carruth et al. (2000b) that use of stock market based measures may reveal cash flow uncertainty
for the firm, but are not relevant indicators of future economic shocks and policy changes.
Moreover, stock prices may be vulnerable to bubbles rather than reflecting fundamentals. Hence
they argue that macroeconomic variables such as price, output and exchange rate volatility are
often preferred instead. One could query this approach since share prices take into account all
information relevant to the future profitability of the firm (or at a macro level the corporate
sector). Furthermore, one can argue that investment is discounted by the long rate plus a risk
premium, where the latter may be linked to equity market volatility (Davis and Madsen, 2001).
Such conflicting views suggest a need to take a comprehensive empirical view of possible
measures and their impact.
As regards measurement of uncertainty, authors that have used ARCH or GARCH
measures of macroeconomic variables when modelling investment include Huizinga (1993) and
Price (1995). Huizinga (1993) considers conditional volatility of US inflation, real wages and real
profits and generally finds a negative effect on investment. Price (1995) utilises the conditional
variance of the growth rate of GDP, and finds a negative effect on UK manufacturing investment
lagged twice. The question of whether there is a different effect from forward and backward
looking measures of uncertainty has been raised by Ferderer (1993). He attempts to produce a
forward looking measure based on the risk premium from the term structure. For the US he finds
that the effect on investment is negative. Driver and Moreton (1991) model uncertainty using the
standard deviation of 12 months forward predictions of output growth and inflation across 12
forecasting teams. They find a negative long-run effect from output growth on investment but no
long-run effect from inflation on investment. Goldberg (1993) and Campa and Goldberg (1995)
6
derived their measure of exchange rate volatility from the standard errors of the residuals from a
moving average representation of the exchange rate using US data.
Looking specifically at exchange rate volatility and investment, Darby et al. (1999) using
a model based on Dixit and Pindyck (1994) suggest that there are situations where exchange rate
uncertainty will depress investment and situations where it will not. In the empirical section of
their paper Darby et al. (1999) find, using a neoclassical model, Tobin’s Q and moving average
exchange rate variance, that uncertainty has a significant and negative impact on investment for
the US, Germany and France. There are additional dynamic effects which are negative for Italy
and the UK. There are negative misalignment effects for US, France, Italy and the UK.
Serven (2003) using GARCH measures of uncertainty, finds a negative and highly
significant impact of real exchange rate uncertainty on private investment in a sample of
developing countries, after controlling for standard investment determinants. The impact is larger
at higher levels of uncertainty – in line with the analytical literature underscoring ‘threshold
effects’. Moreover, the effect on investment of real exchange rate uncertainty is shaped by the
degree of trade openness and financial development: higher openness and weaker financial
systems are associated with a more significantly negative uncertainty- investment link.2
We suggest that there are a number of lacunae in the existing macro literature, that we
shall seek to fill in the following sections. Studies tend to be for a single country, indicator and
measurement method. Unlike many studies cited above, we look both across a range of indicators
and also utilise a panel methodology for the G7 countries, testing for homogeneity in long run
responses across countries. We now go on to describe our approach to estimating the relationship
between investment and uncertainty.
3. Specification and Estimation Issues
3.1 Measures of Uncertainty
2 Some disaggregate studies, including Leahy and Whited (1996), Driver et al. (1996), Guiso and Parigi (1999) and
Temple et al. (2001) also found negative effects of uncertainty on investment in at least some industries. For example
7
As noted above, in empirical applications, uncertainty is typically proxied by measures of
volatility. This section provides a brief discussion of such measures. We conclude the section
with arguments which favour conditional volatility measures such as GARCH, against moving
average standard deviations.
In the GARCH(p,q) model introduced by Bollerslev (1986) we consider the information
set Yt-1, which contains all information on the variable yt until time t-1. Also we assume the time
series yt can be described as
() ()
1,0~,| 2
1
1NIDhYy t
tttt
ηη
=
− (1)
∑∑ =
−
=
−++=
p
i
ti
q
i
tit hyh
1
1
1
2
10
βαα
(2)
Where ht is the conditional variance.3 Given a coefficient on the lagged squared error
α
1 greater
than zero, volatility will tend to cluster, with large residuals following other large residuals, but of
unpredictable sign, while a random, normally-distributed variation in the conditional distribution
(error variance) gives the unconditional distribution (error distribution) fatter tails than the normal
distribution.4
Most of the GARCH studies in the literature, for stock returns, the term structure or
exchange rates, have found a significant degree of both short and long run shock persistence with
high frequency data, thus accounting for the clustering of volatility which is characteristic of such
markets (Bollerslev et al. 1992). Studies of inflation have found similar results (Engel, 1983).
There are numerous variants on the basic GARCH which allow inter alia for a distinction between
short and long run volatility (components GARCH) and asymmetries in effects of volatility
(exponential GARCH). Given our focus here is on testing a range of indicators, we retain
Temple et al. (2001) found that UK industrial survey responses suggested external finance constraints and uncertainty
about demand were factors limiting investment, but that the former depended on concentration in the industry.
3 To ensure a well-defined process, all the parameters in the infinite order AR representation must be non-negative,
where it is assumed that the roots of the polynomial lie outside the unit circle. For a GARCH(1,1), a sufficient lag
length in most applications according to Bollerslev et al. (1992), this amounts to ensuring that both α 1 and β 1 are
non-negative. It follows also that yt is covariance stationary if and only if α 1+ β 1<1.
4 Using the coefficient β1 on the lagged dependent variable and setting the conditional variance constant, GARCH
enables a long run response of the conditional variance to shocks to be calculated. α0/[1 - α1 - β1] is the mean level of
volatility.
8
GARCH (1,1) – in further work on exchange rate volatility (Byrne and Davis 2003a) we set out
complementary results using these alternative approaches. We maintain that they do not
controvert the basic results presented here.
An alternative to the use of GARCH or variants are types of moving average standard
deviations or variances of the series in question. One such measure of volatility based on the
variance was adopted by Kenen and Rodrick (1986) and Darby et al. (1999)
KR =
()
2
1
2
8
1
ln
8
1
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡∆
∑
=
−
i
it
e (3)
Volatility can also be derived using an autoregressive moving average model, the proxy being the
standard deviation of the model’s residuals, as in Goldberg (1993) and subsequent papers:
1111 −−
+
+
=tttt ERER
ε
β
ε
α
(4)
Some authors such as Pindyck and Solimano (1993) also work with the moving average standard
deviation (SD) of the relevant time series.
We argue that the distinction between GARCH (which measures conditional volatility)
and moving average based volatility (a measure of unconditional volatility) is a potentially
important one, and that the shows offers more of the theoretical characteristics of uncertainty than
an unconditional measure would. To motivate this, consider the distinction originally due to
Knight (1921) between risk and uncertainty. Risk can be defined as the danger that a certain
contingency will occur, a measure often related to future events susceptible to being reduced to
objective probabilities, while uncertainty is a term applied to expectations of a future event to
which probability analysis cannot be applied, such as a change in policy regime or a financial
crisis (Shafer, 1986). The response of an uncertain market – and the response of investment
thereto - may appear out of scale with the proximate causes of a given stimulus, if it leads
participants to change the way they form their decisions.
In this context, heightened unconditional volatility alone may merely reflect a greater
incidence of large random and independent shocks, i.e. greater risk, without a change in
underlying perceptions as to the situation on the part of firms considering investment. On the
9
other hand, heightened conditional volatility, besides indicating risk, may also indicate greater
(unmeasurable) uncertainty on the part of the market regarding the direction of the variable and
the intentions of the authorities, including market responses to shocks per se5 (an increased
tendency for shocks to have persistent effects on the market) which may be more likely to affect
investment. More generally, conditional volatility highlights periods of concentrated volatility
which might be expected to maximise uncertainty and hence the option value of waiting to
undertake investment, while the rolling measures could just be capturing background volatility
with occasional outliers that firms learn to live with. A similar point is made by Serven (2003),
who considers use of GARCH essential to measure exchange rate uncertainty as opposed to
“sample variability”.
