POLICY RESEARCH WORKING PAPER 2035
Macroeconomic Uncertainty Economic uncertainty can
affect investment through
and Private Investment different channels, some of
in Developing Countries which operate in mutually
opposing directions, so the
sign of its overall effect is
An Empirical Investigation ambiguous and can be
assessed only empirically. This
paper presents a thorough
Luis Serven empirical assessment of the
investment-uncertainty link in
developing countries.
The World Bank
Development Research Group
Macroeconomics and Growth
December 1998
I POLICY RESEARCH WORKING PAPER 2035
Summary findings
The impact of uncertainty on investment has attracted He then adds these constructed measures to an
considerable attention in the analytical and empirical empirical investment equation that is estimated using
macroeconomic literature. In theory, however, alternative panel data econometric methods, allowing for
uncertainty can affect investment through different simultaneity, country-specific effects, and parameter
channels, some of which operate in mutually opposing heterogeneity across countries.
directions. So the sign of its overall effect is ambiguous The results underscore the robustness of the link
and can be assessed only empirically. between investment and uncertainty. Virtually all of the
To thoroughly assess the impact of macroeconomic volatility measures in the paper show a strong negative
uncertainty on private investment, Serven uses a large association with investment ratios. In addition, the
panel data set on developing countries. He draws a regression estimates indicate that uncertainty has an
distinction between sample variability and uncertainty, adverse direct impact on investment, over and above any
constructs alternative measures of the volatility of indirect effects that might also be at work. This finding is
innovations to five key macroeconomic variables particularly robust in the case of real exchange rate
(inflation, growth, the terms of trade, the real exchange volatility, which invariably has a robust negative effect
rate, and the price of capital goods), and examines their on investment, regardless of econometric specification.
association with aggregate private investment.
This paper - a product of Macroeconomics and Growth, Development Research Group - is part of a larger effort in the
group to understand the determinants of private investment decisions. Copies of the paper are available free from the World
Bank, 1818 H Street NW, Washington, DC 20433. Please contact Hazel Vargas, room 18-138, telephone 202-473 -8546,
fax 202-522-2119, Internet address hvargas@worldbank.org. The author may be contacted at Iserven@worldbank.org.
December 1998. (33 pages)
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about
development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The
papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this
paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the
countries they represent.
Produced by the Policy Research Dissemination Center
Macroeconomic Uncertainty and
Private Investment in LDCs:
An Empirical Investigation
Luis Serven*
The World Bank
1818 H St NW
Washington DC 20433
e-mail: lserven@worldbank.org
* A preliminary version of this paper was presented at the LACEA meetings in BogotA, October 1997. 1 am grateful
to conference participants for useful comments; to Michael Gavin, Norman Loayza and Fabio Schiantarelli for
helpful discussions; and to Charles Chang for excellent research assistance.
1. Introduction
The impact of uncertainty and instability on investment, a topic of obvious concern for
policy-makers, has attracted considerable interest in the analytical and empirical literature. From
the analytical perspective, theoretical work has pointed out a number of different channels
through which uncertainty can impact on investment, under various assumptions about risk
aversion, adjustment costs to investment and other factors (see Caballero, 1991 and Abel and
Eberly, 1994). Some of these effects operate in mutually opposing directions, however, and
hence the sign of the investment-uncertainty relationship is indeterminate on a priori grounds.
Relative to the analytical literature, empirical studies on uncertainty and investment are
less abundant, and mostly confined to a few single-country studies focusing on the U.S. and
U.K., plus a handful of cross-country papers. On the whole, they are not conclusive in their
assessment of the impact of uncertainty on investment, although the majority does find a
negative association between both variables. In most cases, however, these studies use naive
measures of sample variability rather than uncertainty, often ignore important investment
determinants, and/or fail to account for the likely simultaneity between investment and its
determinants-including uncertainty itself As a result, in many cases it is not entirely clear
whether any empirical association between uncertainty an investment reported by these studies
operates directly or through third-variables relevant for investment (such as output growth or the
cost of capital), and whether such association can be given a causal interpretation.
This paper provides a thorough empirical re-examination of the link between
macroeconomic uncertainty and aggregate private investment using a large set of cross-country
time-series macroeconomic data for developing countries. The paper makes an attempt at
separating simple variability from uncertainty, by constructing altemative measures of the
dispersion of the innovations to a number of relevant macroeconomic variables; lays out the
stylized facts concerning their empirical association with investment; and assesses their impact
on private investment performance using an econometric framework encompassing other
conventional determinants of investment, and applying a range of estimation techniques that
control for issues such as simultaneity, country-specific effects and parameter heterogeneity.
2
The paper is organized as follows. Section 2 presents a brief overview of the analytical
and empirical literature on the link between uncertainty and investment. Section 3 assesses the
stylized facts, looking at the co-variation between aggregate private investment and measures of
macroeconomic uncertainty in developing countries using the panel data set mentioned above.
Section 4 goes one step beyond and examines the investment-uncertainty relationship controlling
for other standard determinants of private investment, through estimation of an empirical private
investment equation using various econometric techniques. Finally, section 5 concludes.
2. Investment, Uncertainty and Instability: A Brief Overview
Economic theory has paid considerable attention to the relationship between uncertainty
and investment. Taken as whole, however, the theoretical predictions are ambiguous.
Depending on their underlying assumptions, some approaches predict a positive relationship,
while others predict a negative one.
Much of the theoretical work on uncertainty and investment has been developed in the
framework of risk-neutrality. The impact of uncertainty in standard models of risk-neutral
investors-firms depends basically on the relationship between the expected marginal revenue
product of capital and the uncertain variable(s)-typically the output price or the real wage.
Consider, for example, the familiar scenario of the constant-returns, perfectly competitive firm in
which capital is the only fixed factor, while other productive inputs (e.g., labor) can be costlessly
adjusted in the face of changing output prices. Price shocks then cause the firm to alter the
optimal capital/labor mix, thus making the marginal revenue product of capital rise more (or fall
less) than relative output prices. In such conditions, marginal profitability is a convex function
of output prices, and Jensen's inequality then implies that higher price uncertainty raises the
expected profitability of capital, thereby increasing the desired capital stock and hence
investment (Hartman, 1972; Abel, 1983).
A recent, but rapidly growing, literature has shifted the analytical focus to the adjustment
costs implied by the acquisition and installation of capital, emphasizing the irreversible nature of
most fixed investment projects (see Dixit and Pindyck, 1994 for a thorough discussion). This in
effect makes investment adjustment costs asymmetric-larger for downward than for upward
3
adjustment. Under appropriate conditions, this creates a range of inaction: investrnent takes
place only when expected profitability exceeds a certain threshold.
From the preceding discussion, however, it should be clear that irreversibility per se is
not sufficient to turn around the positive impact of uncertainty on investment following from the
convexity of the profit function. Indeed, even under asymmetric adjustment costs it can be
shown that optimal investment by a competitive firm continues to be a non-decreasing function
of uncertainty (Caballero, 1991; Abel and Eberly, 1994). To reverse this result, it is necessary to
bring in additional assumptions such as imperfect competition or decreasing returns to scale (or
both). When combined with irreversibility, they can create a negative uncertainty-investment
link by making the marginal revenue product of capital a decreasing function of the capital stock
(Caballero, 1991). Under such conditions, the profitability threshold mentioned above rises with
the extent of uncertainty, and if this effect is powerful enough it may outweigh the rise in
expected profitability stemming from the convexity of the profit function, leading to reduced
investment. The intuitive reason is that in this case the asymmetric nature of adjustment costs
makes downside uncertainty more important than upside uncertainty: since disinvestment is
costlier than investment, favorable shocks have a smaller effect on profitability than adverse
shocks, and investors/firms become ex ante reluctant to invest to reduce the risk of being stuck ex
post with unprofitable irreversible projects.
The assumption that the marginal profitability of capital declines with the capital stock
obviously cannot apply to a constant-returns perfectly competitive firm, for which the marginal
profitability of capital is, by construction, unrelated to the level of capital. However, such
assumption does hold for imperfectly competitive firms and, perhaps more importantly, also for a
free-entry perfectly competitive industry taken as a whole. At this industry-wide level, the
distinction between aggregate and firm-specific uncertainty becomes important: the latter
continues to have the positive (or non-negative) effects on firm investment described above, but
aggregate shocks have an asymmetric effect, because the industry-wide impact on profitability of
favorable aggregate shocks is limited by the entry of new firms, while irreversibility prevents
exit when unfavorable shocks occur. As individual perfectly competitive firms are aware of this
fact, higher aggregate uncertainty unambiguously raises their profitability threshold and tends to
reduce firm as well as industry-wide investment (Caballero and Pindyck, 1996).
4
Even in this case, however, the theoretical predictions of the irreversibility approach
concern the ex-ante investment decision, and hence refer mostly to the short-term. In the longer
term, a "hangover effect" (Abel and Eberly, 1995a,b) comes into play: higher degrees of
irreversibility and/or uncertainty make it more likely that firms will ex-post find themselves stuck
with excessive capital, in effect making the long-run capital stock and investment higher than
they would have been otherwise.
