A meta-analysis of the investment-uncertainty relationship.
Koetse, Mark J. ; de Groot, Henri L.F. ; Florax, Raymond J.G.M 等
1. Introduction
The relationship between investment and uncertainty has been
extensively analyzed in both the theoretical and the empirical
literature since the 1970s. One of the most salient features of the
theoretical literature is the inconclusiveness about the direction of
the relationship. Early models can be found in Hartman (1972), Pindyck
(1982), and Abel (1983), among others. Hartman (1972) uses a
neoclassical model without capital stock adjustment costs and analyzes
the relationship between capital productivity and uncertainty. Under the
convexity of this relationship, and by Jensen's inequality, the
incentive to produce and invest increases when uncertainty increases,
implying a positive relationship. Hartman's model is, however,
restricted to markets with perfect competition and relies on assumptions
of constant returns to scale and substitutability of capital for other
input factors. Also, adjustment costs are assumed to be symmetric. In
real-world situations, however, this assumption is likely violated for
most capital investments. Pindyck (1982) allows for asymmetric
adjustment costs and argues that the effects of uncertainty on
investment spending are in fact dependent on the characteristics of the
adjustment cost function. Abel (1983) argues that increased uncertainty
leads to increased investment spending, regardless of the
characteristics of the adjustment cost function, effectively confirming
the results of Hartman (1972). However, he also shows that adjustment
costs matter for the relationship between investment and Tobin's Q.
As a result, uncertainty has a direct effect on investment but also an
indirect effect through Q in the Abel (1983) model. The overall effect
is positive when adjustment costs are convex but is ambiguous when
adjustment costs are concave.
More recent thinking about uncertainty is dominated by the concept
of capital investment irreversibility (see Pindyck 1991; Dixit and
Pindyck 1994). For a neoclassical model with asymmetric capital
adjustment costs, that is, a certain degree of irreversibility of
capital investment, Dixit and Pindyck show that an increase in
uncertainty creates an option value of waiting for new information to
arrive in the future. The central point of the irreversibility or real
options literature is that an increase in uncertainty will ceteris
paribus result in more investment projects being delayed. This argument
has major implications for the timing of investment, implying that
short-run investment levels may be affected but long-run investment
levels will not. We can therefore distinguish between two general
branches of research regarding the investment-uncertainty relationship:
a first branch in which uncertainty is related to the timing of
investment, and a second branch that analyzes the impact of uncertainty
on the investment level. (1) This article focuses primarily on the
second branch.
Given the ambiguity of the theoretical literature, there is no way
to determine the direction of the relationship between investment and
uncertainty a priori, let alone to draw inferences on the magnitude of
the effect and its economic relevance. Various explanations for this
ambiguity have been brought forward. One of the most obvious sources of
heterogeneity is the degree of irreversibility of investment itself;
that is, the smaller the possibilities to disinvest, the larger the
negative impact of uncertainty on investment spending. A similar
argument holds for risk aversion. (2) Numerous attempts have been made
to resolve this issue empirically, but these attempts seem to have added
to, rather than resolved, the existing ambiguity. In Carruth, Dickerson,
and Henley (2000) an excellent overview is provided of the most relevant
topics in the theoretical debate on the investment-uncertainty
relationship. Major issues in the empirical literature are also
discussed, such as the possible consequences of data aggregation and the
differences between operational measures of uncertainty. However,
although the Carruth, Dickerson, and Henley (2000) study is obviously
useful and important in its own right, it is qualitative in nature and
does not attempt to quantify the importance of differences in study
characteristics for the variation in study outcomes.
In this article we therefore perform a meta-analysis on the
relationship between uncertainty and investment spending. Meta-analysis
is a form of quantitative research synthesis originally developed in
experimental medicine and later extended to fields such as biomedicine
and experimental behavioral sciences, specifically education and
psychology. During the last two decades it has also been widely applied
in economics.3 The intuitive appeal of meta-analysis rests on its
ability to combine sometimes widely scattered empirical evidence on a
certain topic and the associated increase in statistical power of
hypothesis testing when combining independent research results.
Moreover, by controlling for variations in characteristics across
studies, meta-analysis provides quantitative insight into which factors
are relevant in explaining the variation in study outcomes. As such,
meta-analysis provides a quantitative analytic assessment in addition to
the more qualitative judgment provided by a narrative literature review
(Stanley 2001).
The remainder of this article is organized as follows: Section 2
discusses the type of estimates used in this study, as well as the way
in which they have been sampled from the literature. We also provide
descriptive statistics for the resulting meta-sample. Section 3
discusses the operationalization of moderator variables, which represent
differences in study characteristics that may systematically affect a
study's outcome. The model and estimation procedure are presented
in section 4, while section 5 discusses the estimation results. Section
6 concludes.
2. Effect Size, Sampling Procedure, and Sample Characteristics
Empirical studies on investment behavior are heterogeneous in many
respects. Studies generally include a wide variety of explanatory
variables; they operationalize investment and uncertainty in different
ways; and they are performed on samples that vary over time and space.
The impact of these sources of heterogeneity will be assessed in the
meta-regression analysis. An important observation is that the
investment-uncertainty literature is focused on the sign of the
relationship and not on its magnitude. In order to focus on the main
issue in the literature we define an effect size that does not include
the magnitude of the relationship. In our attempt to create a sample of
effect sizes that included the magnitude of the relationship we faced
several restrictions, the most important being that most of the study
results are not defined in a common, scale-free metric. In the empirical
literature, three different functional forms are used in primary
studies: linear, semilogarithmic, and double-log specifications.
Coefficients from double-log models can be interpreted as elasticities,
which is generally a good measure for an effect size. Coefficients from
linear and semilogarithmic models, however, are not scale-free, implying
that results from these studies are incomparable and that a
transformation into elasticities is necessary. This was only feasible
for a limited subset of linear and semilog coefficients, implying that
the resulting sample of elasticities would be substantially smaller than
the original sample of study results. In order to use the full sample of
study results, while still focusing on the main issue in the literature,
we focus on the direction and statistical significance of the estimates
rather than on the magnitude of the elasticities. In our empirical
analysis we ultimately distinguish between significantly negative,
insignificant, and significantly positive study results.
In creating our sample of studies we first searched through titles
and abstracts of studies using keywords in standard online search
engines such as Econlit, Picarta, and RePEc. Keywords used were
"investment," "uncertainty," and
"volatility." We subsequently looked for papers and articles
in the reference lists of studies that were collected. Ultimately, we
collected 48 studies that empirically analyze the relationship between
uncertainty and investment. These studies provided a total of 957
estimates, but some studies and estimates had to be excluded from the
database for one of the following reasons. First, as suggested by Abel
and Eberly (1999), among others, one of the potential reasons for the
theoretical ambiguity on the direction of the relationship is that the
relationship is potentially hump shaped. (4) Two studies in our sample
use a model specification in which uncertainty is included in a linear
and a quadratic fashion to test this hypothesis (Lensink 2002; Bo and
Lensink 2005). (5) In these studies, the effect of uncertainty on
investment is conditional on the degree of uncertainty. We therefore
excluded these observations from the analysis (32 estimates). Second,
some studies use a logit or probit model to estimate the relationship.
In these models the dependent variable is either binary or ordered; that
is, the analysis is concerned with estimating the impact of factors
determining the probability that investment actually takes place. As
such, the results from these models do not provide information on the
change in the level of investment and are therefore excluded from the
analysis (59 estimates). Third, standard errors or /-statistics are
essential for constructing our dependent variable, since they are used
to calculate the/)-values (statistical significance) of study estimates.
We exclude the estimates for which no standard errors or t statistics
were provided in the study (24 estimates). Fourth, some models provide
information on the relationship between investment and an uncertainty
measure that was interacted with another variable. For these models
either the isolated effect of uncertainty on investment could not be
extracted, or standard errors for the isolated effect could not be
obtained. These estimates are therefore excluded (32 estimates).
Finally, some studies use alternative dependent variables, such as the
capital-to-labor ratio (Ghosal 1991, 1995; Green, Lensink, and Murinde
2001) or the investment lag (Favero, Pesaran, and Sharma 1994; Hum and
Wright 1994). The former studies provide interesting insights on factor
demand under uncertainty but provide little direct evidence on
investment behavior, and they are therefore omitted (23 estimates). The
latter studies measure the delay of investment instead of the effect on
the investment level, and although this provides interesting insights,
the outcomes of these studies are incomparable to the outcomes of other
empirical studies in our meta-analysis (20 estimates).