3.2 Specification of Investment Functions
We follow Bean (1981), Driver and Moreton (1991) and Darby et al. (1999) in estimating
dynamic error correction models of investment including both short and long-run terms in output,
investment, average Q and the real user cost of capital. Consistent with these authors, we test for
long run homogeneity of output as implied by the CES production function. As background we
briefly introduce empirical investment models, including the flexible accelerator model, the
neoclassical model and Tobin’s Q theory of investment (Caballero, 1999, provides a
comprehensive survey of the investment literature). The flexible accelerator model is represented
by the short and long run relationship between investment and output, such that a desired output
capital relationship is maintained. This ignores, however, possible substitution effects between
factor inputs, unlike the neoclassical model of Jorgensen (1963). To take account of possible
substitution effects the neoclassical model incorporates factor prices in the form of the user cost
of capital (Ct). Hence the desired capital stock (Kt*) is as follows
σ
α
t
t
tC
Y
K=
* (5)
5 See also Kurz and Motolese (2001).
10
where Yis output,
α
is a constant and
σ
is the elasticity of substitution between factor inputs.
As discussed in Carruth et al. (2000a), empirical investigators typically assume either that
net investment is determined as a distributed lag process of changes in the desired capital stock, or
that there are explicit costs of adjustment. Then, it is possible to obtain an investment function for
empirical estimation that equates the level of investment to lags of the change in the level of
output and the user cost of capital. This is illustrated in the following four equations, where
equations (6) and (7) show the evolution of investment in terms of the capital stock, (where δ is
the depreciation rate and g is the steady state growth rate), (8) integrates this into equation (5) and
(9) sets out the equation in logarithmic form.
ttt dKKI
+
=
δ
(6)
()
tt KgI
δ
+
= (7)
()
σ
δα
ttt CYgI /+= (8)
() () ( )
ttt CYI lnlnln 10
ε
θ
θ
−
+
= (9)
A third approach often emphasised in the theoretical investment literature was introduced
by Tobin (1969) and highlights the importance of the stock market valuation of the firm relative
to the capital stock, or Q as opposed to factor prices. Where capital markets are perfect, there is
the implicit assumption that market value will incorporate information on market conditions, for
example, uncertainty of costs or final prices. Accordingly, we test for Tobin’s Q effects as a
variant when assessing the impact of uncertainty on investment.6,7
3.3 Econometric Approach
Turning to our estimation approach, panel methods have become popular in cross
sectional macro data sets since they provide a greater and more varied information set, and hence
6 We use the ratio of the capitalised financial value of the firm to the replacement cost of capital or average Q. This is
equivalent to the theoretically-appropriate but unobservable marginal Q subject to the production and adjustment cost
functions adhering to certain homogeneity conditions (implying inter alia that there is no market power).
11
greater efficiency, than individual country studies. There are two traditional methods for
estimating panel models: averaging and pooling. The former involves running N separate
regressions and calculating coefficient means (see for example the Mean Group Estimator method
advanced by Pesaran and Smith, 1995). A drawback to averaging is that it does not account for
the fact that certain parameters may be equal over cross sections. Alternatively we could pool the
data and assume that the slope coefficients and error variances are identical. However, whereas
there may be theoretical and empirical reasons to presume that the long-run coefficients are
homogenous over the cross-section, there are very few practical cases in which the short-run
dynamics and error variances would be homogeneous too. For example, in the case of investment,
it can be argued that differences in industrial and trade structure such as exists between the G7
would be likely to affect the short run rather than long run responses to uncertainty (given e.g.
their effects on the optimal life of capital goods). We would more generally argue that the PMG
approach may be seen in the investment function as modeling the supply side, whereby firms
have similar long term reactions to economic variables, given a common objective of profit
maximization in the long term, while in the shorter term institutions may play a role (such as
scope of liquidity provided by relationship banking) leading to differing dynamics. We note that
globalization and growth of multinationals makes a common long run response to volatility
plausible (although in the case of exchange rates multinationals might also be sensitive to
covariances between bilateral exchange rates).
Following this point, we consider it appropriate to use the Pesaran, Shin and Smith (1999)
Pooled Mean Group (PMGE) method, which is an intermediate case between the averaging and
pooling methods of estimation and involves aspects of both. The method restricts the long-run
coefficients to be equal over the cross-sections, but allows for the short-run coefficients and error
variances to differ across cross-sections. We can therefore obtain pooled long-run coefficients and
averaged short run error correction dynamics as an indication of mean reversion.
7 See Sensenbrenner (1991), Cuthbertson and Gasparro (1995) and Cooper and Ejarque (2001) for evidence on the
usefulness of Tobin’s Q in empirical modelling. In particular, Cuthbertson and Gasparro find that although Q is
12
The PMGE is based on an Autoregressive Distributive Lag ARDL(p,q,…,q) model
∑∑ =
−
=
−++
′
+=
q
j
itijitij
p
j
jitijit yy
01
εµδλ
x (10)
where it
x (k×1) is the vector of explanatory variables for group i,
µ
i represents the fixed effects,
the coefficients of the lagged dependent variables (
λ
ij) are scalars and
δ
ij are (k×1) coefficient
vectors. T must be large enough so that the model can be estimated for each cross section.
Equation (10) can be re-parameterised as:
itijit
q
j
ij
p
j
jitijitiitiit yyy
εµδλβφ
++∆+∆+
′
+=∆ −
−
=
∗
−
=
−
∗
−∑∑ xx 1
0
1
1
1' (11)
where ⎟
⎠
⎞
⎜
⎝
⎛−−= ∑=
p
jiji 1
1
λφ
, ∑=
=q
jiji 0
δβ
,∑+=
∗−= p
jm imij 1
λλ
and ∑+=
∗−= q
jm imij 1
δδ
In addition we assume that the residuals in (11) are independently distributed across i and t with
zero mean, variance greater than zero and finite fourth moments. Secondly, the roots of equation
(11) must lie outside the unit circle. The latter assumption ensures that
φ
i<0, and hence that there
exist a long-run relationship between yit and xit defined by
()
ititiiit
y
η
φ
β
+
′
−
=x/ (12)
The long-run homogeneous coefficient is equal to
()
iii
φ
β
θ
θ
/
′
−
== , which is the same across
groups. The PMGE uses a maximum likelihood approach to estimate the model and a Newton-
Raphson algorithm. The lag length for the model can be determined using, for instance, the
Schwarz Bayesian Information Criteria.
For our purposes, the key feature of the PMGE is to make the long-run relationships
homogenous while allowing for the heterogeneous dynamics and error variances. We test for
long-run homogeneity using, first, a joint Hausman test (see Pesaran, Smith and Im, 1996, for
details), and second a Likelihood Ratio approach. The Hausman test is based on the null of
equivalence between the PMG and MG estimation. If we reject the null we reject homogeneity of
our cross sections’ long run coefficients, while significant statistical difference between our two
important for long run UK investment, it is not a sufficient statistic.
13
estimators would be indicative of panel mis-specification. The Likelihood Ratio test for long run
parameter heterogeneity is more conventional in this setting and has homogeneity as the null
hypothesis (see Hsiao, 2003). The Likelihood Ratio statistic always rejects homogeneity in the
Pesaran et al. (1999) panel study of aggregate consumption and, as such, can be considered a
much more stringent test for poolability than the Hausman test (which typically accepts
poolability in the Pesaran et al. study). In Pesaran et al. (ibid), it is suggested that the low power
of the Hausman test may be due to the broad standard errors associated with the Mean Group
Estimator. Hence evidence from a Hausman test that Pooled Mean Group is not significantly
different from Mean Group estimation may not be a reflection of PMG being consistent but
merely that MG is inefficient. Given these arguments, we focus largely on the Likelihood Ratio
test in the results section below.