It may be worth noting also that factors analogous to irreversibility can in fact contribute
to apositive investment-uncertainty link, at least in theory. For example, if firms have the option
of abandoning unprofitable projects, investment adjustment costs also become asymmetric, but in
the opposite direction to that implied by irreversibility: with the possibility of abandonment,
upside uncertainty is the one that matters, since the impact of adverse shocks can be mitigated by
shutting down those projects revealed ex-post as unprofitable, and hence higher uncertainty
actually hastens investment-provided the cost of abandoning projects is not too high (Bar-Ilan
and Strange, 1996).
So far the discussion has been limited to the risk-neutral case. An alternative approach
takes as starting point the case of risk-averse agents facing limited diversification possibilities in
the context of imperfect capital markets. Along these lines, Zeira (1990) presents a model of
optimal capital accumulation by perfectly competitive risk-averse investors/firms facing
uncertain relative prices. In this framework, uncertainty has an ambiguous impact on
investment: on the one hand, it tends to raise investment through the convexity of the profit
function mentioned earlier; on the other hand, higher uncertainty discourages investment due to
investors' risk aversion. The net effect depends on the concavity of the utility function
(describing the extent of investors' risk aversion), the convexity of the profit function, and the
distribution of risk.
Both tl - risk-neutral and risk-averse cases above fit into the standard neoclassical
expected utility model. A third approach is that of 'disappointment aversion,' which has been
recently advocated by Aizenman and Marion (1995) and departs from the expected utility
paradigm by assuming that agents' preferences attach more weight to adverse outcomes than to
favorable ones - a notion first developed by Gul (1991) that has had some success in explaining
individuals' behavior in certain controlled experiments. Aizenman and Marion show that, under
5
appropriate assumptions concerning the degree of disappointmnent aversion, this approach can
yield a considerably stronger negative impact of uncertainty on investment than that suggested
by the conventional expected-utility framework.
In turn, the empirical literature on the relationship between uncertainty and investment is
still considerably more scarce than its theoretical counterpart. Most empirical studies,
particularly those using macroeconomic data, adopt a non-structural approach, in which various
uncertainty proxies are appended to otherwise conventional reduced-form investment equations.'
Along these lines, a few studies have examined the impact of uncertainty on U.S. and
U.K. investment. Federer (1993) finds a negative effect on U.S. equipment investment, while
Driver and Moreton (1991) and Price (1995, 1996) find a likewise negative effect on U.K.
manufacturing investment. In turn, Goldberg (1993) explores the impact of real exchange rate
uncertainty on U.S. industry-level investment. She finds basically no effects at the aggregate
level, while at the subsector level her results vary in sign and significance.
Cross-country empirical studies using aggregate data are somewhat more abundant.
Hausmann and Gavin (1995) report a negative association between an index of macroeconomic
volatility-which combines real GDP and real exchange rate volatility-and the aggregate
investment/GDP ratio, using a large sample of developing countries. By contrast, Bleaney
(1996) finds that measures of volatility (including for example the variability of the real
exchange rate) affect adversely growth performance in developing countries, but not aggregate
investment. A similar result is obtained by Ramey and Ramey (1995), who also use aggregate
investment data.
In turn, Aizenman and Marion (1995, 1996) report a negative cross-country correlation
between various indicators of economic instability (such as the volatility of the terms of trade,
inflation and the real exchange rate) and private investment. They further show that these
I Several studies using firm-level data adopt a similar perspective. For example, a recent paper by Bell and Campa
(1997) examines the impact of volatility on investment in a data set comprising US and European firms in the
chemical sector, with volatility measured by the standard deviations of the rates of change of the real exchange rate,
the relative price of oil and overall market capacity. Somewhat surprisingly, the results indicate that the first of
these factors has a negative impact on investment, while the other two have a positive effect. In tum, Ghosal and
Loungani (1996) estimate the impact of output price uncertainty on investment using panel data for US
manufacturing firms. They find a negative effect in those industries characterized by a higher degree of product
market competition. Finally, Guiso and Parigi (1998) examine the effects of future demand volatility on a cross-
section of Italian firms, and they find a consistent negative impact of demand uncertainty on firm investment.
6
volatility measures contribute significantly to explain the performance of private investment
across countries in a reduced-form regression framework. Total (i.e., private plus public)
investment, however, is unrelated to instability indicators in their sample.
A more structural approach is followed by Bertola and Caballero (1994), who implement
empirically a model that follows explicitly from the aggregation of individual firms' irreversible
investment rules; Caballero (1993) applies a similar approach to developing country data. These
two studies illustrate the asymmetric response of aggregate investment to positive and negative
shocks, and its strong dependence on initial conditions: after a deep recession, for example,
many firms are likely to be well below their investment thresholds, and therefore the
responsiveness of aggregate investment to positive incentive changes can be very limited.
Most of the cross-country empirical studies are confined to the cross-section dimension,
ignoring the time-series variation in the data. There are a few exceptions, however. Serven and
Solimano (1993) estimate a private investment equation using panel data on a group of
developing countries including as regressors the standard deviations of inflation and the real
exchange rate; they find some evidence that these measures of variability affect investment
negatively. In turn, Pindyck and Solimano (1993) test for the effects of uncertainty on aggregate
investment following the irreversibility approach. Using panel data for industrial and developing
countries, they construct proxies for the profitability threshold, and examine its relation with the
volatility of profitability itself. They also estimate reduced-form investment regressions
including volatility indicators such as the standard deviations of inflation and the real exchange
rate; the latter is found to have negative impact on investment.2 More recently, Darby, et al.
(1998) estimate simple investment specifications, extended to include a measure of real exchange
rate variability, separately for five OECD economies, and they find a consistently negative effect,
either in the short or the long run or both.
It is worth noting two general problems affecting many of these reduced-form empirical
studies. The first, most basic difficulty concerns the very measurement of uncertainty, which in
2 By contrast, the standard deviation of the inflation rate has no significant effect. Instead, Pindyck and Solimano
find that the inflation rate itself is negatively associated with investment. However, their ensuing interpretation of
the inflation level as a measure of uncertainty is disputed by Eberly (1993).
7
most studies is proxied by sample variability.3 Yet it is quite obvious that variability does not
amount to uncertainty, except when events are unpredictable, and therefore more accurate
measures of uncertainty would be provided by the dispersion of the innovations to the variables
of interest.
The second problem has to do with the interpretation of the empirical link between
uncertainty and investment reported by several studies. In many cases they ignore other
important investment determinants such as real interest rates and/or capital goods prices, which
leaves unanswered the question whether uncertainty has an independent effect on investment-
i.e., if it matters for investment decisions once other standard determinants of capital
accumulation are accounted for. Further, because in a macroeconomic equilibrium the variables
summarizing volatility (typically second moments of inflation and/or relative prices) are
determined simultaneously with investment, it is often unclear whether the empirical results of
these studies identify causation or simply co-movement of the two sets of variables.4 A
particular aspect of this problem concerns the use of cross-section data on countries that may be
very diverse, raising the possibility that the empirical findings could be distorted by
heterogeneity biases-i.e., unmeasured country-specific factors affecting both economic
volatility and private investment.5 I attempt to tackle these problems below.
3. The Empirical Facts
To explore the empirical relation between investment and economic uncertainty, I make
use of a large cross-country time-series data set, comprising 94 developing countries and
spanning the years 1970 to 1995, for a combined total of over 2,200 annual observations. The
3 There are some exceptions, however. For example, Federer (1993) proxies uncertainty by the risk premium
implicit in the term structure of U.S. interest rates, while Driver and Moreton (1991) use the dispersion of growth
and inflation forecasts from a survey of UK forecasting fnns. In this regard, the two papers closest in approach to
the present one are Goldberg (1993), who uses the variability of ARMA-based exchange rate forecasts for the U.S.,
and Price (1995, 1996), who employs a GARCH-based approach to assess growth uncertainty.
4 The main exception is Goldberg (1993), who uses an instrumental-variable estimation procedure.
5For example, economies specialized in labor-intensive agricultural products tend to have lower investment ratios
than manufacturing-based economies, and the relative price structure of the former could be also more volatile that
that of the latter due to their greater exposure to factors like climatic phenomena. This would tend to create the
illusion of a negative cross-country relationship between investmnent and volatility, even if they were in fact
completely unrelated in a causal sense.
8
cornerstone of the data is the information on private fixed investment drawn from the World
Bank databases, which is constructed by subtracting from total fixed investment the part
attributable to the public sector. Whenever the data allow, the latter is defined as inclusive of
local governments and public enterprise investment, drawing from Jaspersen, Aylward and
Sumlinski (1995); investment by these entities is therefore excluded from private investment.
However, this is not always feasible with the existing information, and hence some unavoidable
degree of heterogeneity across countries in the definition of private and public investment is
likely to remain in the data. Finally, the panel is unbalanced, with some countries lacking
observations for several sample years.
Table 1 presents some basic statistics on constant-price private investment as a ratio to
real GDP, for the full sample as well as its separate time-series and cross-section dimensions.