Ultimately, we arrive at a sample of 767 observations taken from 36
different studies (studies are included in the References section,
denoted by *). In Appendix 1 we provide details of each study included
in the meta-analysis. Table 1 presents descriptive statistics of the
sample. The table shows that 66% of the estimates are negative. When a
distinction is made between statistically significant and insignificant
results, using a critical significance level of 5%, the number of
insignificant negative results is of the same order of magnitude as the
number of insignificant positive results. However, a significantly
negative relationship is observed quite frequently (30%), while very few
observations actually find a significantly positive relationship between
investment and uncertainty (4%). Interestingly, the positive estimates
(30 in total) are taken from nine different studies that differ widely
in terms of, for instance, data type, region, time period, and sample
size.
3. Operationalization of Moderator Variables
This section discusses several dimensions along which the studies
that have been included in our analysis differ. We discuss the measure
of investment that has been used, the uncertainty measure, control
variables, data characteristics, spatial and temporal data
characteristics, and estimation technique. This provides the basis for
our meta-regression model specification.
Measures of Investment
Most studies use aggregate investment figures, implying that the
type of investment is fairly homogeneous across studies. Investment is
specified in three distinct ways: unsealed investment, investment scaled
by some measure of income (sales, gross domestic product [GDP]), and
investment scaled by capital. It is not always clear a priori why and in
what way measurement differences affect study outcomes. For example, in
a period of increased uncertainty a firm may decrease investment even
further when it also considers its existing capital stock to be too
large (see the hangover effect discussed in footnote 1). This effect may
have a bigger impact on the outcomes of studies that use
investment-to-capital ratios than on outcomes of studies that use
unsealed investment because in the former case the decrease in
investment is measured as a fraction of a capital stock that is already
considered too high. The effect of uncertainty on investment will still
have the same direction for both investment measures but may seem larger
in magnitude for studies that use the investment-to-capital ratio. This
may systematically affect the statistical significance of study
outcomes, and as such it may show up in our regression results.
Sources and Measures of Uncertainty
The source of uncertainty is heterogeneous across studies. We
distinguish seven sources of uncertainty: uncertainty of
sales/demand/output, profit, output prices, input prices, inflation,
exchange rates, stock prices, and a rest category with variables such as
uncertainty of government expenditures. It is interesting to see whether
these uncertainty differences have an impact.6 We explicitly control for
the fact that in many studies multiple uncertainty sources were included
in the model specification.
There is no clear consensus in the literature on how to construct a
good proxy for uncertainty, mainly because the method of measuring
uncertainty is associated with assumptions regarding the expectation
formation process of decision makers. As a consequence, several measures
of uncertainty are used. Most of the empirical studies on the
investment-uncertainty relationship use historical data on the variable
under investigation to create an uncertainty proxy. However, since
uncertainty is fundamentally a forward-looking phenomenon, historical
data are flawed. A popular approach to measure forward-looking
uncertainty is to ask entrepreneurs or economists for their subjective
evaluations of uncertainty. Five of the articles in our sample use such
a subjective uncertainty measure, avoiding the inherent theoretical
problems associated with historical data. For example, Pattillo (1998)
and Lensink, van der Steen, and Sterken (2005) use a survey in which
they ask entrepreneurs to give a probability distribution of the
development of expected sales over a certain time period.7 Such a
measure comes closest to the ideal measure of an individual's
perceived uncertainty. In our model we account for the difference
between studies that use backward-looking historical data and studies
that use forward-looking subjective evaluations of uncertainty.
Control Variables in Primary Models
Empirical studies on investment can roughly be classified into two
groups according to the underlying theoretical models. The first model,
discussed extensively in Jorgenson (1971), is the accelerator model of
investment. In this model investment spending is driven by income or
sales. These models include sales or GDP as an explanatory variable. The
second investment model is the Q model of investment. In this model
investment takes place if Tobin's marginal Q, the ratio of the
marginal value of capital and the market price of capital, is larger
than 1 (see Tobin 1969; Cuthbertson and Gasparro 1995). Since stock
prices, and therefore Q, reflect expected future profits, the Q model
has an additional feature above and beyond the standard neoclassical
investment model in that it incorporates expected future profits into
current investment decisions. Since Q represents the market value of
capital it should, in principle, incorporate the effects of uncertainty.
It is therefore interesting to check whether explicitly accounting for
uncertainty has power in explaining investment behavior above and beyond
Q. (8)
Apart from the two different types of investment models, there is
substantial variation in the control variables used in the underlying
primary studies. Several studies include one or more of the following:
wages and capital prices, a time trend, debt position, stock prices,
size of a firm, government expenditures, a lagged dependent variable to
control for autocorrelation, and trade flows. Since the explanatory
variables are used in different combinations, we cannot distinguish
between well-defined empirical models and instead resort to including
dummy variables for each of these explanatory variables in the
meta-regression model.
Data Characteristics
An interesting difference in data characteristics across studies
pertains to the distinction between industry-wide and idiosyncratic
firm-level uncertainty. In the models by Abel (1983) and Caballero
(1991), idiosyncratic uncertainty has a positive effect on investment
for firms with constant returns to scale technology, operating in a
competitive environment. Pindyck (1993) argues that if uncertainty is
identical for all firms in an industry, it will be more difficult to
disinvest than if only a single firm experiences increased uncertainty.
He subsequently shows that, under identical technology and market
conditions, industry-wide uncertainty has a negative impact on
investment spending. However, under alternative assumptions,
idiosyncratic uncertainty may be just as important for investment
decisions. In fact, in an empirical analysis using firm-level data. Bo
(2002) finds that idiosyncratic uncertainty has a negative impact on
investment and that it is more important than industry-wide uncertainty.
Because of measurement difficulties, empirical studies that distinguish
explicitly between industry-wide and firm-level uncertainty are scarce.
In the meta-regression analysis the effects may be picked up by
differences in the level of data aggregation. However, differences in
data aggregation may also have other effects on study outcomes, which
may obscure the differential effects of industry-wide and idiosyncratic
uncertainty.
Another distinction in data characteristics concerns the difference
between time-series, panel, and cross-section data. Time-series data
typically reflect short-run changes, while crosssection data are
generally perceived to identify long-run changes. The effect of panel
data is expected to be closer to the short-run effect, the intuition
being that fixed or random effects in a panel data model will generally
control for differences between countries, regions, or industries and
firms, thereby likely picking up at least part of the longer term
effects. The distinction between different types of data in the
meta-regression analysis will therefore shed light on differences
between short- and long-run effects of uncertainty on investment.
Spatial, Temporal, and Econometric Issues
Regional differences across studies may reflect differences in, for
instance, sector composition, the degree of competition, institutional
setting, the functioning of capital markets, and culturally determined
risk aversion. Since many studies use data from several countries and
from several parts of the world, a clear distinction between, for
instance, the United States and Europe, cannot be made. The only clear
distinction is between developed and less developed countries.
Furthermore, in order to control for differences in the time period used
in an underlying study we construct a time trend based on the median
year of the sample data in the underlying studies. (9) Finally, we
distinguish between studies that use ordinary least squares (OLS) or
more sophisticated estimation techniques and control for studies that
correct for possible endogeneity by using instrumental variables.
Remaining Sources of Variation
Clearly, some sources of heterogeneity cannot be accounted for in
the meta-regression analysis. For example, potential control variables
for the investment-uncertainty relationship in underlying studies are
the degree of irreversibility of investment, the degree of risk
aversion, and assumptions with respect to substitution possibilities
between production factors. However, most of the empirical studies do
not provide explicit information on these issues. The only possible way
to account for these sources of heterogeneity is by including fixed
effects for specific sectors with different characteristics on the
abovementioned dimensions. A practical difficulty here is that most
studies that use sector-level data do so for the entire manufacturing
sector, implying that very little sectoral variation is present in the
set of underlying studies. Sector-specific fixed effects are therefore
not included in the meta-regression model, and part of the heterogeneity
in the sample is likely left unexplained. However, the fact that most
studies use aggregate manufacturing data has a potential advantage as
well. Within a single study, sectoral variation may be substantial, but
since most studies use comparable data, sectoral differences between
studies are likely small.