Another reason for utilising the PMGE in our empirical work is that the estimated
coefficients in the model are not dependent upon whether the variables are I(1) or I(0). Pesaran
and Shin (1998) present Monte Carlo evidence that ARDL approach based on the delta method
can be reliably used in small samples to estimate and test hypotheses on the long run coefficients
in cases whether there is a mixture of I(1) and I(0) regressors. We also calculated the Mean Group
(MGE) estimator, which is an average of the individual country coefficients. This provides
consistent estimates of the mean of the long-run coefficients although they are inefficient if slope
homogeneity holds. Under long-run homogeneity, PMG estimates are consistent and efficient.
4. Data
In terms of macroeconomic data needed for the investment function, we use quarterly time
series for the G7 countries, namely US, Canada, Japan, UK, Germany, France and Italy. We use
the OECD Business Sector Database to obtain data for real business investment (IB), real
business sector output (YB) and the real business sector capital stock (KB). Private sector
aggregates often include housing investment and the housing stock, which has a different cyclical
pattern to business investment. Problems of lack of congruence with investment would arise for
14
alternatives to output such as sales growth. Business sector data also overcomes the problems of
transfer from public ownership by including business sector capital, investment and output
irrespective of sector of ownership.
Meanwhile, we use the non-financial corporate sector capital stock for Tobin’s Q (Q),
consistent with the equity stock, which is obtained from sectoral financial accounts published by
the various national statistical agencies (for a discussion of these data see Byrne and Davis,
2003b). Cost of capital data are from the NiGEM macromodel database. The cost of capital is
calculated as the tax adjusted purchase price of a unit of capital, multiplied by the real post-tax
financial cost of capital plus the depreciation rate.
For generating uncertainty proxies, we utilise monthly CPI data, long term interest rate
data, nominal and real trade weighted effective exchange rate8 data, industrial production and the
stock market index for the G7 countries over 1968-2001. These are obtained from Primark
Datastream. In our investment functions we use the conditional variance of the monthly
difference of long rates (DLR ) and the log-difference of the nominal effective exchange rate
(DER), equity prices (DEQ), industrial production (DIP), consumer price9 (DP) and the real
effective exchange rate (DRER) estimated using GARCH as set out in Section 3.2. In the interests
of conciseness, we do not provide details of our GARCH estimates (full details of our conditional
variance estimation results are available in Byrne and Davis, 2002). It suffices to note here that
significant GARCH effects were present in each case and that the GARCH process in the
conditional-variance equation was covariance stationary with
α
1+
β
1<1. We note, following Engle
(1983) that GARCH generates a model-dependent outcome, although the approach adopted, of
differencing the conditional-mean equation sufficiently to avoid autocorrelation, is fairly
standard.
8 We consider the effective rate to be more appropriate than bilateral rates such as those with the US dollar given it
distributes exchange rate changes and related volatility appropriately across the export and import markets of the
country concerned.
9 We note that there are arguments for using volatility of the wholesale price index rather than the CPI. The choice
was based on the wider economic significance of the CPI and status as a target of monetary policy.
15
In evaluating the data prior to estimation, in particular to preclude a stationary regressand,
we use relatively powerful unit root tests using local GLS detrending and adopt the testing
strategy developed in Ng and Perron (2001). Further details and a summary of the test results are
given in Table 1.10 Logs of business investment and business output are I(1), as are most of the
times series in levels used in traditional specifications for investment. So too is Tobin’s Q,
consistent with Cuthbertson and Gasparro (1995). The conditional volatility measures are also
typically I(1), although there is consistent evidence that equity price and real exchange rate
volatility measures are I(0). As noted, Pesaran and Shin (1998) emphasise that the ARDL
approach is robust to the inclusion of I(1) and I(0) regressors. Our results from the unit root test
suggest that we do not start our analysis with evidence of an unbalanced regression. In this regard,
further evidence is provided below of mean reversion to our long run relationships.
5. Results
5.1 Estimation of Investment Functions without Uncertainty
Following the discussion above, in this section we develop a baseline investment function
which can be extended for testing the long run relationship between uncertainty and investment
using GARCH to derive conditional volatility. We assess an accelerator based investment
function (testing for additional cost of capital and Tobin’s Q effects) using the Pooled Mean
Group Estimator (PMGE) for dynamic heterogeneous panels as well as the mean group estimator
(MGE) based on averaging of individual-country results. In addition, we have heterogeneously
estimated dynamics, specified by information criteria. Following the approach of Cuthbertson and
Gasparro (1995) and Carruth et al. (2000a) to modelling long-run investment, the basic equation
is as follows:
() () ()
()
() ()
it
q
j
jti
ji
p
j
jti
ji
it
i
ti
i
it YBIBYBIBIB
εϕδθθφ
+∆+∆+−−=∆ ∑∑ =−
=−−
0
,
1
,
10
1, lnlnlnlnln (13)
10 Full details are available in Byrne and Davis (2002).
16
where in variants we add to the long and short run the additional variables Tobin’s Q, the user
cost and the long term interest rate. Table 2 focuses solely on the long run components of the
estimates. It shows that the long-run coefficient on output is statistically significant, irrespective
of which other variables are included in the specification. Also, the error correction term is always
significant and negative, which is consistent with a non-spurious relationship, although the
implied speed of adjustment is slow, as is typical of estimated investment functions. However, the
neoclassical opportunity cost measures of investment, namely the user cost and long term interest
rates, are often the wrong sign and are always insignificant. Tobin’s Q is significant and positive
in sign using the PMGE. Whereas it can be argued that omission of the long rate is surprising it is
consistent with many other empirical studies such as Chirinko (1993)11. The implication is that
the impact of monetary policy takes place via output and also (when included) Tobin’s Q. We
also acknowledge that the panel specification is fairly simple compared with what is feasible for a
single country equation with individual sources available nationally. However, we do note that in
terms of dynamics the PMG approach allows for marked cross country diversity in responses (not
shown in detail).
It is emphasised in the dynamic panel literature developed in Pesaran and Smith (1995),
Lee et al. (1997) and Pesaran et al. (1999) that it is important to consider whether there is slope
heterogeneity. This potential characteristic of our cross section can affect the usefulness of our
results, in particular, it may bias our pooled estimated coefficients. There is a degree of dispersion
of country slopes in our results: The Likelihood Ratio test rejects homogeneity at the 5%
significance level when we include output and Tobin’s Q in our specification.12 Given that the
most simple specification accepts both tests for poolability, χ2LR{6} = 8.45 [0.21], and slope
11 Also Darby et al (1999) only find a cost of capital effect for France among the G-5, while Cuthbertson and
Gasparro (1995) note that “time series studies have not yet succeeded in yielding robust tax effects via the cost of
capital” (ibid, p938).
12 The Hausman test statistic suggests poolability when we incorporate only Business Output and Tobin’s Q in our
long run specification with a test statistic of 0.20 (p-value=0.90). As pointed out by Pesaran et al. (1999) it is often
difficult to differentiate between panel specifications on the basis of this test unlike the more stringent Likelihood
Ratio test. Indeed the Mean Group coefficient on Tobin’s Q is always insignificant. We hence employ Tobin’s Q only
as a variant to check robustness. The instability of the estimated coefficient on Tobin’s Q when output is excluded
from the specification, and low likelihood, may well be due to different trending components in investment and Q.
17
coefficients are significant for both estimators we initially concentrate on the long run equilibrium
relationship with output only in our panel study. A sensitivity analysis reintroducing Tobin’s Q is
noted in Section 5.3 below allowing us to assess whether inclusion of output and uncertainty only
leads to significant omitted variables bias.