The table reveals considerable variation in private investment ratios, although the variation is
substantially larger across countries than over time, in a proportion of nearly two to one-a likely
reflection of the persistent contrast between the high private investment levels of some East-
Asian countries and the very low levels of many African countries.6 In both dimensions of the
data, the mean investment ratio exceeds its median value, due to the presence of a few large
values at the upper end of the distribution; nevertheless, the difference between mean and median
is rather small.
3.1 Measuring uncertainty
Using the above investment data, I first examine the association between private
investment and measures of economic uncertainty. Unlike in much of the existing
macroeconomic literature, the explicit aim here is to separate sample variation from uncertainty,
because the former may overstate the latter by including not only truly unpredictable innovations
to the variables of interest, but possibly also (cyclical) movements partly predictable from their
own past.
Rather than selecting a priori the 'right' variable whose volatility is to be examined (as
done in much of the empirical literature), I work with five key macroeconomic variables:
6 The full-sample and cross-section minimum values of the private investment ratio shown in Table 1 correspond to
Chad.
9
inflation, the relative price of capital goods, the growth of output (measured by real GDP), the
real exchange rate and the terms of trade.7 The first three variables are related to the aggregate
profitability of capital: inflation is often taken as a summary measure of the overall
macroeconomic stance, and hence the volatility of its unpredictable component can be viewed as
an indicator of overall macroeconomic uncertainty (e.g., Eberly, 1993). In turn, the relative price
of investment goods (measured here by the fixed investment deflator relative to the GDP
deflator) is closely related to the user cost of capital, and hence the volatility of its innovation can
be viewed as a good indicator of the uncertainty on aggregate investment profitability. Thirdly,
the volatility of output growth is taken here to represent the unpredictability of demand, which
has attracted attention in recent microeconomic work on investment under uncertainty (e.g.,
Guiso and Parigi, 1998).
In turn, the terms of trade and the real exchange rate are related to the relative
profitability of investment in different economic sectors-exportables versus importables, in the
case of the terms of trade, and home-market versus foreign-market oriented activities, in the case
of the real exchange rate. Ceteris paribus, increased volatility of these variables makes price
signals less informative about the relative profitability of investment across sectors, likely
hampering investment decisions.8
For each of these variables, I construct seven sample-based measures of dispersion of
their respective unpredictable innovations; for simplicity, I henceforth refer to these constructs as
'uncertainty measures'. The first one is perhaps the closest conceptually to the notion of
uncertainty, and is given by the conditional variance of the innovation to each of the variables of
interest, constructed using the generalized autoregressive conditional heteroskedasticity
(GARCH) specification of Bollerslev (1986). More specifically, I estimate the following
univariate GARCH(l,1) model:
7Some of these variables are also included in the cross-section analysis of Aizemnan and Marion (1995). However,
none of the other empirical studies mentioned in the text takes into consideration the level or the variability of the
relative price of capital as an investment determinant.
8These variables are of course mutually related. Inflation volatility is typically reflected in relative price variability,
while the volatility of the real exchange rate and the terms of trade should both be reflected in the price of capital
goods if investment has a significant import content (as is the case in most developing countries). As will become
clear below, the variables do in fact contain a good deal of common information.
10
yit = aO + alt +A Yi,-1 + -',; t =:1,...,T; (1)
it = Yio +Y,_1 +1 u aI (2)
where aH denotes the variance of et conditional on information up to period t. For each of the
five variables of interest, I estimate the two-equation model (1)-(2) separately for each country i
(where i = 1,...,94) in the sample.9 I take the fitted cra, from equation (2) as the measure of
uncertainty of yi,.
The other six uncertainty measures are more rudimentary. All are based on the 1-step
ahead forecast errors of two univariate auto-regressive models, using respectively one and two
lags of the dependent variable, plus a constant and a time trend. However, to ensure that the
forecasts use no more information than available at the time that they are formnulated, they are
computed from recursive estimation of the auto-regressions. As with the GARCH estimates, this
is done separately for each country in the data set. More compactly, for each of the five variables
above and each country i, I estimate recursively a modified version of (1):
yi, = ato + a, t + O8Y,6l, +2Y,-2 + t = 1,..., r; (1)
in two variants, one imposing /2 = 0 and the other without this restriction. The superscript T on
the coefficients indicates the changing sample size used in the recursive estimation of (1').
Estimations are performed for r = To,..., T, where To denotes the initial sample set aside to start
the recursive procedure. Hence for each country and variable, the above equation is estimated a
total of T-To times in each of the two specifications. It is easy to see that this involves running a
very large number of regressions; however, unlike with the GARCH model, the regressions are
linear and hence the process is computationally feasible.'°
9 For these regressions, the relative price variables were expressed in logs, while the GDP growth and inflation rates
were measured as first differences of the log of GDP and its deflator, respectively. In the case of the inflation rate,
the trend term in (1) was rarely significant and was therefore dropped from the model. For the GARCH regressions,
I used observations for 1965-95 on the five dependent variables.
10 The initial sample size To was set at 6, and hence the recursive estimation starts in 1970. In practice, this amounts
to some 25 regressions per country for each of the five dependent variables and each of the two auto-regressive
specifications-a total around 25,000 regressions. Clearly, this recursive approach would be computationally
prohibitive in the nonlinear setting of the GARCH model.
11
Using the recursive parameter estimates from (1'), I construct the 'naive' uncertainty
measures as (i) the time-varying standard deviation, and (ii) the mean absolute value, of the one-
step ahead forecast errors. Both statistics are calculated over 3 consecutive years-the current
year plus the preceding two; to test the robustness of the results to this choice of time horizon, in
the case of the AR(1)-based forecast errors I also use a 5-year horizon to compute the standard
deviation and mean absolute deviation of the one-step forecast errors. Notice that in each case
this procedure yields annual observations on the relevant uncertainty measure. Combining the
two dynamic specifications (first- or second-order autoregression) with the selected measure of
variability (standard deviation or mean absolute value), and the 3 or 5-year averaging horizon
(with the latter used only for the AR(1) specification, to preserve sample sizes) of the forecast
errors, I thus obtain a total of six 'naive' uncertainty measures in addition to the GARCH-based
one.
3.2 Uncertainty andprivate investment: the empiricalfacts
The next step is to investigate the empirical association between these uncertainty
measures and private investment performance. Table 2 summarizes the correlation between the
different constructed measures and the private fixed investment/GDP ratio, both in the full
sample and its separate cross-section (i.e., using country averages) and time-series dimensions.
Each column in the table corresponds to one economic variable, and each row to one particular
method, among those described earlier, of measuring the volatility of its unpredictable
component. To limit the potential impact of outliers I use Spearman rank correlations rather than
simple correlations; in fact, however, both turn out to present a very similar picture.
The most striking aspect of Table 2 is the remarkable uniformity of the correlations
between investment and the various forms/measures of uncertainty. In the full sample, they are
all negative and significantly different from zero. They are also of roughly similar magnitude,
not only across the different uncertainty indicators associated with a given variable (i.e., along a
given column in the table-which is hardly surprising since the seven volatility measures try to
capture the same phenomenon),"' but also across the different economic variables (i.e., along
11 For each of the five economic variables (inflation, the terms of trade, the real exchange rate, the relative price of
capital and the GDP growth rate), the seven uncertainty measures are very strongly positively correlated with each
12
each row in the table). Interestingly, the GDP growth-related measures seem to be the least
closely associated with investment, while the opposite is true for the real exchange rate and terms
of trade-based measures.
In turn, the cross section correlations (middle section of Table 2) are again all negative,
larger in absolute value than their full-sample counterparts in every case (with the only exception
of the GARCH-based indicator associated with the relative price of capital), and significantly
different from zero with few exceptions. Interestingly, the cross-section correlations of private
investment with the various real exchange rate-based measures are similar to that reported by
Aizenman and Marion (1996), equal to -.34; however, the cross-section correlation of the
investment ratio with inflation- and terms-of-trade uncertainty in Table 2 is much larger than
found by these authors in their sample."2
Finally, the bottom section of Table 2 shows the time-series correlations, derived from
the 'within' dimension of the data (i.e., deviations from country averages)."3 They remain all
negative with only two exceptions, but their size is generally much smaller (under .1 in absolute
value). Further, most of the indicators associated with the terms of trade, and all of those
associated with GDP growth, cease to show a significant association with the private investment
ratio. By contrast, the uncertainty indicators associated with the real exchange rate and the
relative price of capital continue to be negatively related with investment.
The similarity in sign, and even magnitude, of the pairwise correlations between
investment and the volatility of the unpredictable component of each of the five variables under
consideration raises the natural question of the extent to which the latter contain common
information. Table 3 provides an answer. It shows the full-sample correlation between the
volatility of the different economic variables, using two of the seven volatility measures
developed earlier (the others, not reported to save space, offer a similar picture). The general
other, in both the cross-section and the time-series dimensions-although more strongly so in the former (where the
correlations average around 0.8) than in the latter dimension (where they average around 0.6).