4. Model and Estimation Procedure
In the meta-regression model we distinguish between significantly
negative, insignificant, and significantly positive estimates using a
categorical effect size indicator as the dependent variable. The
categories are labeled 0, 1, and 2, respectively, using a 5% critical
significance level. The model that is generally applied in a
meta-analysis with a categorical effect size with more than two ordered
categories is the ordered probit model. (10) In estimating this model we
have to deal with the fact that multiple estimates are derived from a
single study. As shown by Bijmolt and Pieters (2001), a good way to deal
with this is to estimate a model with equal weights per study in which
each observation is weighted with the inverse of the total number of
estimates that is drawn from the same study (see Bijmolt and Pieters
2001). This procedure prevents studies with a large number of estimates
from having a disproportionally large influence on the estimation
results.
Because the ordered probit results reveal the direction of change
rather than the absolute magnitude of changes in observing an effect in
one of the three categories, we transform the ordered probit
coefficients into marginal effects. This implies that for each of the
explanatory variables in our model we calculate the change in the
probability of obtaining a significantly negative, an insignificant, and
a significantly positive estimate. (11) Almost all explanatory variables
are dummy variables, for which marginal effects are obtained by
considering a shift in the dummy value from 0 to 1, keeping the other
explanatory variables (dummy variables included) constant at their
respective means. Since standard errors of the computed marginal effects
are not readily available, they are obtained by linear approximation
using the delta method (Greene 2003, pp. 674-675). (12)
A relevant issue in meta-analysis is publication bias, and
preferably remedial devices should be used to avoid spurious results
(see the contributions in Roberts and Stanley 2005). However, existing
methods to detect and correct for potential publication bias, such as
Hedges' weighted distribution theory (Hedges 1992) and the
meta-regression methods developed by Stanley (2008), have been developed
for effect sizes defined on a ratio scale. The use of standard errors in
defining a categorical effect size indicator in our meta-regression
model precludes use of these techniques. We therefore mitigate the
potential influence of publication bias by including the year of
publication in our model specification.13 Since we also include the
average year of the sample used in a study, the aim here is to correct
for a possible publication time trend (see Goldfarb 1995). Potentially
problematic, however, is the high correlation coefficient between the
two time trend variables (r = 0.68). Ultimately, we present two separate
models: a model with and a model without publication year among the
explanatory variables. (14)
5. Estimation Results
The ordered probit estimates and associated marginal effects of the
model without the publication time trend variable (model 1) are
presented in Table 2. Regarding the measurement of investment, there is
a clear difference in study outcomes between studies that use an
investment-to-capital ratio and studies that use the investment level
(the omitted category in the meta-regression) or the investment-to-sales
ratio. Studies that use the investment-to-capital ratio are
characterized by a relatively high probability of finding a
significantly negative result. The only difference between these
measures is the presence (or absence) of capital in the denominator. The
explanation is therefore most likely related to the impact of
uncertainty on the capital stock. In general, however, it is unlikely
that uncertainty influences the capital stock in the short run, which
means that the result is most probably due to a measurement-related
issue, which may range from error in the measurement of the capital
stock to some statistical artifact. In either case, studies using the
investment level or investment-to-sales ratio as the dependent variable
should be preferred to those that use the investment-to-capital ratio.
Considering the uncertainty measure used, there is a dividing line
between studies that use output price uncertainty (the reference
category) or sales uncertainty and the other uncertainty measures.
Studies using other sources of uncertainty tend to be characterized by a
relatively small probability of finding a significantly negative
outcome. The marginal effects of stock price uncertainty, inflation rate
uncertainty, and exchange rate uncertainty stand out in terms of
magnitude and statistical significance; they increase the probability of
finding an insignificant effect by 20-30%. Using stock price uncertainty
furthermore substantially increases the probability of finding a
significantly positive estimate. Theory states little about the impact
of different sources of uncertainty, so the relevant question is whether
the estimates on the different sources of uncertainty represent true
underlying economic patterns or whether they in fact all represent the
same underlying generic uncertainty that is present in an economy. In
the latter case the differences in estimates may occur because some
sources capture this fundamental uncertainty better than others. Our
results do not point to either interpretation. Strikingly, however, the
three sources that stand out (stock price, inflation, and exchange rate
uncertainty) are clearly more related to macroeconomic processes than
the other sources (sales, output price, input price, and profit
uncertainty), which are more related to firm- or industry-specific
processes. Moreover, fluctuations in stock prices are to a large extent
influenced by macroeconomic fluctuations. The associated interpretation
of the results would then be that uncertainty from macroeconomic
fluctuations matters less than uncertainty from industry- or
firm-specific fluctuations. Although the latter are, of course,
influenced by processes at the macroeconomic level, it is likely that
they better represent the uncertainty that is relevant for individual
investment decisions. Further, when the impact of various sources of
uncertainty is estimated simultaneously, the probability of finding an
insignificant result increases. This result makes sense if the
uncertainty proxies are correlated, in which case including each measure
in isolation would produce, on average, more statistically significant
estimates of the relationship under investigation. Finally, the results
show that using a subjective forward-looking uncertainty measure instead
of a backward-looking uncertainty measure (which is the reference
category) has a small and statistically insignificant effect.
With respect to model specification, it has been argued that
uncertainty can be captured by Tobin's Q (the shadow value of
capital), making it unnecessary to explicitly account for uncertainty in
investment models. Our findings, however, suggest the opposite: using a
Q model actually increases the probability of finding a significantly
negative impact of uncertainty on investment spending by approximately
20%. Ultimately we do not know the true underlying investment model,
which makes the interpretation of this result ambiguous. If the true
underlying investment-uncertainty relationship is negative, the result
suggests that Tobin's Q not only fails to incorporate the full
impact of uncertainty on investment spending, but that its exclusion
from a model specification may even obfuscate a negative relationship.
If the true relationship is that uncertainty does not affect investment
spending, the result suggests that Tobin's Q increases the
probability of a type I error. Not including an accelerator variable,
wages, and capital prices in a model specification has a similar effect;
although, the effects of wages and capital prices are statistically
insignificant. Various other explanatory variables in the underlying
studies appear to be important as well. The impact of including stock
prices in the model substantially increases the probability of obtaining
a significantly negative outcome. These findings reveal potentially
important sources of bias and have clear implications for model building
in future empirical research.
Similar to differences in model specification, the level of data
aggregation is an issue that is of relevance in most meta-analyses in
economics. In our case, studies using industry- and firm-level data
produce around 30-35% fewer statistically significant negative estimates
than studies that use country-level data (the omitted category in the
analysis). Differences in investment between countries depend to a
certain extent on issues that are difficult to account for in a model,
such as institutional settings, culture-related risk attitudes, and
functioning of second-hand markets. It is therefore likely that models
based on country-level data are misspecified and produce estimates that
are off the mark. Moreover, since the effect of uncertainty on
investment spending is a microeconomic phenomenon, the results from
studies that use sector- or firm-level data are generally preferable to
those that use country-level data. In this respect, we find evidence
that industry-wide studies find more negative results than firm-specific
studies. The differences, however, are small. Insofar as these
differences reflect differences between industry-wide and firm-specific
uncertainty, the results provide some evidence for the claim made in
Pindyck (1993) that under industry-wide uncertainty the possibilities to
disinvest are smaller than under firm-specific uncertainty, leading to a
more negative relationship.
Studies that use panel and cross-section data produce more
insignificant estimates (around 20%) than studies that use time-series
data (the omitted category). A difference between these data types is
that cross-section data are better at capturing long-run effects, while
time-series data measure the short-run impact. Our findings therefore
suggest that the impact of uncertainty on investment spending is
negative in the short run but fades out as time progresses. (15)
Finally, the effects of different econometric estimators appear to
be nontrivial. OLS studies display a slightly higher probability to
produce a significantly negative result than studies that use more
sophisticated estimation techniques. A similar but much stronger effect
is found for studies that control for endogeneity, implying that not
taking this issue into account very likely produces an estimate that is
off the mark, at least in terms of sign and statistical significance.
The relationship between uncertainty and investment also appears to have
been constant over the years, as shown by the small and statistically
insignificant marginal effects on the average year of sample variable,
and there appear to be no differences between developed and less
developed countries. The latter is somewhat surprising, since one might
have expected that firms in less developed countries have more
difficulty in coping with uncertainty, for instance, because they do not
have the same possibilities to hedge against or deal with uncertainty.