We cannot accept the hypothesis that the coefficient on output is equal to one in our
investment function, whether or not we include the long rate or Q, since the estimated output
coefficient reported in the second top row of Table 2 is more than two standard errors away from
1.00. Although Carruth et al. (2000a) and Cuthbertson and Gasparro (1995) both find evidence
that the null of output homogeneity is accepted, they also report evidence where the estimated
elasticity on output is greater than one. In particular, Cuthbertson and Gasparro’s preferred
specification has an estimated coefficient greater than one. The fact that we have estimated
coefficients in excess of one can be viewed as due to our wider sample and hence more efficient
estimated statistics. Our results are consistent with Jones (1995) who report evidence of
increasing investment output ratios for the industrialised countries in the post war period and
Davis and Madsen (2001) who note a sharp rise in the capital output ratio in the major industrial
countries. It may also be linked to an increase in the depreciation rate (where the basic model
assumes a constant rate) and increased capital mobility which has made fixed investment more
sensitive.
5.2 Impact of Uncertainty on Investment
We turn now to estimation of uncertainty effects on investment. Developing from equation
(13), Table 3 presents the estimated long run coefficients and short run error correction term from
equation (14).
() () () ()
()
()
() ()
∑∑
∑
=−
=−
=−−
+∆+∆+
∆+−−−=∆
r
j
itjtiji
q
j
jtiji
p
j
jti
ji
itit
i
ti
i
it
CVYB
IBCVYBIBIB
0
,
0
,
1
,
210
1,
.ln
ln.lnlnln
εγϕ
δθθθφ
(14)
18
The long run elasticities on output are always significant and the estimated coefficients are again
slightly larger than one in magnitude. Also, the error correction terms are significant, suggesting
mean reversion to a long-run relationship. In terms of the measures of volatility, we find for the
entire sample period that only the measure of nominal and real exchange rate uncertainty are
significant in influencing long-run business investment across the G7 with a PMG estimated
elasticity of -8.02 and -0.09 respectively (note that the latter is scaled by 100 to aid convergence
of the volatility estimates, so the effects are actually comparable in magnitude). Of the other
measures shown in Table 3, namely the conditional variance of equity prices, CV(DEQ),
industrial production, CV(DIP), and of inflation, CV(DP), there is often a negative effect from
uncertainty of these variables, although we do not find evidence that this is statistically significant
across the G7.
The probability values associated with the Hausman test always accepts the null
hypothesis of equivalence of PMG and MG and hence implies that there is parameter
homogeneity across the G7 as a whole. This is also true for equity prices, industrial production
and inflation for the Likelihood Ratio tests. However, we cannot accept parameter homogeneity at
95% for this test for long rate, nominal and real exchange rate uncertainty, although poolability is
close to acceptance for the long rate and the real exchange rate. Given the greater credence we
attach to the Likelihood Ratio statistic, we pursue the issue of poolability of the exchange rate
results by splitting our sample. One possible poolable combination for the nominal and real
exchange rate, according to the Joint χ2LR statistic, can be obtained by combining Germany,
France, UK and Italy (Tables 4A and 4B). In this instance the uncertainty coefficient increases as
does the significance, to -11.81 (t-statistic=3.31) for the nominal rate and -0.13 (3.23) for the
real rate. We also find a greater speed of mean reversion to the equilibrium relationship.
5.3 Sensitivity Analysis
To assess the robustness of our results, we considered it important to examine the stability
of our estimates across time. We do so by splitting our sample period in the early 1980s, dividing
19
the data roughly in half, and also estimating the uncertainty coefficient recursively. By these
means, we can examine whether the importance of exchange rate volatility has increased or
diminished over time, or indeed whether other types of volatility have important effects which are
obscured in estimation results for the entire sample.
The results from splitting the time series span into two sub-samples (Tables 5 and 6) show
that in the earlier sample period, 1973Q1 to 1983Q4, exchange rate volatility only has a
significant effect on investment at the 10% significance level; the estimated coefficient is -5.16 (t-
statistic = -1.80). However, for the later period the effect from the nominal effective exchange
rate increases in both magnitude and significance (the coefficient becomes -41.59, t-statistic =
2.42). There are similar increases in the real rate coefficient. We note also that the output
coefficients are closer to one in the later sample, in the case of exchange rates and long rates.
These results on the impact of exchange rate volatility are confirmed by the recursively estimated
coefficients (and standard error bands) in Figure 1 and 2.13 From the figures we see exchange rate
volatility becoming more important when we incorporate more recent data into our sample period.
For some of the earlier periods there does not appear to be a strongly negative coefficient. But as
we move into the 1990s the coefficient decreases further below zero (approximately by a factor of
two).
In the latter period, hedging was more common. This may be balanced by greater reliance
on external as opposed to internal finance for investment in the later period, (although note that
the Modigliani-Miller theorem implies that this should be “irrelevant”) and possibly greater
capital mobility. Another explanation may be that firms are less sensitive to exchange rates when
they are in a situation of imperfect competition. In the second half of the sample, financial
liberalisation, anti trust policy, privatisation, removal of trade barriers and other forms of market
opening have been much more marked. There has also been an increase in the range of goods and
13 We believe our approach to recursive estimation of PMGE is highly appropriate for testing the stability of
coefficients in a panel data study.
20
services which are traded. The resulting increased contestability may underlie the shifting
coefficient and growing impact of uncertainty.
There is evidence of cross sectional homogeneity of the long run coefficients for the later
period using both the tests when we incorporate nominal exchange rate volatility into our
specification (Table 5B, columns 2 and 3). The industrialised countries’ macroeconomic
structures with respect to investment, openness and the exchange rate volatility are, this result
suggests, becoming increasingly similar. This may again be linked to increased product market
liberalisation, capital mobility, increased openness and market efficiency.
The G7 results for the volatility of Industrial Production gives us some indication in the
earlier period that there is a significant positive effect on investment in the long run. However, we
discount this result here since the estimated coefficient is only marginally significant, the error
correction term is insignificant at the 5% level (suggesting the absence of a long run relationship
between investment and uncertainty) with no evidence of poolability from either the LR or
Hausman statistic.
With regards to the long-run interest rate, the estimated coefficient in the later period is
significant at the 5% level and negative (-0.23). Although we have a reasonable specification (e.g.
a significant error correction term) panel poolability is rejected using the LR test statistic, and
only accepted with the less stringent Hausman test. Consequently we again consider whether it is
possible to split the cross section sample in Table 6. Germany and Italy appear to behave similarly
with regard to long-run interest rate volatility. This is shown by the likelihood ratio statistic,
which is insignificant by a large margin when we combine these two countries. Our results are
somewhat surprising since Italy is typically considered as an economic “outsider” with regard to
EMU, see for example the evidence from Darby et al. (1999). In this instance it is clearly
exhibiting similar behaviour to the largest European economy. We do not find such clear evidence
for combined significance and poolability for France or the UK with regards the volatility of long
interest rates. But if anything the evidence sides with the UK exhibiting relatively (i.e. compared
21
to France) similar behaviour to Italy and Germany. The LR statistic is borderline significant at the
5% level and uncertainty is having a negative effect, with a coefficient of -0.45.
We seek to calibrate the impact of uncertainty on investment in the context of the
successful results highlighted above. These are shown in Table 7. It can be seen that a 10% rise in
the level of exchange rate or long rate volatility relative to the average leads to a fall in investment
of 1-1.5% except in the case of the later period exchange rate coefficient (5%) and the long rate
for the G7 in the later sample (0.7%). Using the standard deviation of volatility as a benchmark, a
1 standard deviation rise in conditional volatility leads to a 2-4% fall in investment for the full
sample G7 exchange rate and both the long rate estimates. In the case of the EU exchange rate the
effect is 7%, and 20% for the G7 exchange rate in the later sample.