12 Aizemnan and Marion (1996) find no significant cross-section association between the volatility of these two
variables and the private investment ratio.
13 The table presents contemporaneous correlations. However, a very similar picture emerges if the uncertainty
indicators are instead lagged one year.
13
message from Table 3 is that the volatilities of the different variables are all positively correlated
with each other, and very significantly so. Interestingly, inflation volatility seems particularly
strongly associated with the volatility of the remaining variables.
In view of this fact, I finally construct a summary measure of macroeconomic uncertainty
by extracting the first principal component of the five GARCH-based volatility measures
(obviously, the same could be done with the other 'naive' measures). The full-sample rank
correlation of the resulting summary indicator of uncertainty with the private investment ratio is -
.295 (standard error of .025), while its cross-section and time-series counterparts are -.324 (.115)
and -.087 (.025) respectively, all different from zero at any reasonable confidence level.14
In summary, while the various uncertainty measures presented in this section are
admittedly crude-they follow from simple univariate processes-the above results are on the
whole strongly supportive of the negative association between uncertainty and private
investment, and agree with those recently reported for example by Aizemnan and Marion (1996)
using a cross-section of countries. As shown above, however, the negative association remains
present in the time-series dimension of the data, at least for real exchange rate and price of
capital volatility.
In any case, these results refer only to pairwise correlations. They do not clarify whether
economic uncertainty has a direct impact on investment after controlling for standard investment
determinants, or just an indirect one operating through the latter.'5 I turn to this issue in the next
section.
4. Econometric Estimation
To address this issue, I next specify an empirical investment equation and estimate it
using the data just described. I use a simple empirical specification relating the (log of) real
private investment to a set of conventional determinants, to which I add uncertainty measures
introduced in the previous section. Rather than throwing away potentially valuable information
14 The first principal component thus constructed accounts for over 50 percent of the overall variance of the five
individual indicators. The largest weights correspond to the conditional variances of innovations to the real
exchange rate and the relative price of capital goods.
14
by averaging the annual data at an arbitrary phase length (e.g., five or ten year averages), I
choose to work with the yearly information. This means that inertia, very likely present in the
annual data, needs to be taken into consideration. Hence the model to be estimated is of the
form:
I= f(Ii,,1iXi ,agt) + Uj, (3)
where I is the log of private fixed investment at constant prices, x is a set of standard private
investment determinants, a- is a set of uncertainty indicators, u is a random disturbance, and the
subscripts i=l,..., N and t = 1,...,T respectively refer to the cross-section and time-series
dimension of the data.
Among the x variables, I include the current and lagged levels of (the log of) real GDP, to
capture the conventional accelerator effect, and variables measuring the user cost of capital. The
latter include the relative price of capital goods, measured by the (log of the) ratio of the
investment deflator to the GDP deflator, and the real interest rate, measured as Inf(l +i)/(l +z)],
where i denotes the nominal interest rate and 7Tis the GDP inflation rate;`6 both regressors should
exert a negative effect on investment. However, in view of the pervasive role of interest rate
controls and non-price rationing mechanisms in developing-country financial markets-that may
render observed interest rates uninformative as to the true marginal cost of funds-I also add to
the specification a measure of the overall tightness of credit markets, namely the flow of private
credit relative to nominal GDP;'7 for this variable one should expect to find a positive effect on
private investment.
15 For example, Ramey and Ramey (1995) and Aizenman and Marion (1996) find that measures of volatility have a
detrimental effect on output growth. Through the conventional accelerator effect, this would in turn translate into a
negative investment impact.
16 Two alternative measures of the inflation rate were used to construct the real interest rate: (i) the current inflation
rate, and (ii) a simple average of the current plus one-period ahead inflation rates. The empirical results were
similar in both cases; those reported below use the second of these definitions. In both cases, a few outlying
observations showing real interest rates in excess of 50 percent (in absolute value) were discarded from the sample,
as they raised suspicions as to the accuracy of the raw data.
17 Alternative specifications used instead the (log of the) real private credit stock or its first difference, with
qualitative results similar to those reported below.
15
To these conventional investment determinants I add the measures of macroeconomic
uncertainty described earlier. Of the seven alternative sets of measures, I choose the conditional
variances of the innovations to the five variables of interest as obtained from the GARCH
procedure, because they are closest in spirit to the notion of uncertainty. I report empirical
experiments using all five individual measures simultaneously as well as experiments using only
their first principal component. However, some empirical experiments using instead the 'naive'
measures of dispersion of the forecast errors derived from the univariate auto-regressive
estimates presented earlier yielded qualitatively similar results.
4.1 Methodological issues
Empirical implementation of equation (3) using the cross-country time-series data set
summarized earlier raises some methodological issues, which can be easily illustrated using a
linear version of (3):
I1, = ZIJ t-l + Xf,6 + u11, ui, = ai + ej, (3')
where Xit = (x,,', csit') is a (row) vector comprising both the standard regressors and the
uncertainty measure(s), ai denotes a time-invariant country-specific disturbance possibly
correlated with the columns of X, fit is random noise, and the parameters of interest are A and the
(column) vector,l/.
Equation (3') poses a dynamic error-components model whose estimation has been
extensively discussed in the literature. Even in the absence of any other complications, it is well
known that simple-minded estimators such as fixed effects (in panels of short time dimension)
and OLS are inconsistent. The present case poses the added complication of simultaneity, as
some or all of the columns of Xit may be jointly determined with investment. This is particularly
likely for real GDP, but may apply as well to the cost-of-capital variables and the uncertainty
measure(s).
The standard approach to estimation of (3') involves first-differencing the equation to
remove the time-invariant disturbance:
MiA = iAit-I + AiX , + (eit - Eit-1) (4)
16
Estimation of (4) requires an instrumental variable procedure to correct for the
endogeneity of the columns of X as well as the correlation between the lagged difference of the
dependent variable and sit-1 While strictly exogenous instruments are in general hard to come
by, it is often possible to construct 'internal' predetermined instruments using lagged values of
the right-hand side variables. In particular, if one is willing to assume that E[eit I YiS] = 0 for t >
s (but not otherwise), then second and higher-order lags of the columns of X can be used as
instruments in the estimation of (4). This condition is likely to hold if Sit is serially uncorrelated,
in which case the second- and higher-order lags of the endogenous variable are likewise valid
instruments. Using these internal instruments, a GMM estimator can be constructed; see
Arellano and Bond (1991).8 Below I use this approach to compute what I label the difference
estimator.19
This approach, however, has drawbacks. First, differencing the equation removes the
long-run (cross-country) information present in the levels of the variables. Second, if the
columns of X display persistence over time, their lagged levels will be poor instruments for their
differences.
Under additional assumptions, it is possible to construct an alternative GMM estimator
that overcomes these problems. Specifically, if one is willing to adopt the 'stationarity'
assumption that E[ai I YQ = E[aj I Xj,] and E[ai I Iit] = E[ai Iis] for all t and s, then the original
specification in levels (3') can be estimated without need of any transformation, using suitably
lagged differences of the dependent and independent variables as instruments-an almost exactly
symmetric case with the use of lagged levels of the variables as instruments for the first-
difference specification (4) above. Specifically, if - is serially uncorrelated, once-lagged
differences of the regressors are valid instruments. Combining the level and first-difference
specifications, one can construct a system GMM estimator that outperforms the difference
estimator; see Arellano and Bover (1995) and Blundell and Bond (1997). Finally, as long as the
18 Provided the time dimension of the data is long enough, this approach can easily accommodate the case of
serially correlated ti,: witi autocorrelation of order k, regressors lagged k+2 periods become valid instrments.
19 It should be noted, however, that if T is sufficiently large the GMM estimator of A is biased, although less so than
the within estimator as long as T < N (as will be the case here). For a thorough discussion see Alvarez and Arellano
(1998).
17
model is overidentified,20 validity of the assumptions underlying both the difference and the
system estimators can be tested through Sargan tests of orthogonality between the instruments
(i.e., the lagged levels and/or differences of the regressors) and the residuals, and through tests of
second- or higher-order residual autocorrelation (Arellano and Bond, 1991; Blundell and Bond,
1997).21
So far I have ignored the possibility of parameter heterogeneity. However, if the
parameters differ across countries, so that instead of (3') the true relation is of the form
Iit = iIi, t-i + xit/Ji + uit, then pooled estimation imposes the invalid equality restrictions fli =
,8j and Ai = Aj for all i and], and will lead to inconsistent estimates. Specifically, under plausible
conditions the pooled estimate of A converges asymptotically to 1, overstating the true degree of
persistence, while the pooled estimate of ,8 converges to zero, understating the impact of the
remaining regressors (Robertson and Symons, 1992).
In these conditions, and provided the time dimension of the data is large enough, an
alternative is to estimate the model separately for each country and compute an average of the
individual-country estimates (Pesaran and Smith, 1995).22 Thus, to examine the robustness of the
pooled estimation results, below I follow this approach to compute individual-country OLS and
2SLS estimates.