Differences in specialization and functioning of second-hand markets
may, however, counteract this effect.
In Table 3 we present the ordered probit estimates and associated
marginal effects of the model that includes the publication time trend
variable (model 2). The results suggest that the odds of finding a
significantly negative estimate are higher for more recent studies (year
of publication), while the actual underlying relationship has become
more insignificant over the years (average year of sample). The latter
may reflect a true trend in the relationship, which can be explained by
better functioning of second-hand markets, which makes disinvestment of
capital less difficult, and increased possibilities for hedging. Another
explanation may be the move away from heavy manufacturing in developed
countries (the usual source of data), implying lower sunk costs as firms
are more footloose. The results may, however, also be due to studies and
data improving over the years. Arguably, we can only speculate as to the
potential causes for the publication trend. On the one hand, the
publication trend variable may mimic unobserved changes in the research
design and the data and/or econometric techniques over time. On the
other hand, we cannot preclude that the negative coefficient associated
with the publication time trend variable truly identifies publication
bias, in the sense that studies with negative estimates have in more
recent times effectively had a higher chance of being accepted for
publication. In this respect it is interesting to observe that the
publication trend matches the development of theory on the
investment-uncertainty relationship, which has focused more and more on
a negative relationship as time progressed.
The results for the base specification (Table 2) are not entirely
robust to the inclusion of the publication time trend variable. With
respect to the sources of uncertainty, the increased relevance of profit
uncertainty is notable. Using profit uncertainty now decreases the
probability of finding a significantly negative effect and also
substantially increases the probability of a significantly positive
effect. The results for some of the control variables in primary models
have changed as well, but most striking is the change in the relevance
of the level of data aggregation. The differences between country-,
industry-, and firm-level data are less substantial, not statistically
significant, and the pattern has changed. However, the results still
suggest that industry-level data produce more significantly negative
estimates than firm-level data. Finally, the pattern related to
differences in data type is similar to the pattern in model 1 but is
more pronounced, and the probability of finding a negative relationship
now appears substantially larger in developing countries than in
developed countries.
The interpretation of the changes in empirical findings is
ambiguous. As mentioned above, the publication trend variable may merely
mimic and pick up changes in research methods and data over time,
insofar as they are not captured by other explanatory variables in the
model. Alternatively, if the publication trend variable truly picks up
publication bias, the results in Table 3 better reflect the relevant
sources of underlying effect size variation than those reported in Table
2. There is no direct evidence that supports either of these
interpretations.
6. Conclusions
The impact of uncertainty on investment spending has been heavily
debated since the early 1970s. The theoretical insights developed over
the years provide an ambiguous picture as to the direction of the
effect, and many intervening factors have been suggested over time. In
this article we performed a recta-analysis to investigate whether the
body of existing empirical evidence can provide more insight into the
direction of the hypothesized relationship between investment and
uncertainty.
Because most study results were incomparable, a transformation into
elasticities was necessary. This only worked for a limited subset of the
entire sample. Furthermore, the investment-uncertainty literature is
primarily focused on the sign of the relationship and not on its
magnitude. In order to make use of the full sample of study results
while still focusing on the main issue in the literature, we decided to
define a common and scale-free metric that does not include the
magnitude of the relationship. Specifically, we focus on the direction
and statistical significance of empirical estimates, and in our
empirical analysis we estimate an ordered probit model using a
categorical variable with three categories. The coefficients are
transformed into marginal effects representing changes in the
probability of finding a significantly negative, an insignificant, or a
significantly positive study outcome. We estimate a base model as well
as an alternative model that includes a publication time trend variable.
In view of the theoretical ambiguity regarding the direction of the
relationship, our sample shows that very few studies actually find a
significantly positive estimate. The marginal effects from the ordered
probit analyses show that some study characteristics increase the
probability of observing a significantly positive estimate; although,
the effects are generally small. Ultimately, the empirical evidence
supporting a positive investment-uncertainty relationship is limited.
The marginal effects from the model without publication time trend
furthermore show that there is substantial heterogeneity in the
probabilities of observing a significantly negative and an insignificant
study outcome. Differences in the relevant sources of uncertainty, model
specifications, and data characteristics appear to be important sources
of empirical variation in the underlying studies.
Uncertainty related to macroeconomic processes, such as inflation
and exchange rate uncertainty, appears to be far less important for
investment decisions than uncertainty related to sector- or
firm-specific variation, such as uncertainty in output prices, sales,
and profits. With respect to model specification it appears that
excluding Tobin's Q and an accelerator variable, among others,
produces substantially different results. Although we do not know the
true underlying nature of the investment process, this result points to
possible misspecifications in investment models. Furthermore, studies
that use country-level data and studies that use industry- and
firm-level data produce markedly different results. Since variation in
investment between countries depends to a certain extent on issues that
are difficult to account for in a model, it is likely that the observed
differences are related to misspecifications in country-level studies.
We find some evidence for the claim made in Pindyck (1993) that under
industry-wide uncertainty the possibilities to disinvest are smaller
than under firm-specific uncertainty. Although this may reflect a
situation where there is no differential effect, it may also point to
consequences of differences in data aggregation that cannot be
disentangled from the Pindyck effect. The substantial difference between
the short-run (time series) and long-run (cross section) effect of
uncertainty is striking as well. The short-run effect appears to be
negative, but the effect fades away as time progresses.
We also included year of publication among the explanatory
variables in order to control for a possible publication time trend. The
results from this model suggest that the probability of finding a
significantly negative estimate is higher for more recent studies, while
the actual underlying relationship appears to have become more
insignificant over the years. Although most other findings remain
unchanged, some results, such as those on the level of data aggregation,
developing versus developed countries, and sources of uncertainty, are
not robust to the inclusion of the publication time trend variable. One
can obviously debate to what extent the publication time trend variable
captures unobserved underlying real-world trends, unobserved changes in
research design and data (for example, the use of microdata), or a true
publication bias caused by increased odds of publication if a study
documents a negative effect.
In conclusion, there is little empirical evidence for a positive
investment-uncertainty relationship. Our results cannot, however,
provide a definite answer to the question of whether uncertainty matters
for investment spending. In some cases the findings suggest that
potentially misspecified studies, such as those that omit an accelerator
variable and those that do not control for endogeneity, decrease the
probability of finding a significantly negative effect. In several other
cases, however, potential misspecification, such as the use of
country-level data or the use of the investment-to-capital ratio as the
dependent variable, appears to increase the probability of finding a
significantly negative effect. Moreover, if publication year does indeed
represent an actual publication time trend, possibly caused by the
development of theory over time, our results suggest that significantly
negative study outcomes were more likely to be published as the research
field matured.