Two further experiments were undertaken to test the robustness of the main results. First,
we introduced Tobin’s Q to the same equations, following Cuthbertson and Gasparro (1995) and
also to test the empirical finding of Leahy and Whited (1996) that uncertainty proxies may be
irrelevant in the presence of Q – or conversely that exclusion of Q led to omitted variables bias.
Broadly speaking, the results (Table 8) do not controvert our main findings. Q is significant at
around 0.16 in all of the PMG estimates. When including Tobin’s Q, the nominal and real
effective exchange rate remain significant in the PMG estimates. We can only accept poolability
with the Hausman test, however. Second, we removed time means from the variables in the panel
as a means of accounting for residual correlation across cross-sections14 but this again does little
to the results.15 Third, in Byrne and Davis (2002) we reported PMG results for the Kenen and
Rodrick and rolling standard deviation measures of uncertainty. Results are statistically weaker
and economically less plausible than for GARCH,16 thus lending support to the theoretical
arguments we have presented in favour of the superiority of GARCH as a measure of uncertainty.
14 Technically, this deals with common unobserved global factors which have the same effect on individual countries,
as may have occurred in the 1979 oil crisis for example.
15 Results are available upon request.
16 The PMG estimates using these moving-average measures of exchange rate volatility did not provide results in
favour of a negative effect from exchange rate volatility on investment. Of the other variables considered, there is
some indication of a negative effect from inflation. The implications of unconditional volatility for monetary policy
reactions may be a reason why it is moving average and not GARCH measures of inflation volatility that come to the
22
6. Conclusions
This paper contributes to the empirical literature on investment and uncertainty by
assessing a full range of sources of uncertainty and implementing evidence from pooled mean
group estimation. We can thus test explicitly for differences across countries and benefit from
both cross section and time series evidence when testing the relationship between investment and
uncertainty. Our panel approach has the additional benefit that it is based on the ARDL approach
which is robust to the incorporation of I(1) and I(0) regressors. We also suggest that utilising a
GARCH methodology is an improvement on moving average measures of volatility, as the latter
may only measure sample variability rather than uncertainty. The empirical evidence presented
in this paper shows unequivocally that exchange rate uncertainty is harmful to investment, both
for the G7 and for all subsamples. There is evidence of a growing exchange rate effect over the
sample. A long term interest rate effect also emerges for major EU countries other than France
over the 1984-96 period. For most estimates, a one standard deviation rise in conditional volatility of
effective exchange rates or long rates leads to a 2-4% fall in investment although some samples give
much greater declines. Results for uncertainty of inflation, equity prices and industrial production
do not, in contrast, suggest that these variables have a major and consistently negative effect on
investment across the G7.
As regards the implications for EMU, the panel results suggest it is of benefit for all the
large EU countries, including the UK, to reduce effective exchange rate and long rate volatility.
Since it is likely that EMU will reduce trade-weighted nominal exchange rate volatility (where
even for the UK, over 50% of trade is with the euro area), EMU is indicated to favour investment.
This will be the case as long as EMU does not boost real exchange rate volatility (owing to
differential inflation rates), or lead to much greater volatility in the remaining floating bilateral
rates (basically the yen and dollar) that form the trade weighted baskets of the UK and other large
fore. Finally, short, and long interest rates and the equity price series provided little evidence of any effect. In none of
the estimates was poolability accepted with the likelihood ratio test.
23
EU countries. Neither of these seems very likely. Equally, for some countries EMU may also
reduce long rate volatility (given, for example, a deeper and more liquid bond market, lower fiscal
deficits and less volatile short rates). This would compound the beneficial effect.
24
References
Abel, A. (1983). Optimal investment under uncertainty. American Economic Review 73: 228-233.
Abel, A, Dixit, A, Eberly, JC, Pindyck, RS. (1996). Options, the value of capital and investment. Quarterly
Journal of Economics 111: 753-777.
Ashworth, P, Davis, EP. (2001). Some evidence on financial factors in the determination of aggregate
business investment for the G7 countries. National Institute Discussion Paper No. 187.
Bean, C. (1981). An econometric model of manufacturing investment in the UK. Economic Journal 91:
106-121.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics
31: 307-27.
Bollerslev T, Chou RY, Kroner, KF. (1992). ARCH modelling in finance. Journal of Econometrics 52: 5-
59.
Byrne, JP, Davis, EP. (2002). Investment and uncertainty in the G7, NIESR Discussion Paper. No. 198.
Byrne JP, Davis EP. (2003a). Panel Estimation of the Impact of Exchange Rate Uncertainty on Investment
in the Major Industrial Countries, NIESR Discussion Paper No 208.
Byrne JP, Davis EP. (2003b). Financial Structure. Cambridge University Press: Cambridge.
Caballero, R. (1999). Aggregate Investment, in Handbook of Macroeconomics Vol. 1, Eds. J.B. Taylor and
M. Woodford, Elsevier Science.
Campa, JM, Goldberg, LS. (1995). Investment in manufacturing, exchange rates and external exposure.
Journal of International Economics 38: 297-320.
Carruth, A, Dickerson, A, Henley, A. (2000a). Econometric modelling of UK aggregate investment: The
role of profits and uncertainty. The Manchester School 68: 276-300.
Carruth A, Dickerson, A, Henley, A. (2000b). What do we know about investment under uncertainty?
Journal of Economic Surveys 14: 119-153.
Chirinko, R. (1993). Business fixed investment spending: Modelling strategies, empirical results, and
policy implications. Journal of Economic Literature 31: 1875-1911.
Cooper, R, Ejarque, J. (2001). Exhuming Q: Market power versus capital market imperfections, NBER
Working Paper # 8182.
Cushman, D O. (1983). The effects of real exchange rate risk on international trade. Journal of
International Economics 15: 45-63.
Cuthbertson, K, Gasparro, D. (1995). Fixed investment decisions in UK manufacturing: The importance of
Tobin's Q, output and debt. European Economic Review 39: 919-941.
25
Darby, J, Hughes Hallett, A, Ireland, J, Piscatelli, L. (1999). The impact of exchange rate uncertainty on
the level of investment. Economic Journal 109: C55-C67.
Davis, EP, Madsen, J. (2001). Productivity and Equity Returns: A Century of Evidence for 9 OECD
Countries. Brunel University Department of Economics and Finance Working Paper No 01-12.
Dixit, A, Pindyck, R.S. (1994). Investment under Uncertainty. Princeton University Press.
Driver, C, Moreton, D. (1991). The influence of uncertainty on aggregate spending. Economic Journal
101: 1452-59.
Driver, C, Yip, P, Dakhil, N. (1996). Large company capital formation and effects of market share
turbulence: micro-data evidence from the PIMS database. Applied Economics 28: 641-651.
Engle, RF. (1983). Estimates of the variance of US inflation based upon the ARCH model. Journal of
Money Credit and Banking 15: 286-301.
Ferderer, JP. (1993). The impact of uncertainty on aggregate investment spending: an empirical analysis.
Journal of Money Credit and Banking 25: 30-48.
Goldberg, LS. (1993). Exchange rates and investment in United States Industry. Review of Economics and
Statistics 75: 575-589.
Hartman, R. (1972). The effect of price and cost uncertainty on investment. Journal of Economic Theory 5:
258-266.
Hsiao, C. (2003). Analysis of Panel Data. 2nd Edition. Cambridge University Press: Cambridge.
Huizinga, H. (1993). Inflation uncertainty, relative price uncertainty and investment in US manufacturing.
Journal of Money, Credit and Banking 25: 521-554.
Jones, C. (1995). Time series tests of endogenous growth models. Quarterly Journal of Economics 110:
495-525.
Jorgensen,DW. (1963). Capital theory and investment behaviour. American Economic Review 53: 247-
259.
Kenen, P, Rodrick, D. (1986). Measuring and analyzing the effects of short-term volatility in real exchange
rates. Review of Economics and Statistics 68: 311-9.