4.2 Empirical implementation
Data limitations are severe and deserve explicit mention. Even for the simple empirical
specification above, the samples for which all of the regressors are available are substantially
smaller than those underlying the pairwise correlations reported in Section 3 above. Further, for
the instrumental-variable regressions below, each country in the regression sample must possess
20 This is almost invariably the case in a panel of at least moderate time dimension, in which a large number of lags
of the regressors are typically available as instruments. In fact, use of all available instruments (as suggested, for
example, by Anh and Schmidt 1995) leads to severe small-sample downward bias of the GMM estimator and its
estimated standard errors; see Ziliak (1997) and Altonji and Segal (1994). For this reason, the GMM estimates
below use only a subset of the available instruments.
21 Notice that if the original sit are serially uncorrelated, differencing will induce first-order (but no higher-order)
serial correlation of the transformed disturbance in (4).
22 See Baltagi and Griffin (1997) for a recent implementation of this approach in the context of a dynamic model. A
third, intermediate option is to combine the pooled and heterogeneous estimates; see Maddala, et al. (1997).
18
a sufficient number of observations in the time dimension to allow the construction of
instruments for the endogenous variables using their lagged values, along the lines described
earlier. I set this required number of annual observations at 10, which yields an unbalanced
panel consisting of 60 countries and a total of 876 observations. The number of observations per
country ranges from the required minimum of 10 to a maximum of 23.
Using this sample and the empirical specification described above, Table 4 presents
various pooled estimates of the basic model using all five uncertainty indicators
simultaneously.23 The first column reports OLS estimates, which ignore the potential
endogeneity of the regressors as well as the possible presence of country-specific effects. The
estimated coefficient on the lagged dependent variable is close to 1, while the remaining
coefficients are in general very small; both results accord with theoretical predictions.
Nevertheless, the signs of the coefficients on the standard investrnent determinants appear
reasonable: GDP growth and credit availability have a positive effect on investment, while the
opposite applies to the relative price of capital and the real interest rate. Interestingly, the
estimates reveal no long-run effect of the level of real GDP on investment. In turn, neither the
credit measure nor the real interest rate carry significant coefficients. Among the uncertainty
indicators, only that associated with the real exchange rate carries a significant coefficient, of
negative sign.
The second column reports fixed-effects estimates. As noted earlier, in short panels they
are inconsistent due to the presence of a lagged dependent variable, although the severity of the
problem in a heavily unbalanced panel such as the present one is unclear. In any case, their sign
pattern is very similar to that of the OLS estimates, but the coefficient on the lagged dependent
variable is considerably smaller, and the private credit flow now is found to have a significant
positive impact on investment. The estimated coefficients of the uncertainty indicators do not
display any substantive changes.
The third and fourth columns of Table 4 report the difference and system GMM
estimates, respectively, that attempt to correct for both endogeneity and unobserved country-
specific effects. The regressions assume that all the explanatory variables are endogenous, and in
23 All specifications include a set of year dummies, which in general were highly significant.
19
consequence they are all instrumented. For the standard investment determinants, I use their
lagged values as instruments. However, this is not possible with the GARCH-based uncertainty
indicators, because their construction uses future as well as past information; for these, I use as
instruments lagged values of the other 'naYve' uncertainty indicators, which as explained earlier
are constructed from the recursive estimates of the univariate auto-regressive processes and
therefore are based on current and past infornation only. In particular, I use the 3-year standard
deviation of the forecast errors derived from the AR(1) specification. Finally, as additional
instrument I use the log terms of trade.24
The difference-GMM estimates in column 3 again show a sign pattern mostly similar to
that already described, but their magnitude is in most cases considerably larger. Precision is
poor, however, and among the standard investment determinants only the relative price of capital
carries a significant (negative) coefficient. Among the uncertainty indicators, growth volatility
now has a marginally significant negative coefficient (of quite large magnitude), while real
exchange rate volatility retains its negative and significant parameter estimate. The Sargan test
provides no evidence of misspecification, while the serial correlation tests point to first- but no
second-order autocorrelation of the residuals, in accordance with the assumptions underlying the
selection of instruments.
Unlike the estimates in columns 2 and 3, which only make use of the time-series
dimension of the data, the system-GMM estimates in column 4 exploit the cross-section
dimension as well. On the whole, the system estimates are quite precise. All of the conventional
investment determinants carry significant coefficients of the anticipated sign, and their
magnitude is in most cases larger than obtained with the preceding estimation methods.
Interestingly, the system estimates imply a long-run output elasticity of private investment close
to unity (a Wald test shows that the coefficients on current and lagged GDP are significantly
different from each other at the 5 percent level). In turn, the implied long-run elasticity with
respect to the price of capital is of similar magnitude but opposite sign. Another remarkable fact
is that both the credit variable and the real interest rate carry significant coefficients, suggesting
24 Specifically, the difference estirnator in column 3 uses as instruments the second and third lags of the standard
regressors and the 'naive' uncertainty indicators cited in the text, plus the current and three lagged values of the log
terms of trade (all expressed in levels). The system estimator in column 4 adds as instruments for the level
regression the lagged first difference of all these variables.
20
the simultaneous presence of price and quantity effects on investment stemming from financial
markets. Regarding the five uncertainty indicators, four of them carry negative coefficients,
although only those on inflation and real exchange rate volatility are significantly different from
zero at conventional levels. Finally, the diagnostic tests (Sargan and second-order
autocorrelation) reveal no evidence against the validity of the instruments used by the system-
GMM estimator.
Overall, these results show a robust negative impact of real exchange rate uncertainty on
private investment, and less clear-cut effects of the remaining uncertainty indicators. While this
is in broad agreement with results reported by some earlier studies (e.g., Serven and Solimano,
1993; Pindyck and Solimano, 1993; Darby, et al, 1998), it might partly reflect the already-
mentioned fact that all five indicators share a good deal of common information, which suggests
an alternative empirical approach based on a summary measure of uncertainty encompassing the
unpredictable components of all five macroeconomic variables. To explore this track, Table 5
presents pooled estimates of an alternative specification using the first principal component of
the five GARCH-based uncertainty indicators as the summary statistic of macroeconomic
uncertainty.
A quick column-by-column comparison of Tables 4 and 5 reveals very little difference
concerning the estimated coefficients of the conventional regressors for each of the four
specifications. The only exception is the real interest rate, whose system-GMM estimated
coefficient (column 4) is now insignificantly different from zero, in contrast with the last column
of Table 4. As for the summary uncertainty indicator, it carries a negative coefficient regardless
of estimation method, and significantly different from zero in every case except for the OLS
estimate. As before, the diagnostic statistics are, on the whole, supportive of the difference and
system-GMM estimates.
Among the alternative econometric specifications reported in Tables 4 and 5, the GMM
estimates in columns 3 and 4 of the tables are superior on theoretical grounds in view of their
ability to correct for both endogeneity and unobserved country-specific effects. On the basis of
such estimates, the unambiguous conclusion is that uncertainty has a significant negative impact
on private investment.
21
This conclusion, however, is based on the heroic assumption of identical parameters
across countries, which underlies the pooled estimates discussed so far. Thus, in the remainder
of this section I examine the robustness of the above results to allowing for parameter
heterogeneity across countries. To do this, I estimate equation (3') above separately for each
country. Unlike the standard panel context, which typically assumes a fixed time dimension and
a large cross-section dimension, this procedure requires a large value of T, which is unavailable
for many of the sample countries, and thus I restrict the analysis to those countries possessing a
sufficient number of annual observations. Specifically, in order to rerun the regressions in
Tables 4 and 5 at the individual-country level with no less than 5 degrees of freedom in every
case, I set the required number of annual observations at 18. This leaves a total of 30 countries
and 661 data points.25
Table 6 reports the averages of individual-country estimates thus obtained.
Symmetrically with the time effects included in the pooled estimation, the individual estimates
include a time trend, reported at the bottom of the table. To take into account the varying sample
size across countries, the averages are computed weighting the individual-country estimates by
their respective number of observations. Standard errors are in turn constructed under the
(admittedly unrealistic) assumption that the individual-country estimates are independent.
Columns 1 and 2 in Table 6 respectively present average OLS and 2SLS estimates of the
specification including the five uncertainty indicators (i.e., the specification employed in Table
4).26 While the individual-country estimates (not reported to save space) are fairly disperse, the
pattern that emerges appears sensible. Two striking results are the considerably lower degree of
persistence-as measured by the coefficient of the lagged dependent variable-and the much
larger magnitude of several coefficients, both relative to the pooled estimates presented earlier.
Concerning the standard investment determinants, the qualitative pattern of the average OLS
estimates is similar to that obtained with the pooled GMM estimates-positive effects of real
output and credit availability (although the latter is not significant), and negative effects of the
25 Pooled estimates computed on this reduced sample are very similar to those reported in Tables 4 and 5.
26 The 2SLS estimates in columns 2 and 4 of Table 6 use as instruments the current and lagged value of the log
terns of trade, plus the first lags of all the regressors, and the first lags the real exchange rate, the relative price of
capital, the inflation rate and the 'naYve' uncertainty indicators used as instruments in the pooled GMM regressions.