Appendix
Table A1. Characteristics of the Studies Included in
the Meta-Analysis (Ordered Chronologically)
No. of
Study Observations Period Region
Dorfman and Heien 1 1970-1985 USA
(1989)
Driver and Moreton 2 1978-1987 UK
(1991)
Aizenman and Marion 8 1970-1985 LDC
(1993)
Ferderer (1993a) 15 1978-1991 USA
Ferderer (1993b) 16 1969-1989 USA
Goldberg (1993) 174 1970-1989 USA
Huizinga (1993) 73 1954-1989 USA
Pindyck and Solimano 36 1960-1990 LDC
(1993)
Serum and Solimano 4 1976-1988 LDC
(1993)
Aizenman and Marion 7 1970-1993 LDC
(1995)
Episcopos (1995) 5 1947-1992 USA
Price (1995) 3 1961-1992 UK
Bleaney (1996) 8 1980-1990 LDC
Ghosal and Loungani 35 1972-1989 USA
(1996)
Leahy and Whited 64 1981-1987 USA
(1996)
Price (1996) 3 1963-1994 UK
Bell and Campa (1997) 15 1977-1989 USA/Other
Glezakos and Nugent 4 1960-1990 USA
(1997)
Serum (1997) 16 1970-1990 LDC
Brunetti and Weder 3 1974-1989 Various
(1998)
Pattillo (1998) 3 1994-1995 Ghana
Serum (1998) 36 1970-1995 LDC
Aizenman and Marion 6 1970-1992 LDC
(1999)
Darby et al. (1999) 3 1976-1996 Various
Goel and Ram (1999) 12 1974-1992 DECD
Calcagnini and Saltari 11 1970-1995 Italy
(2000)
Ghosal and Loungani 42 1958-1991 USA
(2000)
Ogawa and Suzuki 36 1984-1993 Japan
(2000)
Goel and Ram (2001) 6 1981-1992 DECD
Peeters (2001) 50 1983-1993 Spain/
Belgium
Temple, Urga, and 16 1972-1992 UK
Driver (2001)
Bo (2002) 16 1984-1995 Netherlands
Lensink (2002) 1 1970-1997 Various
Henley, Carruth, and 8 1975-1995 UK
Dickerson (2003)
Bo and Lensink (2005) 8 1984-1996 Netherlands
Lensink, van der Steen, 21 1999 Netherlands
and Sterken (2005)
Uncertainty Aggregation
Study Measure Level Data Type
Dorfman and Heien B F C
(1989)
Driver and Moreton F I T
(1991)
Aizenman and Marion B C C
(1993)
Ferderer (1993a) F F T
Ferderer (1993b) B C T
Goldberg (1993) B I T
Huizinga (1993) B I T, C
Pindyck and Solimano B C T, C
(1993)
Serum and Solimano B C P
(1993)
Aizenman and Marion B C C
(1995)
Episcopos (1995) B C T
Price (1995) B C T
Bleaney (1996) B C C
Ghosal and Loungani B I C
(1996)
Leahy and Whited B F C
(1996)
Price (1996) B C T
Bell and Campa (1997) B I C
Glezakos and Nugent B C T
(1997)
Serum (1997) B C P
Brunetti and Weder B C C
(1998)
Pattillo (1998) F F P
Serum (1998) B C T, C, P
Aizenman and Marion B C C, P
(1999)
Darby et al. (1999) B C T
Goel and Ram (1999) B C C
Calcagnini and Saltari F C T
(2000)
Ghosal and Loungani B I C
(2000)
Ogawa and Suzuki B F P
(2000)
Goel and Ram (2001) B C T, P
Peeters (2001) B F P
Temple, Urga, and B I P
Driver (2001)
Bo (2002) B F P
Lensink (2002) B C T, P
Henley, Carruth, and B F P
Dickerson (2003)
Bo and Lensink (2005) B F C
Lensink, van der Steen, F F C
and Sterken (2005)
No. of observations: number of estimates provided by study. Period:
time period to which study pertains. Region: region or country to
which study pertains. DECD: study uses data for several DECD
countries. LDC: Data are from less developed countries. Uncertainty
measure: B=backward looking, F=forward looking. Aggregation level:
C=country, I=industry, F=firm. Data type: T=time-series data,
C=cross-section data, P=panel data.
Received June 2007; accepted September 2008.
References
(Note: Studies indicated with * have been included in the database)
Abel, Andrew B. 1983. Optimal investment under uncertainty.
American Economic Review 73:228-33.
Abel, Andrew B., and Janice C. Eberly. 1999. The effects of
irreversibility and uncertainty on capital accumulation. Journal of
Monetary Economics 44:339-77.
Abreu, Maria, Henri L. F. de Groot, and Raymond J. G. M. Florax.
2005. A meta-analysis of beta-convergence: The legendary 2%. Journal of
Economic Surveys 19:389420.
Aizenman, Joshua, and Nancy P. Marion. 1993. Macroeconomic
uncertainty and private investment. Economics Letters 41:207-10. *
Aizenman, Joshua, and Nancy P. Marion. 1995. Volatility, investment
and disappointment aversion. NBER Working Paper No. 5386. *
Aizenman, Joshua, and Nancy P. Marion. 1999. Volatility and
investment: Interpreting evidence from developing countries. Economica
66:157-79. *
Bar-Ilan, Avner, and William C. Strange. 1999. The timing and
intensity of investment. Journal of Macroeconomics 21:57-77.
Bell, Gregory K., and Jose M. Campa. 1997. Irreversible investments
and volatile markets: A study of the chemical processing industry.
Review of Economics and Statistics 79:79-87. *
Bijmolt, Tammo H. A., and Rik G. M. Pieters. 2001. Meta-analysis in
marketing when studies contain multiple measurements. Marketing Letters
12:157-59.
Bleaney, Michael F. 1996. Macroeconomic stability, investment and
growth in developing countries. Journal of Development Economics
48:461-77. *
Bo, Hong. 2001. Corporate investment under uncertainty in the
Netherlands. Ph.D. dissertation, University of Groningen: the
Netherlands,
Bo, Hong. 2002. Idiosyncratic uncertainty and firm investment.
Australian Economic Papers 41:1-14. *
Bo, Hong. 2006. An empirical examination of the hangover effect of
irreversibility on investment. Scottish Journal of Political Economy
53:358-76.
Bo, Hong, and Robert Lensink. 2005. Is the investment-uncertainty
relationship non-linear? An empirical analysis for the Netherlands.
Economica 72:307-31. *
Brunetti, Aymo, and Beatrice Weder. 1998. Investment and
institutional uncertainty: A comparative study of different uncertainty
measures. Weltwirtschafiliches Archiv 134:513-33. *
Caballero, Ricardo J. 1991. On the sign of the
investment-uncertainty relationship. American Economic Review 81:279-88.
Calcagnini, Giorgio, and Enrico Saltari. 2000. Real and financial
uncertainty and investment decisions. Journal of Macroeconomics
22:491-514. *
Carruth, Alan, Andrew P. Dickerson, and Andrew Henley. 2000. What
do we know about investment under uncertainty? Journal of Economic
Surveys 14:119-53.
Cuthbertson, Keith, and David Gasparro. 1995. Fixed investment
decisions in UK manufacturing: The importance of Tobin's Q, output
and debt. European Economic Review 39:919-41.
Darby, Julia, Andrew H. Hallett, Jonathan Ireland, and Laura
Piscitelli. 1999. The impact of exchange rate uncertainty on the level
of investment. Economic Journal 109:C5-C67. *
de Dominicis, Laura, Raymond J. G. M. Florax, and Henri L. F. de
Groot. 2008. Meta-analysis of the relationship between income inequality
and economic growth. Scottish Journal of Political Economy 55:654-82.
de Mooij, Ruud A., and Sjef Ederveen. 2003. Taxation and foreign
direct investment: A synthesis of empirical research. International Tax
and Public Finance 10:673-93.
Dixit, Avinash K., and Robert S. Pindyck. 1994. Investment under
uncertainO'. Princeton, N J: Princeton University Press.
Djankov, Simeon, and Peter Murrell. 2002. Enterprise restructuring
in transition: A quantitative survey. Journal of Economic Literature
40:739-92.
Dorfman, Jeffrey H., and Dale Heien. 1989. The effects of
uncertainty and adjustment costs on investment in the almond industry.
Review of Economics and Statistics 71:263-74. *
Driver, Ciaran, and David Moreton. 1991. The influence of
uncertainty on UK manufacturing investment. Economic Journal
101:1452-59. *
Episcopos, Athanasios. 1995. Evidence on the relationship between
uncertainty and irreversible investment. Quarterly Review of Eeonomies
and Finance 35:41-51. *
Favero, Carlo A., M. Hashem Pesaran, and Sunil Sharma. 1994. A
duration model of irreversible oil investment: Theory and empirical
evidence. Journal of Applied Econometrics 9:S95-S112.
Ferderer, J. Peter. 1993a. Does uncertainty affect investment
spending? Journal of Post Keynesian Economics 16:19-35. *
Ferderer, J. Peter. 1993b. The impact of uncertainty on aggregate
investment spending: An empirical analysis. Journal of Money, Credit,
and Banking 25:30-48. *
Ghosal, Vivek. 1991. Demand uncertainty and the capital-labor
ratio: Evidence from the U.S. manufacturing sector. Review of Economics
and Statistics 73:157-61.
Ghosal, Vivek. 1995. Input choices under price uncertainty.
Economic Inquiry 33:142-58.