Knight FH. (1921). Risk, Uncertainty and Profit. Boston: No 16 in series of reprints of scarce texts in
economics, LSE, London.
Kurz, M, Motolese, M. (2001). Endogenous uncertainty and market volatility. Economic Theory 17: 497-
544.
Lee, K, Pesaran, MH, Smith R. (2001). Growth and convergence in a multi-country empirical stochastic
Solow model. Journal of Applied Econometrics 12: 357-392.
26
Lee, J, Shin, K. (2001). The role of variable input in the relationship between investment and uncertainty.
American Economic Review 90: 667-680.
Leahy, J, Whited, T. (1996). The effects of uncertainty on investment: Some stylised facts. Journal of
Money, Credit and Banking 28: 64-83.
Ng, S, Perron, P. (2001). Lag length selection and the construction of unit root tests with good size and
power. Econometrica 69: 1519-1554.
Pesaran, MH, Shin Y. (1998). An autoregressive distributed-lag modelling approach to cointegration
analysis. In Econometrics and Economic Theory in the 20th Century: The Ragnar Frisch Centennial
Symposium, Strøm S (eds). Cambridge University Press: Cambridge.
Pesaran, MH, Shin, Y, Smith, R. (1999). Pooled Mean Group estimation of dynamic heterogeneous Panels.
Journal of American Statistical Association 94: 621-634.
Pesaran, MH, Smith, R. (1995). Estimating long-run relationships from dynamic heterogeneous panels.
Journal of Econometrics 68: 79-113.
Pesaran, MH, Smith, R, Im, K-S. (1996). Dynamic linear models for heterogeneous panels. In L. Matyas
and P. Sevestre (eds.) The Econometrics of Panel Data, Kluwer Academic Publishers: Dordrecht.
Pindyck, R.S., Solimano, A. (1993). Economic instability and aggregate investment, NBER
Macroeconomics Annual, 259-303.
Price, S. (1995). Aggregate uncertainty, capacity utilization and manufacturing investment. Applied
Economics 27: 147-154.
Sensenbrenner, G. (1991). Aggregate investment, the stock market, and the Q model - robust results for six
OECD countries. European Economic Review 35: 769-825.
Serven, L. (2003). Real exchange rate uncertainty and private investment in developing countries. Review
of Economics and Statistics 85: 212-217.
Shafer, JR. (1986). Managing crises in the emerging financial landscape. OECD Economic Studies 8: 56-
77.
Temple, P, Urga, G, Driver, C. (2001). The influence of uncertainty on investment in the UK: A macro or
micro phenomenon? Scottish Journal of Political Economy 48: 361-82.
Tobin, J. (1969). A general equilibrium approach to monetary theory. Journal of Money, Credit and
Banking 1: 15-29.
27
Table 1: Unit Root Tests
US
CN FR GE IT JP UK
ln(IB) I(1) I(1) I(1) I(1) I(1) I(1) I(1)
ln(IB)* I(1) I(1) I(1) I(1) I(1) I(1) I(1)
ln(YB) I(1) I(1) I(1) I(1) I(1) I(1) I(1)
ln(YB)* I(1) I(1) I(1) I(1) I(1) I(1) I(1)
Tobin’s Q I(1) I(1) I(1) I(1) I(1) I(1) I(1)
Tobin’s Q* I(1) I(1) I(1) I(1) I(1) I(1) I(1)
User Cost I(1) I(1) I(1) I(1) I(1) I(1) I(1)
User Cost* I(1) I(1) I(1) I(1) I(1) I(1) I(1)
Long Interest Rate I(1) I(1) I(1) I(0) I(1) I(1) I(1)
Long Interest Rate* I(1) I(1) I(1) I(1) I(1) I(1) I(1)
CV(DER) I(1) I(1) I(1) I(1) I(1) I(1) I(0)
CV(DER)* I(1) I(0) I(0) I(1) I(1) I(1) I(0)
CV(DP) I(1) I(1) I(1) I(0) I(1) I(1) I(1)
CV(DP)* I(0) I(0) I(1) I(1) I(1) I(1) I(1)
CV(DEQ) I(0) I(1) I(0) I(0) I(0) I(0) I(0)
CV(DEQ)* I(0) I(1) I(1) I(1) I(0) I(0) I(0)
CV(DIP) I(1) I(1) I(1) I(1) I(1) I(1) I(0)
CV(DIP)* I(1) I(1) I(1) I(1) I(1) I(1) I(1)
CV(DRER) I(0) I(0) I(1) I(0) I(1) I(1) I(0)
CV(DRER)* I(1) I(0) I(1) I(1) I(1) I(1) I(0)
CV(DLR) I(0) I(0) I(1) I(0) I(1) I(1) I(1)
CV(DLR)* I(1) I(1) I(1) I(0) I(0) I(0) I(1)
Notes: Sample period 1973Q1 to 1996Q4. Asterisk (*) indicates trend in unit root specification. CV(.) is the conditional variance from GARCH
estimation. Table summarises evidence from Byrne and Davis (2002) including: MZa and MZt Modified Phillips-Perron tests; a Modified Sargan-
Bhargava test; Elliott, Rothemberg and Stock feasible point optimal test (ERS Pt); a modified point optimal test; and DF-GLS the Augmented Dickey
Fuller test.
28
Table 2: Basic Investment Function
ln(YB) Tobin’s
Q
User
Cost
Long
Interest Rate
Error
Correction
MLL
χ2
LR{.} χ2
H{.}
PMGE 1.367
(t=24.610) -0.082
(-6.461)
MGE 1.393
(18.322) -0.092
(-5.818)
1648.783
<1653.010>
8.45
[p=0.21]
0.20
[0.65]
PMGE 1.247
(29.271) 0.166
(3.233) -0.091
(-6.908)
MGE 1.250
(10.998) 0.388
(0.743)
-0.103
(-6.043)
1653.317
<1666.087>
25.54
[0.01]
0.20
[0.90]
PMGE 1.247
(26.037)
0.169
(3.243)
-0.005
(0.013)
-0.089
(-6.550)
MGE 1.482
(6.757) 0.292
(0.375)
0.381
(0.319)
-0.117
(-7.371)
1650.481
<1678.085>
55.21
[0.00]
3.34
[0.34]
PMGE 1.346
(24.047) -0.523
(-0.931)
-0.083
(-6.543)
MGE 1.443
(15.648) 0.120
(0.191)
-0.103
(-7.613)
1644.076
<1651.366>
14.58
[0.28]
1.47
[0.48]
PMGE 1.258
(26.526) 0.171
(3.286) 0.002
(-0.468)
-0.092
(-7.018)
MGE 1.218
(11.578) 0.380
(0.968)
0.006
(-0.994)
-0.120
(-7.158)
1653.400
<1673.014>
39.23
[0.00]
na
PMGE 1.359
(24.150)
-0.002
(-0.385)
-0.082
(-6.608)
MGE 1.359
(13.857) -0.002
(-0.253)
-0.102
(5.690)
1651.637
<1659.466>
15.66
[0.21]
na
PMGE 8.948
(5.554) -0.010
(-1.199)
MGE 7.446
(1.688)
-0.033
(-3.023)
1484.742
<1491.177>
12.87
[0.05]
0.13
[0.71]
Notes: The dependent variable is Business Investment and the table shows the long-run coefficients from variants of equation (13). Business output is
ln(YB). Estimation is by Pooled Mean Group (PMGE) and Mean Group (MGE). Sample period 1973Q1 to 1996Q4. The ARDL lag structure is
determined by the Schwarz Bayesian Information Criteria. χ2LR{.} denotes the Likelihood Ratio Statistic, large values reject poolability of our cross
sections. χ2H{.} is the Hausman test for poolability which is a test for the equivalence of PMGE and MGE. If the null hypothesis is accepted (i.e. p-value
greater than 0.05) we can accept homogeneity of cross sectional long run coefficients. As discussed in Pesaran et al. (1996) there is no guarantee that the
difference between the variance of the mean group and pooled estimator used to calculate χ2H{.} is positive definite (such results are denoted here by na).
t-statistics based on asymptotic standard errors are in parentheses (⋅). p-values are in square brackets [⋅]. MLL is the maximised log likelihood, unrestricted
MLL is in brackets <⋅>.