22
real interest rate and the price of capital goods-but the implied long-run output elasticity of
private investment is now well above unity, and lagged GDP ceases to be significant. In turn, the
uncertainty variables carry large coefficients (in absolute value), which are significant and
negative in the cases of real exchange rate and price of capital volatility. The bottom of the table
reports the Wald test of equality across countries of all parameters except constant and trend
(which amounts to allowing for unrestricted country and time effects); the pooling restrictions
are overwhelmingly rejected.
Column 2 presents average 2SLS estimates of the same specification. The qualitative
pattern is similar, although precision is extremely poor-a reflection of overparameterization,
small sample sizes and/or poor instrument quality-and only two of the parameter estimates are
significant at the 10 percent level. The pooling restrictions cannot be rejected in this case,
although in view of the lack of precision of the estimates the result is not particularly
encouraging.
Columns 3 and 4 turn to the condensed specification using the summary measure of
uncertainty given by the first principal component of the five individual uncertainty indicators in
columns 1 and 2. The average OLS results in column 3 are quite similar to those in column 1,
although the coefficient of the lagged dependent variable nearly doubles in size and is much
closer to the GMM pooled estimate. The parameter estimates of the conventional investment
determinants continue to follow the same sign pattern, and they are all significant except for the
credit variable. Most importantly, the summary uncertainty measure carries a large negative
coefficient, significant at the 5 percent level. Finally, column 4 presents the average 2SLS
estimates. They are very similar in sign, magnitude and statistical significance to the average
OLS results in column 3, with the only exception of the uncertainty indicator itself, whose
coefficient becomes considerably larger in absolute value while remaining significantly negative.
Finally, as shown at the bottom of the table, the pooling restrictions are clearly rejected for both
the OLS and 2SLS estimates.
To summarize the empirical experiments in this section, the econometric results
consistently reveal a significant negative effect on private investment of measures of
macroeconomic uncertainty, after controlling for other standard investment determinants.
Moreover, this finding holds when controlling for country-specific effects, simultaneity, and/or
23
parameter heterogeneity (or all simultaneously). Indeed, the only exception are the pooled OLS
estimates in the first column of Table 5, where the uncertainty measure carries a negative but
insignificant coefficient. On the other hand, there are considerable differences across
specifications concerning the estimated magnitude of the impact of uncertainty on investment. In
general, the pooled GMM estimates and, especially, the heterogeneous estimates yield the largest
effects.
On the whole, the various estimates reported here also provide a consistent picture
concerning the more conventional investment determinants. Private investment is found to be
negatively affected by real interest rates and the relative price of capital goods, and positively by
real output and credit availability. As before, the specifics regarding size and significance of the
coefficients vary across estimators, in some cases considerably so. For example, the degree of
persistence appears grossly overstated by the pooled OLS estimates, and is much more modest in
the heterogeneous estimates; both results accord with theoretical predictions. In contrast, the
heterogeneous estimates yield the largest long-run output effects. Further, the real interest rate
coefficient is in most cases significantly negative in the heterogeneous-coefficient regressions,
while in the pooled regressions it is usually insignificant. This suggests that diversity in
financial-market arrangements across countries might be the cause of the frequent failure to
detect significant real interest rate effects in developing-country empirical investment equations.
5. Concluding Remarks
The impact of uncertainty on investment has long intrigued economists. Different
theoretical models make opposing predictions about its sign, which is therefore an empirical
matter. In turn, empirical studies based on aggregate data often use potentially misleading
measures of uncertainty, fail to control for other standard investment determinants, and in many
cases are subject to possible heterogeneity and simultaneity biases.
This paper has re-examined empirically the investment-uncertainty link, using a large
macroeconomic data set for developing countries. Rather than focusing on the sample variability
of any one arbitrarily-chosen variable (e.g., inflation or the real exchange rate), the paper
constructs alternative measures of uncertainty based on the dispersion of the innovations to five
key macroeconomic variables: three related to the macroeconomic environment and the
24
aggregate profitability of capital-growth, inflation and the relative price of investment goods-
and another two more closely related to the relative profitability of different economic sectors-
the terms of trade and the real exchange rate.
Using these constructed measures, the paper has explored the stylized facts concerning
the association between private investment and macroeconomic volatility indicators and has
investigated the impact of uncertainty on private investment in a regression framework,
controlling for other standard investment determinants.
Simple correlations reveal a significantly negative association between the constructed
uncertainty measures and private investment. Further, this applies virtually without exception to
each one of the uncertainty indicators on all five economic variables considered in the paper.
The paper has gone one step beyond by adding these uncertainty measures to an
otherwise standard empirical equation for private investment, which was estimated using
alternative econometric procedures variously allowing for simultaneity, country specific effects
and parameter heterogeneity across countries. On the whole, the estimation yields satisfactory
results concerning the standard investment determinants included in the equation: in general, the
estimates show significant output and cost-of-capital effects, with the latter captured by the
relative price of capital, the real interest rate and/or the availability of credit to the private sector.
Interestingly, allowing for parameter heterogeneity across countries generally reveals a
significant negative impact of the real interest rate on private investment, in contrast with the
lack of significance found with most pooled-estimation methods in the paper. Finally, the results
also show that private investment displays significant persistence over time.
Most importantly, the negative association between macroeconomic uncertainty and
investment remains present in these multivariate regressions, pointing to the existence of a direct
impact of uncertainty on private investment over and above any other indirect effects that may be
at work. In particular, real exchange rate volatility is found to have a robust negative effect on
investment regardless of econometric specification, in broad agreement with some earlier studies
based on both micro and macro data. Likewise, a summary measure of macroeconomic
uncertainty encompassing the volatility of the five variables of interest listed above also has a
consistently negative effect on private investment. The fact that these results hold after taking
account of the likely simultaneity between investment and volatility, as well as after controlling
25
for unobserved country-specific effects and even parameter heterogeneity, indicates that the
relationship is causal rather than merely coincidental-higher uncertainty leads to lower private
investment.
While these results are encouraging, some caveats are in order. The measures of
economic uncertainty used in the paper are based on very simple forecasting procedures that
could be improved considerably, for example by employing multivariate-rather than
univariate-forecasting methods making better use of the available information. Also, the
magnitude of the estimated effects of uncertainty varies considerably across specifications, and
becomes much larger when allowing for parameter heterogeneity-an issue that could deserve
further scrutiny. Perhaps most importantly, the paper has not attempted to disentangle the
mechanisms behind the negative effect of uncertainty on investment, which could reflect
investment irreversibility, limited risk diversification possibilities, or other factors. Some of
these questions could be addressed empirically: for example, if lack of access to risk
diversification is the main issue, then the adverse investment impact of uncertainty should be
consistently larger in countries (and periods) with less developed financial systems. These
issues, however, are beyond the scope of this paper.
References
Abel, Andrew B. 1983. "Optimal Investment Under Uncertainty." American Economic Review 73 (March): 228-
233.
, and Janice C. Eberly. 1994. "A Unified Model of Investment Under Uncertainty." The American
Economic Review (December): 1369-84.
1995a. "Optimal Investment With Costly Reversibility." NBER Working Paper No. 5091.
1 1995b. "The Effects of Irreversibility and Uncertainty on Capital Accumulation." NBER Working
Paper No. 5363.
Aizenman, Joshua and Nancy Marion. 1995. "Volatility, Investment and Disappointment Aversion." NBER
Working Paper No. 5386.
_ 1996. "Volatility and the Investment Response." NBER Working Paper No. 5841.
Alvarez, J. and M. Arellano. 1998. "The Time-Series and Cross-Section Asymptotics of Panel Data Estimators."
Unpublished manuscript, CEMFI.
Altonji, Joseph and Lewis Segal. 1994. "Small Sample Bias in GMM Estimation of Covariance Structures." NBER
Working Paper No. 156.
Ahn, Seung and Peter Schmidt. 1995. "Efficient Estimation of Models for Dynamic Panel Data." Journal of
Econometrics 68, 5-27.
Arellano, Manuel and Stephen Bond. 1991. "Some Tests of Specification for Panel Data: Monte Carlo Evidence
and an Application to Employment Equations." Review of Economic Studies 58, 277-97.
Arellano, Manuel and Olympia Bover. 1995. "Another Look at the Instrumental Variable Estimation of Error-
Components Models." Journal of Econometrics 68, 29-51.
Baltagi, B. and J. Griffin. 1997. "Pooled Estimates vs. their Heterogeneous Counterparts in the Context of the
Dynamic Demand for Gasoline." Journal of Econometrics 77, 303-327.
Bar-Ilan, Avner and William C. Strange. 1996. "Investment Lags." American Economic Review 86, pp. 610-622.
Bell, Gregory and Josd M. Campa. 1997. "Irreversible Investments and Volatile Markets: A Study of the Chemical
Processing Industry." The Review of Economics and Statistics. Vol. LXXIX. No. 1. February.
Bleaney, Michael. 1996. "Macroeconomic Stability, Investment and Growth in Developing Countries." Journal of
Development Economics 48, 461-77.
Blundell, R. and S. Bond. 1997. "Initial Conditions and Moment Restrictions in Dynamic Panel Data Models."