Ghosal, Vivek, and Prakash Loungani. 1996. Product market
competition and the impact of price uncertainty on investment: Some
evidence from U.S. manufacturing industries. Journal of Industrial
Economics 44:217-28. *
Ghosal, Vivek, and Prakash Loungani. 2000. The differential impact
of uncertainty on investment in small and large businesses. Review of
Economics and Statistics 82:338-49. *
Glezakos, Constantine, and Jeffrey B. Nugent. 1997. Relative price
variability, inflation rate uncertainty, and postwar investment of the
United States. Journal of Post Keynesian Economics 19:181-94. *
Goel, Rajeev K., and Rati Ram. 1999. Variations in the effect of
uncertainty on different types of investment: An empirical
investigation. Australian Economic Papers 38:481-92. *
Goel, Rajeev K., and Rati Ram. 2001. Irreversibility of R&D
investment and the adverse effect of uncertainty: Evidence from the OECD
countries. Economics Letters 71:287-91. *
Goldberg, Linda S. 1993. Exchange rates and investment in United
States industry. Review of Economics and Statistics 75:575-88. *
Goldfarb, Robert S. 1995. The economist-as-audience needs a
methodology of plausible inference. Journal of Economic Methodology
2:201-22.
Green, Christopher J., Robert Lensink, and Victor Murinde. 2001.
Demand uncertainty and the capital-labour ratio in Poland. Emerging
Markets Review 2:184-97.
Greene, William H. 2003. Econometric analysis. 5th edition. Upper
Saddle River, NJ: Prentice Hall International, Inc. Hartman, Richard.
1972. The effects of price and cost uncertainty on investment. Journal
of Economic Theory 5:258-66.
Hedges, Larry V. 1992. Modeling publication selection effects in
recta-analysis. Statistical Science 7:246-55.
Henley, Andrew, Alan Carruth, and Andrew P. Dickerson. 2003.
Industry-wide versus firm-specific uncertainty and investment: British
company panel data evidence. Economics Letters 78:87-92. *
Huizinga, John. 1993. Inflation uncertainty, relative price
uncertainty, and investment in U.S. manufacturing. Journal of Money,
Credit, and Banking 25:521-57. *
Hurn, A. Stan, and Robert E. Wright. 1994. Geology or economics?
Testing models of irreversible investment using North Sea oil data.
Economic Journal 104:363-71.
Jorgenson, Dale W. 1971. Econometric studies of investment
behavior: A survey. Journal of Economic Literature 9:1111-47.
Kahneman, Daniel, and Amos Tversky. 1979. Prospect theory: An
analysis of decision under risk. Econometrica 47:263-91.
Koetse, Mark J., Henri L F. de Groot, and Raymond J. G. M. Florax.
2006. The impact of uncertainty on investment spending: A meta-analysis.
TI Discussion Paper No. 06-060/3, Amsterdam: Tinbergen Institute.
Koetse, Mark J., Arno J. van der Vlist, and Henri L F. de Groot.
2006. The impact of perceived expectations and uncertainty on firm
investment. Small Business Economics 26:365-76.
Kulatilaka, Nalin, and Enrico C. Perotti. 1998. Strategic growth
options. Management Science 44:1021-31.
Leahy, John V., and Toni M. Whited. 1996. The effect of uncertainty
on investment: Some stylized facts. Journal of Money, Credit, and
Banking 28:64-83. *
Lensink, Robert. 2002. Is the uncertainty-investment link
non-linear? Empirical evidence for developed economies.
Weltwirtschaftliches Archiv 138:131-47. *
Lensink, Robert, Paul van der Steen, and Elmer Sterken. 2005.
Uncertainty and growth of the firm. Small Business Economics 24:381-91.
*
Nakamura, Tamotsu. 1999. Risk aversion and the
uncertainty-investment relationship: A note. Journal of Economic
Behavior and Organization 38:357-63.
Nijkamp, Peter, and Jacques Poot. 2004. Meta-analysis of the effect
of fiscal policies on long-run growth. European Journal of Political
Economy 20:91-124.
Ogawa, Kazuo, and Kazuyuki Suzuki. 2000. Uncertainty and
investment: Some evidence from the panel data of Japanese manufacturing
firms. Japanese Economic Review 51:170-92. *
Pattillo, Catherine. 1998. Investment, uncertainty, and
irreversibility in Ghana. IMF Staff Papers 45:522 53. *
Peeters, Marga. 2001. Does demand and price uncertainty affect
Belgian and Spanish corporate investment? Louvain Economic Review 67:119
39. *
Pindyck, Robert S. 1982. Adjustment costs, uncertainty, and the
behavior of the firm. American Economic Review 72:415-27.
Pindyck, Robert S. 1991. Irreversibility, uncertainty, and
investment. Journal of Economic Literature 29:1110-48.
Pindyck, Robert S. 1993. A note on competitive investment under
uncertainty. American Economic Review 83:273-77.
Pindyck, Robert S., and Andres Solimano. 1993. Economic instability
and aggregate investment. NBER Macroeconomics Annual 8:259-303. *
Poot, Jacques. 2000. A synthesis of empirical research on the
impact of government on long-run growth. Growth and Change 31:516-46.
Price, Simon. 1995. Aggregate uncertainty, capacity utilization and
manufacturing investment. Applied Economics 27:147-54. *
Price, Simon. 1996. Aggregate uncertainty, investment and
asymmetric adjustment in the UK manufacturing sector. Applied Economics
28:1369 79. *
Roberts, Colin J., and Tom D. Stanley. 2005. Meta-regression
analysis: Issues of publication bias in economics. Oxford: Blackwell.
Rose, Andrew K., and Tom D. Stanley. 2005. A meta-analysis of the
effect of common currencies on international trade. Journal of Economic
Surveys 19:347-65.
Rosenberger, Randall S., and Tom D. Stanley. 2006. Measurement,
generalization, and publication: Sources of error in benefit transfers
and their management. Ecological Economics 60:372-78.
Saltari, Enrico, and Davide Ticchi. 2005. Risk aversion and the
investment-uncertainty relationship: A comment. Journal of Economic
Behavior and Organization 56:121-25.
Serven, Luis. 1997. Irreversibility, uncertainty, and private
investment: Analytical issues and some lessons for Africa. Journal of
African Economies 6:229-8. *
Serven, Luis. 1998. Macroeconomic uncertainty and private
investment in developing countries: An empirical investigation. Policy
Research Working Paper No. 2035. Washington D.C.: The World Bank. *
Serven, Luis, and Andres Solimano. 1993. Debt crisis, adjustment
policies and capital formation in developing countries: Where do we
stand? Worm Development 21:127-40. *
Stanley, Tom D. 1998. New wine in old bottles: A meta-analysis of
Ricardian equivalence. Southern Economic Journal 64:713-27.
Stanley, Tom D. 2001. Wheat from chaff: Meta-analysis as
quantitative literature review. Journal of Economic Perspectives
15:131-50.
Stanley, Tom D. 2008. Meta-regression methods for detecting and
estimating empirical effects in the presence of publication selection.
Oxford Bulletin of Economics and Statistics 70:103-27.
Temple, Paul, Giovanni Urga, and Ciaran Driver. 2001. The influence
of uncertainty on investment in the UK: A macro or micro phenomenon?
Scottish Journal of Political Economy 48:361-82. *
Tobin, James. 1969. A general equilibrium approach to monetary
theory. Journal of Money, Credit, and Banking 1:15-29.
Weichselbaumer, Doris, and Rudolf winter-Ebmer. 2005. A
meta-analysis of the international gender wage gap. Journal of Economic
Surveys 19:479-511.
(1) For theoretical studies that try to unify the timing and level
effects of uncertainty, see Abel and Eberly (1999) and Bar-Ilan and
Strange (1999). The former argue that, under irreversibility, there is a
so-called hangover effect. This results from the fact that when the
marginal revenue product of capital is low, a firm would like to sell
some of its already installed capital but cannot do so because
investment is irreversible. In this case, the desired stock of capital
is lower than the actual stock. Since under reversible investment the
actual capital stock can be instantaneously adjusted to the desired
level, the actual capital stock is higher under irreversibility than
under reversibility when capital productivity is low (see Bo 2006 for an
empirical analysis). Under uncertainty this effect is reinforced.
(2) For theoretical contributions on the role of risk aversion see
Nakamura (1999) and Saltari and Ticchi (2005). Other factors that may
affect the direction and magnitude of the relationship are underlying
market structure (Hartman 1972; Abel 1983; Caballero 1991; Kulatilaka
and Perotti 1998), the discrepancy between industry-level and
firm-specific idiosyncratic uncertainty (Pindyck 1993), and financial
conditions of the firm (Ghosal and Loungani 2000; Peeters 2001).