29
Table 3: Panel Estimation of Investment and Uncertainty: G7 Countries
PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE
ln(YB) 1.346
(24.944) 1.439
(11.637) 1.377
(26.591) 1.466
(13.144) 1.363
(24.610) 1.370
(15.647) 1.348
(27.205) 1.413
(14.931) 1.348
(30.071) 1.390
(18.593) 1.371
(24.720) 1.439
(14.851)
CV(DER) -8.018
(-2.887) -25.198
(-2.097)
CV(DEQ) 0.374
(0.585)
-0.758
(-0.362)
CV(DP) -6.863
(-0.787)
-22.596
(-1.069)
CV(DIP) -0.067
(-1.189) 11.629
(0.817)
CV(DLR) 0.020
(0.302)
0.076
(0.680)
CV(DRER) -0.094
(-2.780) -0.256
(-1.547)
Error
Correction
-0.077
(-5.270) -0.083
(-4.431) -0.084
(-6.403) -0.095
(-4.926) -0.079
(-6.244) -0.093
(-6.132) -0.083
(-6.241) -0.095
(-5.544) -0.086
(-6.510) -0.112
(-5.553) -0.078
(-5.661) -0.081
(-4.429)
MLL 1652.252
<1667.613>
1649.147
<1655.401>
1645.014
<1654.103>
1649.533
<1654.573>
1646.737
<1657.649>
1651.887
<1662.906>
χ2
LR{.} 30.72 {12}
[p=0.00]
12.51 {12}
[0.41]
18.18 {12}
[0.11]
10.08 {12}
[0.61]
21.83 {12}
[0.04] 22.03 {12}
[0.04]
χ2
H{.} 3.44 {12}
[0.18]
1.39 {12}
[0.50]
1.47 {12}
[0.48]
1.64 {12}
[0.44]
0.50 {12}
[0.78]
2.18 {12}
[0.34]
Notes: The dependent variable is Business Investment and the results are the long-run coefficients from equation (14). Estimation is by Pooled Mean Group (PMGE) and
Mean Group (MGE). Sample period 1973Q1 to 1996Q4. ARDL lag structure determined by the Schwarz Bayesian Information χ2
LR{.} denotes the Likelihood Ratio
Statistic, large values reject poolability of our cross sections. χ2
H{.} is the Hausman test for poolability and tests for the equivalence of PMGE and MGE. If the null
hypothesis is accepted (i.e. p-value greater than 0.05) we can accept homogeneity of cross sectional coefficients. CV(.) is the conditional variance from GARCH estimation.
DER is the first difference of the log of the nominal effective exchange rate. DEQ the first difference of the log of the stock exchange index. DP is the first difference of the
log of the consumer price index. DIP is the first difference of the log of industrial production. DLR is the first difference of the long interest rate. DRER is the first
difference of the log of the real effective exchange rate. Underlying GARCH (1,1) estimates are given in Byrne and Davis (2002). t-statistics based on asymptotic standard
errors are in parentheses (⋅). p-values are in square brackets [⋅]. MLL is the maximised log likelihood, unrestricted MLL is in brackets <⋅>.
30
Table 4A: PMGE Investment and Nominal Exchange Rate Sub Samples
PMGE MGE PMGE MGE Individual
Coefficients
χ2
LR{.}
ln(YB) 1.346
(t=24.944) 1.439
(11.637) 1.233
(21.371) 1.202
(63.534) 0.221 {3}
[p=0.97]
CV(DER)
G7
-8.018
(-2.887) -25.198
(-2.097)
CV(DER)
EU4
-11.808
(-3.312) -12.670
(-2.852) 3.826 {3}
[0.28]
Error
Correction
-0.077
(-5.270) -0.083
(-4.431) -0.094
(-3.855) -0.097
(-4.578)
Joint Coefficients
χ2
LR{.}
30.72 {12}
[0.00]
4.19 {6}
[0.65]
χ2
H{.} 3.44 {12}
[0.18]
na
Notes: EU4 consists of France, Germany, Italy and the UK. See also Tables 2 and 3. t-statistics in
parentheses. p-values in square brackets [⋅]. Sample period 1973Q1 to 1996Q4. χ2
LR{.} denotes the
Likelihood Ratio Statistic. χ2
H{.} is the Hausman test.
Table 4B: PMGE Investment and Real Exchange Rate Sub Samples
PMGE MGE PMGE MGE
Individual
Coefficients
χ2
LR{.}
ln(YB) 1.371
(24.720) 1.439
(14.851) 1.265
(21.711) 1.253
(35.919) 0.868 {3}
[p=0.83]
CV(DRER)
G7
-0.094
(-2.780) -0.256
(-1.547)
CV(DRER)
EU4
-0.134
(-3.232) -0.136
(-3.522) 2.132 {3}
[0.55]
Error
Correction
-0.078
(-5.661) -0.081
(-4.429) n.a.‡ n.a.‡
Joint Coefficients
χ2
LR{.}
30.41 {12}
[0.00]
3.343 {6}
[0.76]
χ2
H{.} 2.18
[0.34]
na
Notes: EU4 consists of France, Germany, Italy and the UK. See also Tables 2 and 3. ‡ This is due to a
problem with an Italian dynamic lag length greater than one. Reducing the maximum lag length to
one, reduced the pooled coefficient to -0.142 without changing the individual country results for the
other countries and producing a significant error correction term for PMGE. Sample period 1973Q1
to 1996Q4.
31
Table 5A: Panel Estimation of Investment and Uncertainty: G7 Countries Time Split 1: 1973Q1 to 1983Q4.
PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE
ln(YB) 1.412
(5.164) 1.121
(4.256) 1.446
(10.837) 1.140
(3.378) 1.444
(12.217) 1.473
(6.381) 1.574
(11.845) 1.219
(6.541) 1.436
(10.787) 1.324
(4.874) 1.421
(10.424) 1.252
(4.329)
CV(DER) -5.164
(-1.803)
-13.144
(-1.325)
CV(DEQ) -0.317
(-0.412)
4.300
(0.890)
CV(DP) -0.990
(-0.198)
17.526
(1.318)
CV(DIP) 6.437
(1.994) 35.891
(1.025)
CV(DLR)
0.108
(0.982)
-0.134
(-0.755)
CV(DRER)
-0.041
(-1.480)
-0.044
(-0.632)
Error
Correction
-0.115
(-2.855) -0.083
(-4.431) -0.120
(-3.034) -0.197
(-2.285) -0.079
(-6.244) -0.093
(-6.132) -0.072
(-1.680)
-0.212
(-2.276) -0.123
(-3.053) -0.213
(-2.589) -0.119
(-2.939) -0.210
(-2.632)
χ2
LR{.} 34.65 {12}
[0.00]
29.01 {12}
[0.00] 22.88 {12}
[0.03] 37.32 {12}
[0.00] 23.22 {12}
[0.03] 30.224 {12}
[0.00]
χ2
H{.} 6.88 {12}
[0.03] 1.21 {12}
[0.55]
2.27 {12}
[0.32]
11.77 {12}
[0.00] 3.24 {12}
[0.20]
0.99 {12}
[0.61]
Notes: See Tables 2 and 3
Table 5B: Panel Estimation of Investment and Uncertainty: G7 Countries Time Split 2: 1984Q1 to 1996Q4.
PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE
ln(YB) 1.150
(9.811) 1.562
(2.841) 0.765
(3.953) 1.098
(1.979) 2.227
(7.062) 2.049
(2.398) 0.611
(0.248)
2.382
(1.892)
1.158
(12.919) 2.897
(2.028) 1.131
(8.578) 2.222
(1.739)
CV(DER) -41.593
(-2.422) -49.053
(-1.904)
CV(DEQ) 0.673
(0.499)
-3.271
(-1.195)
CV(DP) 19.001
(1.032)
41.121
(0.285)
CV(DIP)
-9.056
(-0.958)
-24.541
(-0.436)
CV(DLR) -0.227
(-2.268) 0.359
(0.292)
CV(DRER) -0.452
(-2.361) -0.966
(-1.330)
Error
Correction
-0.084
(-6.403) -0.095
(-4.926) -0.043
(-1.807)
-0.098
(-3.904) -0.039
(-1.740)
-0.091
(-3.454) -0.028
(-1.198) -0.093
(-3.775) -0.082
(-4.510) -0.091
(-3.269) -0.036
(-2.535) -0.078
(-3.174)
χ2
LR{.} 18.18 {12}
[p = 0.11]
34.69 {12}
[0.00] 42.73 {12}
[0.00] 38.20 {12}
[0.00] 45.97{12}
[0.00] 33.214 {12}
[0.00]
χ2
H{.} 1.39 {12}
[0.50]
2.78 {12}
[0.25]
0.56 {12}
[0.76]
7.01 {12}
[0.03] 2.79 {12}
[0.25]
0.79 {12}
[0.67]
32
Figure 1: The Long-Run Recursive PMGE
Coefficient for Nominal Exchange Rate Uncertainty
-20
-15
-10
-5
0
5
1983 1985 1987 1989 1991 1993 1995
LR Coeff. +1.96*S.E. -1.96*S.E.
Figure 2: The Long-Run Recursive PMGE
Coefficient for Real Exchange Rate Uncertainty
-0.25
-0.2
-0.15
-0.1
-0.05
0
0.05
1983 1985 1987 1989 1991 1993 1995
LR Coeff. +1.96*S.E. -1.96*S.E.
33
Table 6: Panel estimates for Long Interest Rates: European Countries later Panel
PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE
ln(YB) 1.158
(12.919) 2.897
(2.028) 0.876
(5.049) 1.495
(1.820)
1.105
(9.487) 1.288
(5.286) 0.879
(5.033) 0.767
(1.424)
1.106
(9.584) 2.085
(2.576)
CV(DLR)
G7
-0.227
(-2.268) 0.359
(0.292)
CV(DLR)
FR, GE, IT & UK
-0.215
(-1.114)
1.630
(0.969)
CV(DLR)
GE & IT
-0.473
(-2.026) -0.511
(-4.938)
CV(DLR)
GE, IT & FR
-0.242
(-1.254)
-0.014
(-0.028)
CV(DLR)
GE, IT & UK
-0.452
(-2.023) 1.847
(0.783)
Error
Correction
-0.082
(-4.510) -0.091
(-3.269) -0.068
(-8.340) -0.078
(-4.268) -0.101
(-4.560) -0.105
(-11.409) -0.076
(-11.701) -0.080
(-3.142) -0.077
(-2.709) -0.093
(-7.332)
χ2
LR{.} 45.97{12}
[0.00] 22.00 {6}
[0.00] 1.97 {2}
[0.37]
14.73 {4}
[0.01] 9.51 {4}
[0.05]
χ2
H{.} 2.79 {12}
[0.25]
1.25 {6}
[0.54]
na na na
Notes: Sample period 1984Q1-1996Q4, see Tables 2 and 3.
Table 7: The economic impact of uncertainty on investment
Sample Variable Average
level of cv. Coefficient Base level 10% rise 1 std dev 50% rise
1973-1996 G7 nominal exchange rate 0.013 -8.018 -0.107 -0.011 -0.037
-0.054
1973-1996 EU4 nominal exchange rate 0.012 -11.800 -0.146 -0.015 -0.070
-0.073
1973-1996 G7 real exchange rate 1.41 -0.094 -0.133 -0.013 -0.035
-0.066
1973-1996 EU real exchange rate 1.27 -0.134 -0.171 -0.017 -0.064
-0.085
1984-1996 G7 nominal exchange rate 0.013 -41.600 -0.534 -0.053 -0.202
-0.267
1984-1996 G7 long rate 0.311 -0.227 -0.074 -0.007 -0.021
-0.037
1984-1996 EU3 long rate 0.298 -0.452 -0.144 -0.014 -0.035
-0.072
Notes: base level is equal to the coefficient times the average level of conditional variance (cv) over the sample. Since the investment function is
specified in logs, 0.01 is equivalent to a 1% change.
34
Table 8: Panel Estimation of Investment and Uncertainty with Tobin’s Q
PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE PMGE MGE
ln(YB) 1.247
(29.271) 1.250
(10.998) 1.282
(22.641) 1.245
(10.272) 1.236
(25.336) 1.198
(12.887) 1.255
(27.397) 1.284
(12.029) 1.243
(28.968) 1.248
(12.774) 1.236
(30.469) 1.261
(12.909) 1.259
(28.733) 1.240
(12.577)
Q 0.166
(3.233) 0.388
(0.743)
0.210
(2.218) 0.487
(1.059)
0.166
(3.031) 0.360
(0.662)
0.227
(4.325) 0.338
(0.842)
0.164
(3.155) 0.375
(0.720)
0.173
(3.665) 0.335
(0.871)
0.165
(2.992) 0.506
(1.024)
CV(DER)
-7.135
(-2.717) -23.609
(-2.154)
CV(DP) -4.143
(-0.643)
-40.500
(-1.360)
CV(DEQ) -0.679
(-1.099)
0.238
(0.086)
CV(DIP) -0.047
(-1.203) 13.584
(0.838)
CV(DLR) -0.049
(-1.092)
0.134
(1.045)
CV(DRER)
-0.076
(-3.096) -0.252
(-1.672)
Error
Correction
-0.091
(-6.908) -0.103
(-6.043) -0.079
(-5.542) -0.030
(-2.153) -0.089
(-7.036) -0.100
(-5.822) -0.085
(-7.007) -0.112
(-4.870) -0.091
(-6.953) -0.103
(-5.991) -0.090
(-6.630) -0.117
(-6.097) -0.089
(-6.252) -0.093
(-4.809)
MLL 1653.32
<1666.09>
1654.2597
<1678.1378>
1653.5399
<1668.2952>
1650.51
<1675.51>
1654.07
<1667.37>
1653.72
<1673.95>
1660.75
<1677.56>
χ2
LR{.} 25.54 {12}
[0.01] 47.76 {18}
[0.00] 29.51 {18}
[0.04] 50.00 {18}
[0.00] 26.59 {18}
[0.09]
40.46 {18}
[0.00] 33.62 {18}
[0.01]
χ2
H{.} 0.20{12}
[0.90]
11.15{18}
[0.01] 5.81{18}
[0.12]
0.16{18}
[0.98]
1.01{18}
[0.80]
2.51{18}
[0.47]
2.24 {18}
[0.52]
Notes: See Tables 2 and 3
35
Abstract
Empirical work on uncertainty and investment to date is generally single-country or single-indicator.
We consider the impact of a range of sources of uncertainty on aggregate business investment across
the G7, using Pooled Mean Group Estimation, and GARCH methods to model uncertainty. Significant
negative long run effects from effective trade-weighted exchange rate volatility are found for the G7
and in poolable subgroups including all four larger EU countries. Volatility of long rates also
influenced investment recently. A one standard deviation rise in conditional volatility generally leads
to 2-4% lower investment. The results suggest inter alia that EMU is beneficial to aggregate
investment.
Keywords: Investment; Uncertainty; Exchange Rates, Non-Stationary Panel Estimation.
JEL Classification Numbers: E22, F31.