University College London Discussion Paper 97-07.
Bollerslev, T. 1986. "Generalized Autoregressive Conditional Heteroskedasticity." Journal of Econometrics 31,
307-327.
Caballero, Ricardo J. 1991. "On the Sign of the Investment-Uncertainty Relationship." American Economic
Review 81, No. 1 (March): 279-88.
_ 1993. "On the Dynamics of Aggregate Investment." In Striving for Growth after Adjustment, The Role
of Capital Formation, eds. Luis Serven and Andres Solimano. World Bank Regional and Sectoral Studies.
, and Robert S. Pindyck. 1996. "Uncertainty, Investment, and Industry Evolution." International
Economic Review 37, No. 3 August.
Darby, J., et al. 1998. "The Impact of Exchange Rate Uncertainty on the Level of Investment." CEPR Discussion
Paper 1896.
27
Dixit, Avinash and Robert S. Pindyck. 1994. Investment Under Uncertainty. Princeton University Press.
Princeton, New Jersey.
Driver, C. and D. Moreton. 1991. "The Influence of Uncertainty on UK Manufacturing Investment." Economic
Journal 101, 1452-1459.
Eberly, J. 1993. "Comment on Pindyck and Solimano." NBER Macroeconomics Annual, 303-312.
Federer, J. 1993. "The Impact of Uncertainty on Aggregate Investment Spending." Journal of Money, Credit and
Banking 25, 30-48.
Ghosal, V. and P. Loungani. 1996. "Product Market Competition and the Impact of Price Uncertainty on
Investment: Some Evidence from US Manufacturing Industries." Board of Governors of the Federal
Reserve System International Finance Discussion Paper 517.
Goldberg, L. 1993. "Exchange Rates and Investment in United States Industry." Review of Economics and
Statistics 75, 575-588.
Guiso, L. and G. Parigi. 1998. "Investment and Demand Uncertainty." Mimeo, Banca d'Italia.
Gul, F. 1991. "A Theory of Disappointment Aversion." Econometrica 59, 667-686.
Hartman, Richard. 1972. "The Effects of Price and Cost Uncertainty on Investment." Journal of Economic Theory
5 (October): 258-266.
Hausmann, Ricardo and Michael Gavin. 1995. "Macroeconomic Volatility in Latin America: Causes,
Consequences and Policies to Assure Stability." Inter-American Development Bank, mimeo.
Jaspersen, Frederick, Anthony H. Aylward and Mariusz A. Sumlinski. 1995. "Trends in Private Investment in
Development Countries, Statistics for 1970-94." International Finance Corporation Discussion Paper No.
28.
Maddala, G., et al. 1997. "Estimation of Short-Run and Long-Run Elasticities of Energy Demand from Panel Data
Using Shrinkage Estimators." Journal of Business and Economic Statistics 15, 90-100.
Pesaran, M. and R. Smith. 1995. "Estimating Long-Run Relations from Dynamic Heterogeneous Panels." Journal
of Econometrics 68, 79-113.
Pindyck, R. and A. Solimano. 1993. "Economic Instability and Aggregate Investment." NBER Macroeconomics
Annual 8, 259-303.
Price, S. 1995. "Aggregate Uncertainty, Capacity Utilization and Manufacturing Investment." Applied Economics
27, 147-54.
1996. "Aggregate Uncertainty, Investment and Asymmetric Adjustment in the UK Manufacturing
Sector." Applied Economics 28, 1369-79.
Ramey, G. and V. Ramey. 1995. "Cross-Country Evidence on the Link between Volatility and Growth."
American Economic Review 1 138-1151.
Robertson, D. and J. Symons. 1992. "Some Strange Properties of Panel Data Estimators." Journal of Applied
Econometrics 7, 175-189.
Serven, L. and A. Solimano. 1993. "Debt Crisis, Adjustment Policies and Capital Formation in Developing
Countries: Where Do We Stand ?" World Development 21, 127-140.
Zeira, J. 1990. "Cost Uncertainty and the Rate of Investment." Journal of Economic Dynamics and Control 14,
53-63.
Ziliak, James. 1997. "Efficient Estimation With Panel Data When Instruments Are Predetermined: An Empirical
Comparison of Moment-Condition Estimators." Journal of Business and Economic Statistics 15, No. 4,
October.
28
Table 1
Table 1
Private Investment/GDP Ratio:
Descriptive Statistics
Sample Full Betweena Withinb
Mean 0.1251 0.1236 0.0000
Median 0.1065 0.1140 -0.0068
Standard deviation 0.0715 0.0583 0.0420
Variance 0.0051 0.0034 0.0018
Minimum 0.0014 0.0172 -0.1550
Maximum 0.4488 0.2865 0.2436
Number of observations 2226 94 2226
a. Country averages.
b. Deviations from country averages.
29
Table 2
Rank Correlation between the Private InvestmentlGDP Ratio and Selected Uncertainty Indicators
Uncertainty Source
Inflation Relative Real Terms of GDP
rate price of exchange trade Growth
capital rate
1. Full-sample correlations
Conditional variance from GARCH(1,1) -0.217 -0.212 -0.240 -0.282 -0.150
3-period standard deviation of AR(l) forecast errors -0.157 -0.137 -0.173 -0.213 -0.123
5-period standard deviation of AR(1) forecast errors -0.161 -0.188 -0.187 -0.233 -0.161
3-period standard deviation of AR(2) forecast errors -0.102 -0.146 -0.142 -0.213 -0.094
3-period MAD of AR(1) forecast errors -0.203 -0.160 -0.168 -0.239 -0.129
5-period MAD of AR(1) forecast errors -0.211 -0.196 -0.182 -0.244 -0.144
3-period MAD of AR(2) forecast errors -0.160 -0.177 -0.172 -0.222 -0.113
2. Cross-section correlations
Conditional variance from GARCH(l, 1) -0.256 -0.196 -0.302 -0.394 -0.196
3-period standard deviation of AR(1) forecast errors -0.295 -0.215 -0.345 -0.294 -0.262
5-period standard deviation of AR(1) forecast errors -0.265 -0.241 -0.324 -0.301 -0.211
3-period standard deviation of AR(2) forecast errors -0.141 -0.216 -0.325 -0.282 -0.141
3-period MAD of AR(1) forecast errors -0.301 -0.216 -0.311 -0.327 -0.221
5-period MAD of AR(l) forecast errors -0.291 -0.213 -0.319 -0.309 -0.216
3-period MAD of AR(2) forecast errors -0.179 -0.205 -0.332 -0.313 -0.167
3. Time-series correlations
Conditional variance from GARCH(l,1) -0.078 -0.105 -0.056 -0.053 -0.034
3-period standard deviation of AR(1) forecast errors -0.080 -0.089 -0.075 -0.022 -0.026
5-period standard deviation of AR(l) forecast errors -0.068 -0.086 -0.096 -0.025 -0.041
3-period standard deviation of AR(2) forecast errors -0.028 -0.097 -0.045 -0.060 0.038
3-period MAD of AR(1) forecast errors -0.083 -0.074 -0.084 -0.012 -0.016
5-period MAD of AR(1) forecast errors -0.073 -0.103 -0.103 -0.017 -0.020
3 -period MAD of AR(2) forecast errors -0.053 -0.064 -0.091 -0.032 0.014
Note: Each entry in the table shows the simple correlation between the private investmentlGDP ratio and one specific
measure of the volatility of the innovation to the variable at the top of the column. The correlations shown in bold are
statistically significant at the 5 percent level.
30
Table 3
Full-Sample Rank Correlation Among Measures of Uncertainty
of Various Economic Variables
a. GARCH(1,1) conditional variances
Variable
Variable Inflation Relative price Real exchange Terms of
rate of capital rate trade
Inflation rate 1.000
Relative price of capital 0.397 1.000 -- --
(0.023) -- -- --
Real exchange rate 0.356 0.294 1.000 --
(0.022) (0.024)
Terms of trade 0.396 0.281 0.275 1.000
(0.023) (0.024) (0.024) --
GDP growth 0.344 0.272 0.247 0.259
(0.022) (0.024) (0.022) (0.024)
b. 3-period standard deviations of AR(1) forecast errors
Variable
Variable Inflation Relative price Real exchange Terms of
rate of capital rate trade
Inflation rate 1.000
Relative price of capital 0.342 1.000 -- --
(0.024) -- -- --
Real exchange rate 0.400 0.247 1.000 --
(0.022) (0.024) -- --
Terms of trade 0.331 0.318 0.198 1.000
(0.024) (0.024) (0.024) --
GDP growth 0.278 0.225 0.220 0.212
(0.022) (0.024) (0.022) (0.024)
Note: Standard errors in brackets.