(3) For recent developments in the literature see, for example,
Roberts and Stanley (2005). Good examples of empirical applications of
meta-analysis in macroeconomics are Stanley (1998); Poot (2000); Djankov
and Murrell (2002); de Mooij and Ederveen (2003); Nijkamp and Poot
(2004); Abreu, de Groot, and Florax (2005); Rose and Stanley (2005);
Weichselbaumer and Winter-Ebmer (2005); and de Dominicis, Florax, and de
Groot (2008).
(4) One of the arguments for such a pattern is potential
risk-seeking behavior of economic agents over the domain of small losses
(Kahneman and Tversky 1979). An increase in uncertainty implies an
increase in the trigger value of investment, but it also increases the
probability of hitting this trigger value. Although it is generally
assumed that the former effect dominates, the reverse may be true for
low levels of uncertainty. Another possibility is that firms react
differently to positive and negative shocks, where the inverted U-curve
stems from the notion that negative shocks are generally associated with
high uncertainty (Bo 2001. p. 100).
(5) The main conclusion from these studies is that uncertainty
indeed has a positive effect on investment spending for low levels of
uncertainty and a negative effect for high levels, thereby providing
evidence for a nonlinear investment-uncertainty relationship.
(6) For empirical analyses on the differential impact of different
sources of uncertainty see Huizinga (1993) and Koetse, van der Vlist,
and de Groot (2006).
(7) See Driver and Moreton (1991) and Ferderer (1993a) for
alternative measurements of subjective uncertainty.
(8) See Bo (2001), among others, for an extensive discussion and
empirical investigation of this issue.
(9) The trend variable is rescaled and equals unity for 1970.
(10) A disadvantage of using a categorical variable is that it
discards some of the information on the statistical significance of the
effect. An alternative approach that does make use of all the available
information is a regression on z values. This approach also allows for
the calculation of marginal effects, so the results are comparable to
those obtained using an ordered probit analysis. The patterns and
findings from this analysis are very similar to those presented here.
The most striking difference is that the marginal effects for the third
category, that is, significantly positive estimates, are substantially
smaller (for details see Koetse. de Groot, and Florax 2006).
(11) It is of course possible to distinguish between four or more
categories. For example, we could make a further distinction between
insignificant negative and insignificant positive study outcomes, but we
could also use sample quartiles or deciles. However, for reasons of
tractability and exposition we choose to work with three categories.
(12) A detailed description of the calculation of standard errors
is available upon request from the authors.
(13) The publication trend variable is rescaled and equals unity
for 1989.
(14) In a variant not included here for reasons of space, we also
controlled for differences in publication quality (see Rosenberger and
Stanley 2006). Creating a continuous measure based on impact factors was
not possible since impact factors are not available for working papers
and some of the journals. Furthermore, there are only two working papers
in our sample, making a distinction between working papers and journal
articles unviable. We therefore created dummy variables based on a
categorical ranking of journals used by the Tinbergen Institute (see
www.tinbergen.nl). We distinguish between A, B, and C publications,
where the C category consists of working papers and journals that are
not included in the A or B category. The results turn out not to be
strictly increasing or decreasing in terms of quality of the publication
outlet: A publications produce more insignificant estimates than C
publications, while B publications produce more significantly negative
estimates. The results documented in the main text are, however, robust
to the inclusion of the publication quality dummy variables. Results for
this alternative specification are available upon request from the
authors.
(15) Although one would expect panel data to pick up only some of
the long-run effects of uncertainty, the results show that the effect of
panel data is actually stronger than that of cross-section data.
However, the difference is very small, and statistically we cannot
reject the hypothesis that the coefficients are equal at the usual
significance levels. This holds for the marginal effects as well.
Mark J. Koetse, VU University Amsterdam, Department of Spatial
Economics, De Boelelaan 1105, 1081 HV, Amsterdam, the Netherlands;
E-mail mkoetse@feweb.vu.nl; corresponding author.
Henri L.F. de Groot, University Amsterdam and Tinbergen Institute,
Department of Spatial Economics, De Boelelaan 1105, 1081 HV, Amsterdam,
the Netherlands; E-mail hgroot@feweb.vu.nl.
Raymond J.G.M. Florax, Purdue University. Department of
Agricultural Economics, 403 West State Street, West Lafayette, IN 47907,
USA, and VU University Amsterdam, Department of Spatial Economics;
E-mail rflorax@purdue.edu.
This research was supported through the program "Stimulating
the Adoption of Energy-Efficient Technologies," funded by the
Netherlands Organization for Scientific Research (NWO) and the Dutch
Ministry of Economic Affairs (SenterNovem). We are grateful to Tom
Stanley, John Pepper, and two anonymous referees for useful comments on
an earlier version of this article. The usual disclaimer applies.
Table 1. Descriptive Statistics on Sign of the Investment-Uncertainty
Estimates and Their Statistical Significance Using a 5% Critical
Significance Level (N = 767)
Percentage Percentage
Sign Significance Count (%) Count (%)
Negative Significant 225 30 505 66
Insignificant 280 37
Positive Insignificant 232 30 262 34
Significant 30 4
Total 767 100 767 100
Table 2. Estimates and Associated Marginal Effects of Meta-Analysis
Ordered Probit Model 1 (Standard Errors in Parentheses)
Marginal Effects
Model 1
Ordered Probit Significantly
Model 1 Negative
Constant 0.142 --
(0.379)
Investment measures
Investment-to-sales ratio (a) -0.151 0.056
(0.230) (0.087)
Investment-to-capital ratio (a) -0.791 ** 0.295 **
(0.223) (0.081)
Sources and measures of uncertainty
Input price uncertainty (b) 0.402 -0.136
(0.326) (0.099)
Sales uncertainty (b) -0.019 0.007
(0.285) (0.106)
Stock price uncertainty (b) 1.077 ** -0.302 **
(0.302) (0.057)
Profit uncertainty (b) 0.490 -0.162
(0.362) (0.104)
Inflation rate uncertainty (b) 0.936 ** -0.271 **
(0.310) (0.062)
Exchange rate uncertainty (b) 0.660 -0.226
(0.336) (0.103)
Other uncertainty sources (b) 0.422 -0.143
(0.319) (0.097)
Joint estimation (c) 0.876 ** -0.309 **
(0.170) (0.054)
Subjective uncertainty (d) 0.371 -0.127
(0.328) (0.102)
Control variables in primary models (e)
Tobin's Q included -0.569 * 0.221 *
(0.229) (0.090)
Accelerator variable included -0.573 ** 0.195 **
(0.153) (0.047)
Wages included -0.531 0.207
(0.323) (0.128)
Capital price included -0.058 0.021
(0.230) (0.087)
Time trend included 0.454 * -0.161 *
(0.193) (0.065)
Debt position included 0.639 ** -0.204 **
(0.219) (0.059)
Stock price included -1.434 * 0.505 **
(0.596) (0.141)
Size of firm included 0.167 -0.059
(0.311) (0.107)
Government expenditures included -0.387 0.150
(0.287) (0.114)
Dependent lag included -0.348 * 0.131 *
(0.167) (0.064)
Trade flows included -0.048 0.018
(0.261) (0.097)
Data characteristics-aggregation level
Industry-level data (f) 0.854 ** -0.304 **
(0.231) (0.075)
Firm-level data (f) 1.130 ** -0.358 **
(0.261) (0.065)
Data characteristics-time series vs. cross section
Panel data (g) 0.728 ** -0.240 **
(0.261) (0.076)
Cross-section data (g) 0.672 ** -0.234 **
(0.238) (0.078)
Spatial, temporal, and econometric issues
Average year of sample -0.025 0.009
(0.018) (0.007)
Year of publication -- --
Less developed countries (h) 0.069 -0.025
(0.182) (0.066)
OLS estimation (i) -0.310 * 0.109 *
(0.156) (0.053)
Instrumental variables -0.723 ** 0.263 **
estimation (j) (0.189) (0.068)
Number of observations 767
Prob (chi-squared) 0.000
Log-likelihood -516.8
Log-likelihood restricted -578.1
Marginal Effects Model 1
Significantly
Insignificant Positive
Constant -- --
Investment measures
Investment-to-sales ratio (a) -0.049 -0.007
(0.077) (0.011)
Investment-to-capital ratio (a) -0.257 ** -0.038 **
(0.072) (0.012)
Sources and measures of uncertainty
Input price uncertainty (b) 0.106 0.030
(0.066) (0.034)
Sales uncertainty (b) -0.006 -0.001
(0.091) (0.015)
Stock price uncertainty (b) 0.170 ** 0.132 *
(0.034) (0.066)
Profit uncertainty (b) 0.123 0.040
(0.064) (0.042)
Inflation rate uncertainty (b) 0.163 ** 0.107
(0.028) (0.062)
Exchange rate uncertainty (b) 0.178 0.048
(0.072) (0.034)
Other uncertainty sources (b) 0.112 0.032
(0.066) (0.032)
Joint estimation (c) 0.253 ** 0.056 **
(0.041) (0.017)
Subjective uncertainty (d) 0.099 0.027
(0.072) (0.031)
Control variables in primary models (e)
Tobin's Q included -0.202 * -0.019 **
(0.085) (0.006)
Accelerator variable included -0.152 ** -0.043 *
(0.033) (0.017)
Wages included -0.189 -0.018 *
(0.121) (0.007)
Capital price included -0.019 -0.003
(0.075) (0.011)
Time trend included 0.131 * 0.029
(0.051) (0.015)
Debt position included 0.146 ** 0.058
(0.034) (0.030)
Stock price included -0.482 ** -0.023 **
(0.139) (0.004)
Size of firm included 0.049 0.010
(0.084) (0.022)
Government expenditures included -0.135 -0.015
(0.106) (0.008)
Dependent lag included -0.114 * -0.017 *
(0.056) (0.008)
Trade flows included -0.015 -0.003
(0.084) (0.013)
Data characteristics-aggregation level
Industry-level data (f) 0.252 ** 0.052 *
(0.058) (0.020)
Firm-level data (f) 0.254 ** 0.104 *
(0.036) (0.042)
Data characteristics-time series vs. cross section
Panel data (g) 0.180 * 0.060
(0.048) (0.031)
Cross-section data (g) 0.188 ** 0.046 *
(0.059) (0.021)
Spatial, temporal, and econometric issues
Average year of sample -0.008 -0.001
(0.006) (0.001)
Year of publication -- --
Less developed countries (h) 0.022 0.004
(0.055) (0.011)
OLS estimation (i) -0.089 * -0.020
(0.041) (0.012)
Instrumental variables -0.223 ** -0.040 **
estimation (j) (0.058) (0.013)
Number of observations
Prob (chi-squared)
Log-likelihood
Log-likelihood restricted
(a) Reference category: Unsealed investment.
(b) Reference category: Output price uncertainty.
(c) Reference category: No joint estimation.
(d) Reference category: Objective or backward-looking
uncertainty.
(e) Reference category: Explanatory variable is excluded
from the primary study model.
(f) Reference category: Country-level data.
(g) Reference category: Time-series data.
(h) Reference category: Developed countries.
(i) Reference category: Estimation other than OLS.
(j) Reference category: No instrumental variables estimation.
* p < 0.05.
** p < 0.01.
Table 3. Estimates and Associated Marginal Effects of Meta-Analysis
Ordered Profit Model 2 (Standard Errors in Parentheses) (a)
Marginal Effects
Model 2
Ordered Probit Significantly
Model 2 Negative
Constant 0.962 * --
(0.406)
Investment measures
Investment-to-sales ratio -0.050 0.018
(0.240) (0.086)
Investment-to-capital ratio -0.789 ** 0.284 **
(0.230) (0.082)
Sources and measures of uncertainty
Input price uncertainty 0.439 -0.138
(0.340) (0.092)
Sales uncertainty -0.244 0.089
(0.295) (0.111)
Stock price uncertainty 1.311 ** -0.312 **
(0.315) (0.045)
Profit uncertainty 1.876 ** -0.350 **
(0.413) (0.035)
Inflation rate uncertainty 0.582 -0.176 *
(0.323) (0.079)
Exchange rate uncertainty 0.298 -0.102
(0.348) (0.114)
Other uncertainty sources -0.002 0.001
(0.334) (0.118)
Joint estimation 0.926 ** -0.310 **
(0.175) (0.052)
Subjective uncertainty 0.199 -0.067
(0.339) (0.109)
Control variables in primary models
Tobin's Q included -0.530 * 0.202 *
(0.237) (0.094)
Accelerator variable included -0.304 0.102 *
(0.161) (0.051)
Wages included -0.363 0.136
(0.336) (0.132)
Capital price included -0.397 0.149
(0.238) (0.093)
Time trend included 0.704 ** -0.228 **
(0.204) (0.061)
Debt position included 0.488 * -0.152 *
(0.226) (0.062)
Stock price included -1.430 * 0.515 **
(0.595) (0.155)
Size of firm included -0.117 0.042
(0.323) (0.120)
Government expenditures included -1.055 ** 0.402 **
(0.309) (0.108)
Dependent lag included 0.165 -0.057
(0.184) (0.063)
Trade flows included 0.487 -0.162
(0.280) (0.085)
Data characteristics-aggregation level
Industry-level data -0.147 0.052
(0.268) (0.095)
Firm-level data 0.306 -0.104
(0.286) (0.093)
Data characteristics-time series vs. cross section
Panel data 1.379 ** -0.367 **
(0.280) (0.056)
Cross-section data 1.084 ** -0.338 **
(0.250) (0.069)
Spatial, temporal, and econometric issues
Average year of sample 0.102 ** -0.036 **
(0.024) (0.008)
Year of publication -0.249 ** 0.088 **
(0.031) (0.011)
Less developed countries -0.611 ** 0.232 **
(0.204) (0.080)
OLS estimation -0.571 ** 0.179 **
(0.164) (0.046)
Instrumental variables estimation -0.649 ** 0.227 **
(0.191) (0.067)
Number of observations 767
Prob (chi-squared) 0.000
Log-likelihood -481.9
Log-likelihood restricted -578.1
Marginal Effects Model 2
Significantly
Insignificant Positive
Constant -- --
Investment measures
Investment-to-sales ratio -0.016 -0.002
(0.077) (0.009)
Investment-to-capital ratio -0.256 ** -0.028 **
(0.075) (0.010)
Sources and measures of uncertainty
Input price uncertainty 0.112 0.026
(0.064) (0.029)
Sales uncertainty -0.081 -0.008
(0.103) (0.008)
Stock price uncertainty 0.159 ** 0.153 *
(0.053) (0.074)
Profit uncertainty 0.023 0.326 *
(0.134) (0.146)
Inflation rate uncertainty 0.137 ** 0.039
(0.048) (0.034)
Exchange rate uncertainty 0.088 0.014
(0.096) (0.019)
Other uncertainty sources -0.001 0.000
(0.105) (0.013)
Joint estimation 0.264 * 0.046 **
(0.042) (0.015)
Subjective uncertainty 0.058 0.009
(0.090) (0.019)
Control variables in primary models
Tobin's Q included -0.189 * -0.013 **
(0.090) (0.004)
Accelerator variable included -0.088 * -0.014
(0.043) (0.010)
Wages included -0.126 -0.010
(0.125) (0.007)
Capital price included -0.137 -0.011'
(0.088) (0.005)
Time trend included 0.189 ** 0.039 *
(0.048) (0.016)
Debt position included 0.123 ** 0.030
(0.044) (0.020)
Stock price included -0.500 ** -0.016 **
(0.154) (0.003)
Size of firm included -0.038 -0.004
(0.110) (0.010)
Government expenditures included -0.385 ** -0.017 **
(0.106) (0.004)
Dependent lag included 0.050 0.007
(0.055) (0.008)
Trade flows included 0.137 * 0.024
(0.068) (0.019)
Data characteristics-aggregation level
Industry-level data -0.046 -0.006
(0.085) (0.010)
Firm-level data 0.090 0.014
(0.078) (0.016)
Data characteristics-time series vs. cross section
Panel data 0.234 ** 0.133 *
(0.039) (0.052)
Cross-section data 0.269 ** 0.069 **
(0.051) (0.026)
Spatial, temporal, and econometric issues
Average year of sample 0.032 ** 0.004
(0.008) (0.004)
Year of publication -0.078 ** -0.010
(0.010) (0.009)
Less developed countries -0.216 ** -0.015 **
(0.077) (0.004
OLS estimation -0.144 ** -0.035 *
(0.035) (0.015)
Instrumental variables estimation -0.200 ** -0.027 **
(0.060) (0.009)
Number of observations
Prob (chi-squared)
Log-likelihood
Log-likelihood restricted
(a) See the notes to Table 2.
* p < 0.05.
** p < 0.01.