31
Table 4
Pooled Estimates of the Investment Equation
(dependent variable: log of real private investment)
Equation 1 2 3 4
Estimation method OLS Fixed effects First differences System GMM
GMM
Constant 0.005 -0.001 0.072 -1.841**
(0.005) (0.005) (0.070) (0.455)
Lagged private investmenta 0.868** 0.633** 0.536** 0.566**
(0.053) (0.046) (0.090) (0.075)
Real GDP' 0.091** 0.065** 0.876 1.827**
(0.024) (0.026) (0.756) (0.652)
Real GDP laggeda -0.092** -0.061** -1.002 -1.339**
(0.024) (0.027) (1.047) (0.649)
Relative price of capitala -0.020** -0.040** -0.627** -0.433**
(0.006) (0.010) (0.254) (0.143)
Credit flow to priv. sector/GDP 0.019 0.041* 1.078 3.169**
(0.016) (0.023) (1.029) (0.975)
Real interest rate -0.014 -0.014 -0.194 -0.521**
(0.010) (0.010) (0.334) (0.247)
Inflation uncertaintyb -0.180 -0.118 -0.740 -9.605**
(0.134) (0.091) (5.578) (4.617)
Terms-of-trade uncertaintyb 0.030 -0.024 0.637 -1.955
(0.044) (0.033) (4.643) (2.727)
Real exchange rate uncertaintyb -0.029** -0.059** -2.119** -1.493**
(0.012) (0.018) (0.609) (0.395)
Price of capital uncertaintyb 0.011 0.002 0.104 0.367
(0.008) (0.009) (0.319) (0.249)
GDP growth uncertaintyb 1.205 0.459 -25.855* -19.986
(0.837) (0.576) (15.458) (13.047)
Wald test ofjoint significance (p-value) 0.000 0.000 0.000 0.000
Time effects (p-value) 0.001 0.001 0.002 0.000
Sargan test (p-value) NA NA 0.418 0.851
1st-order autocorrelation (p-value) NA NA 0.002 0.001
2nd-order autocorrelation (p-value) NA NA 0.187 0.256
Number of observations 876 816 816 816
Notes: All regressions include a set of year dummies. Standard errors (in brackets) are heteroskedasticity
consistent. One (*) and two (**) stars denote statistical significance at the 10 and 5 percent level, respectively.
a. Expressed in logs.
b. Measured by the conditional variance from GARCH (1,1) estimates.
32
Table 5
Pooled Estimates of the Investment Equation
(dependent variable: log of real private investment)
Equation 1 2 3 4
Estimation method OLS Fixed effects First differences System GMM
GMM
Constant 0.007 -0.001 0.108** 1.806**
(0.005) (0.005) (0.052) (0.299)
Lagged private investmenta 0.880** 0.636** 0.590** 0.666**
(0.053) (0.046) (0.073) (0.051)
Real GDP' 0.086** 0.063** 0.604 1.778**
(0.024) (0.027) (0.428) (0.335)
Real GDP laggeda -0.087** -0.061** -1.072* -1.377**
(0.024) (0.027) (0.585) (0.353)
Relative price of capitala -0.016** -0.044** -0.682** -0.491**
(0.005) (0.011) (0.201) (0.088)
Credit flow to priv. sector/GDP 0.019 0.050** 0.874 2.931**
(0.017) (0.025) (0.612) (0.640)
Real interest rate -0.011 -0.008 -0.242 -0.019
(0.009) (0.011) (0.267) (0.161)
Uncertaintyb -0.053 -0.152** -2.719** -1.513*
(0.041) (0.060) (1.054) (0.778)
Wald test of joint significance (p-value) 0.000 0.000 0.000 0.000
Time effects (p-value) 0.000 0.001 0.013 0.000
Sargan test (p-value) NA NA 0.388 0.204
1st-order autocorrelation (p-value) NA NA 0.002 0.000
2nd-order autocorrelation (p-value) NA NA 0.443 0.655
Number of observations 876 816 816 816
Notes: All regressions include a set of year dummies. Standard errors (in brackets) are heteroskedasticity
consistent. One (*) and two (**) stars denote statistical significance at the 10 and 5 percent level, respectively.
a. Expressed in logs.
b. First principal component of the conditional variances of inflation, the terms of trade, the real exchange rate, the
relative price of capital, and the GDP growth rate, each obtained from a univariate GARCH(I, 1) model.
33
Table 6
Heterogeneous Estimates of the Investment Equation
(dependent variable: log of real private investment)
Equation 1 2 3 4
Estimation Method Average OLS Average 2SLS Average OLS Average 2SLS
Constant 13.178** -21.805 -13.930** -13.260**
(5.270) (23.785) (3,073) (4.143)
Lagged private investmenta 0.153** 0.047 0.298** 0.292**
(0.060) (0.262) (0.046) (0.058)
Real GDP" 1.874** 2.134* 1.805** 1.825**
(0.365) (1.237) (0.266) (0.375)
Real GDP laggeda -0.318 0.166 -0.205 -0.261
(0.410) (1.171) (0.265) (0.350)
Relative price of capitala -0.506** -0.508 -0.339** -0.361 **
(0.139) (0.463) (0.100) (0.142)
Credit flow to private sector/GDP 0.835 1.641 0.812 1.222
(0.744) (1.995) (0.603) (0.834)
Real interest rate -0.711** -1.472 -0.530** -0.533**
(0.284) (1.509) (0.191) (0.271)
Inflation uncertaintyb 31.440 120.998
(27.269) (138.850)
Terns-of-trade uncertaintyb -7.357 4.830
(15.088) (32.884)
Real exchange rate uncertaintyb -10.538** -19.604*
(3.963) (11.600)
Price of capital uncertaintyb -29.747** -40.835
(11.051) (25.564)
GDP growth uncertaintyb 5.215 18.741
(40.243) (139.515)
Uncertainty0 -15.509** -24.120**
(6.807) (11.515)
Time trend -0.005 -0.010 -0.020** -0.018
(0.014) (0.036) (0.008) (0.011)
Test of pooling restrictions (p-value) 0.0013 0.9997 0.0000 0.0001
Number of observations 661 661 661 661
Notes: One (*) and two (**) stars denote statistical significance at the 10 and 5 percent level, respectively.
a. Expressed in logs.
b. Measured by the conditional variance from GARCH (1,1) estimates.
c. First principal component of the conditional variances of inflation, the terms of trade, the real exchange rate, the
relative price of capital, and the GDP growth rate, each obtained from a univariate GARCH(1,1) model.
Policy Research Working Paper Series
Contact
Title Author Date for paper
WPS2019 The Effects on Developing Countries A. Denny Ellerman December 1998 T. Tourouqui
of the Kyoto Protocol and Carbon Henry D. Jacoby 87431
Dioxide Emissions Trading Annelene Decaux
WPS2020 Trade Liberalization, Fiscal Delfin S. Go December 1998 C. Bernardo
Adjustment, and Exchange Rate Pradeep Mitra 31148
Policy in India
WPS2021 Balance, Accountability, and Anwar Shah December 1998 A. Shah
Responsiveness: Lessons about 37687
Decentralization
WPS2022 The Implications of Foreign Aid Shantayanan Devarajan December 1998 C. Bernardo
Fungibility for Development Vinaya Swaroop 31148
Assistance
WPS2023 Capital Inflow Reversals, Banking Samuel Talley December 1998 M. Geller
Stability, and Prudential Regulation Marcelo M. Giugale 85155
in Central and Eastern Europe Rossana Polastri
WPS2024 Re-Engineering Insurance Lawrie Savage December 1998 H. Arbi
Supervision 34663
WPS2025 Dynamic Capital Mobility, Capital Hong G. Min December 1998 K. Labrie
Market Risk, and Exchange Rate 31001
Misalignment: Evidence from Seven
Asian Countries
WPS2026 Circuit Theory of Finance and the Biagio Bossone December 1998 B. Bossone
Role of Incentives in Financial 33021
Sector Reform
WPS2027 Child Labor: Cause, Consequence, Kaushik Basu December 1998 M. Mason
and Cure, with Remarks on 30809
International Labor Standards
WPS2028 Access to Markets and the Benefits Hanan G. Jacoby December 1998 M. Fernandez
of Rural Roads 33 766
WPS2029 Small Manufacturing Plants, Susmita Dasgupta December 1998 Y. D'Souza
Pollution, and Poverty: New Robert E. B. Lucas 31449
Evidence from Brazil and Mexico David Wheeler
WPS2030 The Political Economy of Financial Cevdet Denizer December 1998 T. Hailu
Repression in Transition Economies Raj M. Desai 44396
Nikolay Gueorguiev
WPS2031 Addressing the Education Puzzle: Ram6n L6pez December 1998 T. Shie~
The Distribution of Education and Vinod Thomas 3e317
Economic Reform Yan Wang
Policy Research Working Paper Series
Contact
Title Author Date for paper
WPS2032 Institutional Investors and Securities Dimitri Vittas December 1998 A. Yaptenco
Markets: Which Comes First? 38526
WPS2033 Diversification and Efficiency of Stijn Claessens December 1998 R. Vo
Investment by East Asian Simeon Djankov 33722
Corporations Joseph P. H. Fan
Harry H. P. Lang
WPS2034 Information, Accounting, and the Phil Burns December 1998 G. Chenet-Smith
Regulation of Concessioned Antonio Estache 36370
Infrastructure Monopolies